Top Banner
The Visible Hand December 2015, Volume XXIV, Number I Spencer Koo Princeton University Comparing Assimilation and Success Rate of Legal First Generation Asian and Hispanic Immigrants in the United States Victor Ghazal Grinnell College CEO Duality and Corporate Stewardship: Evidence from Takeovers Ziyi Yan Bryn Mawr College The Effect of Driving Restrictions on Air Quality in Beijing Sylvia Klosin and Cameron Taylor University of Chicago Parental Employment and Childhood Obesity Damilare Aboaba Cornell University Preserving Financial Stability: Capital Controls in Developing Countries during Times of Finacial Crisis Alex Foster and Adam Sudit University of Chicago Veteran Rehabilitation: A Panel Data Study of the American Civil War Michael Berton University of California - Santa Bar- bara The Origins of the Federal Reserve Wendy Morrison University of Virginia Optimal Liquidity Regulation Given Heterogeneous Risk Preferences and Retrade
95

Visible Hand

Apr 14, 2017

Download

Documents

Wendy Morrison
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Visible Hand

Fall 2015 | The Visible Hand | 1

Volume xxiV. issue i.

The Visible HandDecember 2015, Volume XXIV, Number I

Spencer Koo Princeton University Comparing Assimilation and Success Rate of Legal First Generation Asian and Hispanic Immigrants in the United States

Victor Ghazal Grinnell College CEO Duality and Corporate Stewardship: Evidence from Takeovers

Ziyi Yan Bryn Mawr College The Effect of Driving Restrictions on Air Quality in Beijing

Sylvia Klosin and Cameron Taylor University of Chicago Parental Employment and Childhood Obesity

Damilare Aboaba Cornell University Preserving Financial Stability: Capital Controls in Developing Countries during Times of Finacial Crisis

Alex Foster and Adam Sudit University of Chicago Veteran Rehabilitation: A Panel Data Study of the American Civil War

Michael Berton University of California - Santa Bar-bara The Origins of the Federal Reserve

Wendy Morrison University of Virginia Optimal Liquidity Regulation Given Heterogeneous Risk Preferences and Retrade

Page 2: Visible Hand

2 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

The Visible HandISSN: 1559-8802

Editor-in-Chief:

Nivedita Vatsa

Executive Board:

Kristina HurleyJames O'Connor

Anthony MohammedAlina Dvorovenko

Nabeel Momin

Editors and Referees:

Aleksandre NatchkebiaAnthony Mohammed

Aya AbuosbehCharles Anyamene

Christian CovingtonDan Liu

Daniel AboroahaEdward Chen

Henry Marshall

Jack EllrodtJohn IndergaardJoshua MensahMargaret Wong

Monica CaiRishu Jain

Todd LensmanYasmeen Mahayni

Wendy Li

© 2015 Economics Society at Cornell. All Rights Reserved.The opinions expressed herein and the format of citations are those of the authors and do not represent the view or

the endorsement of Cornell University and its Economics Society.

The Visible Hand thanks:

Jennifer P. Wissink, Senior Lecturer and Faculty Advisor, for her valuable guidance and kind supervi-

sionThe Student Assembly Finance Commissionfor their generous continued financial support.

Table of Contents

3 Editorial Nivedita Vatsa4 Comparing Assimilation and Success Rate of Legal First Generation Asian and Hispanic Immigrants in the U.S. Spencer Koo13 CEO Duality and Corporate Stewardship: Evidence from Takeovers Victor Ghazal22 The Effect of Driving Restrictions on Air Quality in Beijing Ziyi Yan30 Parental Employment and Childhood Obesity Sylvia Klosin and Cameron Taylor 42 Preserving Financial Stability: Capital Controls in Developing Countries during Times of Financial Crisis Damilare Aboaba58 Value-at-Risk: The Effect of Autoregression in a Quantile Process Khizar Qureshi68 Veteran Rehabilitation: A Panel Data Study of the American Civil War Alex Foster & Adam Sudit79 The Origins of the Federal Reserve Michael Berton79 Optimal Liquidity Regulation Given Heterogeneus Risk Preferences and Retrade Wendy Morrison

Issues of The Visible Hand are archived at http://rso.cornell.edu/ces/publications.html

The Visible Hand is published each fall and spring with complimentary copies avaiable

around the Cornell campus. We welcome your letters to the editor and comments! Please

direct correspondence to the Editor-in-Chief at

Cornell Economics Society

Department of Economics, Uris Hall, 4th Floor Cornell University

Ithaca, NY 14853 or

[email protected]

Page 3: Visible Hand

Fall 2015 | The Visible Hand | 3

Volume xxiV. issue i.

Over the past few years, the effects of globalization and the linkages between various international economic and political events have only grown stronger. The negative interest rates set by the European Central Bank have sent waves through the global economy, while the positive and negative effects of an oversupply of crude oil were seen around world. Almost no subject in economics can be analyzed in geographic isolation anymore. This issue of The Visible Hand hopes to provide readers with an appreciation for the complexity of contemporary economic issues. As this issue of our journal goes to print, leaders from 150 countries are meeting in Paris for the 21st Conference of Parties on global climate change. The challenge of reducing greenhouse gas emissions weighs on us now more than ever and Ziyi Yan’s work in “The Effect of Driving Restrictions on Air Quality in Beijing,” addresses the way in which environmental change, enforced on an institutional level, can be effective in lowering carbon emissions. In the run-up to the U.S. primary elections, the topics of immigration and the treatment of immigrants are receiving attention from both presidential candidates and social justice activists. The subject of immigration not only shapes the contemporary discussion of economic development, but also the discussion of economic inequality. In “Comparing Assimilation & Success Rates of Legal First Generation Asian and Hispanic Immigrants in the United States,” Spencer Koo looks into various economic factors such as education and their effect on the wage gap of first-generation immigrants from different ethnicities. This research underscores the importance of identifying policies that ensure progress for all social groups. The recent recession in Brazil and economic slowdown in China have cast a spotlight on the need to find ways of stabilizing economies. In December 2012, the IMF released a statement saying, “In certain circumstances, capital flow management measures can be useful. They should not, however, substitute for warranted macroeconomic adjustment.” A calibrated approach to capital controls has been seen to lead to financial stability, whereas tight controls have been linked to illegal transactions. Therefore, Damilare Aboaba’s “Preserving Financial Stability: Capital Controls

in Developing Countries during Times of Financial Crisis,” comes at an appropriate time as it explores the ways in which the setup of capital controls in developing economies influences their financial stability. The separation of the roles of chairman and CEO roles in corporate leadership has been the subject of debate for decades. Victor Ghazal, in his paper, “CEO Duality and Corporate Stewardship: Evidence from Takeovers,” offers a new take on the much-debated question by studying the nature of negotiation and “aggressiveness” in dual CEOs. The subject of children’s health, particularly the growing obesity epidemic, continues to be an important topic in public debate. Sylvia Klosin and Cameron Taylor explain various causal mechanisms that affect the weight of children in their work, titled “Parental Employment and Childhood Obesity.” Khizar Qureshi’s “Value-at-Risk:The Effect of Autoregression in a Quantile Process,” is more specific study, demonstrating a conditional autoregressive value at risk model. Readers curious about the intersection between finance and mathematics would be particularly interested in this research as it shows how statistical reasoning and various mathematical models contribute to risk management in finance. Regular readers would be interested to know that for the first time The Visible Hand will be publishing an extended online version, which will include additional research that the editorial board was unable to print due to the physical limit on the number of pages in the journal. We are thrilled to have received work of such high caliber and look forward to seeing the journal grow in the future.

Nivedita Vatsa Editor-in-Chief

Page 4: Visible Hand

4 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Unlike previous notable works, which focus pri-marily on comparing immigrant wages against the native white and black populations, this paper will focus solely on the comparison between the United States’ main immigrant groups: Asians and Hispan-ics. Currently, there are no other important prior pa-pers that concentrate exclusively on the economic impact and equality between Asian and Hispanic immigrants.

II. Literature Review

One of the first major papers discussing the earn-ings difference between immigrants and United States natives is Barry Chiswick’s The Effect of Americanization on the Earnings of Foreign-born Men (1978). His influential work suggests the im-portance of years resided as well as worked in the United States as major factors affecting the wage gap between natives and foreign-born. Ultimately, he asserts that the foreign-born labor force’s average

I. Introduction

Despite representing less than five percent of the U.S. population as compared to the near 17% Hispanic population (United States Census Bureau, 2014), Asian-Americans, who are generally looked upon as the “model minority,” have been left out of the diversity conversation. Some believe they are no longer looked upon as minorities due to their finan-cial success and are thus not given certain advan-tages afforded to other minorities (Linshi, 2014).This paper aims to analyze relatively new, very comprehensive data from the Princeton New Immi-grant Survey (NIS) in order to locate the apparent wage gap between Hispanic and Asian immigrants and to isolate the reasons behind said income dif-ferences2/. However, after locating the wage differ-ence and some reasons behind it, many questions still remain: does one race’s propensity to find more financial success in a new country justify programs like Affirmative Action? Or, conversely, since most of the Hispanic immigrants in the survey have re-sided in the United States for much longer than their Asian counterparts (Princeton University, 2006), does the wage gap represent a failure of the United States’ school system and evidence of overall dis-crimination? Putting aside the very complex social and po-litical issues, from an econometric standpoint, this paper’s findings suggest that Asian immigrants not only earn more income, but, more importantly, they also benefit from schooling moreso than Hispanic immigrants. In fact, the data analysis shows that Asian immigrants earn about six percent more in-come than Hispanic immigrants per each additional year of education. Additionally, Asian wages in-crease significantly more than Hispanic wages per each additional year working in the United States. Such clues point to a clear difference with a com-plex explanation behind it, which makes for eco-nomic and policy questions worth exploring.

Spencer Koo1

Princeton University

Comparing Assimilation & Success Rates of Legal First Generation Asian and Hispanic Immigrants

in the United States

Page 5: Visible Hand

Fall 2015 | The Visible Hand | 5

Volume xxiV. issue i.

wage eventually catches up to the native’s average wage, even surpassing it in the long run (Chiswick, 1978). However, at the time of his study, foreign-born United States residents and citizens made up only five percent of the total population, a number which has since been far overshadowed. Building off of Chiswick’s (1978) basic tenets is George J. Borjas’ paper, The Economics of Im-migration (1994). His publication introduces the proposed “aging effect,” which is the rate at which earnings increase over a life cycle. Combined with Chiswick’s (1978) assertion that immigrant wages increase the longer he or she resides and works in the US, Borjas claims that the “aging effect” is greater for immigrant populations as compared to native populations (Borjas, 1994). Similarly to Chiswick (1978), Borjas (1994) studied these effects to com-pare them to the native white and black populations and to make a contention about the effects of immi-grants on those native populations (Borjas, 1994). His regressions on the immigrant wage gap go even further than Chiswick’s, and these key variables that Chiswick (1978) and Borjas (1994) identified and pioneered, such as age and years resided in the US, can help explain the difference between differing immigrant groups’ wages. Since Chiswick (1978) and Borjas (1994) wrote about very broad categorizations of the wage gap,

many economists and sociologists have since pub-lished papers on the earnings gap between much more specific groups such as particular races, gen-der, immigrants, and more. Tienda and Lii (1987) investigated how the size of minority labor markets affects minority wages (separated by group) and white wages. Using data from 1979, they observed that in labor markets with a large share of minority residents, college-educated minorities experienced a significant earnings loss compared to their white college-educated counterparts while poorly educat-

ed minorities did not see as large a wage discrep-ancy with poorly educated whites. Much more rele-vant to this paper, Tienda and Lii (1987) discovered that Asians generally had a much higher education level compared to Hispanics, and, thus, on average earned a higher wage. Additionally, in areas without a large minority presence, the wage gap between Asians and whites was significantly lower than the gap between other minorities and whites (Tienda & Lii, 1987). Expanding on these past research endeavors, this paper will seek to find out whether those economic and educational differences still exist amongst mod-ern-day legal first generation Asian and Hispanic immigrants. The paper will also make use of the key economic characteristics pointed out by both Chis-wick (1978) and Borjas (1994) to identify reasons behind any discovered differences between the two groups.

III. Data

As briefly mentioned, this paper primarily uses the Princeton New Immigrant Survey (NIS), which offers a comprehensive questionnaire focusing on legal first generation immigrants. Though relatively new, the Princeton NIS (2003 and 2007) is the first nationally representative survey of new immigrants and their children (Princeton University, 2006).

This paper will focus on NIS responses from adults between 2003 and early 2004 that covered a well-distributed group of respondents who planned on attaining citizenship. The main reason behind using the NIS rather than the United States Census is the amount of de-tail in and accuracy of the survey. Unlike the broad Census, which tends to suffer from non-response and potential inaccuracies in the responses, the NIS took place with carefully selected families from dif-ferent locations around the US, and the focus of the

Page 6: Visible Hand

6 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

are included to identify the major contributing fac-tors to the wage gap. With the log of wage as the dependent variable, a dummy variable representing the various groups is crucial. Continuing with this method, the first linear re-gression will be run normally and as an entity-fixed regression by state. Additionally, it will only take into account a dummy variable to distinguish the two groups as well as various demographic charac-teristics. It is a simple base-line to show the differ-ence between the groups.

where logwij is the log of earnings in dollars for im-migrant i and state j, Asiani is a dummy variable which equals 1 if Asian and 0 if Hispanic, Agei is the age, Agei

2 is age squared since lifetime wage is qua-dratic, and Yij is a vector term, which represents the demographic data of each immigrant such as gender. Next, the same regression will be modified and a vector of economic characteristics of immigrants will be added. After running the regression again both normally and state-fixed, β1 from the regres-sion (1) can be compared with β1 from regression (2) to show the explained effects of the economic terms on the earnings gap.where the variables are the same as regression (1) with the added vector term Xij, which represents

many of the productive characteristics from Tables I and II such as years of education, English speak-ing ability, English comprehension, years resided in the United States, and years worked in the United States. In the last simple regression, the all-important education variable will be pulled out of the econom-ic characteristics vector, and the same normal and entity-fixed regressions will be run while interacting the education and the Asian dummy variables. This will further explain the wage gap by taking into ac-count race given that immigrant i has a certain level of education.where the variables are the same as regression (2) with the added interaction variables between Asian and education, which is represented by Edui. The final method used for analyzing the dif-ferences between groups is the Blinder-Oaxaca De-composition. The core idea behind this universally used decomposition is “to explain the distribution of the outcome variable in question by a set of factors

data makes up for the lower number of observations. Additionally, one would expect that if a significant difference were found between Asians and Hispan-ics in the NIS, which covers mainly middle class families across the board, there would be an even larger discrepancy in the broader Census. This paper concentrates on a few major variables that are used in related literature on immigration, which include earnings, years of education, age at immigration to the US, language fluency in both speaking and comprehension, age at the time of sur-vey, years resided in the US, and years worked in the US. All the variables from every table only in-clude data for immigrants between “working” ages of 25-603. These data are broken up between three major groups: Asian, Hispanic, and all other immi-grants as a control. As seen in Table I, there exists a significant wage gap between Asian and Hispanic immigrants. The data is further broken up by years of education with-

in each group to see if Asians out-earned Hispanics within education-level categories. Again, in Table II, we see that the Asian popula-tion outpaces the Hispanic one at all levels of educa-tion and even their white counterparts at the highest level. However, it can be difficult to interpret some of the data due to the low numbers of respondents certain categories. Nevertheless, the number of Asian immigrants with a higher education versus the number of Hispanic immigrants with only a high school diploma or less is astounding. As the paper moves forward with the regressions, more variables can be removed to further reduce the number of un-derlying factors and omitted variables.

IV. Methodology

The commonly accepted approach to analyzing wage gap questions between two or more groups

involves multiple stages of regressions. In each sub-sequent regression, potentially underlying variables

Page 7: Visible Hand

Fall 2015 | The Visible Hand | 7

Volume xxiV. issue i.

that vary systematically with socioeconomic status” (O’Donnell, van Doorslaer, Wagstaff, & Lindelow, 2007). In terms of this paper, the Oaxaca Decompo-sition examines the variations in the log of the wage and seeks to find the causal factors or variables that differ systematically based on race, either Asian or Hispanic. More specifically, the income gap be-tween Asians and Hispanics is decomposed into two parts: 1) part that is due to measurable differences in variables, and 2) part that is due to the magnitude of the effect of those variables (O’Donnell, van Doors-laer, Wagstaff, & Lindelow, 2007). For example, the measurable, or explained, difference could be the observed higher level of education of Asian immi-

grants in comparison to Hispanic immigrants. The immeasurable, or unexplained, difference could be a cultural trait for studying more and better work ethic. While the explained results can lead to con-crete policy changes that could potentially lead to greater equality in wages, the unexplained results

do not give a clear answer. However, knowing the magnitude of the effect for those variables can point policy in the right direction. To visualize the decomposition, start with an even further simplified version of regression (1) where xij is a vector of explanatory variables similar to the previous regressions.

As seen in the figure4 , the wage gap can be attrib-uted to both the explained sample means of the x’s (or the endowments, E), the unexplained β’s (or the coefficients, C), and the interaction between the two (CE) (O’Donnell, van Doorslaer, Wagstaff, & Lin-delow, 2007). In more general terms:which shows the individual parts for the explained endowments (E), unexplained coefficients (C), and the interaction between the two (CE). The Oaxaca Decomposition, again, compares the two by com-bining the interaction term with either the explained or unexplained components (O’Donnell, van Doors-laer, Wagstaff, & Lindelow, 2007).

By combining different terms, the single equation can on two different meanings. The first decomposi-tion assumes that Hispanics are paid fairly based on their characteristics, represented by vector x, while Asians earn more with those same characteristics for some unknown reason. The second decomposition assumes the opposite: Asians are paid according to their characteristics, but Hispanics are discriminated against in the work place. Using Stata’s Oaxaca ado-file, it is possible to run a Oaxaca Decomposition to see where the gaps come from and to shed light on possible reasons (Jann, DECOMPOSE: Stata module to compute de-compositions of wage differentials, 2005).

V. Results

Table III only covers the state-fixed regression data for increased accuracy and less potential survey bias. The outputs are very similar to the regular re-gressions, which indicates that the survey data was sufficiently randomized. As seen in column (1) of Table III, under the simplest regression without any detailed productive characteristics controlled, being Asian accounts for a massive 51% increase in income and is statistical-ly significant at the one percent level. Just as Tienda and Lii (1987) found in their research, Asians far outpaced their minority counterparts in terms of in-come. In fact, all of the basic demographic controls account for a statistically significant impact in the wage discrepancy. Clearly, there is more to earning a higher wage than just race and other demographic statistics as seen by the second regression in column (2) of Table III. When the regression includes major productive characteristics, being Asian continues to account for a very high 35% increase in income and is still statistically significant at the one percent level. Un-surprisingly, English comprehension and speaking

Page 8: Visible Hand

8 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

ability both contribute greatly to income as well. One must keep in mind that those two data points were measured on a self-rated scale from one to four (with four being best). Since it is not a uniform rat-ing system, it is possible for subjects to over- or un-derestimate their respective English skills; however, it is doubtful that one group would skew the results to a serious degree. Nevertheless, given that most of the Hispanic survey takers took the survey entirely in Spanish (Princeton University, 2006), it is possi-ble that these two characteristics offer much greater insight into the productive advantage of Asians as opposed to Hispanics, which will be covered later.

From an intuition standpoint, given the sheer amount of resources required to move half-way across the globe versus within the same or con-necting continent, it remains more than plausible that legal Asian immigrants had a better education growing up. Said education would most likely in-clude learning English before coming to the United States. Those differences are apparent in Table I. This would give Asians a significant leg up in terms of job opportunities and wages earned. Just as Chis-wick (1978) and Borjas (1994) claim, the longer the immigrant population resides and works in the United States, the better off they become (Chiswick, 1978) (Borjas, 1994). Simply put, Asians are mul-tiple steps ahead of Hispanics due to greater initial education, English language ability, and resources. Additionally, this theory of better and greater previous education is reflected in the United States’ immigration policies. Currently, US policy allocates many more visas and permanent resident statuses to skilled workers, investors, and persons that hold ad-vanced degrees (Immigration Policy Center, 2014).

Since this study focuses on legal immigrants seek-ing citizenship and the Asians interviewed have resided in the United States for less time and were older at the time of immigration (see Table I), it is reasonable to assume that they are highly educated or skilled workers. Additionally, they most likely had careers and lives in their countries of origin, which further helped with finding jobs of equal stat-ure and pay in the United States. This does not mean that all of the study’s Hispan-ic immigrants were not educated when they arrived. However, based on the data, Hispanic immigrants arrived at a much younger age, and even though they were legal and seeking citizenship at the time of the study, it says nothing about how they got to the United States. Given the proximity and history of Hispanic immigration to the United States, there are most likely many more immigrants of Hispanic origin that came as children or teenagers with little education and less resources who sought legal status after the fact. Given that the average age at the time of the 2003 NIS was about 39 years old (see Table I), arriving illegally in the 1980s would have been easier than in the present, so legal status could have been sought between then and the time of survey. Looking again at Table III, the effect of total years of education, though statistically significant at the one percent level, is smaller than the effect of characteristics like race and gender. In order to find the magnitude of the effect more education has on the Asian immigrant population, the Asian dummy variable must be interacted with the education vari-able as seen in the regression from column (3). This key interaction variable reveals that for each ad-ditional year of education, Asians earn six percent more than Hispanics. This could mean a few things: 1) there is a greater focus on education in Asian countries so the quality of schooling is better and students attend for more years on average, 2) Asians, through some immea-surable characteristics such as work ethic, intelli-gence, or cultural or family pressure, get more out of schooling than Hispanics, or 3) legal Asian immi-grants have more resources than legal Hispanic ones in their home countries and, thus, receive a better education including study skills and other immea-surable traits. At this point, the greatest contributions to a higher wage for Asian immigrants (other than gender and age, which were used just as controls) are education related as seen by the coefficients on years of educa-tion and the interaction variables. In fact, as men-tioned above, even English skills can be attributed to a better education. However, it is still difficult to

Page 9: Visible Hand

Fall 2015 | The Visible Hand | 9

Volume xxiV. issue i.

discern exactly why Asians get more out of educa-tion from these basic regressions. The Blinder-Oax-aca Decomposition and analysis will reveal more about where exactly the gaps lie and how much is still unexplained. In the first step of the Blinder-Oaxaca Decom-position, the Asian dummy variable is removed, and two separate models for each corresponding immi-grant group are run. Unlike the simpler regressions from Table III, which included an Asian dummy variable, Table IV reveals far fewer statistically sig-nifiant variables. Table IV exists mainly to point out the differences between each group independent of the other. The table clearly shows a much larger return to each ad-ditional year of education that Asians receive. This could be due to discrimination in US schools against Hispanics, bias in origin nations against Hispanics that tend to emigrate to the United States, or immea-surable traits that cause Asians to get much more use out of their education. The one variable that leads to a greater wage for Hispanic immigrants is years resided in the United States. Despite the data, it is important to keep in mind that, as seen in Table I, the average number of years resided and worked in the United States is more than double for Hispanic immigrants. There-fore, the small and statistically insignificant magni-tude effect of the number of years resided in the US for the Asian model will most likely have drastically risen since the time of the survey according to Chis-wick (1978) and Borjas (1994). Additionally, despite the lower amount of time resided in the United States, the magnitude at which Asians’ wages increase per year worked in the US

is more than threefold their Hispanic counterparts. Again, these results could suggest discrimination against Hispanic immigrants or advantages for Asians in terms of previous education. Such ad-vantages like resources and a better school system, would explain why Asians make more money per each additional year of school. A major clue that favors the theory of better previous education and resources is that despite having lived in the United States for much less time, Asian survey takers have a far greater understanding of English than their His-

panic counterparts. Again, it must be stressed that, even though the survey was offered in more than 19 languages, 73.1% of Hispanic respondents took the survey in Spanish (Princeton University, 2006). This sets the stage, as discussed in the meth-odology, to combine and compare the two groups on equal terms. Running the second step of the Blinder-Oaxaca decomposition will separate out the cumulative effects of the two groups and compare Asian wages as if they had the endowments and co-efficients of Hispanics. This will reveal how much of the wage gap is explained by their respective en-dowments, coefficients, and the interaction between the two. Column (1) of Table V shows the overall ef-fect of all of the factors on both groups. There is a clear difference between the cumulative effects of the two groups, which is statistically significant at the one percent level. In other words, when the endowments, coefficients, and interaction terms of Hispanics are applied to Asians, there is a statisti-cally significant decrease of 59.5% in Asian wages. According to the overall effect given in column (1), the effect of replacing the Asian coefficients with Hispanic ones accounts for a decline of 49.7% in

Page 10: Visible Hand

10 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Asian wages. This coefficient effect is more statis-tically significant than the effect of changing the endowment values, which decreases Asian wages by 34%. The interaction term, which is the simul-taneous effect of applying Hispanic endowments and coefficients to Asians, is the only factor that fa-vors Hispanic wages. More specifically, it indicates that if Hispanic endowments and coefficients were simultaneously applied to Asians, Asians’ wages would actually rise by 24.2%. However, the interac-tion term is not statistically significant so it can be ignored for the most part. Even though column (4) shows that parts of the interaction effect are signifi-cant, those parts have been countered by the insig-nificance of a majority of the other variables. Next, and most importantly in the Oaxaca De-composition, columns (2), (3), and (4) show the breakdown of each individual variable’s contribu-tion to the effects of the endowments, the coef-ficients, and the interaction. These values provide further insight into whether the main force behind the wage gap is driven by the explained measured data, the unexplained immeasurable traits, or a com-bination of both. Starting with column (2), the statistically sig-nificant variables for the endowment effect, or mea-sured data, are years worked in the United States, English comprehension, and total years of educa-tion. The data for the number of years worked in the United States shows a large effect in favor of the Hispanic immigrant income. Specifically, the decomposition states that if Asians had been work-ing in the United States as long as Hispanics have, their income would increase by 56.5%. Intuitively, this result makes perfect sense. According to the analysis presented in both this paper and the works of Chiswick (1978) and Borjas (1994), the longer an immigrant resides and works in the United States, the more wages he or she will earn. Looking back at Table I, it is clear that Hispanics have both resided and worked in the US for more than twice as long as their Asian counterparts. Despite that advantage, Asian income still far outperforms that of the Hispanics. This means that there are even stronger forces working either against Hispanics in the form of discrimination or in favor of the Asian immigrant population in the form of better work ethic, greater prior resources, or other data not measured in the survey or a mix of all three. Though only significant at the 10% level, English comprehension favors Asian wages by nearly 20%. The decomposition shows that if Asians immigrants from the survey understood English as well as the Hispanic respondents, they would earn 20% less.

This is further proof of potentially better education and resources from the Asian survey takers’ coun-tries of origin. Despite living and working in the US for much shorter periods of time, they understand English at a higher level. Although not statistically significant, the decomposition shows that they also speak English at a higher level as well. By far the most important factor is total years of education, which is significant at the one percent level. The measured endowment effect strongly favors Asian immigrant wages. In simple terms, if Asians had the same number of years of educa-tion as the Hispanic respondents, their wages would drop by a massive 53.2%. As seen in Table I, Asians average 4.1 years more of education than Hispanics, which is the difference between a college graduate and a high school diploma. Nevertheless, despite the huge educational disparity, there is still no hard evidence of discrimination. Both bias against His-panics and better prior education of Asians or some other intangible traits could be the root cause. Beyond the two major control variables of age and gender, the other insignificant variables are years resided in the United States and English speaking skills. While years resided is very closely related to years worked in the United States, it is intuitive that one would have to be working and gaining experi-ence to earn a higher income. Residency itself is not a major factor. In terms of English speaking skills, although very closely associated with English com-prehension, the difference between the Hispanics’ and Asians’ ability to speak does not make a statisti-cally significant impact on wages. In column (3), the statistically significant vari-ables for the unexplained terms, or the magnitude of each variable’s effect, are years worked in the Unit-ed States and total years of education. Significant at the five percent level, the decomposition reveals that if Asians got the same out of each year worked in the United States as Hispanics, their income would decrease by 27.6%. Here, it appears that Asians have much better returns to their income for each year they spend working in the US. The effects that Chiswick (1978) and Borjas (1994) proposed seem to have a much larger effect on the Asian immigrant population. Though not statistically significant, the coefficient on the number of years resided in the United States intuitively favors Hispanics. Since, on average, they have resided in the US for much longer than the Asian immigrants from the survey, they have been able to settle, find jobs, and work their way up for longer. Due to this stability, it makes sense that ap-plying the Hispanic coefficient for years resided on

Page 11: Visible Hand

Fall 2015 | The Visible Hand | 11

Volume xxiV. issue i.

the Asian model would improve Asian wages. Nev-ertheless, running a Blinder-Oaxaca decomposition on the updated data from the 2007 NIS could prove the current decomposition inaccurate in terms of years resided. The major significant variable again concerns education. When the Hispanic coefficient, or how much they get out of each year of school, is applied to Asians, Asian wages decrease by a whopping 92.4%, significant at the one percent level. The co-efficient effect conveys that the overall quality of the education that Hispanics receive does not re-motely compare to that of the Asians. Again, this does not say exactly why Hispanics get much less effectiveness out of each year of education, but it does clearly nudge the conclusion in the general di-rection of education. It seems that once again, there are a couple clear possibilities. One could fairly assume that the rea-son Asians get more return out of each year worked in the United States is due to better and more effec-tive education that they also get a high return from. The other possibility is discrimination against these Hispanic immigrants in the United States and in their respective countries of origin. In reality, it is most likely a mix of the two factors. The last component of the difference is from column (4), the interaction effect. Even though variables within the effect are significant, the over-all effect is not statistically significant. Because of the lack of significance, it is difficult to say how the combination of the endowments and coefficients applied to the Asian immigrants actually affects the wage gap. Therefore, the interaction effect can be ignored for the most part. Clearly, the lack of observations hurts this data set and its analysis. Having a larger data set or com-paring this decomposition using the 2007 NIS could provide more insight into whether the endowment effect of years worked in the United States as well as other values are accurate. By conducting an even more recent survey, the endowment effect of years worked would most likely decrease in favor of Asian wages because both groups have been in the United States for a sufficient amount of time. Another strat-egy would be to take United States Census data to re-run the Blinder-Oaxaca Decomposition. While it would no longer cover only legal immigrants, it may give a more accurate intuition into how years worked in the United States affects the two groups, even if one might expect the wage disparity to be even more in favor of Asians. Ideally, a careful study specifically focused on the education of immigrants, old and new, would need

to be done. The survey must focus on years of edu-cation in the country of origin, years of education in the United States, school district within the United States, time devoted to education in the home, atten-tion from parents concerning education, evidence of bullying and discrimination at school in both the country of origin and the United States, and more along those lines. With those data points in line, it would be possible to more accurately figure out where and why Asians immigrants benefit so much more from the education they receive.

VI. Conclusion

After identifying clear differences between the Asian and Hispanic first generation immigrants from the Princeton NIS, careful statistical analysis and a Blinder-Oaxaca Decomposition were performed to identify the causes behind the discovered wage gap. In summary, the evidence and analysis discov-ered that the wage gap can be attributed to Asian immigrant’s greater amount of education as well as greater return from their years in school. How-ever, the source of the Asian population’s success in school cannot be statistically identified as either discrimination against Hispanics or immeasurable traits that improve Asian school performance and, subsequently, wages. Without that concrete conclu-sion, there cannot be any suggestions toward correct policy changes. Nevertheless, this is a start. More surveys and investigation that focus specifically on education must occur. With newly researched data, the ques-tion can be re-analyzed to not only provide answers to the statistical questions, but also help shape future policy. Rather than arbitrarily giving certain advan-tages to only select minority groups or filling quo-tas through Affirmative Action and guess work, the United States can help fix the education system and, in the process, bolster the work force and the lives of its citizens and future citizens.

VII. References5

Black, D. A., Amelia, H. M., Sanders, S. G., & Tay-lor, L. J. (2008, Summer). Gender Wage Disparities among the Highly Educated. (630-659, Ed.) The Journal of Human Resources, 43(3), 630-659. Re-trieved December 09, 2014

Borjas, G. J. (1994, December). The Economics of Immigration. Journal of Economic Literature, 32(4), 1667-1717. Retrieved December 10, 2014

Page 12: Visible Hand

12 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Chiswick, B. R. (1978, October). The Effect of Americanization on the Earnings of Foreign-born Men. Journal of Political Economy, 86(5), 897-921. Retrieved December 10, 2014

Cotton, J. (1988). On the Decomposition of Wage Differentials. Review of Economics and Statistics, 70(2), 236-243.

Fairlie, R. W. (2003, November). An Extension of the Blinder-Oaxaca Decomposition Technique To Logit and Probit Models. Economic Growth Center. Retrieved December 10, 2014

Immigration Policy Center. (2012, October 25). De-ferred Action for Childhood Arrivals: A Resource Page. Retrieved April 17, 2015, from Immigration Policy Center: American Immigration Council

Immigration Policy Center. (2014, March 01). How the United States Immigration System Works: A Fact Sheet. Retrieved April 17, 2015, from Immigration Policy Center: American Immigration Council

Jann, B. (2005, May 12). DECOMPOSE: Stata module to compute decompositions of wage dif-ferentials. Bern, Switzerland. Retrieved March 26, 2015

Jann, B. (2008). The Blinder-Oaxaca decomposi-tion for linear regression models. (J. Newton, & N. J. Cox, Eds.) The Stata Journal, 8(4), 453-479. Re-trieved December 10, 2014

Linshi, J. (2014, October 14). The Real Problem When It Comes to Diversity and Asian-Americans. Time. Retrieved 03 20, 2015

Neumark, D. (1988). Employers’ Discriminatory Behavior and the Estimation of Wage Discrimina-tion. Journal of Human Resources, 23(3), 279-295.

O’Donnell, O., van Doorslaer, E., Wagstaff, A., & Lindelow, M. (2007). Analyzing health equity using household survey data : a guide to techniques and their implementation. Washington, D.C.: The Inter-national Bank for Reconstruction and Development / The World Bank. Retrieved March 26, 2015

Princeton University. (2006). Study Goals: Why Study Immigration in America? Retrieved Decem-ber 10, 2014, from The New Immigrant Survey

Reimers, C. W. (1983). Labor Market Discrimina-tion against Hispanic and Black Men. Review of Economics and Statistics, 65(4), 570-579.

The Blinder-Oaxaca Decomposition. (n.d.). Re-trieved December 10, 2014, from Washington Uni-versity

Tienda, M., & Lii, D.-T. (1987, July). Minority Con-centration and Earnings Inequality: Blacks, Hispan-ics, and Asians Compared. American Journal of Sociology, 93(1), 141-165. Retrieved December 09, 2014

United States Census Bureau. (2014, December 03). State & Country Quickfacts. Retrieved December 10, 2014

VIII. Footnotes

1. Special thanks to Professor Jessica Pan and Ji Huang for useful comments and assistance. Addi-tional thanks to Monica Cai and Anthony Moham-med for edits and comments.

2. The Princeton NIS covers legal, first-generation immigrants as well as their spouses and children.

3. Major outliers have been removed from the sum-mary statistics and analysis.

4. Graph modified from (O’Donnell, van Doorslaer, Wagstaff, & Lindelow, 2007).

5. APA style used for references and footnotes and American Economic Review style used for general formatting.

Page 13: Visible Hand

Fall 2015 | The Visible Hand | 13

Volume xxiV. issue i.

I. Introduction

When do Chief Executive Officers act in the in-terest of shareholders? A large literature in financial economics asks which specific incentives induce CEOs to act in the best interests of firms and share-holders (e.g., Shleifer & Vishny, 1988). One such incentive, or form of leadership structure, is CEO duality, a situation where a CEO also serves as the Chair of a company’s board of directors. The existing scholarly literature remains divided about whether a CEO should also chair the board of directors. On the one hand, agency theorists argue that Board-Chair CEO’s abuse their excessive un-checked power in a way that destroys shareholder value (e.g., Eisenhardt, 1989; Fama & Jensen, 1983; Roe, 2004; Williamson, 1985). Stewardship theo-rists, on the other hand, argue that excessive moni-toring destroys the CEOs incentive to take risks while also degrading the trust between CEOs and boards, and that good CEOs should be given addi-tional opportunities to take responsibility and initia-tive, which will consequently elevate firm perfor-mance and outcomes (e.g., Brickley & Coles, 1997; Chen, Lin, & Yi, 2008; Jensen & Meckling, 1976; Sridharan & Marsinko, 1997). These two views would make the following predictions about CEO negotiating behavior during an impending acquisi-tion:

H0= Dual CEOs, due to lower levels of monitoring, negotiate less aggressively than single role CEOs, leading to lower levels of value maximization for the firm (Agency Hypothesis).

H1= Dual CEOs, due to lower levels of monitoring, negotiate more aggressively than single role CEOs, leading to higher levels of value maximization for the firm (Stewardship Hypothesis).

However, examination of these hypotheses is complicated by a fundamental epistemological chal-lenge, which is that scholars often test the value-maximizing merits of incentives (like CEO duality) by empirically estimating their effects on outcomes (like deal premia). This trend is common among studies that use ready-made data sets, often includ-ing aggregate performance measures like ROA or ROE to test the effectiveness of a policy within the firm. The prevalent usage of performance measures to test the merits of incentives stems from the lack of suitable proxies for behavior. Since we, as research-ers, cannot monitor CEO behavior by putting cam-eras in offices or tapping phones to directly quantify and measure effort, this kind of day-to-day behav-ioral data typically does not exist. While these stud-ies have made significant contributions to the cor-porate governance literature, it is worth noting that

Victor GhazalGrinnell College

CEO Duality and Corporate Stewardship:Evidence from Takeovers

* Acknowledgement: I would like to thank Professor Caleb Stroup at Davidson College for his excellent mentorship, support, and guidance on this paper.

Many scholars have been quick to criticize the merits of CEO duality, a situation where a company’s Chief Executive Officer is also the Chairman of the Board, by claiming that CEO duality undermines the board’s ability to effectively monitor and constrain self-interested CEOs. These criticisms are often based on em-pirical studies that use firm outcomes—aggregate performance measures—as proxies to evaluate the merits of an incentive structure such as duality on the behavior of CEOs. In this paper, I construct a more direct measure of CEO behavior by gathering information submitted by companies to the Securities and Exchange Commission. The novel variable I introduce in this paper measures how aggressively a CEO whose com-pany is being sold negotiates with a prospective buyer during the pre-announcement sale process. I find that dual CEOs act in the interest of their shareholders by bargaining 16.1% more aggressively in takeover negotiations than do single role CEOs. The paper’s main finding is consistent with the view that top manag-ers, when given higher levels of responsibility, act as good corporate stewards on behalf of their respective firms and shareholders.

Page 14: Visible Hand

14 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

it is not the incentive that influences the outcome, but the behavior of the economic agent the incentive is designed to influence (Finkelstein, Hambrick, & Canella Jr., 2009). In this paper, I advance the understanding of CEO incentive structures by hand-collecting a novel dataset that permits the construction of a more di-rect proxy for a CEO’s behavior, thus allowing me to directly examine the effect of incentives on be-havior, rather than attempting to infer behavioral changes from overall corporate outcomes, such as profitability, which are known to be influenced by many other factors (e.g., regulation, product market competition, international business cycles) that a CEO cannot directly control. I proceed in doing so by exploiting the fact that public disclosure require-ments mandate that proceedings leading to substan-tive changes made public by publically traded com-panies be made available to the public for current and potential investors1. The specific contribution of this paper is to em-pirically examine the link between a particular in-centive, CEO duality, and CEO behavior. Figure 1 illustrates that the relationship between leadership structures and outcomes is indirect, and is mediated by CEO behavior. This implies that regressions of deal premia on CEO duality are unlikely to pro-vide insight about the possible effect of duality on a CEOs behavior.

In Figure 1, leadership structure, an incentive, affects the behavior of the CEO, which then influ-ences corporate outcomes, such as the profitability of the firm. If incentives that lead to good CEO be-havior are chosen, then the positive behavior is ex-pected to lead to a de facto best-case scenario for the firm and its shareholders when an outcome occurs. This is the case because outcomes are determined by not only the behavior of a singular agent, but also by market risk and idiosyncratic firm noise. These

two terms refer to the innumerable factors such as the number of potential bidders for a target, the tar-get’s industry, profitability, the financial suitability of acquirers, government regulations, overall mar-ket forces, among many others, that influence ob-served deal premia. In other words, whether or not the CEO of a company also holds the title of Chair-man of the Board may not necessarily lead to better deal outcomes. The presence of these other features (market risk and firm noise) introduces measurement error that clouds what would otherwise be a clear empirical relationship between leadership structure and cor-porate outcomes. Table 1 illustrates the difficulty of attempting to evaluate the merits of an incentive with an outcome by reporting estimates of the effect of CEO duality2 and CEO Negotiating Aggressive-ness3 (a behavior) on the price received by share-holders when their company is sold (i.e. the deal premium4). Column (1) shows that the estimated effect of CEO duality on the deal premium is statistically in-distinguishable from zero, indicating the presence

of measurement error, described above. The low R-Squared in Column (1) indicates the high amount of econometric measurement error present when at-tempting to measure the merits of an incentive with an outcome. My empirical approach is to hand-collect data on initial bids contained in documents submitted to the Securities and Exchange Commission by merging companies. This information allows me to construct a novel–and direct–measure of a CEO’s behavior. This measure is “Negotiating Aggressiveness,” which proxies for the extent to which the CEO of a selling company bargains with a purchasing com-pany to obtain the highest possible price. Column (2) of Table 1 shows that negotiating ag-gressiveness has a statistically significant and posi-

Page 15: Visible Hand

Fall 2015 | The Visible Hand | 15

Volume xxiV. issue i.

tive estimated effect on the deal premium.Figure 2 illustrates why my empirical approach, which regresses negotiating aggressiveness on CEO duality, circumvents the measurement error chal-lenge: The new measure allows me to directly es-timate the relationship between CEO duality and CEO behavior which is presumed to affect the final price paid. To date, no study that has tested the effect of duality on value maximizing behavior within firms. The present study is the first to evaluate the value maximizing merits of this leadership structure in

in the setting of M&A negotiations, adding another dimension through which the advantages and disad-vantages of monitoring structures can be weighed.

II.A Methodology

I begin by measuring CEO Negotiating Aggres-siveness, NAGGi, as the final bid minus the initial bid over the initial bid in deal i. This equation mea-sures how many percentage-points the buyer’s final bid is removed from the initial bid. Final bid before sale, FBi, is calculated as the price per target share that is accepted as the sale priceby the target upon the execution of the merger agreement in deal i. The initial bid of the future buyer, IBi, is measured as the first bid submitted by the eventual acquirer in deal i. If a range of prices is indicated, the lower of two amounts is assumed.

NAGGi = FBi - IBi IBi

NAGGi: Negotiating AggressivenessFBi: Final Bid before SaleIBi: Initial Bid of Future Buyer

This measure is a tool to shed insight on M&A

negotiating behavior and, thus, value maximizing behavior. Subramanian (2011) describes dealmak-ing as two parties meeting somewhere in a zone of possible agreements (“ZOPA”). Negotiation theory tells us that buyers and sellers approach a negotia-tion table with a range of acceptable prices by which to strike a deal. This is a useful way to think about negotiations when considering value maximization because it is a truly relative term, meaning people have different views on what specifically constitutes maximization. While the law aims to protect share-holders by imposing a fiduciary obligation on man-agers, this does not change the relativity of the term. Therefore, thinking about the existence of a ZOPA allows us to judge value maximization as the extent to which each CEO will negotiate to remove the bidder from his or her initial willingness to pay. High levels of negotiating aggressiveness must be value-adding if one accepts the premise that the buyer does not make an indication of interest or an initial bid with a price that is not truly reflective of his or her willingness to pay. With this in mind, there must be implications for corporate governance if there is a leadership structure that is more condu-cive to these higher levels of negotiating aggressive-ness. An intuitively appealing methodological ap-proach to estimate the effect of duality on negotiat-ing aggressiveness would be Equation (1). NAGGi = β0 + β1 DUALi + εi (1) Where DUALi is a dummy variable that takes a value of one if the CEO of the selling company is also the Chairman of the Board in deal i. The coeffi-cient of interest is β1-hat. Without control variables, a causal interpretation of a positive estimate of β1 would lead to the rejection of the null hypothesis, H0. Conversely, a negative estimate of β1 would cause an idealized econometrician to fail to reject the null hypothesis. However, a potential challenge with this causal interpretation is the possibility that certain firms at-tract more bidders than others. One could imagine that the presence of each additional bidder would detach the CEO from being the chief negotiator, as the bidders would compete with each other. Due to this added sense of competition, the effect of the CEO’s negotiating aggressiveness will be weakened because bidders will shade their initial bids upwards to factor in the presence of additional bids—mean-ing that number of bidders would be negatively cor-related with negotiating aggressiveness. Addition-ally, dual role led firms may attract more bidders

Page 16: Visible Hand

16 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

because the CEO may have more freedom to contact more prospective buyers without having to receive approval from the board, which would cause duality and the number of bidders to be positively corre-lated. If this were true, it would cause the β1-hat co-efficient to be downwardly biased. So, I control for this potential issue with the BDRSi control variable, which is measured as the total number of potential buyers that submit a formal indication of interest or a binding or non-binding letter of intent in deal i. Another factor to consider when observing the takeover process is that failing firms may lack bar-gaining power, causing this status to be negatively correlated with negotiating aggressiveness. Addi-tionally, it may be the case that dual CEOs are at the helm of less failing firms that are for sale because they may have more power to stay in business. If this is the case, failing firms are be negatively cor-related with duality. To avoid downwardly biasing β1-hat, I address the issue with FAILi, a dummy vari-able takes a value of one if firm failure or industry decline are cited as a motivation for the sale in the “Reasons for the Merger” section in the SEC proxy statement in deal i. An important feature of takeovers is that some occur through an auction process. If an auction is present before the final sale, then this may lead to upward bias in the coefficient of interest. Atkas & De Bodt (2009) find empirical evidence that sug-gests that the threat of auction, latent competition, is factored into bidding behavior. Bids occurring be-fore an auction tend to be biased upwards to deter the seller from initiating an auction and introduc-ing new competition. This may create in upwardly biased starting bid, weakening the statistical effect that the aggressiveness variable picks up, which would cause the lack of a previous auction to be negatively correlated with negotiating aggressive-ness. Also, one might imagine that dual role CEOs may be more likely to engage in pure negotiations instead of resorting to an auction because in a one-on-one negotiation, the dual CEO can more effec-tively exercise his or her freedom. This may cause duality and no previous auction to be negatively correlated. To control for the potential upward bias, I include the PAi control variable, which is measured as an indicator variable takes a value of one if an auction has not taken place before a negotiation leading to a sale in deal i. A specific buyer’s over-eagerness to acquire a firm may downwardly bias β1-hat in the bivariate causal interpretation presented in Equation (1). An indication of this factor can be measured as whether the acquirer or the target makes the initial contact in

a given deal. One might imagine that if the acquirer makes initial contact, it may be indicative of a high willingness or ability to pay, which may bias ini-tial bids upwards to ward off the target’s potential impetus to contact additional competitors (Atkas & De Bodt, 2009). Thus, since initial bids would be shaded upwards, this decreases the statistical effect of the NAGGi variable, causing them to be negative-ly correlated. In addition, firms with dual leadership may have CEOs who take more ownership over the sale process and reach out to more bidders—making duality and acquirer initiated contact positively cor-related. If this is true and the coefficient of interest is downwardly biased, I avert the problem by includ-ing the AINITi dummy variable, which takes a value of one if the eventual acquirer initiates first contact in deal i. Another biasing factor that may corrupt the es-timated duality coefficient is the size of the target company. More information about the value of the firm, before due diligence, may be known about bigger firms. If this is the case, initial bids may be shaded upwards, closer to a symmetrically known value, leading to less negotiating on behalf of the CEO. If this is valid, then size is negatively corre-lated with aggressiveness. Further, dual CEOs may tend to lead larger firms at the time of sales, caus-ing duality and size to be positively correlated. This makes sense, given that dual CEOs may tend to start at small firms that grow over time, which are then sold to bigger firms or competitors. Boone and Mul-herin (2004) find that the mean and median target equity value is 56% and 27% of that of the aver-age bidder respectively, which provides evidence in support of this prediction. Because failure to include an adequate control variable would bias my coeffi-cient upwards, I use MKTCi. This variable is taken as the log target market capitalization of the target firm in deal i, in millions of dollars. Moreover, an issue with the causal interpreta-tion of Equation (1) is the possibility that CEOs are also the owners of their firms, which may bias the estimation of β1 in an upward direction. Hermalin & Weisbach (2003) find that CEOs who are also owners behave differently than non-owner CEOs. Due to this, CEOs who are also owners may have an unobservable attachment to the firm that may cause a positive correlation to exist between owner-ship and aggressiveness. Further, one might imagine that duality and ownership are positively correlated because founders or owners of firms may want to place themselves in positions to exercise full over-sight over their firms. To avoid this potential biasing factor, I include OWNi, which is a dummy variable

Page 17: Visible Hand

Fall 2015 | The Visible Hand | 17

Volume xxiV. issue i.

that takes a value of one if the standing CEO at the time of the sale in deal i is also the owner. A final factor that may drive upward bias in the coefficient of interest is post-deal employment perks. One might imagine that if the CEO wants to be retained, he or she may be more likely to exhibit loyalty and competency. If this is the case, post-deal employment is correlated with aggressiveness. Ad-ditionally, dual CEOs may be more valuable to the post-merger firms or may be in a better position to bargain for retention as part of the sale than single role CEOs—causing duality and post-deal employ-ment to be positively correlated. To avoid the up-ward bias in β1-hat, I control for post-deal employ-ment with EMPi, which is an indicator variable that takes a value of one if the CEO of the target com-pany is retained in some capacity at the new firm post deal i. In the previous discussion, I proposed theoretical biases based on my understanding of each variable’s respective relationships with duality and negotiat-ing aggressiveness. To supplement this analysis, I also provide a correlation matrixcorrelation matrix is included in the Appendix to further support the directions and magnitudes of potential biases. Rewriting the main estimating equation and plug-ging in control variables for potentially biasing fac-tors yields Equation (2).

NAGGi = β0 + β1 DUALi + β2 BDRSi + β3 FAILi + β4 PAi + β5 AINITi + β6 MKTCi + β7 OWNi + β8

EMPi + εi (2) NAGGi: Negotiating AggressivenessDUALi: CEO DualityBDRSi: Number of BiddersFAILi: Failing FirmPAi: No Previous AuctionAINITi: Acquirer InitiatedMKTCi: Log Market Cap (M)OWNi: CEO OwnershipEMPi: Post-deal CEO Employment

II.B Data

I examine a sample of 45 M&A deals from the year 2013. These 45 deals are chosen at random from a list of all S&P 1500 mergers completed in 2013 with deal values greater than $1.0 million, where the acquirer gains outright ownership of the target firm. . The sample is not considered small in comparison to other studies with hand-collected data. By only focusing on the year 2013, I attempt to decrease time variant noise effects. I draw data

on dual CEO/Cchair positions, ownership status, post-deal employment, and market cap by investi-gating the company history and financial status on the from the Bloomberg Terminal. I gather confir-matory data to support these variables from SEC Edgar. Additionally, I hand-collect the data on CEO negotiating behavior from the “Background to the Deal” sections of publicly available proxy state-ments (i.e., DEFM-14A, S-4, SC-14D9, and 8-K fil-ings) through the SEC Edgar database (e.g., Boone & Mulherin, 2004; Gentry & Stroup, 2015). I also gather information on the number of bidders, wheth-er the deal occurs pre- or post-auction, whether the acquirer or the target initiates contact, and bid and sale information from the “Background to the Deal” sections in those proxy statements. I gather confir-matorydouble-check details on final sale amount and deal premia from the announcement documents in the 8-K filings (i.e., EX-99.1) on SEC Edgar and the Bloomberg Terminal. I draw further data on firm failure from the “Reasons for the Merger” section provided in these documents, often directly below the “Background to the Deal” sections. I remove deals that are hostile takeovers from my sample, as those deals do not fully reflect the CEOs ability to negotiate. The data I collect for deal premia, used in Table 1 and Figure 3, are taken from the Bloomberg Terminal and depict the percent premium over the closing price from the day before the deal execution.

Table 2 reports descriptive statistics for each vari-able used in this paper. Interestingly, approximately one-third of the randomly selected sample of M&A deals from 2013 involves firms with CEO-Chaired boards. The number of bidders in the sample ranges from 1 to 20, reflecting a large variance in not only

Page 18: Visible Hand

18 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

the number of firms contacted but those that make formal indications of interest. The negotiating ag-gressiveness variable has a minimum value that is negative, which may suggest two possibilities, ei-ther certain CEOs are extremely poor negotiators or the variable is not impervious to firm-specific bias-ing factors. If the second suggestion is valid, then the R-Squared value should increase from Equation (1) to Equation (2) and the usage of control variables is justified in depicting a more accurate estimate of the effect of duality on aggressiveness.

III. Findings

Column (1) of Table 3 reports the main estimating equation in the bivariate case, as shown in Equation (1). Further, in Column (2) the findings from Equa-tion (2), which presents the estimated effect of dual-ity on negotiating aggressiveness in the multivariate case, are reported. Column (1) reports a positive and significant cor-relation between duality and aggressiveness in the bivariate case. When I introduce control variables in Column (2), I find that the coefficient on CEO

duality remains positive and statistically significant at the 1% level. The results suggest that dual CEOs negotiate 16.1% more aggressively than single role CEOs. That is, when the CEO also occupies the role of the company’s Chairman of the Board, the exis-tence of that relationship in the context of an M&A deal negotiation is associated with a 16.1 percent-age point increase from initial to final bid, ceteris paribus. I therefore reject the null hypothesis that dual CEOs bargain less aggressively than single role

CEOs, leading to lower levels of value maximiza-tion. I also find that the presence of each additional bidder is associated with a -1.1% decrease in nego-tiating aggressiveness. This result is significant at the 5% level. In theory, the presence of more buyers should decrease the amount of bargaining the CEO has to do because of the added sense of competi-tion among prospective buyers. I am surprised that none of my other control variables were significant, but acknowledge that this may be due to the small sample size. Figure 3 below reports a 2-dimensional visual representation of the 3-dimensional relationship between the incentive (duality), the behavior (ne-gotiating aggressiveness), and the outcome (deal premium). The contour plot illustrates two important conclu-sions from this study. First, CEO duality and deal premia have no systematic relationship, as shown in Table 1. Observing the chart, the red-shaded area, indicating duality, is not disproportionately to the

north or south of the graph, confirming this asser-tion. Second, the notable divide between the right and left sides of the plot, the red- and blue-shaded areas, confirms the findings in Table 3 in a striking manner; dual CEOs bargain more aggressively than single role CEOs. The results indicated in Figure 3 taken together with Table 2, which suggests that aggressiveness and deal premia are positively correlated at the 1% level, confirms our rejection of the null hypothesis. That is, if we accept the premise that higher deal pre-mia is good for shareholders, then it is clear that the duality structure is more conducive to CEO value maximizing behavior in M&A target negotiations.

Page 19: Visible Hand

Fall 2015 | The Visible Hand | 19

Volume xxiV. issue i.

Boyd, B. (1995). CEO duality and firm perfor-mance: A contingency model. Strategic Manage-ment Journal, 301-312.

Brickley, J., Coles, J., & Jarrell, G. (1997). Leader-ship Structure: Separating The CEO And Chairman Of The Board. Journal of Corporate Finance, 189-220.

Buchanan, R., & Kim, K. (2008). CEO Duality Leadership And Firm Risk-Taking Propensity. The Journal of Applied Business Research, 24(1), 27-42.

Chen, C., Lin, J., & Yi, B. (2008). CEO Duality And Firm Performance—An Endogenous Issue. Corpo-rate Ownership & Control, 6(1), 58-65.

Daily, C., & Dalton, D. (1994). Bankruptcy And Corporate Governance: The Impact Of Board Com-position And Structure. Academy of Management Journal, 1603-1617.

De Bruyn Demivoda, I. (2013). Time lost in nego-tiations: Productive or wasteful? 1-42.

Donaldson, L., & Davis, J. (1991). Stewardship Theory or Agency Theory: CEO Governance and Shareholder Returns. Australian Journal of Manage-ment, 49-64.

Eisenhardt, K. (1989). Agency Theory: An Assess-ment and Review. Academy of Management Re-view, 57-74.

Elsayed, K. (2007). Does CEO Duality Really Af-fect Corporate Performance? Corporate Gover-nance: An International Review, 1203-1214.

Faleye, O. (2014). The Costs of a (Nearly) Fully In-dependent Board. 1-44.

Fama, E., & Jensen, M. (1983). Separation Of Own-ership And Control. The Journal of Law and Eco-nomics, 26(2), 301-325.

FASB. (1980). Statement of financial accounting concepts no. 2: Qualitative characteristics of ac-counting information. Stamford, Conn.: Financial Accounting Standards Board.

Finkelstein, S., & D’aveni, R. (1994). CEO Duality As A Double-Edged Sword: How Boards Of Di-rectors Balance Entrenchment Avoidance And Uni-ty Of Command. Academy of Management Jour-

IV. Conclusion In this paper, I push the frontier of what we know about the role of CEO duality in the firm and pro-vide insight about the role of this leadership struc-ture in the context of takeover negotiations. I expose the fundamental identification problem that occurs when we try to judge the merits of an incentive with an outcome variable, and circumvent the lack of strong behavioral variables by creating my own, CEO Negotiating Aggressiveness, using underuti-lized public information. Using negotiating aggressiveness as the depen-dent variable, I create an estimating equation aimed to measure the effect of duality on CEO negotiat-ing effort. I find that CEO duality has a positive and statistically significant relationship with nego-tiating aggressiveness at the 1% level. I reject the null hypothesis, which states that duality leads to lower levels of negotiating aggressiveness, and thus lower value maximization. I attribute the lack of sig-nificance among my control variables to the small sample size, which is common among studies using hand-collected measures. These conclusions bear implications for corpo-rate governance. Stricter governance does not nec-essarily lead to higher value maximization, at least in the case of negotiations. I believe the added free-dom afforded by duality creates a more conducive atmosphere within the firm for successful negotia-tions. While my results imply that leadership struc-ture matters in M&A dealmaking, the specific aspect of duality that drives these findings remains unclear. Future research can hopefully explore and identify the exact mechanism behind this relationship.

V. References Adams, R., Hermalin, B., & Weisbach, M. (2010). The Role Of Boards Of Directors In Corporate Gov-ernance: A Conceptual Framework And Survey. Journal of Economic Literature, 58-107.

Aktas, N., Bodt, E., & Roll, R. (2009). Negotiations under the threat of an auction. Journal of Financial Economics, 241-255.

Bhagat, S., & Bolton, B. (2007). Corporate Gover-nance And Firm Performance. Journal of Corporate Finance, 257-273.

Boone, A., & Mulherin, J. (2004). How Are Firms Sold? The Journal of Finance, 847- 875.

Page 20: Visible Hand

20 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

ernance. Harvard Law School John M. Olin Center for Law, Economics and Business Discussion Paper Series, 488, 1-28.

Sampson-Akpuru, M. (1992). Is CEO/Chair Duality Associated with Greater Likelihood of an Interna-tional Acquisition? Michigan Journal of Business, 81-97.

Shleifer, A., & Vishny, R. (1988). Value Maximiza-tion and the Acquisition Process. Journal of Eco-nomic Perspective, 7-20.

Smith, A. (1976). An inquiry into the nature and causes of the wealth of nations. E. Cannan (Ed.). Chicago: University of Chicago Press. (Original work published 1776)

Sridharan, U. & Marsinko, A. (1997). CEO Duality In The Paper And Forest Products Industry. Journal Of Financial And Strategic Decisions, 10(1), 59-65.

Subramanian, G. (2011). Dealmaking: New deal-making strategies for a competitive marketplace . New York: W.W. Norton & Co.

Williamson, O. (1985). The economic institutions of capitalism: Firms, markets, relational contracting. New York: Free Press.

VI. Footnotes

1. From the perspective of the law, substantive is defined by the concept of materiality, which is de-fined as “the magnitude of an omission or misstate-ment of accounting information that, in the light of surrounding circumstances, makes it probable that the judgment of a reasonable person relying on the information would have been changed or influenced by the omission or misstatement” (FASB, 1980).2. CEO duality is measured as an indicator variable that takes a value of one if the CEO of the company is also the Chairman of the Board.

3. CEO Negotiating Aggressiveness, the main vari-able I construct in this paper, is defined in the Meth-odology section.

4. Summary statistics on each variable used in Table 1 are provided in the Data section.

For appendices and other notes, please refer to the electronic issue at http://orgsync.rso.cornell.edu/

nal, 37(5), 1079-1108.

Finkelstein, S., Hambrick, D., & Canella Jr., A. (2009). Strategic leadership: Theory and research on executives, top management teams, and boards. New York: Oxford University Press.

Gentry, M., & Stroup, C. (2015). Entry and Compe-tition in Takeover Auctions. 1-57.

Grinstein, Y., & Hribar, P. (2003). CEO Compensa-tion And Incentives: Evidence From M&A Bonuses. Journal of Financial Economics, 119-143.

Hansen, R. (2001). Auctions of companies. Eco-nomic Inquiry, 30-43.

Hartzell, J., Ofek, E., & Yermack, D. (2003). What’s In It for Me? CEOs Whose Firms Are Acquired. Re-view of Financial Studies, 37-61.

Hermalin, B., & Weisbach, M. (2003). Boards of Directors as an Endogenously Determined Institu-tion: A Survey of the Economic Literature. FRBNY Economic Policy Review, 7-26.

Iyengar, R., & Zampelli, E. (2009). Self-selection, endogeneity, and the relationship between CEO du-ality and firm performance. Strategic Management Journal, 1092-1112.

Jensen, M., & Meckling, W. (1976). Theory of the firm: Managerial behavior, agency costs and own-ership structure. Journal of Financial Economics, 3(4), 305-360.

Lam, K., Mcguinness, P., & Vieito, J. (2011). CEO gender, executive compensation and firm perfor-mance in Chinese‐listed enterprises. Pacific-Basin Finance Journal, 1136-1159.

Malmendier, U., & Tate, G. (2003). Who Makes Ac-quisitions? CEO Overconfidence and the Market’s Reaction. 1-64.

Oh, K., Pech, R., & Pham, N. (2014). Mergers and Acquisitions: CEO Duality, Operating Performance and Stock Returns. 1-34.

Rechner, P., & Dalton, D. (1991). CEO duality and organizational performance: A longitudinal analy-sis. Strategic Management Journal, 12, 155-160.

Roe, M. (2004). The Institutions of Corporate Gov-

Page 21: Visible Hand

Fall 2015 | The Visible Hand | 21

Volume xxiV. issue i.

stewards of the firm. Proponents of duality contend that combining the CEO and Chair positions pro-vides a company with a unified command structure and a consistent leadership direction (Chen, Lin, & Yi, 2008). This means the firm incurs fewer costs in decision-making. Jensen & Meckling (1976) and Brickley & Coles (1997) contend that the agency cost of separating CEOs in dual roles is higher than the net benefits gained. Oh, Pech, & Pham (2014) found that duality has a significant and positive effect on the decisions made by Vietnamese firms in mergers, primarily in terms of long-term value added. Hermalin & Weis-bach (2003) argue that board size, an indication of corporate governance strength, is negatively related to corporate performance. This finding supports the idea that top-management does not need to be con-stantly monitored in order to maximize shareholder value (Boyd, 1995). Sridharan & Marsinko (1997) investigated the impact of CEO duality on firm value in the paper and forest products industry and found that dual firms possessed higher market values than firms with baseline leadership structures. Under this line of thinking, duality in firms may also lead to enhanced collegiality and collaboration between board directors and company executives, facilitat-ing smooth and resourceful decision-making.

Correlation Matrix

VI. Appendix

Literature Review

Agency Theory Agency theory predicts that dual CEOs cannot maximize shareholder value because they are not monitored effectively. The prevailing view behind this theory is that CEOs are profoundly selfish and will always put the maximization of their utility over that of the firm and its stakeholders. Proponents of agency theory believe that without proper monitor-ing structures, CEOs pursue selfish interests that are inconsistent with their responsibilities to sharehold-ers (Williamson, 1985). In fact, Adam Smith (1776) theorized:

“The directors of [joint stock] companies, however, being the managers rather of other people’s money than of their own, it cannot well be expected, that they should watch over it with the same anxious vigilance [as owners].... Negligence and profusion, therefore, must always prevail, more of less, in the management of the affairs of such a company.”

Eisenhardt (1989) found that M&A takeover decisions are opportunistic and driven by CEO self-interest unless properly monitored. Such opportu-nistic behavior may include shirking of day-to-day responsibilities and monetary indulgences at the expense of shareholders (Roe, 2004; Williamson, 1985). The model of the opportunistic and individu-alistic economic actor, who is primarily concerned with the maximization of his or her own economic gain, can be traced to the field of organizational psychology through McGregor’s (1960) Theory X (Donaldson & Davis 1991). Daily and Dalton (1994) also found a signifi-cant positive association between CEO duality and firm bankruptcies. Fama and Jensen (1983) warn that allowing duality signals to shareholders that the corporation has failed to separate its decision man-agement from its decision control, which increases the likelihood that the CEO-Chair will take inef-ficient and opportunistic actions that deviate from shareholder interests and reduce shareholder wealth.

Stewardship Theory Stewardship theory predicts that dual CEOs en-hance firm performance and maximize shareholder value to a larger degree than baseline CEOs. The underlying view backing this theory portrays CEOs as benevolent and selfless leaders. The more respon-sibility they are entrusted with, the better they are as

Page 22: Visible Hand

22 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Ziyi YanBryn Mawr College

The Effect of Driving Restrictions on Air Quality in Beijing

I. Introduction

Rapid economic growth and urbanization have dramatically changed China’s transportation, espe-cially in the major cities. For example, in China’s capital, Beijing, unlike before, residents are not only traveling longer distances, but also making more trips and relying more on motorized modes. Rapid motorization has contributed to a series of problems including air pollution, oil price hikes, congestion, and growing greenhouse gas emissions. In Beijing, peak-hour speeds on urban arterials and express-ways often drop below 15 or even 10 kilometers per hour and it is on the list of the World Health Organi-zation’s most polluted cities (Sun, Zheng and Wang, 2014). To address the problems brought on by the ex-ploding growth of car ownership, Beijing has adopt-ed a wide range of policies, including investments in public transportation, ratcheting up vehicle emis-sion standards, as well as restrictions on both driv-ing and new vehicle purchases. Among them, the policy of restricting driving is considered to be an efficient way to alleviate the traffic pressure in Bei-jing and it was first introduced in 2008. Beijing im-plemented the odd-even license plate policy of road space rationing during the 2008 Olympic Games. (BBC, 2008) Due to the success in improving the air quality and the increased road space availability, the government issued a modified version of road space rationing, end-number license plate policy af-ter the 2008 Olympic Games. The odd-even license plate policy limits the use of cars on the alternative days by the parity of the end-number of license plate that if the last digit of license plate is odd, the car is only allowed to be used on the day which has date that is odd and if the last digit of license plate is even, the car is only allowed to be use on the day which has data that is even. For example, if a car has end-number of 7, it is only allowed to be used on 1st, 3rd, 5th, 7th…days of a month and if a car has end-number of 2, it is only allowed to be used on 2nd, 4th, 6th, 8th…days of a month. Similarly, the

end-number license plate policy limits every car on only one day of a week based on the end number of its license plate. If a car has an end-number of 1 or 6, it is only allowed to be used on the Monday of that week. If a car has an end-number of 2 or 7, it is only allowed to be used on the Tuesday of that week, etc. The main objective of the restraint policies is to re-duce the amount of exhaust gas generated by motor vehicles and alleviate the traffic pressure. Moreover, in consideration of the rapid increase in the number of cars, the government implemented a policy that restricts the purchase of small passen-ger cars, started from January 2011. In 2010, the av-erage monthly increase in registered new cars was 66,000, and given the increase, the car ownership is expected to hit 6 million before 2016 (Mu, 2012). The new policy requires the citizens who wish to purchase passenger cars with less than five seats to follow the small passenger car purchase policy to be applicable for purchasing a passenger car. Accord-ing to the policy, the individual purchaser must not already have a passenger car registered under his or her name, and must fulfill various requirements such as having a driving license and living in Beijing; if the purchaser fulfills all of the requirements, he or she could apply for a quota, and then wait for the monthly license plate ‘lottery’. During the 26th of every month the Traffic Management Bureau would take all of the eligible quotas and select a certain amount of them randomly, similar to the way of lot-tery where numbers are drawn randomly. Usually the lottery rate is under 10% (Yang, Liu, Qin and Liu, 2014). In China, pollution is one aspect of the broader topic of environmental issues and as China indus-trialized, various forms of pollution have increased, including air pollution, which is damaging the health condition of Chinese People. According to the stud-ies by World Bank, World Health Organization, and the Chinese Academy for Environmental Planning on the effect of air pollution on health, between 350,000 and 500,000 Chinese die prematurely each year because of air pollution1. The hazardous level

Page 23: Visible Hand

Fall 2015 | The Visible Hand | 23

Volume xxiV. issue i.

of air pollution has urged the government to make efforts to fight the pollution and as a city, which underwent serious condition of ‘airpocalyse’ since 2013, Beijing has desired adequate policies to effec-tively reduce air pollution. Since vehicles have al-ways been claimed as one of the major sources of air pollution, the government has aimed to reduce the level of air pollution by controlling the rapid growth in the use of vehicles, using the different kinds of policies that restrict driving. In this paper, I will evaluate the effect of the driv-ing restrictions on improving the air quality, during the time period from 2008 to 2013. Among the vari-ous measurements of air quality, the level of PM2.5 measures the level of tiny particles that have a width smaller than 2.5 micrometers. PM 2.5 is a type of air pollutant that is a concern for people’s health when levels are high and high levels of PM 2.5 reduce visibility and cause the air to appear hazy. Since these particles are relatively small, they are able to travel deeply into the respiratory tract, reach-ing the lungs. Exposure to PM2.5 can cause health effects such as eye, nose, throat and lung irritation, coughing, sneezing, running nose, and shortness of breath. Studies also suggest that long-term ex-posure to PM2.5 may be associated with increased rates of chronic bronchitis, reduced lung function and increased mortality from lung cancer and heart disease2. Besides, vehicle pollution is found to con-tribute about 22% of PM 2.5 in Beijing3. Therefore, based on this particularity of PM 2.5, in this paper, I will use the level of PM 2.5 as the measurement of air quality and my working hypothesis is: during the time period 2008-2013, policies that target to reduce car pollution, including the restriction on car pur-chases and road space rationing, help to improve the air quality in Beijing, with the indicator as PM2.5. In this paper, Section 2 offers a review of pre-vious studies which examine driving restriction’ effects. Section 3 introduces the description of the empirical strategy used, along with the summary of the data. Section 4 presents the results of the regres-sion and the interpretation. Section 5 discusses the reason and explanation behind the results from Sec-tion 4. Section 6 concludes the paper with policy recommendation, major limitations and interest for future research.

II. Literature Review

Chile is the first country that adopted driving restriction policy that in 1986, it implemented the policy Restriction Vehicular in its capital, Santiago,

to prohibit vehicles to be used on certain days based on the vehicle’s plate number. Since then, similar policies have been used in several large cities across the world. Sun et al. (2014) points out that the popu-larity of these policies is attributed to the policy’s simplicity, perceived quick effects on congestion and air quality, and low cost to regulators. To evaluate the effect of car-related policy on air quality, it is important to know what factors de-termine the traffic emissions in the air. Besides the amount of cars in use, both Viard and Fu (2011) and Sun et al. (2014) state that the determinants of traffic emissions include but not limited to: temperature, humidity, wind speed, supply and cost of public transportation, and fuel price, Furthermore, accord-ing to the geographical location of Beijing, wind direction can also affect the level of air pollution. There are also a small number of empirical studies on the effects on air quality of driving restrictions in some major cities in the world. Eskeland and Feyzi-oglu (1997) shows that Mexico City’s road rationing policy adopted since 1989, Hoy No Circula, did not lead to significant decreases in gasoline demand or car ownership in Mexico City during the period of 1984-1993. Davis (2008) also investigates the case of Hoy No Circula and it compares pollution levels before and after the implementation of the Hoy No Circula to measure the effect of the implementa-tion on air quality, controlling for covariates such as weather condition. It points to behavioral responses as the most likely reason of the Hoy No Circula’s failure to improve air quality. Davis (2008) also concludes that Hoy No Circula led to an increase in the total number of vehicles in circulation through increased household ownership of second cars, as well as a change in fleet composition toward high-emissions vehicles. Lin et al. (2011) focuses on multiple cities, in-cluding Sao Paulo (road rationing policy Operaco Rodizio adopted in 1995), Bogota (road space ra-tioning policy Pico y Placa adopted in 1998), Beijing (the odd-even road space rationing policy adopted in 2008 and end-number road space rationing policy adopted after 2008 Olympics Games) and Tianjin (the odd-even road space rationing policy adopted in 2008) and it indicates no evidence support that over-all air quality has been improved, across different versions of driving restrictions. Nonetheless, they conclude that the effect of the driving restrictions in Sao Paulo and Bogota is primarily on the daily maximum concentrations of CO and PM 10. It also suggests that due to the temporal and spatial shifting of driving, driving restrictions can only be expected

Page 24: Visible Hand

24 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

to alleviate air pollution when implemented with an extended schedule or in an extended schedule that the coverage should be beyond peak hours and beyond city center. For the analysis of Beijing, Lin et al. (2011) uses daily concentration of inhalable particulate matters (PM 10) based on the Air Pollu-tion Index (API) from the Ministry of Environmen-tal Protection for the time period from July 2007 to October 2009. It finds that the odd-even license plate policy during the 2008 Beijing Olympics were associated with at least 38% reductions in PM 10 concentrations, while there is no evidence that after the Olympics, the end-number license plate policy have improved air quality, with the measurement as PM 10 concentration. However, the effects are debatable that research-ers reach to different conclusions by using different data, length of time windows and different orders of polynomial time trend. Salas (2010) argues that reasonable changes in the method used, such as different time windows and polynomial orders can dramatically alter the conclusion that using non-parametric estimates, it identifies a 12-18% reduc-tion in air pollution during the first months of the implementation of Hoy No Circula, followed by a gradual increase in pollutant concentration. Viard and Fu (2011) find that Beijing’s API fell 19% dur-ing the odd-even license plate policy and it fell 8% during the end-number license plate policy. They back their causal inference with the spatial variation in air quality changes that larger changes are found at monitoring stations closer to urban expressways. Furthermore, Sun et al. (2014) points out that the complicated chained process from traffic emis-sions to air pollutant concentrations can be affected by numerous time-varying variables that some of the them may be relatively hard for researchers to include in the study, such as supply of public trans-portation, fuel price, parking cost and taxi fare, etc. All of these factors can potentially enlarge or shrink the “discontinuity” identified around the single time point when a driving restriction scheme is imple-mented. Moreover, there is literature that investigates the effect of driving restrictions on air quality in Beijing, with indictor as PM 10. Sun et al. (2014) concludes although the policies that target to restrict driving have positive effect on alleviating the traffic pressure in Beijing from July 2008 to October 2011, using PM 10 as an indicator of air quality, restricting driving in Beijing with end-plate number policy has little, or even negative impact on air quality. There-fore, the effect of the driving restrictions in Beijing

still seems debatable.

III. Data and the Overview of the Empirical Strategy

i. Data and Variables: Previous literature has discussed the determi-nants of air quality and researchers have identified a number of variables as important factors in the change of air quality indicators. Some variables they have identified include temperature, humidity, wind speed, and precipitation (Onursal & Gautam, 1997). Among them, temperature affects the speeds of evaporation, humidity affects aerosol formation and dust re-suspension, wind speed affects pollutant dilution, and dust re-suspension and precipitation affects wet deposition, or washing-out, of air pol-lutants. (Onursal & Gautam, 1997) With indicator as PM 10 to study the effect of restricting driving on traffic and air quality, Sun et al. (2014) use a set of determinants, including daily mean temperature, daily mean humidity, daily mean wind speed and the presence of policy that restricts driving. Consider-ing the similarity between the sources of PM 10 and PM 2.5 in Beijing, which both have approximate source apportionment including soil dust, automo-bile, secondary source, coal combustion, biomass burning and industrial sources (Zhang, Guo, Sun, Yuan, Zhuang, Zhuang and Hao, 2007), with indi-cator as PM 2.5 in Beijing, in this paper, the set of independent variables is as follows: daily mean tem-perature in Beijing (in Celsius), daily mean humid-ity in Beijing (in %), daily mean wind speed in Bei-jing (in km/h), the presence of policies on restricting car purchases, the presence of polices on road space rationing of end-number license plate, the presence of policies on road space rationing of odd-even number license plate. Here, in consideration of the complexity and the limited availability of the data of wind direction, the wind direction data is obviated. Therefore, in this paper, I use a time series data set that includes the following variables during the time period 2008-2013. I obtain the daily meteoro-logical variables from an online weather base4 and the daily PM 2.5 data from the website of the United States embassy in Beijing5 The data of the presence of the policies is compiled from the website of Bei-jing Traffic Management Bureau6. The dependent variable is the daily average hour PM2.5 concentration (µg/m3) in Beijing. The in-dependent variables are daily mean temperature in Beijing (Celsius), daily mean humidity in Beijing (%), daily mean wind speed in Beijing (km/h), dum-

Page 25: Visible Hand

Fall 2015 | The Visible Hand | 25

Volume xxiV. issue i.

my variable for the presence of policy on restricting car purchases that it is 1 if the policy is being imple-mented at the time and it is 0 if the policy is not be-ing implemented at the time, the presence of polices on road space rationing of end-number license plate such that it is 1 if the policy is being implemented at the time and it is 0 if the policy is not being imple-mented at the time, and the presence of policies on road space rationing of odd-even number license plate such that it is 1 if the policy is being imple-mented at the time and it is 0 if the policy is not being implemented at the time.

ii. Model In this paper, I use a multiple regression model to evaluate the effect of driving restrictions on air qual-ity, with the indicator as PM 2.5 in Beijing. The basic regression model can be written as a lin-ear specification of the following form:

PM 2.5i = α + β1*Temperaturei+ β2*Humidityi + β3*Wind Speedi + β4*Odd Eveni + β5*End Numberi + β6*Purchasei + ei

Where the subscript “i” stands for the time index.The summary statistics for the variables in the re-gression model is presented in Table 1. The independent variables used in the model are grouped into two categories, meteorological variables including temperature, humidity and wind speed, and the other category including the policies that restrict driving. First, the group of meteorological variables is going to be considered. Tai et.al (2010) conclude

that daily variation in meteorology including nine predictor variables (temperature, relative humid-

ity, precipitation, cloud cover, 850-hPa geopotential height, sea-level pressure tendency, wind speed, E-W and N-S wind direction) can explain up to 50% of the daily PM 2.5 variability in US. Addi-tionally, humidity is positively correlated with the level of PM 2.5, while wind speed is negatively cor-related with the level of PM 2.5. It also discusses the effect of temperature on PM 2.5 concentration and that the influence is correlated to other aspects of the environment as well. Therefore, the effect of temperature on the level of PM 2.5 may not be simply determined, which is to say, the direction of the effect of temperature on PM 2.5 concentration may be ambiguous. In this paper, we anticipate the temperature has a positive effect on the air quality that as temperature increases, PM 2.5 concentration decreases. Then, the group of policies that target restrictions on driving is under consideration. Sun et.al (2014) point out that using the indicator as the level of PM 10 shows that although the most stringent driving restrictions had a positive impact on city-wide traf-fic speed, the marginal reduction in the number of usable vehicles may result in little, or even negative impact on air quality. This paper specifies the differ-ence between different types of driving restrictions, pointing out that the end-number license plate poli-cy, which is less stringent than the odd-even license plate policy, does not happen to have any significant positive effect on air quality, unlike the odd-even li-cense plate policy, which helps to both alleviate the traffic pressure and reduce air pollution. However, the implementation of end-number license plate pol-icy is reported to have a positive effect on reducing the use of cars as experts point out that this policy has reduced the number of cars on the public road-space in Beijing by 700,0007. Thus, according to the great amount of pollution contributed by vehicles, in this paper, the end-number license plate policy is expected to have a positive effect on air quality, which is to say that the policy is negatively corre-lated with PM 2.5 concentration. Similarly, based on the positive impact of odd-even license plate policy on the alleviation of traffic pressure and the air quality with indicator as PM 10 concentration (Sun, Zheng and Wang, 2014), odd-even policy is expected to be negatively correlated with PM 2.5 concentration. Yang et al. (2014) conclude that the policy on restricting purchases have significant effect on both vehicles sold and congestion. They point out that with this policy put in place alongside other poli-ices aimed at improving transporation, Beijing ex-

Page 26: Visible Hand

26 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

pericenced a sharp drop in the number of vehicles sold, immediate improvements in congestion, and increases in average driving speeds were also re-ported, which was considerated to be an evidence of the improving traffic condition as well. Accord-ing to the relationship between the use of vehicles and air quality, which suggests that cars are sup-posed to contribute a great amount of particles to air pollution, the policy of restricting car purchases is expected to be negatively correlated with PM 2.5 concentration. Based on the above analysis, the expected sign and the expected significance of each independent variable are presented in Table 2. To identify the presence of autocorrelation, I use the Durban-Watson test. I got the Durban-Watson d-statistic from Stata to be .9930988, which indicates that there exists positive serial correlation and that the PM 2.5 concentration is positively correlated with the PM 2.5 concentration for the day before. However, this serial correlation is easy explain. As

discussed before, the PM 2.5 concentration is close-ly correlated with wind speed. If there is little wind, it may be very hard for PM 2.5 particles to diffuse, instead, they will tend to stay in the air for a long time8. Therefore, if the PM 2.5 concentration is relatively high on one day, it is highly possible that some of the particles will stay in the air during the following day, which may lead to a continued high PM 2.5 concentration. Hence, to correct for the serial correlation in the data, I choose to use the first order autoregressive (AR(1)) model in this paper. The AR(1) model is a linear regression of the current value of the time series on the previous value. Thus, with the linear specification of the regression model that mentioned before as the basis, the difference equation of the regression model is expressed as

PM 2.5i = PM 2.5i-1 + ei

Where the subscript “i” stands for the time index.

IV. Regression Estimates

The estimation results of the regression model are presented in Table 3, and with cut-off level of 0.05 for the p value. Variables have coefficient estimates that are labeled in bold if significant. The regression has an adjusted R-squared to be 0.3548, which is acceptable for the time series data we have and it indicates that the regression model explains a great amount of variation in the data. With the cut-off significance level as 0.05, we have the following interpretations from the results in the above table. First, we see that with everything

else unchanged, for one standard deviation increase in the daily average temperature, the PM 2.5 con-centration is expected to decrease by 6.29 µg/ m3, for one standard deviation increase in daily average humidity, the PM 2.5 concentration is expected to increase by 38.71 µg/ m3 and for one standard devi-ation increase in daily average wind speed, the PM 2.5 concentration is expected to decrease by 14.54 µg/ m3. Besides, with everything else unchanged, compared to the time when the odd-even license plate policy is not being implemented, the imple-mentation of odd-even road space rationing policy leads to PM 2.5 concentration be 47.63 µg/ m3 lower. Similarly, with everything else unchanged, compared to the time when the end-number license plate policy is not being implemented, the imple-mentation of end-number road space rationing pol-icy leads to PM 2.5 concentration be 22.87 µg/ m3 higher. On the other hand, with p value of 0.437, which is larger than the cut-off significance level, 0.05, the coefficient of restrictions on purchases is considered to be insignificant. Therefore, according to the estimates, the implementation of the policy on restricting car purchases does not have significant

Page 27: Visible Hand

Fall 2015 | The Visible Hand | 27

Volume xxiV. issue i.

effect on improving air quality. Thus, we conclude that, in general, during the time period 2008-2013, the odd-even road space ra-tioning policy has a positive effect on improving air quality in Beijing, the end-number road space ra-tioning policy has a negative effect on improving air quality in Beijing, while the policy of restricting car purchases does not have any significant effect on improving air quality.

V. Explanation for the Discrepancy and Unexpected Results

Compared to the expected signs of every inde-pendent variable, the discrepancy and unexpected results mainly come from the sign of end-number li-cense plate policy. Since the end-number road space rationing policy is implemented in order to allevi-ate the traffic pressure and better traffic condition is usually correlated with better air quality, the end-number policy is expected to have positive effect on air quality. However, the regression estimates show that it has a significant negative effect on air qual-ity and under the implementation of the end-number policy, the PM 2.5 concentration significantly in-creases by 22.873 µg/ m3. This may be explained by two reasons. First, about 90% of the observations we obtain are during the time period when the end-number li-cense plate policy is implemented, which may cre-ate bias toward the results because there are not suf-ficiently enough observations from the time period during which the end-number license plate policy is not implemented. Second, although the end-number license plate policy helps to reduce the amount of traffic, which is mentioned the previous section, Sun et. al (2014) point out that this alleviation may be offset by the consistently increasing level of car ownership in Beijing. Meanwhile, there is an unexpected result. Since Beijing experienced a sharp drop in the number of vehicles sold, there were immediate improve-ments in congestion and increases in average driv-ing speeds under the implementation of policy on restricting car purchases (Yang, Liu, Qin and Liu, 2014). Thus, it is expected that the policy of restrict-ing car purchase helps to improve air quality. How-ever, the regression results show that the effect of restriction on car purchases on improving air qual-ity is insignificant. Similar to the reason behind the inefficiency of end-number license plate policy, this insignificance may be explained by the consistently

increase in car ownership in Beijing. As indicated, the policy of restricting car pur-chase was introduced and implemented at the beg-ging of 2011. According to statistics, there are 5.017 million cars registered in Beijing as of Feb 2012, with 173,000 new ones registered in 2011, accord-ing to Beijing Traffic Police Bureau9. This shows that the mandatory lottery for people who wish to buy cars helps to reduce the number of cars bought with a lottery rate less than 10%, because without the lottery, with the annual growth to be 13%, the number of cars registered in 2011 would be ex-pected to be much higher (Yang, Liu, Qin and Liu, 2014). However, this restriction is still insufficient that compared to the statistics by the end of 2010, which there are 4.8 million cars registered (Yang, Liu, Qin and Liu, 2014), the net number of cars reg-istered increased by 0.2 million during 2011. This indicates that in aspects of improving air quality by reducing the total number of car ownership in Bei-jing, the policy of restricting car purchases has been insufficient.

VI.Conclusion

i. Conclusion Corresponding to the Working Hypothesis

Corresponding to the working hypothesis intro-duced in Section 1, I conclude that during the time period 2008-2013, the odd-even road space ration-ing policy has a positive effect on improving the air quality in Beijing, the end-number road space rationing policy has a negative effect on improv-ing the air quality in Beijing, while the policy on restricting car purchases has no significant effect on improving the air quality in Beijing, with indicator as daily PM 2.5 concentration.

ii. Policy Recommendation Based on this conclusion, I would advise gov-ernments that if they desire significant improve-ment in air quality, it is not sufficient to implement end-number road space rationing policy or policy of restricting car purchases. In this paper, we have seen an unexpected negative effect of end-number license plate policy on improving air quality and an insignificant effect of policies of restricting car purchases on improving air quality. Meanwhile, we have seen a significant positive effect of odd-even license plate policy on improving air quality. Therefore, for governments that aim to reduce air pollution by implementing policies on restricting policies, instead of choosing end-number license

Page 28: Visible Hand

28 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

plate policy or policy of restricting car purchases, they should consider a more rigorous version of road space rationing, such as odd-even license plate policy to achieve the desired effect.

iii. Limitations and Future Interest There are some major limitations in my paper. First, since the data I obtained are in daily form, based on the limited accessibility of daily forms for most of data, I was forced to exclude some variables that may also have an effect on PM 2.5 concentra-tion, which may include the amount of factory pol-lution, precipitation, wind direction, cost of public transportation, parking cost, etc. (Sun, Zheng and Wang, 2014) Without these variables that may also have an impact on the dependent variable, the re-gression results may involve omitted variable bias that affect the coefficient interpretation. Second, the PM 2.5 concentration data is only avaiable during the time period 2008-2013 that, as mentioned before, there are not sufficiently enough obervations for the time period when the end-num-ber license plate policy is not implemented. This may lead to a bias in the regression results as well.Last, there are some missing data that I have to ex-clude from my regression, which may also affect the regression estimates. For some dates, I do not have the PM 2.5 data for some periods during the day, therefore I was forced to exclude the date from my regression. An interesting area for future research may be to look at whether the combination of policies on re-stricting driving has better effects on improving air quality. This paper investigates how the three differ-ent policies on restricting driving work to improve the air quality in Beijing individually, but does not discuss the effect on the air quality when some of them are combined to be implemented. Thus, un-der the hypothesis that a combination of policies on restricting driving works better to improve the air quality in Beijing than implementing them individu-ally, an empirical study can be launched to investi-gate this topic.

VII. References

Ardekani, Hauer and Jamei. (2011) Traffic impact models. Transportation Research Board. Washing-ton, D.C. Chapter 7.

Chen, Jin, Kumar and Shi. The promise of Beijing: Evaluating the impact of 2008 Beijing Olympics Games on air quality. NBER Working Paper No.

1690, March 2011.

Davis. (2008) The effect of driving restrictions on air quality in Mexico City. Journal of Political Econ, Vol.116, No.1.

Economy, Elizabeth C. (2007) The great leap back-ward? The costs of China’s environmental crisis. Foreign Affairs, Vol.86, No.5.

Han, Yang and Wang. (2010). Efficiency of the plate-number-based traffic rationing in general net-works. Transportation Research, Vol.46, No.6.

Hu, Tang, Peng, Wang, Wang and Chai. Study on characterization and source appointment of atmo-spheric particulate matter in China. Environment and Sustainable Development, Vol 36, No

Lieberthal, Kenneth. (1997) China’s governing sys-tem and its impact on environmental policy imple-mentation. China Environment Series 1. Washing-ton, D.C. Woodrow Wilson Center.

Lin, Zhang and Umanskaya. (2011) The effects of driving restrictions on air quality: Sao Paulo, Bo-gota, Beijing and Tianjin. 2011 Agricultural and Ap-plied Economics Association Annual Meeting, July 24-25, 2011.

Onursal & Gautam. (1997). Vehicular air pollution experiences from Seven Latin American urban cen-ters. Washington, DC: World Bank Group.

Salas, C. (2010) Evaluating public policies with high frequency data: Evidence for driving restric-tions in Mexico City revisited. Catholic University of Chile Working Paper No.374.

Sun, Zheng and Wang. (2014). Restricting driving for better traffic and clearer skies: Did it work in Beijing? Transport Policy, Vol.32, 34-41.

Viard and Fu. (2011) The effect of Beijing’s driv-ing restrictions on pollution and economic activity. SSRN working paper id2029446.

Wang, H. (2002) Pollution regulation and abatement efforts: Evidence from China. Environment and De-velopment Economics, Vol. 41.

Wang and Xie. (2009) Assessment of traffic-related air pollution in the urban streets before and during

Page 29: Visible Hand

Fall 2015 | The Visible Hand | 29

Volume xxiV. issue i.

the 2008 Beijing Olympic Games traffic control pe-riod. Atmospheric Environment, Vol.43, No.35.

Wang, Zhao, Xing, Wu, Zhou, Lei, He, Fu and Hao. (2010) Quantifying the air pollutants emission re-duction during the 2008 Olympics Games in Bei-jing. Environmental Science Technology,Vol41, No.7.

Xu, Gao and Chen. (1994). Air pollution and daily mortality in residential areas of Beijing, China. Ar-chives of Environmental Health: An International Journal, Vol.49, No.4.

Yang, Liu, Qin and Liu. (2014). A Review of Bei-jing’s vehicle lottery: Short-term effects on vehicle growth, congestion, and fuel consumption. Envi-ronment for Development Discussion Paper Series, January 2014.

Zhang, Guo, Sun, Yuan, Zhuang, Zhuang and Hao. (2007). Source apportionment for urban PM 10 and PM 2.5 in the Beijing area. Chinese Science Bul-letin, 608-615.

VIII. Footnotes

1. The statistics come from http://www.telegraph.co.uk/news/worldnews/asia/china/10555816/Chi-nas-airpocalypse-kills-350000-to-500000-each-year.html.

2. The information of PM 2.5 comes from https://www.health.ny.gov/environmental/indoors/air/pmq_a.htm.

3. The statistics come from http://www.huffington-post.com/2013/01/31/china-pollution-cars-air-prob-lems-cities_n_2589294.html.

4. Daily meteorological variables are obtained from http://www.weatherbase.com/weather/weather-hourly.php3?s=11545&date=2015-04-24&cityname=Beijing%2C+Beijing%2C+China&units=. 5. Hourly PM 2.5 concentration is obtained from http://www.stateair.net/web/historical/1/1.html.

6. The information of the implementation of poli-cies on restricting driving is obtained from http://www.bjjtgl.gov.cn/.

7. The statistics is obtained from http://news.163.com/09/0527/22/5ABT48O1000120GR.html.

8. The information is obtained from https://www.health.ny.gov/environmental/indoors/air/pmq_a.htm.

9. This statistics is obtained from http://news.xinhua-net.com/english/china/2012-02/16/c_122713279.htm.

Page 30: Visible Hand

30 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Sylvia Klosin and Cameron Taylor University of Chicago

Parental Employment and Childhood Obesity

I. Introductory Remarks

Obesity has more than doubled in children and quadrupled in adolescents in the past 30 years (CDC, 2015). Furthermore an increasing amount of attention has been placed on the rising obesity rates of American children. From Michelle Obama’s Let’s Move initiative to the reforms of public school lunches, there is a rally to lower obesity rates among children. In order to create policy that can effective-ly lower obesity rates, we need to examine factors that influence childhood obesity. Thus, we use the 2005 MEPS to examine the impacts of all relevant covariates and parental employment on obesity. Moreover, we expand our analysis to include gen-eral deviations from healthy weight. We do not re-strict ourselves to looking at whether or not children are obese, and we think a general approach such as this provides a fresh perspective on this well-covered area. While we care about the policy im-plications for understanding how to mitigate child-hood obesity, we are also curious how household dynamics, such as parental employment, affect de-viations from healthy weight and being overweight. We want to explore how different uses of time by parents can affect outcomes in weight for children. Overall, we think this paper provides a solid inquiry into important and policy-relevant questions related to household and health economics.

II. Literature Review

There are several studies that have shown the impact of parental employment on the health and obesity of children, which provides information about dependent and main independent variables in our study. Benson (2010) is one of such studies, looking at the impact of parental employment on childhood obesity. She used The Child Develop-ment Supplement (CDS) of the Panel Study of In-come Dynamics (PSID) for her analysis and found that there is a noteworth negative relationship be-tween maternal employment and child BMI. Fur-thermore she found that paternal employment also plays a significant role in child BMI. According to Benson, parent-child activities, like gardening and doing laundry, have an impact on childhood obesity through routines and parent-child interactions. We perform a similar analysis using the MEPS dataset. In another paper Araneo (2008) studies the impact of maternal employment on childhood obesity using The Fragile Families and Child Wellbeing Study, which focuses on low income minority children under the age of three. This study also examines if maternal race and education impact childhood obesity. We similarly use racial and educational factors in our approach, but also look at potential confounders such as family income. Our population is also more diverse than Araneo (2008) because we look at groups beyond minority groups and we

In this paper we study the impact of parental employment on the weight of children. We use the 2005 Medi-cal Expenditure Panel Survey (MEPS) as our source of data. We use several measures of healthy weight for children including body mass index (BMI), deviations from healthy BMI and an indicator for obesity running ordinary least squares and logit regressions. The hypothesis we are testing is whether or not pa-rental involvement with children improves outcomes related to weight. We use parental employment as our proxy for this involvement. We find that the employment status of the father significantly impacts children’s weight outcomes. We analyze this further and propose a causal mechanism for this result. Furthermore, we found that parental education, weight, race, sex and age of the child also had an impact on the weight of the children.

Page 31: Visible Hand

Fall 2015 | The Visible Hand | 31

Volume xxiV. issue i.

look at children of all ages. Thus, while this field of research is fairly dense, we can see that we are still uniquely contributing to this literature through our holistic approach. In another paper Anderson (2012) uses the Early Childhood Longitudinal Sur-vey-Kindergarten Class of 1998-1999 dataset in her research and presents an argument for a negative re-lationship between maternal employment and child obesity. She argues that maternal employment tends to disrupt positive routines like eating meals as a family, and these routines are significantly related to the probability of being obese. Our results are dif-ferent because we find that maternal employment is not statistically significant factor in determining child BMI. We also find that paternal unemploy-ment affects what we know about a child’s BMI, which means that the mechanism that Anderson proposed, i.e. having a parent at home is helpful to child health, does not manifest in our results. We provide a counter mechanism for our results in the discussion section.

III. Data

In our paper we use the MEPS from 2005, which is a set of large-scale surveys of families and in-dividuals, their medical providers, and employers across the United States. We use the Household Component (HC) component of the MEPS data, which collects data from a sample of families and individuals in selected communities across the United States. These individuals are drawn from a subsample of households that participated in the prior year’s National Health Interview Survey. In each interview, MEPS collects detailed informa-tion for all persons within a household. Therefore we have data on household variables as well as on the individuals within the household (MEPS, 2005). In order to make the results of our regressions clear we restricted our analysis to a subsample of the data. For one we limited our sample to two-parent mom and dad household, which led to 3691 obser-vations. To further clarify our results we removed observations of children with BMIs above 50, since such extremely high weights were probably due to a recording error. We also drop all observations of children and parents with a BMI less than 0, since it is not possible to have a negative BMI. While the MEPS dataset has a large number of variables, we needed to add many variables to perform our analysis. For our primary analysis, we defined log family income as the log of the sum of all adults’ income in the household. The second im-

portant variable we needed to recode was the race of the parent. This was slightly problematic since we needed to aggregate race of parent into a single household measure. Thus, we defined an “other” race dummy that took a value of 1 whenever a household had a parent of other (Alaskan, Native American or other in the dataset) race or had parents of different races. All other races (Black, Asian and White) were coded standardly. We needed to add paternal and maternal unemployment too, as this is one of the major focuses of our analysis. First, we identified moms and dads using survey results. Then, we used employment status of those adults to code these dummies. Finally, we needed to find a way to code a variable for parental BMI, since this will undoubtedly be correlated strongly with child BMI due to genetics. Thus, we took the average BMI of both parents in the household in our analy-sis. This ends up being a suitable metric because we restrict our sample size to those households that only have a mother and father. For education, we recoded education to be the max education of any parent in the household3. For our generalization, we needed to recode some dependent variables and code one of our depen-dent variables of interest: deviations from healthy weight. To obtain this metric, we subtracted the BMI of the child from the 50th percentile of the BMI based on age and sex and then took the abso-lute value of that result. These measures are in Table 5 of the Appendix. Finally, we created a variable in-dicating whether a child was obese or not. We did this based on the 95th percentile of BMIs based on age and sex (see Appendix Table 7)4. See Table 1 for relevant summary statistics.

Page 32: Visible Hand

32 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

IV. Methodology

For our analyses, we estimated three main mod-els. The first model is OLS with heteroskedasticity-robust standard errors regressing children’s BMI on parental education, log family income, maternal and paternal employment status, sex of the child, dum-mies for race, and parental weight and age of the child. The model is given by

where the parameters of interest are δD and δM. The second model is OLS with heteroskedastici-ty-robust standard errors regressing BMI deviations based on age on parental education, log family in-come, maternal and paternal employment status, sex of the child, dummies for race, and parental weight. The model is given by where

with the same parameters of interest: δD and δM.

V. Results

Refer to Tables 2, 3 and 4 for all of our regression results. Table 5 is the marginal effects of our logistic regression. In the first model (Table 2), we ran an OLS model in which we regressed child BMI on maximum edu-cation, log family income, paternal employment sta-tus, maternal employment status, sex, race, (black, asian, other), parental weight, and age of child. Our variables of interest were the employment status of the parents, but only paternal employment status was statistically significant. The coefficient on the father’s employment was -.765 and was significant at the 5% level. This means that having an employed father leads to a decrease of -.765 in a child’s BMI controlling for all other factors in our regression. In the next regression (Table 3), we ran an OLS model in which we regressed the deviation of the child’s BMI from healthy weight on maximum edu-

cation, log family income, paternal employment sta-tus, maternal employment status, sex, race (black, asian, other), parental weight, and age of child. Our variables of interest were the employment status of the parents, but only dad employment status was sta-tistically significant. The coefficient on dad employ-ment was -.686 and was significant at the 5% level. This means that having a father employed leads to a lessening of a deviation of -.686 from a child’s BMI from healthy weight controlling for all other factors. Lastly, we ran a logit model in which we re-gressed whether the child was obese or not on maximum education, log family income, paternal employment status, maternal employment status, race (sex, black, Asian), other, parental weight, and age of child (see Tables 4 and 5). Our variables of interest were the employment status of the parents, but neither paternal employment status nor maternal employment status were statistically significant.

VI. Discussion

After running the regressions we worked to un-derstand the causal relationships underlying these results. We wanted to understand why the education of the parents, paternal employments, income, pa-rental BMI, child age and race had significant im-pacts on the obesity of the children. To understand how significant changes in BMI are, we provide a short example. A girl who is 5 feet 7 inches that weighs 115 pounds will have a BMI of 18. In order to have a BMI of 19 she would have to weigh 121.5 pounds, so she would need to gain 6.5 pounds. We found that high education was associated with lower child BMI. This coefficient is highly statistically and economically significant, since one additional year of schooling leads to a reduction of about 0.25 BMI. Thus, the difference between a par-ent with a high school diploma and a parent with a college diploma is approximately 1 BMI, which is quite significant. As well, having a parent with an extra year of schooling decreases the probability that the child is obese by 2 percentage points, which aggregated over 4 years plus is quite substantial. We believe that this is due to the fact that parents who are more educated have more knowledge about nutrition and healthy lifestyles. They can then pass this knowledge onto their children, which leads to healthier BMIs. Furthermore, it is possible that par-ents that have had more education are surrounded by other people who plan to be parents and practice positive parenting techniques, and thus there will be positive externalities associated with spending more

Page 33: Visible Hand

Fall 2015 | The Visible Hand | 33

Volume xxiV. issue i.

Page 34: Visible Hand

34 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

time in school. We suspect that paternal employment has a more complicated causal mechanism than education. We found that having a father at home made the chil-dren gain weight, but logit models revealed that having father at home was not a significant factor in increasing their chance of being obese. In the two OLS models, the coefficient on father’s employ-ment is around -0.7, meaning that children who have an unemployed father are associated with a 0.7 increase in their BMI, which on the margin is quite important. As well, it is associated with a less-ening of deviation from healthy weight by 0.7. The fact that having a father at home makes the children heavier (or more unhealthy in general) may be due to the fact that the father may not be as attentive to the physical activity and lifestyle choices of the chil-dren as a mother who spends time at home would. Cultural norms and the history of gender segrega-tion in work in the household in the States encour-age women to be good mothers and place the needs of their children first while same pressure does not exist to the same extent for fathers. This may lead to less attentive fathers. Another explanation for these results is that in the States it is still the norm for fathers to work and the mothers to stay at home. Therefore, it is likely that if a father stays at home it is because he cannot find employment. These men may be unemployed because they have unfa-vorable traits, like short tempers or impatience that lead them to be both unemployed and bad fathers. Indeed the inputs to being productive as a father and an employee are probably similar in many aspects – raising a child well takes dedication, hard work and significant investment. We could not control for factors such as these in this study, but we suspect there is a positive relationship between the employ-ability of a father and his parental skills. However, it is possible to argue that education could control for some of these factors since more educated parents are more likely to be better parents simply through education. The coefficient of interest for our other main variable, maternal employment, was not significant. We think that this is due to the fact that in the United States women still face a lot of pressure to be good mothers, and the health of their child is a large re-flection of their abilities as individuals. Therefore even if a woman works, she is under pressure to make sure that her children are well off. Hence, her employment will not change the BMI of her child, which explains why none of these coefficients were significant. The only place where maternal employ-

ment was statistically significant was in a logistic re-gression where we looked at the children who were underweight (1) or not (0). However, since the focus of this paper is obesity, we leave further discussion of this result to the Appendix (see Tables 13 and 14) and note that even though maternal employment is statistically significant, it is not very substantively significant. We also looked at the impact of log family in-come on the obesity of children. We found that there was negative relationship between the two. In our OLS regression we found that a one percent change in income leads to a -.291 change in BMI and a -.421 lessening in deviation. Similarly, an analogous change in income leads to a 2.5% decrease in the probability of the child being obese. This is likely because the wealthier the family the easier it is for them to afford healthy food and athletic activities for their children. Another important variable in our regressions that is noteworthy is race. We found that being black was significant in the first two regressions and the logit models. The coefficient was around 0.6 to 0.75 in the OLS regressions, meaning that children with black parents had a substantially higher BMI than those with white parents. As well, the logistic marginal ef-fects show us that children with black parents have a 4% higher risk of being obese. We think that this significance is due to genetic factors. Some current genetic literature argues that black people are genet-ically predisposed to obesity (Monda, 2013). This would explain our positive coefficient. Yet another important independent variable is parental BMI - it was significant in all the regres-sions that we ran. In our OLS regressions it had a coefficient of around 0.15, meaning that an increase in average parental BMI of 1 leads to an increase in 0.15 BMI of the child or variability in the deviations of the child by 0.15 BMI. This is not strikingly sig-nificant, but is still important in large disparities be-tween different household’s average BMI’s. There are both genetic and cultural factors that are cap-tured in this parameter. For one, parents will pass their genetic predisposition for obesity or unhealthy weight to their child. Second, parents with unhealthy life habits themselves will pass these patterns onto their children. While we would like to separate these factors, it is almost impossible in this study. Another significant variable was child age. Like we suspected there was a positive relationship be-tween age and BMI in the first OLS, since older children tend to, on average, have higher BMI’s (see Appendix Figures 2 and 3). Increasing a year

Page 35: Visible Hand

Fall 2015 | The Visible Hand | 35

Volume xxiV. issue i.

in age increases BMI by about 0.5 points, a substan-tial margin. This provided a necessary control in our regression. However, in the logistic regression, we saw that increasing age decreases the probabil-ity of having obesity by about 2 percentage points, which indicates the older children are less at risk for childhood obesity. This is surprising since BMI has greater variance with age, but may be due to harsher standards for defining obesity in younger children (see Appendix Figure 2 and 3). As well, it could be possible that as children get older, they become more independent and thus can change their own diet and exercise habits. Thus, pressure at school or from society to lose weight may cause them to change their diet and exercise habits, lowering their risk for obesity. We wanted to control for the sex of the child in our regressions as well. Though the coefficient in our first OLS regression was only significant at the 10% level, it was fairly high at -.2955. This means being that a girl leads to a decrease of -.295 in BMI. In the logit model, sex was highly signifi-cant and showed that girls are less likely to be obese by about 6 percent. These results are valid because the prevalence of obesity is higher among boys than girls (18.6% of boys and 15.0% of girls were obese) (NCHS 2012 ) and there are most likely social pres-sures for girls to be less heavy, leading them to be less at risk for being obese, whether by independent will or parenting.

VII. Limitations

While the results we have found are very interest-ing, there are potential endogeneity problems that we were not able to fully address in our research. The first we will address is the coefficient on the employment status of the father. As stated, the caus-al mechanism is problematic because those fathers who stay at home may be unemployed because of traits that affect both their employability and their parenting ability. Thus, we have a general omitted variable bias in controlling for and looking at par-enting ability. If available, we would have liked to have variables and measurements of some personal-ity traits of the fathers. This is very hard to measure and find in a dataset. Thus, even something as basic as crime records would help us understand the dy-namics of the father’s parenting style allowing us to uncover the effects of their involvement better. We also have potential for measurement error - some of the children’s BMI results that we found were so extraordinary that we did not include them

in the data we analyzed. Thus, we worried about general measurement of children’s BMI in the whole dataset, and other variables measured. Because this was survey-collected data, there is the potential for some surveyors to be more careless than others and report data differently. With measurement error in BMI, our coefficient estimators are still unbiased and consistent. However, if there is measurement error in any of the independent variables of inter-est, which we found no direct evidence for, then the problem is more substantial. We proceeded as if there is no such error as we found no evidence for it. Another potential problem was misspecification of functional form. When plotting the data, we not-ed that there was a funnel relationship between the two income measures (see Appendix Figure 1). For this reason, a linear relationship seemed somewhat unsuitable and this was part of the motivation for measuring deviations from BMI 50th percentile as opposed to just standard BMI. The results from this regression were fairly similar to the original specifi-cation, and thus we are not worried about the func-tional form. Finally note that this scatterplot also suggests heteroskedasticity. Thus, in all regressions, we used heteroskedasticity-robust standard errors (as discussed earlier). A general problem that we may have with our re-sults is that they are not very general. While we look at BMI levels, BMI deviations and obesity, there are many potential measures of child health and weight. Thus, as previously discussed, we also included a logistic regression looking at severely underweight children. However, as a robustness check for our claims about the effects of paternal employment on children’s health, we also regressed perceived health status (coded 1 as “Excellent” up to 5 as “Poor”) as an ordinal variable on all our covariates of interest (see Table 15 in the Appendix for results). In the re-sults, maternal employment gains significance, but paternal employment has more than double the sub-stantive significance of maternal employment. Thus, we believe our results and claims related to paternal employment remain robust. Because the substantive significance of mothers is quite small, we do not think our causal mechanism is jeopardized.

VIII.Conclusion

In our paper we found that parental employ-ment has an impact on childhood obesity and healthy weights of children. Specifically, having an unemployed father made children more obese and also affected deviations from healthy BMI ranges.

Page 36: Visible Hand

36 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Moreover, we saw that factors that we would ex-pect to be highly correlated with BMI, such as age of child, parental weight and race, are significant in explaining the BMI of the child. We found that fam-ily income and education of parents are significant in explaining BMI of children, too. Overall, these results are fairly consistent with the literature (Ben-son, 2010) as a whole. However, the uniqueness and singularity of the parental employment status effect result was novel, and allows for an interesting take on how we think parenting affects childhood obesity and healthy weight. It also is directly contradictory to Anderson (2012) who finds that mother employ-ment status matters. In this paper we presented al-ternative mechanisms that could explain this dif-ference. We would be interested in continuing this analysis with further controls for parental ability such as personality or legal records as it would al-low us to cure some of our most important endo-geneity problems and undertake the analysis with different data sets, as this would give us a better understanding of exactly how paternal employment status affects healthy weights of children.

IX. References

Anderson, Patty. “Parental Employment, Family Routines and Childhood Obesity.” National Center for Biotechnology Information. U.S. National Library of Medicine, 10 Dec. 2012. Web. 22 May 2015.

Araneo, Jackie. “The Effects of Maternal Employ-ment on Childhood Obesity in the United States.” Diss. Princeton U, 2008, Print.

Benson, Lisa. “The Role of Parental Employment in Childhood Obesity.” Digital Repository at the University of Maryland (DRUM), University of Maryland, 2010. Web. 23 May 2015.

“Childhood Obesity Facts.” Centers for Disease Control and Prevention, 24 Apr. 2015. Web.

Monda, Keri L. “A Meta-analysis Identifies New Loci Associated with Body Mass Index in Indi-viduals of African Ancestry.” Nature Genetics, 14 Apr. 2013. Web.

“NCHS Data on Obesity.” Centers for Disease Control and Prevention, pag. 12 Jan. 2012. Web.

X.Footnotes

1. Note that each observation is a child, not a household. 2.We also defined a log income per child variable, but saw that it did not perform as well as our log family income variable in the regression. These two have different economic interpretations: the difference is essentially between households who need to allocate income among children and house-holds who just have large income overall. See the Appendix Tables 7-10 for the regressions with log income per child.

3. This is suitable since higher education of one parent would probably have positive externalities on the other parent’s parenting. Thus, the highest education of the parent would bring the other par-ent’s parenting ability up.

4. We also created a variable for severly under-weight chidlren. Find the cutoffs, results and dis-cussion in the Appendix in Tables 8, 13 and 14.

5. It is not relevant for the second OLS since the deviations are based on both sex and age.

XI. Appendix

Page 37: Visible Hand

Fall 2015 | The Visible Hand | 37

Volume xxiV. issue i.

Page 38: Visible Hand

38 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Moreover, we saw that factors that we would ex-pect to be highly correlated with BMI, such as age of child, parental weight and race, are significant in explaining the BMI of the child. We found that fam-ily income and education of parents are significant in explaining BMI of children, too. Overall, these results are fairly consistent with the literature (Ben-son, 2010) as a whole. However, the uniqueness and singularity of the parental employment status effect result was novel, and allows for an interesting take on how we think parenting affects childhood obesity and healthy weight. It also is directly contradictory to Anderson (2012) who finds that mother employ-ment status matters. In this paper we presented al-ternative mechanisms that could explain this dif-ference. We would be interested in continuing this analysis with further controls for parental ability such as personality or legal records as it would al-low us to cure some of our most important endo-geneity problems and undertake the analysis with different data sets, as this would give us a better understanding of exactly how paternal employment status affects healthy weights of children.

IX. References

Anderson, Patty. “Parental Employment, Family Routines and Childhood Obesity.” National Center for Biotechnology Information. U.S. National Library of Medicine, 10 Dec. 2012. Web. 22 May 2015.

Araneo, Jackie. “The Effects of Maternal Employ-ment on Childhood Obesity in the United States.” Diss. Princeton U, 2008, Print.

Benson, Lisa. “The Role of Parental Employment in Childhood Obesity.” Digital Repository at the University of Maryland (DRUM), University of Maryland, 2010. Web. 23 May 2015.

“Childhood Obesity Facts.” Centers for Disease Control and Prevention, 24 Apr. 2015. Web.

Monda, Keri L. “A Meta-analysis Identifies New Loci Associated with Body Mass Index in Indi-viduals of African Ancestry.” Nature Genetics, 14 Apr. 2013. Web.

“NCHS Data on Obesity.” Centers for Disease Control and Prevention, pag. 12 Jan. 2012. Web.

X.Footnotes

1. Note that each observation is a child, not a household. 2.We also defined a log income per child variable, but saw that it did not perform as well as our log family income variable in the regression. These two have different economic interpretations: the difference is essentially between households who need to allocate income among children and house-holds who just have large income overall. See the Appendix Tables 7-10 for the regressions with log income per child.

3. This is suitable since higher education of one parent would probably have positive externalities on the other parent’s parenting. Thus, the highest education of the parent would bring the other par-ent’s parenting ability up.

4. We also created a variable for severly under-weight chidlren. Find the cutoffs, results and dis-cussion in the Appendix in Tables 8, 13 and 14.

5. It is not relevant for the second OLS since the deviations are based on both sex and age.

XI. Appendix

Page 39: Visible Hand

Fall 2015 | The Visible Hand | 39

Volume xxiV. issue i.

Page 40: Visible Hand

40 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Discussion of Logistic Regression with Underweight

In the logistic regression employment status of the mom becomes relevant along with weight of the parent and age of the child. While statistically sig-nificant, the economic significance of the marginal effect of having a mom at home is not as strong since a mom not being employed increases the risk of children being underweight by 1%. While the re-sults of this regression are interesting, we focus on the general deviations and obesity, since these are more pressing health matters, and we leave this for future studies.

Page 41: Visible Hand

Fall 2015 | The Visible Hand | 41

Volume xxiV. issue i.

Page 42: Visible Hand

42 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

double-edged sword. There are many benefits to be gained from international trade in assets, hence-forth referred to as ‘capital flows’. For example, allowing foreign firms to invest in domestic mar-kets engenders greater competition and efficiency (Eichengreen, et al. 1999; Forbes, 2008). Yet, for all its benefits, the international movement of capital also comes with its own unique risks. For instance, Ostry et al. (2010) note that capital inflows can make countries more vulnerable to financial crisis. Herd behavior, where investors act in concert, and excessive optimism on the part of foreign lenders combined with shortsighted borrowers who under-estimate risk, can lead to suboptimal borrowing that endangers financial stability. Thus, in a world with increasing financial integration openness to capital flows also means openness to the ramifications of a global economic downturn, as in the case of the 2008 financial crisis. While developing countries stand to benefit most from integration into the global economy, they are

I. Introduction It has been 7 years since the 2008 financial crisis. During the last quarter of 2008, as the economies of rich nations contracted, real GDP also plummeted in economies like Singapore and Brazil, economies re-nowned for dynamism (The Economist—Counting their blessings). Understandably, most commenta-tors believed that emerging and developing markets would suffer excessively due to their close links with Western countries. In fact, The Economist re-ported that, “Exports in that dreadful last quarter of 2008 fell by half in the Asian tigers (Hong Kong, Singapore, South Korea and Taiwan) at an annual-ized rate; capital flows to emerging markets went over a cliff as Western banks ‘deleveraged’” (The Economist— Counting their blessings). The 2008 financial crisis, which saw economic downturns in developed economies trigger reces-sions in other less developed economies, highlights the burgeoning interconnectivity of global finance. In many ways, global financial integration is a

Damilare Aboaba Cornell University

Preserving Financial Stability: Capital Controls in Developing Countries

during Times of Financial Crisis

Developing countries stand to benefit highly from international finance because it engenders transparency, competition and efficiency in domestic markets. Yet, as the Great Recession showed, they also stand to lose the most from financial integration. Although unpopular in the 1980s and 90s, capital controls—measures that discriminate between how foreign and domestic capital are treated—have begun to enjoy greater popu-larity in academic and policymaking circles, particularly since the 2008 financial crisis. While some argue that capital controls represent a useful macroprudential tool during economic crisis, others argue that capital controls are, at best, ineffective. Additionally, because much of this debate has focused on developed countries, scholars have paid insufficient attention to the countries least able to bear the cost of a global financial crisis, and thus most likely to use capital controls. By using data on the presence and intensity of capital controls in 29 developing countries and their real historical GDPs from 1990 to 2012, this paper seeks to address this gap and demonstrate whether temporary capital controls represent a viable policy option for developing countries to preserve financial stability during times of financial crises. I hypothesize that, during periods of global financial crisis, the temporary use of capital controls can help developing countries preserve financial stability. To test this, I measure capital controls based on severity, which I argue offers more definitive results than one predicated on class. However, the results from this analysis suggest that capital controls—regardless of class or intensity—do not help developing countries preserve financial stability.

Page 43: Visible Hand

Fall 2015 | The Visible Hand | 43

Volume xxiV. issue i.

also those most negatively affected by financial cri-ses. For example, during the Asian Financial Crisis of 1997, the overreaction of market actors, coupled with regional financial integration triggered a mas-sive economic downturn that wiped out the financial gains of the prior years (Corsetti, Pesenti, & Rou-bini, 1999). As a result, it is possible that capital controls might be a beneficial option for develop-ing countries that wish to minimize their financial risk. Corsetti et al. (1999) conclude that an ‘ideal’ government dealing with one clear distortion could theoretically improve welfare by imposing capi-tal controls. However, the fundamental debate is whether any such ideal government exists, and whether it would be able to isolate one clear dis-tortion. Although developing countries rarely pos-sess ideal governments, improving welfare would clearly be in their interests. More recently, authors like Korinek (2011), in considering how episodic capital controls can enhance financial stability, have noted how inherent characteristics of financial mar-kets, such as incomplete information, pose an ex-ternality problem. In economics, an externality is a byproduct of the production or consumption process whose cost is not factored into the decision-making process of consumers and/or producers. Essentially, much like a tax on pollution forces drivers to inter-nalize the costs of the pollution externality, a tax on capital inflows could be used to limit foreign bor-rowing by making borrowers internalize the costs of the financial instability caused by excessive borrow-ing (Korinek, 2011). Thus, a worthwhile question to ask is; in periods of financial crisis, do temporary capital controls represent a viable policy option for developing countries hoping to preserve financial stability?

II. A Short History

A capital control is any restriction on capital transactions—the movements of international capi-tal in and out of domestic markets. In other words, a capital control can be thought of as any policy that discriminates between how residents and foreigners can move capital into and out of a country. Capi-tal controls take a multitude of forms ranging from taxes on capital movements to complete prohibi-tions on the purchase of domestic factors of produc-tion by foreign companies or persons. Additionally, capital controls can serve a range of functions, such as revenue generation, credit allocation, correction of international balance of payments deficits or sur-pluses, protection of domestic financial firms, and prevention of potentially volatile inflows (Neely,

1999). Capital controls have a surprisingly robust histo-ry both on a theoretical and empirical level. During the 1930s, the original objective of capital controls was to curb mass outflows of capital from domestic market to international markets in negative econom-ic periods, such as the Great depression (Nurkse, 1944). After implementing capital controls in 1931, Hungary, Greece, and Bulgaria did not experience mass capital outflows. Nurkse (1944) notes the prevailing economic rationale that when foreign or domestic investors become skittish, capital controls are the only instrument for preventing capital flight. Thus, capital controls in the ‘thirties represented a technique to check the disequilibriating nature of capital movements. One example of this nature is the fact that withdrawals of foreign credit from debtor countries usually occur when foreign credit is most needed. At the time of the Great Depression, such withdrawals occurred on a massive scale. On a theoretical level, the oft-cited touchstone for the contemporary iteration of the idea is Tobin’s 1972 proposal. Tobin argued that countries might be best served utilizing a market based capital control to counteract the possibility of international capital movements causing wild swings in currency val-ues (Klein, 2012). More formally, Tobin (1978) advocated a small tax on foreign exchange trans-actions to “throw some sand in the wheels of our excessively efficient international money markets”. However, as a result of the subsequent rise of neo-liberal economic theory, which argued that the ben-efits of international capital movement far exceeded any costs, views on the efficacy of capital controls changed. Consequently, many countries dismantled capital controls. However, this trend began to reverse again in the early 2000s and was sped up by the Great Re-cession. According to Klein (2012), capital controls were reintroduced mainly to address worries about capital inflow-fueled exchange rate appreciations and asset booms, which could lead to asset bubbles as in the case of the Great Recession. It is impor-tant to note as well that this shift in practice has also been mirrored by shifts in general economic opin-ion. Crucially, the International Monetary Fund had begun to change its views on capital controls. For instance, in 2002, Kenneth Rogoff in his capacity as the then Chief Economist and Director of Research at the IMF wrote a piece titled Rethinking capital controls: “When should we keep an open mind?”, in which he argued that a more eclectic approach (one that made use of capital controls albeit temporarily) was required. (Rogoff, 2002). This shift continued

Page 44: Visible Hand

44 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

till the IMF more recently published a paper arguing that capital controls represent a viable policy option if all other options have been exhausted (Ostry et al., 2010).

III. The Literature

Whereas most economists agree that in devel-oped nations the benefits of freely moving capital outweigh the costs, there has been an inconclusive debate on the appropriateness of capital controls in developing economies (Forbes, 2008). Economists have explored the relationship between capital con-trols and indicators of macroeconomic health in de-veloping countries on a country or regional specific basis. Crucially, research on the suitability of capital controls in developing economies in inconclusive because different studies have measured for very specific types of capital controls and often extrapo-lated to the general population of capital controls. For instance, Alfaro, Kalemli-Ozcan, & Volosovych (2007) analyze the pattern and determinants of capi-tal flow directions and capital flow volatility using data for developed and developing economies from 1970-2000. The authors analyze data on flows ex-tracted from balance-of-payment statistics. Essen-tially, they analyze changes in the liabilities and as-sets of a country’s financial account, and use these as proxies for inflows and outflows respectively. This methodology is problematic for two reasons. First, the term capital control is representative of a large list of measures that can vary in multiple ways. For example, capital controls can be imposed on a short-term or long-term basis, and their measure does not explicitly differentiate between these. Sec-ond, in not focusing on developed economies alone, many scholars have paid insufficient attention to the countries least able to bear the cost of a global fi-nancial crisis, and consequently the countries most likely to use capital controls. This study attempts to addresses this gap in the literature by focusing on the use of capital controls in developing countries during the Great Recession. The literature on capital controls is both rich and diverse. For instance, the literature covers top-ics ranging from in-depth analysis of the impact of capital controls in Chile (Forbes, 2007b), to general analysis of the costs of capital controls at micro-eco-nomic levels (Forbes, 2007a). This study organizes the literature on two dimensions: (1) scope–whether studies focus on one or a few specific countries or whether they employ a cross-country analysis—and (2) focus—whether studies focus on the impact of capital controls or on the impact of financial liber-

alization. The paper discusses case studies (small country analyses of capital controls), cross-country analyses of the impact and/or determinants of capi-tal controls, and cross-country analyses of capital liberalization.

i.Case studies The first branch of the capital control litera-ture is populated by studies analyzing the effect of capital controls on a case study basis (De Gregorio, Edwards, & Valdes, 2000; Edwards, 2000; Forbes, 2008; Gallego & Schmidt-Hebbel, 1999; Valdés-Prieto & Soto, 1998). These studies generally serve two purposes; (1) adding detailed knowledge on specific capital controls, and (2) testing causal rela-tionships in specific instances. Partially because of this, however, they are unable to generalize about the effects of capital controls beyond specific cases. For instance, Forbes (2007a) examines Chile’s use of capital controls from 1991 to 1998 (the Encaje), a case that is often used as an example of success-ful capital controls because of the concurrent strong economic performance. Contrary to widespread perceptions, Forbes argues that good macroeco-nomic policy—not capital controls—were respon-sible for the country’s economic performance dur-ing that period. While the costs of the “Encaje” are clear, the same cannot be said of its benefits. For instance, there is little evidence that controls were successful in protecting the Chilean economy from shocks, even though there is some indication that it altered the composition of capital inflows, (Forbes, 2007a). Yet, the controls clearly caused a marked rise in the financing costs of small to medium-sized firms (Forbes, 2007b). Thus, because the effects of the Encaje were mixed, the strong economic perfor-mance of Chile during the 90s cannot be solely, if at all, attributed to its use of capital controls. On the other hand, other studies in this vein (Ariyoshi, 2000; Edwards, 1999; Rodrik & Velasco, 1999) suggest that capital controls might be effec-tive in restraining capital flight by moderating the inflow of short-term speculative capital (hot mon-ey) and changing the maturity composition of in-flows towards longer-term less volatile flows. For instance, Edwards (1999) argues that controls on capital inflows increased the maturity structure of Chile’s foreign debt. Using Chile as his focus, the author suggests that the positive impact of capital controls in Chile has been somewhat exaggerated. To be sure, the controls did appear to have altered the maturity of Chile’s foreign debt considerably, as indicated by a dramatic decline and increase in short-term flows and long-term flows simultane-

Page 45: Visible Hand

Fall 2015 | The Visible Hand | 45

Volume xxiV. issue i.

ously between 1988 and 1997. However, if maturity is measured residually rather than contractually a different situation emerges. Residual maturity is the value of the stock of a country’s liabilities in hands of foreigners that come due within a year. Conse-quently, prior to the use of capital controls more than 40% of Chile’s debt to banks had a residual maturity of less than a year indicating that financial flows to Chile were perhaps not as mature as some claim. Additionally, Edwards (1999) notes that the total volume of aggregate flows into Chile did not decline, and the capital controls have come with significant costs, such as significantly increasing the cost of capital for small and medium sized Chilean companies. Similarly, Ariyoshi (2000) associates capital con-trols with a reduction in the full level of net capital inflows to Thailand, but not to Chile, or Colombia. His analysis suggests that capital controls were at least particularly successful in changing the com-position of inflows from short-term to longer-term inflows. In contrast to Ariyoshi (2000), Rodrik and Velasco (1999) find that while capital controls on short-term debt flows implemented by both Malay-sia and Chile did not influencing the overall volume of capital flows, they did affect the maturity compo-sition of flows. Despite this branch’s focus on detail, a key criticism of it and the conclusions that arise from it is that most studies use different definitions and measures for capital controls. For example, Ed-wards (1999) looks at changes in the gross foreign capital flowing into Chile, while Forbes (2007a) studies changes in domestic cost of capital in Chile. Similarly, different studies also use different defi-nitions when measuring the success or failure of capital controls. Following from the example above, while Edwards (1999) would consider a substantial change in the gross foreign capital flowing into Chile accompanied by improving macroeconomic conditions as successful, Forbes (2007a) would consider improving macro conditions coupled with little adverse change in the domestic cost of financ-ing as indicative of the positive impact of capital controls. Thus, there is usually little overlap in the definitions, measures, or conditions of success used by different capital control case studies.

ii.Cross-country studies In contrast to scholars that primarily employ case studies, a second group of scholars examine the ef-fects and/or determinants of capital controls on a multi-country basis. (Montiel & Reinhart, 1999; Edison & Warnock 2003; Magud & Reinhart, 2006; Ostry et al., 2010). This branch of the literature is

unified by topic, but not necessarily by method. Edison and Warnock (2003) propose a monthly measure of the intensity of capital controls. Specifi-cally, they use restrictions on foreign ownership of equities to show the effects of financial liberaliza-tion—or the removal of capital controls—on several economic variables across 29 emerging countries. This study is important because it focuses specifi-cally on emerging economies, and devises a more nuanced means of measuring how severe the rules restricting the flow of capital are. As an example, the authors are able to say that countries like Ko-rea and Thailand were relatively closed in the early 90s, but greatly relaxed capital controls during the Asian Financial Crisis (Edison & Warnock, 2003). Overall, Edison and Warnock find that a complete liberalization causes a much more intense decline in the cost of capital than previous studies have report-ed. Additionally, they find that the more complete financial liberalization is, the greater the magnitude of exchange rate appreciation and capital inflows. Similarly, Magud and Reinhart (2006) construct two indices of capital controls, Capital Controls Ef-fectiveness and Weighted Capital Control Effective-ness, to summarize the results of over 30 empirical studies in the capital controls literature from 1985 to 2003. Their main premise is that besides the Ma-laysian and Chinese experiences, little evidence of capital control effectiveness exists. As such, their primary goal is to measure how effective capital controls are generally. It is important to note that, the WCCE index is compiled as a way to control for the differing degree of rigor used in each of the stud-ies they consider. Magud and Reinhart’s analysis suggests that while capital controls were successful in making monetary policy independent by alter-ing the composition of capital flows towards longer maturities, they were not successful in reducing the volume of net flows or capital outflows. Cross-country studies analyzing effects of classes of capital controls Smaller subsets of the cross-country effect of capital controls branch of the literature focus par-ticularly on distinctions between classes of capital controls (Korinek, 2011; Klein, 2012). This paper will bear greatest resemblance to the studies in this group. Klein (2012) examines the pattern of con-trols on capital inflows and their association with key financial variables. A focal point of his paper is the dividing line between long-standing controls on a wide range of assets, henceforth referred to as walls, and periodic controls on a narrower set of as-sets that can and are often imposed and removed, henceforth referred to as gates. For instance, while

Page 46: Visible Hand

46 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

a country like China has a wall, other countries like Brazil have a gate because they have only imple-mented capital controls when they appear necessary. Thus, walls tend to be wide and high, limiting all manner of capital flows, including those that could have positive benefits such as providing cheaper capital. In contrast, gates can open during good times to allow the economy to benefit from inter-national finance, but shut in the face of potentially destabilizing capital. Additionally, they are more ef-ficient than the former because they are transitory and targeted towards particular categories of assets. The author’s premise is that using capital controls as temporary, pro-cyclical, focused tools is an effective strategy for enhancing financial stability. Klein (2012) comes to the conclusion that coun-tries closed to capital inflows do have higher an-nual rates of GDP growth and lower annual rates of growth of financial variables linked to financial instability. Additionally, his initial regression results indicate a statistically significant lower rate in the growth of financial instability indicators in coun-tries that use walls, in comparison to those that use gates. However, because financial liberalization is well correlated with development, when GDP per capita is controlled for the relationship disappears, and neither class of capital control appears to be bet-ter than the other. Klein’s paper is important to this study because I build on his description of gates to measure capital controls.

iv. Cross-country studies analyzing effects and de-terminants of capital controls

Another subset under the cross-country effect of capital controls branch of the literature is exclusive-ly interested in the analysis of the effects and deter-minants of capital controls (Epstein, 1992; Grilli & Milesi-Ferretti, 1995; Quinn & Inclán, 1997; Mile-si-Ferretti, 1998; Bai & Wei, 2000). For example, Grilli & Milesi-Ferretti (1995) conduct a theoretical and empirical study of the effects and determinants of capital controls using panel data from 61 devel-oping and developed countries. More specifically, they investigate whether specific political and struc-tural characteristics have an effect on capital control imposition or removal. From a political economy point of view, the relationship between political stability, government preferences, and credibility is paramount. The authors come to several conclu-sions. First, the likelihood of capital control imposi-tion increases with a lower income level, larger gov-ernment, less independent central bank, and more closed economy. Additionally, some other determi-nants of capital controls are the exchange rate re-

gime, current account imbalances, and the openness of the economy. Second, and particularly relevant to the case of developing countries, governments with lower initial credibility may be those with the most to gain from implementing capital controls. Third, as to the effects of capital controls, Grilli & Milesi-Ferretti (1995) find that capital controls tend to be associated with increased inflation and lower inter-est rates, but no relationship with economic growth. Similarly, Alesina & Tabellini (1989) study the political economy of capital controls. Their paper attempts to explain the rising external debt in sev-eral developing economies, particularly those in Latin America, and why governments have not at-tempted to curb this debt by using capital controls, avoiding rapid appreciations of their exchange rates, or restricting the public sector’s external borrowing. Interestingly, it is suggested that capital controls do not allow domestic residents to avoid inflation taxes on domestic money balances. Alesina and Tabellini argue that left-wing governments are more inclined to use capital controls than right-wing ones. They conclude that generally, capital flight and large gov-ernment debt are more likely in countries that are politically unstable.

v. Cross-country studies analyzing the impact of Capital Account Liberalization

The last branch is made up of cross-country stud-ies analyzing the impact of capital control removal (capital liberalization) on several factors, most prominently growth (Prasad, Rogoff, Wei, & Kose, 2003; Kose, Prasad, & Terrones, 2003; Mody & Taylor, 2003). For example, Prasad, Rogoff, Wei, & Kose (2003) focus on the impact of capital control removal on economic growth and macroeconomic volatility in developing countries. The authors use a measure of financial openness based on the esti-mated gross stocks of foreign assets and liabilities as a percentage of GDP. They reach the conclusion that financial integration is best undertaken cau-tiously, and with functional institutions and good macroeconomic structures as the foundations of any such undertaking. Additionally, they note that their evidence does not provide clues as to the ideal se-quencing and pace of capital control removal. Kose, Prasad, and Terrones (2003) examine the impact of capital control removal on macroeco-nomic volatility. The authors’ starting premise is that economic theory is unclear as to how financial integration should affect volatility, thus this topic is primarily an empirical question. Looking at the changes in macroeconomic volatility in a large sample of industrial and developing countries from

Page 47: Visible Hand

Fall 2015 | The Visible Hand | 47

Volume xxiV. issue i.

1960 to 1999, the authors come to three major con-clusions (Kose, Prasad, & Terrones, 2003). First, the volatility of output growth has appeared to decline over time. Second, on average the volatility of con-sumption growth relative to the volatility of income growth has increased for more financially integrated developing countries over time. Third, an increase in financial integration, or a decrease in capital con-trols, is connected to the rising relative volatility of consumption up to a certain threshold. However, be-yond this threshold, benefits of financial integration seem to build up. Generally, this branch suggests weakly that capital control liberalization may posi-tively impact growth. As Kose, Prasad, Rogoff, and Wei (2009) note, while it remains difficult to find strong evidence that the removal of capital controls increases growth once other variables are controlled for, the weight of the evidence points to positive marginal effects.

IV. Criticisms of the literature and unanswered questions

Overall, the literature on capital controls suffers from a lack of clarity with respect to proper measure of the effects and success of capital controls. Put more succinctly, most studies use different defini-tions and measures of capital controls. For instance, while Edison and Warnock 2003 calculate the re-strictions on foreign ownership of equities as a mea-sure of capital controls, Klein (2012) uses scores of 0 or 1, no capital controls and capital controls re-spectively, for six categories of assets from the IMF. Additionally, most measures of capital controls sim-ply measure the presence or lack of capital controls, rather than the intensity of capital controls. This is true even for the subset measuring classes of capi-tal controls, Klein (2012). Although this gives some indication of the scope of capital controls, it does not really give an indication of the degree to which capital flows are regulated. The intensity of capital controls is particularly crucial because capital con-trols can be very difficult to enforce especially in developing countries because of underdeveloped markets. As a result, the same capital control may have different degrees of effectiveness depending on the country in which it is implemented. For in-stance, imagine two countries. While country A has severe restrictions on just one category of capital controls, country B has less severe but more numer-ous restrictions on different categories of capital. In this scenario, country A’s capital controls could potentially be less severe than country B’s. This is particularly true if country A is a developing country

with poor implementation capacity limiting the ef-fectiveness of capital controls. As a result of using different definitions and mea-sures of capital controls, studies also differ in how they assess the success or failure of capital controls. For instance, studies that focus on case studies and cross-country analysis of capital controls are par-ticularly prone to these flaws. As mentioned above, Forbes (2007a) would consider improving macro-economic conditions coupled with little change in the domestic cost of financing as indicative of the success of capital controls. In contrast, Magud and Reinhart (2006) base capital control success on the achievement of 4 key objectives; reduction of the volume of capital flows, modification of the com-position of capital flows, reduction of real exchange rate pressures, and creation of space for a more au-tonomous monetary policy. Clearly, consensus on the success of capital controls is elusive because the conditions for success are not common across studies. Research on the effect of capital control liberalization on macroeconomic variables also suf-fers from this flaw considering the fact that different types of capital controls may have markedly dif-ferent effects on growth and other macroeconomic variables. For instance, Kose, Prasad, and Terrones (2003) run regressions on macroeconomic volatility using multiple measures of financial liberalization such as current account restrictions, capital account restrictions, and financial openness. They find that the impact of capital liberalization on macroeco-nomic volatility varies significantly depending on what type of capital controls are being removed. Additionally, the fact that the removal of capital controls can depend on factors that are difficult to measure, such as other reforms that often accom-pany capital control liberalization, makes it difficult to ensure that reported effects are solely caused by capital controls. Finally, in terms of branches utilizing cross-coun-try analysis, a major flaw is the lack of difference (or heterogeneity) across the countries and times used. Developing countries used in these studies are of-ten not characteristic of the larger populations. For instance, Montiel and Reinhart (1999) look at sev-eral developing countries, but the majority of those countries are either Latin American or Asian. Only two sub-Saharan and one Middle East and North African countries are included. Similarly, Ostry et al. (2010) look at six different selected cases of con-trol measures on capital inflows. Of these six cases, three are Latin-American, two are Asian, and 1 is European. The lack of cohesiveness in research strategies and

Page 48: Visible Hand

48 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

the definition and measurement of capital controls is problematic for several reasons. Firstly, the literature as of yet does not contain cross-country analyses of the impacts of specific types of capital controls, par-ticularly capital controls that would theoretically be useful in the case of a global financial crisis, such as episodic controls on capital outflows. Secondly, the literature does not have a standardized method of measuring the intensity and/or success of capital controls; this is particularly relevant in deciding if capital controls ever represent viable options for de-veloping countries. A standardized holistic method of measuring the intensity of capital controls would allow the effects and success of different types of capital controls to be adequately compared. Lastly, the literature as of yet does not pay sufficient atten-tion to developing countries, especially those devel-oping countries that are characteristic of the larger population. This is crucial because, on a theoretical basis, developing countries would seem to be the countries most in need of capital controls, and, on a theoretical basis, developing countries are statisti-cally more likely than developed countries to have some kind of capital control in place Klein (2012). Consequently, this paper will attempt to fill these gaps by explicitly specifying what capital controls will be studied, developing a standardized method for measuring the intensity and success of capital controls, and analyzing several developing countries broadly representative of most developing nations.

V. Methodologyi.Hypothesis

Although the literature on capital controls lacks consensus, some studies indicate that capital con-trols can have some impact on the volume and com-position of capital flows. This is particularly true in the cross-country studies. For example, Ostry et al. (2010) note that empirical studies are more success-ful at finding an effect of capital controls on capi-tal inflows—especially in terms of lengthening the maturity of inflows. Similarly, Magud and Reinhart (2006) conclude that capital controls can change the composition or maturity of capital flows and, in some special cases, also reduce the volume of capi-tal flows. That said, there are several case studies that suggest a link between capital controls and ma-turity of capital flows. For instance, Forbes (2008) notes that there is some evidence to suggest that capital controls in Chile lengthened the maturity of capital inflows. The literature seems to suggest that the benefits of capital controls might be found primarily in their ef-fect on the volume and composition of capital flows.

It is also important to note that during a financial crisis, the volume and composition of capital flows are crucial in determining the financial stability of a country. Thus, it is possible that during times of financial crisis developing countries can enhance short-term financial stability by using a tool (capital controls) that speaks directly to those factors. This paper provides a test for this hypothesis. As HI and H2 outline, I expect that:

H1: During a global financial crisis, developing countries that temporarily use capital controls to (1) restrict sudden capital outflows (capital flight) and (2) change the composition of capital inflows are more financially stable, and are more macro eco-nomically healthy than those who do not. H2: A measure of capital controls analyzing severity will give more definite and conclusive results, than a measure of capital controls based on the class of capital controls.

To be sure, past research has looked at the re-lationship between capital controls and financial variables, such as macroeconomic stability. For instance, Klein (2012) studies the effectiveness of capital controls on the growth of several financial variables, including the Real exchange rate and GDP. However, results are not entirely convincing or conclusive. As discussed in my literature review, most studies use different definitions and measures to analyze capital controls because it is difficult to capture the different types of capital controls (Edi-son & Warnock, 2003). Yet because of this varia-tion in technique, scholars are actually measuring different types of capital controls even though they claim to all measure capital controls in general. Consequently, the inconclusiveness of the literature on whether capital controls are appropriate is un-derstandable, and the generalizability of individual studies is somewhat disputable. An approach that could bridge this gap is mea-suring the severity of capital controls. Beyond the fact that there are many types of capital controls, an important caveat highlighted by the experiences of countries that have instituted capital controls is that the effectiveness of controls is crucially dependent on a country’s implementation capacity. Effective-ness depends on whether the country already has capital controls and thus an already functioning ad-ministrative infrastructure. Thus, effectiveness of capital controls is partially a function of how often a country uses them (Ostry et al., 2010). Put simply then, not all capital controls are created equal. Thus, to fairly measure capital controls it is crucial to use a

Page 49: Visible Hand

Fall 2015 | The Visible Hand | 49

Volume xxiV. issue i.

measure that can accurately reflect the multitude of capital controls and their severity on a case-by-case basis. I believe that capital control intensity fits this criterion. Assessing capital control intensity adds to the literature because intensity is an easily generalizable measure that allows for the fair and accurate evalua-tion of the financial impact of varied types of capi-tal controls. In addition, it also represents a more nu-anced technique of measuring capital controls when compared to simply scoring cases based on whether they use capital controls or not. It takes into account the quality of capital controls, both their strength and scope. Lastly, even if this study’s hypothesis proves false, the study will at least demonstrate how much the impact of capital controls varies by mea-sure, and whether measures of presence, intensity, or a combination of both are best suited to analyzing capital controls.

ii. Data & Method I run random effects regressions to estimate the impact of capital controls on financial stability, while controlling for inflation and the U.S. interest rate. I will use two models with two different mea-sures of capital controls. The first will use the class of capital control i.e. whether the control is a gate or wall, and the second will use the intensity of capital controls. A key contribution of this research study is determining whether using different measures of capital controls yields different results. I expect that there will be a significant difference and this can be used, as noted above, to partially explain the varia-tion that exists in the literature regarding whether capital controls enhance financial stability or not. Developing countries stand to gain and lose the most from burgeoning financial integration (Eichen-green, et al. 1999; Ostry et al. 2010). Thus, this pa-per will focus on the relationship between capital controls and financial stability in developing coun-tries. Specifically, this paper will draw on the expe-riences of 29 developing countries during the Great Recession. I should note that the term “developing” can often be perceived as contentious in the social sciences. However by developing country I simply refer to nations that fall into two of the World Bank’s Country Lending Groups. The first group are 5 low-income which are defined as countries with a Gross National Income (GNI) per capita of $1,045 or less as of 2013. The second group are 23 lower-middle-income economies—countries with a GNI per capi-ta of between $1,046 and $4,125 (The World Bank). These countries are Armenia, Bangladesh, Bolivia, Cote d’Ivoire, Egypt, El Salvador, Georgia, Ghana,

Guyana, Indonesia, India, Kenya, Kyrgyz Repub-lic, Malawi, Mongolia, Morocco, Nigeria, Paki-stan, Papua New Guinea, Paraguay, Philippines, Sri Lanka, Tanzania, Uganda, Ukraine, Vietnam, West Bank & Gaza, Zambia, and Zimbabwe. The financial crises studied will be the Great Recession. The Great Recession is arguably the first truly global financial crisis in terms of having an im-pact that was felt in the financial markets of many physically distant developing countries. The Great Recession can be roughly divided into 2 smaller financial crises the Global Financial & Subprime Mortgage Crisis from 2007 to 2008 and the Euro-pean Debt Crisis from 2010 to 2012. The method for selecting the countries used in the statistical analysis is not entirely random. First, it is important to note that data collection in develop-ing countries is generally poor. Additionally, several countries, such as Libya and Sierra Leone, have so much domestic instability that macroeconomic in-dicators, even when available, are unreliable. Con-sequently, I specifically selected somewhat stable countries for which data is readily available. By somewhat stable, I mean countries that are not, or have not recently experienced, substantial conflict or crisis. For instance, while I omit Sierra Leone due to the recent Ebola crisis, I include Nigeria because it had significantly less cases. Even then, I include several countries, such as Ukraine, which are pres-ently quite unstable. Thus, I avoid the problem of selecting on my dependent variable by including most developing countries for which data is readily available.

iii. Measurement of Variables a. Independent Variable In this paper, the independent variable is the use of capital controls. I define capital controls as any government sanctioned restriction on the in-ternational movement of capital that discriminates between residents and foreigners. Accordingly, two indicators will be used to gauge this variable. These indicators are (1) the class of capital controls, and (2) the intensity of capital controls. In addition to the main goal of measuring the effect of capital con-trols on financial stability, I will also use these in-dicators to examine whether results these measures produce are consistent or not. To measure the presence and class of capital controls, I use data from the International Monetary Fund’s Annual Report on Exchange Arrangements and Exchange Restrictions (AREAER) database. The database, going as far back as 1950, provides information on different types of capital controls

Page 50: Visible Hand

50 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

used by the IMF’s 187 member countries. The pres-ence of capital controls (POC) forms the basis of both of these indicators. It can be measured on a 0 (capital controls are not used) to 1 (capital controls are used) scale. This method is particularly popu-lar in the earlier capital controls literature (Alesina, Grilli & Milesi- Ferreti, 1994; Montiel & Reinhart, 1999). I consider several classes of capital transac-tions; repatriation requirements, controls on capital and money market instruments, controls on deriva-tives and other instruments, controls on credit op-erations, controls on direct investment, controls on liquidation of direct investment, controls on real estate transactions, and controls on personal capital transactions. The class of capital controls can be divided into three groups; open, gated, and walled, as in Klein (2012). Open refers to countries without capital controls. This group is demarcated by a POC av-erage of 0 to 0.25. Gated refers to countries using periodic controls on a narrower set of assets. This group is demarcated by a POC average of 0.26 to 0.69. Walled refers to countries using long-standing controls on a wide range of assets. This group is de-marcated by a POC average of 0.7 to 1. The intensity of capital controls will be mea-sured using the foreign ownership restrictions tech-nique developed by Edison and Warnock (2003). Although Edison and Warnock use this technique primarily to provide information on the degree of financial openness, their measure can also be used to indicate the degree of restrictions on capital mobil-ity. Edison and Warnock use two indices measured and computed by Standard and Poor’s Ratings Ser-vices and the International Finance Corporation (the latter computed these indices until 1999, the former since then). The first index is a global index designed to rep-resent the total available market. The global index is formed by taking a subsection of all traded stocks of domestic companies deemed representative of 60 to 75% of a country’s total market capitalization (the total dollar market value of all of a company’s outstanding shares). After being selected, the con-stituent parts of the index are analyzed and market capitalizations are accordingly adjusted to take ac-count of factors such as government ownership. The second index is an investable index designed to represent the share of the total market available to foreign investors. The investable index is com-posed of a segment of the global stocks available to foreign investors and passing screening require-ments for minimum size ($50 million in investable market capitalization) and minimum liquidity ($20

million in annual trading). Beyond this, further as-sessments of openness are conducted first at a mar-ket level, such as the ability of foreign investors to buy and/or sell shares, then at an institutional level, such as the magnitude of corporate charter limita-tions on ownership. This assessment gives rise to a calculated investability factor that is essentially an indication of the amount of outstanding shares open to foreigners; it is applied as a weight to the stock market’s capitalization when determining its weight in the Investable index (Edison & Warnock, 2003). The ratio of the market capitalizations (the total dollar market value) of a country’s investible index to its global index provides a rough quantitative standard for the availability of a country’s equi-ties to foreign investors. One minus this ratio can be considered an adequate measure of the intensity of capital controls (Edison & Warnock, 2003). Spe-cifically, a measure of nation i’s foreign ownership restrictions at time t is:

FORi,t=1- MCi.tIFCI

MCi.tIFCG

Thus, the two indices are used to create one indi-cator, foreign ownership restrictions, and that vari-able should accurately reflect the intensity of capital controls. The investable index proved too difficult to procure. As a result, I calculated my own investable index market capitalization using the average POC score for each year in each country. These capital-izations were then used as above. FOR can take any value between zero and one. Zero represents a com-pletely open market, without any capital controls, and one represents a completely closed market, one with complete capital controls.

b. Dependent Variable Financial stability can be defined in multiple ways. For instance, Nelson and Perli (2007) de-fine financial stability as a situation where markets function well, financial institutions operate without difficulty, and asset values do not deviate greatly from real prices. This definition while very compre-hensive is quite difficult to operationalize. Conse-quently, this study follows Gadanecz and Jayaram (2008) in defining financial stability as a situation characterized by the absence of excessive volatility, stress or crises. Using this admittedly simple defini-tion of financial stability, several indicators can be used. Some of these indicators are; 1) Real GDP growth rate, 2) household assets or debt, and 3) Real exchange rates. In the paper, I will only use the Real GDP growth rate indicator as a measure of financial stability. There are two main reasons for this. First, in order

Page 51: Visible Hand

Fall 2015 | The Visible Hand | 51

Volume xxiV. issue i.

to have a workable model, it is crucial that it is both thorough and yet not unnecessarily overcompli-cated. Whereas use of several different indicators to determine financial stability would create a very complicated model, it is not clear that these three indicators differ sufficiently to offer a payoff greater than the complexity they would introduce (i.e. they capture the same variation). Second, when coupled with the average real historical GDP growth rate (ARHG), real GDP growth rate represents a good measure of financial stability. This is because the Real GDP growth rate is indicative of the strength of the whole macro-economy, its ability to create wealth, and its risk of overheating. Thus, negative or low positive values would indicate a slowdown. On the other hand, excessively high values may indicate unsustainable growth (Gadanecz & Jayaram, 2008). Additionally, Real GDP describes the real sector of the economy made of tangible goods and services. Given that the phenomenon of capital controls is primarily financial, use of this dependent variable does not run the risk of overestimating the effect of capital controls. Thus, the Real GDP growth rate is a sufficient measure for the task at hand. I follow Ostry et al.’s (2007) simple method of using the Real GDP growth rate as a measure of fi-nancial stability. Specifically, I use data from a com-bination of the World Bank and the International Monetary Fund to measure the difference between the Real GDP growth of the current year and the ARHG from 1990 to 2005. Using the difference be-tween real and average real historical GDP as the final measure of financial stability is important for two reasons. First, it allows historical trends in a country’s real GDP growth to be taken into account. This is crucial because inherent financial stability in any country in the sample can skew the results. Sec-ond, the financial stability measure is relevant with regards to deviation from the ARHG. Essentially, for most countries a financial crisis is expected to cause financial instability, reflected by a large deviation in the financial stability measure. Thus, if the real GDP growth rate remains fairly close to the ARHG even after the financial county, in a country with capital controls, I would take that as evidence of the suc-cess of capital controls. Conversely, if the real GDP growth rate deviates largely from the ARHG, that would provide evidence against the effectiveness of capital controls during a financial crisis.

c. Control Variables This study includes five control variables: (1) Inflation, (2) the U.S. interest rate, (3) income level, (4) region, and (5) specific financial crisis.

Inflation is best thought of as an increase in the price level. Stated simply, when inflation rises, each unit of money, such as one dollar, is worth less. Thus, goods and services essentially become more expen-sive. Inflation is important because it exhibits a sig-nificant effect on equity (in a sense the independent variable), net debt (in a sense the dependent vari-able), and short-term debt. Campion and Neumann (2004) find that by decreasing foreign investment substantially, inflation has a significantly positive and significantly negative effect on portfolio equity and net equity respectively. Further, they find that inflation has a moderately positive impact on both volume of portfolio equity relative to GDP and net debt relative to GDP. Similarly, Domowitz, Glen, and Madhavan (2001) find that because inflation is correlated with uncertainty for fixed-income instru-ments, there are lower levels of debt issuance during periods of inflation. I am primarily concerned with preventing inflation trends, such as persistently high inflation in developing countries, from distorting my results. For inflation data I will draw from the World Bank dataset that measures inflation by the annual growth rate of the GDP implicit deflator—the ratio of GDP in current local currency to GDP in constant local currency. The percentage change in this ratio shows the rate of price change in the entire economy. The interest rate can be thought of as the re-turns to lending, or the cost of borrowing. I use the U.S. Interest Rate (USIR) as a proxy for the world interest rate. According to Campion and Neumann (2004), USIR has a significantly negative impact on portfolio equity and portfolio debt rise. It is crucial to note that portfolio debt flows tend to be more sensitive to this decline than are portfolio equity flows. In addition, net debt flows in comparison to total inflows show a sharp and positive increase in response to a decline in USIR. On the other hand, due to a reduction in direct investment—which outweighs the correspondent increase in portfolio equity—net equity flows decline as USIR declines (Campion & Neumann, 2004). Lastly, Montiel and Reinhart (1999), find that a decreasing USIR skews the composition of flows away from foreign direct investment and towards a mixture of portfolio and short-term flows relative to total flows. I will be us-ing yields on short-term U.S securities, primarily a 3-month treasury note, as in Montiel and Reinhart (1999), to measure the USIR. This measure is avail-able from the United States Treasury’s website. The last three variables can all be considered as secondary control variables. This paper uses developing countries that vary on multiple dimen-

Page 52: Visible Hand

52 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

sions. However, the most important characteristics by which they differ are income level, region, and experience in specific financial crisis. Consequently, to ensure that the variation on any of the three listed parameters does not skew my results, I control for each of them. For income level, I simply use the two World-Bank Country and Lending groups—low-income economies and lower-middle-income econ-omies—spoken about earlier in this paper. For re-gion, I divide the countries into 6 regions also based on World Bank classifications. These regions are Europe and Central Asia, South Asia, Latin America and the Caribbean, Sub-Saharan Africa, the Middle East and North Africa, and East Asia and the Pacific. Lastly, I divide the Great Recession into two smaller financial crises, the Global Financial & Subprime Mortgage Crisis from 2007 to 2008 and the Euro-pean Debt Crisis from 2010 to 2012 as described above. Although I have termed this the secondary group of control variables, it is important to note that con-trolling for them is crucial. For instance, if Sub-Sa-haran Africa experienced systemic macroeconomic instability unrelated to the Great Recession, and the region was not controlled for, it would significantly bias the results towards capital controls having a negative effect. Thus, controlling for regions, stops regional characteristics from skewing results.

iv. Generalizability I believe that this paper will exhibit substantial generalizability because the sample is well repre-sentative of the broader set of cases. First, the sam-ple captures tremendous variation. For example, dif-ferent regions capture different experiences during the most recent financial crises. Similarly, different regions also capture vastly different experiences. To deal with this variation I will be controlling for in-come level, region, and specific financial crisis as discussed above. As such, they cover most of the obvious variations existing within developing coun-tries. One obvious shortcoming however, is that all the variation cannot be accounted for. Consequently, there might be specific historical or path dependent features, such as the impact of Japanese colonialism in Asia, which this study does not capture. However, I contend that even if this proves to be the case, such factors are similar enough to those already account-ed for. Additionally, it is my view that they would not exhibit a significant impact on either the inde-pendent variable or the dependent variable. Further, the control variables will serve well in minimizing the impact of any of these factors on the study. For instance, if a particular region of developing coun-

tries tends to favor high inflation or high interest rates, the control variables will ensure that this does not distort the study. Lastly, generalizability of the study will also be significantly enhanced by the way the dependent variable will be measured. This paper will measure deviations of real GDP growth rates from the av-erage real historical GDP growth rate. Measuring financial stability in this way is crucial because de-veloping countries vary markedly with regards to this variable. Thus, attempting to simply measure absolute levels of financial stability could skew results significantly. Measuring deviations of real GDP growth rates from the average real historical GDP growth rate solves this issue because the inher-ent financial instability/stability of any country vis-à-vis other countries will not be used to judge the effectiveness of capital controls. More importantly, this aids generalizability because when financial sta-bility is measured in this way, the study does not se-lect against, and is thus relevant to, any developing country regardless of inherent stability or instability. It is important to note that my analysis of the impact of capital controls in developing countries will suffer from one limitation. The limitation is that I do not draw a distinction between controls on inflows and controls on outflows. This is because I hypothesize that capital controls will enhance finan-cial stability through both channels. As a result, I run the risk of underestimating the impact of capital controls. However, although a distinction between controls on capital inflows and controls on capital inflows might help to further explain the phenom-enon, I argue that it is not essential in finding a re-lationship between capital controls and financial stability.

VI. Data Analysisi. Correlation

I ran a function to look at the correlation between each of my predictive variables to ensure that the correlation between my independent and control variables was not substantial. Most of the results are reassuring in that most of my variables are not closely correlated. For instance, neither of my independent variables (class and in-tensity of capital controls) is highly correlated with my control variables. However, the correlation test does show that both of my independent variables are quite correlated. Testing these relationships by regressing the variables will shed some more light on the issue.

ii.Regression The first important thing to note is that when asking Stata to handle panel data, it returns the reply

Page 53: Visible Hand

Fall 2015 | The Visible Hand | 53

Volume xxiV. issue i.

strongly balanced meaning that nearly the full set of data for each country is available. Panel data is usually analyzed using either fixed effects or Ran-dom effects regressions. A fixed effects regression assumes that something within the entities may im-pact or bias the independent or dependent variables and this needs to be controlled for. A random effects regression assumes that the variation across entities is random and uncorrelated with the independent variables. To decide between the two techniques, I run a Hausman test. b = consistent under Ho and Ha; obtained from xtreg

B = inconsistent under Ha, efficient under Ho; ob-tained from xtreg Test: Ho: difference in coefficients not systematic chi2 (2) = (b-B) ‘ [ (V_b - V_B) ^ (-1) ] (b-B) = 0.10 Prob>chi2 = 0.9519

The test returns a result stating that the coefficients are not systematic, and the “Prob > chi2” value is 0.9519, far above the 0.05 cutoff at which the use of fixed effects would have been appropriate. Con-sequently, a random effects regression is used be-cause the primary objective is to analyze the impact of capital controls on financial stability, two vari-ables that do not seem to vary much over time. Es-sentially, the regression tests the effect of a 1-unit value change in the independent variable within and across cases on the deviation from the average real historical GDP growth rate. The advantage with this type of regression and analysis is the ability to sta-tistically control for other differences across coun-

tries that might explain the change in the ARHG. Given that each country has its own individual char-acteristics that may or may not influence the use of capital controls, this regression assumes that there are factors within the country, apart from the control variables, that may impact or bias the independent and dependent variables.

a. Regression 1 The results of my first regression can be found in Table 3 below. For this regression I ran the model with class of capital controls as my independent variable. Looking at the coefficients of the regressors, when the class of capital controls changes by one unit, fi-

nancial stability only changes by -1.056 percent. Al-though this initially seems small, a 1% deviation in ARHG is potentially significant. The p value, is sup-posed to test the hypothesis that each coefficient is different from 0. To reject this hypothesis, the values have to be greater than less than 0.05 respectively. This gives a confidence level of 95% for both tests. According to my results, the class of capital controls does not have a significant effect or influence on fi-nancial stability. However, it seems that the USIR (World interest rate) and the specific crisis have a significantly positive influence on financial stabil-ity. When the USIR changes by 1 percent, financial stability changes by 0.87%. Similarly, as the specific crisis changes from the Global financial crisis to the European Debt Crisis (a change of 1 unit), financial stability changes by approximately 3%.

b. Regression 2 The results of my second regression can be found in table 4 below. For this regression I ran the model with intensity of capital controls as my independent variables.Looking at the coefficients of the regressors, when

Table 1: Correlation between Predictive Variables

Table 3: Effect of class of capital controls on Macroeconomic Stability

Table 2: Hausman Test

Page 54: Visible Hand

54 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

the intensity of capital controls changes by one unit, financial stability only changes by -1.568 percent. The p value, is supposed to test the hypothesis that each coefficient is different from 0. To reject this hy-pothesis, the values have to be greater than less than 0.05 respectively. This gives a confidence level of 95% for both tests. According to my results, the in-tensity of capital controls does not have a significant effect or influence on financial stability. However, it seems that the USIR (World interest rate) and the specific crisis again have a significantly posi-tive influence on financial stability. When the USIR changes by 1 percent, financial stability changes by 0.805%. Similarly, as the specific crisis chang-es from the Global financial crisis to the European Debt Crisis (a change of 1 unit), financial stability changes by approximately 2.56%.

VII. Conclusion

In conclusion, my regressions suggest that capital controls do not play a significant role in determin-ing a country’s macroeconomic health, measured by the average real historical GDP growth rate. In addition, my regression indicates that the class of capital controls measure gives more definitive re-sults. Furthermore, and more importantly, the class of capital controls measure is also the measure that comes closest to being statistically significant. Thus, my regressions prove both of my hypotheses false. During a global financial crisis, developing coun-tries that temporarily use capital controls are not more financially stable than those who do note. Secondly, a measure of capital controls analyzing severity does not give more conclusive results than a measure of capital controls based on the class of capital controls. The results of my regression are consistent with the economic consensus on the appropriateness of

capital controls in developed countries, and seem to corroborate what some of the literature on capital controls in developing countries document. In terms of case studies, the results of my analysis align quite closely with the studies performed by Forbes (2007a) and Magud and Reinhart (2006). That is, preserving financial stability during a financial crisis is primarily the result of good macroeconomic pol-icy, but that policy is not about capital controls. For instance, Forbes categorically denied that Chile’s strong economic performance even in the face of financial crisis was the result of the successful use of capital controls. Rather, she argues that the econ-omy’s performance is the result of adroit macroeco-nomic policy (Forbes, 2007a). Additionally, as Ma-gud and Reinhart (2006) note, capital controls are ineffective at modifying the volume of capital flows during a financial crisis. On the other hand, my re-sults disagree with the suggestions of studies, such as those of Ariyoshi (2000), Edwards (1999), and Rodrik and Velasco (1999). Although I did not ana-lyze capital inflows or the maturity structure of debt explicitly, my results indicate that even if it were the case that capital controls were effective in restrain-ing capital flight by moderating and changing the maturity composition of inflows, this had no effect on financial stability. Thus, during a financial crisis countries would not use capital controls. However, this scenario is unlikely. It is more likely that be-cause capital controls were unsuccessful in altering capital inflows as detailed above, financial stability was also unaffected. In terms of cross-country studies, the results of my research agree most strongly with those found by Klein (2012). My results indicate that during a financial crisis, capital controls do not have an im-pact on financial stability, and the result is quite consistent whether the impact of capital controls is measured using intensity or class. Similarly, Klein (2012)’s final conclusion is that neither class of capital controls has an effect on a country’s mac-roeconomic health. The conclusions I have drawn from my analysis also share slight similarities with those reached by Magud and Reinhart (2006). The authors find that capital controls were not success-ful in reducing the volume of net flows or capital outflows. However, my results are inconsistent with much of the conclusions reached by the rest of the cross-country capital control literature reviewed earlier. For instance, although I use a similar method to Edison and Warnock (2003), their conclusion that the use of capital controls reduces capital inflows, is not well supported by my results. Similarly, Kose, Prasad, Rogoff, and Wei (2009) suggest that the

Table 4: Effect of Intensity of capital controls on Financial Stability

Page 55: Visible Hand

Fall 2015 | The Visible Hand | 55

Volume xxiV. issue i.

weight of the evidence indicates that the use of capi-tal controls reduces growth. However, my results indicate that capital controls have largely minor ef-fects on growth and macroeconomic health.

i. Limitations & Future Research That said, further research on capital controls is still needed. First, although I attempt to mirror the capital control intensity measure created by Edi-son and Warnock (2003), I settle for an imperfect facsimile of the investible index used to calculate the measure. Hence, the results of my analysis are not conclusive but preliminary, especially where in-tensity is concerned. The capital control literature would benefit from applying the intensity measure to a wider set of cases to test how it compares with other common measures of capital controls. Sec-ondly, I do not draw a distinction between controls on inflows and controls on outflows. This is because I initially hypothesized that capital controls would enhance financial stability through both channels. Considering that my results indicate that capital controls have no significant influence on the finan-cial stability of developing countries, I would argue that rerunning this analysis while distinguishing between controls on capital inflows and controls on capital outflows is crucial. Third, I would argue that of the two measures, the measure distinguishing between classes of capital controls holds the most promise in terms of finding a relationship between capital control use and macroeconomic stability. This measure came very close to statistical signifi-cance and it fared much better than the measure of intensity of capital controls. Lastly, a key criticism in my literature review is the fact that the capital controls literature generally does not focus sufficiently on the impact of capi-tal controls on developing countries. For instance, I criticize Alfaro, Kalemli-Ozcan, & Volosovych (2007) saying that they pay insufficient attention to the most vulnerable countries. After going through the rigors of finding workable data, my conclusions back up the capital controls literature in that the pau-city of data explain the literature’s focus on devel-oped countries. In conclusion the implications of my initial find-ings are the same for both academia and policy. Essentially, my regressions indicate that capital controls are not a valid macroeconomic tool even during a financial crisis, as much of the literature has already decided. Rather, during a financial cri-sis, developing countries would be wise to heed the advice of Forbes (2007a), and focus on using skill-ful macroeconomic policy to preserve or enhance

financial stability.

VIII. References

Alesina, A., & Tabellini, G. (1989). External debt, capital flight and political risk. Journal of Interna-tional Economics, 27(3), 199-220.

Alesina, Alberto, Vittorio Grilli, and Gian Maria Milesi-Ferretti. 1994. “The Political Economy of Capital Controls.” In Leonardo Leiderman and As-saf Razin (eds.), Capital Mobility: The Impact on Consumption, Investment and Growth. Cambridge: Cambridge University Press.

Alfaro, L., Kalemli-Ozcan, S., & Volosovych, V. (2007). Capital flows in a globalized world: The role of policies and institutions. In Capital Controls and Capital Flows in Emerging Economies: Policies, Practices and Consequences (pp. 19-72). University of Chicago Press.

Ariyoshi, A. (2000). Capital controls: country expe-riences with their use and liberalization. Intl Mon-etary Fund.

Bai, C. E., & Wei, S. J. (2000). Quality of bureaucra-cy and open-economy macro policies (No. w7766). National bureau of economic research.

Campion, M. K., & Neumann, R. M. (2004). Com-positional effects of capital controls: evidence from Latin America. The North American Journal of Eco-nomics and Finance, 15(2), 161-178.

Corsetti, G., Pesenti, P., & Roubini, N. (1999). What caused the Asian currency and financial crisis?. Ja-pan and the world economy, 11(3), 305-373.Counting their blessings. (2010, January 2). The Economist. Retrieved February 27, 2015]

De Gregorio, J., Edwards, S., & Valdes, R. O. (2000). Controls on capital inflows: do they work?. Journal of Development Economics, 63(1), 59-83.

Domowitz, I., Glen, J., & Madhavan, A. (2001). In-ternational evidence on aggregate corporate financ-ing decisions. Financial Structure and Economic Growth: A Cross-Country Comparison of Banks, Markets, and Development. MIT Press, Cambridge and London, 263-295.

Edison, H. J., & Warnock, F. E. (2003). A simple measure of the intensity of capital controls. Journal

Page 56: Visible Hand

56 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

of Empirical Finance, 10(1), 81-103.

Edwards, S. (1999). How effective are capital con-trols? (No. w7413). National bureau of economic research.

Edwards, S. (2000). Interest rates, contagion and capital controls (No. w7801). National bureau of economic research.

Eichengreen, B., Mussa, M., Dell’Ariccia, G., De-tragiache, E., Milesi-Ferretti, G., & Tweedie, A. (1999). Liberalizing Capital Movements: Some Analytical Issues. IMF Economic Issue, (17). Re-trieved February 27, 2015

Epstein, G. A. (1992). Structural determinants and economic effects of capital controls in the OECD.

Forbes, K. J. (2007a). The microeconomic evidence on capital controls: no free lunch. Capital Controls and Capital Flows in Emerging Economies: Poli-cies, Practices and Consequences (pp. 171-202). University of Chicago Press.

Forbes, K. J. (2007b). One cost of the Chilean capital controls: increased financial constraints for smaller traded firms. Journal of International Eco-nomics, 71(2), 294-323.

Forbes, K. J. (2008). Capital Controls. In The New Palgrave Dictionary of Economics, Second Edition.

Gadanecz, B., & Jayaram, K. (2008). Measures of financial stability–a review. Irving Fisher Commit-tee Bulletin, (31), 365-383

Gallego, F., Hernández, L., & Schmidt-Hebbel, K. (1999). Capital controls in Chile: effective? Effi-cient? (No. 59). Banco Central de Chile.

Grilli, V., & Milesi-Ferretti, G. (1995). Economic Effects and Structural Determinants of Capital Con-trols. IMF Staff Papers, 42(3), 517-551. Retrieved February 27, 2015

Klein, M. W. (2012). Capital controls: Gates versus walls (No. w18526). National Bureau of Economic Research.

Kose, M. A., Prasad, E., Rogoff, K., & Wei, S. J. (2009). Financial globalization: A reappraisal. IMF Staff Papers, 8-62.

Korinek, A. (2011). The new economics of pruden-tial capital controls: A research agenda. IMF Eco-nomic Review, 59(3), 523-561.

Kose, M. A., Prasad, E. S., & Terrones, M. E. (2003). Financial integration and macroeconomic volatility. IMF Staff papers, 119-142.

Magud, N., & Reinhart, C. M. (2006). Capital con-trols: an evaluation (No. w11973). National Bureau of Economic Research.

Milesi-Ferretti, G. M. (1998). Why Capital Con-trols? Theory and Evidence. Positive Political Econ-omy: Theory and Evidence. Cambridge: Cambridge University Press.

Mody, A., & Taylor, M. P. (2003). International cap-ital crunches: the time-varying role of informational asymmetries.

Montiel, P., & Reinhart, C. M. (1999). Do capital controls and macroeconomic policies influence the volume and composition of capital flows? Evidence from the 1990s. Journal of international money and finance, 18(4), 619-635.

Neely, C. J. (1999). An introduction to capital con-trols. Federal Reserve Bank of St. Louis Review, 81(November/December 1999).

Nelson, W. R., & Perli, R. (2007). Selected indica-tors of financial stability. Risk Measurement and Systemic Risk, 4, 343-372.

Nurkse, R. (1944). International currency experi-ence: lessons of the interwar period (No. 4). League of Nations.

Ostry, Jonathan, Atish Ghosh, Karl Habermeier, Marcos Chamon, Mahvash S. Qureshi, and Dennis B.S. Reinhardt. 2010. “Capital Inflows: The Role of Controls.” IMF Staff Position Note, SPN/10/04.

Prasad, E., Rogoff, K., Wei, S. J., Kose, M. A. (2003),“Effects of financial globalization on devel-oping countries: some empirical evidence”. IMF Occasional paper, 220.

Quinn, D. P., & Inclan, C. (1997). The origins of financial openness: A study of current and capital account liberalization. American Journal of Political Science, 771-813.

Page 57: Visible Hand

Fall 2015 | The Visible Hand | 57

Volume xxiV. issue i.

Rodrik, D., & Velasco, A. (1999). Short-term capi-tal flows (No. w7364). National bureau of economic research.

Rogoff, K. S. (2002). Straight talk—Rethinking capital controls: When should we keep an open mind?. Finance & Development, 39(4).

Rogoff, K., Wei, S. J., & Kose, M. A. (2003). Effects of financial globalization on developing countries: some empirical evidence (Vol. 17). Washington, DC: International Monetary Fund.

The World Bank, Country and Lending Groups. (n.d.). Retrieved April 19, 2015, from http://data.worldbank.org/about/country-and-lending-groupsTobin, J. (1978). A proposal for international mon-etary reform. Eastern economic journal, 4(3/4), 153-159.

Valdés-Prieto, S., & Soto, M. (1998). The effective-ness of capital controls: theory and evidence from Chile. Empirica, 25(2), 133-164.

Page 58: Visible Hand

58 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Khizar Qureshi Massachusetts Institute of Technology

Value-at-Risk: The Effect of Autoregression in a Quantile Process

I. Introductory Remarks

Several recent financial disasters have made clear the necessity for a diverse set of risk management tools. Traditional models of risk management often rely on trivial probabilistic tools and often fail to relax key assumptions for the underlying statistics. An effective tool for risk management should be a withstanding measure of uncertainty robust to a large set of situations. Moreover, it should be suit-able for its users, adaptive to complex situations, and compatible for various sample sizes. Risk man-agement is not simply a tool to establish the upper bound on a loss, but also a preventative measure that should lead to the development of an informed decision-making process. Perhaps the most well known tool for risk man-agement amongst finance practitioners is Value atRisk (VaR). Conceptually, VaR measures the supre-mum of a portfolio’s loss with a particular level ofconfidence. Consider a portfolio of unitary value with an annual standard deviation of 15%. The 95%daily VaR is simply the product of the daily stan-dard deviation and the total value, or $12:350:15 =$0:3525. A 95% daily VaR of 0.3525 means that, if our day were hypothetically conducted an infinitenumber of times, the loss of our one dollar portfolio would be greater than 0.3525 with probability 0.05.We say with 95% confidence that on a given day, the maximum loss implied by VaR is 0.3525 for theone dollar portfolio. A more rigorous definition of VaR is a particular quantile of future portfolio values, conditional on

current information. In particular, we say that

where yt is a time t return, is the set of available information in a weak sense, and is the confidence level or probability. The immediate considerations for a functional mod-el include a closed-form representation, a set of welldefined intermediary parameters, and a test to validate the proposed model. In advance of our model proposition(s), we will review and evaluate existing models for VaR. The remainder of the paper is organized as fol-lows. Section II will introduce and evaluate existingmodels for Value at Risk, all of which will guide us in constructing CAViaR. Section III will cover the notion of Conditional Autoregression, the under-standing of which is critical to realistic non-i.i.dprocesses. Section IV will introduce various meth-ods of testing quantile regression, which will enable us to compare the set of well-known models with ours. Section V will focus on the empirical test of CAViaR on IBM, GM, and SPX time series data. We will conclude with section VI.

II. Existent Models VaR has become a quintessential tool for portfo-lio management because it enables funds to estimate the cost of risk and efficiently allocate it. Moreover, a growing number of regulatory committees now requires institutions to monitor and report VaR frequently. Such a measure discourages excessive leverage and increases transparency of the “worst-

Value-at-Risk (VaR) is an institutional measure of risk favored by financial regulators. VaR maybe interpreted as a quantile of future portfolio values conditional on the information available, wherethe most common quantile used is 95%. Here we demonstrate Conditional Autoregressive Value atRisk, first introduced by Engle, Manganelli (2001). CAViaR suggests that negative/positive returnsare not i.i.d., and that there is significant autocorrelation. The model is tested using data from 1986-1999 and 1999-2009 for GM, IBM, XOM, SPX, and then validated via the dynamic quantile test.Results suggest that the tails (upper/lower quantile) of a distribution of returns behave differentlythan the core.

Page 59: Visible Hand

Fall 2015 | The Visible Hand | 59

Volume xxiV. issue i.

case scenario”. While VaR , methods of estimation vary across both markets and firms. For ease and convenience of terminology, we will refer to all in-stitutions and funds concerned with monitoring VaR as “holders”. One component that often varies between differ-ent VaR models is the method by which the distribu-tion of portfolio returns is estimated. A rudimentary example was readily introduced at the beginning of the paper, in which returns were assumed to be in-dependently and identically distributed, or i.i.d. As we will see, however, returns almost never follow a martingale process, but rather, are Markov. A port-folio’s performance on any given day almost always effects the performance on subsequent days. Thus, the probability of observing a specific return or vari-ance as an event is dependent on the probability of observing the same event one period prior. The cal-culation of returns falls within two categories:

1. Factor Models: Here, a universe of assets are studied for their factors, all of which are correlated. Thus, the time variation in the risk of the portfo-lio is derived from the volatility of the correlations. A well known example is the Fama-French four-factor model. The approach, however, assumes that negative returns follow the same approach as non-negative returns. Perhaps an even more alarming assumption by such models is the homoscedasticity between returns per unit risk. 2. Portfolio Models: Here, VaR is instantaneously constructed using statistical inference of past port-folios. Then, quantiles are forecast via several ap-proaches, including Generalized Autoregres-sive Conditional Heteroscedasticity (GARCH), expo-nential smoothing, etc., most of which incor-rectly assume normality. Moreover, the set of models as-sume that after a certain amount of time, a particular historical return has probability zero of recurring.

We can deduce without an empirical demonstra-tion that that portfolio models will underestimate VaR after time T, and factor models will fail to account for autoregression. Interestingly enough, the last decade has motivated the introduction of extreme quantile estimation, and the notion of as-ymptotic tail distributions. Many of these models, however, are only representative of especially low (< 1%) quantiles, and do not relax the weak i.i.d. assumption.

III. Conditional Autoregressive Value-at-Risk We now address many of the concerns above with

CAViaR. In particular, we will study the asymptotic distribution, account for autocorrelation, and do so under various regimes. Suppose that there exists an observable vector of returns, {yt}

Tt=1. We denote θ

as the probability associated with the Value-at-Risk. Letting xt be a vector of time t observable variables (i.e. returns), and β be a vector of unknown param-eters, a Conditional Autoregressive VaR model may take the following form:

i. Interpretation The quantile of portfolio returns at a time t is a function of not only the past period returns, but also the past period quantile of returns. That is, ft(β) = ft(xt-1,βθ. The lag operator in the third term, l, links the set of available information at t-j to the quantile of returns at t. The autoregressive function, ftt-1(β), creates a smooth path between time-oriented quan-tiles. The first term, β0, is simply a constant. The ex-ample provided at the beginning of the paper, which does not account for autoregression, would simply remain a constant: ft(β) = β0. Now that we have de-veloped an understanding of the basic form for a CAViaR model, we will explore a few examples.

ii.Adaptive Model In general, an adaptive model follows the form

The adaptive model successfully accounts for in-crease in expected VaR. Whereas the traditional model would change only with a change in port-folio value, the adaptive model increases the Val-ue-at-Risk by unit one whenever it is exceeded. Moreover, it decreases VaR by unit one if initial estimates proved to be too high. It is clear to see that such a conditional adjustment, in the form of a step function, would provide for a more accurate myopic estimation. However, because all changes are of magnitude one, the adaptive model overlooks large deviations in returns upwards or downwards. For example, consider a state in which the portfolio halved in value for three consecutive days. While the portfolio has been left at an eighth of its value, the VaR only increased by three units from its value at t=0.

iii. Symmetric Absolute Value

Page 60: Visible Hand

60 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

The SAV model takes the form

The SAV is an autoregressive model in which a change in returns, regardless of direction, results in a change in VaR. What is particularly useful about this model is its ability to generalize movement in portfolio value. However, it is because of this very feature that SAV should not be used as a primary tool for measurement. Consider a series of large de-viations in portfolio value, alternating upwards and downwards. While the long-run change in value is zero, the VaR implied by SAV would be unrealisti-cally high. Similarly, a series of small deviations would imply an unrealistically low VaR.

iv. Asymmetric Slope The AS model takes the form

The AS model is intended to capture the asymmetric leverage effect. Specifically, it was designed to de-tect the tendency for volatility to be greater follow-ing a negative return than a positive return of equal magnitude. The model relies on magnitude of error, rather than squared error, as in GARCH.

v. Indirect GARCH(1,1) The Indirect GARCH model takes the form While

the GARCH model is estimated by maximum like-lihood, Indirect GARCH is estimated via quantile regression.

vi. Regression Quantiles Thus far, we have understood the general form of a Conditional Autoregressive Value at Risk model, and have also seen several possible forms. The pri-mary dfference between any pair of CAViaR models is the organization and treatment of βi, the regressive parameter. In the case of SAV, we were interested in a β that reflected magnitude of change in portfolio returns, where in AS, we were interested in only extreme ends of a series of returns. However, how are the underlying parameters actually measured? Koenker and Basset (1978) introduced the notion of a sample quantile to a linear regression. Consider a sample of observations y1,...,.yT generated by the linear model

where xt is a length p vector of regressors (i.e. re-turns), and Qθ(εθt|xt) is the θ-quantile of εθt condi-tional on xt. Consider the linear representation of an adaptive process: ft(β) = xtβ. The θ regression quan-tile is to satisfy the objective

Qualitatively, we adjust beta until VaR is no lon-ger exceeded, or “hit” by a certain amount. Such a condition is satisfied when the observed indicator variables are 0. To account for the set of available information,

Theorem 1 (Consistency). For generalized model

(above), in probability, where solves

Proof. Please see AppendixTheorem 2 (Asymptotic Normality). In testing for asymptotic normality, for statistic T,

where

Proof. Proof left as exercise for reader

IV. Testing Quantile Models

While the expanded set of models available now account for the autocorrelation of returns, as well as large deviations, they must be tested. Given a new observation, the model remains valid if P[yt < ft(β

0)] =θ. If such a condition holds for the entirety of the time series, then it is proven valid. As shown by Christoffersen (1998), such a method is equivalent to testing for the independence of indica-tor variables for the same condition. In other words,

Page 61: Visible Hand

Fall 2015 | The Visible Hand | 61

Volume xxiV. issue i.

[I(yt < ft(β0))]T

t=1={It}T

t=1.While this provides for a natural test of forecasting models, it does not fully assess the validity of quantiles. To test conditional quantile models, we introduce a representative indi-cator variable that changes with the quantile itself. Define a sequence of independent random variables {zt}

Tt=1=1 such that

P(zt = 1) = θ P(zt = -1) = 1 - θExpressing positive or negative autocorrelated re-turns in terms of zt indeed accounts for the prob-ability of exceeding a quantile. However, whilst the unconditional probabilities are uncorrelated, the conditional probabilities for a hit still depend on one another. Because these tests evaluate the lower bound of the VaR in the weakest sense, we work towards defining a dynamic quantile. Let

Hitt(β0) = I(yt < ft(β

0)) - θHitt(β

0) assumes a value of (1-θ) for underestima-tions of VaR, and -θ otherwise. Notice that theexpected value is zero, and that there should be no autocorrelation in the values between successive hits. For our first test, we determine whether the test statistic, T-1/2X’( )Hit( ) is significantly differentfrom zero where X t ( ), t = 1,..., T may depend on , and is q-measurable information (i.e. returns).Suppose we wish to test the significance of an entire set of data along several β simultaneously. Then,let MT = (X’(β0))-E[T-1X’(β0)H 'f(β0)]D-1

T x 'f(β0) Here, H is a diagonal matrix with binaryindicators conditioned on available information. Such a test would be run both in-sample, as well asout-of-sample. We will define and prove the condi-tions of each.

Theorem 3 (In-Sample Dynamic Quantile Test). If the assumptions made (see appendix) are valid, thefollowing holds

Moreover,

where is the difference between and a function of the gradient of f

Proof. Proof left as exercise for reader Essentially, the DQ test above tests whether or not the test statistic follows a normal distributionin the sense of an identity matrix, and whether the set of all dynamic quantiles in-sample follow a

chi-squared distribution. We now shift our focus to a test statistic for dynamic quantiles out-of-sample.

Theorem 4 (Out-of-Sample Dynamic Quantile Test). Let TR denote the number of in-sample obser-vations and NR denote the number of out-of-sample observations. Then

Proof. Proof left as exercise for reader Use of the dynamic quantile tests allows for an estimation of the independence in "hits". The idealquantile test would be one in which all hits (yt < V aRt) are independent. Regulators would be able tochoose between different measures of VaR when evaluating a portfolio.

V. Empirical Results

The historical series of portfolio returns were studied for four different regimes of CAViaR. A total of 2,553 daily prices from WDRS/CRSP for General Motors (GM), International Business Ma-chines (IBM), and the SP 500 (SPX) were retrieved. Daily returns were computed as 100 times the dif-ference of the log of the prices. Two sets of ranges were used: one from April 7, 1986 to April 7, 1999, and the other from April 7, 1999 to June 1, 2009. From the total set of daily prices, 2,253 were used in-sample, and the last 300 were used out-of-sam-ple. Both 1% and 5% day-end VaR were estimated for each of the four regimes. The optimization was completed using MAT-LAB R2015 and Gurobi Optimizer under a Quasi-Newton Method. The loops for recursive quantile functions (i.e. SAV) were coded in C.

i. Optimization Methodology Using a random number generator, n vectors were generated, each with uniform distribution in [0; 1].The regression quantile function was computed, and from the n vectors, m n vectors with the lowestregression quantile criteria were selected as initial values for optimization. For each of the four re-gimes, we first used the simplex algorithm. Follow-ing the approximation, we used a robust quasi-New-ton method to determine new optimal parameters to feed into the simplex algorithm. This process was repeated until convergence, and tolerance for the re-gression quantile was set to 10-10.

Page 62: Visible Hand

62 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

ii. Simplex Algorithm The algorithm operates on linear programs in standard form to determine the existence of a feasi-ble solution. The first step of the algorithm (Phase 1) involves the identification of an extrema as a guess.Either a basic feasible solution is found, or the fea-sible region is said to be empty. In the second step7 of the algorithm (Phase 2), the basic feasible solu-tion from Phase 1 is used as a guess, and either anoptimal basic feasible solution is found, or the solu-tion is a line with infinite (unbounded) optimal cost.

iii. Quasi-Newton MethodQN methods are used to locate roots, local maxima, or local minima if the Hessian is unavailable at each step. Rather, the Hessian is updated through ana-lyzing gradient vectors. In general, a second order approximation is used to find a function minimum. Such a taylor series is

for a gradient Δf and a Hessian approximate B, the gradient of the approximation is

with root

We seek the Hessian Bk+1=argminB||B-Bk||V where V is a positive definite matrix defining the norm

VI. CAViaR Results

We now review Conditional Autoregressive Value-at-Risk, methodology introduced by (Engle, Manganelli, 2001). First we review results from 1986-1999. Then we extend to 1999-2012. To test the significance of autoregression, we study values of βt-1, as well as p-values, for three regimes. Finally, we compare results of conditional VaR to uncondi-tional VaR., the traditional risk measurement tool.

i. Interpretation As previously discussed, the arrival of informa-tion creates uncertainty, and increases VaR. While this is true in both the conditional and unconditional case, it is emphasized in the former. It is well-known that the period between 1986 and 1999 was volatile, and was affected by events such as LTCM and Glob-al crises in Russia and Asia. If VaR truly increases

with the arrival of information, then it is sensible tosee the peak in 1987, where all positive beta assets faced an increase in systemic risk with the marketcrash. The adaptive regime adjusts for changes in VaR, so the momentary increase in 1987 was givenless weight. It is also interesting to note that the autoregres-sive models capture non-systemic risk. While themarket crash in 1987 affected all positive-beta as-sets, we also see less severe increases in VaR during1999. With high probability, this is due to an id-iosyncratic event–a total inventory recall by GM. Such a recall likely created uncertainty in expected cash flows, thus increasing the periodic volatility of the stock. The persistent volatility was exponen-tially weighted in the adaptive regime, resulting in a large increase towards the end of the period of study.

ii. Interpretation Information in the 2000’s was much more read-ily available than it was during 1980-2000. Conse-quently, a rapid digestion and reflection of infor-mation in asset prices may have resulted in more dynamic expectation. We see immediately that conditional VaR is much lower, indicating that the arrival of information did not induce as much uncer-tainty as it did in the decade prior.

We are aware from the previous iteration that VaR increased in 1999 across all adaptive regimes. Fromthis prior, it is sensible to see high VaR from the very first year of data, given that time is continuous.While the increased natural filtration of information

Page 63: Visible Hand

Fall 2015 | The Visible Hand | 63

Volume xxiV. issue i.

suggests lower VaR, we infer sources based onbackward-looking bias. Aside from the vehicle re-call in 1999, the market faced a "mini-crash" in theearly 2000’s. This is better known as the "bubble

burst". Further, the financial crisis in the 2008-2009period resulted in an increase in conditional VaR. It is interesting to note the remarkable similarity across the four regimes, indicating an increase in the mean reversion coefficient of volatility.

VII. Broader Interpretation

The arrival of news results in an expansion of

information available. If the market is truly effi-cient, asset prices will reflect the expectations of those exposed to this information (Fama, 1997). However, expectations are not always dynamic, and integration of information may not be continuous. Consequently, volatility, a form of uncertainty, on a particular day will be autoregressive, and depend on volatility from previous days. It becomes necessary to use autoregression when calculating the expected loss within a p-quantile. We demonstrate several adaptive models, includ-

ing Symmetric Absolute Value, Asymmetric Slope,and Indirect GARCH (1,1). We recognize that while all of these models satisfy the requirements ofautoregression, they differ in their treatment of βt, the autoregressive parameter. When evaluating Value-at-Risk, conditional on information, we must carefully choose the model, and understand theunderlying assumptions. It is well known that for a monotonically increas-ing cumulative distribution function, an increase in p will cover a larger portion of the distribution of risk. An immediate consequence of this is that the 99% VaR exceeds the 95%. We re-confirm the no-tion. We show that conditioning on the arrival of information may increase or decrease VaR, depend-ing on the change in expectation. We contrast con-ditional against unconditional VaR, and autoregres-sive against independent VaR. In both contrasts, the former is a more proper treatment.

VIII.Conclusion

We have demonstrated the autoregressive prop-

Page 64: Visible Hand

64 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

erties of VaR conditioned on information available. Existing methods for calculating VaR fail to relax several assumptions including (i) the i.i.d. behavior of returns and (ii) the large deviations at the tail of

the return distribution. CAViaR addresses both is-sues, and suggests that there is significant autocor-relation for the data used (GM, IBM, SPX 1999-2009). Regressive parameters (βi) were estimated by minimizing the regression quantile loss function,

and the models were tested via dynamic quantiles. The worst performing method was the adaptive method, which failed to detect poor returns in 1999.The best performing method was Asymmetric Slope, which captured the larger effect of negative returnsvs. positive returns on VaR. Symmetric Absolute Value (SAV), which does not segregate positive and negative returns, illustrated general portfolio move-ment. Indirect GARCH, which does not distinguishtails of distributions, overemphasized movement in 1999, 2001, and 2009. These very same results are reflected in the 1 % SPX News curves, which study the effect of news-driven returns on the SP 500. Standard, non-conditional quantile regressions were studied for XOM, a large-cap stock similar to GM. Without autoregressive properties, 95 % VaR was underestimated relative to the CAViaR case. Such a result motivates the use of multiple risk mea-surement tools, each carrying different treatment of (βi) and underlying assumptions. The most typical use of VaR involves determin-ing the expected loss with 95 % certainty. While

useful as a bare approximation, a more proper anal-ysis of risk should be carried across various quan-tiles and multiple distributions. Moreover, the loss function for a given day should be clustered within time frames. We have re-confirmed that volatility is autoregressive and conditional on the arrival of information. Treating value at risk independently

Page 65: Visible Hand

Fall 2015 | The Visible Hand | 65

Volume xxiV. issue i.

Page 66: Visible Hand

66 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

will almost surely underestimate maximum losses during periods of high volatility and overestimate during periods of low volatility. The use of multiple tools may be beneficial to financial institutions, the individuals they may represent, and the health of market participants as a whole.

IX.Extensions

Future applications include smaller time frames, different sets of stocks, an extension to the multi-variate case, and quantile sensitivity. It is also worth investigating the implied volatility of out-of-the-money options calculated under the conditional and non-conditional regime. We expect the skew to car-ry more weight in the conditional regime to account for disaster risk. There also exists potential to improve the meth-odology used within the paper. Namely, -Cost Regularization of Autoregressive parameters

-Local inference of Hessian matrices to employ in-terior point methods for optimization

X. References

Buchinsky, M., Estimating the Asymptotic Covari-ance Matrix for Quantile Regression Models., Jour-nal of Econometrics, 303-338, 1995.

Chernozhukov, V., Specification and Other Test Processes for Quantile Regression., Stanford Uni-versity, Mimeo, 1999.

Christoffersen, P.F., Evaluating Interval Forecats., International Economic Review, 841-862, 1998.

Engle, R.F., New Frontiers for ARCH Models, Jour-nal of Applied Econometrics, 425-446, 2002.Engle, R.F. and Ng. V, Measuring the Testing the Impact of News On Volatility, Journal of Finance,1749-1778, 1993.

Fama, E.. Market Efficiency, Long-Term Returns, and Behavioral Finance, University of Chicago,1997.

Koenker, R. and Basset, G., Regression Quantiles, Econometrica, 33-50, 1978.

Manganelli, S., and Engle, R.F., A Comparison of Value at Risk Models in Finance, Risk Measures

for the 21st Century, U.K. Wiley, 2004.

Weiss, A. Estimating Nonlinear Dynamic Models Using Least Absolute Error Estimation, Economet-ric Theory, 46-68, 1991.

For appendices and other notes, please refer to the electronic issue at http://orgsync.rso.cornell.edu/org/ces/Publication

XI. Appendix

Page 67: Visible Hand

Fall 2015 | The Visible Hand | 67

Volume xxiV. issue i.

Page 68: Visible Hand

68 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

in veterans, abilities to work. We consider two fac-tors in particular: government disability pensions and residence in a large city. Both reflect ways of providing assistance - the former with an emphasis on decentralized empowerment of veterans to reha-bilitate themselves, and the latter with an emphasis on access to centralized services and communities that could assist rehabilitation. We develop on exist-ing literature - which focuses on measures of real-ized labor participation - by examining change in the veterans’ ability to participate in labor (or their ”rehabilitation”) as measured by medical examin-ers. Existing literature has found that government assistance may incentivize veterans to not enter the labor force, and terminating assistance programs could increase labor participation. Measuring ability to work embodies far less moral hazard than labor participation; regardless of whether a veteran choos-es to remain unemployed, most will not, or cannot, choose to remain ill or wounded. If public intervention succeeds in rehabilitating veterans, it may be necessary, even if it also disin-centivizes work, because enabling veterans to work is the crucial first step to integrating them into the workforce. If public intervention does not succeed in rehabilitating veterans, then we can safely assume that the affect of aid programs on veteran unemploy-ment is the same as the affect of aid programs on general unemployment, and it may be counterpro-

I. Introduction

Among the damages of war, which include such devastating afflictions as famine, political instabili-ty, and the destruction of national infrastructure, one of the most persistent is the cost of forcing many of the nation’s most fit laborers into unemployment due to disabilities acquired while serving in the mil-itary. Veteran disability has moral and economic im-plications, and, in recognition of its importance, the United States government is deeply invested in ad-dressing it. In 2013, about 3.5 million of the Ameri-cas 22 million veterans had some disability prevent-ing them from working at full capacity, and Veterans Affairs disability payments amounted to about $54 billion, adjusted for inflation to 2014 dollars (”Vet-erans’ Disability Compensation”). While some veterans valued for the skills they acquired in the military, many veterans suffer from post-traumatic stress disorder, drug and alcohol abuse, feelings of isolation and depression, and physical injuries sus-tained during engagements. The unemployment rate for veterans in the U.S. averaged 7.2% in 2014, ap-proximately one percentage point above the rate for the civilian population (Mutikani). This paper examines the significance of factors that might contribute to veteran reintegration into the workforce by looking at whether these factors have historically had a significant effect on changes

Alex Foster & Adam SuditUniversity of Chicago

Veteran Rehabilitation: A Panel Data Study of the American Civil War

Every year, millions of technically-trained, working-age Americans are rendered unemployable by dis- abilities they acquired while serving the nation’s military. In hopes of alleviating this moral and economic crisis, the US government spends tens of billions of dollars on rehabilitation programs, but the comparative effectiveness of these treatments is hardly known. This paper measures the historical effectiveness of two rehabilitation treatments: having access to the resources of a large city and receiving government pensions in cash. We examine history in order to see a full picture; panel data from the American Civil War allows us to track direct indicators of rehabilitation, as measured by the change in a veteran’s doctor-evaluated abil-ity to work (rather than labor participation, which is subject to moral hazard) in the decades following the war. We find that both treatments were significant in promoting rehabilition for Civil War veterans over the two decades following the end of the war. We hope that this result can serve as both a fascinating historical insight and a potential lesson for policymakers today.

Page 69: Visible Hand

Fall 2015 | The Visible Hand | 69

Volume xxiV. issue i.

ductive. An obstacle to studying rehabilitation is the need for longitudinal data on each wounded vet-eran. With this data, we can measure treatment ef-fects over a long period of time. Such data on vet-erans returning from recent engagements are not yet available; non-aggregated national census data is still private for censuses after 1940 according to the Federal Census Burueaus policy of witholding per-sonally identifiable information until 72 years after it was collected (”The ’72-Year Rule’”). Therefore, we study rehabilitation in the context of the Ameri-can Civil War. We use the Union Army Data Set, which provides panel data, allowing us to map each soldier’s ability to work throughout his lifespan, in-cluding in the decades following the war. With the dataset, we can also relate the change in his abil-ity to work with variables describing health, demo-graphics, economics, and geography. The American Civil War was fought from 1860 to 1865 and engaged an enormous proportion of the nation’s workforce. 60% of all northern white men born between 1822 and 1845 served in the Union Army (Costa 1998, ”Appendix A”). To accommo-date the generation of survivors seeking to reinte-grate themselves into society, the first national pen-sion program of its kind was formed. Through the records of this pension program and matches from it to other data sources, we obtain unique insight from lifespan panel data, which research on more recent military engagement lacks. We find that both of our independent variables of interest - receiving a pension and living in a large city - had a significant and positive effect on reha-bilitation rates for Civil War veterans over the two decades following the end of the war.

II. Literature Review

Several existing papers guided our research, covering topics including: the relationship between general welfare programs and labor supply; the re-lationship between disability insurance and labor supply; and the relationship between proximity to urban areas and access to health care services. The literature review will be organized based on these topics. We begin by reviewing some of the literature on the relationship between welfare and labor supply. Since the program AFDC (Aid to Families with De-pendent Children), until its dissolution in 1996, was one of the largest programs in the American wel-fare system, many studies prior to 1995 focused on the effects of the AFDC on labor supply. Almost all

of these studies found that the AFDC program de-creases labor supply by 10% to 50% of non-AFDC levels (Card, p.27). Since these results have been widely accepted, many more studies have analyzed other welfare programs. Hoynes (1996) analyzed the program AFDC-UP (Assistance to Families with Dependent Children-Unemployed Parent). He used data from a Survey of Income and Program Par-ticipation from 1983 to 1986 to measure dependent variables of labor force participation and AFDC par-ticipation, and independent variables of an AFDC guarantee and a guaranteed transfer amount after tax. It should be noted in this context that the stan-dard static model in labor economics at its simplest treats the benefits from a specific welfare system as equal to the guaranteed transfer of the program, minus a marginal tax on earned income and an ex-ogenous endowment of income (Card, p.2403). He found that overall the program had a very large neg-ative effect on the supply of labor (Card, ch 34 Table 4). Hagstrom (1996) analyzed the American Food Stamp program using a Survey of Income and Pro-gram Participation from 1984, with participation in the labor force and the Food Stamp program as his dependent variable, and the usual guaranteed Food Stamp level and accompanying marginal tax rate as his variables of interest. He found the negative ef-fect on labor force participation to be much lower, perhaps illustrating that in-kind transfers may not be as distortionary as direct cash transfers. Keane and Moffit (1998) analyzed the AFDC, Food Stamps, and subsidized housing programs among poor sin-gle mothers. Similar to other studies, they found that labor supply participation and degree of use in the three programs were the dependent variables with the programs’ corresponding guarantees as the variables of interest. They calculated an elasticity of substitution between these program guarantees and labor force entry of 1.82, and a negative income elasticity of -.21. Lastly, Meyer and Rosenbaum (2001) examined the AFDC, Food Stamps, and the Earned Income Tax Credit (EITC), with ”probabil-ity of working” as the dependent variable, and guar-antees from the AFDC and Food Stamp programs as the variables of interest, along with any potential incentives that the sample population would face in chosing to work. They found that these guarantees decrease the chances that those in the sample popu-lation would be employed. Analyses of TANF (Temporary Assistance to Needy Families) have also been done due to the negative relationship between AFDC guarantees and labor supply. TANF gradually replaced AFDC as the key program for government transfers to

Page 70: Visible Hand

70 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

low-income families in 1996 due to the Personal Responsibility and Work Opportunity Reconcilia-tion Act (Card, p.2423). Unlike AFDC, it has been tough to determine the effects of this new legisla-tion on labor supply because the program was man-dated in all 50 states. Thus, it has been quite dif-ficult to find exogenous variation in the program in order to avoid omitted variable bias. This has led to the use of a difference-in-differences approach in relevant literature to create variation among dif-ferent groups affected by the program. For one of the more novel uses of this approach in labor eco-nomics, see Card and Krueger (1994), who used the same method to show that an increase in New Jersey’s minimum wage had no negative effect on employment. Indeed, even this method has proven to be insufficient in robustly understanding the ef-fects of TANF on labor force participation. Ellwood (2000) analyzed the treatment effect of TANF on employment and earnings by using a difference in difference methodology with high-income individu-als in his control group, but could reach no defini-tive conclusion about the effects of TANF because of the confounding effects of the EITC (Card, ch 34 Table 5b). McKernan, Lerman, Pindus, and Valente (2000) found that TANF had a positive effect on employment by using women without children as a control. These results were confirmed by ONeill and Hill (2001), who used variation in the date of imple-mentation of the program across different states as a source of exogenous variation. We now turn to a brief review of literature on government compensation for disability specifical-ly, and the effects on labor supply. It should be men-tioned that many notable studies on the subject oc-curred during the 70s and 80s, since that was when many World War II veterans were fully enjoying the benefits of disability insurance. The two main disability insurance programs in the United States are the SSDI and the SSI. One particular subject of research interest to many in the field has been try-ing to quantify the magnitude of moral hazard that is caused by the two programs, since disabled indi-viduals who are able to enter the labor force may still try to enroll in the welfare program and succeed in getting benefits (Card, ch 51 p.3472). Gastwirth, in his unprecedented study in 1972, tried to provide counterfactual evidence as to how many individu-als receiving benefits from SSDI would actually be in the labor force without the program by studying labor force participation by men who are disabled but not receiving benefits (Card, p.3472). Swisher (1973) pointed out the severe measurement error in Gastwirths study; Gastwirth included disabled indi-

viduals with all degrees of severity in injury when in reality the overwhelming majority of men par-ticipating in SSDI were severely disabled. While Swisher’s argument was valid, he also made some mistakes in his own study. For example, Swisher classified men’s disability severity based on forms that they filled out asking if they were too disabled to work. Thus Swisher may have under estimated the moral hazard problem that he criticized Gastwirth for overestimating, albeit for a different reason. Also, both scholars failed to realize or fully com-pensate for the fact that men in the disabled control group who did not receive any benefits from SSDI still had access to other government programs, even if direct transfers were not always part of these pro-gram (Card, p.3474). Many of these issues shaped the subsequent methodologies of other economists studying this issue. Bound (1989) attempted to alle-viate this issue by using a group of disabled individ-uals who applied to the SSDI but were rejected. Be-cause one would think that those who were accepted into the SSDI system are less able to enter the labor force than those who were rejected, it is possible that the percentage of rejected disabled applicants who work may be the ”upper bound” on how much disabled people would work without SSDI (Card, p.3475). Bound then proceeds to show, in confirma-tion of conclusions reached by Goff (1970), Smith and Lilienfield (1971), and Treitel (1976), that that there is actually a low labor force participation rate among those disabled persons rejected from the program. This might indicate that disability welfare programs cannot account for all, or even the major-ity of a decrease of labor force participation within the population of disabled citizens. Of course it is important to keep in mind that there are other programs besides the SSDI that as-sist disabled people, so distortionary effects on la-bor force participation may still exists without the SSDI. Also, it is difficult to isolate confounding fac-tors when doing these studies over a long period of time due to evolving economic conditions and the initiation and termination of various government disability programs. Yet, a unique issue often arises when studying the effect of government transfers on the labor force participation of disabled citizens. It is difficult to create variation in the amount of ben-efits received by similar samples, which is quite es-sential to many studies in labor economics, because most people that are of a similar age often have the same access to programs, with the same probability of being accepted to these program and receiving specific transfers. Also variation across states, an oft used remedy for this issue, cannot be used for the

Page 71: Visible Hand

Fall 2015 | The Visible Hand | 71

Volume xxiV. issue i.

SSDI and SSI programs. Therefore, time series data is the most useful way to create variation within a program because most variation within samples occurs within income earned over a long period of time anyway (Card, p.3475-6). In this vein, Bound and Waidman (1992) concluded that at least in the 1970’s, older men entering disability welfare pro-grams in lieu of retiring can explain why so many older men that still were young enough to work left the labor force (Card, p.3476). Perhaps the most re-cently well-known researcher in the relevant litera-ture is Parsons (1980a,b). With the use of National Longitudinal Survey of Older Men, he calculated a ”SSDI replacement ratio” as a variable of interest in order to eventually estimate the elasticity of exit-ing the workforce with respect to how much benefits were given out to be somewhere between .49 and .93 (Page 3478). These results, assuming of course that the rationale used for the variable of interest are correct, imply that the SSDI might be able to explain the significant decline in labor supply by men be-tween 45-54 after World War II (Page 3479). Many other papers address the subject, but these studies mentioned give the main intuition that underlies the relationship between government transfers and la-bor participation by disabled people. We briefly examine some literature in urban economics by looking at one paper in particular by Blazer, Landerman, Fillenbaum, and Horner (1995). They study the use of health care services among the elderly in North Carolina, and find, among other things, that cost was a much greater barrier to get-ting medical care in rural areas than for elderly peo-ple in urban areas. As we see in the interpretations section, this is line with our results.

III. Data Source

The data used in this paper are drawn from the Union Army Data Set, a longitudinal data set of American Civil War veterans compiled by a team led by Robert Fogel at the University of Chicago beginning in 1981. Fogel and his team selected 332 companies of Union Army soldiers using a one-stage cluster sampling procedure from the complete list of Union Army regiments found in A Compen-dium of the War of the Rebellion by Frederick H. Dyer (1908). They then extracted names and identi-fying information from the Regimental Books (Re-cord Group 94) housed at the National Archives and Records Administration in Washington, D.C. Each soldier was linked to his Compiled Military Service Record, Pension record, Carded Medical Record, and U.S. Federal Census submissions. A unique

10-digit recruit identification number identifies each soldier throughout Union Army Data Set. The sample contains 40,000 northern, white soldiers serving in 332 companies of volunteer in-fantry, and it is representative of the northern, white male population of the United States at the time (”Union Army”). The Union Army as a whole was a representative cross-section of the northern, white population; mean wealth of the households of sol-diers is similar to that found in a random sample of the population, suggesting that military service was not very selective of men of lower socioeconomic status. In fact, 95% of soldiers in the sample were volunteers (Costa 2000, ”Understanding”). Condi-tions in the war were random, so survivors are a rep-resentative sample of the population as well (Costa 1998, ”Appendix A”). The data relevant to our study come primarily from Pension records. Congress began the pension program in 1862 to provide pensions both to drafted and volunteer recruits who were severely disabled as a direct result of military service (for a history of the Union Army pension program, see Costa 1998, ”The Evolution of Retirement”:197212). This initial program, which came to be known as the General Law pension system, was formed under pressure by the Grand Army of the Republic (GAR), a union of northern veterans and one of the most notori-ously powerful lobbies in the history of the United States. By 1870, $29,000,000 (9.4%) of government spending went towards veteran pensions, and as ad-dendums to the General Law system significantly extended and increased coverage, this expenditure grew through the end of the century (Fogel). 85% of soldiers who survived the war have a pension record. Desertion is the primary non-random rea-son soldiers might survive the war and not receive a pension. Furthermore, no other pension program existed concurrently with this program from 1862 until 1890 - the time window which we examine (Costa 1998, ”Appendix A”). Before 1890, pensions were awarded exclu-sively on the basis of disabilities incurred by Union Army veterans while serving in the war. A standard-ized national scale prescribed pension amounts for various degrees of disability with an emphasis on how the disability would affect the veteran’s ability to work. For example, under the 1862 General Law system, a veteran who lost an index finger would be deemed 3/8ths disabled and, at least under the 1862 General Law system, awarded three dollars a month in pension (Fogel). Pension officials soon found that many veterans suffered from war related ailments not represented in the 1862 General Law.

Page 72: Visible Hand

72 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

These disabilities were recorded, but did not warrant pensions until subsequent addendums to the pension program were instated. These addendums through-out the century incentivized veterans to return for medical examinations after their initial pension ap-proval, since they would often qualify for compen-sation for ailments that were initially disregarded. The strong incentive to report disabilities for com-pensation makes it unlikely that many veterans had disabilities that went unreported. Disability then had to be verified by a board of three medical examiners. An inspector from the pension bureau would speak with the applicant, and also anyone who may have known the applicant, in-cluding neighbors, soldiers from the same unit, and personal doctors. Through this extensive process, the Pension Bureau went to great lengths to confirm or disprove that a veteran’s ailments were real and directly caused by the war. Indeed, by 1900 the most common ground for rejection (in 24% of cases) was judgment that the veteran’s disabilities were unre-lated to the war (Costa 1998, ”Appendix A”). This system makes it unlikely that many veterans suc-cessfully reported fraudulent disabilities.Records of these medical examinations were kept both for men whose pension applications were re-jected and for men whose applications were accept-ed (Costa 2000, ”Understanding”).

IV. Data Cleaning

Values for all variables in the Union Army Data Set are indexed by the recruit identification num-ber of the soldier to whom the observation belongs and also an instance number. Instance numbers note each time a value was observed for the same recruit and variable throughout the many documents exam-ined by researchers. The instance numbers do not necessarily coincide with the chronology of the doc-uments (for example, the first instance of a pension amount being observed for recruit X may be from an 1890 pension ruling, though the pension amount listed under their second instance was from an 1870 ruling). To produce cleaned spreadsheets of panel data with variable values indexed by recruit iden-tification number and year, we match observations with the years of their original document. We then select data from particular years as needed to create variables that are associated with observations from particular times. Lastly, we align all of our vari-ables according to each soldier’s recruit identifica-tion number. After these treatments, an observation might be, for example, $10, labeled as the pension that recruit #52402 was awarded in 1870.

V. Methodology

We develop a linear OLS model of change in a veteran’s ability to work as a function of how much he received in pension, whether he lived in a large city, and control variables. The base model is as fol-lows:

D.AbilitytoWork = β0 + β1PensionTransfer + β2BodyWound + β3Illness + β4MildSeverity +

β5SeriousSeverity + β6Literate + β7YearsSince1865 + β8InitialAbilitytoWork +

β9FromTop20City + U

Where we do not assume U is homoskedastic - please see our limitations section for a detailed dis-cussion of our residual error. Change in Ability to Work is our measure of re-habilitation. We calculate it by taking the difference between a veteran’s ability to work according to their latest pension application submitted between 1865 and 1870 and their ability to work according to their latest pension application submitted between 1880 and 1890. Ability to work is a percentage value determined by the medical examiners. The variable denotes the soldier’s ability to work at full productivity. For example, an ability to work value of 30 implies that the veteran could work at 30% full productivity due to his disability, or he was 70% disabled from working. Veteran disability com-pensation in the United States is still today gradu-ated according to a percentage quantification of the veteran’s disability in this way (”Compensation”). Thus if a veteran’s 1870 pension application reports an ability to work of 30, and his 1890 pension ap-plication reports an ability to work of 90, we would record a 60% change in ability to work. Pension Transfer is the monthly pension re-ceived by the veteran averaged over all years of the rehabilita- tion period, with each year’s monthly pension level adjusted for inflation to 2014 dollars according to historical Consumer Price Index con-version factors proposed by Robert Sahr at Oregon State University (Sahr). In years before the veteran’s first pension ruling, we infer his pension amount to be zero. From Top 20 City indicates whether the veteran lived in one of the twenty largest American cities by population according to the 1870 federal census data (”Population”). The cut-off of twenty cities was found to be slightly more explanatory to rehabilita-tion than using a cut-off of the top five, ten, or fifty cities, though the significance of living in a large city was similar regardless of the cut-off. We control for whether the veteran’s disability

Page 73: Visible Hand

Fall 2015 | The Visible Hand | 73

Volume xxiV. issue i.

was a wound and/or an illness and its severity ac-cording to medical records. Qualitative descriptions of severity are categorized as unsubstantial (if de-scribed in the data: ”slight” or ”no wound”), mild severity (if described: ”acute,” ”fracture,” ”flesh,” or ”slight fracture”) and serious severity (if described: ”chronic,” ”severe fracture,” ”severe,” or ”fatal”) in order to use indicator variables. These descrip-tors characterize the vast majority of the data; any other descriptions were discarded. Along a similar thought, Initial ability to work controls for the initial degree of the disability from which the veteran is recovering. Literacy is added as a control for consistency with existing literature, which has found significant in- teraction between education and responses to government transfers and participation in urban ser-vices. A designation of 1 implies the veteran is liter-ate. We consider the veteran illiterate if he was ever designated illiterate on any pension application; this presumably reflects their first pension application (before the rehabil- itation period) since it is un-likely the applicant forgot how to read in later years. Years since 1865 reflects how many years past 1865 our initial observation of their ability to work was made (the observations were made at various years from 1865 to 1870 since not all veterans applied for pensions in the same year). We suspect that veterans will naturally demonstrate significant rehabilitation in the years immediately following the war, so this variable controls for how much of that rehabilitation we observe and how much occurs before our obser-vation window.

Descriptive statistics for all variables can be found in Table 3.

VI. Results

We now turn to our results. Interpretations and discussion of these results will be reserved for the following section. Please refer to table 1 and 2 for a statistical summary of our results. We begin by examining our full model speci-fication, which is displayed in column (1) of Table 1. When all the regressors are included, we find the key results of our study. First and foremost, we find that pension transfers have a positive and very significant effect on changes in a recruit’s ability to work. An increase of one dollar (in 2014 dollars) in monthly pensions correlates with .04 increase in the degree of the recipient’s ability to work. Our other economic variable of interest, the veteran being lo-cated in one of the twenty most populous cities of

that time, also displays a positive and reasonably significant effect on the recruit’s change in ability to work. All else equal, veterans living in large cities could expect to experience a 10.764 degree higher change in ability to work than veterans outside of these twenty cities. We thus confirm our hypothesis that veteran rehabilitation is improved by both pen-sions and access to urban centers. The dummy variable indicating a wound of mild severity also has a significant and relatively large negative effect on the recruit’s change in his ability to work. It is worth noting the R2 measure of fit, which at .364 is reasonably high relative to our other specifications. In turning to column (2) of Table 1, we can-not emphasize enough the importance of controlling for initial ability to work. Intuitively, a veteran who comes out of the war already quite able to work has less opportunity to demonstrate positive change in his ability to work, so initial ability to work is cor-related with our dependent variable. Furthermore, medical examiners considered veterans’ initial ability to work in order to determine their pension amounts, so initial ability to work is correlated with our variable of interest. Therefore, failing to control for initial ability to work leads to omitted variable bias. This problem is also indicated by the higher residual standard error in the specification without initial ability to work. Table 3 shows our model specification without using controls for severity of ailments. One might think that these variables are irrelevant to the model with the inclusion of initial ability to work. Indeed, this inuition seems to be the case, as the exclusion of the severity controls does not dramatically alter our results.

VII. Discussion and Interpretation

First of all, we can say with certainty that pen-sion transfers after the Civil War had a slightly posi-tive and very significant correlation on improving a veteran’s ability to work. We therefore believe that government transfers toward disabled veterans aid-ed veterans in rehabilitation after the war. A disabled veteran could have used such transfers to pay for medicine or medical services, aiding in his recov-ery. In addition, the transfers would have allowed returning, injured veterans to more easily integrate themselves back into society; they could have used the extra money to lower their initial labor force par-ticipation, allowing them to focus on recuperating instead of worrying about paying for rent or other basic necessities.

Page 74: Visible Hand

74 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

There are some reasons why the coefficient on pension transfer may be lower than policymakers might have hoped. For one, while ability to work is indeed a measure of a veteran’s health, perhaps some veterans preferred to use some of the pension transfer for things not directly conclusive to reha-bilitation, like investing in land or supporting de-pendents. Moral hazard may still be present: with the knowledge that they are receiving regular trans-fers, some veterans may not believe they have an in-centive to improve their ability to work, and simply prefer to receive the same transfer amount instead of taking steps to become healthier. Critics of pension programs might also suggest that transfers were not a very efficient way to improve a veteran’s health because even with cash, many wounds do not heal. However, given our interpretation of the estimator, we believe that an estimated increase in ability to work of .04 per dollar in monthly pension validates using pension transfers to help veterans. We now turn to our other independent vari-able of interest, location in one of the twenty most populous cities of the time. Being in a top twenty city had a significant and highly positive effect (in magnitude) on a veteran’s change in ability to work. This could be for a number of reasons. First of all, it could be that veterans in urban areas have better ac-cess to services conclusive to improving their health and ability to work. Most services, like medical services for wounds, bodily illnesses, and psycho-logical afflictions were and continue to be better in urban centers than in rural areas, because typically cities attract talented professionals and have more resources at their disposal. Veterans seeking medi-cal treatment would surely encounter better services in a large hospital in New York City than in a small clinic in a rural area. Furthermore, the Federal gov-ernment has a much greater presence in cities than rural areas. Therefore it is possible that veterans had greater access to governmental programs for reha-bilitation if they were closer to a city. Lastly, regard-less of services a city can better provide, cities have much larger markets and are hubs of commerce, so any medical supplies or goods that a veteran could purchase on his own to improve his ability to work would be available in greater quantity and quality in the city. This is especially true during the Civil War Era, when the second industrial revolution was still gripping the nation and the infrastructure for mass commerce beyond big cities had yet to be es-tablished. Thus, it was harder for goods to travel to rural areas in as great number as in cities, and many of these services would have been much more ex-pensive to procure (as mentioned in our literature

review). It should be noted that while the city variable was slightly little less significant than the pension transfer variable, the magnitudes of their effects are difficult to compare given the difference in units. We believe that residing in a large city has a larger effect on rehabilitation than a one dollar per month increase in pension transfer does (if we assume a causal relationship). If we assume that we can ex-tend the linear relationship between pension dollars and rehabilitation for increases in pension amounts, we might interpret our coefficients to suggest that the effect of living in a large city is comparable to earning about $308 in monthly pension (since the estimated coefficient on living in a large city is 308 times the coefficient on pensions). Thus centrallized services, especially those prevalent in urban cen-ters, may be more effective than unconditional cash transfers in improving a veteran’s ability to work and overall rehabilitation. Though not an economic variable of interest, measures of the veterans’ wound severity and initial ability to work correlated with both our independent and dependent variables, and thus were crucial con-trols. Wound severity systematically correlates with pension amount and also influences how much the veteran can recover from his wounds. But we be-lieve that our severity variables are not sufficient controls by themselves to account for how bad a re-cruit’s injury actually was.This is because we have, as is shown in our column, a severe case of omitted variable bias with respect to our pension transfer variable when both severity controls are kept but the initial ability to work is not present. Whereas wound severity evaluates the effect a disability has on everyday life, initial ability to work evaluates the effect a disability has on work life specifically. Like wound severity, initial ability to work is in its own way correlated to recruits’ early pension transfers. Initial ability to work, as mentioned before, also significantly affects a recruit’s change in ability to work. One could imagine that the relationship be-tween the two variables resembles a logistic func-tion: when his initial ability to work is very low, a re-cruit is probably severely injured and has no chance of significantly improving his ability to work over time due to the permanent and severe nature of his injuries. If a recruit is initially somewhat disabled, it may indicate his initial injury is severe enough to hamper his ability to work but not severe enough to make him more unable to recover. A veteran in this condition can display a significant improvement in his ability to work. If a veteran’s initial ability to work is very high, his wound from the war was

Page 75: Visible Hand

Fall 2015 | The Visible Hand | 75

Volume xxiV. issue i.

probably not very bad to begin with, and he cannot improve his ability to work significantly. The high explanatory power of the initial ability to work vari-able also shows why it is so important in improving the fit of our overall regression. With the results of our main control thoroughly explained, it is now time to move to what we believe to be a more puzzling result. Both with and without the presence of the initial ability to work control, our mild severity control, which indicates whether a veteran had a mild ailment or not, had a significant and highly negative effect on the veteran’s change in ability to work. We initially thought that the op-posite would be the case due to the ”logistic func-tion” rationale discussed above. One would think that those with mild ailments stood to improve the most in their ability to work. And yet, the opposite seems to be the case. Having a mild wound actually was a detriment to improving one’s ability to work. The only explanation we can provide for this is that the wound severity data is generally a highly error prone measure of the actual severity of a veteran’s wound. This may be due in large part to the fact that some wounds that in actuality were highly detrimen-tal to a veteran’s ability to work, like concussions or post-traumatic stress syndrome, were not properly understood at the time and thus unaccounted for in determining the health of a veteran. Of course, this rationale undermines the explanatory power of the initial ability to work control, so we doubt this is true. More likely, the mild wound severity might be endogenous. We still have a reasonably high re-sidual error after adding in our main control in col-umn (1), and it is very possible that some significant omitted variables highly correlated to mild severity are not being accounted for.

VIII. Limitations

One limitation that we encounter is the fact that being located in a large city after the war may not be exogenous. People may deliberately move to the city after the war in order to receive more care for a debilitating injury. Of course, we tried to control for the severity of a person’s wound making them more likely to move to a city, but is difficult to tell if the city variable is truly exogenous or not. Looking at only recruits who returned to their home towns after the war might solve this issue, but we found that this was difficult to do. We tried two matching schemes to confirm whether a veteran’s pre- and post- war city of residence were the same, neither of which worked reliably. First, we matched the city a recruit was born in to the city from which he applied for

a veteran pension. We found that the many of the cities did not match, but also that surprisingly many mismatches were the same city referred to by differ-ent names. Thus, it was extremely difficult to devise a way to tell if the recruit returned to the city of his birth after the war. This was exacerbated by the fact that we had thousands of observations, so manually correcting for labeling inconsistencies was unfea-sible. In our second approach, we matched the re-cruit’s city right before they went to war to the city they were in when they applied for pensions, but ran into the exact same problem - and to begin with had fewer recruits with observations for both times. Without more resources at our disposal, we cannot know for certain if being in one of the twenty most populous cities in the U.S. was exogenous or not. Another issue we had was heteroskedasticity. We plotted our residuals against our fitted values (Plot 3 below), and found some bizarre linear pat-terns occurring, where certain residual values de-creased at a constant rate for certain intervals of fitted values. We then ran the following Breusch-Pagan Test to test for heteroskedasticity (for a gen-eralized OLS regression model used for notational simplicity):

y = β0 + β1x + U; u-hat2 = γ0 + γ1x + v H0 : γ2 = . . . = γp = 0 vs H1 : otherwise

Where u-hat2 are our squared residuals, p is the number of linear restrictions in our test, and γi are our parameters from regressing the squared residual on our independent variables. Thus we have a null hypothesis of ho- moskedasticity. Our resultant test statistic is 38.3644 with a p-value pf 1.499x 10-5. Therefore, we can reject our null; we have hetero-skedasticity. As you can see from our code, we ac-knowledge this issue by calculating adjusted HAC standard errors. These are present in our regression results tables. Another issue that we have already discussed at great length is that the sign of the estimate for our mild severity coefficient lacks economic intuition. We believe this explanation for this perplexing re-sult is sufficient, but not ideal. We invite further adjustments to our model that would relax the assumption of linearity on the rela-tionship between pension money and rehabilitation. Cash transfers in reality may have diminishing mar-ginal effects on rehabilitation, since the most cost-effective treatments to veterans’ disabilities are tak-en up first; however, since pensions were only part of veteran’s total wealth, large changes in pensions reflect smaller changes in veteran’s wealth, so we suspect that the effect of even these large changes can be approximated by a linear function.

Page 76: Visible Hand

76 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

In addition, it was difficult to control for socio-economic status since we found that the indepen-dent variable for wealth was extremely noisy and probably endogenous. For one, since wealth could include multiple unknown factors like hereditary wealth and personal property, wealth is probably correlated with unobserved factors in the error term. There might also be a simultaneity issue; wealth may influence one’s ability to work since wealthier individuals can seek better care, but at the same time one’s change in ability to work could influence his wealth via income. Income was also difficult to use, since we found a high probability of measurement error; veterans had an incentive to misreport their income on their pension application form with the hope that they would receive more money. Further-more, income reporting was heavily sporadic in the sample: most veterans’ records did not include any income reports, and those that did did could be year-ly, weekly, or monthly income. Thus, we suspected that income was not a well-observed or well-report-ed variable. While we used literacy to try to control for socio-economic status, this control was clearly not sufficient, as literacy was not significant in our results. As discussed in our Data Source section, re-applications to the pension program seemed to most fre- quently occur in response to additions or changes in the General Law Pension system. That being said, a clear issue that must be addressed is the fact that some veterans might have re-applied because their condition deteriorated, which means veterans with deteriorating conditions might be overrepresented in the sample. We believe that most reapplications responded to changes in the law, but we cannot perfectly confirm this belief within our sample: we can tell the dates that veterans reapplied, but there is no accompanying information anywhere in the data set as to why they reapplied. Therefore, we cannot know for sure if the correlation between pension increases and deteriorating health is signifi-cant or not. Some measurement error may be present in our change in ability to work variable because there is considerable heterogeneity in the time observed for each recruit’s reported early ability to work and late ability to work. While we tried to restrict the ob-served instances of ability to work to specific time intervals, heterogeneity is still inevitable. Our study has some limitations that we could avoid if we were able to conduct a similar study with modern data. For example, advances have been made in the diagnosis of various ailments that were likely to affect veterans’ abilities to work, but

would not have been detected by civil war era phy-sicians; specifically, mental illness is now consid-ered to be perhaps the largest obstacle to veterans’ reintegration into the workforce. A study that took this into account would be able to more comprehen-sively define the effect of treatments on veteran re-habilitation.

IX. Conclusion

Our study finds that in the aftermath of the American Civil War, both pensions and residence in a large city had significant, positive effects on veter-an rehabilitation. We believe that this finding carries implications for policymakers today: it provides his-torical validation for the government’s large invest-ment in cash transfers and centralized services for disabled veterans. Clearly, the government’s large investment in cash transfers to disabled veterans has historical validation, as do centralized govern-ment services. While much of the current literature is concerned with getting veterans back to work, as it should be, readers should recognize that giving veterans the ability to work is a crucial step in this process.

X. References

Blazer, D. G., L. R. Landerman, G. Fillenbaum, and R. Horner. "Health Services Access and Use among Older Adults in North Carolina: Urban vs Rural Residents." American Journal of Public Health, Oct. 1995. Web. 03 June 2015.

Card, David, and Alan B. Krueger. Minimum Wages and Employment a Case Study of the Fast Food In-dustry in New Jersey and Pennsylvania. The Ameri-can Economic Review, Vol. 84, No. 4. (Sep., 1994), 772-793. Print.

Card, David, and Orley Ashenfelter. "Handbook of Labor Economics." Elsevier, Nov. 2010. Web. 03 June 2015.

"Compensation." Benefits.va.gov. US Department of Veterans Affairs, 22 Oct. 2013. Web. 04 June 2015. <http://www.benefits.va.gov/COMPENSA-TION/types- disability.asp>.

Costa, Dora L. 1998. The Evolution of Retirement: An American Economic History, 1880– 1990. Chi-cago: University of Chicago Press

Costa, Dora L. "Appendix A: Union Army Pensions

Page 77: Visible Hand

Fall 2015 | The Visible Hand | 77

Volume xxiV. issue i.

and Civil War Records." National Bureau of Eco-nomic Research, Jan. 1998. Web. 02 June 2015.

Costa, Dora L. "Understanding the Twentieth-Century Chronic Conditions Among Older Men." Demography 37.1 (2000): 53-72. Springer. Web. 4 June 2015.

Fogel, R.W. (2000) Public Use Tape on the Aging of Veterans of the Union Army: Military, Pension, and Medical Records, 1860-1940, Version M-5. Center for Population Economics, University of Chicago Graduate School of Business, and Department of Economics, Brigham Young University. 323-343. Moffitt, Robert. Means-tested Transfer Programs in the United States. Chicago: U of Chicago, 2003. Na-tional Bureau of Economic Research. Web. 3 June 2015.

Mutikani, Lucia. "U.S. Military Veteran Unemploy-ment Easing, but Still High." Reuters.

Thomson Reuters, 18 Mar. 2015. Web. 04 June 2015. "Population of the 100 Largest Urban Places: 1870." Census.gov. US Bureau of the Census, 15 June 1998. Web. 04 June 2015. <http://www.cen-sus.gov/population/www/documentation/twps0027/tab10.txt>.

Sahr, Robert C. "Consumer Price Index (CPI) Con-version Factors for Years 1774 to Estimated 2024 to Convert to Estimated Dollars of 2015." (n.d.): n. pag. Oregon State University, 15 Feb. 2015. Web. 4 June 2015. <http://liberalarts.oregonstate.edu/sites/liberalarts.oregonstate.edu/files/polisci/faculty-re-search/sahr/inflation-conversion/pdf/cv2015.pdf>.

"The "72-Year Rule"" The US Census Bureau, 6 Jan. 2015. Web. 04 June 2015. <https://www.cen-sus.gov/history/www/genealogy/decennial_census_records/the_72_year_rule_1.html>.

"Union Army." Union Army Data, n.d. Web. 04 June 2015.<http://www.uadata.org/about/>.

"Veterans' Disability Compensation: Trends and Policy Options." Congressional Budget Office, 07 Aug. 2014. Web. 03 June 2015.

XI. Appendix

Table 1

Table 2

Table 3

Page 78: Visible Hand

78 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Plot 1

The residual from regressing change in ability to work on all covariates besides early ability to work is plotted against the residual from regressing early ability to work on the other covariatesPlot 2

The residual from regressing change in ability to work on all covariates besides pension amount is plotted against the residual from regressing pension amount on the other covariates

Plot 3

The residual from our base model regression is plot-ted against the fitted values of our regression. A pat-tern seems to exist, suggest heterogeneity; however, the plot does confirm linearity in our model.

Page 79: Visible Hand

Fall 2015 | The Visible Hand | 79

Volume xxiV. issue i.

larger corporations to outperform their rivals. As a result, the Gilded Age is characterized by the rise of large trusts and financial institutions. For example, the railroad required steel, giving rise to large rail-road monopolists like Andrew Carnegie. The other two major corporate powers in the United States were John D. Rockefeller and J.P. Morgan. John D. Rockefeller owned Standard Oil while J.P. Morgan was the leading financial power in the United States. Furthermore, the Progressive Era established the credibility of the social sciences such as eco-nomics, political science, and sociology, as well as the growing legitimacy of academic institutions such as Harvard University and Yale University. Faculty at these schools shifted ideological power away from the government and toward experts. One of the most notable developments of the Progressive Era was the role that economists played in establish-ing the legitimacy of a central bank. Arthur Twining Hadley, President of Yale University, confirmed the support of the academic community when he stated, “Economists largest opportunity in the immediate future lies not in the theories, but in practice, not with students but with statesmen, not in the educa-tion of individual citizens, however widespread and salutary, but in the leadership of an organized politi-cal body.” This statement epitomizes educational control over economic issues that were formerly considered within the realm of political control. While many historians note that the Federal Reserve Act was a result of the Panic of 1907, the latter half of the 19th century was marked by four significant financial panics—in 1857, 1873, 1893, and 1896. These panics directly signaled the need for a central bank among the public as Americans felt they could no longer trust the banks to hold their deposits. Experts began to realize that recessions could be cut short if there was a reserve system that could restore liquidity to the market. By 1896 politi-cians, businessmen, and economists began meeting

I. Introduction/Background

The origins of the Federal Reserve System por-tray the balance of powers that existed in the United States during the Gilded Age and the Progressive Era. This balance of power between academic, cor-porate, and political interests established the Federal Reserve System in December 1913. An analysis of how these interests were incorporated into the Fed-eral Reserve Act reveals the influence of both the Gilded Age and the Progressive Era. During the Gilded Age, powerful businessmen such as John D. Rockefeller, J.P. Morgan, and Andrew Carnegie gained vast amounts of wealth and assumed an in-credible amount of control over the U.S. economy. Furthermore, the rise of the social sciences bolstered the reputation of academia that brought credibility to the establishment of the central bank. Ultimately, the Federal Reserve originated not only from financial panics, but also from the com-petition and cooperation of political, corporate, and academic interests. In short, the origins of the Fed-eral Reserve reflect the balance of power between political, academic, and corporate interests that existed in the United States in 1913. Furthermore, the strict guidelines of the Federal Reserve Act per-manently guaranteed this balance would continue perpetually. In effect, this balance now operates as a system of checks and balances to ensure that the Federal Reserve is not fully controlled by one group or interest. To truly understand the origins of the central bank debate, it is essential to understand the histori-cal background of the late 19th century, symbolized by the Gilded Age. The Gilded Age was fueled by the Industrial Revolution as economic inequalities grew from 1870 to 1930. Economies of scale played a huge role in the establishment of this new Ameri-can bourgeoisie. Economies of scale states that pro-duction cost per unit decreases as the total number of units increases. This development allowed for

Michael BertonUniversity of California - Santa Barbara

The Origins of the Federal Reserve

Page 80: Visible Hand

80 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

to discuss the implementation of a central bank. The election of 1896 was an extremely impor-tant election because it laid the foundations for the gold standard and the Indianapolis Monetary Com-mission. The biggest issue of the election was the issuance of free silver. Due to the inequalities in wealth engendered by the Gilded Age, the Populist Party emerged rapidly in the United States. This group, led by Democrat William Jennings Bryan, primarily consisted of farmers who wanted free sil-ver to inflate the value of the dollar to make it easier to pay back debts. This was because farmers were constantly going into debt by purchasing land and capital. These farmers would also receive higher prices for their crops if the money supply increased through the issuance of silver. Not surprisingly, large creditors, such as the financial institutions of the Northeast, were opposed to free silver because it decreased the value of the money debtors had to pay back. While Republican presidential candidate Wil-liam McKinley represented financial interests and the gold standard, Democratic presidential candi-date William Jennings Bryan represented the Pop-ulists and free silver. In one of the highest voter turnout elections of the century, William McKinley defeated William Jennings Bryan by 95 electoral votes. By winning the election, President McKinley insured that the United States continued to rely on the gold standard. Due to the panic of 1896, Presi-dent McKinley and many experts realized that the gold standard ought to be reformed so that it could restore liquidity to the market during times of reces-sion. The main implication of this reform was that elasticity could only be achieved through a central bank.

II. The Indianapolis Monetary Commission

To this end, Henry Hugh Hanna, president of the Atlas Engine Works of Indianapolis, sent letters to experts, businessmen, and economists to orga-nize a banking reform movement in Indianapolis, Indiana. Henry Hugh Hannah strategically chose to form the Indianapolis Monetary Commission (IMC) in the Midwest to prove that the movement did not have ties to Wall Street. Overall the IMC established the roots of the Federal Reserve Act because it or-ganized businessmen, politicians, and economists to convince the public that a central bank was neces-sary. In 1897 the IMC urged President McKinley to “continue the gold standard and create a new system of elastic credit.”3 The leader of the IMC was James Laurence Laughlin, head professor of Political

Economy at the University of Chicago. The IMC’s first task was to distribute questionnaires to experts, which were then broadcasted to the public. The overall purpose of these questionnaires was to influ-ence citizens to support a central bank and also to show transparency to the public. 4 After collecting questionnaires, the IMC published its recommenda-tion for a central bank in many financial newspa-pers. To convince the public of the necessity of a central bank, Henry Hugh Hanna hired financial journalist Charles Arthur Conant to develop pro-paganda surrounding the report conducted by the IMC in 1898. In these reports, Charles Conant em-phasized that one of the key consequences of the Industrial Revolution was a capital surplus. To this end, he chronicled international banking history in a History of Modern Banks of Issue in 1896. Further-more, Charles Conant was also a heavy advocate of the gold standard and believed that silver should be exchanged for a fixed amount of the gold standard. Globally, Charles Conant envisioned rich countries owning gold, developing countries using silver, and advanced countries using paper. The implication of the advanced countries using paper was that this required a highly organized central bank that could control the proper inflation of the paper currency. Strategically, Charles Conant also partnered with Jules Guthridge, Secretary of the IMC, to ad-vertise the questionnaires into over 1,000 newspa-pers. The purpose of these advertisements was to prove to the public that experts supported the cen-tral bank. Ironically, the actual reforms that the IMC proposed, which included the exclusive use of the gold standard and more elastic credit, were not as important as the question of banking reform itself. As a result of Charles Conant’s journalistic efforts, the Indianapolis Monetary Commission met again on January 25, 1898. The goal of this second meet-ing was to convince businessmen and bankers to agree with the IMC’s recommendations for a central bank. For example, Secretary of Treasury Charles Fairchild stated, “If men of business give serious at-tention and study to these topics, they will substan-tially agree upon legislation and their influence will be prevailing.” The report eventually made its way to the House Banking and Currency Committee in May 1898, but it was rejected in the Senate. The second report of the IMC directly called for new monetary policy, a system of elastic credit (es-tablished through reserves), and a central bank. The report stated: “To insure that they (dollars) shall at all times be equivalent to the gold dollar, they must be at all times directly or indirectly convertible with

Page 81: Visible Hand

Fall 2015 | The Visible Hand | 81

Volume xxiV. issue i.

that dollar. This plainly necessitates that there shall be kept somewhere in the country a stock of standard dollars, in other words, a standard money reserve.” In the closing paragraph of the report, F.M. Taylor discussed the legacy of the IMC when he wrote, “At least the business classes will have gained in aptitude to undertake the influencing of legislation. The public will have gained in the knowledge of monetary and banking matters necessary to prepare them to pass judgment upon any future project of reform.” Furthermore, the IMC also introduced bills into Congress based upon the IMC’s report. These bills were not approved until 1907 because politi-cians felt that they lacked public approval. From 1898 to 1907, public support for the cen-tral bank continued to grow in strength as Charles Conant continued advocating for the Federal Re-serve. During the early 20th century, intellectuals, politicians, and the corporate elite were continuous-ly meeting to discuss the intricacies of establishing a central bank. The overall sentiment of politicians was that they could not introduce a bill for a cen-tral bank without public approval. While Charles Conant was influential in exposing the public to the central bank, politicians did not receive the amount of support they needed until another financial panic erupted in 1907, revealing a historical theme: peo-ple mobilize in the face of economic depression and poverty. As a result, many people argue that public support for a central bank really accelerated after the Panic of 1907.

III. J.P. Morgan’s Response to the Panic of 1907 The Panic of 1907, like all financial panics, was a result of overspeculation. The stock market fell by 50% and many smaller banks began to go bankrupt as people lost confidence in the banking system. In October 1907, the Knickerbocker Trust of New York collapsed, signaling the Panic of 1907. The New York Stock Exchange was collapsing, and without a central bank, the financial system had no method of restoring liquidity to the market with ex-cess reserves. As a result, J.P. Morgan led a group of bankers that gave 23 million dollars to the NYSE in October 1908, which today would be worth 600 million. Additionally, with the approval of President Theodore Roosevelt, J.P. Morgan also purchased the highly leveraged Tennessee Railroad and Authority. Through these maneuvers, J.P. Morgan guided the United States out of the recession. This was the first time in American history in which a single mortal figure had the ability to pull the United States out of a recession.

While many Americans heralded J.P. Morgan as a financial hero, critics believed that only the govern-ment should have this tremendous amount of power. These critics did not want to have to rely on a lead-ing financial interest like J.P. Morgan in the future. They wanted to disperse the responsibility among the government, banks, and academic institutions. Furthermore, in the aftermath of the Gilded Age, Americans had a deep distrust for corporations and trusts. After the Panic of 1907, Americans were convinced that a reserve system was necessary, and that it should not be in the hands of major financial leaders. As a result, Americans began to realize that they needed a central bank to restore liquidity to the markets. In 1912 the “Money Trust Hunters” subpoenaed Ransom H. Thomas, the ex- president of the New York Stock Exchange to testify about J.P. Morgan’s involvement in the bailout of the NYSE, which oc-curred on October 24, 1908. Ransom H. Thomas testified,

“I went to see James Stillman, the president of City Bank. He recommended to me to go and see Mr. Morgan and tell him the exact story I told him. Mr. Stillman wanted to know how much would be need-ed to relieve this stringency. I told him 23 million. Mr. Stillman said he would advise Mr. Morgan that I was coming. I went over to Mr. Morgan’s office but found him in a conference. I was obliged to wait 20 minutes before I could see Mr. Morgan. Mr. Morgan finally came out of his private office and said we are

going to let you have the 23 million. Go over tothe Exchange and announce it.”

This testimony provides a substantial amount of evidence that a group of bankers bailed out the NYSE. Even Ransom Thomas verified that Mr. Morgan stated, “We are going to let you have the 23 million.” Furthermore, Mr. Morgan was in a private conference for twenty minutes, most likely consulting a team about this decision. Ultimately, this 23 million dollar stimulus package symbolized the turning point of the recession because it restored consumer confidence in the market. The fact that J.P. Morgan could almost single-handedly pull the United States out of a financial panic alarmed many Americans. Although J.P. Morgan cut the panic short, Americans did not want to have to rely on one man to pull the economy out of the next recession. As a result, the public exerted a significant amount of influence on politicians to institutionalize this power and create a central bank.

Page 82: Visible Hand

82 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

IV. The National Monetary Commission

Politically, the government responded to the Panic of 1907 by creating the National Monetary Commission (NMC) with the Aldrich-Vreeland Act in 1908. The chairman of the NMC was Senator Nelson Aldrich, a Republican whip in the Senate representing Rhode Island. Nelson Aldrich wrote a 72-page report to the Senate in order to discuss the problems associated with the current banking system and to propose a central bank. Senator Al-drich supported the relationship between the estab-lishment of the Commission and the Panic of 1907 when he wrote, “the act of May 30th, 1908, provid-ing for the appointment of the National Monetary Commission was a direct consequence of the Panic of 1907.” Ultimately, the National Monetary Com-mission symbolized the government’s response to the Panic of 1907. Nelson Aldrich led the National Monetary Commission and hired Charles Eliot, President of Harvard University, A. Piatt Andrew, a Harvard economist, and Charles Conant. On September 22, 1909, Charles Conant wrote a 14-part series of ar-ticles on the central bank issue in the Wall Street Journal. The purpose of the NMC was to propose monetary legislation for a central bank. To justify the central bank, the NMC discussed the flaws with the current system. The NMC emphasized the fact that “we have no power to enforce the adoption of uniform standards with regard to capital, reserves, examinations, and the character and publicity of reports of all banks in the different sections of the country.” Besides justifying a central bank to the public, the NMC was also responsible for drafting bills that would be introduced into Congress. By 1910, Nelson Aldrich was desperate to draft a bill that had the support of bankers, politicians, and uni-versity experts.

V. The Secret Meeting at Jeckyll Island

One of the most speculated topics about the Federal Reserve Act is the secret meeting that took place at Jeckyll Island. Believers say that the purpose of this meeting was to organize financial leaders to write the Aldrich Plan. Frank Vanderlip, President of the National City Bank, confirmed that he attended a secret meeting at Jeckyll Island in his biography From Farm Boy to Financier in 1935 two years before his death. Furthermore, Frank Vander-lip also advised Congressmen Robert Owen and Carter Glass while they were drafting the Federal Reserve Act. Since Frank Vanderlip was an integral

architect of the Federal Reserve, it is important to consider his testimony about the meeting at John D. Rockefeller’s vacation home at Jeckyll Island off the coast of Georgia. In October 1910, Senator Nelson Aldrich, father-in-law to John D. Rockefell-er Jr, used his private railroad car in New Jersey to transport seven of the world’s financial leaders to John D. Rockefeller’s vacation home at Jeckyll Is-land. While disguised as a duck-hunting trip, Frank Vanderlip says that the purpose of this meeting was to write the Aldrich Plan. According to Frank Vanderlip, the parties rep-resented were the Morgans, Rockefellers, Roths-childs, and the Warburgs. The Rothschilds and the Warburgs were the two largest financial powers in Europe at the time. Frank Vanderlip stated, “Those had been asked to go were Henry Davison, Paul Warburg, Ben Strong, and myself. From Washing-ton came A. Piatt Andrew, who was then an Assis-tant Secretary of the Treasury.” Henry Davidson was a senior partner at J.P. Morgan, Paul War-burg represented the Warburgs and Rothschilds, and Ben Strong would later become the 1st President of the Federal Reserve Bank of New York. Frank Vanderlip described the motivations for this meeting when he wrote,

“At the time for the assembling of Congress drew near, Senator Aldrich became increasingly con-cerned about the report he must write on behalf of the joint monetary commission; likewise, there ought to be, he knew, a bill to present to the new Congress and none had been drafted. This was how it happened that a group of us went with him to Je-

kyll Island Club on the coast of Georgia.”

Frank Vanderlip stated that the meeting was se-cretive because the Aldrich Plan would have never been approved if the public knew about the meeting. In reference to this secrecy, Frank Vanderlip wrote, “Discovery, we knew, simply must not happen, or else all our time and effort would be wasted. If it were to be exposed publicly that our particular group had gotten together and written a banking bill, that bill would have no chance whatsoever of passage by Congress.” After the meeting at Jeckyll Island, Charles Conant also played an essential role by in-fluencing the public to support the Aldrich Plan. In the Banker’s Magazine, Charles Conant endorsed the Aldrich Plan when he wrote, “In my opinion, the plan presented by Senator Aldrich will, with some modifications of detail, cure most of the defects in our currency and banking mechanism which were disclosed by the Panic of 1907.” Upon returning

Page 83: Visible Hand

Fall 2015 | The Visible Hand | 83

Volume xxiV. issue i.

to Washington D.C., Nelson Aldrich and the NMC were now in a position to draft the Aldrich Plan.

VI. The Aldrich Plan

Under the Aldrich Plan, the National Reserve Association (NRA) would have capital equal to twenty percent of the reserves of its subscribing banks. The NRA would primarily be responsible for administering bank reserves, issuing elastic curren-cy, and depositing federal money. The NMC even called for the establishment of local associations, or districts, that were “authorized in serious emergen-cies to guarantee the direct obligation of subscrib-ing banks with adequate security.” This new system also had a complex, highly organized leadership led by a Board of Directors made up of members from the local associations. In many ways, the National Monetary Commission’s proposal for the National Reserve Association laid the seeds for the Fed-eral Reserve System because it showed members of Congress that dramatic banking reforms were needed to avoid financial panics. Additionally, Nel-son Aldrich’s proposed system was favorable to the bankers because it called for a 46-member board with only 6 appointed by the government. Senator Theodore Burton introduced the Al-drich Plan into Congress in January 1912. The members of the House and Senate were divided over three core dilemmas about the implementation of a central bank. The first question was whether man-agement of the new system would be controlled by both businessmen and government, or just by cabi-net members and political appointees. The second debate was whether the new banking reform would function as a true central bank or as a cooperation of regional banks with a central bank. The third question was who the currency would be obligated to—the government or the banks. The Democratic House and Senate wanted cabinet supervision to control the regional central banks as well as gov-ernment obligation of issued currency. By contrast, the banking interests wanted to manage the regional branches with issued currency obligated to banks. Ultimately, the Aldrich Plan failed to gain traction because both political and business interests were not satisfied. One reason why the Aldrich Plan was not effective was because of the American prefer-ence for liberalism and a strong separation between government and business. Overall, the major im-pact of the Aldrich Plan was that it set in motion the debate for major banking reform and gave rise to future discussions about the Federal Reserve Act. In his biography, Aldrich Plan architect Frank A.

Vanderlip wrote,

“Now although the Aldrich Federal Reserve plan was defeated when it bore the name of Aldrich, nev-ertheless its essential points were all contained in the plan that final was adopted. The law as enacted provided for twelve banks instead of the one, which the Aldrich plan would have created; but the intent of the law was to coordinate the twelve through the Federal Reserve Board in Washington, so that in ef-fect they would operate as a Central Bank. There can be no question about it: Aldrich undoubtedly laid the essential, fundamental lines which finally

took the form of the Federal Reserve Law.”

VII. The Federal Reserve Act

To address the failures associated with the Aldrich Plan, two Aldrich Plan architects Frank Vanderlip of Citibank and Paul Warburg of Kuhn, Loeb & Co advised Congressmen Robert Owen and Carter Glass on a new bill. These congressmen were both chairmen of the Senate and House com-mittees on Banking and Currency. With the guid-ance of Frank Vanderlip and Paul Warburg, Robert Owen and Carter Glass drafted the Federal Reserve Act. To gather public support for the new bill, Frank Vanderlip formed a commission, and elected Jacob Schiff, J.P. Morgan, George Baker of the First Na-tional Bank of New York, and Lyman Gage, Presi-dent of Rockefeller-controlled U.S. Trust Company. Two members of this commission represented Mor-gan—J.P. Morgan and George Baker – while Lyman Gage represented Rockefeller. The commission was tasked with gathering support from experts in favor of a central bank. To this end, the commission chose to use the Indianapolis Questionnaire Technique to convey expert approval. In July 1912 Robert Owen and Carter Glass coauthored the Federal Reserve Act, which would simultaneously be introduced into both the House and the Senate. They wanted to reshape the Aldrich Plan into a new bill that would satisfy both government and businesses interests. In July 1912, they decided to add to the bill the estab-lishment of an advisory council of men elected by regional reserve banks and large financial leaders. The Federal Reserve Act was controversial be-cause it dictated that all of its notes were obligated to the federal government, not the banks. In the Al-drich Plan, the notes were obligated to the banks. Frank Vanderlip discussed the dangers of letting notes be obligated to the government in the Wall Street Journal. Frank Vanderlip stated, “The really grave danger lies in leading the public to accept the

Page 84: Visible Hand

84 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

fallacy that the Government can print money for which it provides within itself no metallic means for redemption, and have that paper successfully perform all the functions of a proper circulating note.” Frank Vanderlip accurately predicted that the government would abandon the gold standard and issue fiat money. Congressmen Owen and Glass ar-gued that it was the U.S. Treasury’s responsibility to ensure that the government backed the purchasing power of notes. To this end, Robert Owen stated,

“We have been charged with making a great and se-rious error in having those notes obligations of the Government. Yet, I remind you that thirteen years ago, the Secretary of Treasurey was required by law to maintain the parity between all forms of money

issued by the U.S.”

Robert Owen was referring to the Gold Stan-dard Act of 1900, which made the government responsible for ensuring the value of the money. Although Robert Owen believed that the bankers wanted the notes to be obligated to the government, the bankers were still dissatisfied with the Federal Reserve Act because they were not represented on the Board of Governors. The fact that the President elected these governors meant that the government could exercise full control over the leadership of the Federal Reserve. In the end, Congress approved the bill because it gave the government control over the notes and the Board of Governors. Furthermore, the Democrats controlled both the House and the Sen-ate by December 1913, allowing for the smooth pas-sage of the Federal Reserve Act on December 23, 1913.

VII. Conclusion

Still today, the Federal Reserve reflects the competing political, corporate, and academic forces of 1913. The Gilded Age and the Progressive Age gave rise to both corporate and academic power, re-spectively. The leading academic institutions sided with corporate and political interests to revolution-ize the monetary system in the United States. It is crucial to understand the forces that influenced the Federal Reserve because they still play a huge role in the United States economy today. By being able to set interest rates, reserve requirements, and dis-count rates, the Federal Reserve exhibits an incred-ible amount of control over lending policies, ex-change rates, and the growth of the money supply. Corporate, political, and academic interests have always controlled the Federal Reserve—

Paul Volcker was a financial economist for Chase Manhattan Bank and a financial analyst at the US Treasury, Alan Greenspan was President of the in-vestment bank Townshend-Greenspan Inc., Ben Bernanke was an economist at Princeton, and Janet Yellen was a Professor of Economics at the Univer-sity of California, Berkeley. The incorporation of academic institutions and corporations into the Fed-eral Reserve shows just how much influence these interests have in our economy today as many would argue that the Federal Reserve influences lending and inflation in a more direct way than the federal government. The balance of powers in 1913 that created the Federal Reserve is still extremely similar today for many reasons. The corporations still profit from the Federal Reserve. In September 2008, the Fed-eral Reserve gave the American International Group (AIG) 85 million dollars because the federal govern-ment ruled that the insurance company was essential for the economy. AIG would have collapsed without the support of the Federal Reserve because it held insurance on risky mortgage backed securities. The government knew that AIG was too big to fail, and would cause a rippling effect in the economy if it failed. As a result, the Federal Reserve stepped in and injected money into the system. The role the Federal Reserve played in 2008 is extremely similar to the role J.P. Morgan played in 1907. Due to the strict guidelines set forth by the Fed-eral Reserve Act, the Federal Reserve System still deeply reflects the balance of powers that existed in 1913. This is due to the writing of the Federal Reserve Act in 1913. Since the Federal Reserve is the lender of last resorts, banks and insurance com-panies continue to be bailed out by the central bank during financial panics. The first groups to profit from reserves are banks and insurance companies because they are too big to fail. This makes it so that corporations and banks can rely upon the Federal Reserve during financial panics. Instead of the bur-den falling upon J.P. Morgan, the burden now falls upon the taxpayer. Furthermore, the Board of Gov-ernors typically consists of economists because the President wants to prove that he has elected quali-fied experts to lead the nation’s monetary policy. Political power is still prevalent because the Presi-dent gets to decide the Board of Governors while the Senate is responsible for approval. Economists play a huge role in the Federal Reserve because they determine the actions and policies of the Federal Reserve as Board of Governors. Furthermore, the Federal Reserve Act secured a permanent system of checks and balances between

Page 85: Visible Hand

Fall 2015 | The Visible Hand | 85

Volume xxiV. issue i.

corporate, political, and academic groups in 1913. This balance of power serves as a system of checks and balances because it ensures that the central bank is not fully controlled by one group or interest. As a result, corporate, political, and academic influences compete with one another to set monetary policy. These checks and balances are rigid and do not fluc-tuate due to the strict guidelines set forth by the Fed-eral Reserve Act regarding the Federal Reserve’s or-ganization. For this reason, the Federal Reserve Act guaranteed that corporate, political, and academic groups would continuously be able to exert their in-fluence on the United States economy.

IX. References

Conant, Charles. “A History of Modern Bank of Is-sues.” Bankers Magazine Nov. 1911 “Currency Bill Debate by Glass, Vanderlip, and Owen.” Wall Street Journal 12 Nov. 1913. Retrieved 17 November 2015

Livingston, James. Origins of the Federal Reserve. 1st ed. Ithaca: Cornell UP, 1989. “Money Trust Hunters Hear How J.P. Morgan Checked the Panic of 1907: R.H. Thomas

Tells the Pujo Committee about Pouring $23,000,000 into the New York Stock Exchange.” Wall Street Journal 13 June 1912. Retrieved 17 November 2015

Rothbard, Murray. The History of Money and Bank-ing in the United States. Auburn: Ludwig Von Mises Institute, 2010. Print.

Rothbard, Murray. The Origins of the Federal Re-serve. Auburn, Ala.: Ludwig Von Mises Institute, 2009. Print.

Taylor, F.M. “The Final Report of the Indianapolis Monetary Commission.” Journal of Political Econo-my. Vol. 6, No. 3 (June 1898), p. 293-322

Vanderlip, Frank A. From Farm Boy to Financier. 1st ed. Appleton-Century, 1935

Page 86: Visible Hand

86 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

cial system in the years leading up to the crisis was characterized by “too much” risk, and that the tur-moil that occurred during the crisis and the resulting economic downturn was unacceptable and avoid-able.1 This sentiment makes the theoretical study of financial market failures and potential regulatory solutions extremely timely and important; certain features of the 2007-8 financial crisis particularly motivate the study of overall liquidity provision. Leading up to the crisis, banks, “increasingly fi-nanced their asset holdings with shorter maturity in-struments”, which left them, “particularly exposed to a dry-up in funding liquidity.”2 This trend was thought to have directly contributed to the “liquidity squeeze” during the crisis in which financial insti-tutions were unable to meet their short term fund-ing needs, a phenomenon known as the ‘liquidity phase’ of the crisis.3 Again, however, just because the tightness occurred and losses ensued doesn’t necessarily imply an inefficiency. The general consensus that a failure had indeed occurred led policymakers to institute new regula-tions on liquidity in Basel III.4 One of the major components of Basel III was the Liquidity Cover-age Ratio, which required banks to have a certain stock of “unencumbered high quality liquid assets” that could be converted into cash easily and imme-diately.5 In the regulation, the authors argue that, “During the early “liquidity phase” of the financial crisis that began in 2007, many banks – despite adequate capital levels – still experienced difficul-ties because they did not manage their liquidity in a prudent manner. The crisis drove home the impor-tance of liquidity to the proper functioning of finan-cial markets and the banking sector.” The liquidity squeeze and the losses experienced, argue the Ba-sel Committee, was a result of disregarding basic principles of sound liquidity risk management. In a properly functioning, competitive banking sector, one would expect market discipline to prevent this sort of mismanagement. This claim then motivates the question, which market failures or constraints on incentives caused this disregard. One suspected source of market failure is known as the fire-sale externality, in which financial

I. Introduction

This paper will extend a line of research rel-evant to the events of the recent financial crisis. The line of research surrounds theories of under-provi-sion of liquidity in financial markets as a whole, and subsequent justifications for liquidity regulation. It begins with the “workhorse” Diamond-Dybvig banking model and shows how, if agents are able to engage in an arbitrage opportunity known as re-trade, a free-riding problem emerges and the market under-provides liquidity. In the model, the demand for liquid assets is motivated by the presence of the chance of facing an economic shock and needing access to liquid assets immediately, as well as by the risk aversion of consumers. The optimal level of liquidity in the market in previous models depended on the level of risk aversion of a single “represen-tative depositor; this paper relaxes the assumption that all depositors in the market have the same risk aversion preference, and risk preference is allowed to, perhaps more realistically, vary along a distribu-tion. The major question I intend to explore is the way in which the optimal level of liquidity regu-lation (if any at all) that the policy maker should choose depends on the nature of the distribution of risk preferences. To do this, I will assume a familiar undergraduate-level functional form for risk aver-sion, use this to reformulate many of the established conclusions, then construct a function that shows that while heterogeneous risk-preferences prevent the possibility of a strictly Pareto-improving liquid-ity regulation, an optimal liquidity regulation can be chosen based off of the distribution of risk prefer-ences in an economy. It is clear that the presence of risk and the observation of losses in a financial system do not themselves necessitate intervention into a market. It is only when market failures can be clearly identi-fied that regulators should intervene, and even then, if not properly calibrated, regulation may have the potential to do more harm than good. In light of the recent crisis, however, there is a general sense among policymakers and the public that the finan-

Wendy MorrisonUniversity of Virginia

Optimal Liquidity Regulation Given Heterogeneous Risk Preferences and Retrade

Page 87: Visible Hand

Fall 2015 | The Visible Hand | 87

Volume xxiV. issue i.

institutions fail to fully internalize the costs of their own under-liquidity when a lack of short-term fund-ing forces them to quickly sell off assets to cover their obligations.6 During the crisis, market partici-pants saw “asset prices drop, financial institutions’ capital erode and at the same time, lending stan-dards and margins tighten. Both effects cause fire-sales, pushing down prices and tightening funding even further.”7 In this way, the funding decisions of major financial institutions seem to be intertwined with the fate of other institutions and the economy as a whole, suggesting the presence of a classic ex-ternality in which social marginal costs associated with the risk of a liquidity crunch far exceed an in-stitution’s private marginal risk of failure. Largeness and interconnectedness alone, how-ever, do not themselves imply a failure and neces-sitate regulation. In order to fully justify and design effect regulation, one must also clearly demonstrate how market participants fail to internalize the costs of their financial decisions. The line of research I intend to further has identified another type of pe-cuniary externality in relation to liquidity provision known as ‘retrade’ that may also explain why mar-kets under-provide liquidity. It‘s proponents argue that financial actors are able to free-ride off of mar-ket liquidity by holding insufficient levels of liquid assets themselves, and counting on the ability to be able to purchase liquidity from someone else in the market in the event of a shock. Like the classic ‘pub-lic goods’ market failure, if all actors in the econo-my attempt this free riding strategy, liquidity will be under-provided in the economy at large. Liquidity regulation would force institutions to hold the ap-propriate level of liquid assets much in the same way a state might tax individuals to provide for a public good collectively. While this would prevent individual institutions from benefitting from a free riding strategy, such regulation would ensure the provision of the socially optimal level of the public good, in the case, market liquidity. The model expanded upon in this paper is of particular interest because it studies the optimal amount of resources set aside to deal with economy-wide liquidity shocks, and in that way has a ‘macro-prudential’ flavor. In other words, instead of exam-ining incentive constraints on the ‘micro-prudential’ or individual firm level , this line of research ex-plores the way in which incentives constraints affect an overall market’s ability to provide optimal liquid-ity.

II. Literature Review

The Diamond Dybvig (DD) Model: In the mod-el, financial institutions provide two services for their depositors: investment and insurance. In other words, financial institutions provide both attractive return and guaranteed availability of one’s funds on demand.8 In the DD framework, there are two time periods over which individuals can invest and con-sume, and two goods available to invest in, one liq-uid and one illiquid. The liquid asset, “cash,” can be spent at any time, but offers no net return, whereas the illiquid 5 investment, “capital,” offers a high re-turn but requires waiting until the second period to collect it. These assumptions mimic the properties of real world assets, and banks help depositors at-tenuate this tradeoff between liquidity and return.To do this, financial institutions simultaneously ac-cept deposits that are more liquid than the assets they hold. This can be viewed as an insurance ar-rangement for depositors. The demand for liquid assets (or first period consumption) is modeled by individuals facing a risk of economic shock, where some may be “unlucky” and unable to wait until the last period for the high return. Banks are able to take deposits and invest them in the appropriate ratio of low return cash and high return capital based on the general risk of being unlucky and the risk prefer-ences of the depositors.While this may mean that the return banks offer lucky depositors in period two is smaller than if they had invested their entire endowment in capi-tal, banks become able to offer unlucky depositors a slightly higher return. This liquidity provision is a highly valuable service because individuals are as-sumed to be risk averse, and they would be willing to accept this smaller return in the event that they are lucky in order to ensure a higher return if they are unlucky. In this way, financial institutions perform an insurance function for depositors and provide the ‘optimal’ level of liquidity in the market.Algebraically, assume the probability of facing a shock in the first period is α and the probability of being able to wait until the second period is (1-α). Then the optimal amount of liquid investment for a risk neutral individual with an endowment of y in-vesting on their own is:αy in cash and (1-α)y in capital. It was assumed by the original authors of the model that any second period consumption does not factor into the utility functions of any consumer facing a shock.9 The in-tuition behind such an assumption is that the shock imposes an immediate financial need so great that, if not met, may result in disastrous, irreparable, or per-manent consequences for the consumer. This makes a dollar spent in the first period so much more valu-

Page 88: Visible Hand

88 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

able to a consumer than a dollar spent in the second period that utility from second period consumption can be assumed away without consequence. This as-sumption becomes less reasonable however, as the consequences of the shock become less serious. The net return on cash is zero while the net two period return on capital is ry where 0 < r < 1. A consumer ex ante can expect to have a α chance of consum-ing αy and a (1-α) chance of consuming αy + (1-α)yr. The law of large numbers ensures that from the financial intermediary’s point of view, α is also the proportion of depositors in the overall economy who end up facing a shock. The expected value for a consumer ex ante is:

EV = α2 y+ (1-α)(αy + (1-α)yr) The primary insurance function of the bank then comes from their ability to allow individuals to pool risk as depositors and take advantage of the law of large numbers to offer depositors higher returns than they could otherwise achieve. In other words, banks, who know the long run probabilities of de-positors facing shocks, can invest α of their total deposits in cash and (1-α) in capital, thus ensuring a higher expected value for their depositors. The new expected value is:

EV = αy + (1-α)yr In order to make my work within this model manageable, I depart from the established proofs in form only, and reformulate existing conclusions us-ing a functional7 form for risk aversion I am com-fortable working with. This functional form has the attractive property of being simple and convex, meaning that consumers in my economy experience risk aversion. I assume all consumers have the util-ity function:

U(x1,x2 )= (x1d)/d+(x2

d)/dwhere x1 and x2 are the expected value of gross re-turn (consumption) in period 1 and 2, and d, set be-tween [-30,1], is 1 minus risk aversion coefficient. Risk aversion was set at or below 1 to ensure that all depositors were modeled as either neutral or averse to risk, and was expanded as low as -30 to provide additional variation in order to accommodate Mat-lab’s level of precision. As d approaches 1, the de-positor approaches risk neutrality. The bank collects a total of Y in deposits from all of its depositors, and can invest in cash and capital at a ratio it chooses. The bank is, however, still under the same li-quidity constraints as the individual, so its invest-ments will directly determine the rates of return it can offer its depositors in period 1 and period 2. In other words, if a bank acting in its function as an insurer deviates from maximizing expected value by investing less than (1-α)Y in capital, it can offer it’s

unlucky depositors a gross rate of return in the first period greater than 1. More precisely, assume an in-stitution reduces their capital investments by π per-cent. By reducing the second period rate of return by investing only (1- π)(1-α)(Y) in capital and α(Y)+ π (1-α)(Y) in cash, financial institutions can increase the first period rate of return by πα/(1-α). Thus the expected utility of a depositor as a function of π and d is:

EU= α(y(1+πβ/α))d+β((1-π)ry)d

By taking the derivative of this function and setting it equal to zero, the bank can solve for the optimal amount of liquidity, which will straightfor-wardly depend on risk preference, d. The fact that the optimal amount of liquidity depends on risk preference was first noted formally by Allen and Gale (2004).10 If a competitive profit-maximizing banking industry is assumed, banks will invest in the ‘optimal’ amount of liquidity on behalf of their depositors and the outcome will be efficient. Opti-mal liquidity within the context of my formulation is the following, a decreasing function of d (Appen-dix B).

If banks in a competitive banking sector know the risk preference of their consumers, d, and there-fore invest according to this relationship, they can offer gross rates of return r_1,r_2 where 1 ≤ r_1 ≤ r_2.

r1=(1+πβ/α) r2=(1-π)rand the optimal set of period one and period two consumption faced by a depositor with risk aversion coefficient, d, is x1,x2 where y≤x1 and x2≤yr

x1=(1+(π*β)/α)y x2=(1-π*)ryIn this way, the bank offers insurance to their de-positors by allowing the consumption set they face to be less variable. By pooling risk, banks are able to offer higher expected returns than individuals act-ing alone could achieve, and by investing accord-ing to the risk preferences of their depositors, banks maximize their expected utility by offering period 1 and period 2 consumption bundles x1, x2 such that y≤x1≤x2≤yr.

III. The Market Failure

The idea that a market failure exists within this context that could lead to under provision of liquidi-ty or insufficient risk sharing was first introduced by Charles Jacklin (1987) and expanded upon in depth

Page 89: Visible Hand

Fall 2015 | The Visible Hand | 89

Volume xxiV. issue i.

in Farhi et al (2007). Jacklin noted how, “if ex post trading is possible and new assets can be introduced, on the margin individuals have no incentive...to de-posit their funds in a bank”, due to a type of arbi-trage opportunity that exists ex ante.11 Consider a ‘rogue’ actor, be they institution or depositor, who has the same risk preference, d, as the representative consumer. Consider what would happen if they, or the institution acting on their behalf, invested their entire endowment in capital. A properly function-ing market would presumably deter them from this strategy unless of course the representative consum-ers are risk neutral, as it would result in the set of ex-pected gross rates of return (1,r), rather than (r1,r2). In other words, though the second period rate of return and the total expected value of investment is higher under this strategy, the depositor would be incurring excessive risk given their preference. If, however, in the face of a shock, this rogue actor could find a ‘lucky’ depositor with whom to trade ex post, this actor could on the margin achieve the ul-tra-high second period return of r without incurring any of the associated risk as long as prices (returns) stayed the same.12 Would we expect to see these trades occur? In the face of a shock in the first period, the rogue ac-tor could trade the withdrawal rights to their period two returns from capital, ry, to their partner for her first period liquidity. In other words, the value of the period two investment to a consumer facing a shock is her ability to trade the claim to her investment for liquidity for liquid assets in period 1. In this way, the rogue actor could count on the expected first and second period returns (r1,r) and thus can expect a period 1 and period 2 consumption bundles (x1,yr). Because y<x1 and x2<yr, and utility is an increasing function of consumption, an individual pursuing the arbitrage strategy is unambiguously better off.If all market participants foresee this ability to “free ride” off of market liquidity, an unravelling of the system occurs, as all participants will want to take advantage of this strategy and none will remain in the more liquid position. This incentive failure causes liquidity to be under-provided in the market if consumers are not risk neutral. This unraveling problem, notes Jacklin, is, “a result of the prefer-ence assumptions of Diamond and Dybvig. Since (unlucky) individuals have no utility for period 2 consumption, marginal rates of substitution are not equalized across types at the social optimum.”13 Formally, the problem as explained both by Jacklin and by Farhi et al is that the marginal rates of substi-tution of depositors in this economy between period 1 and period 2 consumption does not equal the ratio

between their prices (rates of return) and thus there is over-investment in capital.

IV. Proposed Solutions to the Market Failure

Taking their lead from Jacklin (1987) and Allen and Gale (2004), Farhi et al (2007) solve explicitly for optimal liquid investments within the context of the model, and show how the aforementioned incentive constraint causes the MRS between first and second period consumption to be misaligned. They then show how a government regulation in the form of a simple properly calibrated liquidity requirement “solves” this free rider problem and delivers the optimal amount of liquid investment and risk shar-ing. Their regulation, “stipulates a minimal portfolio share to be held in the short-term asset by interme-diaries. The liquidity floor increases the amount of the first period aggregate resources and drives the (second period) interest rate on the private markets down.”14 By regulating the amount of illiquid investment any institution can make, the planner implicitly regulates the second period interest rate offered, as without the ability to invest in more than the optimal amount of capital, institutions are unable to offer second period rates higher than r2. This eliminates the arbitrage opportunity because it precludes the possibility of profiting by pursuing the super high second period return, r. Farhi et al show how, “the liquidity floor can be chosen to implement the opti-mal solution” and note that, “this simple regulation resembles the different forms of reserve require-ments imposed on banks. However, our requirement is concerned with regulating the aggregate amount of liquidity.”15 In other words, the Farhi et al solu-tion is not aimed at the solvency of any individual institution, but rather the systemic stability of the market as a whole. A similar and equally effective solution to the retrade market failure is a Pigouvian tax imposed on financial institutions and set to be proportional to their deviation from the optimal amount of liquid-ity.16 Here the optimal level of liquidity should remain the same function of risk preference, and as with other market failures in which actors fail to internalize the aggregate or spillover effects of their decision making, taxing retrade behavior such that the incentives of all market participants become aligned with what is socially optimal is a standard solution to this problem. The level of tax or fee needed to deter free riding and ensure the optimal amount of liquidity can and has been be solved for explicitly as a function of an institution's deviation

Page 90: Visible Hand

90 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

from the optimal investment. The tax would in theory allow for any amount of liquid investment, but would be set such that the marginal gains to the arbitrage strategy were exactly offset by the amount of the tax, therefore forcing any rogue institutions to internalize the social cost of their risk taking on aggregate liquidity levels. Be-cause the tax can be calibrated to achieve a given optimal liquidity level, one that remains dependent on risk preference, whether the regulatory solution chosen is a Pigouvian tax or a direct liquidity floor, the following analysis will be unchanged. In other words, knowledge of the risk preference of the rep-resentative consumer is necessary both in calibrat-ing a tax or setting a mandate, so one solution has no obvious advantage over the other. Thus, the solution to the optimal regulation problem depends only on the level of risk preference and can be achieved with either a tax or a mandate.

V. The Nature of Risk Preferences

Current formulations of the re-trade market failure treat the level of risk aversion and the corre-sponding desire for liquidity as a homogenous char-acteristic across all actors in the economy that the planner need only mandate in order to solve the mar-ket failure. There is substantial evidence, however, that risk preferences, and thus liquidity preferences, are both heterogeneous across individuals and vary-ing over time. Differences in risk preferences across individuals with heterogeneous characteristics have been well documented and have been associated with a wide variety of plausible traits ranging from an individual’s age and household wealth18 (Bucci-ol and Miniaci 201119, Honda 2012) to sex (Byrnes et al 1999)20 and testosterone levels (Apicella et al 2008). Specifically, risk aversion has been found to increase with age and whether or not an individual is female, and decrease with their household wealth and testosterone levels. Relative risk aversion can be imputed using a consumer’s observed willing-ness to pay for risk reduction, for example buying insurance. As demographics in financial markets shift, say with increased financial participation of lower income households, women, or the elderly, so too will the composition of risk preferences in financial markets change. Milan finds empirical evidence that revealed preferences for liquidity in United States financial markets declined between the 1950s and 1980s, but has since risen (Milan 2014). For the conclusions of the retrade model to have greater real world relevance then, a more realistic conception of

risk preferences is needed.

VI. The Model My paper considers what would happen if you relaxed the assumption of a representative depositor with a single risk preference coefficient, d. Instead, I construct a world with 1000 different possible val-ues of risk aversion, with the coefficient, d, rang-ing from -30 to 1. In the original model, the planner would need only know the single risk preference of all depositors in the market in order to calibrate the regulation to the appropriate level of liquidity. When the risk coefficient is allowed to vary along a distribution, however, the utility maximizing level of liquid investment also varies as optimal liquidity is a function of d. The original market failure is preserved in my formulation, as I restrict my range of values for d to preserve the convexity assumption, meaning that all depositors in my economy are either neutral or averse to risk. This ensures that for each depositor type d, their utility maximizing one and two period return, rd1 and rd1 respectively, will be between 1 and r such that 1< rd1<rd2< r. This outcome would not hold for risk-loving depositors, and the consider-ation of such depositors would complicate the mod-el considerably. Because the market failure causes the realized first and second period rates of return in the market to be (1, r), those least averse to risk and who want no added liquidity will be unaffected by the market failure. Optimal liquidity is a decreas-ing function of the risk aversion parameter d, as the more risk averse a depositor, the more liquidity they prefer. There is therefore, no one unambiguously, Pa-reto-improving liquidity regulation. Any liquidity requirement chosen by the plan-ner corresponds to the preference of only one “rep-resentative depositor”, and any regulation would be considered ‘too high’ by depositors who were less risk averse than the “representative depositor”. Similarly, all depositors who are more risk averse than the representative depositor will be forced to accept insufficient liquidity as ‘re-trading’ will still occur, just on a smaller scale, pushing the market level of liquidity down to the required level. This would mean that, given the re-trade assump-tion, planners will not observe higher levels of li-quidity than the mandated level. Thus, any liquid-ity requirement will make depositors who are more risk-averse than the representative depositor worse off than if the liquidity requirement had been set higher.

Page 91: Visible Hand

Fall 2015 | The Visible Hand | 91

Volume xxiV. issue i.

To see this, recall the market conditions that led to the original market failure. Market participants face an arbitrage opportunity in which an actor can count on earning a second period return of r by in-vesting his endowment only in capital. In the event of a shock, he counts on a more liquid partner to trade him their higher first period return x1 for his period two withdrawal rights to yr. Thus all that is needed for this arbitrage opportunity to still occur is for the first and second period rates of return asso-ciated with the mandated level of liquidity r1* and r2*, and the corresponding levels of first and second period consumption x1L and x2L to lie above and below the rates of return associated with a more liq-uid investment, r1 and r2, such that r1<r1

* and r2*< r2

(Appendix A). Because of this, we can expect the mandated level of liquidity, L, to be the realized level of li-quidity, and r1L and r2L to be the rates of return faced by all depositors in the economy. I will use a utili-tarian social welfare function to represent aggregate utility in my economy, in which the total level of utility for each depositor will be equally weighted. Social Welfare is represented by an aggregate ex-pected utility function that weights the expected utility of a depositor with risk preference, d, facing the mandated level of liquidity, L, by the total num-ber of depositors of type d associated with the prob-ability distribution function, pdf(d).

Recall that π is a function of d:

and α, y, and r are all constants. The planner’s problem is to select a liquidity requirement, π *, that maximizes social welfare, and as π is a function of d, the planner’s problem can also be thought of as picking the representative depositor type (or the value of coefficient d) cor-responding to the level of liquidity, L, that would maximize aggregate utility. In other words, while heterogeneous risk preferences prevent the planner from picking a single liquidity level that is optimal for all depositors, they can attempt to minimize the aggregate difference between the liquidity require-ment they set and the preferences of the depositors in their economy.

VII. Method

To begin, I examine the relationship between greater risk aversion (a smaller or more negative risk aversion coefficient) and optimal liquidity. I first assign theoretically plausible numerical values to all of the constants in my social welfare function, a y of 1, a gross two period rate of return of 1.2, and an α of 40 percent. As risk aversion increases, optimal liquidity is increasing at a decreasing rate. The relationship is shown in Figure 1. Figure 1 also depicts the relationship between risk aversion and optimal liquidity for two different levels of risk (probabilities of facing a shock), holding all other parameters constant. π * is an increasing function in α. This result is unsurprising, as the probability of facing a shock increases, so too will the demand for liquidity. I then explore how the socially optimal liquid-ity regulation changes with changing distributions of risk preference. Recall that aggregate social wel-fare is the sum of the utility of all risk preference types, weighted by the proportion of depositors in an economy of each type, given a certain mandated level of liquidity, π (d). Because doing this involves weighting my social welfare function with a distri-bution, the welfare function no longer has an ana-lytical solution. Instead, I use an iterated nonlinear optimization method to solve for the level of liquid-ity (or the representative consumer associated with a level of liquidity) that maximizes aggregate social welfare. I experiment with different distributions of risk aversion to weight the function, and then us-ing Matlab’s unconstrained nonlinear optimization algorithm to find the local maximum. Specifically, I start with risk preference distributed ~N(-15.5,7) . This will serve as my ‘base’ population for further comparison.Using this method, I examine several relationships. I first start by examining how the optimal liquidity, π *, throughout the results expressed in terms of the ‘representative depositor’ with whose preferences π * is associated, changes with the variance of the dis-tribution of risk aversion. I then attempt to capture a measure of the value to depositors of the liquid-ity regulation. Because changes in the numerical estimates of utility have no cardinal significance, I instead calculate the proportional increase in con-sumption, λ,, needed to make each depositor indif-ferent between their expected utility under the real-ized liquidity associated with the market failure and the regulated liquidity, essentially a measure of their willingness to pay. The derivation of λ, can be found

Page 92: Visible Hand

92 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

in Appendix C. The advantage of this approach is that is allows us to see that often, the optimal liquid-ity regulation is not Pareto improving, but instead there are “winners” and “losers”. More risk neutral depositors will be ‘over regulated’ and would need to be compensated for the costs the regulation im-poses on them in order to make the regulation Pareto Optimal. I then repeat this analysis by raising the prob-ability of a shock, α, keeping all else constant, and comparing the representative consumer and the av-erage λ,, with that of the base case. I repeat this com-parison again, raising the two period rate of return. Finally, I repeat this comparison with two skewed normal distributions. Intuitively, these two distribu-tions can be thought of as two economies in which most of the depositors are either highly averse to risk or mostly neutral towards risk.

IV. Results

I begin by considering the base case, where ~N(-15.5,7). Figure 2A shows the distribution of risk preference and Figure 2B depicts the willingness to pay for the optimal regulation, λ,, as a function of risk preference. Figure 3A shows the distribution of risk preference with decreased variance. Decreasing the variance of the risk preference distribution can be interpreted as representing an economy in which there is greater homogeneity of risk preference. In figure 2A, the representative depositor whose preferences correspond to the optimal liquidity floor is denoted by the red circle. It is clear from figure 3A that decreasing the variance of the distribution leads the optimal liquidity floor to be closer to the optimal liquidity level associated with the mean depositor. In figure 2B, the willingness to pay, λ,, is graphed for the base case liquidity regulation. The mean λ, is denoted in green. This can be interpreted as the average proportional increase in consumption, or the average amount that consumption would need to increase, to make the depositor indifferent between the regulated liquidity and the market failure. For example, if the mean λ, was .04, this would imply that consumers would need their consumption to increase by 4% to be as well off under the market failure liquidity level as under the regulation. For most depositors, λ, is positive, meaning that the reg-ulation is advantageous to them. A red line is drawn at λ, = 0, however, to demonstrate the presence of a small number of depositors with negative values for λ, meaning that consumers would need their consumption to be increased by 4% to be as well of under the regulation as they were before. This

implies the possibility of over-regulating such that the amount by which the disadvantaged depositors would need to be compensated to make the regula-tion Pareto-Improving would exceed any gains to the advantaged depositors. This group of ‘worse off ’ depositors was present in every variation explored in this paper, meaning that under no circumstances considered was the optimal regulated liquidity Pare-to-improving. Table 1 reports the complete list of results from all variations with the base case highlighted. The representative consumer associated with an in-creased chance of shock is more risk averse than the base case representative consumer, and the optimal level of liquidity regulation, π*, is higher. This un-surprisingly implies that optimal liquidity regula-tion is increasing with risk. The mean λ in this case is lower than the mean λ in the base case, meaning that the average proportional increase in consump-tion needed, or the average willingness to pay for the regulation is lower. This is likely due to a scal-ing difference. In an economy with increased risk of a shock, overall expected utility is lower than it would be if the shock risk was lower. Expected util-ity is lower in the world with greater risk both with and without the liquidity regulation, meaning the proportional increase in consumption needed to be indifferent, λ, will also be smaller. When the two period gross rate of return on capital, r, increases, the representative consumer becomes less averse than the base case, and the op-timal liquidity regulation drops. The higher the two period rate of return, the higher the opportunity cost of holding low-return liquid assets. Therefore, as r increases, say with increased productivity of capital, the optimal liquidity regulation decreases. In this case, mean λ is higher than the base case. This is likely also due to a similar scaling effect, as overall expected utility is higher as r increases, both with and without regulation. Figures 4 depicts the case in which, all else held equal, α is increased to 45 per-cent and Figures 5 depicts the case in which, all else held equal, r is increased to 1.3. Both figures can be found in Appendix D and E along with Figures 6A and 7A, which depict the two skewed normal distri-butions. 6A represents an economy with mostly risk neutral depositors. 7A represents an economy with mostly risk averse depositors. The optimal liquidity regulation adjusts closer to the mean of each skewed distribution. The mean λ’s associated with these two cases are similar to the base case mean λ. The mean λ of the left skewed distribution is lower than the base, and the mean λ of the right skewed is higher than the base case.

Page 93: Visible Hand

Fall 2015 | The Visible Hand | 93

Volume xxiV. issue i.

VIII. Conclusion

The events leading up to and characterizing the most recent financial crisis make the theoreti-cal study of financial market failures, particularly ones involving liquidity provision, extremely timely and important. My aim was to innovate and expand upon a promising line of research that has identified a plausible and coherent way of modeling the mar-ket wide under provision of liquidity. The demand for liquid assets is motivated by the presence of eco-nomic shocks and the risk aversion of depositors, both reasonable and plausible assumptions. The ex-isting research does not, however, allow for diver-sity of risk preference, and as a result concludes that liquidity regulation is straightforward to calibrate and unambiguously beneficial. The results of allowing risk preference to vary along a distribution demonstrate two important points. The first is that, even if we assume the re-trade market failure does indeed characterize real world financial markets, liquidity requirements were found to be at best, Kaldor-Hicks efficient, as mandating higher levels of liquidity causes a portion of the more risk neutral depositors to be worse off than they would have been under the market failure. The second is that they demonstrate how optimal liquidity regulation varies as the economic environ-ment changes. The level of optimal regulation was found to increase as the risk of a shock increased, decrease as the rate of return to capital increased, and be sensitive to changes in the variance and mean of the distribution of risk preference. These results have real world implications for policy makers cali-brating liquidity regulation. If the re-trade market failure does characterize real world financial mar-kets, phenomenon like lower volatility in the macro-economy and greater capital productivity may call for lowered liquidity requirements, while the chang-ing distribution of risk preference toward greater aversion may call for tightening requirements. As is the case with the provision of any pub-lic good, calibrating the provision of (regulation of) market liquidity is an extremely difficult task. This theoretical exploration was meant to serve a similar function as other models of markets and the effects of market regulations. Knowing how optimal liquid-ity varies with the economic environment, and being conscious of the possibility of over-regulating may help policy makers better understand the implica-tions of their decisions. While perfect knowledge of the values of the parameters of any model is unrealistic, methods to

impute estimates, albeit admittedly imprecise ones, of relative risk aversion exist (Chetty 2006). If such methods were applied to census data, it may be pos-sible to construct time varying distributions of rela-tive risk aversion. Translating a changing distribu-tion of relative risk aversion into a direct rule for liquidity regulation is no more prudent or feasible than any other exact rule translating parameter es-timates into policy responses, however recognizing the theoretical dependence of optimal liquidity reg-ulation both on the existence of the retrade market failure and the distribution of risk aversion of the population suggests that regulators should take into account metrics of societal risk aversion when mak-ing policy decisions.

IV. Sources

Nicolaci Da Costa, P (2014, August 26). Bernanke: 2008 Meltdown was Worse than Great Depression. Wall Street Journal

Markus K. Brunnermeier, 2009. "Deciphering the Liquidity and Credit Crunch 2007-2008," Journal of Economic Perspectives, American Economic Asso-ciation, vol. 23(1), pages 77-100, Winter.

Basel Committee on Banking Supervision, Basel III: The Liquidity Coverage Ratio and liquidity riskmonitoring. (2013) p. 7

Basel III Legislation, p.7

Basel III Legislation: The Liquidity Coverage Ratio and liquidity monitoring tools

Chetty, Raj. 2006. "A New Method of Estimating Risk Aversion." American Economic Review, 96(5): 1821-1834.

Dávila, Eduardo. 2011. “Dissecting Fire Sales Ex-ternalities”

Brunnermeier, (2009)

Diamond, Douglas W. 1984. “Financial Interme-diation and Delegated Monitoring.” Review of Eco-nomic Studies 51 (July): 393–414.

Diamond and Dybvig (1984)

Allen, F. and Gale, D. (2004), Financial Intermedi-aries and Markets. Econometrica, 72: 1023–1061.doi: 10.1111/j.1468-0262.2004.00525.x

Page 94: Visible Hand

94 | The Visible Hand | Fall 2015

Volume xxiV. issue i.

Jacklin, C. J. 1987. “Demand Deposits, Trading Re-strictions, and Risk Sharing.” In ContractualArrangements for Intertemporal Trade, eds. E.C. Prescott and N. Wallace. Minneapolis, MN: Univer-sity of Minnesota Press: 26–47.

Emmanuel Farhi & Mikhail Golosov & Aleh Tsyvinski, 2009. "A Theory of Liquidity and Regu-lation of Financial Intermediation," Review of Eco-nomic Studies, Blackwell Publishing, vol. 76(3), pages 973-992, 2007

Farhi et al (2007)

Mikhail Golosov, 2007. "Optimal Taxation With Endogenous Insurance Markets," The Quarterly Journal of Economics, MIT Press, vol. 122(2), pag-es 487-534, 05.

Borys Grochulski, 2014. “Pecuniary Externalities, Segregated Exchanges, and Market Liquidity in a

Diamond-Dybvig Economy with Retrade,” Eco-nomic Quarterly, Federal Reserve Bank of Rich-mond, Volume 99, Number 4, Fourth Quarter 2013 - Pages 305-340

Toshiki Honda, 2012. “Dynamic Optimal Pension Fund Portfolios When Risk Preferences areHeterogenous Among Pension Participants” , Inter-national Review of Finance, Vol.12, Issue 3, pp.329-355

Bucciol, Alessandro, and Miniaci, Raffaele, “House-hold Portfolios and Implicit Risk Preference,” Re-view of Economics and Statistics, Vol . 93, Issue 4, pp. 1235-1250

Byrnes, James P., Miller, David C., and Schafer, William D. 1999, “Gender Differences in Risk Tak-ing:A Meta-Analysis,” Psychological Bulletin, Vol 125, Issue 3, pp. 367-383

Apicella, Dreber, Campbell, Gray Hoffman, and Little, 2008. “Testosterone and Financial RiskPreferences,” Evolution and Human Behavior, Vol. 29, Issue 6, pp.384-390

Milan, Marcelo, 2014. “Macrofinancial Risks and Liquidity Preference,” International Journal of Po-litical Economy, Vol. 43, Issue 1, pp.43-64

Grochulski (2014)

Page 95: Visible Hand

Fall 2015 | The Visible Hand | 95

Volume xxiV. issue i.

rso.cornell.edu/ces