Impure Altruistic Model: Empirical Investigation of Crowding-out Effects of Private Donations by Government Grants Josephine Xu 20th March, 2012 Abstract A large literature examines the interaction of private and public funding of public good provision and charities. Most of it tests whether private donations are crowded out by pub- lic funding, or more specifically, government grants. This essay seeks to make the following two contributions to the existing literature. First, the impure altruistic model developed in the 1980s is confirmed by empirical testing through the use of tax data. Second, the existing hy- potheses of other determinants of donations, such as fund-raising expenditure, efficiency and programme service revenue, are tested using more recent data, which serves as an update to existing results. I test for crowding-out or crowding-in effects using a large panel data set gathered from charitable organisations’ tax returns. I find strong evidence that government grants do not crowd out private donations and the crowding-in effects found are small, con- sistent with the impure altruistic model. Regression point estimates indicate that fund-raising plays a significant role in determining the level of private donations, consistent with exist- ing literature. In the meanwhile efficiency and programme service revenues are generally not statistically significant. 1
57
Embed
Impure Altruistic Model: Empirical Investigation of ...Impure Altruistic Model: Empirical Investigation of Crowding-out Effects of Private Donations by Government Grants Josephine
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
Impure Altruistic Model: Empirical Investigation ofCrowding-out Effects of Private Donations by
Government Grants
Josephine Xu
20th March, 2012
Abstract
A large literature examines the interaction of private and public funding of public goodprovision and charities. Most of it tests whether private donations are crowded out by pub-lic funding, or more specifically, government grants. This essay seeks to make the followingtwo contributions to the existing literature. First, the impure altruistic model developed in the1980s is confirmed by empirical testing through the use of tax data. Second, the existing hy-potheses of other determinants of donations, such as fund-raising expenditure, efficiency andprogramme service revenue, are tested using more recent data, which serves as an update toexisting results. I test for crowding-out or crowding-in effects using a large panel data setgathered from charitable organisations’ tax returns. I find strong evidence that governmentgrants do not crowd out private donations and the crowding-in effects found are small, con-sistent with the impure altruistic model. Regression point estimates indicate that fund-raisingplays a significant role in determining the level of private donations, consistent with exist-ing literature. In the meanwhile efficiency and programme service revenues are generally notstatistically significant.
If the donor is motivated purely by altruism, qi1 > 0 and qi2 = 0. If the donor is motivated purely
by joy-of-giving, qi1,qi2 > 0 and qi1 + qi2 = 0 . That is to say, altruistically motivated donations
are completely crowded out by government grants while donations motivated by joy-of-giving face
almost zero crowding-out effects.
To analyse the aggregate effects, summing this over all donors, we have:
dH =
NΣ
i=1
qi1qi1+qi2
(dyi−dτi)
1+NΣ
i=1
1−(qi1+qi2)qi1+qi2
+κNdHG (7)
where κN ≡ −AN1+AN
, AN ≡NΣ
i=1
1−(qi1+qi2)qi1+qi2
. Here κN represents the extend of crowding-out and it
depends on AN’s behaviour as N→ ∞.
This model predicts that in a large economy with many donors, under the assumption that
donors’ utility functions are concave, then the effect of government grants on private donations
would asymptotically approaches zero. If the marginal donor is not purely altruistically motivated.
Otherwise, the effect asymptotically approaches 11.
1For proof of this prediction, see Ribar and Welhelm (2002) Appendix
15
4 Data
The data on charitable organisations come from IRS tax returns filed by eligible organisations.
They are based on Form 990 that must be filed by all 501(c)(3) organisations, except religious
groups and organisations with less than $25,000 in gross receipts. Private foundations are not
included as their main source of revenue is investment or endowment and they do not use their
revenue directly in charity activities. These organisations are required to file form 990-PF. The
data I use were sampled by Statistics of Income Division (SOI) of the IRS. Data from 1994-2007
are contained in micro-data files made available on SOI’s website, containing 268,455 observations
from all public charities filed within that fiscal year.
Organisations included in the data set are categorised using National Taxonomy of Exempt
Entities (NTEE). In this system, charities are categorised into 645 level codes, grouped in 26
major groups and 10 major categories. In order to test the model across different categories and
yet maintain a level of comparability, I include 7 out of 10 major categories, excluding religious
groups, mutual benefit organisations and membership based organisations2.
One concern that surfaces in the initial stage of this research is the reliability of the data from
IRS Form 990. It is believed that filing 990 is not a priority, Froelich et. al. (2000)report that
data contained are fairly consistent with audit data, especially in basic items such as donations,
programme service revenue, fund-raising expenditure and so on. I still chose to clean the data.
After dropping 501(c)(4-9) organisations, the data set contains 204,193 (76.1% of the initial
data set) observations on 28026 (68%) organisations. I first drop observations with clear reporting
error, i.e. organisations whose reported revenue by category does not add up to reported total
revenue or whose reported expenditure by category does not add up to reported total expenditure.
This drops 11823 (5.8%) observations and 924 organisations (3.2%). Then I exclude organisations2For description of all then major groups, see Appendix A
16
that report a negative amount for private donations, government grants, programme service revenue
or fund-raising expenditure, eliminating a further 56 organisations.
This data is a very unbalanced panel. I include in my regression only those organisations
present for all 14 years, that is 63951 observations (33.2% of the total) and 4900 organisations
(30.2%). Below I consider how this would affect both the summary statistics and how it will affect
regression results. This balanced data set is used for baseline regressions, and regression results
based on the unbalanced panel are included as well for comparison.
During the initial data coding process, I realised that there are possible discrepancies in NTEE
code and state. In the balanced panel, none change their NTEE code over the 14 year period.
Some of them (0.8%) do change states. I suspect this to be due to data error which could present a
problem for my estimation, because I use state-year level and state-industry level control variables
and instruments. It is also possible that the organisation actually relocated. To distinguish one case
from the other, I consider organisations whose states are not the same for consecutive years to be
errors in reporting as I think it is highly unlikely that an organisation relocates in one year and
move back the next.
4.1 Descriptive Statistics
The charities’ revenue sources are represented in Figure 3, which considers different categories.
The first category is direct public support, which is the main category of private donations. Second
is indirect public support, including donations given to federated fund-raising agencies such as the
United Way. The third category is government grants, which include grants from federal, state
and local governments. Programme service revenue is the amount of revenue collected through
services under which the organisation is exempt from tax. For example, a school would count all
the charges from education services as programme service revenue. Dues collected include only
17
amounts collected that are not donations, such as amount of membership fees going towards access
to facilities or subscription of newsletter. Investment income includes dividends and interest on
savings and cash accounts; rents and sales include revenue from sales of assets. The last category
includes all other forms of revenue, including income from special events.
Table 1 presents and compares summary statistics from the full, unbalanced data set to the
balanced panel. Private donations are calculated as the sum of direct and indirect public support.
Government grants and programme service revenues are taken directly from the data. It also reports
statistics on fund-raising expenditure, management expenditure3, programme service provision
expenditure and efficiency price4. We could observe from this table that means are much higher
than medians and even higher than the 75th percentile, suggesting that the data set is skewed
towards high-revenue firms. Thus, performing econometric analysis on this data set may over-
emphasise the effect from larger charities. Trends in these values from the unbalanced as well as
balanced sample are presented in Figure 2. The presence of crowding-out would imply that peaks
in government grants would be seen together with dips in private donations and vice versa. No
such pattern can be observed from Figure 2 as both seem to be increasing in the time period under
consideration. However, the trends are different across different categories. Figure 4a shows the
trneds in government grants for each category and 4b shows the trends in private donations for
each category.
3It is the amount of money spent on salaries, office administration, conferences and other activities relating to theday-to-day operation of an organisation.
4It is a concept representing the efficiency of an organisation. Numerically, it is the reciprocal of the proportion ofrevenue spent on providing public goods and services.
18
5 Econometric Methods
Fundamentally, to test the validity of the impure altruistic model, two empirical questions are
examined. First, do government grants affect private donations? Second, what is the size of
crowding-out or crowding-in effects? While the majority of papers agree on the presence of
crowding-in or crowding-out effects, the extent of the effect varies largely from one to another.
In an attempt to address these two questions, I run regressions in which the level of private dona-
tions to a charity is the dependent variable and the level of government grants is an explanatory
variable. My initial attempts are to test whether it is appropriate to include a warm-glow preference
in the impure altruistic model. Following these initial regressions, I decide to include instrumental
variables in my analysis to address the endogeneity issues and hope to gain some understanding of
the causality between government grants and private donations.
The prediction of the impure altruistic model to be tested is that given a large economy with
many donors, crowding-in or crowding-out effects asymptotically approach zero. This prediction
can be translated into two hypotheses: First, it could be the case that the effects of government
grants on private donations are not statistically significant. Or it could be the case that these effects
do exist, but the size is economically small.
My estimating equation takes the form:
yit = βxit +BXit +αi + γt +δi +uit (8)
In this estimation model, I use the natural logarithm of my dependent and explanatory vari-
ables. The dependent variable yit is the level of private donations for charity i in year t . The right
hand side variable of interest xit is government grants. The net crowding-in or crowding-out effect
is β , or the elasticity of private donations in response to a change in the level of government grants.
19
A vector of other explanatory variables Xit is included to account for the variation in private dona-
tions. I include a year fixed effect γt , state fixed effects δi and organisation fixed effects αi. These
fixed effects control for specific unobservable effects using panel data econometric methods. The
Hausman specification test rejects the assumption that the unobservable effect is uncorrelated with
the other regressors, so a fixed-effects model is used instead of a random-effects model. The error
term is uit .
The regression model I used for organisation level analysis is
where, in this particular specification, the dependent variable lnDONit is the natural logarithm of
the dollar amount of private donations received by organisation i in year t. lnFUNDit−1 measures
the organisation’s expenditures on fund-raising in the previous year. lnGRANTit−1 and lnPSRit−1
denote respectively, the natural logarithms of government grants and programme service revenues
the organisation received in the previous year.
PRICE−it represents the average of efficiency price at state- ndustry level of organisation i in
year t. It is constructed by calculating the weighted averages of PRICE for all organisations of the
same industry and located in the same state as organisation i. Similarly, EXP−it−1 is the average
of service expenditure, i.e. amount of revenue spent on providing public goods.
lnPRICEit is the natural logarithm of the "price” of contributing a dollar of output to the
organisation and is defined as PRICEit =1
1−Fit−Mit, Fit =
FUNDitREVit
and Mit =MEXPitREVit
, MEXPit is the
management expenditure and REVit is the total donation revenue of the organisation. The use of
lagged variables is common in the literature (Okten and Weisbrod, 2000; Payne, 2001; Andreoni
and Payne, 2003; Ribar and Wilhelm, 2002), though not typically well justified. It may suggest
20
that individual donors are not well-informed enough to speculate on the current year level of grant
that will be received by an organisation, nor would they have good information on the fund-raising
or revenue generation behaviours of an organisation in the current year.
The inclusion of PRICE , i.e. the efficiency price as it is termed by Weisbrod (1986), is intu-
itively justified under a classical altruistic model. From the above construction we can see that it is
the reciprocal of the proportion of revenue spent on providing collective goods and services. Under
the assumption the donors care about the provision of public goods and services, it represents the
"price” of donating. That is, the higher proportion an organisation spent on providing public goods
and services, the lower the "price” of donation would be. Thus it is reasonable to say that the clas-
sical model predicts a negative effect of PRICE on private donations. In the meanwhile, a donor
motivated by reasons other than altruism would not care much about the level of public goods and
services and the above-mentioned effects would not be significant statistically or economically.
One concern in the specification is that the estimate of β3 will be biased if private donations
and government grants are determined endogenously. If there is a common shock, for instance
the increase in demand for disaster relief after an earthquake, both government grants and private
donations will increase, hence biasing β3 upwards. On the other hand, if a charity chooses to real-
locate funding between donations and grants, there will be a negative relationship between these
two variables, thus biasing β3 downwards. To address this problem, I use instrumental variables in
my regression.
An appropriate instrument must be relevant (affecting the level of government grants) and
excludable (not affecting private donations directly). I use government transfers to individuals
from Supplemental Security Income programme (SSI)5 and the state government budget6. SSI’s
basic level is determined at the federal level, but in reality many state governments choose to5This is the overall level of transfers and government giving in a state in a particular year. Data available at US
Social Services Administration6State budget data obtained from National Association of Budget Officers
21
supplement the basic level. Thus, SSI is an instrument used to proxy for the level of generosity
of a state government, which is determined in a political process. This may be directly correlated,
however, with private donations as generous donors tend to elect generous governments. The state
budget is included as it positively correlates with level of government grants to organisations but
is unlikely to directly influence private donations.
Since all the data for instruments are at the state level, I collapsed my data set into one with
state-industry level averages, and treat each industry in each state as a unit of observation and run
the regression with the above mentioned instrumental variables. The only difference between this
state-level regression and the full panel regression is that I no longer include relative efficiency
price and relative service provision expenditure. These two variables were initially constructed as
the state-industry level averages and upon aggregating, they will simply be the level of efficiency
price and service provision expenditure in the new state-industry level data set. At the state-level,
I added in two control variables to the regression. I obtain the state-year level unemployment rate
from the Bureau of Labour Statistics and the state-year level of GDP from the Bureau of Economic
Analysis. Thus, my regression model for this analysis is
lnDON jt = β1 lnFUND jt−1 +β2 lnPRICE jt +β3 lnGRANTjt−1 +β4 lnPSR jt−1
+β4 lnGDPjt−1 +β5 lnUE jt−1 +u jt (10)
GRANTjt−1 = γ1BUDGETt−1 + γ2SSIt−1 (11)
where GDP is the state-level gross domestic product and UE is the state-level unemployment rate.
The regression results and test statistics suggest that identification of the effect of government
grants on private donations is successful. The F-statistics for the significance of the instruments in
the first stage are all greater than 10. Other papers have used lagged variables in their first stage
regression. It could eliminate part of endogeneity but it still remains a concern. I include the results
22
from such models in the appendix.
Lastly, I also run regressions by industry, as previous literature suggests that different indus-
tries exhibit different behaviours which can already be observed from Figure 4. This seems to have
been a reason that many of the papers on crowing-out effects do not agree with each other. As an
alternative approach, I also use the interaction between industry dummy variables and government
grant at the state level to examine how the effects of government grant on private donations would
differ between industries. In this regression with interactions, I use a simultaneous equations ap-
proach to estimate the coefficients. First I constructed variables of the form Industry×GRANT
, Industry×Budget and Industry× SSI , the latter two serve as instruments of the interaction
involving the endogenous variable GRANT . This is valid asIndustry variables are exogenous.
I choose to exclude religion related groups in all of the above-mentioned regressions, because
there are only 32 such organisations in the full panel data set and 14 in the balanced data set7.
Given such a small number of observations present, the regression results will not be representative.
Moreover, most of the religion related groups do not generally receive government grants and thus
the regression results will not be very relevant to this analysis.
For all instrumental variable regressions, I report the first stage regression results in the Ap-
pendix. In the main regression tables, I present the F-statistic in the instruments from the first stage
regressions, Hansen over-identification test J-statistic and the Cragg-Donald F-statistic for weak
instruments.7balanced data set, henceforth, refers to the data set containing only organisations that have data available for all
14 years.
23
6 Regression Results
The results are presented in Table 2 through 8. Table 2 and Table 3 present results with different
fixed effects. Table 2 reports results from the unbalanced panel and Table 3 reports results from
the balanced panel. Regressions by industries are presented in Table 4 and Table 5. These regres-
sions are done using first-difference panel regressions, i.e. all fixed effects are being accounted for.
Table 4 is based on an unbalanced panel while Table 5 is based on a balanced panel. Instrumented
regression results, using specification 10, are reported in Table 6 (first stage results in Table B1).
The regression results with specifications involving interactions between industries and govern-
ment grants are presented in appendix tables C1 and C2, with first stage F-statistic summaries.
Table 7 and 8 provides robustness checks by limiting the panel to different sub-samples. Table 7
is done using first-difference panel regression based on balanced panel, and results are compared
with Column 4 of Table 3. Table 8 is done using specification 10 with instruments based on the
balanced panel, and results are compared with Column 2 of Table 6. First stage results for Table
8 can be found in appendix Table B2. All reported standard errors are robust to heteroskedasticity
and auto-correlation with a Newey-West estimator of the co-variance matrix. The analysis of re-
gression results focuses mainly on those obtained from the balanced data set. The results from an
unbalanced panel are not explicitly discussed, though they are very similar to those obtained from
the balanced data set.
The main focus of Table 2 and 3 is to compare how the magnitude and level of significance
change as more fixed effects are added in. In the first column, only time fixed effects are included
and all coefficients except relative efficiency price are significant. At this early stage, it appears
that there exists no crowding-out effects and only crowding-in effects. Programme service revenue
seems to crowd out private donations, fund-raising expenditure brings in more donation as expected
(as it provides more information to donors) and the efficiency price is negatively correlated with
24
private donation, which is consistent with the classical model. State fixed effects are added in
the second column and coefficients’ magnitude do not vary much, though the relative efficiency
price appears to be more significant in this column. In the third column, industry fixed effects
are included as well. In this column, programme service revenue ceases to be significant and the
magnitude and significance level of the efficiency price drops. For the last column, all fixed effects
are being considered through the use of a first-difference panel regression and we notice a drastic
decrease in the magnitude of government grants, fund-raising expenditure, the efficiency price and
relative service provision. The government grant and fund-raising expenditures remain statistically
significant at 1% levels. The Programme service revenue coefficient becomes positive in the fixed-
effects model and statistically significant at the 1% level. This suggests that although programme
service revenue can theoretically have both positive and negative effects, the negative effects are
more likely to be due to unobservable organisation or industry level effects. Once all fixed effects
are accounted for, the positive effect prevails and this finding is consistent with previous literature
which suggests no negative effect of PSR on private donations and some positive impact (Khanna
et al., 1995).
6.1 Government Grants
After all fixed effects are accounted for, the coefficient for government grants remains positive and
statistically significant at 1% significance level (Column 4 of Table 3). This is consistent with
earlier work by Khanna et al. (1995) and Payne (2001). This is also consistent with the impure
altruistic model which predicts that additional increases in govrnment support have very limited
effects on private donations in a large economy. A statistically significant yet economically small
estimate indicates the donors’ preferences are not guided at the margin by altuisim.
In Table 4 and 5, I use panel regression with fixed effects to explore the difference in coef-
25
ficient for government grants in different industries. Column 1 of Table 5 is taken directly from
Column 4 of Table 3 for comparison and the next 7 columns represent one industry each. The
coefficients for government grants are positive and significant in general, with the exception being
the education industry. One explanation for the lack of significance in estimator for government
grants in the education industry would be that almost all donors are well aware of the presence of
these organisations and there are many other ways for donors to understand the quality and quant-
ity of goods and services provided by organisations in education. Thus, government grants would
not be as strong a signal of either the presence or the quality of these organisations.
Another interesting thing to note from this table is that in international and foreign affairs
organisations, the magnitude of the government grants’ coefficient is almost twice as large as
in other industries. This means that in this particular industry, government grants have a much
stronger effects on private donations. This is reasonable as one would expect these organisations
to rely heavily on government grants. Government grants would send a much stronger signal in
the industry of international affairs than in other industries, possibly because an average donor has
no thorough understanding of the needs of people in a foreign country or are not even aware of the
existence of these organisations.
In Table 6 I run regression based on specification (10) at the state level. Column 1 is based on
the unbalanced panel while Column 2 is based on the balanced panel. The coefficient of govern-
ment grants remains positive and statistically significant at 5% significant level. Tables 6 presents
the F-statistics for the joint significance of the instruments in first-stage regression and they are
strongly significant for all specifications. They also present Hansen J-statistics from test of the
over-identification restriction, possible because the number of instruments exceeds the number of
endogenous regressors. The null hypothesis of this test is that the instruments are valid, so a rejec-
tion calls into question the validity of the instruments. The null hypothesis cannot be rejected in
Table 6. Cragg-Donald F-statistics from a test for weak instruments are also reported. It should be
26
noted that the Cragg-Donald F-statistics are quite low compared to their critical values.
In these regression results, I found no evidence of crowding-out. To the contrary, there exist
crowding-in effects. However, these effects are of very small magnitude for 6 out of 7 industries.
This small magnitude confirms the prediction of the impure altruistic model, and suggests that
at the margin, the donor is not purely motivated by altruism. In addition, this rather consistent
effect for all the major groups of nonprofit organisations suggests that the difference in estimates
of government grants’ effects cannot be due to the types of organisations included in a sample.
Overall, regression results suggests that a 1% increase in government grants to a charity increases
the charity’s private donations by about 0.08%. These results are generally significant at the 1%
level.
Though fund-raising activities could be helpful, government grants to these organisations
make donors aware of potential recipients and needs of which they may not have been aware.
6.2 Fund-raising Expenditure
As discussed in Section 3.1, fund-raising expenditure has two opposing effects. On one hand it
provides more information to potential donors and increases private donations and on the other
hand, it can be seen by concerned donors as a "waste" of revenue and thus discourages them from
donating. In my regression, the first effect is captured through the coefficient on fund-raising
expenditure and the second effected captured by the coefficient of efficiency price. In Table 3, it
is clear that the coefficient on fund-raising expenditure is positive and significant both statistically
and economicaly. In the meanwhile, the coefficient on efficiency price is no longer significant once
all fixed effects are accounted for. Thus I would conclude that the overall effect of fund-raising
expenditure is positive, meaning that it increases private donations by providing donors with more
information.
27
In Table 5, the cross-industry analysis shows that efficiency price is not significant except
for environment and animal support organisations. One explanation would be that many of these
organisations’ financial situations are under more scrutiny and the public tends to criticise these
organisations’ "waste" of revenue more. However, without further data and study to substantiate
this explanation, the result remains puzzling. Across all 7 industries in Table 5, coefficients on
fund-raising expenditure are all positive and significant at 1% significance level. There, however,
are differences in magnitudes of these coefficients, and I would speculate that the more well-known
an organisation is, the smaller the impact fund-raising would have on private donations marginally.
This can be confirmed if a proxy for reputation of an organisation can be identified and included
in the specification. 8. Once instrument variables are introduced, efficiency price is not siginificant
and fund-raising expenditure becomes more siginificant economically.
The lack of statistical significance of the coefficient on efficiency price seems to be in lieu with
the prediction of the impure altruistic model. The foundation of the idea of efficiency price and
inclusion of it in econometric estimation models are based on the assumption that donors care about
the output of public goods and services. If donors do not care what amount of quality of public
goods are provided, then they have no reason to see expenditure on fund-raising or management
as a "waste" of revenue. However, this is not to say that donors do not care about the provision
of public goods and services at all, it only suggests that donors do not care as much as previous
literature conjectured and donors do not donate just because they would like to see more public
goods provided.
One interesting thing to note here about fund-raising expenditure is that it appears to have a
much stronger impact on private donations then government grants do. In Table 3, the effect of
fund-raising is about 5 times stronger as suggested by regression results and in Table 6 the effect
8Many other papers suggest that the age of an organisation, i.e. the number of years since the establishment of anorganisation, would serve as a good proxy for reputation. However, such a variable is no longer available in the SOImicrodata file and thus excluded from the specification
28
of fund-raising is estimated to be about ten times the size of government grants’ effect. Since both
fund-raising and government grants can serve as signals to potential donors, this big difference in
magnitude suggests that donors react more to changes in fund-raising than changes in government
grants. One possible reason for this could be that fund-raising events have a much higher visibility,
meaning that when there is a fund-raising campaign, many donors would have direct access to it.
In the meanwhile, though government grants could be a more credible signal, these are not easily
accessible. The amount and type of grants are usually released through government bulletins or
organisations’ press releases, both tend to be lengthy and not many donors would actually read
through all of them.
6.3 Other Determinants
Programme servie revenue is a source of income for non-profit organisations and thus theoretically
under the classical assumption of donation, it could crowd out private donations. On the other
hand, it can also be seen as a signal of self help, indicating that an organsation is serious about its
cause. This could increase private donations. From Table 3, programme service revenue has a pos-
itive impact on private donations after all fixed effects are accounted for. Before fixed effects are
included, its coefficients are negatively significant. This change in sign suggests that the crowding-
out effect programme service revenue has on private donations is due to unobservable effects at
the organisation level. In Table 5, programme service revenue is significant in 4 out of 7 industries
which is not surprising. When we look at the 3 industries in which it is not significant, we realise
that these industries such as environmental groups, international aid organisations and public bene-
fits organisations (such as research institutes) do not generate revenues under the activities they are
exempt from taxes, or at least the generation of programme service revenue by these organisations
are not visible to the public. To the contrary, the other organisations generate programme service
revenue through much more explicit and visible manners. For example, arts organisations could
29
hold exhibitions and performances with tickets sales, educational organisations do charge for the
education they provide, health services charge for medical services and human services organisa-
tions such as legal assistance groups charge fees for services as well. Though these fees may be
below market rate, they are still considered revenues generated and donors to these organisations
are well aware of the existence of these incomes. In the instrumented regressions, programme ser-
vice revenue are no longer significant. It is worth noting the signs, magnitude and significance of
this estimator vary widely in different specifications and with different industries, which I suspect
could be due to an underlying mechanism involving interactions of programme service revenue
and other explanatory variables in each speficiations 9.
The inclusion of relative efficiency prices and relative service provision aims to examine
whether there exists a substitution effect when donors decide which organisation to donate to.
Intuitively, if donors know there are more efficient and better organsations within proximity and
offering similar services, they would be more inclined to donate to these better organisations. When
fixed effects are accounted for, both terms are negatively significant, suggesting the existence of
substitution effects. However, the actual substitution effect may not be as significant because the
relative efficiencies and provisions are aggregated at NTEE major group level and within a major
group, organisations still provide a wide range of public goods and services. As the industries are
divided into smaller sub-groups, I would expect the substitution effect to decrease significantly.
This would possibly explain why when the regression is performed within each major group, the
significance level drops drastically.
State level control variables GDP and unemployment rates are important to be included in
specification (10) because the aggregate level donations would be more closely related to the over-
all economic conditions. I omitted them in specification (9) following the previous literature in
9For instance, one possibility would be that in an economic recession, organisations would generate much lessrevenue. Thus programme service revenue would be correlated with GDP level and unemployment rate in specification10
30
which state level macroeconomic indicators do not have a significant impact on individual donors
or organisations (Okten and Weisbrod, 2000; Heutel, 2009; Payne, 2001).
6.4 Robustness Checks
To examine the robustness of above estimated coefficients, I limit the data set to smaller sub-
samples and perform regressions on them. The results are reported in Table 7 and Table 8. In each
table, the first column excludes organisations that has never received government grants, the second
column excludes organisations that ever have zero fund-raising expenditure and the third column
excludes organisations with revenues in the top 5%. Table 7 is produced using panel regression
with first difference, in its last column is results from Column 4 of Table 3. Table 8 is done using
specification (10) and its last column is Column 2 of Table 6.
From both tables, it appears that the estimated coefficients are robust to these restrictions.
Of special interests is the results in Column 3 of both tables. While cleaning up the data, I have
concerns that the estimates would be strongly biased towards the larger organisations and suspect
that if a sample with small organisations was used, the results would be quite different. However,
we can see from Column 3 that though the coefficients on government grants increase, the changes
are rather small and these coefficients remain significant at 1% level. This finding seems to suggest
that the impact of government grants has some correlation with the size of an organisation, or if
size can proxy reputation of an organisation, then the extent of crowding-in effects would depend
on the reputation. Coefficients on fund-raising expenditure decrease by about 50%, though still
significant at 1% level. Again, if size is a good proxy for reputation, I would suspect that the
success of fund-raising events depends on reputation as well. The fact that fund-raising changes
more than government grants do suggests that again, fund-raising activities appear to be more
visible and accessible to donors than the information regarding government grants.
31
The by industry regression also provides a means of robustness checks. Some types of non-
profits may not truly be providing public goods, though they are of nonprofit status. For instance,
it is debatable whether a theatre company or an orchestra provide output that can be considered a
public good. It also gives us an opportunity to examine how certain characteristics only pertinent
to one industry would affect the coefficients.
7 Conclusion
7.1 Contribution to existing literature
Warr and Roberts proposed crowding-out of private donations by government grants. This was ex-
tended to include a warm-glow effect in Andreoni. Since then a large body of empirical studies has
followed. Many of them, including Kingma and Payne find significant evidence of partial crowding
out effects, and other papers such as Khanna et. al and Payne find evidence of crowding in effects.
Many researchers believed that the differences arose due to econometric specifications or the type
or organisations involved. Only until recently, Ribar and Welhelm suggested that the difference
lies within the impure altruistic model and could be related to the size of samples used. Here I ex-
tended the literature by using a large data set including all types of charitable organisations except
religious groups. The results I obtained above appear in favour of the impure altruistic model and
question the explanation that different types of organisations exhibit very different crowding-in
or crowding-out effects. I extended Ribar and Welhelm’s empirical study from international aid
organisations to all types of organisations and validate the prediction that crowding-out effects are
small in large samples.
My findings confirm the effect of fund-raising as a way to provide more information to donors
and thus increase donations. In addition, the fact that fund-raising brings in more private donations
32
than government grants would suggest that organisations spend relatively more on fund-raising
campaigns and less on lobbying activities. I found no evidence of efficiency playing a significant
role in determining private donations as Okten and Weisbrod (2000) suggested, it seems to confirm
the assumption that donors do not care much about how raised funds are spent, thus suggesting
that their preference includes more factors than just providing public goods and services. It is
reasonable to think that government could increase funding to charitable organisations in order
to increase social welfare. However, it is worth noting that the efficiency of an organisation, i.e.
where it is spending its funds, cannot be monitored very well, and if donors do not care much
about efficiency either, there is need for new mechanisms to fund these organisations and keep
them accountable.
7.2 Limitations and possible future research
According to the impure altruistic model, if for an industry the crowding-out effect is close to
zero, one possible explanation would be that at the margin altruistic motivation is weak. On the
other hand, if the crowding-out effect is close to one, it might be because the marginal donor is
motivated by altruism. If the difference in motivations can be justified by putting them in the con-
text of policy and legal framework specific to an industry, it could explain the differences in the
extent of crowding-out effects across different industries. For this conjecture to be possible, the
model requires a large economy with almost infinitely many donors. The panel data set to be used
fulfills this condition: it contains over 160,000 observations for 14 years. However, unless rela-
tionship between number of donors and crowding-out effects is examined, the proposition cannot
be explicitly confirmed by empirical evidence.
One possible extension would be to consider a model involving different types of public
goods, rather than just one public good. How would government and individuals respond to
33
changes in contributions among different types of organisations is an interesting question to study.
The regression results for different industries in Table 4 and 5 have suggested that different in-
dustries exhibits different behaviours, but it does not consider the possibility that donors donate to
more than one type of organisations and government must allocate finite funding between different
types of organisations.
In addition, the finding of this research is limited to the extent that there were no major or
drastic changes in government policies towards charitable organisations between 1994 and 2007.
Though government grants to the charitable sector has increased significantly, it happened gradu-
ally throughout the years rather than over a short period of time. Thus the results are limited to
situations where changes in governmental support are modest. The validity of the model in the
context of drastic changes in government grant requires further study using simulation and/or field
experiments.
7.3 Policy implications
Government supports the provision of public goods, whether by providing them itself or by fund-
ing organisations that provide them. The government’s aim would be to increase social welfare by
overcoming the free rider problem inherent in public goods. Prior literature on crowding out sug-
gests that governments ought to acknowledge the effects of their grants on private donations. This
paper suggests that the government may not have to, as the effects of government grants on private
donations are very limited. Instead of being concerned with the direct effect it has on private dona-
tions, it should focus more on how grants affect organisations’ fund-raising behaviour, which has
a bigger impact on private donations.
34
References
Andreoni, J. (1989). Giving with impure altruism: applications to charity and ricardian equival-
ence. Journal of Political Economy, 97:1447–1458.
Andreoni, J. (1990). Impure altruism and donations to public goods: a story of warm-glow giving.
Journal of Political Economy, 100:464–477.
Andreoni, J. (1993). An experimental test of the public goods crowding-out hypothesis. American
Economic Review, 83:1317–1327.
Andreoni, J. and Miller, J. H. (2002). Giving according to garp: an experimental test of the
consistency of preferences for altruism. Econometrica, 70:737–753.
Andreoni, J. and Payne, A. A. (2003). Do government grants to private charities crowd out giving
or fund-raising. American Economic Review, 93:792–812).
Cloward, R. and Epstein, I. (1965). Private social welfare’s disengagement from the poor: the
case of family adjustment agencies, pages 623–624. John Wiley and Sons, New York.
Froelich, K., Knoepfle, T., and Pollak, T. (2000). Financial measures in nonprofit organisation
research: comparing irs 990 return and audited financial statement data. Nonprofit and Voluntary
Sector Quarterly, 29:232–254.
Heutel, G. (2009). Crowding out and crowding in of private donations and government grants.
National Bureau of Economic Research Working Paper.
Khanna, J., Posness, J., and Sandler, T. (1995). Charity donations in the uk: new evidence based
on panel data. Journal of Public Economics, 56:257–272.
Kingma, B. R. (1989). An accurate measurement of crowd-out effect, income effect, and price
effect for charitable contributions. Journal of Political Economy, 97:1197–1207.
35
Kolm, S. C. (1969). The optimal production of social justice, pages 145–200. London.
List, J. A. (2011). The market for charitable giving. Journal of Economic PErspectives, 25:157–
180.
Okten, C. and Weisbrod, B. A. (2000). Determinants of donations in private nonprofit markets.
Journal of Public Economics, 75:255–272.
Palfrey, T. R. and Prisbrey, J. E. (1996). Altruism, reputation and noise in linear public goods
experiments. Journal of Public Economics, 61:409–427.
Palfrey, T. R. and Prisbrey, J. E. (1997). Anomalous behavior in public goods experiments: how
much and why? American Economic Review, 87:829–846.
Payne, A. A. (2001). Does the government crowd-out private donations? new evidence from a
sample of non-profit firms. Journal of Public Economics, 69:323–345.
Ribar, D. C. and Wilhelm, M. O. (2002). Altruistic and joy-of-giving motivations in charitable
behavior. Journal of Political Economy, 110:425–457.
Roberts, R. D. (1984). A positive model of private charity and public transfers. Journal of Political
Economy, 92:136–148.
Rodgers, J. D. and Hochman, H. M. (1969). Pareto optimal redistribution. American Economic
Review, 59:542–557.
Rose-Ackerman, S. (1982). Charitable giving and excessive fund-raising. Quarterly Journal of
Economics, 97:193–212.
Steinberg, R. (1991). Does government spending crowd out donations? interpreting the evidence.
Annals of Public and Cooperative Economics, 62:591–617.
36
Straub, J. D. and Manzoor, S. H. (2005). The robustness of kingma’s crowd-out estimate: evidence
from new data on contributions to public radio. Public Choice, 123:463–476.
Warr, P. G. (1982). Pareto optimal redistribution and private charity. Journal of Public Economics,
19:131–138.
Weisbrod, B. A. and Dominguez, N. D. (1986). Demand for collective goods in private nonprofit
markets: can fundraising expenditures help overcome free-rider behavior? Journal of Public
Observations 29786 4905 11103 941 4641 332 1321 6511R2 0.114 0.128 0.157 0.265 0.103 0.217 0.110 0.088Industries (1)-(7):Arts, Education, Environment and Animal Support, Health Services, International Affairs, Public and Societal Benefits and Human ServicesStandard errors in brackets∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
45
Table 5: First Difference Regression by Industry on Balance Panel
All Organisations Arts Edu Env Health Intl Public Human
Observations 16167 2281 7848 458 2432 174 624 2336R2 0.162 0.178 0.191 0.446 0.136 0.285 0.138 0.169Industries (1)-(7):Arts, Education, Environment and Animal Support, Health Services, International Affairs, Public and Societal Benefits and Human ServicesStandard errors in brackets∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
46
Table 6: IV Regression
(1) (2)
Fundraising Expenditure 0.615∗∗∗ 0.696∗∗∗
[0.0348] [0.0311]
Government Grant 0.0848∗∗∗ 0.0727∗∗
[0.0255] [0.0276]
Programme Service Revenue -0.0268 -0.0129[0.0180] [0.0158]
Efficiency Price -0.0678 -0.0949[0.142] [0.108]
State GDP 0.172** 0.169**[0.0900] [0.0842]
State Unemployment 0.132 0.147[0.0908] [0.0981]
Constant 2.173∗∗∗ 2.196∗∗∗
[0.0746] [0.0687]
Observations 1941 1567
F-statistic for instruments 31.09 35.17Hansen test statistic 1.081 1.032Cragg-Donald Statistic 1.793 1.115Column 1 on unbalanced panel, column 2 on balanced panelOnly observations from 1999-2007 are included as SSI data are only available in those years.Standard errors in brackets∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
47
Table 7: Robustness Checks with Fixed Effects Using Balanced Panel
Exclude if Never Receive Exclude if Zero Exclude if in Column 4 ofGovernment Grants Fund-raising Top 5% of Revenue Table 3
First Stage F-Statistic 28.71 24.19 32.08 31.09Standard errors in brackets∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
49
Appendices
A Data description
A.1 Sample Methods
The sample has two parts by design of SOI. The first sampling frame contains all returns filed by
organisations exempt under Internal Revenue Code (IRC) section 501(c)(3). The second frame in-
cludes a pool of all returns filed by organisations exempt under IRC sections 501(c)(4)-(9). These
two samples are then divided into strata based on the size of end-of-year total assets, with each
stratum sampled at a different rate. The exempt organisation sample is aiming to provide reliable
estimates of total assets and total revenue. This is done by sampling 100% of 501(c)(3) organ-
isations with total assets of $50 million or more, as they represent the vast majority of financial
activities in the charitable sector. 501(c)(4) through (9) organisations with assets of $10 millions or
more are included in the sample. The remaining population is randomly selected at various rates,
ranging from about 1% to less than 100% depending on asset size.
50
A.2 NTEE Major Categories
Table A.1: Description of NTEE Major Categories
Arts, culture, and humanitiesA Arts, Culture and Humanities
EducationB Education
Environment and animalsC EnvironmentD Animal-Related
HealthE Health CareF Mental Health and Crisis InterventionG Voluntary Health Associations and Medical DisciplinesH Medical Research
Human servicesI Crime and Legal-RelatedJ EmploymentK Food, Agriculture and NutritionL Housing and ShelterM Public Safety, Disaster Preparedness and ReliefN Recreation and SportsO Youth DevelopmentP Human Services
International and foreign affairsQ International, Foreign Affairs and National Security
Public and societal benefitR Civil Rights, Social Action and AdvocacyS Community Improvement and Capacity BuildingT Philanthropy, Voluntarism and Grantmaking FoundationsU Science and TechnologyV Social ScienceW Public and Societal Benefit
Religion-relatedX Religion-Related
Mutual/membership benefitY Mutual and Membership Benefit
Programme Service Revenue 0.249*** 0.203***[0.020] [0.024]
State GDP 0.557*** 0.478***[0.173] [0.223]
State Unemployment 0.231* 0.278**[0.113] [0.133]
SSI 0.384*** 0.206**[0.108] [0.142]
Budget 0.310** 0.686**[0.136] [0.164]
Observations 1941 1567
F-Statistic 31.09 35.17Government Grant is instrumented with SSI and BudgetColumns 1-2 correspond to Columns 1-2 in Table 6Standard errors in brackets∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Observation 1552 1541 1560F-Statistic 28.71 24.19 32.08Instrumented variable is Government Grant. Instruments are SSI and Budget.Columns 1-3 correspond to Columns 1-3 in Table 8Standard errors in brackets∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
53
C IV regression with interactions
Table C.1: IV with Interactions on Balanced Panel
OLS OLS with fixed effects IV Panel IV Simultaneous Equations