Does poor health affect productivity? Define labor in terms of efficiency units 8 per time period, so that “effective” labor L * for a farmer is (T - l)8 . f f Let the efficiency per time-period of a laborer depend on his/her a food consumption X , so that i a 8 = 8(X ), where 8N>0, 8O<0 i i The production function now, including hired labor, is a Q = F(L *, L *, V, A) f h In a competitive, full information market there would be a price (wage) for each unit of efficiency labor, call it w*. Then the budget constraint (ignoring the cash crop) is: a a v m m a a w* 8(T - l) + [p Q - w*L * - w*L * - p V] - p X - p X = 0 f h The marginal cost of consuming an additional unit of food is a L* p [1 - (T - l)F 8N] This is less than the market price per-unit of food: Implications: 1. The poor will eat more per unit of income than the rich (lower savings rate than the rich). 2. As long as hired labor is not a perfect substitute for family labor, the poor (small) farmer will get less output per unit of land than the rich (large-land) farmer. Thus poverty is a cause of low savings and low productivity!
141
Embed
Does poor health affect productivity?cru2/econ731_files/Mark/LectureNotes 1a.pdfNumber of observations 2328 2328 2328 2328 2328 2328 2328 2328 2328 Mean of dependent variable 0.41
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
Does poor health affect productivity?
Define labor in terms of efficiency units 8 per time period, so that
“effective” labor L * for a farmer is (T - l)8 .f f
Let the efficiency per time-period of a laborer depend on his/her
afood consumption X , so thati
a8 = 8(X ), where 8N>0, 8O<0i i
The production function now, including hired labor, is
aQ = F(L *, L *, V, A)f h
In a competitive, full information market there would be a price
(wage) for each unit of efficiency labor, call it w*. Then the
budget constraint (ignoring the cash crop) is:
a a v m m a aw*8(T - l) + [p Q - w*L * - w*L * - p V] - p X - p X = 0f h
The marginal cost of consuming an additional unit of food is
a L*p [1 - (T - l)F 8N]
This is less than the market price per-unit of food:
Implications:
1. The poor will eat more per unit of income than the rich (lower
savings rate than the rich).
2. As long as hired labor is not a perfect substitute for family labor,
the poor (small) farmer will get less output per unit of land than the
rich (large-land) farmer.
Thus poverty is a cause of low savings and low productivity!
(use prices of foods, non-labor income as instruments).
2. Are time wages the same as productivity?
3. Is it plausible that employers know calorie consumption of
workers?
Estimated Effects of Calorie Consumption on the log Wage,
by Payment Method:
Bukidnon, Philippines Harvest Workers over Four Seasons
All Harvest
Wages
Harvest Time
Wages
Harvest Piece-
Rate Wages
Calories consumed
(x10 )-3
.211
(2.36)
.0153
(0.22)
.438
(3.13)
Height 1.04
(2.37)
.040
(0.30)
.446
(1.41)
Male -.366
(2.37)
.285
(0.61)
-.357
(2.66)
N 327 291 136
Health and Schooling
Poor health, related to bad endowments - geography - often thought to be one reason for lack of
growth
One linkage - poor health among children reduces school attendance, performance
Examples: 1.3 billion people infected by hookworm and roundworm; whipworm affects 900
million people; schistosomiasis affects 200 million people
Educational impact of de-worming key issue
A. De-worming is relatively cheap - single-dose oral treatment reduces
infections by 99%, but need annual application because of reinfection
B. Children
1. Account for bulk of infections: 85-90% of all heavy schistosomiasis
infections in Eastern Kenya
2. Children most likely to spread disease, worm infections, because have
worst hygiene practices (less likely to use latrines)
3. Believed to affect schooling - anemia, malnutrition, illness
Relates to question: how much does geography matter? Poor health endowment: worms
What is evidence on effect of deworming on schooling outcomes?
Prior studies find little affect of de-worming treatment on school performance - test scores
Methodology seems correct: randomization of individual treatment of children
within schools
Two problems, however:
1. Neglects externalities: non-treated children are less likely to be infected infected
A. Understates effect of treatment for two reasons:
1. “Control” group (placebo) actually affected by the treatment:
difference between control-group and treatment-group outcomes
underestimates effectiveness of treatment on the treated
2. The reduction in the infections of the untreated - the externality - is
part of the benefits of the treatment and is ignored
2. The studies have not looked at a range of outcomes, including attendance in
school, enrollment test scores, promotion rates
New study (Miguel and Kremer, 2003): Primary School Deworming Project in Busia, Kenya
Methodology: Randomized phase-in of treatment at the school level, not individual.
75 schools randomly divided in three groups:
Group 1: free deworming in 1998 and 1999 (treatment in 1998 and 1999)
Group 2: free deworming in 1999 (control in 1998, treatment in 1999
Group 3: free deworming in 2001 (control in 1998 and 1999)
How infected were the children prior to treatment?
Survey of pupils: 92% had at least one helminth infection (understated -
why?)
Treatment: All students in school in treatment schools given drug for geohelminth
infection, except girls 13 and above by public health officials (nurses,
officers)
Also, health education on infection prevention
Treatment rates: 67% of eligible in 1998 (80% attendance); 57% in 1999
Results monitored: Survey one year after first round of treatment
Estimation strategy to take into account externalities:
Takes advantage of the fact that many close neighbors go to different schools: people in
proximity get different treatments
Children attending school or living nearby treatment schools have different exposure to
risk of infection: exposure is number of pupils in treatment school that are nearby, or
distance times treatment pupil density
Thus estimating equation is:
ijt 1 1it 2 2it d d dit d d dit i ijtY = a + $ T + $ T + E (( N ) + E (( N ) + u + eT
ijtwhere Y = outcome for student j in school i at time t
T = assigned treatment in year 1 or 2
ditN = total number of pupils in primary schools at distance d from school
i
ditN = number of pupils in treatment schools at distance d from school iT
(d = 1 is 1 kilometer, d=2 is 2 kilometers, etc.)
iu = school effect
1t d d ditSo, average treatment effect is $ + E (( N ), where N is the average number ofT’ T’
pupils in treatment schools located at d from the school
Results
58
Table 7: Deworming health externalities within and across schools, January to March 1999†
Any moderate-heavyhelminth infection, 1999
Moderate-heavyschistosomiasis infection,
1999
Moderate-heavygeohelminth infection, 1999
(1) (2) (3) (4) (5) (6) (7) (8) (9)Indicator for Group 1 (1998 Treatment) School -0.25***
(0.05)-0.12*
(0.07)-0.09(0.11)
-0.03(0.03)
-0.02(0.04)
-0.07(0.06)
-0.20***
(0.04)-0.11**
(0.05)-0.03(0.09)
Group 1 pupils within 3 km (per 1000 pupils) -0.26***
(0.09)-0.26***
(0.09)-0.11(0.13)
-0.12***
(0.04)-0.12***
(0.04)-0.11**
(0.05)-0.12*
(0.06)-0.12*
(0.07)-0.01(0.07)
Group 1 pupils within 3-6 km(per 1000 pupils)
-0.14**
(0.06)-0.13**
(0.06)-0.07(0.14)
-0.18***
(0.03)-0.18***
(0.03)-0.27***
(0.06)0.04
(0.06)0.04
(0.06)0.16
(0.10)Total pupils within 3 km (per 1000 pupils) 0.11***
(0.04)0.11***
(0.04)0.10**
(0.04)0.11***
(0.02)0.11***
(0.02)0.13***
(0.02)0.03
(0.03)0.04
(0.03)0.02
(0.03)Total pupils within 3-6 km (per 1000 pupils) 0.13**
(0.06)0.13**
(0.06)0.12*
(0.07)0.12***
(0.03)0.12***
(0.03)0.16***
(0.03)0.04
(0.04)0.04
(0.04)0.01
(0.04)Received first year of deworming treatment, whenoffered (1998 for Group 1, 1999 for Group 2)
-0.06*
(0.03)0.03**
(0.02)-0.04**
(0.02)(Group 1 Indicator) * Received treatment, when offered -0.14*
(0.07)-0.02(0.04)
-0.10***
(0.04)(Group 1 Indicator) * Group 1 pupils within 3 km (per1000 pupils)
-0.25*
(0.14)-0.04(0.07)
-0.18**
(0.08)(Group 1 Indicator) * Group 1 pupils within 3-6 km (per1000 pupils)
-0.09(0.13)
0.11(0.07)
-0.15(0.10)
Grade indicators, school assistance controls, districtexam score control
Yes Yes Yes Yes Yes Yes Yes Yes Yes
Number of observations 2328 2328 2328 2328 2328 2328 2328 2328 2328Mean of dependent variable 0.41 0.41 0.41 0.16 0.16 0.16 0.32 0.32 0.32
†Grade 3-8 pupils. Probit estimation, robust standard errors in parentheses. Disturbance terms are clustered within schools. Observations are weighted by totalschool population. Significantly different than zero at 99 (***), 95 (**), and 90 (*) percent confidence. The 1999 parasitological survey data are for Group 1 andGroup 2 schools. The pupil population data is from the 1998 School Questionnaire. The geohelminths are hookworm, roundworm, and whipworm. We use thenumber of girls less than 13 years old and all boys (the pupils eligible for deworming in the treatment schools) as the school population for all schools.
60
Table 9: School participation, direct effects and externalities†
Dependent variable: Average individual school participation, by yearOLS OLS OLS OLS OLS OLS IV-2SLS(1) (2) (3) (4)
May 98-March 99
(5)May 98-March 99
(6)May 98-March 99
(7)May 98-March 99
Moderate-heavy infection, early 1999 -0.028***
(0.010)-0.203*
(0.094)Treatment school (T) 0.051***
(0.022)First year as treatment school (T1) 0.062***
(0.015)0.060***
(0.015)0.062*
(0.022)0.056***
(0.020)Second year as treatment school (T2) 0.040*
(0.021)0.034*
(0.021)Treatment school pupils within 3 km(per 1000 pupils)
0.044**
(0.022)0.023
(0.036)Treatment school pupils within 3-6 km(per 1000 pupils)
-0.014(0.015)
-0.041(0.027)
Total pupils within 3 km(per 1000 pupils)
-0.033**
(0.013)-0.035*
(0.019)0.018
(0.021)0.021
(0.019)Total pupils within 3-6 km(per 1000 pupils)
-0.010(0.012)
0.022(0.027)
-0.010(0.012)
-0.021(0.015)
Indicator received first year of dewormingtreatment, when offered (1998 for Group 1,1999 for Group 2)
0.100***
(0.014)
(First year as treatment school Indicator)*(Received treatment, when offered)
† The dependent variable is average individual school participation in each year of the program (Year 1 is May 1998to March 1999, and Year 2 is May 1999 to November 1999); disturbance terms are clustered within schools. Robuststandard errors in parentheses. Significantly different than zero at 99 (***), 95 (**), and 90 (*) percent confidence.Additional explanatory variables include an indicator variable for girls < 13 years and all boys, and the rate ofmoderate-heavy infections in geographic zone, by grade (zonal infection rates among grade 3 and 4 pupils are usedfor pupils in grades 4 and below and for pupils initially recorded as drop-outs as there is no parasitological data forpupils below grade 3; zonal infection rates among grade 5 and 6 pupils are used for pupils in grades 5 and 6, andsimilarly for grades 7 and 8). Participation is computed among all pupils enrolled at the start of the 1998 schoolyear. Pupils present during an unannounced NGO school visit are considered participants. Pupils had approximately3.8 attendance observations per year. Regressions 6 and 7 include pupils with parasitological information from early1999, restricting the sample to a random subset of Group 1 and Group 2 pupils. The number of treatment schoolpupils from May 1998 to March 1999 is the number of Group 1 pupils, and the number of treatment school pupilsafter March 1999 is the number of Group 1 and Group 2 pupils.
The instrumental variables in regression 7 are the Group 1 (treatment) indicator variable, Treatment school pupilswithin 3 km, Treatment school pupils within 3-6 km, and the remaining explanatory variables. We use the number ofgirls less than 13 years old and all boys (the pupils eligible for deworming in the treatment schools) as the schoolpopulation for all schools.
Calculations:
1. What is the average spillover gain: the cross-school externality reduction in infection?
0-3 0-31 3-6 3-61( N + ( N = .26*454 + .14*802 = .23 (23 percentage points drop)T’ T’
2. What is the average spillover gain: the cross-school externality on enrollment?
2 percentage points increase in non-treatment schools
3. What is the direct effect of the treatment in treatment schools?
7.5 percentage point gain (one quarter drop in absenteeism)
4. What is the total increase in school years from treating one child?
1*.075 + .5*.075 + 1.5*.02 = .14 years
Based on the fact that:
1/3 of children in treated schools not treated
2/3 of all pupils not in treatment schools in 1998, 1/3 in 1999
What is the rate of return from treating one child?
1. Benefit:
A. From wage regression: one year gives a 7% increase in wage (males)
So 14 years gives .14*7 = 1% gain
B. Output per worker in Kenya = $570
C. Wages are 60% of output per worker
D. People work 40 years and wages do not increase
E. Use discount rate of 5%
2. Cost:
A. Opportunity cost of schooling: ½ of adult wage
B. $0.49 for administering the drug
Treating one child gives a net present value increase in lifetime wages of $30
Given externalities, subsidy is warranted but existence of infectious parasites does
not account for schooling or income deficit across Kenya and the developed world.
How Much Does Increasing Health Increase Development?
Observe: Low-Income Countries are less healthy - almost all measures of ill health negatively
correlated with per-worker GDP
Worker productivity lower in countries with low health
Why does this association not tell us the impact of health on development?
1. Low incomes may cause low health
2. Unmeasured factors causing ill health may also be limiting productivity
A. Poor delivery of health services may be due to bad governance, which also
affects ability to invest
3. Also, specific indicators of ill health may be proxying for more general health issues
A. Prevalence of malaria, or anemia may be correlated with other health
problems - so getting rid of malaria may not have a big impact
Two examples: Anemia (low levels of hemoglobin in blood due to worms, malaria, iron
deficiency in diet)
Low birthweight (low maternal nutrition)
Both have strong negative association with incomes by country
Available micro estimates indicating causal effects (from randomization) of:
A. Eliminating anemia
B. Eliminating the birthweight gap between rich and poor countries
Micro evidence:
Anemia: from randomized trials of iron supplementation (Shastry,
2002)
Birthweight: from differences in birthweight among identical twins
(Behrman and Rosenzweig, 2004)
Figure 1Mortality, Anemia, and Income per Capita
20
40
60
80
100
-4.0 -3.5 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5
Ln (Income per Worker relative to US)
Perc
enta
ge o
f Wom
en w
ho a
re n
ot A
nem
ic
400
500
600
700
800
900
1000
Adult Survival Rate
Table 1
Anemia Prevalence and Productivity Loss by Country
Country
Anemia Presence, Non-
pregnant Women
Average Productivity relative
to No Anemia: Randomized-
Design Estimate
India 76.7 .896
Tanzania 69.7 .914
Haiti 55.8 .942
Senegal 48.0 .954
Sri Lanka 45.1 .958
Algeria 36.7 .969
Brazil 29.7 .976
Japan 17.6 .988
United States 5.0 .997
From Kartini Shastry G.; Weil D.N., “How Much of Cross-country Income Variation is Explained
by Health?” Journal of the European Economic Association, Volume 1, Numbers 2-3, 1 April 2003,
pp. 387-396(10)
Figure 1
Low Birthweight and Log Per-worker GDP Around the World
Table 4
Percent Low-Birthweight Births and Mean Birthweight for Selected
Countries
Country % Low
birthweight
Mean
birthweight
Number
of births
Source
India 22.6 98.3
(24.5)a
8650 1998/99 DHS
Philippines 16.2 111.0
(26.3)
4337 1998 DHS
Pakistan 16.0 111.9
(28.8)
607 1991 DHS
Malaysia 14.3 108.3
(19.2)
3941 1976/77 MFLS
Senegal 11.2 110.9
(26.5)
2239 1997 DHS
Uganda 11.2 113.5
(28.2)
1889 1995 DHS
Brazil 9.1 114.3
(21.4)
4427 1996 DHS
Peru 9.0 114.6
(23.4)
10654 1996 DHS
United
States -
whites
8.9 118.2
(20.9)
1711 NLSY79, 1988
round
Standard errors in parentheses.a
Effects of birthweight on adult productivity are non-linear:
Greater impact at lower levels of birthweight
So aggregate effect depends on country-specific birthweight distribution
The formula for computing the percent earnings gain for a country j from closing the
birthweight gap between it and some target country is:
j i ij 1i j%earnings gain = G f á [birthweight gap ].
Examples: Malaysia (10-ounce gap) and India (20 ounce gap) relative to the US
Use no weights, Malaysia weights, India weights based on the birthweight distributions
Table 5
Within-MZ Estimates of Increasing Birthweight on the Ln Wage,
by Country-Specific Sample Weights
Weights No weights Malaysia weights India weights
Sample Full
sample
L.B. Full
sample
L.B. <50% Full
sample
L.B <50%
Birthweight
(ounces)
.00478
(2.41)a
.00643
(2.35)
.00413
(2.04)
.00612
(2.23)
.00502
(2.33)
.00430
(2.24)
.00736
(2.45)
.00452
(2.10)
N 812 404 812 404 632 812 404 616
Abolute value of t-ratio in parentheses.a
Figure 2
Distribution of the Absolute Value of Birthweight Differentials (oz.) Among identical Twins
Cooking with Biomass and Health: The Short- and Longer-run
Effects of Indoor Air Pollution
Facts
One half of the world’s population relies on biomass and coal as their primary source of
household energy, as much as 95% in Bangladesh
Exposure to indoor air pollution (IAP) from solid fuels is alleged to cause several diseases
including acute respiratory infections (ARI) (also chronic obstructive pulmonary disease
(COPD), asthma, cataracts and blindness, and low birth weight.)
ARI accounts for 6 percent of worldwide disease and mortality, mostly in the developing
countries.
ARI is also the most common cause of illness and mortality in children in the developing world
Acute lower respiratory infection accounts for 20 percent of the annual deaths of children
under five, with nearly all of these deaths occurring in the developing countries.
This paper investigates how and to what extent the division of household responsibilities and
ventilation, fuel and stove type affect both child and adult health in rural Bangladesh and India
Bangladesh: Households use biomass for cooking fuel; no substitution possibilities: only
variation in housing and in who cooks
India (national): Households use both biomass and clean fuels, different stove types
Average particulate concentrations of 300 ìg/m or higher are common in Bangladeshi3
households. [California indoor air quality standard = 50 ìg/m .]3
Empirical challenges:
1. Identification of exposure, fuel, stove, ventilation effects on adults and children when all are
endogenous (optimally chosen) - focus on choice of exposure, conditional on technology.
2. Identification of optimizing behavior: do households behave as if they are informed about the
health consequences of IAP?
Theory
A simple heuristic model to show how households seeking to minimize the burden of an unhealthy but
necessary activity will allocate the unhealthy task among its members.
Consider a separable household utility function of the form
where
m m1 m2 mJ H =(h h ,...,h ) is the set of health statuses of the J males in the household,
f f1 f2 fK H =(h ,h ,...,h ) is the set of health statuses of the K females in the household,
m m1 m2 mJ f f1 f2 fK X =(x x ,...,x ) and X =(x ,x ,...,x ) are sets containing the allocations of composite
consumption goods.
A single unit of time is allocated between a productive activity that is deleterious to health (“cooking”)
c a(t ) and productive employment (t , “agriculture”) that is not.
cOnly women devote time to the cooking activity and the total quantity of cooking time t that women
c1 c2 cmust provide is fixed, so that time spent cooking must be allocated such that t + t = t .
Health for women is produced with technology
where
fiì = the exogenous component of health (health endowment).
The productivity of time spent in agriculture is sensitive to health
cibut the productivity of time spent cooking, t , is not.
Households maximize the utility function subject to the health technologies, the productivity functions,
the time constraints, and the budget constraint
where
mi ai miv = the sum of non-earnings income and male earnings net of their consumption (3w t -p3x )
p = the price of consumption good x.
Implications:
Because health only affects income through agricultural work, then even if all women are
f1 f2 cidentical (ì =ì ) and t # 1 (household cooking time is less than the time available to any one
women), one woman may optimally do all of the cooking and the other woman will specialize in
agriculture if cooking requirements are sufficiently high.
The rationale is clear – any time spent in cooking by a woman reduces her health and
productivity in agriculture without affecting her productivity in cooking.
f1 f2Moreover, if one woman is innately less healthy than another (ì <ì ) she will always do more of
the cooking, while the other will specialize more in agriculture.
Thus, because the less healthy cook, the observed association between time exposed to an unhealthy
environment and ill-health will be an upward-biased measure of the effect of exogenously increasing
exposure.
In the Bangladesh, relative productivities and health endowments may not be the only factors that
allocate women to tasks, but also the identity of women in terms of their relationships to the household
head.
The anthropological literature suggests that mothers-in-law dominate daughters-in-law, mothers
dominate daughters, and elder brothers’ wives dominate younger brothers’ wives. A young wife
submissively follows the lead of her husband’s mother, and is rarely involved in decision-making.
This”social hierarchy” component of time allocation is exploited in the estimation strategy we adopt
(and is assessed).
Bangladesh Data: 2000-2003 Nutrition Survey of Bangladesh
The survey sample has two components:
(1) a random sample of all households in 14 villages carried out in 2000, and
(2) a panel survey, consisting of the households of all surviving individuals included in the 1981-
82 Nutrition Survey of Bangladesh (Ahmed and Hassan, 1983) (originally sampled from the
same 14 villages) regardless of their residence during the interval 2000-2003, some of whom
were also included in the 14 village random sample frame of year 2000.
This data set provides multi-level (individual, household and village) survey information on health
status, activities, food consumption and resources for over 4000 men and women.
The panel component of the survey, by following individuals who departed from the original 14
villages, is characterized by very low (3%) household and individual attrition rates despite the
approximately 20-year interval between rounds.
The questionnaires elicited information on:
(i) detailed activities, including those for pay and not for pay, as well as the earnings from paid
activities, in a 24-hour period for every household member. Activities were coded into more than
150 categories. A key feature of this module is that activities are anchored around the five prayer
times observed during the day, salient features in the lives of the Muslim respondents.
(ii) detailed food consumption for every member of the household observed over the same 24-
hour period
(ii) home dimensions and construction materials including the location of the kitchen (outside or
not) as well as roof and wall material , and
(iii) 23 health symptoms for all household members, provided to a trained clinician by each
household respondent over age 10, and for children less than or equal to 10 by the relevant
mother. Of the 23 symptoms, three symptoms are respiratory-related (coughs, difficulty
breathing, with or without fever).
Table 1
Sample Characteristics
Variable Adults Aged 16+ Children 2-9
Respiratory symptoms .154
(.361)
.221
(.415)
Intestinal symptoms .0489
(.216)
.0694
(.254)
Age 36.0
(15.8)
6.09
(1.98)
Education (years) 3.51
(4.14)
.515
(.939)
Female .477
(.500)
.485
(.500)
Mother with children<5 .242
(.563)
-
Wife of head .278
(.448)
-
Daughter-in-law of head .0517
(.221)
-
Total hh expenditures (taka
per month)
4797
(4176)
4558
(4856)
Permeable roof .205
(.403)
.224
(.417)
Permeable walls .326
(.469)
.377
(.484)
Kitchen outdoors .262
(.440)
.292
(.455)
Number of individuals 4026 1365
Number of households 1198 780
Does cooking time cause respiratory illness symptoms among adults?
The equation estimated is given by
ij t ij A ij j z j X hj hij hij(9) h = á t + á A + Z á + X á + ì + ì + e ,
where
ijh is the incidence of any respiratory symptom for person i in household j;
ijt is the time spent cooking;
ijA is a set of person-specific attributes (age, sex, education);
jZ is a vector of household-level smoke-related factors that reflect ventilation (permeability of
walls and roof, and whether cooking is carried out outdoors);
jX is a vector of other household-level characteristics that may affect health, such as income;
t A z Xá , á , á and á are the corresponding vectors of coefficients.
Because cooking time may be correlated with unmeasured household and individual-specific health
variables, we also use a household fixed-effects procedure, which differences across women in the same
household
ij t ij A ij hij hij(10) Ä h = á Ä t + á Ä A + Ä ì + Ä e ,j j j j j
where Ä is the across-person difference operator.j
tBut estimates of á from (10) will not be consistent if:
(1) the within-household distribution of women’s chores is related to the differences in individual
health endowments (upward bias according to the model), or
ij(2) there is measurement error in the time exposed to smoke (t ) variable (bias to zero).
The net effect of the two sources of bias is of opposite direction and unknown a priori
Two strategies:
1. Instrumental variables applied to (10)
2. Within mother estimates: sweeps out individual endowment, identifies exposure effects by child age
Instrumental Variables
For instruments, variables are required that affect the allocation of cooking time across women in the
same household but do not, given a women’s time allocation, otherwise affect her health.
Households in rural Bangladesh contain sexually segregated spheres of influence in which gender-
specific hierarchies, based in large part on relationship to the head of household, operate to allocate
women to tasks based both on the gains from specialization and on rank in the household hierarchy.
Differences across women in their relationships to the household head are unlikely to directly affect
differences in respiratory health or be correlated with individual health endowments or productivity net
of age.
The instruments are dummy variables indicating whether the person is a wife of the head or a daughter-
in-law, the interaction of wife and daughter-in-law with the number of daughters-in-law, and the
interaction of daughter-in-law with the presence of any wife of the head in the household.
Diagnostics
Figure Y: Life-Cycle Status Changes: Proportion of Married Womenby Status Category and Age
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
20-29 30-39 40-49 50-59
Wife of HeadDaughter-in-law of HeadMother of HeadHead
Table 2: The Effects of Cooking Time on the Incidence of Respiratory
Symptoms, All Adults and Women Only,
by estimation Procedure
All Adults Aged 16+ Women Aged 16+
Estimation
procedure
RE FE-
HHold
FE-IV RE FE-
HHold
FE-IV
Cooking time
(x10 )-3 a
.152
(2.77)
.0871
(1.56)
.349
(2.72)
.159
(2.68)
.170
(2.51)
.455
(2.58)
Age .00107
(2.84)
.00113
(2.77)
.00146
(3.37)
.00069
(1.12)
.00178
(2.41)
.00284
(2.96)
Education -.0036
(2.09)
-.00106
(0.50)
-.00011
(0.05)
-.0028
(0.95)
.00465
(1.14)
.00892
(1.87)
Female .00093
(0.06)
.0157
(0.92)
-.046
(1.43)
- - -
Total expend
(x10 )-5
-.380
(1.88)
- - -.511
(1.78)
- -
Permeable
roof
-.0115
(0.66)
- - -.0384
(1.59)
- -
Permeable
walls
.00225
(0.15)
- - .0125
(0.59)
- -
Kitchen
outdoors
.015
(0.94)
- - .00948
(0.42)
- -
Number of
individuals
4590 4590 4590 2202 2202 2202
Number of
households
1371 1371 1371 1368 1368 1368
Endogenous variable. Absolute value of asymptotic t-statistic in parentheses.a
Diagnostics for Identification Strategy
1. Overidentification tests:
F=1.07 (p = 0.37) for adults and 0.99 (p = 0.42) for women Pass
but a) the tests may have no power
b) cooking time may be highly correlated with other health inputs: i.e., calorie
consumption
2. Estimate effect of cooking time on health symptoms not implicated as being affected by particulates.
i.e., intestinal symptoms if so, spurious relationship
Use same specification and identification strategy: should find no effect of cooking time on
intestinal symptoms if not spurious relationship
Also apply overidentification test
3. Include calorie consumption in both health equations: control for (endogenous and mis-measured)
calorie allocation
Apply overidentification tests, test cooking time effects net of calories
4. Assess anthropological literature observation: does household status have special relationship with
allocation of cooking?
Table 3
FE-IV Estimates: The Effects of Cooking Time on the Incidence of Intestinal Symptoms, All Adults and
Adult Women Only
Variable/Sample All Adults Women
Cooking time (x10 ) .0391-3 a
(0.49)
.0852
(0.76)
Age .000755
(2.80)
.000808
(1.32)
Education .00108
(0.80)
.000748
(0.25)
Female -.00226
(0.11)
-
Number of individuals 4590 2202
Number of households 1371 1368
Endogenous variable.a
But overidentification test fails: F = 1.81 (p = 0.10)
Test evidently has power - what is wrong?
Omitted input: calorie consumption
Table 4First-Stage FE-Household Estimates: The Determinants of Cooking
Time and Calories Consumed: All Adults
Cooking Time Calories
Variable (1) (2) (3) (4)
Age -1.33(10.2)
-.792(6.73)
-.178(0.20)
.129(0.14)
Education -3.43(4.72)
-2.13(3.35)
-10.7(2.22)
-9.18(1.90)
Female 256.3(72.2)
160.1(33.8)
-728.8(30.9)
-843.6(23.5)
Wife of head - 144.8(28.0)
- 129.0(3.28)
Daughter-in-law ofhead
- 197.6(5.47)
- -163.3(0.80)
Wife x number ofdaughters-in-law in thehousehold
- -67.3(10.2)
- 104.0(2.08)
Daughter-in-law xNumber of daughters-in-law
- -19.0(1.79)
- 10.3(0.13)
Daughter-in-law x Isthere wife of head?
- -35.4(1.58)
- 460.3(2.70)
R .640 .720 .168 .1682
F-statistic (5, 2647) - 171.5 - 5.67
Table 5
FE-IV Estimates: The Effects of Cooking Time and Calories on the
Incidence of Respiratory Symptoms, Body Weight and the Incidence of Intestinal Symptoms: All Adults
Dependent Variable Respiratory
Symptoms
Body Weight (kg) Intestinal Symptoms
Cooking time (x10 ) .392-3 a
(2.35)
11.5
(3.09)
-.00296
(0.02)
Calories consumed (x10 ) .00467-3 a
(0.04)
7.78
(2.92)
.170
(2.01)
Age .00156
(3.10)
.0217
(1.89)
.000787
(2.10)
Education .00205
(0.75)
.245
(3.98)
.00398
(1.97)
Female -.0671
(0.57)
-5.22
(2.07)
.128
(1.50)
Number of individuals 3878 3878 3878
Number of households 1371 1371 1371
Endogenous variable.a
Table 6
FE-IV Estimates: Do House Materials or Kitchen Location Ameliorate
the Effects of Cooking Time on the Incidence of Respiratory Symptoms?
All Adults and Women Only
Variable
All Adults Aged
16+
Women Aged 16+
Cooking time (x10 ) .418-3
(2.89)
.343
(1.59)
Cooking time x permeable roof
(x10 )-3
-.0857
(0.78)
.638
(1.47)
Cooking time x permeable walls
(x10 )-3
.0517
(0.52)
.0634
(0.16)
Cooking time x kitchen
outdoors (x10 )-3
-.0212
(0.20)
-.0183
(0.03)
Age .00168
(3.66)
.00336
(2.94)
Education .000749
(0.32)
.0122
(2.08)
Female -.0506
(1.46)
-
Test statistics, no house effects
÷ (3),( p-value)2
0.93 (.818) 2.43 (.488)
Number of individuals 4590 2202
Number of households 1371 1368
Table 7
The Effects of Mother’s Cooking Time on the Incidence of Respiratory
and Intestinal Symptoms of Her Children 2-9,
by Estimation Procedure
Respiratory Symptoms Intestinal Symptoms
Estimation
procedure
FE-HHold FE-
HHold
FE-
Mother
FE-
HHold
FE-
HHold
FE-
Mother
Mother’s
cooking time
(x10 )-3
.608
(1.93)
.524
(1.65)
- -.0909
(0.38)
-.102
(0.42)
-
Cooking time
x Child<5
(x10 )-3
- .311
(1.55)
.460
(2.19)
- .0415
(0.27)
.00858
(0.05)
Child age .0241
(0.52)
.119
(1.55)
.173
(2.18)
-.0114
(0.32)
.0012
(0.02)
.-.0121
(0.20)
Child age squ.
(x10 )-1
-.0404
(1.06)
-.105
(1.86)
-.143
(2.47)
.00461
(0.16)
.00395
(0.09)
.00633
(0.14)
Child female -.0595
(1.93)
-.0594
(1.93)
-.0541
(1.72)
-.0142
(0.60)
-.0141
(0.60)
-.0303
(1.28)
Mother’s
education
.0149
(0.59)
.0152
(0.61)
- .0319
(1.67)
.0320
(1.67)
-
Number of
children
1365 1365 1365 1365 1365 1365
Number of
mothers
889 889 889 889 889 889
India REDS 1999
Sample of 7,474 rural households in 17 states of India
7,015 households with at least one ever-married woman aged 15-60 (9,208 with children)
Advantages:
1. Variation across households in fuel use and stove type (5 categories)
2. Information on venting - house materials, windows, chimneys
3. Information on health symptoms for all children residing in each household
4. Detailed daily (by ½ hour) time use data for all women aged 15-60
5. Information on principal activities of all household members
Disadvantages:
1. No information on health symptoms for adults
2. Categories of time use lump together cooking, cleaning and child care
Thus, analysis of effects of maternal smoke exposure on her youngest children’s health as
mediated by fuel, stove type, ventilation
Figure 1Rural India, 1999: Household Distribution of Stoves
0
5
10
15
20
25
30
35
40
45
50
TraditionalChulah
OtherTraditional
Stove
Improved(Smokeless)
Chulah
Gas/KeroseneStove
Biogas Stove
Figure 2Rural India, 1999:
Mean Per-Household Expenditures (Including Imputed Self-Collected) on Fuel (Rupees)
0
100
200
300
400
500
600
700
800
900
Firewood Dung Charcoal Soft Coke Kerosene Gas
Figure 3Rural India, 1999:
Percentage of Children with Respiratory Illness in the Past Year, by Age and Stove Type
0
5
10
15
20
25
30
Age 2-4 Age 5-9
Traditional StoveSmokeless Stove
0
10
20
30
40
50
60
70
80
Chimney Kitchen Window
Traditional StoveSmokeless Stove
Figure 4Rural India, 1999:
Percentage of Households with Chimneys and Kitchen Windows, by Stove Type
Table 1
Conditional Logit Estimates: Effects of Maternal Household Work Specialization,
Cooking Fuel and Venting on the Incidence of Respiratory Symptoms for her Children 2-9
Variable/estimation
procedure
FE-
Household
FE-
Household
FE-Mother FE-Mother FE-Mother
Mother at home .444
(1.18)
-.261
(0.34)
- - -
Mother at home x traditional
biomass stove
- .947
(1.07)
- - -
Mother at home x child aged
2-4
-.0446
(0.24)
-.289
(1.36)
-.279
(1.24)
-.279
(1.24)
-.279
(1.24)
Mother at home x child aged
2-4 x biomass stove
- .435
(2.13)
.423
(1.95)
.549
(2.36)
.509
(1.79)
Mother at home x child aged
2-4 x biomass stove x
chimney
- - - -.485
(1.38)
-.437
(1.15)
Mother at home x child aged
2-4 x biomass stove x
kitchen window
- - - - -.0411
(0.12)
Mother at home x child aged
2-4 x biomass stove x thatch
roof
- - - - .163
(0.48)
Child aged 2-4 -.0906
(0.46)
-.0796
(0.40)
-.0859
(0.42)
-.0886
(0.43)
-.0868
(0.42)
Age of child -.00727
(2.57)
-.00727
(2.56)
-.00805
(2.72)
-.00812
(2.74)
-.00808
(2.72)
Age of mother .357
(2.40)
.369
(2.53)
- - -
Age of mother squared -.00530
(2.33)
-.00548
(2.44)
- - -
Number of mothers 4468 4468 2431 2431 2431
Number of observations 2503 2503 2070 2070 2070
Absolute value of robust t-statistics in parentheses.
Table 2
Conditional Logit Estimates: Effects of Maternal Household Work Specialization,
Cooking Fuel and Venting on the Incidence of Intestinal Symptoms for her Children 2-9
Variable/estimation
procedure
FE-
Household
FE-
Household
FE-Mother FE-Mother FE-Mother
Mother at home -.593
(0.69)
12.6
(11.2)
- - -
Mother at home x traditional
biomass stove
- -13.8
(9.86)
- - -
Mother at home x child aged
2-4
-.462
(1.11)
-.567
(1.16)
-.263
(0.50)
-.263
(0.50)
-.262
(0.50)
Mother at home x child aged
2-4 x biomass stove
- .154
(0.36)
-.00616
(0.01)
-.0270
(0.06)
-.152
(0.24)
Mother at home x child aged
2-4 x biomass stove x
chimney
- - - .221
(0.23)
-.102
(0.10)
Mother at home x child aged
2-4 x biomass stove x
kitchen window
.446
(0.61)
Mother at home x child aged
2-4 x biomass stove x thatch
roof
-.0216
(0.03)
Child aged 2-4 .346
(0.75)
.360
(0.78)
.398
(0.84)
.400
(0.85)
.389
(0.82)
Age of child -.0136
(2.05)
-.0136
(2.05)
-.00809
(1.18)
-.00804
(1.17)
-.00831
(1.20)
Age of mother .237
(1.18)
.247
(1.21)
- - -
Age of mother squared -.00326
(1.08)
-.00333
(1.07)
- - -
Number of mothers 4468 4468 2431 2431 2431
Number of observations 596 596 454 454 454
Absolute value of robust t-statistics in parentheses.
Are there longer-term health effects of childhood smoke exposure?
1. Can construct for each mother a history of her status in the household based on the retrospective
(a) marriage histories of all immediate family members of the head and (b) household division
histories in the 1999 India data.
2. Given 1, know the status of any child’s mother at any age of the child.
3. With a predicting equation based on the mother’s status, can predict at each age of the child
whether the mother is the principal homemaker in the household.
4. Issues:
1. Is the relationship between maternal status and work assignments stable over time?
A. Use 1999 and 1982 rounds to estimate the predicting equation, test if
stable structure.
2. Predicting equation comes from the 1999 and 1982 rounds of data using all women,
while the child health equation are estimated using the 1999 round for mothers with
children aged 6-10, so need to correct standard errors of the coefficients to take this into
account.
B. Use Multiple Imputation method, with 1000 bootstrap iterations.
Table V
Within-Household Estimates of the Determinants of Specialization in
Household Care (India), by Survey Year
Variable 1999 1982 Combinedb
Relationship of woman
to heada
Mother .193
(5.10)
.146
(1.61)
.154
(4.11)
Sister-in-law .117
(3.83)
.0610
(1.60)
.105
(3.74)
Daughter-in-law .0320
(1.17)
.0235
(0.66)
.0669
(3.31)
Oldest son daughter-
in-law
-.0250
(1.53)
-.0208
(1.06)
-.0307
(2.01)
Age -.0197
(4.90)
-.0284
(4.46)
-.0139
(4.90)
Age squared .000256
(4.84)
.000394
(3.97)
.000182
(4.84)
Year of survey - - .0163
(26.8)
Number of households 6,733 3,948 10,681
Number of women 9,208 5,539 14,747
Left out categories: Wife of head and heada
Test statistic for equality of coefficients across survey years,b
restrictions on mobility (preserve integrity of network)
Social restrictions on mobility welfare-enhancing and efficient
B. for 2 $ 1 + PL + PM and free mobility, W2 > W1, jati welfare increases
At some point, social restrictions on mobility reduce jati welfare and
efficiency
Empirical question: do we see caste-based restrictions on mobility in blue-collar jatis,
and thus non convergence, due to network externalities in the labor market?
Data: 2002 Dadar, Mumbai Student Survey
1. Random sample of 4700 student records for students residing in the 29 schools in Dadar
A. Enrolled in grades 1 through 10 in fall 2001 or
B. Enrolled in grade 10 over the period, from 1982-1991.
Thus, covers enrollment decisions over the period 1982-2001
2. In-home interviews of parents of students completed February 2002
A. Information on parents, grandparents, siblings
B. Information on schools attended, scores on secondary-school-leaving exams for
students; earnings histories, schools attended and how found job for parents;
parental and sibling occupation, remittances and transfers; and sub-caste (jati). There
are 59 sub-castes represented.
3. Survey of the school principals in the 29 schools
A. Medium of instruction, class sizes, teacher qualifications, average test scores of
students, facilities.
B. English is the medium of instruction in 10 schools; Marati (local language) in 9.
Table 1Secondary Student Quality and School Quality in Dadar, by School
Language of Instruction
School type EnglishMedium
MarathiMedium Difference
Student exam results (1998-2001)
Percent passed 92.3(6.19)
52.5(24.6)
39.8(t=6.33)
Percent first class amongpassed
36.4(5.38)
24.7(13.8)
11.7(t=3.02)
Percent distinction amongpassed
25.3(12.3)
7.16(7.77)
18.2(t=4.00)
School characteristics
Student-faculty ratio 36.7(7.60)
35.8(8.96)
0.956(t=0.28)
Class size 61.9(3.69)
62.3(3.16)
-378(t=0.08)
Students per desk 2.40(0.316)
2.36(0.479)
0.039(t=0.23)
Proportion of teacherswith B.Ed.
0.725(0.221)
0.701(0.203)
0.024(t=0.28)
Proportion of teacherswith higher degree
0.0786(0.0925)
0.0971(0.147)
-0.0185(t=0.36)
Computers per student 0.0174(0.0138)
0.0176(0.0192)
-0.0002(t=0.03)
Number of schools 10 18 28
Enrollment per school 1528 1029
Percent of Men Receiving Job Referrals and Speaking English, by Occupation
Table 1
Occupational Distribution (%), by Caste: Mumbai Men
Occupation Low Castes Middle Castes High Castes
No work 2.71 2.76 0.97
Unskilled
manual
10.9 7.69 4.38
Skilled manual 16.8 13.4 10.4
Organized blue
collar
22.1 18.5 2.80
Clerical 27.3 35.5 20.7
Professional 8.25 8.5 42.9
Business 7.70 8.86 15.2
Petty trade 3.93 4.24 2.56
Farming 0.33 0.48 0.12
Number 1806 1885 821
Table 2
Occupational Distribution (%), by Caste: Mumbai Women
Occupation Low Castes Middle Castes High Castes
No work 79.7 80.5 49.1
Unskilled
manual
6.06 3.24 1.18
Skilled manual 1.81 1.60 3.17
Organized blue
collar
0.90 1.03 0.35
Clerical 6.38 7.88 23.4
Professional 3.46 4.53 20.3
Business 0.90 0.51 1.88
Petty trade 0.80 0.72 0.59
Farming 0 0 0
Number 1881 1942 851
Table 3
Determinants of the Choice of English-Medium Schooling, by Gender
Sample Boys Girls All
Variable/estimation
procedure OLS
FE-
occup. OLS
FE-
occup. FE-caste
Jati-level job
assistance
-.378
(2.55)
-.334
(2.21)
.116
(0.69)
.169
(1.00)
-
Jati-level job
assistance x boy
- - - - -.404
(5.59)
Age (cohort) -.0090
(4.51)
-.0112
(6.83)
-.0099
(5.17)
-.012
(5.34)
-.00992
(6.64)
English medium
schooling - father
.234
(7.13)
.208
(5.33)
.309
(12.0)
.285
(10.1)
.246
(11.9)
English medium
schooling - mother
.211
(7.38)
.175
(6.01)
.263
(5.98)
.240
(6.60)
.232
(7.52)
Years of schooling
- father
.0222
(5.63)
.0193
(5.33)
.0199
(6.64)
.0158
(4.85)
.0209
(8.85)
Years of schooling
- mother
.0242
(7.21)
.0193
(6.38)
.0262
(8.75)
.0222
(6.84)
.0244
(9.96)
Father’s income
(x10-5)
.566
(1.21)
.271
(0.84)
.818
(2.78)
.601
(3.16)
.557
(1.76)
Boy - - - - .253
(6.13)
N 2240 2240 2046 2046 4286
Table 4
Determinants of the Choice of English-Medium Schooling for Boys, by
Time Period
1992-2001 1982-1991 1982-2001
Jati-level job assistance -.394
(2.19)
-348
(1.63)
-
Jati-level job assistance
- 1995-2000
- - -.438
(2.22)
Jati-level job assistance
- 1990-1994
- - -.439
(2.32)
Jati-level job assistance
- 1980-1989
- - -.315
(1.50)
Age (cohort) - 1995-
2000
- - -.00065
(0.05)
Age (cohort) - 1990-
1994
- - -.0177
(2.56)
Age (cohort) - 1980-
1989
- - -.00317
(1.14)
1995-2000 - - .121
(0.62)
1990-1994 - - .27`
(1.67)
Age (cohort) -.0160
(4.48)
-.00294
(1.04)
-
N 1209 1031 2240
Table 5
Change in School Selectivity by Caste-type, Post-1990 Period:
Student’s Father’s Schooling
Sample Boys in Marathi-Medium
School
Boys in English-Medium
School
Variable/Estim
ation procedure
OLS FE-Caste OLS FE-Caste
Age (cohort) .708
(3.29)
.548
(3.51)
-.331
(2.10)
-.392
(2.49)
Age x caste-
level job
assistance
-1.43
(3.88)
-1.14
(4.53)
.706
(2.27)
.806
(2.54)
Caste-level job
assistance
6.54
(1.58)
- -15.2
(4.17)
-
Constant 7.63
(3.06)
- 20.4
(10.9)
-
If They Build Schools, Will They Come? The Indonesia INPRES Program
Mobilization of oil revenues to finance a school building program in 1973
1. Between 1973-74 and 1978-79, 61,807 new schools built - fastest primary school
building program in the history of the world
2. Recruited and trained new teachers, but number grew at a slower pace (school quality?)
3. Suppressed primary school fees (so not just a school building program)
4. Schools built where non-enrollments rates high - placement was non-random
What was the impact - on schooling attainment and earnings?
Approaches:
A. Compare schooling attainment for those in areas with intense school construction
program and less intense program when they started school (aged 2 through 6 in
1974)
But, school construction placed in low demand areas - could get
negative effect!
B. Compare difference between children exposed differentially to the new schools
with older children from the same areas - difference in difference. Cross-area
differences for older children not due to the INPRES.
Based on Table 1, Duflo (2001)
Table 3 results from Duflo (2001):
Comparing across areas with low and high building intensity:
1. Among children exposed to the program, educational attainment highest in low-
exposure areas (program placement bias)
2. Among children not exposed from same areas (born before the program was in
place) educational attainment highest in low construction areas
3. Difference between the differences is positive
A. School attainment went up by .12 years = .13 years from adding one school
per 1,000 children (not stat. significant)
B. Wages up by 2.6% = 2.9% years from adding one school per 1,000
children (not stat. significant)
C. Implied return to schooling is B/A = 2.9%/.13 = 22%! (but not stat.
significant)
Regression analysis exploits variation in building intensity (not just categories), so more
precise results
Regression estimates:
1. Adding one school per 1,000 children increases schooling attainment by .12 -.19 years
(stat. significant) based on whole sample
2. Adding one school per 1,000 children increases wages by1.5-2.7%
But, only 45% of earners work for wages!
3. Including the self-employed (imputing incomes based on occupations and earnings
from another source), the return on schooling is 3.5% (compared with 6.8-10.6% for
wage earners)
Demand-side Intervention: The Mexican Progresa Program
Alternative direct (non-growth) poverty-reducing programs:
A. Means-tested transfer program: provide income grants to
poor people (condition on income)
1. Creates disincentive to work or upgrade skill
2. Effect on school enrollment of children -
income effect
3. Non targeted with respect to schooling - single
men included
B. Means-tested transfer program to women with children
(condition on income and fertility)
1. Same as 1 above
2. Same as 2 above
3. Encourage larger families, reduce schooling -
subsidy to numbers of children, not “quality”
C. Price supports for agricultural commodities
1. No work disincentives
2. Raises income and wages of child workers
3. Benefits largest farmers most
4. Hurts net consumers
D. Progresa Program: promised transfers over three years
that condition on pre-program income only and children
enrolled and attending school
1. Avoids disincentive work on work - post-program or
current income does not affect income transfer
2. Creates subsidy to child schooling
Progresa design and administration
1. Identify set of poor rural communities also unlikely to
benefit from NAFTA - 495 identified
2. Based on 1997 census, identify poor households in poor
communities - 2/3 of households “poor”
3. Randomly phase-in: randomly select 314 of 495 to receive
the program for first two years, then remainder receive
program in third year (181 “controls”)
4. Grants
A. Eligible households with children enrolled in grades
3-9 with 85% or better attendance record get three-year
grant
B. Amount of grant
1. How does grant compare to school costs?
2. What are school costs?
A. Tuition (=0), books, travel: direct costs
B. Foregone earnings: opportunity costs
39
Table 1
Monthly Payments for Progresa Program Eligible Familiesfor Children who attend at least 85 Percent of Daysa
Educational Levels of StudentsEligible for Payments July - December 1998b
Primary School - both sexes3rd Year4th Year5th Year6th Year
7080105135
Secondary School 1st Year Males
Females2nd Year Males
Females3rd Year Males
Females
200210210235225255
Source: Progresa Staff
a Excluding those days for which medical or parent excuses were obtained,accumulated over the last two months.
b Corresponds to school year first-term, September to December, 1998.
60
Table A-2
All Children in October 1997 Household Censusof All 500 Progresa Evaluation Villages
AgeProportion (Samples Size)
In Paid Labor ForceAverage Monthly Wage
Pesos (20 Days)
Female Male Female Male
8 .003 (1751) .006 (1888) 178 353
9 .004 (1686) .007 (1699) 99 350
10 .008 (1802) .014 (1920) 184 373
11 .007 (1782) .021 (1745) 607 346
12 .022 (1710) .053 (1898) 387 420
13 .040 (1674) .098 (1737) 467 413
14 .066 (1612) .187 (1721) 538 482
15 .115 (1604) .305 (1706) 584 593
16 .151 (1518 .438 (1564) 637 599
Thus, grants equal about 2/3 of child wage and 44% of male adult
agricultural wage (635 pesos per months)
Evaluation: difference in differences (Schultz, 2001)
1tS = eligible household in Progresa village: gets
transfer
2tS = eligible household in non-Progresa village
(control): no transfer
t = program in place, t-1 = pre-program year
1. Compare difference in outcomes:
t 1t 2tD1 = S - S
But assumes program placement is random (supposed to
be in this case!)
2. Compare differences over time in both places:
t 1t 2t 1t-1 2t-1DD1 = (S - S ) - (S - S )
Did the difference in schooling change in treatment and
control villages among the eligible households after the
program was in place?
Results
1. Impact - how much did schooling increase? Other effects?
2. What was the rate of return on the program?
41
Source: Estimated by the author based on the two pre-program rounds of the survey for only children who are matched in all five rounds or the PanelSample.
Table 3Differences Between Enrollment Rates Between Progresa and Non-Progresa Poor Children and Over Time.
(Significance Levels in Parentheses Beneath Differences)b
Year ofSchooling
Completed inPrevious Year
Pre-Program Difference of PoorProgresa - Non-Progresa
D1
Post-Program Difference of PoorProgresa - Non-Progresa
D1
Post-Preprogram Difference in Differences
DD1
All Female Male All Female Male All Female Male
0.009
(.351).010
(.433).007
(.615)-.002(.854)
-.010(.564)
.006(.742)
-.011(.482)
-.021(.353)
-.001(.969)
1.001
(.410)-.009(.816)
.010(.376)
.022(.008)
.007(.418)
.036(.002)
.020(.136)
.016(.652)
.025(.070)
2-.004(.276)
-.013(.386)
.006(.506)
.020(.009)
.018(.796)
.021(.001)
.023(.226)
.031(.693)
.015(.030)
3.015
(.278).025
(.162).005
(.882).032
(.008).013
(.679).049
(.001).017
(.219)-.012(.508)
.044(.014)
4.008(.500
-.016(.836)
.030(.266)
.041(.001)
.038(.261)
.044(.001)
.033(.053)
.055(.335)
.013(.064)
5.015
(.129).005
(.544).025
(.125).047
(.001).055
(.232).041
(.000).032
(.146).050
(.647).017
(.077)
6.024
(.345).048
(.433)-.019(.002)
.111(.002)
.148(.001)
.065(.317)
.087(.004)
.100(.070)
.085(.005)
7-.012(.894)
-.005(.854)
-.015(.958)
.013(.147)
.025(.533)
.003(.006)
.025(.378)
.030(.583)
.018(.062)
8-.030(.913)
-.051(.932)
-.016(.836)
.001(.162)
.015(.575)
-.010(.100)
.031(.347)
.066(.687)
.006(.235)
9 or More
.103(.534)
.327(.001)
-.156(.006)
.066(.317)
.111(.042)
.026(.813)
-.037(.914)
-.216(.044)
.182(.020)
Notes: a For definition of D1 and DD1, see Figures 1 and 2 and text
46
Table 7
Cumulative Expected Enrollment Years for Birth Cohort of Poor Children who Enroll and Complete Grade 1
Grade Completed
PreprogramRounds 1 and 2
Post-Program Rounds 3, 4, and 5
Difference inDifferences
ProgresaNon-
Progresa ProgresaNon-
Progresa DI DDI
1 .977 .975 .975 .953 .022 .020
2 .936 .938 .939 .899 .040 .042
3 .896 .884 .904 .837 .067 .041
4 .856 .838 .866 .768 .098 .080
5 .816 .786 .825 .695 .130 .100
6 .464 .428 .511 .352 .159 .121
7 .436 .407 .484 .330 .154 .125
8 .414 .399 .450 .306 .144 .129
Expected TotalYears Enrolledfor Both Sexes
6.80 6.66 6.95 6.14 .81 .66
Years EnrolledFemales 6.66 6.62 6.95 6.19 .76 .72
Years EnrolledMales 6.93 6.72 6.96 6.11 .85 .64
1. Impact: elasticity of enrollment to school cost = -.2
50% reduction in opportunity costs
10% increase in schooling attainment (.66/6.8)
2. Rate of return calculation = 8%
Equate discounted Progresa grant costs to discounted stream of additional earnings over
lifetime of child - but what are additional earnings?
Assumptions:
1. All children end schooling at age 16
2. All children migrate to urban areas
Too few wage earners in rural areas - self-employed again!
3. Urban wages 12% higher for each additional grade completed (Census estimates)
4. Rural migrants in urban areas earn 20% less than natives (Census estimates)
5. Children start work at 18 and work until they are 65
The Progresa program evaluation ran for a short period - can we do
better in anticipating its long run effects?
But in the longer run, fertility might change - and too little time to
estimate impact of the actual program.
What about alternative programs - creating prizes for schooling
accomplishments? Would these have been more effective?
How can we know from this one, short-term experiment?
Estimate a structural model - obtain estimates of fundamental parameters
(preferences, technology, constraints) and use them to carry out policy
experiments of any type.
But how do we know the model is a good one - what is validation?
Todd-Wolpin study:
1. Estimate dynamic, structural model using pre-program data
(baseline plus control).
2. Assess model by comparing it predictions for Progresa effects
on schooling with actual effects from the randomized evaluation
— A valid structural model can be used to eval-
uate counterfactual policies.
∗ consider variations in program parameters:amounts paid, program structure, changes
outside scope of program
∗ consider impacts of long-run programs. Al-most always, social experiments are short-lived. If they seem to work, they have to be
extended to the control group. If they seemnot to work, they are abandoned. You can’t
usually use randomized evaluations of longterm programs (n.b., this does not mean
you can’t use these methods to look at long-
run impacts).
— the model makes it possible to consider mech-anisms through which the impact occurs. e.g.,
to interpret the externalities in the worms pa-per, Miguel and Kremer had to have a model.
• Why not?
— Answers only sensible if the model is right.
— pre-testing bias: there are approximately an
infinite number of auxilary assumptions that
need to be made on the shapes of preferences,
technologies, distributions of unobserved ran-
dom variables. A clever researcher can choose
these to match the data quite nicely.
• This is why Todd and Wolpin is interesting. Theyestimate the structural model using the control
group and pre-intervention treatment group, then
compare predictions with the information from
the randomized evaluation. (Some danger here).
The Todd-Wolpin Model
The problem: before the intervention, and for the con-
trol group, there is no variation in cost of schooling.
How to identify impact of subsidy?
Use variation in wages. Suppose for example:
u = C + (α+ ε)s
s = 1 if kid in school.
C = y + w(1− s)
ε˜N(0, σ2). Hence s = 1 iff
y + α+ ε > y +w
ε > w − α
So
Pr(s = 1) = 1−Φ(w − α
σ)
where Φ is the standard normal density. So as long
as we observe all the wage offers and the schooling
choices, we can identify α and σ. Then we also know
the impact of a subsidy on schooling choices.
Now the real model:
• Discrete time, chose to have a kid next period,choose school for kids 6− 15, choose work (kids> 12) (alternative, sit at home).
• parents income is given exogenously. no saving,
no borrowing.
• p(t) indicator of becoming pregnant→ n(t+1) =
1. (they differentiate between b(t+ 1) and g(t+
1).) Stock of kids is N(t+1) = N(t)+n(t+1).
A child born to a women when she is τ years old
is t − τ in t. s(t, τ) = 1 if that kid is in school
at t.
• c(t, τ) = 1 if a year of school is completed at t.
So the stock of schooling of a kid born at τ is
S(t, τ) = S(t − 1, τ) + c(t − 1, τ). Probability
of completing a year of schooling once started is
πc(t− 1, τ , S(t− 1)|s(t− 1) = 1, µc) where µc isa family fixed effect.
• h(t, τ) = 1 if the kid works
• There is a specific utility function given in theappendix. The general idea is: