Selection, Gender and the Impact of Schooling Type in the Dhaka - - PowerPoint PPT Presentation

Selection, Gender and the Impact of Schooling Type in the Dhaka Slums

John C. Ham 1 Saima Khan 2

1New York University, Abu Dhabi 2National University of Singapore

May 30, 2018

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 1 / 58

Overview

Overview I: What we do

In joint work with Saima Khan, between 2015-2016, we collected

unique data set on 4 - 12 year olds in two Dhaka slums.

◮ First such study of education for Urban Bangladesh. ◮ First to study JAAGO - new school-type with a framework previously available to elites only. ◮ First to examine relative impact of JAAGO, Govt. and NGO schools in Dhaka slums.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 2 / 58

Overview

Overview I: What we do (contd)

Allow for selection across 3 different school-types (JAAGO,

• Govt. and NGO schools), based on detailed background

information and 2 measures of fluid intelligence (IQ).

◮ Controlling for IQ is especially important for Treatment Effects; Not usually done.

Estimate the impact of the different school-types on student

achievement, by gender, after controlling for these variables.

◮ Focus on gender since the literature documents substantial gender difference.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 3 / 58

Overview

Overview II: Results

Evidence of Selection

◮ On average, JAAGO and Govt. schools get ‘better’ students than NGO schools in their student body, in terms of family expenditure, parents’ schooling, & child’s fluid intelligence.

Controlling for Selection

◮ Only controlling for family background (as done in most studies) does not eliminate the selection problem; ◮ Including measures of fluid intelligence reduces impact of school-type by 20 − 50 %.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 4 / 58

Overview

Overview II: Results (contd)

Strong evidence of Gender Heterogeneity across school-types.

◮ JAAGO vs. Govt: Girls are better off at JAAGO, but boys perform equally well at both school-types. ◮ Govt. vs. NGO: Boys are better off at Govt. schools, but girls perform equally well at both school-types. ◮ JAAGO vs. NGO: Both genders better off at JAAGO.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 5 / 58

Overview

Overview II: Results (contd)

Evidence of gender disparity within each school-type.

◮ JAAGO schools: girls and boys perform equally well. ◮ NGO schools: girls and boys perform equally well. ◮ Govt schools: boys substantially outperform girls.

Helps to explain the gender difference in aggregate data.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 6 / 58

Overview

Presentation Outline

1

Motivation

2

Context, Research Question and Data Collection

3

Evidence of Selection

4

Methodology

5

Results

6

Conclusion

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 7 / 58

Motivation

Motivation

In the context of Bangladesh:

• 1. Little evidence about impact of school-type on achievement and

gender differential (especially for urban Bangladesh).

• 2. High enrolment but poor learning outcomes.

◮ Students who have completed Grade 5: 25% & 33% acquired grade 5 level competency in Bengali & Maths -based

• n a 2011 nationwide survey of Government Primary School

students [World Bank, 2013]

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 8 / 58

Motivation

Motivation (contd)

• 3. Gender parity in primary school enrolment, but not achievement.

[World Bank (2013), ADB Country Gender Assessment Bangladesh (2010),

Amin and Chandrasekar (2009)].

◮ Girls lag behind boys in terms of learning outcomes at both primary and secondary level. ◮ We find similar pattern even after controlling for family background and fluid intelligence in government schools.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 9 / 58

Motivation

Motivation (contd)

In a wider context:

Similar low student achievement and wide gender gap in

Pakistan [Das, Pandey and Zajonc (2012), The Economist (Jan 4, 2018)]

Pakistan has started to undertake reforms that mainly include:

◮ privatization/ outsourcing to charities; ◮ school voucher programs; ◮ addressing the problem of ”ghost” teachers.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 10 / 58

Motivation

Motivation (contd.)

While there is some overlap with the JAAGO framework, to the

best of our knowledge, Pakistan has not tried out the JAAGO school-type.

Given this situation in both Bangladesh and Pakistan (combined

population of 356 million), it is important to consider alternative schooling models like JAAGO. Most field experiments consider

• nly one deviation from current system. JAAGO is a dramatic

break from the existing system.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 11 / 58

Motivation

Motivation (contd)

Given our results, introducing JAAGO helps to equalize gender

• utcomes.

◮ Within school-type comparison (after controlling from X’s):

Boys and girls do equally well (badly) in NGO schools; But boys outperform girls at Govt schools.

◮ On an average, boys do better because some of them go to government schools. ◮ However, at JAAGO girls do better than if they attended government or NGO schools. ◮ Thus, JAAGO may help attain gender parity in terms of both enrolment and achievement.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 12 / 58

Context, Research Question and Data Collection

Presentation Outline

1

Motivation

2

Context, Research Question and Data Collection

3

Evidence of Selection

4

Methodology

5

Results

6

Conclusion

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 13 / 58

Context, Research Question and Data Collection

School-type Characteristics

We examine the impact of 3 types of schools:

• 1. JAAGO schools;
• 2. Government schools;
• 3. NGO (including charity schools).

No madrassahs (can be very different from above 3

school-types).

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 14 / 58

Context, Research Question and Data Collection

School-type Characteristics

What does JAAGO do?

◮ Its medium of instruction - English (comparable to English medium private schools available to the country’s elites). ◮ Addresses problems of ”ghost teachers” - strict administrative monitoring. ◮ Hires teachers based on merit. ◮ Has higher share of female teachers. ◮ Has longer school days and a longer school year. ◮ Does not use rote learning. ◮ Does not use corporal punishment.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 15 / 58

Context, Research Question and Data Collection

Characteristics Across the 3 School-types

Characteristics JAAGO School Govt. School NGO School

Instruction in English

• ×

× Minimum teacher qualification - Bachelors Degree

• ×

× Teachers require strong command over English

• ×

× High level in-service training ×

• ×

High share of female teachers

• ×
• High teacher absenteeism (low teacher effort)

×

• NA

High headmaster absenteeism (low monitoring) ×

• NA

High teacher salary ×

• ×

Small class size

• ×
• Longer school days
• ×

× Longer school year

• ×

× Corporal punishment ×

• ×

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 16 / 58

Context, Research Question and Data Collection

Research Questions

• 1. What type of students are being drawn to, and accepted, by

each school-type?

◮ Is there selection across school-types for boys and girls?

• 2. What is the impact of school-type on test scores by gender, after

controlling for selection?

(i). JAAGO vs. Government (ii). JAAGO vs. NGO (iii). Government vs. NGO

• 3. Within each school-type, how do boys and girls perform after

controlling for selection? We will use propensity score matching, adjusted for choice based sampling, to control for selection.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 17 / 58

Context, Research Question and Data Collection

Data Collection

Between 2015-2016, collected our own stratified (by school

type) data on 1816 slum children (aged 4 - 12) attending the 3 types of schools.

◮ JAAGO schools - 576 children; ◮ Government schools - 598 children; ◮ NGO schools - 642 children.

Used choice based sampling to ensure sufficient number of

JAAGO students show up in the sample (common for sampling

• f rare events).We collect the data by sub-neighbourhoods. We

start with a street with a JAAGO student, then collect other

• students. We adjust the standard errors for this cluster sampling.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 18 / 58

Context, Research Question and Data Collection

Data Collection (contd)

Took many steps, including 100% audio auditing, to insure data

quality.

We are now collecting Wave 2 so we can do d-in-d matching.

◮ Hope to do several more waves to create a panel dataset

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 19 / 58

Evidence of Selection

Presentation Outline

1

Motivation

2

Context, Research Question and Data Collection

3

Evidence of Selection

4

Methodology

5

Results

6

Conclusion

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 20 / 58

Evidence of Selection

School Type and Selection: Sorting?

Investigate selection across school-types in terms of 5 key

variables :

• 1. Monthly Family Expenditure (deflated by equivalence scale);
• 2. Father’s Schooling;
• 3. Mother’s Schooling;
• 4. K-BIT (IQ/Fluid Intelligence);
• 5. Raven’s Coloured Progressive Matrices (IQ/Fluid Intelligence).

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 21 / 58

Evidence of Selection

Selection: Means Across School-types for Boys

Table 1: Means Across School Types (Boys)

JAAGO Govt NGO Monthly Family Expdt 5822.8 6306.8 5450.7 (in BDT adjusted by equivalence scale) [117.9] [102.6] [94.52] Father’s schooling 3.933 4.045 3.084 [0.227] [0.223] [0.204] Mother’s schooling 3.685 3.377 2.633 [0.209] [0.186] [0.172] K-BIT (IQ) 0.348 0.0524

• 0.317

[0.0620] [0.0513] [0.0628] Raven’s CPM (IQ) 0.241 0.215

• 0.249

[0.0688] [0.0598] [0.0568] Observations 262 286 285

Notes: (a) Standard errors in parentheses; (b) For both the IQ scores, we report their respective age adjusted Z-scores. Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 22 / 58

Evidence of Selection

Selection: Means Across School-types for Girls

Table 2: Means Across School Types (Girls)

JAAGO Govt NGO Monthly Family Expdt 5863.2 6060.3 5303.3 (in BDT adjusted by equivalence scale) [100.1] [96.97] [78.10] Father’s schooling 3.448 3.605 2.813 [0.205] [0.202] [0.176] Mother’s schooling 3.889 3.288 2.506 [0.180] [0.182] [0.148] K-BIT (IQ) 0.195 0.0897

• 0.283

[0.0569] [0.0568] [0.0468] Raven’s CPM (IQ) 0.0618 0.0316

• 0.231

[0.0566] [0.0569] [0.0435] Observations 314 312 357

Notes: (a) Standard errors in parentheses; (b) For both the IQ scores, we report their respective age adjusted Z-scores. Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 23 / 58

Evidence of Selection

Selection: Mean Differences Across School-types for Boys

Table 3: Mean Differences Across School Types (Boys)

J vs. G J vs. N G vs. N Monthly Family Expdt

• 484.0∗∗∗

372.1∗∗ 856.1∗∗∗ (in BDT adjusted by equivalence scale) [155.6] [150.0] [139.5] Father’s schooling

• 0.112

0.849∗∗∗ 0.961∗∗∗ [0.319] [0.305] [0.302] Mother’s schooling 0.308 1.052∗∗∗ 0.744∗∗∗ [0.279] [0.269] [0.254] K-BIT (IQ) 0.295∗∗∗ 0.665∗∗∗ 0.370∗∗∗ [0.0799] [0.0885] [0.0811] Raven’s CPM (IQ) 0.0256 0.489∗∗∗ 0.464∗∗∗ [0.0908] [0.0887] [0.0825] Observations (no. of boys in the 548 547 571 two school-types being compared)

Notes: (a) Standard errors in parentheses; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) For both the IQ scores, we report their respective age adjusted Z-scores. Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 24 / 58

Evidence of Selection

Selection: Mean Differences Across School-types for Girls

Table 4: Mean Differences Across School Types (Girls)

J vs. G J vs. N G vs. N Monthly Family Expdt

• 197.1

560.0∗∗∗ 757.0∗∗∗ (in BDT adjusted by equivalence scale) [139.4] [125.5] [123.3] Father’s schooling

• 0.157

0.635∗∗ 0.792∗∗∗ [0.288] [0.269] [0.267] Mother’s schooling 0.601∗∗ 1.382∗∗∗ 0.782∗∗∗ [0.256] [0.231] [0.232] K-BIT (IQ) 0.105 0.477∗∗∗ 0.373∗∗∗ [0.0804] [0.0731] [0.0730] Raven’s CPM (IQ) 0.0302 0.293∗∗∗ 0.263∗∗∗ [0.0803] [0.0705] [0.0707] Observations (no. of boys in the 626 671 669 two school-types being compared)

Notes: (a) Standard errors in parentheses; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) For both the IQ scores, we report their respective age adjusted Z-scores. Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 25 / 58

Evidence of Selection

Evidence of Selection Across School-Types

J vs. G

◮ Boys at Govt schools are significantly wealthier. ◮ Girls with better educated mothers going to JAAGO schools. ◮ Boys with higher fluid intelligence (K-BIT) go to JAAGO schools. ◮ Raven’s fails to pick up any difference between JAAGO and Govt students for both genders.

J vs. N and G vs. N (both genders)

◮ NGO boys and girls are “weaker” in terms of family wealth, father’s schooling, mother’s schooling and fluid intelligence.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 26 / 58

Methodology

Presentation Outline

1

Motivation

2

Context, Research Question and Data Collection

3

Evidence of Selection

4

Methodology

5

Results

6

Conclusion

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 27 / 58

Methodology

Methodology: Least Squares Estimation

School-type is endogenous - children with different abilities &

family background sorting into different school-types. Achi = c + γ Malei + α1 DJi + α2 DNi

+β1 [DJi × Male] + β2 [DNi × Male] + ǫi

◮ where:

Achi: child’s z-score in the Woodcock Johnson Test; Govt (female) is the reference group; DJi × Male : JAAGO × Male ; DNi × Male : NGO × Male.

Due to selection, we are worried that:

◮ cov (DJi, ǫi) > 0 ◮ cov (DNi, ǫi) < 0

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 28 / 58

Methodology

Methodology: Least Squares Estimation (contd.)

One way to control for this selection - run OLS with family

background and fluid intelligence. Achi = c + γ Malei + α1 DJi + α2 DNi + π1 Xi

+β1 [DJi × Male] + β2 [DNi × Male] + π2 [ Xi × Male] + νi

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 29 / 58

Methodology

What is Matching (without Choice Based Sampling)?

Due to this selection - to find treatment effects, cannot simply

compare mean outcomes of students from two school-types (say J vs. N);

◮ Since JAAGO has smarter boys and girls with richer and more educated families.

However there may be observable X’s such that once we

condition on such X’s who goes to JAAGO and NGO school is like a coin toss.

◮ This is called the Conditional Independence Assumption or CIA.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 30 / 58

Methodology

What is Matching (without Choice Based Sampling)? (contd)

Condition on all X’s that affect both school sorting and

achievement: child and family characteristics, and in particular, child fluid intelligence (IQ). One difference with other work - conditioning variables are not collected pre-treatment.

By Rosenbaum and Rubin (1985) conditioning on P(X) is

equivalent to conditioning on X.

◮ where P(X) - the propensity score - is defined as probability of going to school i. ◮ P(X) usually estimated by using Logit/Probit.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 31 / 58

Methodology

What is Matching (without Choice Based Sampling)? (contd)

Let’s take the example of JAAGO vs. NGO student We know achievement (Yi) of a JAAGO student attending

JAAGO school, but we do not know how the same student would perform if he were to go to an NGO school (the counterfactual).

Assuming that the CIA holds, for each JAAGO student, use

students of NGO with similar propensity score to estimate how JAAGO students would have performed had they attended NGO

• schools. This gives us an Average Treatment Effect (ATT).

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 32 / 58

Methodology

What is Matching? (contd)

Of course, we can always reverse the process and ask what

would have happened to an NGO student if they went to

• JAAGO. This gives us an alternative Average Treatment Effect

that we call the ATU.

In general ATU = ATT because NGO students differ from

JAAGO students in observable and unobservable ways.

We can average the ATT and ATU to get ATE, the effect of

going to JAAGO for a randomly chosen individual.

Consistent estimation of the regression model forces

ATU = ATT = ATE.

This is testable within the Propensity Score Matching Model if

use the bootstrap to get Cov(ATU, ATT).

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 33 / 58

Methodology

Common Support Condition

If a treatment has ˆ

P(X) = 0.95, then we can only find comparison for this child if we can find members of the comparison group with a similar ˆ P(X). If not, we must delete that treatment since this situation violates the common support condition.

One Issue: because we have choice based sampling, we match on

LOR = log ˆ P(X)/

• 1 − ˆ

P(X)

• But 0 < LOR < ∞ while 0 < ˆ

P(X) < 1, so it is hard to impose a constant support condition with regard to the LOR.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 34 / 58

Methodology

What is Matching? (contd)

How to use these X′s we have?

◮ Simplest example is“Nearest Neighbour Matching”. ◮ For JAAGO student, i, find the closest estimated propensity score from the Control group.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 35 / 58

Methodology

What is Matching? (contd)

Note that for JAAGO student i, there are a number of additional

comparisons in NGO group.

◮ Nearest neighbour matching ignores these other comparisons.

We use local linear matching to utilize more data than nearest

neighbour matching.

◮ Pick a bandwidth, h. ◮ Only use NGO students whose ˆ pj fall in the range ˆ pi ± h. ◮ Place weight on ˆ pj based on how close it is to ˆ pi.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 36 / 58

Methodology

Diagnostics

Can we get a signal if matching is appropriate here, i.e. the CIA

holds - balancing tests.

Given P(X), look for a treatment effect on the Xs shouldn’t see

• ne.

Problem - getting standard errors for LLR is very time intensive,

so don’t want to use LLR to form diagnostics. Instead use Nearest Neighbour or 5 Nearest Neighbours for diagnostics since there is an analytical solution for the standard errors.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 37 / 58

Methodology

Why do Matching?

In addition to OLS - why do we do matching?

◮ Matching allows us to get better comparisons for each treated individual.

OLS is the crudest form of matching where relatively dissimilar individuals may be compared to each other. All NGO students considered to be comparisons to all JAAGO students.

◮ OLS imposes functional form restriction as opposed to

• matching. As noted with OLS, no difference between ATE and

ATU by assumption.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 38 / 58

Methodology

Why do Matching? (contd)

Another way to control for this endogeneity - Instrumental

Variable Approach.

◮ But IV estimates are inconsistent in the presence of choice based sampling [Solon et al. (2015)]. ◮ Adjusting IV estimator to make it consistent infeasible given our sample; cannot get back to the original probit/logit in our case.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 39 / 58

Methodology

Methodology:Propensity Score Matching

Do matching while accounting for choice based sampling;

◮ Standard implementation of matching does not yield consistent estimates given choice based sampling, again because we cannot get back to the original probit/logit in our case. ◮ Instead match on log odds ratio of the estimated propensity

• score. [Heckman and Todd (2009)].

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 40 / 58

Methodology

Defining Treatment Effects

Recall that we have 3 Treatment Effects for JAAGO vs. NGO:

◮ Average Treatment on the Treated (ATT); ◮ Average Treatment on the Untreated (ATU); ◮ Average Treatment Effect (ATE).

Of course, we have analogous effects for Govt vs. NGO, JAAGO

• vs. Govt.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 41 / 58

Methodology

Defining Treatment Effects

Average Treatment on the Treated (ATT)

◮ ATT captures the effect of taking a group of students going to JAAGO schools and placing them in NGO schools.

Average Treatment on the Untreated (ATU)

◮ ATU captures the effect of taking a group of students attending NGO schools and placing them in JAAGO schools.

Average Treatment Effect (ATE)

◮ ATE refers to randomly assigning a student to JAAGO schools and then randomly assigning the said student to NGO schools; ◮ Essentially, the ATE refers to doing this for the whole population of JAAGO and NGO students.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 42 / 58

Methodology

Defining Treatment Effects (contd.)

All 3 treatment effects measure the difference in outcomes

between different students attending one school versus the exact same group of students attending a different school-type as an (imaginary) counter-factual.

Essentially, the difference between the three effects only comes

from differences in the sample of students being considered.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 43 / 58

Results

Presentation Outline

1

Motivation

2

Context, Research Question and Data Collection

3

Evidence of Selection

4

Methodology

5

Results

6

Conclusion

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 44 / 58

Results

Impact of School-Type by Gender: OLS Results

Table 5: Impact of School-type by Gender (OLS Results)

Dependent Variable - Achievement Test Z-Score (1) (2) (3) (4) (5) No IQ No IQ Raven’s CPM K-BIT Both IQ J vs G (girls) 0.242∗∗∗ 0.233∗∗∗ 0.225∗∗∗ 0.194∗∗ 0.198∗∗ (0.0679) (0.0734) (0.0721) (0.0751) (0.0748) J vs. G (boys) 0.090 0.086 0.079

• 0.033
• 0.012

(0.0861) (0.0903) (0.0717) ( 0.0855) (0.0774) J vs. N (girls) 0.489 ∗∗∗ 0.431∗∗∗ 0.348∗∗∗ 0.254∗∗∗ 0.246∗∗∗ (0.0931) (0.0848) (0.0962) (0.0962) (0.0815) J vs. N (boys) 0.581∗∗∗ 0.522∗∗∗ 0.374∗∗ 0.266∗∗ 0.239∗ (0.1411) (0.1335) (0.1255) (0.1343) (0.1310) G vs N (girls) 0.247∗∗ 0.198∗∗ 0.123 0.0601 0.0478 (0.0986) (0.0917) (0.0935) (0.0907) (0.0915) G vs. N (boys) 0.490∗∗∗ 0.436∗∗∗ 0.294∗∗ 0.299∗∗ 0.251∗∗ (0.1192) (0.1173) (0.1213) (0.1222) (0.1230) Mother & Father’s schooling No Yes Yes Yes Yes Observations 1816 1816 1816 1816 1816

(a) Standard errors in parentheses clustered at street/area level; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) This table is derived from regressing achievement on family and child characteristics; school-type effects for the six comparison cases is calculated from the estimated coefficients of the OLS results; (d) All regressions include the standard controls (age, gender, father absent and family size).

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 45 / 58

Results

Impact of School-Type by Gender: Matching Results Excluding IQ Measures

Table 6: Matching Results with No IQ)

(1) (2) (3) (4) (5) (6) J vs G (F) J vs G (M) J vs. N (F) J vs. N (M) G vs. N (girls) G vs. N (boys) ATT 0.230∗∗ 0.0701 0.382∗∗∗ 0.448∗∗∗ 0.179∗ 0.335∗∗∗ (0.0961) (0.105) (0.0910) (0.123) (0.0975) (0.129) ATU 0.220∗∗∗ 0.112 0.529∗∗∗ 0.579∗∗∗ 0.220∗∗ 0.556∗∗∗ (0.0767) (0.0980) (0.0914) (0.152) (0.104) (0.139) ATE 0.225∗∗∗ 0.0921 0.460∗∗∗ 0.516∗∗∗ 0.201∗∗ 0.445∗∗∗ (0.0826) (0.0957) (0.0874) (0.133) (0.0944) (0.125)

Notes: (a) Bootstrapped standard errors in parentheses clustered at street/area level; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) Other matching covariates include child’s age, family size, father absence dummy, father’s schooling and mother’s schooling; Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 46 / 58

Results

Impact of School-Type by Gender: Matching Results with Raven’s (IQ)

Table 7: Matching Results with Raven’s (IQ))

(1) (2) (3) (4) (5) (6) J vs G (F) J vs G (M) J vs. N (F) J vs. N (M) G vs. N (girls) G vs. N (boys) ATT 0.215∗∗ 0.0699 0.312∗∗∗ 0.308∗∗∗ 0.110 0.233∗ (0.0879) (0.0905) (0.112) (0.109) (0.104) (0.123) ATU 0.211∗∗∗ 0.112 0.453∗∗∗ 0.432∗∗∗ 0.118 0.320∗∗ (0.0744) (0.0814) (0.102) (0.145) (0.102) (0.141) ATE 0.213∗∗∗ 0.0920 0.387∗∗∗ 0.372∗∗∗ 0.114 0.275∗∗ (0.0767) (0.0786) (0.102) (0.123) (0.0947) (0.117)

Notes: (a) Bootstrapped standard errors in parentheses clustered at street/area level; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) Other matching covariates include child’s age, family size, father absence dummy, father’s schooling and mother’s schooling; Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 47 / 58

Results

Impact of School-Type by Gender: Matching Results with K-BIT (IQ)

Table 8: Matching Results with K-BIT (IQ))

(1) (2) (3) (4) (5) (6) J vs G (F) J vs G (M) J vs. N (F) J vs. N (M) G vs. N (girls) G vs. N (boys) ATT 0.165∗

• 0.0101

0.164∗ 0.220

• 0.00148

0.226∗ (0.0926) (0.0984) (0.0980) (0.144) (0.101) (0.128) ATU 0.198∗∗∗

• 0.0346

0.386∗∗∗ 0.282∗∗ 0.0615 0.373∗∗∗ (0.0748) (0.0917) (0.0785) (0.140) (0.107) (0.137) ATE 0.181∗∗

• 0.0230

0.282∗∗∗ 0.252∗ 0.0321 0.297∗∗ (0.0799) (0.0903) (0.0835) (0.137) (0.0966) (0.117)

Notes: (a) Bootstrapped standard errors in parentheses clustered at street/area level; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) Other matching covariates include child’s age, family size, father absence dummy, father’s schooling and mother’s schooling; Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 48 / 58

Results

Impact of School-Type by Gender: Matching Results with Both IQ

Table 9: Matching Results with Both IQ

(1) (2) (3) (4) (5) (6) J vs G (F) J vs G (M) J vs. N (F) J vs. N (M) G vs. N (girls) G vs. N (boys) ATT 0.171∗ 0.0314 0.154 0.173

• 0.0153

0.169 (0.0911) (0.0935) (0.103) (0.132) (0.101) (0.125) ATU 0.200∗∗∗

• 0.00949

0.367∗∗∗ 0.269∗ 0.0308 0.270∗ (0.0754) (0.0858) (0.0871) (0.145) (0.109) (0.151) ATE 0.185∗∗ 0.0101 0.268∗∗∗ 0.222∗ 0.00934 0.218∗ (0.0782) (0.0818) (0.0909) (0.132) (0.0966) (0.118)

Notes: (a) Bootstrapped standard errors in parentheses clustered at street/area level; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) Other matching covariates include child’s age, family size, father absence dummy, father’s schooling and mother’s schooling; Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 49 / 58

Results

Impact of School-Type by Gender: Matching Results

Table 10: Matching Results & Controlling for Selection: KBIT vs. Raven’s CPM

(1) (2) (3) (4) Only Family Family Background Family Background Family Background ATE Background & Raven’s & K-BIT & Both IQ J vs. G (girls) 0.225∗∗∗ 0.213∗∗∗ 0.181∗∗ 0.185∗∗ (0.0826) (0.0767) (0.0799) (0.0782) J vs. G (boys) 0.092 0.092

• 0.023

0.0101 (0.0957) (0.0786) (0.0903) (0.0818) J vs. N (girls) 0.460∗∗∗ 0.387∗∗∗ 0.282∗∗∗ 0.268∗∗∗ (0.0874) (0.102) (0.0835) (0.0909) J vs. N (boys) 0.516∗∗∗ 0.372∗∗∗ 0.252∗ 0.222∗ (0.133) (0.123) (0.137) (0.132) G vs. N (girls) 0.201∗∗ 0.114 0.032 0.009 (0.0944) (0.0947) (0.0966) (0.0966) G vs. N (boys) 0.445∗∗∗ 0.275∗∗ 0.297∗∗ 0.218∗ (0.125) (0.117) (0.117) (0.118)

Notes: (a) Bootstrapped standard errors in parentheses clustered at street/area level; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) Family background matching covariates refers to child’s age, family size, father absence dummy, father’s schooling and mother’s schooling. Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 50 / 58

Results

Understanding the Gender Heterogeneity

What school-type characteristics may be driving the gender

difference in achievement between JAAGO and Govt schools?

◮ Pro-male gender bias at Govt. schools; ◮ Share of female teachers lower at Govt. schools; ◮ Corporal punishment common at Govt. schools.

Find evidence of gender differential for all above components in

the literature.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 51 / 58

Results

Impact of School-Type by Gender: Matching Results with Both IQ

Table 11: OLS with Gender Difference for Each School-Type

(1) (2) (3) (4) (5) No IQ No IQ Raven’s CPM K-BIT Both IQ M vs. F (JAAGO) 0.125∗ 0.110 0.0579 0.0510 0.0349 (0.0681) (0.0687) (0.0636) (0.0629) (0.0616) M vs. F (NGO) 0.0347 0.0218 0.0325 0.0421 0.0436 (0.0844) (0.0826) (0.0796) (0.0754) (0.0753) M vs F. (Govt) 0.277∗∗∗ 0.261∗∗∗ 0.204∗∗∗ 0.281∗∗∗ 0.246∗∗∗ (0.0808) (0.0782) (0.0735) (0.0710) (0.0700) Father’s & Mother’s schooling No Yes Yes Yes Yes

Notes: (a) Standard errors not clustered; (b) ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01; (c) Regression includes standard set of controls: gender dummy, child’s age, family size, father absence dummy, father’s schooling and mother’s schooling. Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 52 / 58

Results

Preliminary Evidence: Gender Difference with-in School-type

Without schools like JAAGO, on an average, girls are at a

disadvantage.

◮ At present, 2 school-types available to slum children - Govt. and NGO schools. ◮ At NGO schools, boys and girls perform equally well; but at

• Govt. schools, boys outperform girls.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 53 / 58

Conclusion

Presentation Outline

1

Motivation

2

Context, Research Question and Data Collection

3

Evidence of Selection

4

Methodology

5

Results

6

Conclusion

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 54 / 58

Conclusion

Conclusion

• 1. Students in urban slums of Dhaka sorted across school-types.

◮ ‘Better’ students sorted into JAAGO and government schools; ◮ ‘Weaker’ students sorted into NGO schools.

• 2. Fluid Intelligence plays a crucial part in controlling for selection.

◮ Including fluid intelligence, which most developing country studies fail to account for, substantially reduces bias.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 55 / 58

Conclusion

Conclusion (contd)

• 3. Preliminary Results of gender difference with-in school-types

sheds light on why girls do worse in aggregate data.

• 4. School-types like JAAGO should reduce the gender gap in

achievement.

◮ If all girls going to Govt. schools could switch to JAAGO, gender gap would disappear or fall substantially.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 56 / 58

Thank You

Appendix

Distributions of Schools by School Type in Our Sample

Table 12: Summary Statistics of Schools by School Type

(1) (2) (3) (4)

• No. of schools

Mean

• Std. Dev

Total (no. of students) (no. of students) Govt 16 37.37 75.11 598 JAAGO 2 288 80.61 576 NGO 31 20.71 42.11 642

(a) Note that due to unavailability of administrative data, we are unable to present distribution

• f schools per school-type in the greater population.

Ham & Khan (2018) Selection, Gender & School Type May 30, 2018 58 / 58