Science: 20 GPs in a single District

One of the quieter successes of post-Independence India is that the gender gap in education has closed. The trajectory is plain in the historical data. In 1951, fewer than three in ten girls were enrolled at the primary level, against six in ten boys; at upper primary, girls were virtually absent. By the mid-2000s the primary gap had closed. By 2021–22, the Gender Parity Index in higher education was 1.01, and at the postgraduate level women were already the majority, at more than 55% of enrollment.

How did this happen? On the candidates list is the 73rd Amendment, which in 1993 reserved one-third of seats on village councils for women. The seats rotate across election cycles, and the researchers who study them have generally treated the rotation as exogenous. That design feature, repeated across hundreds of thousands of villages over three decades, has produced what is probably the cleanest natural experiment in the developing world on what happens when a woman holds political office.

In 2012, Lori Beaman, Esther Duflo, Rohini Pande, and Petia Topalova published a paper in Science arguing that female leadership raises adolescent girls' aspirations and educational attainment. The setting was Birbhum district in rural West Bengal. The design exploited the randomized reservation of gram panchayat pradhan seats for women under the 73rd Amendment. West Bengal first reserved seats for women in 1998 and again in 2003; the surveys were conducted in 2006 and 2007, so the relevant exposure period was up to nine years. The result was striking. In villages reserved for a woman pradhan in two consecutive elections, the gender gap in parents' aspirations for their children closed by 25%, the gender gap in adolescents' own aspirations closed by 32%, and the gender gap in educational attainment was, in the paper's phrasing, erased.

What moved, and what didn't

Closing 32% of a gap depends on what the gap was. In never-reserved villages, the picture across the four items the paper averages into its aspirations index is uneven. On the housewife item, 60% of girls say they do not wish to be a housewife or have their in-laws choose their occupation, against 99.8% of boys. On marrying after 18, 66% of girls say yes, against 98.0% of boys. These are large gender gaps, and the boys' baselines on both items are at the ceiling. On wanting to be pradhan someday, 48% of girls say yes, against 50% of boys. On wanting a high-education job, 16.6% of girls say yes, against 15.7% of boys. These two are statistical ties. On wishing to graduate secondary school, 19.5% of girls against 29.6% of boys, a moderate gap with both genders free to move.

In never-reserved villages, girls trail boys on the four-item normalized aspirations index by 0.507 standard deviations. Table 2 reports the change in this gap in twice-reserved villages: +0.166 standard deviations, with a standard error of 0.057. The ratio, 0.166 over 0.507, is 32.7%. Treatment is assigned at the gram panchayat level, not the village level, so this estimate rests on 20 twice-reserved GPs, about 60 villages and 377 adolescents in the analysis sample.

The decomposition of that closure across the four items, taken directly from the paper's Table 2, is:

The index is a normalized sum, so each item contributes in proportion to the standard deviation of its baseline gap. The two items where the gap is largest, housewife at 40 percentage points and marry-after-18 at 32 points, dominate the average. The two items where the gap is essentially zero contribute almost nothing. The one item with a moderate and substantively interpretable gap, graduation aspiration at 10 points, contributes some, and its coefficient is +0.001.

The paper's Results section walks through three of the four items. Girls become more likely to reject housewifery or in-law-chosen occupations (8.3 percentage points, p=0.053), more likely to want to marry after 18 (8.8 points, p=0.034), and more likely to want a job requiring an education (8.6 points, p=0.02). The fourth item, the desire to graduate secondary school, goes unmentioned. The wish to be pradhan goes unmentioned too. This is the item that maps most literally onto the role-model claim. Girls who grew up watching a woman hold the office should be more likely to want it themselves. The twice-reserved Difference-in-Gap on this item is −0.008, with a standard error of 0.053.

The parents' results in Table 1 follow the same pattern. The paper foregrounds one striking number: the share of parents who want their daughter's occupation, but not their son's, decided by in-laws falls from 76% to 65%. The result one column over goes unmentioned. The coefficient for whether parents want their daughter to graduate from secondary school or higher is 4.8 percentage points with a standard error of 4.8, a t-statistic of one. The parental aspiration measure that maps most directly onto the paper's title is indistinguishable from zero.

Educational attainment is the other headline result. The abstract describes the gender gap there as erased. In Birbhum, the baseline gap in whether a child attends school was 6 percentage points, against a girls' base rate of 68%. The baseline gap in whether a child can read and write was 3.9 percentage points, against a girls' base rate of 91%.

What links these two halves of the paper, aspirations and attainment, is an explicit causal claim. The Discussion attributes the closing of the educational gap principally to the change in aspirations. That is the mechanism the decomposition has to carry. And the component of the aspirations index that most directly predicts educational attainment, the desire to graduate secondary school, is the one that does not move.

A step, not a curve

The dose-response question turns on the once-reserved cell. Of the 165 GPs in the sample, 71 were reserved in exactly one of the two prior cycles. Table 2 reports the change in the gender gap on the four-item index in once-reserved villages relative to never-reserved villages: −0.005 standard deviations (SE 0.052). The gap moves by essentially zero. On the housewife item, the change is −0.023 (SE 0.036). On marry-after-18, +0.014 (SE 0.034). On wishing to graduate, −0.003 (SE 0.028). On wishing to be pradhan, +0.003 (SE 0.039). On every item except the high-education job (+0.009 with SE 0.028, also indistinguishable from zero), the once-reserved coefficient is small, statistically null, and in some cases pointing away from the role-model story. After one cycle of reservation, with a female pradhan in office for up to nine years, the gender gap in adolescent aspirations is statistically indistinguishable from where it was before.

The paper anticipates this. Table S6 splits once-reserved villages into those reserved in 1998 only and those reserved in 2003 only, and reports that the two are indistinguishable from each other (p = 0.28, 0.79, 0.84, 0.82 across the four outcome families). The paper reads this as evidence for a dose-response: what matters is the number of cycles, not how long ago the first one happened. But two nulls being equal to each other does not establish that the dose matters. It establishes only that one cycle produces no effect, whenever it happened. A dose-response requires a curve; the data show flat at zero through one cycle, then a jump at two. The supplementary material concedes as much in the sentence just after introducing S6: the gap closure is "significant only in villages reserved for a woman leader in both election cycles." On the index, the trajectory runs (0, −0.005, +0.166); on housewife (0, −0.023, +0.075); on marry-after-18 (0, +0.014, +0.085); on high-education job (0, +0.009, +0.099). What the design produces is a step localized to the twice-reserved cells.

In name only

The Science paper argues that the effect is a role model effect. Girls and their parents see a female pradhan, update their beliefs about what women can do, and raise their aspirations accordingly. The story requires that the pradhan is seen: indexed in the respondent's mind as a real person occupying a real office.

The QJE paper, drawing on the same villages and the same survey wave, reports the relevant baselines directly. In never-reserved villages, 33% of women correctly named the current pradhan; among men, 67%. Two thirds of adult women in rural Birbhum could not produce the name of the elected village head on demand, in villages where the head had been male the entire time. This is the baseline against which a role-model story would have to operate. Reservation does not raise this baseline. The 2009 paper reports that men are 11 to 12.5 percentage points less likely to know the current pradhan's name in currently-reserved GPs, and 10 to 14 percentage points less likely to know the previous pradhan's name in previously-reserved ones. The authors describe both men and women as less likely to have interacted with female pradhans, and less likely to know their names. More striking still, the 2009 paper finds that the entire attitudinal effect of reservation operates only among the subset of villagers who can name the pradhan: the impact is "concentrated among those who know the Pradhan's name." For women, the paper concludes that the attitudinal effect of reservation is absent, and explains that women are simply too unaware of local politics for exposure to operate. The exposure channel the 2012 paper relies on is the one the 2009 paper documented as gated by recognition and null for women.

Even when citizens try to interact with the office, the person they reach is often not the elected woman. A 2026 phone audit by Varun Karekurve-Ramachandra and Gaurav Sood scraped self-reported mobile numbers for the universe of 2020 Sarpanch winners in Rajasthan, sampled 500 from female-reserved seats, and called them. Of the 377 women who answered, 9% were the elected representative. Eighty-five percent of calls were intercepted by male relatives, with 46% answered by spouses and 12% by sons. Asked to transfer the call, 86% of the male relatives refused. In a parallel sample of 500 male-held open seats, 76% of calls reached the elected representative directly. The same pattern shows up in direct measurement of decision-making authority. In a 2026 paper covering 1,927 GPs and over 9,600 respondents in Gujarat, Madhya Pradesh, and Jharkhand, Alba Huidobro, Soledad Prillaman, and Deepak Singhania find that 89% of male sarpanches report holding final authority over governance decisions, against 43% of female sarpanches. Citizens are correspondingly less likely to bring claims to female sarpanches and report greater uncertainty about who actually governs the GP. The role-model story requires the elected woman to be the person citizens transact with and the person making decisions. Across the largest available samples, in four states, she is often neither.

All of this makes arithmetic for the treatment effect awkward for the role-model account. The ITT on the adolescent aspirations index is 0.166 standard deviations. The Karekurve-Ramachandra and Sood phone audit puts the share of citizens who reach the elected woman at 9%, which implies a treatment-on-treated effect around 1.84 SD. Take the more generous denominator from the QJE companion paper — the 33% of women in never-reserved Birbhum villages who could name the current pradhan — and the implied TOT is still 0.50 SD. Cohen et al.'s self-affirmation interventions, which the Science paper cites approvingly, run at around 0.3 SD. The role-model story has to clear effect sizes that direct, structured interventions struggle to reach, through an exposure channel that on the broadest plausible reading touches one woman in three.

Larger than the bicycle

Karthik Muralidharan and Nishith Prakash studied the Bihar Cycle program, which gave every girl entering Class 9 a grant of roughly ₹2,000 to buy a bicycle. The program raised girls' age-appropriate secondary enrollment by 5.2 percentage points, against a baseline of 16.3% (Table 3), and "reduced the corresponding gender gap by 40 percent." This is one of the most cost-effective girls' education interventions ever documented in India, and it is a direct material transfer. The reservation effect on girls' school attendance is 9.8 percentage points. A decade of female pradhan exposure, on this metric, produces roughly double the effect of giving every girl a bicycle.

The comparison runs the same way when we shift from enrollment to marriage. The Haryana government ran Apni Beti Apna Dhan from 1994 to 1998, promising eligible families a bond that would mature to roughly $400 when their daughter turned 18, conditional on her remaining unmarried. The International Center for Research on Women's evaluation (Nanda et al., 2016) found no effect on the probability of marriage before 18; girls who participated were if anything more likely to be married once they turned 18. A later re-analysis by Biswas and Das (Economica 2026), exploiting birth-cohort variation in a difference-in-differences design, found a 9 percentage point reduction in the prevalence of marriage before 18, "around 34%" of baseline. The reservation effect on adolescent girls' desire to marry after 18 is 8.8 percentage points. ABAD measured marriage behavior; the Science paper measured stated wish, so the two are not strictly comparable. But aspirations are usually easier to shift than behavior, and an 18-year conditional bond moving actual marriage by roughly the same amount as pradhan exposure moves stated wish is a surprise that cuts the same way.

Bicycles and bonds are material transfers. Structured persuasion is closer in spirit to what role modeling, if it works, ought to look like. Diva Dhar, Tarun Jain, and Seema Jayachandran evaluated the Breakthrough classroom curriculum (AER 2022, 112(3): 899–927) in 314 government secondary schools in Haryana. Over two and a half school years, trained facilitators ran 27 forty-five-minute sessions on gender equality, household chores, women's employment, and related themes. The effect on a gender attitudes index was 0.18 standard deviations, which the authors describe as having "converted 16 percent of regressive attitudes." The reservation effect on the adolescent aspirations index is 0.166 standard deviations. Twenty-seven sessions of direct, structured, explicitly persuasive classroom discussion produce effects statistically indistinguishable from what the Science paper attributes to exposure to a pradhan most village women cannot reliably name.

The largest comparator is the most generous one. Mexico's PROGRESA paid eligible households monthly, conditional on school attendance. From 18 months of experimental data, Schultz (2004, Journal of Development Economics 74(1): 199–250) simulated that full program participation across ages 6–14 would lift schooling by roughly 0.7 years. The reservation effect on girls' grades completed is 0.59 grades, from a treatment that involved no cash at all.

Postscript: replication

The argument above relies on the paper's published numbers. A replication using the public data reproduces the baseline means in never-reserved villages within rounding tolerance and returns the twice-reserved coefficients essentially unchanged. The seven primary outcomes across Tables 2 and 3, with standard inference and multiple-testing corrections:

The wild cluster bootstrap corrects for the small number of clusters in the twice-reserved cell. It returns the housewife item at p = 0.051 from a raw p of 0.036. Randomization inference, which permutes the treatment assignment to construct an exact null distribution, returns p = 0.10 on the same item. Bonferroni correction pushes every p-value above 0.25. Benjamini-Hochberg, the more permissive false discovery rate procedure, leaves the three best outcomes tied at p = 0.16. Under any of these corrections, no outcome reaches p < 0.05.

The result is also fragile to the composition of the twice-reserved sample. Dropping each of the 20 twice-reserved GPs in turn and re-estimating, the housewife coefficient fails to clear p < 0.05 in 8 of the 20 subsamples, and the marry-after-18 coefficient in 15 of 20.