Ceci N'est Pas une École
Asher, Jha, Novosad, Adukia, and Tan's February 2026 NBER working paper is a major descriptive contribution (NBER). It links three national datasets at the level of the enumeration block, a unit of roughly 100 to 125 households (about 500 people), and builds a dataset covering 1.5 million neighborhoods and about 63% of India's population. The paper is conditionally accepted at the American Economic Review. It documents that Muslim and Scheduled Caste neighborhoods in Indian cities are less likely to host public facilities and have worse household infrastructure than other neighborhoods in the same town. The segregation findings are real. The data work is impressive. The paper deserves to be read carefully, which is another way of saying it deserves to be read as the paper it actually is, not as the paper public discussion has turned it into.
The authors are explicit about scope: "Our analysis is descriptive. We document patterns of segregation and public service access, but do not disentangle discrimination, homophily, or other sorting mechanisms" (p. 2). In the current version, the education-outcome results that anchored earlier drafts have been removed. Footnote 7 explains: "In the working version of this paper, we showed that people growing up in marginalized group neighborhoods—regardless of their social group—have systematically worse educational outcomes. These results are left for future work, because we lack the data to distinguish whether these outcome disparities are caused by unequal service access, discrimination, or just sorting of marginalized people into poor and underserviced neighborhoods" (footnote 7, p. 4).
This essay makes three arguments. First, the paper's facility-presence variables do not measure access, and the gap between hosting and access is larger than readers typically realize. Second, the paper's most-cited Muslim result, the urban infrastructure gap, is also the result that least survives the paper's own robustness checks, a fact the paper buries in the appendix. Third, the paper's public life, including the authors' own summary materials, is substantially disconnected from its current content.
Hosting is not access
The paper's facility-presence variables measure whether a school or clinic is located inside a 500-person enumeration block. The paper does not and cannot measure distance to the nearest school. The authors state this directly: "we observe neighborhood identifiers, but their geographic coordinates are not available to us" (p. 10). Geocoding would have cost roughly $2 per block for 1.5 million blocks, plus georeferencing (footnote 15). The paper itself notes that urban residents "can travel across many enumeration blocks for work or access to public services" (p. 10).
The denominators make this concrete. The mean probability that an urban block hosts a public secondary school is 0.02, and the mean for a public health facility is also 0.02 (Table 5). The paper's summary line, that a 100% Muslim neighborhood is "only half as likely to have a secondary school" (paper, p. 3), is a halving of 2%, or about 1 percentage point. A 2% hosting rate implies roughly one school per 25,000 people, a catchment spanning dozens of blocks. The urban public primary school gap is smaller still: −0.4 pp on a 7% base. Rural public primary schools are a different case: the Muslim-share coefficient is −8.5 pp on a 33% base, a much larger absolute gap. Since in rural areas blocks can correspond to entire villages separated by kilometers, the hosting-access gap cuts differently there.
The 0-to-100 framing itself needs qualification. The paper describes the distribution of block-level Muslim share as "notably bimodal" (Section 4): 26% of urban Muslims live in blocks above 80% Muslim, and over half of urban Muslims live in blocks above 50% Muslim, but only about 10% of neighborhoods exceed 50% Muslim share. With a bimodal distribution, a linear regression is comparing mean outcomes between two clusters: low-Muslim-share blocks and high-Muslim-share blocks. The coefficient captures that comparison reasonably well. But the 0%-versus-100% scaling presents it as a precisely estimated contrast between endpoints that contain very little data. A 50 pp change in Muslim share, which stays closer to where the data have support, implies gaps about half the headline size.
What the paper's own appendix does to its headline results
The paper's introduction advertises the Muslim infrastructure gaps as among its most dramatic findings: "Compared with a 0% Muslim neighborhood, a 100% Muslim neighborhood in the same city is 10% less likely to have piped water" (p. 3). No media write-up I found fails to repeat this number. It sounds like direct evidence of unequal service delivery, because piped water and closed drainage are household-level measures that do not suffer from the hosting-versus-access problem that plagues the school results.
But the baseline regressions in Tables 5 through 7 control only for town or subdistrict fixed effects and log neighborhood population (Section 5). They do not control for neighborhood income, land values, building density, or any spatial covariates. The paper justifies this in footnote 5: "Since our interest is in how ostensibly universal government services are allocated, we view the uncontrolled estimates as more relevant for our study" (footnote 5, p. 3). That is a defensible framing choice for a descriptive paper about service allocation patterns. But it is a framing choice, and the appendix reveals what happens when you add controls.
Appendix Table A.4 adds slum indicators; the results are "virtually unchanged." But Appendix Table A.5 adds neighborhood consumption, and the Muslim infrastructure coefficients collapse. The piped water coefficient moves from −0.082 to 0.008. Closed drainage moves from −0.099 to 0.017. Electric light moves from −0.019 to 0.001 (Appendix Table A.5). In other words, once you account for the fact that Muslim neighborhoods are poorer, the Muslim infrastructure "gap" is close to zero. The SC coefficients, by contrast, remain large and negative under the same control. This asymmetry between Muslim and SC results is not flagged in the abstract, in the introduction, or in any of the media summaries. The paper's most-cited Muslim finding is also the one that least survives its own robustness check.
The authors may reasonably argue that neighborhood consumption is a "bad control" in the Angrist and Pischke sense: if discrimination causes both segregation and poverty, then conditioning on poverty removes part of the effect you want to measure. That is a fair point in a causal framework. But the paper explicitly disclaims causal identification. In a descriptive paper, showing that Muslim blocks have less piped water and then showing in the appendix that this is almost entirely explained by those blocks being poorer is not a robustness check. It is the main result. It tells you that the Muslim infrastructure gap is a poverty gap, not an independent religion gap, at least at the level of precision the data can support. That the paper presents it the other way around, with the raw gap in the introduction and the consumption-controlled result in the appendix, is a presentation choice that deserves scrutiny.
The private provision puzzle
The private-facility results are, to my eye, the most diagnostic part of the analysis, and they get less attention than they deserve.
In urban areas, the patterns for private facilities track the public ones: private primary schools are 3.7 pp less likely to be present in a 100%-Muslim block (on a 14% base), private secondary schools 5.5 pp less likely (8% base), and private health facilities 9.3 pp less likely (30% base) (Table 6). The private health facility gap is not a small-baseline artifact: a 9.3 pp reduction on a 30% base is a 31% relative decline, and it describes an outcome entirely outside government control. In rural areas, the signs flip: private primary schools are 1.6 pp more common (18% base), private health facilities 2.8 pp more common (13% base), and private secondary schools are roughly flat (Table 6).
This urban-rural split matters. Private facilities respond to demand and profitability, not to government allocation rules. If the story were purely about discriminatory government allocation, you would expect the private sector to at least partially fill the gap where demand existed. Instead, in urban areas, private and public provision move together: both are lower in high-Muslim-share blocks. That is the signature of a shared confounder, not of a mechanism specific to government discrimination. The paper itself offers "limited ability to pay" as one possible explanation (footnote 5), which is consistent with the Appendix A.5 finding: once you control for neighborhood consumption, the Muslim infrastructure gaps mostly vanish.
The second candidate confounder is the built environment. Muslim neighborhoods in Indian cities are disproportionately old-city neighborhoods: the lanes behind Jama Masjid in Delhi, the old quarters of Lucknow, the inner wards of Ahmedabad. These are dense, congested areas with narrow streets, limited open parcels, and high construction costs. A secondary school or private clinic needs land, road access, and physical plant. The probability that any given 500-person block in such an area hosts one of these facilities could be mechanically lower for reasons of urban form. The SECC controls available in the paper (household asset indices, education of household head) cannot capture building density, street width, land availability, or distance from the commercial center of town.
The rural results are consistent with this reading. In villages, the built-environment constraint is weaker: land is more available, density is lower, and the physical barriers to placing a facility are smaller. And in rural high-Muslim-share blocks, private primary schools and health facilities are in fact more common. If Muslim demand were uniformly insufficient, you would not see positive rural coefficients. The urban-rural sign flip is what you would predict if the dominant confounder were something about urban form or urban land markets, not something about Muslim communities as such.
Other concerns
Two additional issues are worth flagging. The facility data come from the 2013 Economic Census, which the paper acknowledges is an imperfect source: it cites evidence of undercounting health providers by roughly a factor of two in Madhya Pradesh, "likely because it misses birth attendants and untrained providers" (footnote 20). If undercounting correlates with informality, and informality differs across Muslim and non-Muslim blocks, the presence variables could be biased in ways the paper cannot assess.
Muslim identity is inferred from names using an LSTM neural network with about 97% out-of-sample accuracy (though labeled set is not representative so it is not clear what 97% maps to in SECC data; footnote 1). This is impressive engineering, but in a regression setting, random misclassification of the right-hand-side variable attenuates slopes toward zero. Non-random misclassification across regions or languages (where naming conventions vary) could introduce subtler biases that are harder to sign.
Finally, the paper's within-town framing is doing more work than many readers realize. The town fixed effects force every comparison to be within the same local government jurisdiction. This means it is entirely possible for Muslims to live in better-resourced towns on average while still showing worse within-town outcomes. The decomposition figures make this logic explicit: for SCs, district-level advantages are almost entirely offset by within-town disadvantages. A reader who walks away thinking the paper proves "Muslims are worse off than Hindus nationally in every sense" is over-reading the estimand. The paper is showing something narrower: conditional on being in the same town, higher-Muslim-share blocks tend to be associated with worse raw outcomes, an association that, for infrastructure at least, is largely explained by neighborhood poverty.
How the paper is cited, and what remains on the authors' own website
The paper's public life is substantially disconnected from its current content. The pattern is visible in both media coverage and author-affiliated materials.
The Print, in its February 2026 write-up, says the study shows segregation "directly affects" who gets services, and its lede describes residents who "send children miles away for school" (The Print, February 2026). The paper measures whether a school sits inside a 500-person block, not how far a child walks. The Wire says facilities "tend to bypass" minority neighborhoods (The Wire, February 2026). CJP asserts that the study "proves" Muslim neighborhoods have limited access (CJP, December 2024). Policy Circle describes the findings as demonstrating the "impact" and "consequences" of segregation and foregrounds education results that the current paper no longer reports (Policy Circle, July 2023). In academic citations, Saba and Gupta (2025) cite the paper while titling their own work "Residential Apartheid in India" (Urban Studies), an escalation from "segregation" that the paper does not support. The Pearson Institute's own summary states that facilities are "systematically allocated away from" minority neighborhoods (Pearson Institute), using the passive voice to imply intentional allocation decisions the paper does not identify.
The most consequential gap between the paper and its public life involves the withdrawn education results. The 2022 working paper's abstract reported that children in minority neighborhoods attain less schooling even after controls (2022 working paper). The February 2026 NBER version removes these results and explains why in footnote 7. But the Development Data Lab website (devdatalab.org/segregation), dated June 2023, still prominently states: "Children in Muslim neighborhoods fare even worse, getting 2.2 fewer years of schooling than children in non-marginalized neighborhoods." The Wire's June 2023 coverage reported these numbers (The Wire, June 2023). CJP built its write-up around them. Muslim Network TV and Madhyamam both reported the "two fewer years" figure in February 2026 (Muslim Network TV; Madhyamam). None of these outlets indicate that the authors withdrew the results.
The paper also reports that Indian segregation is "only slightly lower than Black-White segregation in the U.S." The comparison is carefully hedged in the text: the authors aggregate Indian blocks upward toward roughly 4,000-person units to better match U.S. census tracts, and they note that segregation measures move mechanically with neighborhood size (paper, Section 4). No coverage I found reproduces the scaling caveat. Scroll's headline converts it to Indian "ghettoisation" being "as bad as racial segregation in US" (Scroll.in, June 2023).
The gap between paper and reception is not solely the fault of careless journalists. The authors' own public summary advertises withdrawn results. Their most-quoted finding ("half as likely to have a secondary school") is technically accurate at the block-hosting level but corresponds to about 1 percentage point, a context no outlet provides. And the infrastructure result that anchors the public narrative ("10% less likely to have piped water") is the one that collapses to near-zero in the paper's own appendix once neighborhood consumption is controlled. When researchers present their least-robust finding as the headline and their most-robust robustness check as an appendix table, they bear some responsibility for how the work travels. The media bear the rest.
The paper documents high segregation and real within-town differences in block-level facility presence and household infrastructure. It does not identify whether those patterns come from discrimination, self-sorting, neighborhood poverty, land markets, urban form, or some combination of all of them. For Muslim neighborhoods specifically, the household-infrastructure results are much less stable than they appear once neighborhood consumption is controlled, and the facility-hosting results remain easy to overread as direct measures of access. The paper is important. It also needs to be read with more care than it usually gets.