← All insights

When admin data is the only data

For low-frequency outcomes in remote Indian districts, such as rural employment days, housing completion status, foodgrain offtake and school attendance, survey data is expensive, dated and often unavailable at the geographic scale where policy operates. Administrative data is the alternative: NREGA muster rolls and the MGNREGA MIS, PMAY-G housing records, PDS dealer records, UDISE+ school information and the Aadhaar-enabled Public Financial Management System. These sources are large and granular to the district or block. They update more often than the National Sample Survey, and anyone can download them from a public portal.

We use them all the time, and most of our district-level analysis would not be possible without them. Over the years we have also kept a private list of the ways they go wrong without anyone noticing. Usually the data was collected for one purpose and we are using it for another. The most consequential errors we have seen came from reading an administrative record as a direct measure of what it was filed about.

The first problem is with NREGA muster rolls. They record work as the worksite operator reported it, signed off by the panchayat and entered into the MIS by a block-level data entry operator, and each of those steps can introduce variation that has nothing to do with the work itself. We have seen blocks where MIS attendance figures are systematically rounded to multiples of five. We have seen districts where disputes over payment delays led to attendance being entered retroactively, weeks after the work, from whatever the operator had to hand. The MIS attendance variable is a bureaucratic output, and we do not use it as a direct count of person-days worked.

The second is that job card holders are a selected population. Households without job cards do not appear in the NREGA data at all. Whether a household has a job card is correlated with caste, literacy, awareness of the scheme, distance from the panchayat office and the household's political position in the village. NREGA participation can be a sensible measure of rural distress if the analyst remembers that the population it shows is the one that completed an administrative enrolment step. The most distressed households are not always in that population, and sometimes they are systematically missing from it.

The third is definitional drift. The MIS defines a "person-day generated" in a specific way. That definition has changed over time, states have applied it differently, and it has been gamed where the official metric and conditions in the field diverged. The published time series looks comparable across years, and the definitions behind it do not support that. To read NREGA trend data accurately you need to know which years are being compared and what the definition was in each.

The fourth is geographic coverage. Districts where the MIS is poorly maintained look like districts with little NREGA activity. Sometimes the opposite is true, because the administrative state there is too thin to record the activity that is happening. This is the standard problem with administrative data, where a missing record can mean that nothing happened or that nothing was recorded. In an outcome variable, that ambiguity is a serious problem.

PMAY-G shows a related problem. Housing scheme records are often used as a proxy for poverty or housing deprivation, and they have their own selection pipeline. A household has to apply, be found eligible by the local administration, be approved, receive the first instalment, complete construction and have the final geo-tagged photograph uploaded to the MIS. Some households drop out at every step. The published count of "houses sanctioned" is consistent with very different counts of houses actually inhabited, depending on which step the analyst reads.

Our practical advice has three parts. First, document the administrative process that produced the record. Read the scheme's operational guidelines, ask a block-level functionary how the data entry actually happens, and find at least one case where the official record and the field reality diverged. This tells you what the data is.

Second, where the stakes are high, check the records against a small survey in a randomly drawn subsample. Comparing the administrative record with direct observation in twenty randomly selected villages costs perhaps fifteen days of fieldwork. That is worth more than any amount of further analysis of the original records, and the match rate between the two is an important finding in itself.

Third, do not report the administrative variable in a headline as if it were the outcome. Report it as an administrative record and, where the gap to the outcome is known, quantify it. "Of the 4,217 PMAY-G houses sanctioned in this district between 2022 and 2024, our ground-truth audit of 80 randomly selected geotagged completions found 71 inhabited, 6 incomplete, 3 not located" says something different from "this district sanctioned 4,217 PMAY-G houses." Both sentences are true, and they support different policy conclusions.

The administrative state in India produces more data per district than any external research team will ever match, and we depend on it. Using it well means knowing how each record was made, so we treat the published MIS variable as the place where our work begins.

Useful references are the NREGA MIS for muster roll and works data, the PMAY-G MIS for housing scheme records and UDISE+ for school-level data. For method, Card and DellaVigna's JEP essay on the use of administrative data in applied research is a useful cross-reference.