Smooth operator: Too little variation in India’s demographic estimates

Abstract: A survey’s design determines how much its estimates should vary. If they vary too little, they cannot have been produced as described. India’s Sample Registration System publishes annual demographic rates for a sixth of the world’s people but releases no underlying records. I find that its recent estimates are implausibly smooth: Over 2014–2024, 13 of 19 states’ birth-rate series and 15 of 19 states’ infant-mortality series fall below the fifth percentile of what sampling alone would produce. The shortfall has gotten worse over time but varies by indicator: Recent crude death rates vary more than the sampling benchmark. These results are robust to publication rounding, the choice of trend, and a variety of sampling-error correlations, including those measured in India’s Demographic and Health Surveys. They are evidence not of bias but that the published numbers are incompatible with the account of how they are made.