Do German Census Corrections Recur in the Same Districts? Evidence from the 2011 and 2022 Censuses

Census 2022 lowered Germany's population by about 1.4 million (1.6%) against the register-based estimate carried forward from Census 2011. Census 2011 had made a correction of similar size. If register error builds up in the same places between censuses, per-capita statistics for those districts are biased in a known direction for a decade. This paper tests whether the districts corrected most in 2022 are the ones corrected most in 2011, and which district characteristics predict the 2022 correction. For 395 harmonised districts, the 2022 correction is computed from the official back-calculated estimate on the 2022 basis, relative to the published 2011-basis estimate, both at 31 December 2021. The 2011 correction is computed from the Destatis comparison of both bases at 31 December 2011; this direct measure replaces the pre-registered extrapolation proxy, which rested on a mistaken claim that the comparison was not public. The rank correlation between the two corrections is negative, not positive: Spearman rho = -0.24 (95% bootstrap CI -0.34 to -0.14, permutation p = 0.0001). The persistence hypothesis is not supported. The registered proxy gives rho = -0.23 (CI -0.33 to -0.13), and the sign holds under its two registered variants and, post hoc, when Länder are resampled as blocks (CI -0.35 to -0.07). Weighted by population, the correlation is close to zero (r = -0.05). Of seven pre-registered covariates, only the 2014-2016 non-German inflow (+0.46 percentage points per standard deviation, HC3 CI 0.12 to 0.80) and population size (+0.27, CI 0.02 to 0.52) are distinguishable from zero under independent-district errors; neither remains so when errors are clustered by Land (post hoc). Together the seven explain 6% of the district variance. The coefficient on the foreign population share is not distinguishable from zero. The registration's timing cannot be verified.