Calibrating Change-Point Inference in Regional Rainfall Extremes: Evidence from Mizoram, Northeast India

Abstract Regional rainfall studies often detect a change point and test the resulting periods as if the break had been specified in advance, a procedure that severely inflates false positives when the break is instead selected from the same data being tested. We quantify this inflation and provide a general calibration framework, applied to ten ETCCDI indices from 0.25° gridded daily rainfall across 29 cells in Mizoram, northeast India (1950–2025). The procedure was calibrated using multivariate AR(1) surrogates that preserve temporal persistence and spatial covariance but contain no change point. In these signal-free series, the uncorrected procedure gave p < 0.05 in 83% and p < 0.001 in 27%, a 17-fold and 265-fold inflation over nominal levels. Applying the calibration to the observed record, only two of ten epoch contrasts survived correction. Consecutive wet days increased 44.9% across the 2006 break (p = 0.003) and was the only spatially coherent signal across all sensitivity settings. SDII declined 24.7%, significant only under the primary null (p = 0.037; sensitivity p = 0.128). Monsoon totals were unchanged (− 2.9%, p = 0.629), while wet days increased 17.5% and intensity declined 16.5% (both p ≤ 0.001), indicating longer, less intense spells rather than a supply change. The 232 mm decline in annual rainfall was not distinguishable from correlated noise (p = 0.058). These results show that regional rainfall analyses relying on searched change points must calibrate the full detection-and-testing procedure. Without it, evidence for regional rainfall change is almost entirely an artifact of break selection.