Country lead/lag cross-correlation

Inputs. Daily citable-censorship incident counts (incident_type IN ('censorship','mixed')) for the top-50 countries by 365-day volume. IODA disruption rows are excluded because they conflate fiber cuts, BGP outages, DDoS, and weather with censorship — the same correction shipped in the May-21 forecast-labels fix.

Smoothing. 7d centered rolling mean. Daily incident series are spiky point events; without smoothing the cross-correlation is dominated by accidental single-day overlap.

Cross-correlation. For each ordered pair (A, B) and lag L ∈ {-30, …, +30} \ {0}, we compute Pearson r between A[t] and B[t+L] over the overlap window. Two-sided p-value from the t-statistic on r with n_eff = 365 − |L|. Lag 0 is intentionally excluded: simultaneous correlation already lives at /atlas/correlation-matrix.

Direction convention. If the best lag is positive for (A, B), A leads B; if negative, we swap to (B, A) and flip the sign so reported lag_days is always non-negative.

Multiple-comparison correction. With 20 countries × 60 lags per pair, the raw p-values vastly overstate significance. We apply Benjamini–Hochberg FDR control at α=0.05 and require |r| ≥ 0.4. BH-FDR (rather than Bonferroni) is chosen because the analysis is exploratory and we tolerate a controlled false-discovery rate rather than the catastrophic loss of power Bonferroni would impose.

Sparse-country filter. Countries with fewer than 5 nonzero days OR fewer than 5 total incidents are dropped — they produce artifactual r=1.0 correlations after smoothing because two delta spikes trivially align. We started at 50 countries and kept 20 that clear the support floor.