EXPLORATORY EVENT STATISTICS

This is a post-discovery analysis of 17 already selected episodes. No estimate is a causal effect or a confirmatory significance test. The registry records independently sourced dates where available; this does not make event selection independent.

The primary descriptive contrast compares the selected month with the prior 3 calendar months pooled. Each shared outlet receives equal weight in the paired tone contrast. Positive/negative/neutral labels remain in the denominator. Ten thousand paired outlet resamples yield percentile sensitivity ranges; few and dependent outlets mean these are not calibrated confidence intervals. Leave-one-out and outlet-sign counts are reported. Full post-event months 1–3 and 4–6 are compared with the same pre-event baseline where all requested months exist.

Attention is the share of country collected news and is descriptive. The absence of outlet-specific compatible denominators prevents an outlet-resampling attention interval. Adjusted copy membership and geography follow the parent correction ledger. Adjusted labels are partial audited corrections, not full-corpus validation.

Interrupted series exclude the anchor month because it mixes pre/post days and anticipation. Separate linear pre/post trends use up to 12 months on each side. At least 9 pre and 6 post nonzero-coverage months, with no other interior gaps, are required. The main reported departure averages months 1–3 relative to the extrapolated pretrend. Bartlett HAC at 2 actual calendar-month lags uses n/(n-k) correction and a t reference. Lag 3 and an annual harmonic are sensitivity checks, not specifications selected by favorable results. Monthly composition, contemporaneous events, limited time points and selective corrections remain confounders.

Sources: Lopez Bernal et al. https://pmc.ncbi.nlm.nih.gov/articles/PMC5407170/ ; Cameron and Miller https://faculty.econ.ucdavis.edu/faculty/cameron/research/Cameron_Miller_JHR_2014_July_09.pdf ; statsmodels https://www.statsmodels.org/stable/generated/statsmodels.regression.linear_model.RegressionResults.get_robustcov_results.html .

Run: statistics/.venv/bin/python statistics/run_event_statistics.py from corrected_analysis, or use absolute paths. Protocol and register are separate from estimates and were written before this implementation estimated the results.