Results

For each treated market we compare the observed search index to its synthetic-control twin (a weighted blend of untreated metros fit to track it). With no game taking place the two should match — chart ①. During game periods, any gap is the estimated effect — chart ②. Difference-in-differences (chart ③) nets out whatever baseline gap exists, and is the number to trust.

The term group buttons select sets: Non-Confounded is the wellbeing signal the study is about; Confounded (game, win) rises mechanically around the event itself and acts as a positive control. Pick a single term to view it alone; multi-term views show the average, each term indexed to its own pre-event baseline = 100, with negative-affect terms inverted so ↑ always means better.

The treated window chooses what counts as treated time; the placebo windows (untreated years, random month) should show no effect and exist to prove the method honest. The baseline defaults to matched-season (same calendar window of other years) because search behavior drifts seasonally.

Error bars are 95% moving-block bootstrap CIs and the p-values are two-sided bootstrap tests — both robust to the week-to-week autocorrelation in search data (naïve standard errors would be too narrow). Where a window covers only 1–2 weekly data points we show the point estimate without inference — daily-granularity data is the planned fix.