Analysis Lag

Lagged correlation between two metrics: r/p/q/n across a window of lags.

Pro only (402 for free plans), alongside the rest of the Signal surface. A free caller has no way to name a pair here anyway: the outline they get from /analysis/ecosystem carries opaque indices, not metric ids.

For each lag L in [-max_lag, +max_lag] (in resample buckets), metric_b is shifted by L and correlated with metric_a over the overlapping points. A positive lag means changes in metric_b tend to follow metric_a — temporal precedence, not proof of causation.

Two guards against the profile's multiple-comparisons problem (up to 61 correlated tests whose tallest bar the UI invites the user to read off):

• each p is corrected for autocorrelation via Bartlett's effective sample size n_eff = n·(1−ρ₁ρ₂)/(1+ρ₁ρ₂) (daily health series are strongly AR(1), which makes raw p wildly anti-conservative), and
• q is Benjamini–Hochberg across the lag family, so the client can emphasize only lags with q ≤ 0.05 and render the rest as context.

resample="None" is rejected: a row-shift over irregular raw timestamps is not a time shift, so lags are only meaningful on a regular grid.