Fixed bootstrap percentile confidence intervals to match Stata fuzzydid: failed replications (carrying a sentinel value) are dropped when computing the bootstrap standard error but retained when computing the percentile interval, so their extreme values widen the interval under weak support. The previous behavior discarded them for both, producing intervals that were too narrow whenever share_failures > 0.
W_CIC is now reported as non-estimable (an informative error, rather than a finite estimate over the remaining strata) when a positive-mass treated baseline treatment stratum has no control-group support and thus an undefined quantile map. The previous behavior silently dropped the unsupported stratum, changing the estimand and reweighting the supported strata.
Fixed multi-period aggregation of W_DID, W_TC and W_CIC to match Stata fuzzydid. Contributions are now weighted by the post-period switcher-cell count (n11, not n11 + n10) and the design sign of the switcher arm, and W_TC/W_CIC are weighted averages of the per-subdesign Wald ratios (anchored on the DID denominator) rather than ratios of summed numerators/denominators. Results now reproduce Stata to 1e-6 on unbalanced multi-period designs; the previous behavior differed on any unbalanced or multi-arm panel.
Fixed the sieves = TRUE basis for two or more continuous covariates: it now uses a total-degree polynomial (including cross-products such as x1 * x2), matching Stata’s Legendre sieve span. The previous additive-only basis omitted interactions and could not recover interaction effects.
The covariate-adjusted stable-control (special_case) path now returns a missing TC estimate when the adjusted DID components are non-finite, instead of silently falling back to a treated-only before/after change.
numerator = TRUE now matches Stata by erroring unless the design has exactly two time periods and two treatment groups (the only case in which the reduced-form numerator is defined).
Added frozen Stata-parity regression tests for the unbalanced multi-period and two-covariate sieve designs.