Troubleshooting

Treat a failed analysis as unavailable. Do not substitute a zero estimate or choose another method solely because it produces significance.

Incomplete or duplicated panels

Parallel analysis expects one finite outcome per geo and observed date in each selected window. Resolve duplicates using a prespecified aggregation rule. Compare the supplied geo set with the retained allocation. A geo absent from the entire input cannot be detected automatically.

Do not drop geos after examining their treatment outcomes. Missing-data analysis requires a separate specification; it is not implemented by automatic omission.

Undefined normalised outcomes

Lift requires strictly positive baseline means. A zero baseline does not imply minus 100 per cent growth. Choose an outcome-unit estimand such as DiD when it fits the original decision, or resolve a measurement error. Do not select a new estimand merely to obtain a favourable result.

Too few geos in a stratum

The supported parallel variance calculation requires at least two geos in each arm of each stratum. Before randomisation, reduce the number of strata or revise the prespecified grouping. Do not merge strata retrospectively based on effects. K-means can create small strata even when the overall geo count is large.

Unidentified CRT

If all geos switch together, calendar time can be indistinguishable from treatment. Additional outcome observations do not fix that identification problem. Revisit the design before launch. If the trial has already run, report the limitation; do not remove period effects solely to obtain an estimate.

Bootstrap or variance unavailable

Constant outcomes can leave analytic variance or BCa acceleration undefined. CRT bootstrap draws can lose treatment estimability. The library raises an error rather than returning a misleading interval. Check whether the proposed data and inference procedure support the intended claim.

Incomplete power scenarios

Check n_valid, n_failed and failure_reasons. The input calendar must be regular daily or weekly data, long enough for both requested windows. Calendar weeks have the same meaning for both frequencies. Power is unavailable when a replicate fails; the conditional rejection rate is not a replacement for it.

Reproducing a result

Retain the original allocation, outcome data, declared windows, strata, seed and dependency versions. Seed zero is valid. Use a fixed random_seed for resampling. Changing the number of workers does not change the power simulation’s seeds.

These repairs intentionally change legacy numerical behaviour. Preserve existing allocations; never re-randomise a completed experiment when upgrading the software. See methodology and validation.