Data quality for causal inference
Causal methods cannot rescue broken exposure logs, post-treatment controls, or identity drift. Start with the data audit before choosing the estimator.
Exposure logs
Confirm when a user became eligible, assigned, exposed, and able to act.
Covariates
Use only variables observed before treatment when adjusting for confounding.
Missingness
Check whether missing events are random or concentrated in a treatment, device, region, or plan.
- Keep raw event counts next to modeled outputs.
- Audit bot and internal traffic filters.
- Version metric definitions when product logging changes.