Diagnostics and sensitivity

Review diagnostics before interpreting a precise-looking estimate. They answer different questions from the effect table.

Check	What to look for	What to do next
Sample and event counts	Sparse groups, few events, exclusions	Review coding, missingness and eligible scope
Propensity overlap	Unsupported treatment comparisons	Review the target population and contrast
Weights and ESS	Dominant observations or low effective sample size	Inspect raw and truncated weight diagnostics
Covariate balance	Residual standardized differences	Review the matching/weighting design
PH diagnostics	Evidence against proportional hazards	Review Cox interpretation and eligible survival contrasts
Censoring and horizon	Weak follow-up at the target time	Check the estimand’s support and assumptions
Method status	Skipped versus attempted-and-failed	Read the specific reason instead of treating a blank as zero

The SMD 0.1 line, PH P-value references and other defaults are review aids. None is a universal certificate of identification. See parameter control for explicit settings and record sensitivity choices separately from the primary model.

Sensitivity summaries

E-values, tipping-point summaries, matching sensitivity and related supplements provide method- and scale-specific context. They do not establish absence of unmeasured confounding. A deterministic null-distance calculation is not a model-based hidden-bias threshold.

Read which sensitivity methods were requested, eligible, run, skipped and failed. If a method is ineligible for a risk difference or other additive target, keep that status instead of borrowing another estimator’s numbers.

Keep the audit trail

Save the input definition, selection history, settings, full testing family, diagnostics and original method tables. A primary display row is a presentation choice. It does not mean other requested methods were run successfully or support the same target. Runtime support and method validation are documented separately in validation and support.