Confidence interval methods (ci_method)

CausalStress distinguishes between a user-facing CI intent (ci_method) and the CI implementation chosen by a specific estimator.

Details

Allowed values

• "none": no confidence intervals are computed.

• "default": use the best supported method for the estimator.

• "bootstrap": compute bootstrap confidence intervals (requires a deterministic seed).

• "native": use the estimator's native/Wald-style CI, when supported.

Precedence and defaults

• In estimator functions (e.g., est_lm_att()), a missing config$ci_method is treated as "none" to avoid surprising, random, or expensive computation.

• Runner functions (e.g., cs_run_single()) may set ci_method when it is missing, based on runner flags such as bootstrap=TRUE. Runners also inject config$seed from the run's seed.

Estimator-specific mapping for "default"

• Defaults to "bootstrap" for: lm_att, ipw_att, gengc, gengc_dr.

• Defaults to "native" for: grf_dr_att, bart_att, tmle_att.

Reproducibility rule (bootstrap)

Bootstrap confidence intervals require config$seed. If bootstrap CIs are requested explicitly (ci_method = "bootstrap") or indirectly (ci_method = "default" mapping to bootstrap) and no seed is available, CausalStress fails with a "causalstress_config_error".

Recommendation

• For one-off exploratory calls, either set ci_method = "none" (point estimate only) or provide a seed in config.

• For campaigns, prefer the runner APIs (cs_run_seeds(), cs_run_campaign()) which handle seeds deterministically.

Introspection

Estimator outputs record the resolved CI method in res$meta$ci_method and the origin of that choice in res$meta$ci_method_source (e.g., "explicit", "default_mapped", "runner_bootstrap").