Parameters and analysis settings

Decide the endpoint, exposure contrast, covariates and testing family before changing performance or screening settings. Keep the effective request with the result so another reader can see what was actually used.

The shared request

The workflow examples use explicit route, features where applicable, endpoints, methods and parameters. A complete workflow sets parameters$complete_workflow = TRUE; use plan_only = TRUE to inspect the request before an execution you have not yet agreed.

Setting	Purpose
`workflow`	Choose `pathway_first` or `exposure_first`
`features`	Exact named features for exposure-first
`endpoints`	Binary endpoint and/or time, event status, event code and units
`covariates`	Agreed baseline adjustment variables
`methods`	Requested estimator set; support filtering remains explicit
`parameters$seed`	Reproducible random settings
`parameters$profile`	A supported settings profile
`parameters$complete_workflow`	Request the full analysis-to-delivery route

Direct R control objects

Advanced direct workflow APIs can use iobr_causal_control():

ctrl <- iobr_causal_control(
  profile = "standard",
  screening = list(top_n = 30, fdr_threshold = 0.15),
  estimation = list(boot_R = 500, boot_parallel = TRUE, boot_ncpus = 4)
)

Use the relevant function’s help for its control argument. This object is not an instruction to run a second backend after a complete shared request.

Profiles provide defaults, not study-specific justification. A broad screening FDR gate is not a universal threshold for final claims. Formal OS mediation under a frozen project protocol has its own fixed model and family rules.

Exposure and survival contrasts

Direct estimator wrappers expose selected options. The core iobr_causal_estimate() additionally accepts effect_spec for bootstrap, continuous intervention and survival target settings. For example, forest_num_trees is the core setting behind the forest wrapper’s num_trees. The default survival target is RMST difference; Cox summaries require an explicit hazard_ratio target. A fixed-horizon survival probability contrast is separate.

Specify policies, dose contrasts and horizons from the scientific question and observed support. A wrapper does not expose every advanced setting. Keep primary settings fixed and label sensitivity analyses separately.

Detailed control dictionary

Parameter	Controls	Increase means	Decrease means
`top_n`	how many ranked candidates move forward	broader review	fewer candidates
`fdr_threshold`	FDR gate for candidate highlighting	more permissive	more conservative
`p_threshold`	nominal P reference	more exploratory signals	fewer nominal signals
`missing_threshold`	allowed missing fraction	keeps more incomplete variables	removes incomplete variables earlier
`min_abundance`	value required for one matched sample to count as detected	requires more abundance	counts lower values as detected
`min_prevalence`	fraction of matched samples that must be detected	removes more rare features	keeps rarer features
`min_complete_rate`	complete-case fraction	stricter data support	more permissive
`min_orr_n`	ORR complete-case N	blocks small ORR scans	allows smaller ORR scans
`min_survival_n`	survival complete-case N	blocks small survival scans	allows smaller survival scans
`min_group_n`	each binary group size	stricter group support	allows rare groups
`min_event_n`	survival event count	stricter survival evidence	allows event-sparse scans
`boot_R`	bootstrap repetitions	more stable, slower	faster, noisier
`boot_parallel`	multiple CPU workers for candidate screening and bootstrap-backed estimators	faster screening/formal estimation, more memory use	single-process execution
`boot_ncpus`	requested CPU worker count	usually faster, more memory use	lower memory use, slower
`ipw_truncate`	IPW weight trimming	wider range trims less	narrower range trims more
`psm_caliper`	matching tolerance	more matches, looser	fewer matches, tighter
`ph_p_warn`	PH warning threshold	usually keep at 0.05	usually keep at 0.05
`censor_rate_warn`	censoring warning cutoff	more tolerant	stricter
`ess_ratio_warn`	effective sample warning	more tolerant if lower	stricter if higher