Concepts represented more than once among selected covariates

concept_map and prune_to_one_form act on a candidate pool before a screen. This is the after-the-fact view, and it is the only one available for a job run without pruning – which is the case for any faithful reproduction of a SAS %macro model, since SAS had no notion of a concept.

Usage

selection_crowding(selected, aliases = character(0), ...)

Arguments

selected

Character vector of phase-qualified selected parameter names.

aliases

Named character vector, as for concept_of. Only those whose target was selected in a given phase are forwarded, because concept_map refuses an alias pointing outside its pool – right for a candidate pool, wrong here, where a parent form may legitimately be absent (in_arin can be selected while area_int is not).

...

Passed to concept_map.

Value

A data frame with columns phase, concept, n_forms and forms, ordered by phase then decreasing n_forms. Empty when nothing is crowded.

Details

Why it is worth reporting. A step budget is shared: max_steps caps total steps across both phases. So a slot spent on a second form of a concept already in the model is a slot no other variable could have. When the selection also stopped on the cap, the two facts together say the model is not merely budget-limited but budget-limited by redundancy – and neither fact is visible in a coefficient table.

Grouping is concept_map's, so it is conservative by construction. Under-grouping only understates crowding; over-grouping would assert two clinical concepts are one. This errs the safe way, and the result is a floor: a concept nobody listed is still counted as separate variables.

selected is phase-qualified, as setdiff(names(coef(sel)), names(coef(base))) returns it: "late.age", "early.zexp2". Phases are counted separately – one form of age in each phase is two slots but not a crowded concept.

See also

concept_map, prune_to_one_form

Examples

selection_crowding(c("early.age", "early.ln_age", "late.bsa"))
#>   phase concept n_forms       forms
#> 1 early     age       2 age, ln_age