Complete-case frame for a set of model-selection candidates

Drops every row that is missing a variable used by any candidate, so that all candidates are fitted to identical records.

Usage

selection_frame(data, candidates, formula = NULL, extra = NULL)

Arguments

data

A data frame.

candidates

A named list of candidates, as accepted by sdmTMB_selection(): each element a formula, or a list that may contain a formula element.

formula

Optional default formula for candidates that do not carry one.

extra

Character vector of further columns that must also be complete (e.g. the coordinate columns, the time column, or an offset column).

Value

data with incomplete rows removed. The number of dropped rows and the columns that were checked are attached as the "dropped" and "columns" attributes.

Details

This is the precondition that makes a selection table mean anything. Information criteria and cross-validated log-likelihoods are only comparable between models fitted to the same response data; if one candidate carries s(depth) and another does not, and depth has gaps, sdmTMB() silently drops different rows for each and the resulting AIC column compares two different quantities.

Build the selection frame once, select on it, then fit the chosen specification to the full data: there is no reason to keep discarding records because a covariate that did not enter the model is missing for them.

See also

sdmTMB_selection()

Author

Mikko Vihtakari

Examples

cands <- list(
  "Full" = density ~ 0 + as.factor(year) + depth_scaled,
  "No depth" = density ~ 0 + as.factor(year)
)
d <- selection_frame(sdmTMB::pcod, cands, extra = c("X", "Y"))
attr(d, "dropped")
#> [1] 0