Resamples individuals with replacement and re-estimates the selection gradients on each resample to obtain bootstrap standard errors and percentile confidence intervals for every coefficient (beta, gamma, gamma_ij).
Usage
bootstrap_selection(
data,
fitness_col,
trait_cols,
fitness_type = c("auto", "binary", "count", "continuous"),
standardize = TRUE,
group = NULL,
use_relative_for_fit = TRUE,
n_boot = 1000,
conf = 0.95
)Arguments
- data
A data frame containing fitness and trait measurements.
- fitness_col
A string specifying the name of the fitness column.
- trait_cols
A character vector of trait column names.
- fitness_type
A string indicating the fitness type:
"auto","binary","count", or"continuous". Binary and count fitness take their p-values from a GLM (logistic, or Poisson and negative binomial) on the raw values; the gradients always come from OLS on relative fitness.- standardize
Logical indicating whether to standardize traits to mean 0 and SD 1. Default is
TRUE.- group
Optional string specifying a grouping variable (e.g., "year", "site"). Traits and relative fitness are standardised within each group, and for binary and count fitness the GLM that supplies the p-values gets a separate intercept for each group.
- use_relative_for_fit
Logical; if
TRUE(default) the gradients are estimated on relative fitness \(W / \bar{W}\) for every fitness type, as in Lande & Arnold (1983). SetFALSEonly to reproduce coefficients on the absolute fitness scale.- n_boot
Integer number of bootstrap resamples. Default is 1000.
- conf
Confidence level for the percentile interval. Default is 0.95.
Value
A data frame with one row per coefficient and columns Term,
Type, Estimate (point estimate on the full data),
Boot_SE, CI_lower, CI_upper, P_Value, and
N_Boot (usable resamples for that coefficient). P_Value is
the parametric p-value from the point-estimate fit (OLS t-test, or the
logistic-GLM Wald test for binary fitness), not a bootstrap p-value.
Details
Each resample goes through the whole procedure again: with
standardize = TRUE the traits are restandardised, and fitness is made
relative to the resample's own mean, before both models are refitted, so
the intervals carry the uncertainty in the standardisation as well as in
the fit. Resamples whose fit fails are dropped, and N_Boot counts
the ones that were used.
When group is given, individuals are resampled within each
group so every resample keeps the original group sizes. A resample in
which a trait has no variance within a group (so it could only be centred)
is discarded rather than fitted on a degenerate value. The point estimate
is the ordinary selection_coefficients() fit on the full data and
its warnings are reported as usual.
With continuous fitness the intervals for beta can be too narrow when the
traits have heavy tails (skewed or not) and the linear model leaves out
curvature, though less so than the parametric ones (see
check_selection_assumptions()). For survival and counts in the
package's simulation they covered a little less than the parametric
ones, 92 to 93% for beta. Refitting
on a transformed trait changes the scale of the gradients and is only a
check.
Examples
set.seed(1)
bootstrap_selection(bumpus, "survival", c("total_length", "weight"), n_boot = 50)
#> Warning: Collinear traits (VIF above 5) may inflate the standard errors
#> Term Type Estimate Boot_SE CI_lower
#> 1 total_length Linear -0.17811220 0.10303749 -0.4021452
#> 2 weight Linear -0.09992466 0.08007062 -0.2573204
#> 3 total_length² Quadratic -0.23342057 0.26857154 -0.7271855
#> 4 weight² Quadratic 0.02096490 0.17479658 -0.2530171
#> 5 total_length × weight Correlational -0.06949785 0.16260489 -0.3050220
#> CI_upper P_Value N_Boot
#> 1 0.005323691 0.07276336 50
#> 2 0.032591335 0.30287778 50
#> 3 0.174355313 0.24402273 50
#> 4 0.329534214 0.93873420 50
#> 5 0.271905000 0.57236384 50