This function combines estimates, standard errors and p-values across imputed
datasets, for a single model (i.g., one vertex).
The function was largely taken from mice::pool() and
mice::summary.mipo() code. It averages the estimates of the
complete data model, and computes relevant statistics, following Rubin's
rules (Rubin, 1987, p. 76).
P-values estimation Default: P-values are estimated using the *t-as-z* approach. This is known to the anti-conservative for small sample sizes but provides a computationally efficient solution. Type I error control is addressed more rigorously at the cluster-forming stage. Wald Chi-square test Satterthwaite Approximation
The residuals of the model are currently simply averaged across imputed datasets, for lack of a better idea of how to combine them.
Arguments
- out_stats
Output of
single_lmm, i.e.: a list containing:"stats": a dataframe with estimates and SEs for each fixed effect term)"resid": a vector of residuals for the given model."warning": a character vector with warning messages (if any)
Note that if a model failed for that dataset
out_statshas formlist("error"="Error message")- m
Integer indicating the number of (imputed) dataset
- pvalue_method
String indicating which approximation methods should be used to compute p-values. Options:
"t-as-z","wald-chi2","satterthwaite"(see Details). Default:"t-as-z".- min_pvalue
Float, used to avoid pvalues == 0L for which log10 is Inf. Set this to 0L if no trimming should be applied. Default:
2^-149( == 1.401298e-45) the smallest positive subnormal float.
Note
Used inside run_vw_lmm.