Process Special Multiple Imputation Output From fit.mult.impute
# S3 method for fit.mult.impute
processMI(
object,
which = c("validate", "calibrate", "anova"),
plotall = TRUE,
nind = 0,
prmi = TRUE,
...
)an object like a validate, calibrate, or anova result obtained when no multiple imputation was done. This object is suitable for print and plot methods for these kinds of objects.
a fit object created by fit.mult.impute
specifies which component of the extra output should be processed
set to FALSE when which='calibrate' to suppress having ggplot render a graph showing calibration curves produced separately for all the imputations
set to a positive integer to use base graphics to plot a matrix of graphs, one each for the first nind imputations, and the overall average calibration curve at the end
set to FALSE to not print imputation corrections for anova
ignored
Frank Harrell
Processes a funresults object stored in a fit object created by fit.mult.impute when its fun argument was used. These objects are typically named validate or calibrate and represent bootstrap or cross-validations run separately for each imputation. See this for a case study.
For which='anova' assumes that the fun given to fit.mult.impute runs anova(fit, test='LR') to get likelihood ratio tests, and that method='stack' was specified also so that a final anova was run on the stacked combination of all completed datasets. The method of Chan and Meng (2022) is used to obtain overall likelihood ratio tests, with each line of the anova table getting a customized adjustment based on the amount of missing information pertaining to the variables tested in that line. The resulting statistics are chi-square and not $F$ statistics as used by Chan and Meng. This will matter when the estimated denominator degrees of freedom for a variable is small (e.g., less than 50). These d.f. are reported so that user can take appropriate cautions such as increasing n.impute for aregImpute.