R/F_deviances.R

In CenterForStatistics-UGent/RCM: Fit row-column association models with the negative binomial distribution for the microbiome

#' A function to extract deviances for all dimension,
#' including after filtering on confounders
#'
#'@param rcm an object of the RCM class
#'@param squaredSum a boolean, should total deviance be returned?
#'
#'Total deviances can be deceptive and not correspond to the differences in
#' log-likelihood. As the dispersion is different for each model.
#'  To compare models it is better to compare likelihoods.
#'
#'@return If Sum is FALSE, a named array of deviance residuals of the
#'  independence model and all models with dimension 1 to k, including after
#'  filtering on confounders. Otherwise a table with total deviances (the sum of
#'  squared deviance residuals), deviance explained and cumulative deviance
#'  explained.
deviances = function(rcm, squaredSum = FALSE) {
    vec = if(length(rcm$confModelMat))
        c(0, 0.5, seq_len(rcm$k)) else c(0:rcm$k)
    outnames = c("independence",
        if (length(rcm$confModelMat)) "filtered" else NULL,
        paste0("Dim ", seq_len(rcm$k)))
    if (squaredSum) {
        tmp = vapply(FUN.VALUE = numeric(1), vec, FUN = function(i) {
            sum(getDevianceRes(rcm, i)^2)
        })
        names(tmp) = outnames
        # Also make cumulative comparisons
        cumDevianceExplained = round((tmp - tmp[1])/(tmp[length(tmp)] -
            tmp[1]), 3)
        out = rbind(deviance = tmp,
                    devianceExplained = c(0, diff(cumDevianceExplained)),
            cumDevianceExplained = cumDevianceExplained)
    } else {
        out = lapply(vec, function(i) {
            getDevianceRes(rcm, i)
        })
        names(out) = outnames
    }
    return(out)
}