R/ma.wtd.kurtosisNA.R

In miceadds: Some Additional Multiple Imputation Functions, Especially for 'mice'

## File Name: ma.wtd.kurtosisNA.R
## File Version: 0.14



#--- weighted kurtosis
ma.wtd.kurtosisNA <- function( data, weights=NULL, vars=NULL,
        method="unbiased" )
{
    #*** pre-processing
    res <- ma_wtd_stat_prepare_data(data=data, weights=weights, vars=vars )
    data <- res$data
    weights <- res$weights
    M <- length(data)
    #*** weighted kurtosis
    res <- matrix( NA, nrow=M, ncol=ncol(data[[1]]) )
    for (ii in 1:M){
        data1 <- data[[ii]]
        dataResp <- 1 - is.na( data1 )
        data1[ is.na(data1) ] <- 0
        data1 <- as.matrix( data1 )
        # calculate means
        sumweight <- colSums( dataResp * weights )
        M_vars <- colSums( data1 *  weights ) / sumweight
        M_varsM <- matrix( M_vars, nrow=nrow(data1), ncol=length(M_vars), byrow=TRUE )
        data1adj <- ( data1 - M_varsM ) * dataResp # take care of missings
        w1 <- colSums( dataResp * weights )
        sdx <- sqrt( colSums( data1adj^2 * weights ) /  w1 )
        # adjustment of covariance
        if (method=="unbiased"){
            wgtadj <- w1 - colSums( dataResp * weights^2 ) / w1
            wgtadj <- w1 / wgtadj
            sdx <- sqrt(wgtadj) * sdx
        }
        sdxM <- matrix( sdx, nrow=nrow(data1), ncol=length(sdx), byrow=TRUE)
        data1adj <- ( ( data1 - M_varsM ) / sdxM )^4 * dataResp
        M1 <- colSums( data1adj *  weights ) / sumweight
        res[ii,] <- M1 - 3
    }
    res <- colMeans(res)
    names(res) <- colnames(data[[1]])
    return( res )
}