R/mice.impute.pmm4.R

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

## File Name: mice.impute.pmm4.R
## File Version: 2.22

mice.impute.pmm4 <- function (y, ry, x, donors=3, noise=10^5, ridge=10^(-5), ...)
{
    res <- mice_imputation_factor_pmm_prepare(y=y)
    y <- res$y
    y_aggr <- res$y_aggr
    is_factor <- res$is_factor

    x <- cbind(1, as.matrix(x))
    res <- mice_imputation_norm_draw(y=y, ry=ry, x=x, ridge=ridge, ...)
    yhatobs <- res$yhatobs
    yhatmis <- res$yhatmis
    yobs <- res$yobs
    GG <- 1000* max( abs( yhatobs[,1] ), abs( yhatmis[,1] ))
    dfr <- cbind( 1, 1:nrow(yhatobs), yhatobs[,1], y[ry] )
    dfr0 <- cbind( 0, 1:nrow(yhatmis), yhatmis[,1], NA)
    dfr <- rbind( dfr, dfr0 )
    dfr <- data.frame(dfr[ order(dfr[,3] ), ])
    colnames(dfr) <- c("obs", "index", "yhat", "y")
    # add some small noise to create unique entries in matrix d0
    d00 <- abs(diff(dfr$yhat))
    fg1 <- min( d00[ d00 > 0 ] )
    dfr$yhat <- dfr$yhat + stats::runif( nrow(dfr), 0, fg1 / noise )
    dfr <- data.frame(dfr[ order(dfr[,3] ), ])
    dfr$sortindex <- seq( 1, nrow(dfr))

    #******************
    # create donors
    dfr1 <- dfr[ dfr$obs==1, ]
    N1 <- nrow(dfr1)
    for (dd in 1:donors){
        dfr1[, paste("donor.p",dd,sep="") ] <- c( dfr1$yhat[-seq(1,dd)], rep(NA,dd) )
        dfr1[, paste("donor.m",dd,sep="") ] <- c( rep(NA,dd),
                                            dfr1$yhat[-c( seq( N1,N1+1 - dd) ) ]  )
    }
    dfr1$yhatdonor <- dfr1$yhat
    #**** look at appropriate donors
    dfr$sortindex.obs1 <- cumsum(  dfr$obs )
    dfr$sortindex.obs2 <- dfr$sortindex.obs1 + 1
    dfr$sortindex.obs1[ dfr$sortindex.obs1==0 ] <- 1
    dfr$sortindex.obs2[ dfr$sortindex.obs2 > N1 ] <- N1
    dfr2 <- dfr[ dfr$obs==0, ]
    dfr2a <- dfr1[ dfr2$sortindex.obs1, grep( "donor", colnames(dfr1) ) ]
    dfr2a <- cbind( dfr2a, dfr1[ dfr2$sortindex.obs2, grep( "donor", colnames(dfr1) ) ] )
    dfr2 <- cbind( dfr2, dfr2a )
    gc <- grep( "donor", colnames(dfr2))
    d1 <- dfr2.abs <- abs( dfr2$yhat - dfr2[, gc ] )
    dfr2.abs[ is.na( dfr2.abs ) ] <- GG
    dfr2.gc <- dfr2[,gc]
    yhatmin <- wmin <- matrix( 0, nrow=sum(!ry), ncol=donors )
    for ( dd in seq( 1, donors ) ){
        wmin[,dd] <- apply( dfr2.abs, 1, min )
        I1 <- ( dfr2.abs==wmin[,dd] )
        dfr2.abs <- dfr2.abs + GG * I1
        yhatmin[,dd] <- rowSums( dfr2.gc * I1, na.rm=T) / rowSums(I1)
    }
    yhatM <- matrix( t(yhatmin), ncol=1, byrow=FALSE )
    wmin2 <- match( yhatM, dfr1$yhat )
    wmin2 <- dfr1[ wmin2, "index" ]
    wmin2 <- matrix( wmin2, ncol=donors, byrow=T )
    wmin <- wmin2
    samp <- sample( 1:donors, sum(!ry), replace=TRUE )
    ind <- wmin[,1] * ( samp==1 )
    if (donors >=2){
        for ( dd in seq( 2, donors )){
            ind <- ind + wmin[,dd] * (samp==dd )
        }
    }
    # create imputations
    imp <- (y[ry])[ ind ]
    imp <- imp[ order(dfr2$index) ]
    imp <- mice_imputation_factor_pmm_convert_factor(imp=imp,
                    is_factor=is_factor, y_aggr=y_aggr)
    return(imp)
}