mice.impute.pmm3.RdThis function imputes values by predictive mean matching like
the mice::mice.impute.pmm
method in the mice package.
mice.impute.pmm3(y, ry, x, donors=3, noise=10^5, ridge=10^(-5), ...) mice.impute.pmm4(y, ry, x, donors=3, noise=10^5, ridge=10^(-5), ...) mice.impute.pmm5(y, ry, x, donors=3, noise=10^5, ridge=10^(-5), ...) mice.impute.pmm6(y, ry, x, donors=3, noise=10^5, ridge=10^(-5), ...)
| y | Incomplete data vector of length |
|---|---|
| ry | Vector of missing data pattern ( |
| x | Matrix ( |
| donors | Number of donors used for imputation |
| noise | Numerical value to break ties |
| ridge | Ridge parameter in the diagonal of \( \bold{X}'\bold{X}\) |
| ... | Further arguments to be passed |
The imputation method pmm3 imitates
mice::mice.impute.pmm imputation method
in mice.
The imputation method pmm4 ignores ties in predicted \(y\) values.
With many predictors, this does not probably implies any substantial problem.
The imputation method pmm5 suffers from the same problem. Contrary to
the other PMM methods, it searches \(D\) donors (specified by donors)
smaller than the predicted value and \(D\) donors larger than the
predicted value and randomly samples a value from this set of \(2 \cdot D\)
donors.
The imputation method pmm6 is just the Rcpp implementation
of pmm5.
A vector of length nmis=sum(!ry) with imputed values.
See data.largescale and data.smallscale
for speed comparisons of different functions for predictive mean
matching.
if (FALSE) { ############################################################################# # SIMULATED EXAMPLE 1: Two variables x and y with missing y ############################################################################# set.seed(1413) rho <- .6 # correlation between x and y N <- 6800 # number of cases x <- stats::rnorm(N) My <- .35 # mean of y y.com <- y <- My + rho * x + stats::rnorm(N, sd=sqrt( 1 - rho^2 ) ) # create missingness on y depending on rho.MAR parameter rho.mar <- .4 # correlation response tendency z and x missrate <- .25 # missing response rate # simulate response tendency z and missings on y z <- rho.mar * x + stats::rnorm(N, sd=sqrt( 1 - rho.mar^2 ) ) y[ z < stats::qnorm( missrate ) ] <- NA dat <- data.frame(x, y ) # mice imputation impmethod <- rep("pmm", 2 ) names(impmethod) <- colnames(dat) # pmm (in mice) imp1 <- mice::mice( as.matrix(dat), m=1, maxit=1, method=impmethod) # pmm3 (in miceadds) imp3 <- mice::mice( as.matrix(dat), m=1, maxit=1, method=gsub("pmm","pmm3",impmethod) ) # pmm4 (in miceadds) imp4 <- mice::mice( as.matrix(dat), m=1, maxit=1, method=gsub("pmm","pmm4",impmethod) ) # pmm5 (in miceadds) imp5 <- mice::mice( as.matrix(dat), m=1, maxit=1, method=gsub("pmm","pmm5",impmethod) ) # pmm6 (in miceadds) imp6 <- mice::mice( as.matrix(dat), m=1, maxit=1, method=gsub("pmm","pmm6",impmethod) ) dat.imp1 <- mice::complete( imp1, 1 ) dat.imp3 <- mice::complete( imp3, 1 ) dat.imp4 <- mice::complete( imp4, 1 ) dat.imp5 <- mice::complete( imp5, 1 ) dat.imp6 <- mice::complete( imp6, 1 ) dfr <- NULL # means dfr <- rbind( dfr, c( mean( y.com ), mean( y, na.rm=TRUE ), mean( dat.imp1$y), mean( dat.imp3$y), mean( dat.imp4$y), mean( dat.imp5$y), mean( dat.imp6$y) ) ) # SD dfr <- rbind( dfr, c( stats::sd( y.com ), stats::sd( y, na.rm=TRUE ), stats::sd( dat.imp1$y), stats::sd( dat.imp3$y), stats::sd( dat.imp4$y), stats::sd( dat.imp5$y), stats::sd( dat.imp6$y) ) ) # correlations dfr <- rbind( dfr, c( stats::cor( x,y.com ), stats::cor( x[ ! is.na(y) ], y[ ! is.na(y) ] ), stats::cor( dat.imp1$x, dat.imp1$y), stats::cor( dat.imp3$x, dat.imp3$y), stats::cor( dat.imp4$x, dat.imp4$y), stats::cor( dat.imp5$x, dat.imp5$y), stats::cor( dat.imp6$x, dat.imp6$y) ) ) rownames(dfr) <- c("M_y", "SD_y", "cor_xy" ) colnames(dfr) <- c("compl", "ld", "pmm", "pmm3", "pmm4", "pmm5","pmm6") ## compl ld pmm pmm3 pmm4 pmm5 pmm6 ## M_y 0.3306 0.4282 0.3314 0.3228 0.3223 0.3264 0.3310 ## SD_y 0.9910 0.9801 0.9873 0.9887 0.9891 0.9882 0.9877 ## cor_xy 0.6057 0.5950 0.6072 0.6021 0.6100 0.6057 0.6069 }