What are the consequences of replacing missing data with median?
The conclusion is that it artificially reduced the variability of the correlation coefficient. It is bad practice. But it is much better than doing nothing ! cor.original <- NULL cor.na <- NULL cor.median <- NULL for (i in 1:10000) { A <- rnorm(100, mean=100, sd=20) B <- rnorm(100, mean=100, sd=20) Bprime <- ifelse(sample(c(0,1), 100, replace = TRUE), B, NA) Bter <- ifelse(is.na(Bprime), median(B, na.rm = TRUE), Bprime) cor.original <- c(cor.original, cor(x=A, y=B, method = "spearman")) cor.na <- c(cor.na, cor(x=A, y=Bprime, method = "spearman", use="complete.obs")) cor.median <- c(cor.median, cor(x=A, y=Bter, method = "spearman", use="complete.obs")) } layout(1:3) hist(cor.original, xlim=c(-0.6, 0.6), breaks=seq(from=-0.6, to=0.6, by=0.05)) hist(cor.na, xlim=c(-0.6, 0.6), breaks=seq(from=-0.6, to=0.6, by=0.05)) hist(cor.median, xlim=c(-0.6, 0.6), breaks=seq(from=-0.6, to=0.6, by=0.05)) quanti