Clip outliers in columns in df2,3,4... based on quantiles from columns in df.tr

Question

I am trying to replace the "outliers" in each column of a dataframe with Nth percentile.

n <- 1000
set.seed(1234)
df <- data.frame(a=runif(n), b=rnorm(n), c=rpois(n,1))
df.t1 <- as.data.frame(lapply(df, function(x) { q <- quantile(x,.9,names=F); x[x>q] <- q; x }))

I need the computed quantiles to truncate other dataframes. For example, I compute these quantiles on a training dataset and apply it; I want to use those same thresholds in several test datasets. Here's an alternative approach which allows that.

q.df <- sapply(df, function(x) quantile(x,.9,names=F))
df.tmp <- rbind(q.df, df.t1)
df.t2 <- as.data.frame(lapply(df.tmp, function(x) { x[x>x[1]] <- x[1]; x }))
df.t2 <- df.t2[-1,]
rownames(df.t2) <- NULL
identical(df.t1, df.t2)

The dataframes are very large and hence I would prefer not to use rbind, and then delete the row later. Is is possible to truncate the columns in the dataframes using the q.df but without having to rbind? Thx.

Answer 1

So just write a function that directly computes the quantile, then directly applies clipping to each column. The <- conditional assignment inside your lapply call is bogus; you want ifelse to return a vectorized expression for the entire column, already. ifelse is your friend, for vectorization.

# Make up some dummy df2 output (it's supposed to have 1000 cols really)
df2 <- data.frame(d=runif(1000), e=rnorm(1000), f=runif(1000))

require(plyr)
print(colwise(summary)(df2))  # show the summary before we clamp...

# Compute quantiles on df1...
df1 <- df
df1.quantiles <- apply(df1, 2, function(x, prob=0.9) { quantile(x, prob, names=F) })

# ...now clamp by sweeping col-index across both quantile vector, and df2 cols
clamp <- function(x, xmax) { ifelse(x<=xmax, x, xmax) }
for (j in 1:ncol(df2)) {
  df2[,j] <- clamp(df2[,j], df1.quantiles[j])  # don't know how to use apply(...,2,)
}

print(colwise(summary)(df2))  # show the summary after we clamp...

Reference: [1] "Clip values between a minimum and maximum allowed value in R"