Given a histogram obtained using given data points, how do I randomly sample from the distribution predicted by the histogram?

Any conceptual comment / R code would be welcome.

  • 2
    $\begingroup$ Your title seems to be asking how to sample from a histogram-as-population-pdf while the body text seems to be asking how to sample from a kernel-density-estimate-as-population-pdf (two different problems, the second of which is solved here and here). $\endgroup$
    – Glen_b
    Commented Jan 21, 2016 at 9:50

2 Answers 2


Since the sampling from a kernel density estimate is solved once or twice already, I'll focus on sampling from a histogram-as-population-pdf.

The idea is simply

For each observation in the new sample

  1. choose a histogram bin according to the proportions of 
     the original sample (treated as a discrete pmf)

  2. sample uniformly from that bin-interval

For example in R:

#create an original histogram

#sample from it
bins=with(xhist,sample(length(mids),samplesize,p=density,replace=TRUE)) # choose a bin
result=runif(length(bins),xhist$breaks[bins],xhist$breaks[bins+1]) # sample a uniform in it

Just for completeness, (since sampling from the kernel density estimate* is very simple):

repeat nsim times:
  sample (with replacement) a random observation from the data
  sample from the kernel, and add the previously sampled random observation

* note that some kernels - like fourth order kernels - are not densities and this assumes that the kernel is a density

In R, for a Gaussian kernel and bandwidth h, with data in x:

 dnorm(nsim,m=sample(x,nsim,replace=TRUE), s=h)

I think what you want is:

y <- density(x)
x.new <- rnorm(length(x), sample(x, size = length(x), replace = TRUE), y$bw)
lines(density(x.new), col = "blue")

Note that the density function uses kernel = "gaussian" as default. So to generate gaussian random numbers from a particular density you just need to use rnorm with the mean values equal to the original series and the standard deviation equal to the smoothing bandwidth. In this example I am resampling those mean values to generate different simulations. You can see this example in the density function documentation here.

  • $\begingroup$ Can you provide some text to explicate this code? $\endgroup$ Commented Jan 21, 2016 at 11:01
  • $\begingroup$ Sorry, I've added some clarification. Note that this example is taken from the density function documentation. $\endgroup$ Commented Jan 21, 2016 at 12:13
  • $\begingroup$ As far as I can see you're producing a sample from a particular Gaussian. That's not in general the distribution predicted by any histogram (or kernel density estimate). There seems to be confusion here between the kernel and the data. $\endgroup$
    – Nick Cox
    Commented Jan 21, 2016 at 17:41
  • $\begingroup$ I had the same feeling when I read the documentation of the density package, but they seem to justify it by saying that a kernel density fit is an equally-weighted mixture. $\endgroup$ Commented Jan 21, 2016 at 18:24
  • $\begingroup$ I don't use R except very occasionally, so I am guessing what the code. But the bandwidth of a kernel will usually be much less than the SD of the data. Either way, the assumption that the data are Gaussian will in general be quite false. $\endgroup$
    – Nick Cox
    Commented Jan 21, 2016 at 18:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.