4
$\begingroup$

I have previously asked about a way to test whether two samples are drawn from the same distribution (Non-parametric test if two samples are drawn from the same distribution). I was very glad to learn about the Kolmogorov–Smirnov test. This seems excellent for hypothesis testing with p-values.

Now I am interested in generalizing this to a visual comparison of samples using confidence bands, and I wonder if my idea is valid and if so then how I can implement it with R, and if not then what a better approach might be.

The idea is to visually plot CDF confidence bands for each of N samples. This should show two things. First, the "tightness" of the confidence bands will suggest how adequate the sample sizes are for drawing inferences. Second, anywhere the confidence bands for a pair of samples do not overlap will indicate a significant difference in the KS-statistic i.e. a point where one CDF seems to be "greater than" or "less than" another.

Does this make sense? I was hoping to find an example graph on Google Images but the nearest I found is with only one sample:

ECDF with confidence band

My graph would be similar except that I would show N samples and look for blank spaces between the confidence bands.

If indeed this sounds reasonable then I have a few specific questions:

  • Is this a well-known technique? (Reference to a description?)
  • How should I compute the CDF confidence bands? (Is there an R package that I should use?)
  • How valid is the visual comparison of confidence bands? (If I plot two samples then how closely will a non-overlapping point correspond with a significant result from the KS test?)
  • What pitfalls should I keep in mind when comparing more than two samples? (Should I adjust the parameters of my confidence bands in the spirit of the Bonferroni correction?) Note: I am okay with the comparison being more informal than a p-value.

Thanks!

UPDATE

I found an example of the kind of graph that I have in mind now:

graph

I will try to work out how to create such a graphic using ggplot2 and the ecdf.ksCI() function recommended in the comment.

UPDATE 2

Thank you for the answer below. I was able to generate my graph and here is a sample. I am pleased!

CDF with confidence bands

$\endgroup$
  • $\begingroup$ See: stat.ethz.ch/pipermail/r-help/2003-July/036643.html $\endgroup$ – kjetil b halvorsen Aug 16 '17 at 21:33
  • 1
    $\begingroup$ Thanks for the tips! I read a presentation (web.as.uky.edu/statistics/users/pbreheny/621/F10/notes/8-26.pdf) that talks about the difference between pointwise confidence intervals verses confidence bands/envelopes for whole CDFs. They seem to advocate the latter as being more appropriate. Is this relevant to the discussion? (Is ecdf.ksCI providing a pointwise confidence interval and will this have practical consequences e.g. more type I/II errors?) $\endgroup$ – Luke Gorrie Aug 17 '17 at 9:49
  • $\begingroup$ I have tested the ecdf.ksCI function now. It's pretty neat. Can it also somehow display N samples together for comparison? (Or is there a straightforward way to adapt it to ggplot2 and geom_ribbon?) $\endgroup$ – Luke Gorrie Aug 17 '17 at 14:42
  • $\begingroup$ You would have to get the code for the function and use it as a starting point. As it is written now it only produces the plot, it returns nothing. $\endgroup$ – kjetil b halvorsen Aug 17 '17 at 19:44
4
$\begingroup$

You can use the Kolmogorov-Smirnov test, and invert it to get a confidence band. Let $X_1, X_2, \dotsc, X_n$ be iid observations from some continuous distribution function $F$. Then the KS test statistic is given by $$ D_n = \sup_x \mid \hat{F}_ n(x)-F_0(x) \mid = \max_{i=1,2,\dotsc,n} \max \{\frac{i}{n}-F_0(x_{(i)}),F_0(x_{(i)})-\frac{i-1}{n} \} $$ where $x_{(1)} \le \dotso \le x_{(n)}$ is the order statistics. What is remarkable is that the distribution of $D_n$ do not depend on the assumed null distribution $F_0$ (which must be prespecified). Now we can invert this hypothesis test to get a confidence band. WE can calculate $$ P_{F_0}(D_n \le d) = P_{F_0}( \sup_x \mid \hat{F}_ n(x)-F_0(x) \mid \le d) = \\ P_{F_0}( \hat{F}_n(x)-d \le F_0(x) \le \hat{F}_n(x)+d, \quad \text{for all $x$}) $$ this calculation shows that this is indeed a simultaneous confidence band, valid simultaneously for all $x$.

An implementation of this can be found in the R package (on CRAN) sfsmisc, in the function ecdf.ksCI. (Disclaimer: That was originally my code)

$\endgroup$
  • 1
    $\begingroup$ Thank you very much. Reading the source code of the ecdf.ksCI function was very instructive and lead me to an implementation that suits my own purposes. $\endgroup$ – Luke Gorrie Aug 20 '17 at 8:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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