8
$\begingroup$

There has been talk on other questions of how one might use the Two One-Sided Tests (TOST) approach for the Kolmogorov-Smirnov (KS) test, but I was wondering whether it was possible to directly use the test statistic to show that two distributions were similar?

As far as I understand it, the KS test statistic represents the biggest difference between two CDFs, with the one-sample version being used originally as a goodness-of-fit test. This is shown in [1] as being when the empirical distribution crosses outside the confidence interval (i.e. any one point is too far from the hypothetical distribution they are testing against).

If the two-sample version is often used to show that two distributions are significantly different to one-another, in a similar way to the one-sample version, can we invert the calculation of the confidence intervals from using $(1-\alpha) = 0.05$ to instead use $(1-\alpha) = 0.95$, as a way of showing that the maximum difference between the two distributions is significantly similar?

[1] Massey, F. "The Kolmogorov-Smirnov test for goodness-of-fit", Journal of the American Statistical Association, vol. 46, no. 253, pp. 68-78, Mar 1951

$\endgroup$
3
$\begingroup$

When conducting the Kolmogorov-Smirnov test, we assume $H_0:$ the two distributions are equivalent. We then calculate a test statistic and, if the corresponding $p$-value is small enough, we reject $H_0$ and conclude $H_A:$ the two distributions are different.

As far as hypothesis tests go, we use a $p$-value to quantify the amount of evidence we have to reject the null hypothesis. A $p$-value of 1 indicates that we have gathered no evidence to reject the null hypothesis. A $p$-value close to 0 indicates there is overwhelming evidence to reject the null hypothesis.

Let's assume we have data and calculate a $p$-value from the K-S test where $p=0.99.$ This indicates there is very little evidence to reject the null hypothesis. However, we cannot establish a standard of $\alpha=0.95$ such that $p>\alpha$ implies that we conclude the null hypothesis is correct. Further, I don't believe there is an alternate test that would allow us to conclude that the two distributions are the same.

What I believe you can do is to be entirely honest in the write-up or discussion. Mention that you ran a K-S test, report a $p$-value, and if the $p$-value is sufficiently high, then articulate that there is very little evidence to suggest that the two distributions are different. So, while you cannot conclude that the distributions are identical, you should be able to note that there is no evidence suggesting that the two distributions are different. As your sample size $n$ increases, the more faith you'll have in this answer.

It's not quite the answer that you were probably looking for, but it's not a total wash, either. Hope this helps!

$\endgroup$
  • $\begingroup$ (1/2) I get that we can't just use a p value of > 0.98.. etc. -- what I was wondering is that, if the KS test is actually measuring the maximum distance between the two distributions, then it feels like it's in a unique position to perform equivalence testing directly. For example, if the test statistic D = 0, then surely the two distributions have to be identical? And if that's the case, what about D < 0.0001? It feels like there should be a critical point (like when p crosses 0.05), where we can no longer be certain that they are the same distribution, and $\endgroup$ – jamesyjamesjames Nov 16 '15 at 14:32
  • $\begingroup$ (2/2) ... so can flip that on its head and use 0 < D < 0.05 (for example) to say that the two distributions are significantly similar. $\endgroup$ – jamesyjamesjames Nov 16 '15 at 14:32
  • 1
    $\begingroup$ Even if $D=0$, we cannot conclude that the distributions are identical. We can only conclude that our sample gives us no evidence against the statement that the two distributions are the same. Perhaps our sample merely doesn't include observations that would indicate a difference in distribution. This is the same as all hypothesis tests - a test statistic that corresponds to a $p$-value of 1 provides no evidence against the null hypothesis but cannot definitively prove that the null hypothesis is correct. Practically speaking, as $n$ increases, you can become more confident that the... $\endgroup$ – Matt Brems Nov 17 '15 at 0:11
  • 1
    $\begingroup$ ...distributions are identical but you cannot conclusively establish this fact. $\endgroup$ – Matt Brems Nov 17 '15 at 0:11
  • 2
    $\begingroup$ Ah of course -- because it's an ECDF. Yeah that makes sense. $\endgroup$ – jamesyjamesjames Nov 18 '15 at 9:35

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.