1
$\begingroup$

I am using the project jerzy to run a Two-sample Kolmogorov-Smirnov test in Javascript, regarding another question I asked on stats.SE: Timing attacks: When the time to complete two different tasks are statistically indistinguishable.

Not knowing anything about statistics, I would be grateful for some assistance interpreting the results. The code for the test being run is on Github (pieterprovoost/jerzy).

I am in particular running the test like this (and the numbers are contrived):

 var diff = [750, 740, 790]; // array of nanosecond results
 var equal = [750, 610, 960]; 
 results.diff = new jerzy.Vector(diff);
 results.equal = new jerzy.Vector(equal);
 var ks = new jerzy.Nonparametric.kolmogorovSmirnov(diff, equal);
 console.log(ks);

The output I get from the real data is something along the lines of:

 { d: 0.032657926102502954,
   ks: 0.6660700343005954,
   p: 0.7667168595417211 }

In the real tests the diff and equal are arrays of nanosecond timings. I would like to establish with some confidence that the arrays are effectively from the same distribution, with a difference of around 15ns.

How would one interpret the above result of the kolmogorovSmirnov function of jerzy, in terms of how strongly one might state the probability and confidence that the two arrays are from the same distribution?

$\endgroup$
4
$\begingroup$

The null hypothesis of the two-sample Kolmogorov–Smirnov test is that the two datasets are coming from the same distribution. The test is essentially trying to reject the null hypothesis, and, if it fails to do so, the alternative hypothesis is accepted.

The decision (to reject the null hypothesis) is based on the p-value computed for the given data; this is what the $p$ attribute of your JavaScript object stands for. Before performing the test, one typically decides on the significance level suitable for the problem at hand; this level is conventionally denoted by $\alpha$ and quite often chosen to be 0.05. Then the null hypothesis is rejected if

$$ \alpha \geq p. $$

So, in your case, the test, having at its disposal only three points in each dataset, has failed to reject the null hypothesis at significance level 0.05. In order to reject the hypothesis, the test might need more points.

The $d$ attribute provided by jerzy is the uniform distance (maximal pointwise distance) between the empirical CDFs computed for the two datasets, and $ks$ is $d$ multiplied by a factor related to the two-sample Kolmogorov–Smirnov test.

Lastly, the Kolmogorov–Smirnov test does not provide any confidence intervals that you are asking for. Some other test might be better suited if you need confidence intervals.

$\endgroup$
  • $\begingroup$ Many thanks @Ivan. So the p value is the ordinary kind - good to know (not sure why I'd have expected otherwise...). I am still hazy on the d and ks attributes - as soon as I figure them out I'll likely mark this answer as correct; that said, I believe it would improve this answer if their explanation was expanded upon, too. Cheers $\endgroup$ – Brian M. Hunt Feb 18 '15 at 16:35

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.