1
$\begingroup$

I have two, large, two- or three-dimensional samples. I have used implementation of multidimensional Kolmogorov-Smirnov test in R (named Peacock.test) to compare these two samples and calculate test statistic. Is it any method to check null and alternative hypothesis without using p-value? How can I determine if the test statistic is in the critical area of the test?

Via Wikipedia, I can reject null hypothesis in two sample, one-dimensional KS test when: $$ D_{n,m} > c(\alpha) \sqrt{\frac{n+m}{nm}},$$ where $D_{n,m}$ is test statistic, $c(\alpha)$ I know from table, $n$ and $m$ are the sizes of samples. Is it works well for multidimensional Kolmogorov-Smirnov test?

$\endgroup$
  • $\begingroup$ You can never confirm it... $\endgroup$ – Tim Oct 5 '17 at 8:53
  • $\begingroup$ OK, so can I reject null hypothesis using this criterion? $\endgroup$ – Patryk Oct 5 '17 at 8:55
1
$\begingroup$

The Kolmogorov-Smirnov test uses an approximating null distribution called the Kolmogorov distribution, which can be used to obtain an approximation to the p-value. This is the same distribution that is used to construct the tabular values of critical points for the test.

Let $X_1, ..., X_n$ and $Y_1, ..., Y_n$ be independent random variables with respective continuous distribution functions $F_X$ and $F_Y$. Let $E_X$ and $E_Y$ be the respective empirical distribution functions of the samples. The Kolmogorov-Smirnov test statistic is:

$$D_{n,m} \equiv \sup_{t} | E_X(t) - E_Y(t) |.$$

Let $K(t) \equiv 1 - \sum_{k=1}^\infty (-1)^{k-1} \exp (-2k^2 t^2)$ be the CDF of the Kolmogorov distribution. Then it can be shown that the Kolmogorov-Smirnov statistic obeys the following asymptotic result:

$$\lim_{n \rightarrow \infty} \lim_{m \rightarrow \infty} \mathbb{P} \Bigg( \sqrt{\frac{mn}{m+n}} \cdot D_{n,m} \leqslant t \Bigg| F_X = F_Y \Bigg) = K(t).$$

Hence, the p-value can be approximated by the large sample distribution as follows:

$$\begin{equation} \begin{aligned} p \equiv p (d_{n,m} ) &\equiv \mathbb{P}(D_{n,m} \geqslant d_{n,m} | F_X = F_Y) \\[6pt] &= 1 - \mathbb{P}(D_{n,m} < d_{n,m} | F_X = F_Y) \\[6pt] &\approx 1 - K \Bigg( \sqrt{\frac{mn}{m+n}} d_{n,m} \Bigg) \\[6pt] &= \sum_{k=1}^\infty (-1)^{k-1} \exp \Bigg( -2k^2 \frac{mn}{m+n} d_{n,m}^2 \Bigg). \end{aligned} \end{equation}$$

This expression can be evaluated numerically by truncation at a large value of $k$, yielding an approximate p-value for the two-sample Kolmogorov-Smirnov test. Statistical software will calculate this probability for you.

$\endgroup$

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.