In the $\chi^2$ test of independence, the test statistic is defined over the contingency table as such:

$\chi^2 = \sum_{cells} \frac{(observed - expected)^2}{expected}$

Why is it defined like that? I know that the $\chi^2$ distribution is the sum of $N$ squared standard normal random variables. But I'm not sure the variable $\frac{(observed - expected)^2}{expected}$ is going to be normal. And if it won't be normal, then does the $\chi^2$ test even apply?

Although, the variable $\frac{(observed - expected)^2}{expected}$ looks suspiciously similar to $\frac{(x - \mu)^2}{\mu}$, which looks like the distribution of "normalized squared deviation from the mean". And that one is supposed to be a normal distribution, right?

From what I understand, the $\chi^2$ test of independence is trying to see if the distances between the $observed$ data points and the $expected$ data points follow normal distributions (because of the central limit theorem, maybe?). And because we define $expected$ to be data points generated by independent variables, then the $observed$ data points should only slightly deviate from $expected$, and the distances are following normal distributions. This means that the $\chi^2$ distribution can be used for testing. Is my understanding described here correct?

  • 1
    $\begingroup$ Am I asking too many questions into one? $\endgroup$
    – CamilB
    Jul 17, 2016 at 18:25
  • $\begingroup$ The question is probably answered here: stats.stackexchange.com/questions/82260/… $\endgroup$
    – CamilB
    Jul 17, 2016 at 19:10
  • $\begingroup$ I believe that's the same question, but I'm not so sure it (yet) has complete and correct answers. $\endgroup$
    – whuber
    Jul 17, 2016 at 19:19


Your Answer

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