I am learning about chi tests and I've understood everything so far except how to calculate critical values. I don't want to take number of degrees of freedom and probability of exceeding the critical value(alpha) and look it up in the table. I want to learn how to calculate the critical value.

  • $\begingroup$ As Robert says, trying to compute the critical values by hand is a tall order, unless you've somehow devised a nice method to easily approximate the incomplete gamma function. Nevertheless, you might be interested in looking up the Wilson-Hilferty transformation, which approximates the $\chi^2$ distribution with a normal distribution. $\endgroup$ – J. M. is a poor mathematician Sep 20 '12 at 18:32
  • $\begingroup$ +1 @J.M. for pointing to a classic paper from the 1930s. Good ideas never die! $\endgroup$ – Michael Chernick Sep 20 '12 at 19:54

The chi-square distribution with $\nu$ degrees of freedom has cumulative distribution function $$F(x) = 1 - \frac{\Gamma(\nu/2,x/2)}{\Gamma(\nu/2)}\ \text{for} \ x > 0$$ For each positive integer $\nu$ this can be written in terms of powers of $x$, $e^{-x/2}$, and (if $\nu$ is odd) $\text{erf}(\sqrt{x/2})$. The upper-tail critical value at level $\alpha$ is the solution of $F(x) = 1 - \alpha$, while the lower-tail critical value at level $\alpha$ is the solution of $F(x) = \alpha$. This can't generally be done in closed form, but you can use numerical methods.

EDIT: In the case $\nu = 1$, $F(x) = \text{erf}(\sqrt{x/2})$, so the solution of $F(x) = \alpha$ is $x = 2\; (\text{erf}^{-1}(\alpha))^2$.

In the case $\nu = 2$, $F(x) = 1 - e^{-x/2}$ for $x>0$, so the solution of $F(x) = \alpha$ is $x = - 2 \ln(1-\alpha)$.

In the case $\nu = 4$, $F(x) = 1 - (1 + x/2) e^{-x/2}$ for $x > 0$, and the solution of $F(x) = \alpha$ is $x = -2\; \text{LambertW}(-1,-(1-\alpha)/e) - 2$.

Those are the only cases, as far as I know, where the solution can be done in any sort of "closed form".

  • $\begingroup$ The \Gamma(a,x) function in Robert's answer is known as the incomplete gamma function; one would be hard-pressed to numerically evaluate this function, or its inverse, by hand. $\endgroup$ – J. M. is a poor mathematician Sep 20 '12 at 18:29

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.