0
$\begingroup$

For coursework, I am doing a two-sided test ($H_0 \beta = 0, H_a \beta \neq 0$). The test itself is a generalized likelihood ratio. The test statistic is the ratio LR:

$$LR=\frac{L(\beta=0)}{argmax_{\beta \in R}L(\beta)}$$

Where L is the likelihood function.Then $-2ln(LR)$ follows a $\chi_1^2$.

I am trying to calculate the p-value for a specific value of LR. Say I look for at a $\chi_1^2$ table (or online calc) and find out that

$$ P(\chi_1^2 \geq -2ln(LR)) = \alpha $$

Is $\alpha$ my final p-value?

Or do I need to account for the fact that the test is two-sided and set the p-value $= 2*\alpha$?

I think I should do the later but I am a bit unsure. Some extra intuition would help.

Edit: added the LR function.

$\endgroup$
  • $\begingroup$ @LinAlg, tks. I realized that in general tests (of a transformation, standardization of X) we need to look at both sides. But in this case, we have a likelihood ratio of a constrained likelihood (in the numerator) divided by an unconstrained likelihood (in the denominator). Thus the smaller the value the greatest the evidence against H0. $\endgroup$ – LucasMation Nov 17 '18 at 20:39
  • $\begingroup$ By the way, I forgot to mention that -2ln(LR) is what has Chi^2 distribution, not LR itself. Will edit that now. $\endgroup$ – LucasMation Nov 17 '18 at 20:40
  • $\begingroup$ @LinAlg, please incorporate my point avobe into your answer so I can mark it as a solution to the question. $\endgroup$ – LucasMation Nov 17 '18 at 20:41
  • $\begingroup$ could you explicitly add what likelihood was in the denominator? $\endgroup$ – LinAlg Nov 17 '18 at 21:05
  • $\begingroup$ @LinAlg see above $\endgroup$ – LucasMation Nov 19 '18 at 2:48
0
$\begingroup$

The $p$ value is related to the rejection region, so the answer depends on how you divide $\alpha$ over both sides for the test. If you reject H0 when the test statistic is less than $F(\alpha/2)$ or more than $F(1-\alpha/2)$, then $2\alpha$ is the right answer. However, you can use any rejection region $(-\infty,a] \cup [b,\infty)$ as long as $a$ and $b$ satisfy $F^{-1}(b) - F^{-1}(a) = 1-\alpha$. For skewed unimodal distributions you can impose $f(a)=f(b)$, which leads to a different answer.

In your example of a generalized likelihood ratio test, the rejection region is an interval $[b,\infty)$, even though the test is two sided. So $\alpha$ is the p-value.

$\endgroup$

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.