In the book "All of statistics" (Wasserman 2005) they present the likelihood ratio hypothesis test, using the likelihood ratio statistic: $$\lambda=2\log{\Lambda\left(x^{n}\right)\triangleq2\log {\frac{\sup_{\theta\in\Theta}L\left(\theta\right)}{\sup_{\theta\in\Theta_{0}}L\left(\theta\right)}}}=2\log {\frac{\mathcal{L}(\hat\theta)}{\mathcal{L}(\hat\theta_0)}}$$ They also present the Wilks theorem stating that under $H_0:\theta\in\Theta_0$: $$\lambda\left(X^{n}\right)\triangleq2\ln\Lambda\left(X^{n}\right)\rightsquigarrow\chi_{d}^{2}$$ where $d$ is the dimension difference of $\Theta$ and $\Theta_0$.

However, it seems extremely odd that a Chi-square RV can take negative values (when the likelihood for $H_0$ is larger than the likelihood of $H_1$).


1 Answer 1


Checked the original paper "The Large-Sample Distribution of the Likelihood Ratio for Testing Composite Hypotheses" (Wilks 1938).

Turns out they limit this theorem to cases where $\Theta_0$ is a subset of $\Theta$, implying $\frac{\sup_{\theta\in\Theta}L\left(\theta\right)}{\sup_{\theta\in\Theta_{0}}L\left(\theta\right)}\ge 1$ and thus $\lambda(X^n)\ge 0$.


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.