Take a linear model with a block random effect ($b_i$) $$ y_{ij} = \mu + b_i + \beta x_{ij} + e_{ij} $$ with $b_i \sim N(0, \sigma_b^2)$, etc etc.

If $\sigma^2_b = 0$, then the model reduces to a very simple linear model. To test the null hypothesis $H_0: \sigma_b^2 = 0$ versus the alternative $H_1: \sigma_b^2 > 0$, one can use a likelihood ratio test. However, the limiting distribution under the null is not $\chi_1^2$, but rather $\tfrac{1}{2} \chi^2_0 + \tfrac{1}{2} \chi^2_1$, i.e., a 50:50 mixture of a point mass at zero and a chi square distribution with one degree of freedom.

How do we determine the critical value of (or the p-value using) that mixture?


The area to the right of any point above 0 is half that of a $\chi_1^2$. So to get a level $\alpha$ test, look up the $2\alpha$ point of a $\chi_1^2$.

.... as long as $\alpha < 0.5$.

Of course, p-values work similarly. Look the value up as if it were a $\chi_1^2$ and halve the resulting p-value.`

  • $\begingroup$ OUUUHHH, this was obvious!!! $\endgroup$ – ocram Mar 28 '13 at 10:30

Some R code if interested:

Using TcGSA package:

ss <- 0.3
sample_mixt <- TcGSA:::rchisqmix(n=1e5, s=0, q=1)
TcGSA:::pval_simu(s=ss, sample_mixt)

Using base:


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.