I am analysing data from a small study. There are two treatment groups and for each, one notes success and failures. For one of the groups, the number of successes is 0. I did an exact significance test, but reviewers of the paper want a confidence interval for the relative risk - in other words, a confidence interval for $p_1/p_2$, where $p_1=0$.

What should I do? The usual formula assumes that neither proportion is 0 or 1.

| cite | improve this question | | | | |
  • 1
    $\begingroup$ Just to be clear, it seems you want a CI for $p_1/p_2$ where the estimate $\hat{p}_1$ is $0$. Because you did an exact significance test, you should be able to use essentially the same calculations to obtain an exact confidence interval. What formula, then, did you use? $\endgroup$ – whuber Sep 12 '13 at 20:57
  • $\begingroup$ I used Fisher's exact test ... but I don't see how I would turn that into an expression in p1/p2. I am tending to the opinion that it makes no sense to put a CI around a relative risk where one of the risks is 0. $\endgroup$ – Placidia Sep 13 '13 at 14:05
  • $\begingroup$ Why should a basic concept like a confidence interval fail to apply for specific values of the data? The answer is that it cannot: a CI makes perfect sense no matter what you observe. $\endgroup$ – whuber Sep 13 '13 at 14:21
  • $\begingroup$ Good point ... but I am still trying to put a confidence interval around the true porportion p1/p2, whatever that is, when the estimate of p2 is 0. $\endgroup$ – Placidia Sep 13 '13 at 16:32
  • $\begingroup$ That's right (which is why I have upvoted this question). You can compute a CI, and there are many ways to do it ranging from using a generalized linear model to a permutation test. I therefore anticipate seeing some interesting and useful answers appear. $\endgroup$ – whuber Sep 13 '13 at 16:55

Thanks to @whuber's suggestions, I think I know what to do. The confidence interval (or region) contains those parameter values that would not be rejected were they the subject of a null hypothesis. I used a Fisher exact test and since the data represent the most extreme possibility (no failures in one arm), the test statistic for a given pair of parameters is simply the likelihood evaluated at those points. I can evaluate the likelihood over a grid of parameter pairs and select those with likelihood > 0.05 (i.e. not in the rejection region, were I doing a test). From these, I can calculate the ratio and the difference. The spread of values gives me the confidence intervals I need for the parameter functions of interest: relative risk and absolute difference.

| cite | improve this answer | | | | |

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.