Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I have two matched groups consisting of 22 observations each for a Poisson distributed variable. I am interested in evaluating the power to detect a 20% decrease in the outcome for the second group. I have no idea of the correlation between matched groups and hence will have to assume independence. I am planning on evaluating the power using a simulation study and have the following ideas in mind.

1) Define rate0 and rate1 = rate0-(0.20*rate0)

2) Simulate X1~Poisson(22,rate0) and X2~Poisson(22,rate)

3) Compute pairwise differences X1-X2 and conduct a paired Wilcoxon test or a paired ttest.

4) Conduct simulation nsim times and compute the times pvalue < 0.05

I would appreciate some feedback/suggestions regarding step 3. Would anyone have any other ideas?

share|improve this question

The power analysis by simulation is ok, I think what you’re really asking for is a way to compare matched Poisson variables, other than Wilcoxon test or paired t-test.

  • A brute force approach would be: use as test statistic $\sum_i X_{1i} - X_{2i}$ ; assume H0 (same rate in two groups), estimate the common rate $\lambda$ using pooled data, and simulate N (N = big) times two groups of 22 variables $\sim \mathcal P(\lambda)$ to get an empirical distribution of you test statistic.

  • If your rates are big enough, you could also use the fact that if $X\sim\mathcal P (\lambda)$, then the distribution of $2 \sqrt X$ is approximated by a normal $\mathcal N(2\sqrt\lambda,1)$. This leads to normal based test (the distribution of $X_{1i} - X_{2i}$ is known under $H_0$, independently of $\lambda$), and can lead to a nice paper-pen computation to the power. The figure below illustrate the (in)accuracy of this approximation for various $\lambda$ (in black, the cdf of $2 \sqrt X$, in red, the normal approximation).


share|improve this answer
+1 Good answer. Simulations suggest that a paired t-test of $\log(X+1/2)$ may be slightly more powerful still, especially for small rates $\lambda$. (The result is not very sensitive to the offset $1/2$; it's there to handle zero values of $X$.) – whuber Jan 2 '12 at 15:33
An additional comment is that in all these tests, the null hypothesis is in fact $X_{1i}, X_{2i} \sim \mathcal P(\lambda_i)$: the rate is the same for the two members of the pair, but may differ from one pair to an other. – Elvis Jan 2 '12 at 16:59
That seems to be a generalization of the question itself, Elvis, and is much more problematic to solve. The question refers to "22 observations each for a Poisson distributed variable." Although that could be read as allowing the parameter to vary from one observation to another, that would be an unusual interpretation. – whuber Jan 2 '12 at 17:51
Thank you all. This has been extremely helpful. I am concerned about transforming mainly because of the difficulty in determining the standard errors for the mean difference estimate in its original scale. – user1106711 Jan 2 '12 at 21:50
@user1106711 I am glad this helped. I have been a little cryptic on the construction of the test: please consider that as each $2\sqrt{X_{1i}} - 2\sqrt{X_{2i}} \sim \mathcal N (0, 2)$, the sum $2\sum_i \sqrt{X_{1i}} - \sqrt{X_{2i}} \sim \mathcal N (0, 2n)$ and you can do a $z$-test with known variance $2n$, you don’t need to estimate the variance from the data! In my opinion, this is the main strength of this variance stabilizing transformation. – Elvis Jan 3 '12 at 10:15

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.