I'm interested in using a Poisson approximation for modeling a common (~40% of the time) binary outcome. But my data has some clustering in it, having come from three different sites, which suggests the use of a random effects model.

Getting the actual estimates are easy enough, using something like lme4 in R:

model <- glmer(Outcome ~ Sex + Age + (1|Site), data=study, family=poisson)

The problem is that Zou G. A modified Poisson regression approach to prospective studies with binary data. Am J Epidemiol. 2004;159(7):702–706 suggests that a robust variance estimator is needed using this method, but the standard sandwich library doesn't work (by, as I understand it, design) with lme4.

Which suggests two questions:

  1. Is a robust variance estimator still needed in this circumstance?
  2. How does one obtain one?

I would model the sites as fixed effects: you should have the power to do so and it theoretically achieves the same effect as including a random effect. Adjusting for site as either a fixed or random effect makes participants within the same site conditionally independent of each other. The random effects variance estimates becomes a little unwieldy when nearly 50% of the off-diagonal covariance matrix entries are non-zero.

If the model has entirely fixed effects it is amenable to estimation with a GEE and that solves the problem easily.

If you had more sites, another approach would be using the robust sandwich- variance estimate to the generalized least squares model using an exchangeable correlation structure for site-based clusters. Random intercepts and exchangeable correlation lead to equivalent estimates when the intracluster correlation is positive and that's usually the case. Sandwich-variance calculation can be done by-hand as shown here.


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.