I'm trying to estimate the design effect of a series of relatively small sample size surveys ($n\sim 70$) with multiple responses. Design effects roughly correspond to how much larger actual sample variance than would be expected from naive random sampling. The simplest way to parametrize this is for Effective Sample Size to be equal to SampleSize/D, where D is an unknown parameter.

Throwing it naively into BUGS is impossible due to the inability to handle multinomial models with unknown sample size. I tried to do a multinormal approximation with a structured covariance matrix, but I can't figure out/find a closed form for the precision matrix of a multinomial variable.

Does anyone have any ideas about the best way to proceed? This seems like it should be a common problem.


Better late than never...

The covariance matrix has diagonal entries $np_i(1-p_i)$ and off-diagonal entries $-np_ip_j$. JAGS and BUGS allow you to invert a matrix numerically (sigma[1:K, 1:K] <- inverse(tau[,]) in WinBUGS), so you don't actually need a closed-form expression for the precision matrix.

Your approach doesn't sound unreasonable, if your sample size is large enough. An alternative would be to just run a bunch of models with different specified effective sample sizes, and pick the one with the best deviance. A manual implementation of golden section search, if you will.

| 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.