I have a dataset with 14 clusters. Each cluster is a time series of 80 periods with autocorrealtion, and I am doing maximum likelihood estimation of a structural multinomial logit model. I suspect there is across-cluster correlation. Therefore, I am thinking two options:

  1. Driscoll-kraay standard errors. While I have read about how to calculate it in OLS (from Stata Journal), I am struggle thinking what the se formula is in ML using information matrix. Any learning material is much appreciated!

  2. Bootstrapping. I am not sure how I should do resampling when I have correlated clusters. I am wondering if the following resampling makes any sense:

    a. I first sample the 14 clusters with replacement.

    b. Then I divide the time series into $x$ blocks and sample (without replacement) $y$ continuous blocks for each selected clusters.

    I have some concerns about such resampling: my dataset consists of 14 subjects who repeatedly make group decisions. Theoretically group size and types of subjects in a groups can make a difference. So I am wondering if the first step will compromise the structural of original dataset.


1 Answer 1


It is important to explicitly model the autocorrelation. You might model cluster effects as random intercepts, but within-cluster effects need to be handled with something like a continuous-time AR(1) correlation structure.

  • $\begingroup$ But random intercepts cannot solve the cross-cluster correlation, can it? Maybe I understand it wrong.. $\endgroup$
    – jasmine
    Commented Nov 30, 2023 at 13:44
  • $\begingroup$ I haven’t seen that addressed. I would want to explicitly model the cross-cluster correlations using what is known about the study design. $\endgroup$ Commented Dec 1, 2023 at 12:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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