1
$\begingroup$

I am trying to choose between GEE and hierarchical linear regression for analysis of experimental vignette (2x2 factorial (0/1) design) data. Each respondent (N=160) filled in 2 vignettes, thus the data is nested within respondents (within-subjects design). I want to test the effect of the two factors (and their interaction) on three dependent/outcome variables (scale, 1-10).

I am finding it difficult to decide as in my view hierarchical linear modelling would work, but my supervisor prefers GEE (as there are no distributional assumptions). However, as far as I understand, using GEE I won't get goodness-of-fit measure and no subject-specific estimates.

Do you have any views on which approach would be better?

Improve this question
$\endgroup$
2
  • $\begingroup$ Thank you! I've now managed to fit an HLM model with clustering at the respondent level, working with intra-class correlation coefficients to estimate the respondent-level variance (rather high, but stable across model levels). The notes were helpful! $\endgroup$
    – GabrieleC
    Commented Dec 15, 2021 at 10:33
  • $\begingroup$ If you found this answer helpful, then please consider upvoting and/or accepting it. $\endgroup$
    – kjetil b halvorsen ♦
    Commented Dec 15, 2021 at 12:05

1 Answer 1

Reset to default
3
$\begingroup$

GEE makes distributional assumptions and is an asymptotic approach, i.e., may require very large N to be accurate. It's usually worth the effort of carefully specifying the within-subject correlation pattern and using a full maximum likelihood estimation procedure. Such procedures include mixed effects models and marginal models (marginal over subjects but modeling serial correlation patterns, for example, such as generalized least squares or Markov models). Tradeoffs of different approaches are described in a table in the longitudinal modeling chapter of RMS course notes.

Improve this answer
$\endgroup$

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.