0
$\begingroup$

I am running multilevel models for panel data on a binary outcome (mixed logistics regression) and on a ordinal outcome (mixed ordinal logistic regression). I am aware that for example with a mixed linear model, the level-2 variances should be normally distributed. What are such assumptions that should be checked for mixed logistic and ordinal regressions?

$\endgroup$

1 Answer 1

1
$\begingroup$

The most basic assumption to check is whether the correlation pattern your model assumes is consistent with the study design. If your response variables are serially collected over a time span than is long with respect to what's happening, then you need to take into account the forward flow of time by modeling serial correlation (e.g. Markov binary or ordinal logistic model). Multilevel models traditionally incorporate only exchangeable random effects, so they assume compound symmetry correlation patterns that are bidirectional in time and assume that the correlation between two measurements within subject are the same no matter how far apart the measurements are taken.

Random effects models are more appropriate for pure repeated measures without a time component, e.g., re-test someone every 5 minutes for 4 tests in all with a total time span of 20 minutes.

For more about Markov vs. random effects models see links in https://hbiostat.org/proj/covid19 especially the VIOLET 2 and ORCHID studies, and see these references and these.

$\endgroup$
2
  • $\begingroup$ Hi Frank. So, for my particular design I've got a group of individuals who following a baseline agreed to be re-contacted from 2018 to 2020, whereby a follow up was conducted. Am I correct in thinking that a multilevel model would be suitable for this, such that the data is clustered by participant. In which case how would I go about checking that "compound symmetry correlation patterns that are bidirectional in time and assume that the correlation between two measurements within subject are the same no matter how far apart the measurements are taken."? $\endgroup$
    – gogo123
    Commented Jan 22, 2021 at 16:58
  • 2
    $\begingroup$ My guess is that infrequent measurements collected over a long time span such as yours are more likely to be consistent with the use of random effects to capture correlation patterns. You can examine correlations to help check this, e.g. if you assess the subject yearly for 3 years as Y1 Y2 Y3 check correlation between Y1 and Y2 against correlation between Y1 and Y3 to see if they are roughly the same magnitudes. $\endgroup$ Commented Jan 22, 2021 at 18:39

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.