Multilevel modeling when observations are clustered but clusters are not random

Question

I am working on a project where I am trying to model a continuous outcome, y, in an organizational setting. In this particular organization, employees are nested within teams, which are nested within division. I am willing to concede that employees are random because many join and leave the organization. However, the organizational structure is not random. In that, the organization - team structure both has and will remain constant. Given all of this, I would still like to make an inference with regards to the amount of variance present at each cluster. Additionally, I have plenty of observations at each cluster, so the cluster level "sample" is not an issue. From what I have read there appears to be a blurry line with regards to when MLM should be used.

There are 33 divisions with an average of roughly 5 teams per division.

So in sum, my question is whether or not an MLM would be appropriate in this situation given that observations are clearly clustered, yet the clusters are not technically random.

Answer 1

It is often remarked that clusters should be modelled with random intercepts when we can consider the clusters to be samples from a wider population of such clusters. However, this is just one of several criteria to be used when deciding whether to model a factor as fixed or random. Others include:
- the number of levels of the factor
- whether there is interest in the "effect" of the factor itself

With 5 teams on average per division, and with 33 divisions, the first of these is not an issue.

As for the second, nothing in the OP suggests that the "effects" of team and division are of interest.

I see nothing wrong in fitting a mixed model to these data, with team nested within division and computing a variance partition coefficient (which is stated as an objective of the study).