Does it make sense to fit a nested model if some nested factors only have one level?

Question

I am trying to build a meta-analytic model (using metafor package in R) to analyze effect sizes extracted from multiple studies (I have about 300 effect sizes from 50 studies). Studies may report effect sizes from different experiments and from different years within experiments, and I would like to analyse effect sizes using a model with the following random structure:

(1|Study/Experiment/Year)

However, not all of the studies report results from several experiments - some include only one experiment. Similarly, some experiments report results only for one year. In other words, for some levels of Study nested factors sometimes have only one level. Hence my question:

- A) Does it make sense to fit a nested random-effect model if some of the nested factors only include one level?

From what I understood from this post, I would say that this is not a problem, as long as the number of levels for the "top" factor (Study) is sufficiently high. However, I really am not sure of this.

Moreover, some Study:Experiment:Year combinations only include one observation (for some years in given experiments, only one effect size is reported). Thus, I have a second question:

- B) Does it make sense to fit a nested random-effect model if some levels only include one observation?

Answer 1

Your reasoning based on that other post seems fine to me.

It is perhaps worth adding that the package you are using is robust and will not fall over unless seriously provoked. If you include a random term which it cannot estimate this is usually fairly obvious from the values of $\sigma^2$. I would strongly suggest looking at the profile likelihoods for the random effects to see if they have in fact been based on adequate data. The profile.rma.mv() function is your friend here.