I am working on a problem which revolves around individual axons / nerve cells in two treatment conditions. All relevant questions are clearly on the axon level, and dependent on properties (size, myelination, etc) of individual axons. However, I think that treating each axon as an independent observations without considering which animals they belong to would be cheating. Hence I would like to use a two-level approach with axons nested within animals.

Reading up on this it seems to me that I could use either a nested anova (aov in R with animal_id as a random effect) or a multilevel linear model (nlme / lme4 with a random intercept). But I'm having a hard time deciding between them (in practise, both actually produce fairly similar results).

Here's what I've learned so far:

  • I understand that nlme / lme4 use maximum likelihood methods which should make them more suitable for unbalanced designs but I do not think that this is a particularly big problem in my case.

  • I have also heard suggestions that nested anovas may be more suitable for smaller sample sizes. Is this true?

  • Finally, the developers of lme4 seems to be vehemently against the idea of assigning p values to their results. if I understand correctly this is because "degrees of freedom" is not a very straight forward concept in mixed linear models. But whenever I see people make this argument they also seem to be opposed to p values and null hypothesis testing in general. Is there a better case for trusting the p values in nested aov compared to lme4?

I am very grateful for any answer or comment.

Thank you


After having spent a few days in the library, I still feel far from understanding all implications of the two options. However, I have learned a few things and if anyone else is in the same boat, I would very much recommend the book "Multilevel Analysis" by Snijders and Bosker to guide your decision. It is by far the best text on this topic that I could find and spans from nested ANOVAs to very complicated multilevel linear models.

For me, it provided some clarification for my three points above:

1) Even though I have fairly balanced groups, my group sizes are small enough that changing a few values on the micro level can have a significant effect on the group variance. Multilevel linear models should be more stable in this situation because group effects of one group inform those of other groups. S&B emphasize that this effect may be particularly noticeable for groups with <50 individuals. [+1 MLM]

2) S&B suggest having at least 10, ideally 20+ groups in an MLM analysis. I am definitely at the low end of this spectrum. [+1 ANOVA?]

3) Yes, F-tests for nested ANOVAs are definitely much more straight-forward. S&B seem to believe that there is a case for p-values in MLMs, too but reading the book makes very clear why it is a lot more complicated. [+1 ANOVA]

I hope this helps someone. I am still very much unsure how to proceed so if anyone has a comment I'd be really happy to read it.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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