When using a mixed effect model the rule of thumb seems to be that you need at least 5 levels to use a random factor . Is this still True when you have a hieachical model. i.e A - 4 level factor B - 3 level factor C - 6 level factor

with a formula of


Will the first 2 levels just cause issues?


1 Answer 1


In a nested design, it is the combinations of the levels of the factors that is important.

y ~ 1 + (1|A/B/C)

expands to

y ~ 1 + (1|A) + (1|A:B) + (1|A:B:C)

So, you need to apply the rule of thumb to the levels of A, the levels of the interaction A:B and the levels of the interaction A:B:C. In the example you give, A has 4 levels, A:B has 12 levels and A:B:C has 72 levels.

It is debatable whether 4 levels is sufficient, and at the end of the day, pragmatism is probably the best approach:

  1. Does the data support that random structure? That is, does the model converge and is it non-singular ?
  2. Are the random effects plausibly normally distributed ?
  3. Does the model fit adequately ?
  • 1
    $\begingroup$ Thanks, I think the expansion was what I was missing. I'll investigate the other points. $\endgroup$
    – sam
    May 23, 2019 at 10:17

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.