I think this question is similar to that regarding the minimum number of studies required for running a meta-analysis, because when you're running categorical moderation, you compute two (or more) effect sizes (depending on the modalities of the moderator), then you compare them. Of course that when one have a small number of studies (i.e., at least two) included in each moderator category, the computed effect sizes can be unstable, and thus the information obtained based on moderation analysis can have limited utility...

I don't know that you'll find a specific citation for this, but it is not uncommon for researchers to set a minimum number. The problem is that this number is essentially arbitrary. I think the main consideration is the power of the moderator test. I also think that the extent to which the moderator covaries with other study characteristics also matters a lot (e.g., if level A of a moderator tends to occur in studies of one type, and level B of the moderator tends to occur in studies of another type, this is problematic, and it might be more of a problem as the number of studies decreases).

I think this question is similar to that regarding the minimum number of studies required for running a meta-analysis, because when you're running categorical moderation, you compute two (or more) effect sizes (depending on the modalities of the moderator), then you compare them. Of course that when one have a small number of studies (i.e., at least two) included in each moderator category, the computed effect sizes can be unstable, and thus the information obtained based on moderation analysis can have limited utility...

Later edit: here you have a reference where the author recommends 4 studies per moderator category - http://www.jclinepi.com/article/S0895-4356(11)00029-1/abstract