Theoretically, you can use a three way PERMANOVA analysis for an unbalanced design. Thar is my answer. However, I am quite confuse in your case. Could you explain it better? At a first glance, it seems you have two factors: habitat and hydrological phase. How many replicates you have for each combined factor (e.g. lake1, low-water phase)?

Hi Vinicius,

I have three factors: locality: 5 levels (pond 1, pond2, pond3, pond4, stream1), habitat: 3 levels (temporal ponds,  permanent ponds, streams) and finally hydrological phases: 4 levels (low waters, rising waters, high waters and falling waters). 

The problem is that for the habitat factor, I have two replicates for temporal ponds, two for permanent ponds but only one for streams. For each hydrological phase, I have 5 replicates in each locality.  

I would say you cannot directly compare the three habitat levels, as you just have 2 cases of temporary ponds, 2 permanent ponds and one stream. I would suggest you to run the analyses with locality (5 ecosystems) and hydrological levels (4 levels) as the predicting factors. You can try some pos-hoc tests to assert the major differences within the locality level. 

Thank you Vinicius, I agree with your suggestion. 

Kind regards.