You should include all levels where the ICC is significant as random effects in your melogit model (if they're all significant, the clustering at each level significantly impacts the total variance)

Thank you Marius Ole Johansen for your response. To give you detail information, stage 1 is district (two categories with more than 500 cases each), stage 2 comprises local admin structure (9 categories with about 180 cases each) and third stage is final cluster (55 clusters with about 30 cases). Is it okay to add a level with only two clusters (stage 1) in model? Plus, is it okay to add stage 2 and 3 variables in random effect model equation and consider stage 1 as independent predictor?

Thanks!

Ah ok, if stage 1 only has only 2 clusters (categories) you'll run into some issues as you can't really estimate the random effects variance, you should as you say treat stage 1 as a fixed effect (you could even use a binary indicator/variable representing the district in this part of the model) and include stage 2 & 3 as random effects instead in your model

Many thanks Marius Ole Johansen !!