This is a weakness of the PPS sampling scheme when used in emergencies in which there has been considerable displacement. In these contexts I would not use PPS but use instead a stratified spatial sample of villages and collect populations or hut counts during the survey and weight the sample during analysis (as is done in RAM). Now ... to your question ... there are three (probably more) approaches: (1) Just take the households that you can and leave it at that. You PPS weightings will be wrong (they will be anyway if there has been displacement not accounted for in your village populations) but there is little that you can do about that unless you want to weight during analysis. The overall sample size will not be too badly hit unless you find a lot of abandoned / decimated villages because we usually sample one or two extra children from clusters. (2) Take the households that you can and then take a top up sample from the nearest village not otherwise sampled. This also fouls up the PPS weightings (but they might be fouled up anyway). (3) Plan the sample with contingency clusters. If you need (e.g.) 30 clusters then plan / locate 33 (or more) clusters. Clusters 11, 22, and 33 will be contingency clusters and these are only sampled if needed. This used to be common practice with the 30-by-30 survey method on which SMART is based. The problem with this approach is that you end up excluding the decimated villages and that might bias results. Better, perhaps, to use this approach to cope with insecurity or to replace completely abandoned villages. You could use any of these approaches. If I had to choose I think I would do (1) and (3) if a village was completely empty. Whatever method you choose you should document it in reports. I hope this helps.