The sample size should not be too much larger than SRS. I thing about twice the size might be OK particularly if you use something other than proximity sampling in the cluster. This increase in sample size is usually far outweighed by the difficulty in using SRS in many settings. If you choose the approach of SRS from clinic registers then you will probably introduce a selection bias and be limited to saying thing about complimentary feeding in the registered population (which might not be representative of the whole population). It your "health development army" is doing its job then you'd reasonably expect better IYCF practices in the registered population than in the non-registered population. The data "time of initiation of complementary feeding" is a very narrow IYCF component. I would use something like the ICFI index and then decompose / diagnose it. You'd get a fuller picture of IYCF practices with a small sample size. I would combine this with with coverage of IYCF programming. See [url=https://dl.dropboxusercontent.com/u/73741635/meSL.pdf]this document[/url] from Sierra Leone where they take this approach. I am not sure what you mean by "As another alternative can I change the design from community based to health facility based?" but this will also risk a selection bias (as above). I hope this helps.

I believe that the list of households kept at the HPs are pretty exhaustive in most cases, particularly if they also include the list of the HDAs they belong to; which in effect might mean that the data is recent. You can also ask the HEWs whether they have updated them so that you do not miss any household. If you want to cover different kebeles or a whole woreda you should verify the acceptability of the data from each kebele. Since all HEWs do a mapping of all the households to prepare the list, the probability of households to be missed from the roster- here considered the sampling frame- is going to be very low to have effect on your overall findings. You should also look at the other domains like logistics and resources including time, in your decision making. Even if you believe that you will have a good/ representative data through the SRS, you might end up having the trouble to identify and include the households, as they will be randomly distributed. If you do not have the resources, I think you can also consider undertaking a facility based survey- with all its limitations.