When sampled cluster doen't have enough HH demanded by sample size because of urgent and sudden population movement

I would definitely recommend that you drop this cluster and use a RC (replacement cluster).

SMART has some solutions for such case. You can combine a nearby village which was not select as cluster initially with the village with fewer households than required. The second option is that you can do the few households you are able to get in that village and move on with other clusters, the non respondent percent you added at the sample size calculation will take care of the shortcoming.

Your decision is good Nicky.

Here are some thoughts on the issues....

If the team arrives at a cluster and there are not a sufficient number of households or individuals, one option is to assess those that are there and then use sample weights in the analysis. In the example the goal was 21 HHs and the cluster had only 20 HHs - one less HH will not make affect the estimates in any meaningful way. Therefore I would agree with Nicky on this issue to just sample the 20 HHs

What if there were even fewer HHs, say 10 or fewer? Should a replacement cluster be used or visit a neighboring cluster or other options? Some of these were discussed by others in their responses.

In general, I would not recommend having contingency clusters. When clusters are selected using PPS, they have probability of selection but if all of the clusters are not visited (i.e., the "contingency" clusters), this probability of selection is violated. My recommendation is that if it is thought that some clusters may be may not be accessible due to insecurity or other reasons, the total number of clusters should be increased - same in concept that when households are visited in a cluster, the number of households is increased to account for nonresponse. For example, if a 30 cluster survey is desired, but it seems likely that two clusters may not be accessible, then select 32 clusters. Every reasonable effort should be made to visit all 32 clusters - the teams should not stop once 30 clusters have been visited. It may be that the actual number of clusters assessed could be 32 (all accessible), 31 (one not accessible), 30 (two not accessible), or fewer. The important issue is that the results may be biased if clusters not accesible differ meaninfully in the survey indicators from those assessed. Less of an issue is that when fewer than the target number of clusters are visited, there may be a slight loss in precision.

Most surveys (e.g., DHS and MICS) generally do not recommend replacement clusters for the same reason that most household-based surveys do not recommend replacement households - the potential for introducing more bias to the survey. While not being able to assess a cluster may introduce some bias, cluster replacement has the potential to introduce even more bias. If 30 clusters are to be assessed but one is not accessible, selecting a replacement cluster violates the selection probabilities and may even introduce more bias into the survey results than the loss of one cluster.

Whatever approach is used, I agree with Mark that all procedures in the selection of clusters and households be described in detail so that the reader can make their own judgement on the quality of the methodology and potential affect on the interpretation of the results. The further the a survey drifts from acceptable survey methods, the greater the potential for biased estimates.

Given the constraints of attempting to adhere to the statistical aspects while faced with less than desirable field conditions, there is no such thing as a "perfect" survey but we should strive for a reasonable attempt to attain high quality survey results.

Kevin

Based on the information provided I don't think it is necessary to replace the cluster with 20HHs with a replacement cluster (RC) as it can't meaningfully affect your final sample size as stated by Kevin above. I would definitely proceed with surveying the cluster with 20HHs as it is. After all the target is children 6-59 months old and the number of HHs/Cluster are calculated based on the proportion of under-5 years/6-59months old children per HH. Hence as in most cases with surveys on the ground that same proportion could be less or higher per HH. So the non-response in HHs will cancel out each other by this fact and your Survey will still maintain the basic principle of probability sampling, there by avoiding more bias being introduced due to this reason.

However I prefer increasing the number of Clusters to be surveyed instead of the number of HHs (eg. instead of using 30 clusters of 20HHs it is better to use 42 clusters of 14HHs) as this will reduce the variability of responses in a clustered sample. Statistically speaking increasing the number of clusters enhances power more efficiently than does increasing the number of subjects within a cluster.

On the other hand as most of these surveys are emergency nutrition surveys time and logistics as well as financial issues are best maintained within acceptable levels.(so using contingency/replacement clusters would save the much needed time allowing us to target on meeting only our representative sample size with minimal bias being introduced). So these decisions are usually best made at the ground level taking into account the geographic/spatial population distribution/arrangements of the survey area as well as the context, while maintaining its statistical soundness in sampling.