I'll deal with each question in turn ... (a) Yes. The active and adaptive case-finding procedure is only about 50% sensitive form MAM cases (we did some work on this in Ethiopia and Malawi). This means that you would be better surveying fewer villages (OK because MAM prevalence is usually much higher than SAM prevalence) using house-to-house and door-to-door sampling. (b) Yes (probably) because the SMART survey sample will be concentrated in towns and larger villages This is what PPS is supposed to do. This usually means that the SMART coverage estimate will be higher then the CSAS coverage estimate. (c) The issues with SMART are (i) sample size and (ii) spatial representativeness. At 10% prevalence with a sample size of 900 you will get a sample size of about 90. This is pretty good ... good enough to estimate coverage with better than about 10% precision. The PPS sample concentrates the sample in population centres. This means that coverage may be poor in more rural settings but the survey will tell you that coverage is good. For some, this PPS estimate is fine. I prefer a more spatially representative sample. I hope this helps.

Hi Mark, A couple of comments on your response: "(b) Yes (probably) because the SMART survey sample will be concentrated in towns and larger villages This is what PPS is supposed to do." A two-stage cluster survey using PPS to select clusters is a method useful for estimating prevalence or coverage in a targeted geographic area. For example, if 40% of the population lives in larger villages, then ~40% of the clusters will fall in larger villages. There is no "concentration" per se. "The PPS sample concentrates the sample in population centres. This means that coverage may be poor in more rural settings but the survey will tell you that coverage is good. For some, this PPS estimate is fine. I prefer a more spatially representative sample." In the above statement is the mention of coverage in rural settings. If there is a need to estimate coverage in a rural area, then two possible choices are: 1. Define the target area as the rural area, or 2. Perform a stratfied survey - one in the rural area and the other in the urban area. No one type of assessment can answer all questions - the primary questions to be answered need to be determined first and then a selection of the design that balances the needs for validity and precision with costs.

It is a matter of how you define "representative". PPS defines it one way and CSAS defines it another way. From the perspective of an even spatial sample (as CSAS tends to give) then a PPS sample will be "concentrated". I prefer CSAS-like samples for coverage because they can detect and map heterogeneity. CSAS is a stratified sample and we can (theoretically) weight the sample when calculating an overall coverage estimate. I have tended to not to do this because weights are not easy to derive. WRT weights being difficult to derive ... this also applies to the PPS weights. We shouldn't use raw population estimates alone (as we usually do with PPS). The correct weight is population times local prevalence. Using population alone assumes homogeneity of prevalence. SAM and MAM are often associated with infection and infection (almost by definition) is a clustered phenomena. This (IMO) means that weights created under a homogeneity of prevalence assumption might be way off.

You bring up a number of different issues in your response. Again, there needs to be a decision as to the primary question(s) to answer in a survey or assessment, and then pick a reasonable design that balances validity/precision issues with resources available. If a primary goal is to map heterogeneity, then pick a design that will reasonably answer that question. If a primary goal is to estimate overall coverage, use a design that will reasonably answer that question. I think it confuses some by stating "... we shouldn't use raw population estimates ..." and "... the correct weight is population time local prevalence..." because these statements assume specific goals in mind for a survey and may not be the best choice for all goals. It usually concerns me when I see statements that basically say "there is only one way to do this ..." As a secondary issue, a two-stage cluster survey does not assume homogeneity of prevalence or coverage - the level of heterogeneity is built into the variance structure and can be measured in terms of the ICC and DEFF. No one survey type is perfect for all purposes - all we can hope to do is select a reasonable design to answer the primary questions. There are a lot of tools in the survey toolbox - simple random sampling, cluster survey, LQAS, CSAS, purposeful sampling, and many, many others - one should pick the tool that is most appropriate for the task.

The weights proposed (i.e. population multiplied by local prevalence) are suited to coverage surveys since only it is the number of cases in each location (rather than cases and non-cases as in SMART surveys) that should be used to weight the sample. If you use raw populations (as is common with SMART type surveys) then you are making a homogeneity of prevalence assumption. It is often said that "a two-stage cluster survey does not assume homogeneity of prevalence or coverage - the level of heterogeneity is built into the variance structure and can be measured in terms of the ICC and DEFF" but am not convinced since (1) the method yields a single wide-area estimate and (2) the DEFF captures underlying variance and the variance introduced (or lost) due to the use of (e.g.) proximity sampling in the second stage. This are all tangled up together. I agree that "no survey is perfect for all purposes" and we "should pick the tool that is most appropriate for the task". I am not convinced that a PPS sampled survey as with SMART is appropriate to coverage assessment.

From Nikki Blackwell: hi mark, thank you for your commentary - just to clarify, when you say " I am not convinced that a PPS sampled survey as with SMART is appropriate to coverage assessment" do you mean that you DON'T believe that SMART is the appropriate tool for coverage assessment? i thought that is exactly what it was designed for? thanks nikki

SMART was designed to assess prevalence not coverage. It does that (within limits) well enough. SMART is not an appropriate tool for coverage assessment. There are many reasons for this ... I think the sample size issue is the main problem ... SMART samples children between 6 and 59 months. Very few (e.g. 1%) will be SAM cases. If we assume a sample size of 900 and prevalence of 1% then you have a sample size of about 9 to estimate coverage. That is a small sample size. I think you need an effective sample size (i.e. after correction for design or prior information for a Bayesin approach) of about n = 96. To get that needs a sample size for a SMART type survey of about 10,000. Too big top be feasible.