This is an older thread, but I've just come across it, so will add a response.

The clustering of cases of kwashiorkor is very common in central Africa, and as Dr Golden noted, the clusters can be as small and close to each other as neighboring villages. This clustering of cases makes the usual cluster-based survey sampling method inappropriate for detecting kwashiorkor and has probably led to underestimations of kwashiorkor burdens. Interestingly, instead of differences between areas with and without kwashiorkor being the focus of causation in research, often location is used as something to control for in the search of the "real" cause, when maybe an environmental factor is a cause.

Having observed this, we did a census survey and mapped out prevalence by village. In this case, it appeared that historical differences in settlement patterns and associated social disadvantages followed the differences in kwashiorkor prevalence.

Here is a link to the article. It is open source, so hopefully the link will work.

https://pubmed.ncbi.nlm.nih.gov/30136596/

Have a good day,
Merry

Merry, thank you for the reference to your interesting paper demonstrating heterogenous distribution of kwashiorkor in eastern DRC.  Just a quick comment on your statement that "This clustering of cases ... has probably led to underestimations of kwashiorkor burdens." Use of the word "underestimations" implies the presence of bias; that is, a survey estimate of the outcome does not match the true outcome in the population because of some problem with measurement, sampling, or some other mistake. Cluster sampling, when done correctly, is free of such biases and will not lead to underestimations, regardless of how heterogenously the outcome is distributed.  Of course, the heterogeneity of the distribution of kwashiorkor in your villages would doubtless lead to a large ICC which produces a large design effect, thus decreasing the precision of any estimates based on a random sample of the population. Such poor precision may also result in a substantial difference between the point estimate from a survey and the true population value, but we can calculate the precision (often expressed as confidence intervals around a prevalence estimate) in order to judge how useful the survey's estimates are. We must always keep in mind the difference between poor precision and bias because they have very different implications. 

Dear Bradley,

Thank you for the correction. Indeed, this would be an error of precision rather than bias - so depending on how the selected clusters fell, cluster sampling might also over estimate the prevalence.

In the case of the area studied in the article, the selection of clusters happened to fall into the villages with the lower prevalence, leading to the conclusion that the prevalence of the entire zone was so low the NGO pulled their services and redirected them elsewhere. If the cluster selected had been just a half km west, they would have had come to very different conclusions. When the MoH saw the results of this survey, they tried to redirect their own resources to address this cluster, but they lacked the resources to do it properly.

I'm certainly no expert at the nuances of survey design. My understanding is that the current 30x30 cluster sampling is based on accurately detecting prevalences more typical of moderate wasting, that it is not really appropriate for accurate SAM prevalence when the prevalence is fairly low. And if kwashiorkor does cluster more tightly than wasting, as annecdotally appears to be the case, and if prevalence is often as high as the area in this study, then perhaps we need to be working on a different survey design for regions where kwashiorkor is the majority of the SAM cases?

Through out eastern DRC, local health staff can direct you to those villages or groups of villages where kwashiorkor is clustering. In those cases where I have followed up on their directions, they've been correct. In these areas, people are pretty good at detecting not only cases with bipedal pitting edema, but also children that are in earlier stages. Perhaps in cases where the overall prevalence is estimated to be relatively small, say under 4%, would active case-finding provide a more accurate approach? Something like # cases found through active case-finding/total estimated population?

Thanks for the commentary and input,

Merry 

I think Woody is right that we might expect poor precision (wide confidence intervals) on prevalence estimates of clustered conditions when using a cluster sample. Kwashiorkor is a rare condition and we have little use for wide confidence intervals and may end up with confidence intervals that extend (impossibly) below zero ... if we see only one case then prevalence must be above zero so a lower confidence limit <= 0 is nonsensical). This might be addressed using more exact analytical methods which are available now we have more powerful computers ... when I was a young epidemiologist in short trousers we used to crudely "trim" nonsenical intervals (i.e. replace the <= 0 limit with a very small number such as on tenth of one percent) ... but could also be addressed using different survey designs and sampling strategies.

I think there may be a bias issue here or maybe it is because we are making a category error. We might be mistaking prevalence for "burden" or "need" when we might better using incidence. If a condition is acute (i.e. cases resolve or die quickly) then the number of prevalent cases will be far fewer than the number of incident cases and will underestimate the burden of disease. We see this sort of thing frequently when we use SAM prevalence estimates (multiplied by population) to predict OTP numbers and we end up with caseload predictions that are well below observed admissions even when coverage is poor. These are problems of cross-sectional survey rather then problems specific to cluster samples.
 

I hope this make some sense.

It is mainly related to the type food consumed I.e. areas where populations eat more of cereals with low intake of protein sources.