the third hypothesis is not correct, i.e everything must be approximate.

There are a few points on you question that I would like to make.

(1) SMART when done well can estimate GAM prevalence with both accuracy and precision. It is not no good with SAM as relative precision is poor. If (e.g.) we have an effective sample size (i.e. after account for design effects) of 400 and prevalence of 1% then (assuming accuracy) the 955 CI on the prevalence estimate will run from 0.27% to 2.54% that is very poor relative precision (at 154% rather than 33.7% on a 10% GAM estimate). I am also concerned about the use of PPS sampling which will place the bulk od a sample in the largest communities.

(2) I am not sure about "most stakeholders". First it can ignores the prime stakeholder which is the malnourished child). Large national programs such as the Nigerian and the Sudanese CMAM programs are MUAC only (i.e. MUAC and oedema) programs. This is not uncommon as height boards are not often present in clinics (not part of essential clinic supply packs) and W/H is not covered by IMCI.

I think the main issues here are mortality risk and coverage.

The mortality risk associated with W/H is highly variable and varies from place to place. The mortality risk associated with MUAC is not so variable. Most studies using populations cohorts (i.e. not clinical cohorts subject to all sorts of selection biases) find W/H to be very weakly predictive of mortality and MUAC to be strongly predictive of mortality. There are all soty of issues with W/H. See this review.

I am not convinced that high spatial and temporal coverage of screening using W/H can be achieved. Without this we cannot achieve treatment coverage. Without treatment coverage we cannot achieve effectiveness. We know we can achieve goos coverage of MUAC screening and recent innovations by ALIMA and AAH have show that mothers can do MUAC very well. Having both W/H and MUAC as admission criteria seems safe and sensible but W/H may damage MUAC coverage because it is complicated and expensive. IF W/H is used then we need to be very careful that it does not damage MUAC screening and case-finding.

I fear that the (probably well-intentioned) continued advocacy for W/H is confused and a cause of confusion that undermines program effective. Donors have little problem with MUAC only programming.

Dear Mark,

I appreciate your explanation on MUAC vs WH admission criteria for CMAM programming at community level. An I am aware on the international debate on it.

But my concern is about estimating a unique GAM/SAM prevalence considering both criteria together (+ edema off course) in a given context where a CMAM protocol is already established with the 3 admissions criteria (WH and/or MUAc and/or Edema).?

Thanks
Regards

Damien Pereyra

I think this is just a reporting issue. It is possible to report estimates based on different criteria:

MUAC or OEDEMA will be essential in setting where "MUAC only" CMAM programs operate.

WHZ or OEDEMA may be useful for historical time series when MUAC has not been used.

MUAC or WHZ or OEDEMA will be essential in setting where MUAC and WHZ are both used.

Other criteria may be used (e.g. "marasmic kwashiorkor").

Note that these will only form part of a burden / caseload calculation. Burden will require an estimate of incidence from prevalence. Caseload will require an estimate of incidence from prevalence and an estimate of expected coverage.

Since we can expect different coverages for different case-definitions we might need to have prevalence reported in a more granular manner to predict caseloads.

Just to correct one small misconception. Selecting primary sampling units with probability proportional to size (PPS), then selecting the same number of basic sampling units in each selected primary sampling unit does not result in a disproportionate fraction of the sampling being located in larger primary sampling units. If you multiply out the probability of selection of a) a given primary sampling unit, and b) a certain basic sampling unit within a selected primary sampling unit, you will see that each basic sampling unit in the sampling universe has exactly the same probability of selection.

For example, in a survey of a district, if the primary sampling unit is village and the basic sampling unit is household, a household in a village of 1000 households has exactly the same probability of selection as a household in a village of 100 households. And every household in that district has the same probability of selection. As a result, the proportion of households in the selected sample coming from large villages is exactly the same as the proportion of the whole population living in large villages. If 50% of the district's households are located in the district capital, 50% of the sample households will be from the district capital.

This is one way of looking at this issue. Switching from community to houseshold is, I think, an intelectual sleight-of-hand.

If the selection probability of a PSU is proportional its population then larger PSUs will be more likely to be sampled than smaller PSUs. If services such as health centres and schools are easier to access in larger PSUs than in a smaller PSU then all sorts of things should be better in the larger PSUs. If your sample is located in larger PSUs then survey results will reflect that. That is, the survey will be biased towards painting a good picture. The problem will be most grave with coverage indicators. With PPS, the primary sample is biased and that bias persists in later stage samples.

If we imagine sampling proportional to some other factor ... if this is altitude then then we will tend to select communities on tops of hills rather than in the bottoms of valleys. Would we pretend that such a sample was not, in some way, biased?

As a cadet epidemiologist in sandals and short trousers and with a full head of hair I was taught that Time, Place, Person (TPP) were basic variables to always consider. A PPS sample ignores place as a variable of interest. We regularly present SMART results by age and sex (person variables) and by time (i.e. surveys at different times of year) but never by place. I think we fool ourselves if we think that place is not an important variable. That is, I think, a huge misconception.

Just my tuppence.