Dear Caroline, I don't have a way to get a proxy on the global results, but what we do since 2 years in our MUAC assessment is to divide the tally sheet between children < 85 cm and >= 85 cm because we have seen that the correlation between MUAC and W/H is better in small children. In analysis of our nutritional surveys, I realised that often, the GAM & SAM as expressed by MUAC for the height group < 85cm was quite similar to the GAM & SAM as expressed by W/H for the all sample. So, to have an approximation of expected GAM & SAM in W/H, when I do rapid MUAC assessment, I take the prevalence of the group < 85cm. Hope this help Pascale

First a warning ... unfortunately "rapid assessment" often means "poorly done assessment" using a convenience sample such as a call to a central location in a village chosen because it is not too far to travel. Another approach is an "optimally biassed" sample where the sample is taken from a location where you suspect there might be a problem. Such samples are subject to all sorts of biases which make them less than ideal for making inferences about the general population. That said ... Since you quote 110 mm and 125 mm I will use those thresholds. There are reasonable threshold for SAM and GAM based on near-term mortality risk. There are a few approaches to this problem. The simplest is estimate a proportion in the usually way. That is apply case-definitions and calculate: prevalence (%) = (Number of cases / Sample size) * 100 Applying the SAM case-definition: MUAC < 110 mm We get: (13 / 120) * 100 = 10.8% Applying the GAM case-definition: MUAC <= 125 mm (should be < 125 mm but you do not give that) We get: (79 / 120) * 100 = 65.8% You can calculate a 95% CI using the formula: p +/- 1.96 * sqrt((p * (1 - p)) / n) I'l show a worked example for SAM: 0.1083 +/- 1.96 * sqrt((0.1083 * (1 - 0.1083)) / 120) 0.1083 +/- 0.556 = 0.053; 0.164 or 5.3%; 16.4% Another approach is to use the mean and SD of all the MUAC data and then use the PROBIT function to estimate prevalence. This can be done quite easily in a spreadsheet. I will not show this here but if anyone is interested I will post this in a second message. Let me know. Now a second warning ... These figures (i.e. GAM of 65.8% with 10.8% SAM .. these are extremely high) suggest to me that you have a biassed sample which favoured the sick over the healthy. How were these data collected? BUT ... WAIT ... Pascale's reply makes me think that I have misunderstood the question. Translating between MUAC prevalence and W/H prevalence is difficult since W/H is strongly influenced by body-shape and this varies with diet, climate, altitude, genetics &c. I would stick with MUAC case-definition and report prevalence using them. You could derive a conversion formula from LOCAL survey data with both MUAC and W/H and apply that. There is merit in Pascale's suggestion of using only the younger children since the effect of body shape occurs after the age of about 2 years (that corresponds to about 85 cm). The problem is that you will only have a sample size of about 40. PASCALE : What is the nature of the difference between prevalence by MUAC and prevalence by W/H that you see in the older age-group? My work suggests that this should go both ways. In agrarian / mountain folk there should either be little difference or MUAC < W/H. In warm-climate pastoralists the difference should be W/H > MUAC. Is this what you see?

The current WHO recommendations are that children with MUAC <11.5 cm should be considered SAM whereas those between 11.5 and 12.5 cm are moderately acutely malnourished. Those above 12.5 cm are considered normal. Among those with SAM you can go further to examine if it is SAM with complications or not using the presence of oedema grade +++, loss of appetite, clinically unwell or not alert etc. You can find further information from WHO's website or from the International Malnutrition Task Force website www.imtf.org

Hi Mark, I am interested in learning more about using the PROBIT function for estimating GAM and SAM prevalence rates. The NGO I work for plans to implement a CMAM program in two districts. We are getting very conflicting GAM prevalence rate estimates--extremely low according to HMIS data (though health workers supposedly do routine screening) and relatively high for a non-emergency setting according to a household survey done a few years ago. We plan to do a SMART survey in the target area, but before investing significant resources in a survey I think we need to get some rough estimates of the prevalence rate, perhaps from mass screening using MUAC in randomly selected villages. Any guidance you could provide on the PROBIT function would be much appreciated. Thanks!

I am not sure how much detail you need so I'll start with a "nuts and bolts" introduction. To use the PROBIT technique we usually need data that can approximate a standard probability distribution (we can use resampling if we know how to do that). We most often use the Normal distribution. If our data are not normal then we can often transform then towards the normal. In the case on small deviations from the Normal (e.g. slight asymmetry, heavy tails) we might use robust estimators of the mean and standard deviation. we can do this with MUAC data provided it is not truncated (as it would be using MUAc at admission to OTP which is truncate at MUAC = 115mm). use of robust estimators for the mean and SD should also be used with small sample sizes. I use the median (rather than the mean) and: SD = Interquartile-range / 1.34989 other robust estimators can be used but these work pretty well. Having got your estimate of the mean and SD you can calculate the probability that a child picked at random will (e.g.) have a MUAC < 125 mm. This is the same as prevalence of MUAc < 125 mm. You can do this by standardising and using a set of statistical table or using a computer. Most stats packages and spreadsheets will be able to do this. In OpenOffice Calc (e.g.) we can ask the question "What is the prevalence estimate for MUAC < 125 mm given mean = 145mm and SD = 12 mm" with: =NORMDIST(125;145;12) which returns: 0.0477903523 which is about 4.8%. Doing this in Excel is similar - just use "," instead of ";" in the cell formula. Post back here if you need more detail or need to do more. I hope this helps.

Hi Mark, Thanks for your helpful reply-- this all makes sense. As I mentioned, our plan is to do a SMART survey in the target area, but before investing significant resources in a survey we hope to get rough estimates of the prevalence of acute malnutrition from MUAC data from mass screenings in randomly selected villages. Is there any guidance on how many children should be included in the mass screening to get a reasonable prevalence estimate? In other words, what the required sample size would be?

Dear Jennifer Will see the Mark technical view and recommendations later but what i am thinking and suggesting for mass screening to know the exact situation of malnutrition through MUAC without conducting SMART survey is, you will screen whole under 5 children (6-59 months) of targeted village/community within the village rather than calculating sample size to have a clear picture of prevalence of malnutrition. Recently we had done the same procedure for estimating malnutrition prevalence among IDPs and Host community. Hope this will make some sense Thanks

This is, I think, a very good idea especially if your program is using mass-screening anyway. This way you combine "service" with "survey" and will very probably find and refer and (hopefully) cure a lot more cases than you would with a SMART survey. There are some things to consider ... (1) The screening should be of all eligible people in the screened communities. In some settings central location screening will not find all cases as they may not attend due to stigma or shame. Also, sick kids tend to be both lethargic and irritable and carers may be reluctant to attend. If you get this wrong then you will bias your prevalence estimate downwards. If the opposite is the case and only sick children get brought for screening then you will bias you prevalence estimate upwards. You can avoid this. Either go house-to-house for screening or "mop-up" after a central location screening using something like active and adaptive case-finding (avoiding double-counting). In some settings people are reluctant to move outside their immediate neighbourhoods. If you use central location screening then make sure you use several "central" locations. (2) Make sure that your sample of communities is not a convenience sample (e.g. villages close to centres or roads). This would also introduce a bias (probably downwards). I would use a spatially representative sample so the sample comes from all over your program area. (3) You need to have a fair number of communities to be sure that you have a representative sample of communities. The general rule is "more is better" but there is little point in going above about 24 communities. I think 12 or 16 should be sufficient. If you do this every month or so in a small sample of communities (changing the sample each round) then you would have the makings of a nutritional surveillance system. (4) Data analysis should treat the sample as a stratified sample and a weighted analysis should be performed. In the case of mass-screening that meets (1) above the community-specific weights would be the number of children screened in each community. If you get this right (and it is not too hard) then you will have no need for a SMART survey. The [url=http://www.fantaproject.org/monitoring-and-evaluation/squeac-sleac]SQUEAC / SLEAC Technical reference[/url] has material on spatial sampling (pages 93, 94, 95, 96, 100, 101, 102) and analysis of data from stratified samples (pages 127, 128, 129). I hope this is useful.

Thanks for the helpful reply, Mark. This info is very useful. In terms of sampling, I think the best approach in our case will be to stratify by clinic catchment area and select villages systematically from a complete list of villages sorted by clinic catchment area (information that we already have available). Then screen all children 6 to 59 months in the sampled villages, using active and adaptive case finding. You mentioned that sampling 12 to 16 villages/communities should be sufficient. Should I be calculating a target sample size (using the estimated prevalence rate for the region and desired precision) in order to determine the number of villages to sample?

This approach looks fine but ... just to be clear ... if you use active and adaptive case-finding alone you will bias the sample towards finding cases and any prevalence estimate will be an overestimate. Another way of describing these types of method is "optimally biased". I assume that you mean that you will do mass-screening at one or more "central locations" and then do active and adaptive case finding ti "mop up" cases that may have been "hidden". If you use only active and adaptive case-finding to find cases than you estimate prevalence if you have a good estimate of the populations of the sample villages. WRT sample size ... there are two approaches to estimating prevalence. (1) The classic approach is to recode MUAC data to binary variables (e.g. GAM not GAM, MAM not MAM. SAM not SAM) and then estimate by dividing the number of cases by the sample size. To get useful precision you often need a large sample size because (a) you need quite fine precision and (b) you will have a design effect because of the nature of your sample. This needs a sample size of about n = 500 or more. You can use the sample size calculator in the ENA software. The main problem with this approach is that the SAM estimate will lack precision. (2) The PROBIT approach makes more use of the MUAC data and does not need such a large sample size. A sample size of n = 192 is usually sufficient to give similar precision to a classic approach with a sample size 3 or 4 times larger for GAM. The estimates for SAM are more precise than a classic estimator with a sample size 3 or 4 times larger. To illustrate ... a classic estimator with n = 544 (largest sample size in the SMART manual) has relative precisions of about 27% and 65% for GAM and SAM respectively whereas a PROBIT estimate with n = 192 has relative precisions of about 24% and 34% respectively (see [url=http://www.brixtonhealth.com/updateRAM01.pdf]here[/url] and [url=http://www.brixtonhealth.com/updateRAM01.pdf]here[/url] for more details). I do not think sample size will be an issue as you will be screening all children in your sampled communities. You need to be sure that you keep track of the number of children in the sampled communities. I hope this helps.