You could consider doing a cluster survey which samples using PPS with the 3 livelihood zones as your primary sampling units, but of course your results would then not be representative of each livelihood zone, but of the whole programme catchment. You could also do simple random or systematic random sampling within each livelihood zone to get results which are representative of each zone if depending on the availability of lists of household and the arrangement of households.

This sort of problem is one reason for developing the RAM (RAM = Rapid Assessment Method) methodology. Given the information above I'd guess you'd want a sample of 12 children (selected using a map-segment-sample method) from each of 16 villages (using a spatially stratified sampling method) from each livelihood zone would fit your needs (i.e. three RAM type surveys). This is the approach used (e.g.) in Sudan for nutritional surveillance by repeated cross-sectional survey, in Sierra Leone for M&E of nutrition programs, and in SQUEAC and SLEAC surveys. A PROBIT estimator could be used. This gives better precision than a classical estimator with small sample size. Three RAM surveys will probably cost a little more than a single SMART type nutrition assessment.

Thanks Blessing and Mark, Let me clarify further regarding the baseline assessment. Due to the nature of the program, the assessment will include IYCF (0-24 months), anthropometric and Household questionnaire. So, my question to Mark is, can we use RAM for such setup?

We can do IYCF, WASH, EPI, Vitamin A coverage, DDS, HHS, GAM, SAM, MAM, &c. inside a RAM sample. The ENGINE project in Ethiopia (e.g) collect a very full set of indicators inside RAM type surveys. The CNS in Sudan collects a full set of nutrition indicators using a RAM sample. The indicators that we use are (have to be) a little different from the "standard" indicators because of the small sample size. We try to use "all sample" indicators and novel (to emergency nutrition) approaches such as PROBIT for GAM and SAM. This may seem to be a disadvantage but many of the "standard" indicators are designed for surveys such as MICS and DHS which have very large sample sizes and are unsuited to the sorts of sample sizes we might use in (e.g.) SMART type surveys - this is made plain in the first pages of the WHO IYCF manuals. You'd have to be careful with the sample design to ensure you get sufficient numbers in the 0-24 month age-group if you have a need for a 6-59 month sample. I usually asses GAM in the 6-24 month period as this is the risk age-group for wasting. [url=https://dl.dropbox.com/u/73741635/meSL.pdf]Here[/url] is a guide outlining a RAM type approach to routine M&E surveys for IYCF, WASH, &c. from Sierra Leone. Alternative first stage sampling is CSAS. Confidence intervals can be made using a bootstrap estimator or a model-based approach as used in EpiInfo CSAMPLE and many other statistics packages.

Dear Mark, Thank you very much for your comprehensive feedback. I do have the following follow up questions 1. What make this different from the conventional 30X30 or SMART principles – if we are selecting 16 villages out of the list of communities, it seems to me the same to that of cluster sampling method regardless of the number of sites selected and PPS. Why do you select 16 villages? Why not let say 20? Or 14? Or 30? 2. Why PPS is not used? As per the guide, the sixteen villages are selected randomly regardless of their population size. 3. It seem to me that the guide/tool is mainly a monitoring tool for the ongoing programs. do you think it can be used taking baseline assessments with multiple indicators? 4. Total sample size – you are recommending 192 (16*12). Is this 192 for single indicator or for multiple indicators? For example in the document you shared us, for IYCF assessment (0-24 months) the sample size was 192. If I want to include anthropometry survey (6-59 months), do I need to add another 192? If not, do you think the sample size is enough to capture enough information? Of the total 192, the estimated number of infants under six is about 19 children. Due to think 19 kids are enough to get better information on EBF? The same is true to calculate complementary feeding for narrow age groups (6-8 months). 5. Is this RAM method for small population like my case or can be used for high population? thanks

It is not a 30-by-30 (or SMART) survey. It is a 16-by-12 survey (i.e. 16 clusters of 12 children). It also uses a different within-cluster sampling method. This method has been found in tests (during the development and testing of the EPI method) to retain variation. The conventional (proximity) method tends to lose variation (hence the need for a large number of clusters and large sample sizes used when proximity sampling is used). In fact, the proximity method was tested for a variety of indicators and found to be unsuitable for many. Sixteen clusters are selected to reduce costs and time. The RAM was developed to be quick and a lot of time is often spent in travelling and a lot of money spent on vehicles, fuel &c. The use of a smaller number of clusters and a smaller overall sample size is enabled by the within-cluster sampling method, the small within cluster sample size, and the use of indicators suited to small sample size. PPS was not used because the original client (a UNO) request a method that did not need to have population data in advance of data collection (i.e. it was for rapid emergency assessment where displacement and mortality may be high) and was quick and cheap. We also do not use PPS in SQUEAC and SLEAC because "population" (i.e. number of children multiplied by prevalence of SAM) is unknown. Also, PPS locates the sample in most populous communities. This may not be what you want (particularly for coverage). We can collect approximate population data (e.g. door counts) for clusters and use that to weight data during analysis. The guide is for an M&E tool. I provided this as it gives a words and pictures guide to the sample design and also because it shows the small sample IYCF and WASH indicators. It is a general purpose survey method and can be used for a broad range of indicators. We have used it with EPI, WASH, IYCF, GAM, MAM, SAM, stunting, dietary diversity scores, hunger scales, as well as for coverage of a variety of programs. We and others have generally used it for multiple indicators. The sample size of n = 192 is given because it (just about) guarantees a 95% CI of plus or minus 10% or better on a proportion of 50%. It is also gives equivalent or better precision for GAM, MAM, and SAM as a SMART survey with a sample size three times larger when using the PROBIT estimator. We also selected 192 because it has a large number of whole number divisors (i.e. 1, 2, 3, 4, 6, 8, 12, 16, 24, ... and so-on) making it easy to distribute the sample among clusters. You may need a different sample size (e.g. 24 clusters of 8 or 16 clusters of 12). The sample size needed will depend upon the indicator, estimation method, required precision, &c. It is common to pick a sample size that gives useful precision for all indicators. WRT your example, I would, I think, go for a single sample that would give you about n = 192 (assuming this is big enough for your needs) in the smaller eligible group. Assuming uniform age distribution this would be about n = 450. WRT your IYCF example, the sample for the EBF component of the small sample IYCF indicator would be something like: 192 * 0.25 = 48 This is a diagnostic indicator and the whole sample is used for the main IYCF indicator. I think the sample is sufficiently large for diagnosis. I think you may be confused. We do not use the large sample approach. You cannot sensibly use those "standard" indicators unless you have very large sample size (i.e. much bigger than RAM or SMART). This is stated in the WHO documentation for these indicators. The proposed IYCF indicator looks at "adequacy" or "goodness" across the entire sample. EBF, 6-8 months, &c. refer to a different indicator scheme than the one we propose for use with RAM. What we have is (probably) closest to MAD in the "standard" indicator set but with some diagnosis. The indicator has been well tested and is based on work done for DHS and by FANTA / USAID. The RAM method is general and can be used in large and small populations. If the population size is less than about N = 5000 you may (as with any survey) need a smaller sample size for estimation. If you want to classify instead of estimate then a smaller sample is required. We use a maximum of n = 40 in SLEAC with a three class LQAS classifier. Our principal testing has been using PROBIT in populations of n = 17000 (i.e. 17% of 100000 total population may be 6-59 months). In the size of population we get good precision with n = 192. I think that covers your questions. Please post more here if you need more clarification or documentation or if I have missed anything.

In response to the question on sampling for a survey estimating IYCF practice rates, the Care Step by Step Guide is an excellent aide to help with such practical questions. You can find it on the ENN website, under tools for (other) assessment: http://www.ennonline.net/resources/tag/152

Take a look at [url=http://www.brixtonhealth.com/ennIYCF.pdf]this draft paper[/url]. It works quite well with small sample sizes (e.g. n = 192). I hope this helps.