Dear Montse:

The simple answer is "It depends".  It depends on what assumptions you make about the population to be assessed.  The minimum sample size needed to make an estimate of prevalence depends on: a) the assumed prevalence of the condition, b) the precision needed for the final estimate of prevalence, c) the type of sampling done, d) the heterogeneity of distribution of the condition for which you are calculating sample size, and e) whether you are calculating sample size for a single survey or to compare 2 or more surveys. I will assume that you are calculating sample size for a single survey only.  

The formula for calculation of sample size in order to achieve a minimum precision for a single survey is:

n = DEFF * 3.84 * (p * (1-p)) / d squared. 

3.84 is the Z value for 1.96 which assumes that you will be calculating 95% confidence intervals (or using p<0.05 as the threshold for statistical significance).  p is the assumed prevalence and d is the 1/2 confidence interval you want for your final estimate of prevalence (for example, if you want a confidence interval of +/- 3 percentage points, d = 0.03).  DEFF is the design effect which reflects the loss of precision from doing cluster sampling instead of simple random sampling; it depends on the average cluster size and the inherent heterogeneity of distribution of the condition.  

So you can see that there is no universal "minimum sample size".  If we assume that the design effect for a cluster survey will be 1.5 (a reasonable assumption for many nutrition outcomes), the sample needed to achieve a precision of +/- 5 percentage points is 578 for a condition with a prevalence of 50% (perhaps stunting) and only 209 for a condition with a prevalence of 10% (perhaps wasting). Other outcomes which may be measured in nutrition assessment surveys, such as health care coverage or water source, have much higher design effects (as high as 4 or 5), so estimating the prevalence of these outcomes requires a much larger sample size given a given desired precision. If your survey will measure more than one important condition, you should calculate the sample size for each condition, then include the largest number in order to ensure adequate precision for all conditions measured.   

This is an entirely inadequate summary of a somewhat complicated topic. If you do not have training and/or experience in calculating sample size, I urge you to seek assistance from someone who does. It would be shame to carry out a survey, and then find that the results are uninterpretable because of inadequate precision. 

Dear Montse,

Bradley answer's present the complexity of sampling design and sample size...

It is difficult to advise propertly with low information, but the rapid smart guideline could be of some help.

NB: The rapid SMART method (which only apply for GAM evaluation in emergency context with low accessto the field) proposed an pre-define sample size (for cluster sampling) to consider the nutrition outcome (GAM prevalence in case of rapid SMART):

• (1) if the GAM<15% a minimum of 10 households in 25 clusters (250HH/strata) or
• (2) if the GAM>15% a minimum of 12 households in 25 clusters (300HH/strata)

A minimum of 25 clusters, should be reached in order to catch the heterogeity of the nutrition outcome over the target population. Having 10 or 12 hh per cluster, permits to limit the time in the field during the data collection (considering 1 cluster/day/team). This design is particularly adapted for frontline humanitarian data collection worker when access to the field is limited.

Hope this is from some help

Regards

Thanks so much Bradley for your comprehensive answer, very appreciated.

And thanks also Damien for your contribution.

My question was related to this statement of  the Rapid SMART guide :

The field tests of the method have suggested that when proper sampling is respected, measurements are accurate, and the precision is within accepted limits, Rapid SMART surveys can provide GAM and SAM estimates which are representative for the studied population (children from 6 to 59 months) living in the delimited zone.

I wanted to be sure that we can apply the same formula for sample size calculation even if the age group is different, for instance 0 to 6 months old or for PLW.

Could you confirm it please?

Thanks a lot, very very useful.

Dear Montse:

The formula for sample size to estimate a proportion or prevalence is the same regardless of the unit of analysis.  It applies to children 0-59 months of age, pregnant and lactating women, households, widgets being sampled from a factory's production line, or any other unit of analysis. 

Regarding the Rapid SMART Guide, I would still recommend that you calculate your own sample size. Cookbook guides for carrying out surveys are fraught with potential problems, and many of them, including the Rapid SMART Guide, contain incorrect information and recommendations. Moreover, each population to be surveyed is different, and cookbook guides do not the allow flexibility necessary to adapt your methodology to specific situations. Understanding the epidemiologic and statistical principles behind survey methodology makes it much more likely that any questions which arise during planning or execution are answered correctly.  

For future reference, you may be able to use this web-based course in emergency epidemiology to answer some questions about surveys in emergency-affected populations: http://conflict.lshtm.ac.uk/index.htm.  

I hope this is helpful, and I wish you luck in your work.

Dear Montse,

Building on the suggestions from Bradley and Damien, it is no easy task conducting a Rapid Nutrition Assessment (which generally implies collecting data swiftly at the expense of a smaller sample size and possibly less representative data) while also collecting sufficient data among infants 0-5 months because they make up such a small proportion of the population. In most cases, if the sample size is based on children 0-59 months but you wish to examine infants 0-5 months separately, your sample size will be insufficient to render meaningful results.

Let’s say an agreed sample size for infants 0-5 months has been calculated, following the calculation shared by Bradley. If conducting a representative household survey, you will then need to find updated population information to help estimate what proportion of the population is composed of infants 0-5 months. From there, you can estimate the number of households that will likely be needed to achieve your sample size.

Agreed with comments above that this is a complex topic. If ultimately you are in need of technical support for sample size calculation and/or planning, I am the Assessment Advisor with the GNC Technical Alliance and we are happy to support in this area. To request support, visit https://ta.nutritioncluster.net/ and click the “Request Support” button at the top of the page.

Best of luck with your survey planning!

Woody's response of "it depends" is good advice.

It also depends on the estimator you use. I have been using the PROBIT approach because it usually requires a smaller sample size then the classic estimator which tends to squander the information in a sample by collapsing anthropometric data into binary categories. The RAM and RAM-OP survey methods use PROBIT estimators. These methods frequently use a sample size of n = 192 (collected as m = 16 clusters of n = 12 childen) and report prevalence with usful precision. They use efficient spatially stratified sample designs. Blocked and weighted bootstrap estimators are used.

There is more on RAM-OP here and here.

I hope this is of some help.