Language: English Français

# Is it possible to use LQAS methodology to assess the prevalence of stunting in non-emergency settings?

This question was posted the Assessment and Surveillance forum area and has 10 replies.

### Mark Myatt

Epidemiologist at Brixton Health

Frequent user

23 Jul 2015, 11:05

You will be able to classify prevalence as likely to be above or below a given threshold in each of the n = 19 samples. The m = 5, n = 19 samples can be combined. This should be done by treating each sample as a stratum in a stratified survey (see here) rather than by just joining the samples end-to-end. The final effective sample size will likely not be less n = 95 so any estimate will have a precision (width of 95% CI) of +/- 10%. or better. You may get better precision using a PROBIT estimator. You could use a RAM type sample. This is usually an m = 16, n = 12 sample (i.e. overall sample size is n = 192). This costs about half the cost of a SMART survey. I hope this helps.

### Mark Myatt

Epidemiologist at Brixton Health

Frequent user

23 Jul 2015, 13:29

Taking each question in turn ... Baseline / end-line comparison : The utility of the combined LQAS sample (or any sample) will depend on what you want to do with it. In the case of a baseline / end-line comparison you will, I guess, want to be able to detect a difference (with a stated direction) between the two survey results. If the desired or expected difference is small then you will need a large sample size. If the difference is large then you will need a small sample size. A rough rule-of-thumb is that the difference you will be able to detect will be about the same as the width of the 95% CI. With n = 95 the width of the 95% CI on a 50% prevalence will be about 20% (i.e. +/- 10%). This means that you will be able to detect a difference between 50% at baseline and 30% at end-line. If you have a stated direction of difference then you can do better (the term for this is "one-sided hypothesis" because the direction is in one direction). You can calculate sample sizes for comparing percentages here. Since you will be working with a continuous measure (i.e. HAZ) then you may want to consider testing for a difference between means rather than proportions. This is usually more efficient. PROBIT : This is a way of estimating a proportion using a threshold case-definition applied to a continuous measure (e.g. HAZ < -2). It is usually more efficient than the classic estimator (i.e. the number of cases divided by the sample size) because it makes better use of the available data. See this Field Exchange article for a description. RAM : RAM is a type of two-stage cluster sample. It is better than a SMART type sample because it uses implicit stratification in the first and second stage sample. This is also discussed in the RAM-OP article (see above). The mechanics of sampling is shown in this manual from Sierra Leone. RAM and PROBIT : Put together RAM and PROBIT provide estimate with a similar precision to a typical SMART survey of about three times the size (e.g. in tests n = 192 for RAM/PROBIT is at least as good as SMART with n = 545). LQAS 33x6 and 67x3 methods : My view on these methods is that the bulk of survey costs are in travelling to and from cluster locations. This means that the methods offer little savings over SMART type surveys UNLESS there is a very large data collection overhead fro individual samples. I hope this is of some use.

Normal user

29 Jul 2015, 16:51

Thanks for this Mark. From what you have said it looks like if we want to measure the prevalence of stunting we either need to do a full SMART or the RAM PROBIT survey approach you mention. However, if measuring stunting thresholds and not prevalence an LQAS of 16 x 12 would be suitable for anthropometric indicator and more cost effective then the 33x 6 or 67 x 3 which add cost by increasing clusters but do not allow estimation of point prevalence. To put this in context we're trying to assess whether there has been a positive trend in stunting reduction and if possible, to what degree, in our target population group. We assume rates will decrease (a one-sided hypothesis), but obviously recognise that a plethora of external variables (such as rains, disease patterns etc.) will also play a role in dertmining trends and rates. Thus it is completely feasible that rates of stunting may increase from baseline to endline (although hopefully the intervention will have contributed to minimising any increases). With this in mind and what you have said in your post would you agree that either the SMART or the RAM PROBIT are the most suitable and cost effective options available to our programme? Thanks, Sinead

### Mark Myatt

Epidemiologist at Brixton Health

Frequent user

2 Aug 2015, 10:26

The sample size for an LQAS survey will depend on the LQAS sampling plan (decision rule) that you use. This will depend (mostly) on the prevalence thresholds that you use to define high, low, and moderate levels of prevalence (i.e. you prevalence classes). It is not common to need sample sizes much greater than n = 50. SLEAC (e.g.) uses LQAS to classify coverage into one of three classes and does this with acceptable levels of error with n = 40. A sample size of m = 16, n = 12 (i.e. overall n = 192) is probably overly large. For you application you could use either SMART or RAM PROBIT. The RAM PROBIT method would be more cost-effective than SMART (e.g. about half the cost for similar precision). You may also want to consider using mean HAZ as your indicator. This would allow you to use a much smaller sample size (i.e. n <= 100). You could then do many surveys (e.g. 4 or 6 per year) so you can create a time-series showing seasonal effects and longer term trends). I hope this helps.

Normal user

4 Aug 2015, 11:57

Hi Mark - thanks so much for this reply it is very helpful. With regards to analysing the findings of a RAM PROBIT survey, is there a data analysis package available for this (I note in your article on RAM - OP in Field XChange it states that progress is being made in developing one)? If there isn't one available can you point us in the direction of instruction on PROBIT analysis methods so we can consider what is involved in this before making a decision between SMART and RAM PROBIT? Thanks again, Sinead

### Mark Myatt

Epidemiologist at Brixton Health

Frequent user

5 Aug 2015, 09:17

We have software for the full RAM-OP survey. This covers a wide range of indicators and not just GAM, MAM, and SAM by PROBIT. This could be (quite quickly) adapted for your purpose. A simpler approach would be to use standard statistical software (e.g. SAS, STATA, SPSS, EpiInfo) that can analyse data from complex samples such as a two stage cluster sample. The ability to handle complex sample data is important for the calculation of confidence intervals. The procedure is: (1) Use your statistical software of choice to calculate the mean HAZ with the associated lower and upper 95% confidence limits and standard deviation (I'll call these meanHAZ, lowerHAZ, upperHAZ, and sdHAZ below). (2) Use a speadsheet such as Excel. The formulae: ``` =NORMDIST(-2,meanHAZ,sdHAZ,TRUE) =NORMDIST(-2,upperHAZ,sdHAZ,TRUE) =NORMDIST(-2,lowerHAZ,sdHAZ,TRUE) ``` yield the prevalence and lower and an upper 95% CI for the prevalence of stuntedness. Use "-3" for severe stuntedness. I hope this is useful. BTW : I would still consider many frequent smaller surveys using the mean HAZ as the main indicator.

Normal user

5 Aug 2015, 16:44

Thank you so much for this Mark.