Menu ENN Search
Language: English Français

Coverage prevalence survey for blanket SFP for chronic malnutrition

This question was posted the Coverage assessment forum area and has 6 replies. You can also reply via email – be sure to leave the subject unchanged.

» Post a reply

Anonymous 2232

Normal user

18 Mar 2013, 06:51

Our programme is a blanket supplementary feeding programme (BSFP) that is targeting very high rates of chronic malnutrition, the largest nutrition problem in the country. We are targeting all children 6-23 months in selected districts with RUSF and have been doing so since June 2011 in the first district that the programme started in.

Before the programme began, a baseline was done to get the rate of stunting for children under 24 months. Now we want to do a second cross-sectional survey, including nutrition indicators and programme coverage.

Specifically, the survey objectives are to: 1) determine rate of stunting among children under 59 months (to compare to future surveys and to include children who have participated in the BSFP but are now over 24 months of age) and rate of stunting of substratum of children under 24 months (to compare to baseline); 2) to determine programme coverage prevalence; and 3) to record other standard nutrition and food security indicators for context.

Since it is a BSFP and all children are eligible, assessing them for enrolment/participation should give us coverage prevalence while we are also surveying stunting prevalence. However, unlike acute malnutrition programmes, enrolment is not enough to gauge coverage since the RUSF is given once a month for a total of 18 months at village level, assuming that the child is enrolled at 6 months of age.

The question then is how to define "coverage". Is there any established standard for this type of BSFP? So far it is proposed to define a child as "covered" who has received the RUSF for at least 50% of the months in which he/she has been eligible since the programme started.

On sampling, it is proposed to stratify by 0-23 months and 24-59 months. As long as the 0-23 month stratum sample is large enough, and then the sample expanded proportionately to include the 24-59 month stratum, it should be possible to determine stunting rates for 0-23 and 0-59 months of age. As with the earlier baseline, the survey will use a two-stage cluster sample with village and household layers.

Recognizing that BSFP for chronic malnutrition and assessment of such programmes is an area still very much in development, your comments are most appreciated.

Mark Myatt

Consultant Epideomiologist

Frequent user

18 Mar 2013, 17:41

WRT “Rate of stunting” . . . I think you mean prevalence of stunting. The term “rate of stunting” is ambiguous as it could refer to prevalence, incidence, or to the angle between the observed and reference growth curves (i.e. how fast a child is becoming stunted).

Stunting and stunted : I think you need to be careful WRT to using the terms “stunting” and “stunted” (i.e. between process and outcome). Children who are stunted at about 24 months tend to remain stunted and track the growth curve until the pre-adolescence growth spurt.

0-23 months : I would avoid length measurement in the 0-6 month age-group. this has been discussed elsewhere on these forums.

Sampling (1) : I would probably simplify sampling by having an overall sample size large enough to give sufficient sample size in each age-group.

Sampling (2) : I’d also consider using a PROBIT estimator for prevalence of low H/A since this will reduce the sample size requirement considerably. I am not clear why you want to work with prevalence when you could work with the HAZ directly and test for effects using t-tests, ANOVA, or linear models. All of these approaches make more efficient use of data than the “recode and tabulate” approach.

Sampling (3) : I would not use PPS sampling for coverage since PPS places the sample in the most populous centres and tends to overestimate coverage. See this paper.

Coverage : The problem WRT the definition of coverage is interesting. I think you will need to be careful about the maternal recall over such long periods. There will likely be a lot of error which may increase with increasing age. You might look to try a shorter recall period and define “covered” as (e.g.) two of the previous three months. I’d be tempted just to use the current month. I suggest some piloting of this.

I am not sure if this is help or just hand-waving.

Anonymous 2232

Normal user

7 Jun 2013, 04:57

Hi Mark, very belated thanks, but your comments were a huge help to shape the final survey design and protocol.

We are almost a week into the survey and have run into an urgent sampling problem. As a quick background, we used CSAS. Our district was divided into 8 quadrants, 8 villages selected by proximity to the center of the quadrant. We expected about a sample of about 300 children 6-59 months (seeking every single child in this age range in each village), but in fact after using every method of identification including door to door visits and cross checking against available village lists, the total number is going to be considerably lower -- around 200.

The multivariate analysis looks to be challenging because as noted earlier, the definition of 'coverage' is not simple. We have several indicator questions for this, but so far we are seeing sharing of food plus overall compliance with daily ration not looking as good as expected. So the overall definition of coverage (food received + ration eaten) is a) problematic and b) possibly low, so we are running the risk of not having a large enough 'covered' sample. We also have a good number of confounders measured per child: wealth index, WASH, food consumption/IYCF score, maternal education, household size, child health, maternal health, maternal education, low birthweight.

So the question is:
Do we need to increase the sample size? If yes, to what? How do we do that with CSAS, as we can't possibly visit another 8 villages? We have already randomly ordered the second-nearest quadrant center villages, but time, resources, logistics and survey fatigue mean that it would be far more feasible and strategic to select 1-2 of them.

Thanks (in advance!) this time.

Anonymous 2232

Normal user

7 Jun 2013, 06:09

One more thing: even though we don’t expect to see any change in prevalence of stunting we would still like the survey to give prevalence of stunting for children 6-23 months and children 6-59 months for future use. We earlier calculated this sample size at 306 children, however as you previously suggested, maybe there are ways to reduce the required sample size by using an estimator??

Mark Myatt

Consultant Epideomiologist

Frequent user

7 Jun 2013, 08:31

To address your second problem first ...

The alternative estimator is "model based". It relies on the frequentist "dogma" that:

    probability = proportion

We can reverse this so that we have:
    proportion = probability

This is a perfectly legitimate reversal.

The problem now becomes one of estimating probability. This required us to specify an appropriate probability model for our variable of interest. The appropriate model for H/A is the normal distribution. We then estimate the probability of a child will have (e.g.) a HAZ < -2. If we can do that we have prevalence of HAZ < -2 since:

    probability = proportion = prevalence

The normal distribution is fully described by two parameters. These are the mean and the standard deviation (SD). With small sample sizes we prefer to use robust (i.e. resistant to outliers) estimators for the mean and SD. These are:
   estimate of population mean = sample median   
   estimate of population SD = sample IQR / 1.34898

where:
    IQR = upper quartile - lower quartile 

Once we have these we can use the cumulative normal probability function (available in most stats packages and spreadsheets) to calculate the probability (prevalence) we are interested in.

Here is an example ... assuming (from our sample), we calculate:

    median = -1.6
    IQR / 1.34898 = 1.4

In Microsoft Excel (e.g.) we would use:
    =NORMDIST(-2;-1.6;1.4;TRUE)

which gives:
    0.3875484811

which is an estimated prevalence of 38.8%.

In R (e.g.) we would use:

    pnorm(-2, mean = -1.6, sd = 1.4, lower.tail = TRUE)

which gives:
    [1] 0.3875485

which is (also) an estimated prevalence of 38.8%.

If you need to calculate a 95% CI (we usually need to do this) then we first calculate a 95% CI on the median and use that. An approximate formula for a 95% CI on the median which is safe for long-tailed (i.e. somewhat non-normal situations) is:

    median +/- 1.58 * (IQR / sqrt(n))

This formula is from:
    Velleman PF, Hoaglin DC, Applications, Basics, and Computing
    of Exploratory Data Analysis, Duxbury Press, Boston, 
    Massachusetts, USA, 1981

with the example data and a sample size of n = 200 we have:
    -1.6 + 1.58 * (1.9 / 14.14) = [-1.4;  -1.8]

In Microsoft Excel we have:
    =NORMDIST(-2,-1.4,1.4,TRUE)
    =NORMDIST(-2,-1.8,1.4,TRUE)

which give:
    0.4432015032
    0.3341175709

We have an estimate of 38.8% (95% CI = 33.4%; 44.3%).

You should check my arithmetic here.

BTW : This procedure is (somewhat confusingly) referred to as PROBIT and is is used in RAM and S3M type surveys.

I hope this helps.

Mark Myatt

Consultant Epideomiologist

Frequent user

7 Jun 2013, 09:15

Anonymous 2232

Normal user

9 Jun 2013, 07:26

Thanks very much for your help!

Back to top

» Post a reply