Firstly, I would like to explain about the geographical area. Two upzilas covering 34 Refugee camps of which two are registered camps. The approximate population of the two registered camps is around 20,000 for each. On the other hand, all the makeshift camp hosting around 800,000 population of which 25 camps located in one upazila and the rest of the 7 camps located in another upazila. In terms of accessibility, some camps are far, which might lead to the lack of provision of services to the beneficiary. The population is homogeneous and arrived in the country at same time.
Considering the above situation, we have designed 3 SMART Surveys: two independent SMART for two registered camps; third one is for all Makeshift Camp. We are following the same over the past three rounds.
During first round, we were able to implement the surveys in all three areas. However, due to refusal in next two rounds we had to exclude one registered camp.
So, is there any necessity to divide makeshift camps as population is so high, to conduct two /three SMART surveys for this large area? Will the GAM prevalence vary if we divide the MS camps?
Also, can we design 1 SMART for two registered camps combined? Please note that these two registered camps are situated in two upazila and the population of one registered camp didn't take participate in the survey and the other registered camp has chronic water crisis?
If you think that the prevalences in the two registered camps is likely to be similar (did you see this in the first round of surveys?) then you could combine these into a single survey. It seems, from what you write, that you will only be able to a good survey in one of the registered camps.
You are assuming that a single estimate for all "makeshift" camps is reasonable and useful. I am not sure that is likely as these camps may be very different from each other in terms of populations and/or the presence, coverage, and quality of services that might act to control malnutrition. I'd be tempted to make strata for camps based on levels of service available, or by TFP and SFP workload, or by upazila and perform separate surveys in each stratum or upazila.
I am not much in favour of wide-area surveys without spatial sampling and small area estimates because I think that it is unlikely that the resulting estimate will apply over the whole survey area and the survey may leave pockets of worryingly high prevalence undetected.
I hope this is of some use.
Answered:
5 years agoThanks for your explanation. Very useful. Why SMART doesn't consider population for sample size calculation? Is it valid for large population? If yes why?
Answered:
5 years agoSMART surveys almost always use populations to locate clusters proportional to population size (PPS). This is a broadly accepted method.
I think that you mean using the total population aged 6-59 months in sample size calculations.
I had a quick look at the SMART METHODOLOGY (2006) manual which covers sampling in considerable detail. This does not mentions adjusting sample sizes to account for population sizes. I also looked at the ENA for SMART manual (2012) and see that the calculated sample size is adjusted (corrected) for population size as an option in the ENA software and suggests that his should be doen when the population is less than 10,000 children.
The term "small population" needs some definition. Sample size calculations often assume that n observations are taken from a population of size = N with replacement. We usually violate this assumption be sampling without replacement. This violation is not usually considered important if:
Sampling proportion = n / N
is below about 0.1. Since we tend to sample from large populations we do not usually need to worry about small population sizes. We typically sample from a population of N = 100,000 adults (n ≈ 20,000 children) so would only worry if the sample size was above:
n = 20000 * 0.1 = 2000
Most sample sizes calculations assume a very large population. If you have a small population (e.g. <= 2,000 people) then the calculated sample size will be slightly larger than is needed. This is OK as the sample size will be more than sufficient to meet the desired precision.
The uncorrected sampel size will be something like:
n = p(1 - p) / (e / 1.96)^2
where n is the sampel size, p is the expected prevalence, and e is the desired precision. Using an example from Table 3 on page 46 of the SMART METHODOLOGY (2006) manual:
n = (0.1(1 - 0.1)) / (0.03 / 1.96)^2 = 384
which is the same sample size as is given in the SMART manual.
For a small population we can correct this using a "finite population correction" (FPC):
n = (n * N) / (n + (N - 1))
where n is the uncorrected sampel size and N is the total population size. Continuing the example above with a population of N = 2,000 children ... the sampling prportion is:
Sampling proportion = 384 / 2000 = 0.192
so we have a small population and should use
n = (384 * 2000) / (384 + (2000 - 1)) = 322
If we sampled n = 384 children we would have a sample size of at least n = 322.
We don't often bother with using a FPC with SMART surveys because (i) we seldom sample from small populations, (ii) we need to apply a sparate FPC during data analysis which can complicate data analysis, and (iii) the uncorrected sample size works well enough. It is, however, common to use an FPC for SQUEAC and SLEAC coverage surveys as there a usually small numbers of SAM cases (i.e. a very small population) in any population.
I hope this is of some help.
Answered:
5 years ago