# Coverage estimate calculation - Final stage - Help again

This question was posted the Coverage assessment forum area and has 14 replies. You can also reply via email – be sure to leave the subject unchanged.

### Géraldine LE CUZIAT

ACF

Normal user

14 Dec 2011, 18:56

Dear all,

I hope you are doing fine.

Well, I still have an issue and need a bit of guidance. We are in the very final stage of the SQUEAC investigation. I have just compiled the No of current SAM cases attending the programme (n=59), No of current SAM cases not attending the programme (n=82). The total no of current SAM cases makes 141 - which is way too high for a denominator and can not be entered in the Bayes software !!

any suggestion ?

Thanks in advance for your support,

Géraldine

### Mark Myatt

Consultant Epideomiologist

Frequent user

15 Dec 2011, 10:39

We can change the software to handle larger sample sizes. It is open-source so you (or anyone else) can change that. The source code is here.

You need to change line 963 which starts:

scale .frameControls.likeN -variable likeN -from 1 -to 96

change the 96 to anything you like and recompile using a tclkit.

If there is a demand for a larger limit then I will make the change and compile for Windows and OS-X.

In the mean time I suggest that you use the hand calculation method as outline in this section of the SQUEAC handbook.

If you post the data here we could go through it as a "worked example".

### Saul Guerrero

Director of Nutrition

Technical expert

16 Dec 2011, 12:45

I think the idea of doing these calculations on paper here would helpful.

Having said that, I also think that this is not an isolated event. I know that MSF-S had this exact same issue in India about a year ago. As we continue to do more and more of these in places with high caseloads, the need for a higher denominator may come up again.

Just a thought.

S

### Mark Myatt

Consultant Epideomiologist

Frequent user

16 Dec 2011, 16:58

OK. Give me an upper limit and I will make the changes, compile for Windows and OS-X and put it all up on the web.

So ... let me know and I will do what needs to be done.

### Saul Guerrero

Director of Nutrition

Technical expert

16 Dec 2011, 17:06

Hi Mark - thanks for offering to do this (I have spent some time today trying to find a tclkit to do the changes..)

The highest number of cases found by MSF in India (in one area) was 140. In Myanmar it was 141. I think that setting the upper limit at 150 could be sufficient, but to avoid having to do it yet again in the future, perhaps we could do 200?

You know best though.

Hope that helps

S

### Lio

CMAM advisor

Technical expert

19 Dec 2011, 10:59

A quick thought,

Yes, when the caseload is likely to be high, it is quite frequent that the final sample is larger than the required sample size; therefore adjustment in the software is welcome. However, having a very large sample size is not necessary and will not really improve the precision. In my opinion, when large caseload is expected (large number of SAM Kids or Kids in the programme) are foreseen, it would save time and resources to reduce the number of villages to survey in the same quadrant or in the same area (depending what is the sampling method)

### Mark Myatt

Consultant Epideomiologist

Frequent user

19 Dec 2011, 18:19

I have made a new version (2.01) with an upper limit of 192 for the denominator sample size. You can get it from here. Let me know if this is OK,

### Mark Myatt

Consultant Epideomiologist

Frequent user

20 Dec 2011, 09:51

Finding a sample size larger than original required is not uncommon in a first survey using active and adaptive case-finding with a given set of survey staff. This is because the survey staff get better (quicker and better) quite quickly over time. Another reason for this is that you may do small-area surveys in phase II to investigate possible low coverage areas and these areas are not low coverage but low prevalence (i.e. low program case-numbers from these areas reflect low prevalence not low coverage). This can lead you to underestimate prevalence over a wide-area and to choose to select a larger number of villages than is needed to get the required sample size.

Probably the most common reason for getting a much larger sample size than required is unrecognised low coverage. People get to thinking that their program has good coverage. If they have low case numbers they can easily believe that this is due to low prevalence rather than to low coverage. From "inside" a program the two situations:

Low prevalence = Low case numbers High prevalence + Low coverage = Low case numbers

can be difficult to distinguish from each other and wishful thinking can cloud judgement so that evidence of low coverage is ignored (there is an example of this in the case-studies in the SQUEAC handbook).

The sample size prompting this question is four or five times larger than required. This suggests a large miscalculation on the part of the investigator. Everyone makes mistakes. It would be interesting to hear Géraldine's take on this.

A sample size such as *n* = 96 is conventional for coverage surveys of child survival interventions (e.g. EPI surveys are design to have an effective sample size, after accounting for expected design effects, of *n* = 96). I am not in favour in increasing the sample size limit in the BayesSQUEAC software since SQUEAC is supposed to be a low resource method. A sample size of much about *n* = 50 will usually be a waste of resources (I have done a few SQUEACs and cannot recall ever going above that). I have, however, increased this to *n* = 192 at Saul's request.

**A note on sample size and precision :** A larger sample size will improve precision but not as much as you might think. Moving (e.g.) from a sample sise of 100 to a sample size of 200 does ** not** double precision (or half the with of the credible or confidence interval). Instead, precision increase with the square root of sample size. This means that the doubling the sample size increases precision by about 1.4 times (i.e. the square root of 2). Moving (e.g.) from a sample size of 100 to a sample size of 1000 improves precision by only about 3.2 times (i.e. the square root of 10).

### Mark Myatt

Consultant Epideomiologist

Frequent user

20 Dec 2011, 10:28

I have made a new version with an upper limit on the denominator of 256. You can download it it from here. Check if this is OK.

### Ernest Guevarra

Valid International

Technical expert

10 Jan 2012, 15:44

Dear Géraldine,

Thanks for your update. Your posts launched the coverage assessment forum well and started a lively discussion.

I just have two questions I wanted to ask you regarding the SQUEAC you implemented and that we are currently discussing about in this forum.

1. I am curious, what do you think is the reason why you suddenly got more cases than you expected? I remember, this post started when you asked for help on what to do if you don't reach your sample size and then after a few days, you shared good news that you were actually finding more than what you needed.

2. What is your case definition for this SQUEAC survey that you just did? Is it MUAC = 115 only or does it include a weight for height criteria? If the latter, was your case finding methodology active and adaptive case finding?

Thanks again for sharing.

### Mark Myatt

Consultant Epideomiologist

Frequent user

11 Jan 2012, 16:39

Géraldine,

Thanks for your detailed response. Have you considered writing this up as a case-study for publication in Field Exchange? I think you should.

### Mark Myatt

Consultant Epideomiologist

Frequent user

12 Jan 2012, 12:07

Géraldine,

Thanks for the clarification. Most of the SQUEAC material available to data has come from people involved with the development of the method. The major exception to this is this Field Exchange article which was written before the addition of the "stage 3" survey method. I think we need more user-produced documentation. I think, therefore, that it would be very useful for you to document your experiences using the new SQUEAC method in a Field Exchange article.

Mark