Exercise · Advanced
The two programmes that never met
A note claims TSFP cures better than OTP — 75.3% against 72.6%. Adjust for admission MUAC and nothing changes. Then look at where the two programmes overlap, and find 106 children with no comparator at all.
A programme note is being circulated to the nutrition cluster:
Cure rates in 2024 were 75.3% in TSFP against 72.6% in OTP. TSFP is the more effective modality and should be expanded.
The arithmetic is right. The recommendation does not follow, and the reason is not the one most reviewers would reach for first.
The file
cmam-admissions-2024.v1.csv — 1,100 treatment episodes across six sites, with the
admission measurements and the discharge outcome.
What to do
One: put an interval on the claim. Compute the cure-rate difference between the two programmes with a confidence interval, restricted to the OTP and TSFP episodes. Say in one sentence whether the note’s claim survives it.
Two: adjust for severity. TSFP and OTP admit children on different criteria. Fit a logistic model for cure with programme, admission MUAC, age, sex and site, and report what adjustment did to the estimate.
Three: check whether the adjustment was possible. Tabulate admission MUAC in 5 mm bands by programme, and count the children in each band who have no comparator in the other programme. This is the step the note skipped and it is the one that decides the answer.
Four: re-fit on the overlap only, and say what question the trimmed model answers that the full one does not.
Check yourself
| Expected | |
|---|---|
| OTP and TSFP episodes | 526 and 559 |
| Crude cure rates | 72.6% and 75.3% |
| Crude difference, TSFP − OTP | +2.7 points, 95% CI −2.5 to +7.9 |
| Median admission MUAC, OTP and TSFP | 119 mm and 129 mm |
| Children in the 110–115 mm band | 106 OTP, 0 TSFP |
| Adjusted odds ratio, OTP vs TSFP | about 0.85, 95% CI 0.61 to 1.18 |
| Episodes in the 115–150 mm overlap | 898 of 1,085 |
| Overlap-only adjusted odds ratio | about 0.86, 95% CI 0.61 to 1.22 |
If your crude difference comes out much larger, the stabilisation-centre episodes are still in the comparison — there are fifteen of them, they are the oedema cases, and they belong to a third question.
Three questions
Two sentences each.
1. Adjusting for admission MUAC barely moved the estimate. Explain why that is not evidence that severity is unimportant here, using the band table rather than the coefficient.
2. One hundred and six children were admitted to OTP at a MUAC between 110 and 115 mm and no child in TSFP was. State what the model is doing when it produces an adjusted comparison for those children, and why the number it produces is not an observation.
3. The trimmed model answers a narrower question than the full one. Write that question out in the form a cluster coordinator would need to read it, and say which children the answer does not cover.
The point
The two programmes admit different children by design. OTP treats severe acute malnutrition and TSFP treats moderate; the admission criterion is the assignment rule, and the ranges barely overlap where it matters most.
Regression will produce an adjusted coefficient anyway. It fills the gap by extrapolating the MUAC slope into a region where one programme has no children at all, and nothing in the output says it did — no warning, no diagnostic, no drop in fit.
Adjustment cannot compare groups that do not overlap. The honest report says the register cannot rank the two modalities, names the 187 episodes outside the common range, and offers the within-overlap comparison as the narrower thing it can support — which, on these data, is also null.