Project · Advanced
SMART survey analysis and plausibility report
Estimating acute malnutrition prevalence from raw anthropometry, then producing the plausibility report that decides whether the estimate can be published at all.
Context
- Organisation
- Nutrition cluster commissioning a 30-cluster SMART survey
- Problem
- The survey came back with a global acute malnutrition estimate that would trigger a scale-up, but one measurement team's clusters are markedly worse than the rest. Nobody could say whether that is a real geographic difference or a mis-set scale, and publishing the wrong answer either wastes a response or misses one.
- Audience
- Nutrition cluster and the national nutrition coordination body
- Decision it informs
- Whether to accept the survey, exclude a team's data, or repeat the fieldwork before the estimate informs a scale-up.
Datasets used
Deliverables
- Analysis notebook
notebooks/whz-prevalence.ipynb - Dashboard
dashboard/index.html - Technical documentation
docs/plausibility-report.md - Final report
reports/smart-survey-report.pdf - Presentation slides
slides/nutrition-cluster.pdf
The estimate is the easy half
Weight-for-height z-scores against the WHO 2006 standards, oedema overriding anthropometry, flags applied before prevalence — done carefully that gives global acute malnutrition near 10.3% and severe near 1.1%. Serious on WHO thresholds, below the 15% emergency level.
Two things go wrong before that number is trustworthy. Children under 87 cm were measured lying down, and length reads about 0.7 cm longer than height for the same child; skip the adjustment and every z-score for the younger half of the sample is biased. And SMART flagging (±3 SD from the survey mean) and WHO flagging (fixed −5 to +5 bounds) exclude different children, so the report has to say which rule produced the published figure.
The plausibility report is what the decision rests on
Team 3’s clusters show global acute malnutrition near 15.5% against 7 to 11% elsewhere, with a mean z-score about 0.4 lower across every cluster it visited. A real geographic effect does not follow a team around. Team 2 rounds: roughly 69% of its height readings end in .0 or .5 against about 20% for everyone else. Ages heap on whole years, with 80 children recorded at exactly 36 months against 28 and 17 either side.
None of that is visible in the prevalence estimate. All of it determines whether the estimate should be believed, and the recommendation this project delivers is about the survey, not about the children.