Sampling and weighting
Designs, sample size, weights, and the claims each supports.
Who this is for
Research and M&E officers, and analysts reading somebody else’s design.
Designs and what they support
Eleven designs are available. What matters is which ones support a population inference: probability designs do, non-probability designs do not, and the platform enforces that rather than trusting the label on the study.
Representativeness is not a setting
Sample size
Sample-size calculation runs Cochran, then the finite-population correction, then the design effect, then non-response — in that order — and returns every step, so the number can be checked rather than trusted.
Weights
Design weights, post-stratification and normalisation are available. Zero and negative weights are errors, not warnings. Every result states whether it is UNWEIGHTED, WEIGHTED or NOT_APPLICABLE.
Weighted variance uses the Kish effective sample size approximation. This is not design-based variance, and every weighted interval carries that as a stated assumption — it will understate uncertainty for a clustered design.
What this does not do
Stated here rather than discovered later
- Weighted variance is an approximation, and the platform says so on every weighted interval rather than in a footnote. For a clustered design it will understate uncertainty.
- No design is offered that would let a non-probability sample produce a population estimate. That is not a missing feature.