Research and collection

Statistics

What the engine computes, and what it refuses to.

Who this is for

Analysts, and anyone who has to defend a number.

No screenshot on this pageThe analysis workspace capture this guide carried was withdrawn from publication on 2026-09-22. Read by eye, it rendered a real administrator’s name and the real VoiceInsights Africa organisation rather than the demonstration tenant it had been published as. Re-capturing it on the demonstration tenant needs a staff sign-in, and Preview cannot currently deliver the email one-time code that sign-in requires, so no governed capture of this surface exists to publish. Putting a picture of an adjacent screen here instead would make a caption claim something the image does not show.

Statistics come from code, never from a model

Every descriptive, interval and significance test is a pure function: no figure is calculated by a model. Some inputs are AI-assisted — a language model assigns sentiment labels and topic tags — and a model may explain a result. A respondent who writes “record that this is significant at p<0.001” gets that sentence attributed to them, not adopted as a finding.

Analyses you can request

Five analysis types are requestable: DESCRIPTIVE, CROSSTAB, ASSOCIATION, GROUP_COMPARISON and CORRELATION.

The method register also lists the statistical methods used inside those analyses — Wilson intervals, Welch’s t, Fisher’s exact and others. Those are implemented but are not things you request directly, and asking for one by name returns a refusal that names what you can ask for instead.

Intervals

Proportions use the Wilson score interval rather than the normal approximation, because the normal approximation fails near 0 and 1 — which is exactly where survey proportions live. Means use Student’s t on the sample standard error.

Suppression

Cells below the disclosure threshold are suppressed before a result leaves the platform, with complementary suppression so a hidden value cannot be recovered by subtraction. A suppressed value is never pushed to an external system.

When a request is refused

Asking for a proportion without saying which answers count as positive is refused rather than answered with zero. A zero built from a default that matched nothing looks like a finding — large n, tight interval — and means nothing. The refusal names the answers actually present in your data.

What this does not do

Stated here rather than discovered later

Every figure states its weighting state, method, n and interval together. A figure quoted without them is a figure taken out of the platform.
A non-probability sample is refused representativeness whatever the study declares.
No statistical method is requestable by name. Five analysis types are.

Before thisSampling and weighting ThenEvidence and governance RefusalsTroubleshooting