Data Analyst

Analysis and interpretation

You turn a reviewed dataset into results that will survive being challenged. The platform is opinionated about what it will and will not compute, and knowing where those lines are is most of doing this job well here.

What you are responsible for

The analysis and its stated method. You are not responsible for approving publication, and you are responsible for refusing to produce a figure the design cannot support — the platform will refuse with you.

No screenshot on this pageThe three captures this guide carried — the analysis workspace, the evidence trail and the report planner — were withdrawn from publication on 2026-09-22. Read by eye, each rendered a real administrator’s name and the real VoiceInsights Africa organisation rather than the demonstration tenant they had been published as. Re-capturing them 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 an analysis surface exists to publish. Showing an adjacent screen here instead would make a caption claim something the image does not show, and that is worse than showing nothing.

The analysis workspace

An analysis runs against a dataset snapshot rather than against live rows, so a result can be reproduced later. Every run carries its method, its n, its weighting state and its interval.

What the platform will compute

Descriptives, proportions and means, with Wilson intervals for proportions and Welch–Satterthwaite for means.
Significance tests, disaggregation and cross-tabs where the design supports them.
Weighted and unweighted estimates, with the weighting state stated on every figure.

Every reported statistic is calculated by deterministic codeNo language model calculates a number here. Sentiment labels and topic tags are AI-assisted inputs, and a model may explain a result or draft an interpretation.

What it will refuse, and why that is the feature

A non-probability sample is refused representativeness, regardless of what the study declares and however large it is.
Such a design cannot be post-stratified. Weights that imply a sampling frame which never existed are refused rather than produced.
A figure cannot be quoted without its method and weighting state — the four travel together.
An undeclared methodology field reports NOT_PROVIDED rather than being omitted, so the gap is visible instead of invisible.

This will occasionally block youA refusal here is not a bug and is not something to work around by exporting and computing elsewhere. If the platform will not weight your citizen-feedback responses, the reason is that they have no sampling frame — and the number you would get outside the platform would be just as wrong, only without the warning.

Connecting results to evidence

An analysis result becomes usable when it is linked to the evidence it rests on. That link is what lets a reviewer follow a claim in a report back to the responses behind it.

Getting results out

Datasets and results export in the formats a team already works in. DOCX, PPTX and XLSX were compared inside the decompressed documents and carry identical facts. PDF renders as a valid document; its content was not machine-verified, and we do not claim it was.

When something goes wrong

The failures that actually happen, and what to do about each.
What you see	What to do
An analysis will not run	Check the study’s declared design. A mismatch between design and requested analysis is refused at validation with the reason.
Weights are refused	The design is non-probability. This is the refusal working, not a configuration problem.
A figure looks wrong against the field’s own count	Check the snapshot date. An analysis runs against a snapshot; live collection has moved on since.
An export is refused	Something the output claims does not resolve to an evidence record. The refusal names it.

Limits worth knowing before you plan around them

What this role cannot do, and what the platform does not offer

There is no automated triangulation or joint-display engine for mixed-methods work. Bringing a quantitative and a qualitative strand into one narrative is your work, supported by the platform rather than performed by it.
The platform does not detect the language of a response. A declared language is what the caller supplied.
No speech-recognition accuracy figure exists for any African language, because none has been benchmarked. Do not quote one.

NextStatistics RelatedSampling and weighting Who acts on itExecutive guide