Averages hide who is being left behind. This walks through disaggregating an outcome by subgroup in Sopact Sense, grading each cut, and finding the one input that makes a weak subgroup result reportable.
In short: A single average can look healthy while one subgroup quietly falls behind. In Sopact Sense you disaggregate the outcome by subgroup, see the gap to the mean for each cut, and grade every cut green, amber, or red — so equity shows up in the data instead of hiding in the average.
Point the Assistant at a clean dataset with persistent contact IDs so every response stays tied to one person across the cohort. Load your Decision Brief first — the decision, the audience, the outcome, the indicator, and your evidence standard — so the disaggregation answers a real reporting question rather than slicing data for its own sake.
Break [METRIC] change for [COHORT] by [DIMENSIONS]: change per subgroup, gap to mean; flag left-behind; small-n caution. Grade green/amber/red.
The prompt carries five elements. The dataset is the loaded cohort. By subgroup names the demographic dimensions you cut on. Gap to mean measures each subgroup against the cohort average. Small-n caution tells the model to flag any cut too thin to report. Grade G/A/R forces a green, amber, or red verdict on each cut so you can act.
Run on the Workforce Cohort dataset (DEMO 03) already loaded in Sopact Sense.
GRADE: green | 1 | Strong subgroup; amber | 1 | Small-n subgroup; red | 1 | Left-behind East site
Sense returns a graded cut for each subgroup. A well-evidenced subgroup comes back green. A subgroup resting on too few responses comes back amber — the result might be real, but it cannot be reported with confidence. The site sitting well below the cohort mean with no explanation comes back red: that is the group being left behind.
Take the lowest-graded element above and fix it using only what the program could realistically measure. Show the before → after grade and the single indicator/edit that moves it to green.
For the left-behind East site, the fix is small: add a short why-question at exit for East-site participants. That single indicator turns an unexplained red gap into an evidenced amber you can act on — and eventually a green you can report.
Create a 'missing & incomplete' report from this analysis in Sopact branding [or paste your website URL / brand guideline to apply your own]. List every element graded amber or red, what is missing, and the one input that fixes each. Lead with the decision this report informs.
Create a shareable link for this report and open it in a new tab.
Decide your small-n floor first. Set a minimum subgroup size before you run the cut, so the model flags thin groups consistently instead of you deciding case by case after the fact.
Gap to mean beats raw scores. A subgroup can post a decent number and still trail the cohort badly. Always read each cut against the mean, not on its own.
Don't over-slice. Cutting by five dimensions at once produces dozens of tiny, unreportable cells. Pick the two or three dimensions tied to your decision.
Pool before you bury. A thin subgroup isn't useless.
Pool it with an adjacent cycle or report it as indicative, not conclusive — then collect more before the next report.
Disaggregate the outcome by the dimensions that matter to your decision, measure each subgroup's gap to the cohort mean, flag any cut too small to report, and grade each one green, amber, or red. Sopact Sense does this in one prompt over a clean dataset with persistent contact IDs.
A small-n subgroup is a cut with too few responses to report confidently. It is graded amber because the result may be real but cannot stand on its own — you either pool it with another cycle or label it indicative until you collect more.
Read each subgroup's gap to the cohort mean. The group sitting well below the mean with no explanation is the left-behind one — and the fix is usually a single why-question added at exit for that group.
ChatGPT, Claude, and Gemini are fine for a quick test — but not for an answer you'll put in front of a funder or board. When it has to hold up, run it in Sopact Sense.
Try it in Sopact →
