Tools · People analytics
Survey Signal Finder
Stop guessing what moves the needle — see what actually does, with the math to back it.
The method
Key-driver analysis (correlational driver ranking with multiplicity and collinearity discipline)
The engagement survey closes and the executive team wants to know what to fix first. The default move — rank the items by score and attack the lowest — answers the wrong question. The lowest-scoring item is often just the hardest to please. The item that actually moves retention may be sitting quietly mid-pack.
Key-driver analysis is the corrective. In People Analytics For Dummies, Mike West's sequence is deliberate: measure fuzzy ideas with well-designed surveys, then prioritize with key-driver analysis — relate each item to the outcome you care about instead of admiring the items' averages. An item's score tells you where people are unhappy; its relationship to the outcome tells you where unhappiness costs something. Those are different lists, and only the second one deserves budget.
Edwards, Edwards, and Jang's Predictive HR Analytics walks the engagement and turnover cases click by click — correlation first, then regression — with the explicit aim of pulling back what they call the magic curtain of people analytics: none of this is exotic, and all of it is checkable. Keith McNulty's Handbook of Regression Modeling in People Analytics supplies the honesty rules the checkable version needs. People data sets are small and consequential, so inference discipline beats prediction bravado: check the assumptions, respect what the sample has the power to detect, and never read two collinear drivers as two independent levers when they are carrying the same variance.
The limits are where this method earns or loses its keep. Correlations are not causes — a driver analysis nominates suspects, it does not convict. Testing dozens of drivers manufactures false positives unless multiplicity is corrected. And a driver nobody can act on is trivia, however strong its coefficient.
Here the math runs in code — pairwise correlations, Fisher confidence intervals, Bonferroni multiplicity checks, collinearity flags, and noise called noise — while the language model writes only the one-paragraph story the evidence earns, plus the do-not-conclude list your executive readout needs. Raw rows or precomputed correlations both work, so the responses never have to leave your organization.
The books behind this tool
- People Analytics For Dummies — Mike West
- Predictive HR Analytics — Martin R. Edwards, Kirsten Edwards & Daisung Jang
- Handbook of Regression Modeling in People Analytics — Keith McNulty
How it works
Deterministic driver analysis (MF-182): zero-order correlations per driver (pairwise-complete), 95% Fisher CIs, Bonferroni multiplicity checks, signal/weak/uncertain/noise verdicts, collinearity flags, and variance shares — all in code (the LLM never does arithmetic). The LLM writes the exec-ready one-paragraph story earned by the evidence and the do-not-conclude list (causality, failed-multiplicity verdicts, collinear drivers as independent levers). Accepts raw respondent rows or precomputed correlations when data can't leave the org. Grounded in the people-analytics corpus.
You bring
{ outcome_key, records?|precomputed?, driver_keys?, context?, cluster? }
You get
{ computation (ranked drivers · CIs · verdicts · warnings), interpretation (narrative · do_not_conclude · next_steps), grounded_in, provenance }
Use it for
- →The Friday exec readout: engagement survey in → the one paragraph that matters, noise called noise
- →Privacy-preserving mode: send only per-driver correlations (r, n) — raw responses never leave the org
- →Top signal → engagement-action-planner for the intervention design
See it work
example output300-respondent engagement survey with four candidate drivers — which are real signals and which are noise, before the Friday exec readout.
| Driver | r | 95% CI | Verdict | Survives multiplicity |
|---|---|---|---|---|
| manager_support | 0.824 | [0.784, 0.857] | signal | ✓ |
| growth | 0.418 | [0.32, 0.507] | signal | ✓ |
| snack_quality | 0.067 | [-0.046, 0.179] | noise | — |
| commute | 0.019 | [-0.094, 0.132] | noise | — |
Two things reliably track with engagement in this survey, and one of them dominates. Manager support is by far the strongest thing we measured (r=.82), accounting for roughly 79% of the explained variance — it stands apart from everything else and clears both statistical significance and the multiple-comparison bar. Growth opportunity is the clear second signal (r=.42, a moderate association explaining about 20% of variance) and also survives adjustment. The two perennial 'perks' hypotheses did not pan out: snack quality (r=.07) and commute (r=.02) are statistically indistinguishable from zero and read as noise. Important caveat: this is correlational survey data from 300 respondents — these are strong candidates for where to focus, not proof that changing them will move engagement.
Do not conclude
- Do not claim manager support or growth cause engagement — this is one wave of correlational survey data; the association is a candidate for intervention design, not proof an intervention works.
- Do not report snack quality (r=.07) or commute (r=.02) as even minor levers — both are noise, failed significance, and failed multiplicity; a marginal-looking number is not a finding here.
- Do not treat manager support and growth as fully independent, additive levers — drivers intercorrelate up to .38, so their variance shares overlap and cannot simply be summed into a total 'engagement budget.'
- Do not over-read the 79% variance share for manager support as a causal ceiling; shared-method bias (same survey measuring both support and engagement) can inflate correlations between self-reported constructs.
Next steps
- Design and pilot a manager-support intervention (e.g., structured 1:1 cadence, coaching for frontline managers) with a control/comparison group, since manager quality is the dominant signal — then measure whether engagement actually shifts rather than assuming it.
- Instrument growth opportunity more precisely for the next wave (career pathing, internal mobility, development access) to confirm the r=.42 signal replicates and to separate it from manager support given the observed intercorrelation.
- Run a relative-weights or dominance analysis on manager support and growth together to properly partition their overlapping (r up to .38) contribution before presenting either as a standalone lever.
- Stop investing survey real estate and budget rationale in snack/commute perks as engagement drivers — reallocate that attention to the manager-quality and development hypotheses that earned it.
Run it on your data
Call it on your own inputs — over the API, or hand it to your AI agent via MCP. Discovery is open; running it is metered.