Skip to content
Playbook
AdvancedHRPeopleOps

Causal inference for people analytics — A/B, diff-in-diff, and the questions correlation can't answer

Most 'people analytics' is correlation in a dashboard. The questions leaders actually want answered — does our new manager program reduce attrition?

11 min read
On this page
60-Second Summary
  • Correlation ≠ causation, but you can earn causation with the right design.
  • A/B tests are gold-standard but rarely possible in HR (ethics, scale).
  • Difference-in-differences is the workhorse: compare change in treated vs control over time.
  • Regression discontinuity exploits sharp cutoffs (promotion thresholds, tenure bands).
  • Instrumental variables handle confounding when randomisation isn't possible.

Boards no longer accept 'engaged teams have lower attrition' as analysis. They want to know: does our intervention cause the outcome? That's a causal question, and causal questions need causal methods.

Why this matters

Every people initiative competes for budget against revenue-generating projects. 'We did this and the number moved' is not enough — the number might have moved anyway. Causal methods are how you defend the spend.

Four methods, plain English

MethodWhen to useExample
A/B test (RCT)You can randomise who gets the treatmentTwo onboarding flows assigned at random
Difference-in-differencesTreatment rolls out to some groups, not othersNew manager program in 2 BUs; compare attrition trends
Regression discontinuityThere's a sharp cutoff in eligibilityPromotion bonuses above a calibration score
Instrumental variablesNon-random assignment + a clean 'instrument'Office reopening timing as instrument for in-person days

Building the capability

  1. Hire or partner with one analyst with statistics training (not just SQL).
  2. Pick 2-3 high-stakes questions per year that deserve real causal analysis.
  3. Pre-register: write down the hypothesis and method before you look at the data.
  4. Publish internally with limitations explicit — credibility compounds.
  5. Connect to the evidence-based management loop.
The single biggest mistake

Running 12 correlations, picking the one that confirms your narrative, and presenting it as insight. That's not analytics — it's confirmation bias with a chart.

Written by Pawan Joshi.Sources cited inline.
First published 16 Jun 2026See site changelog →