Causality in people decisions
12 min read
Lesson objectives
- Construct a causal diagram and identify confounding before recommending a workforce intervention.
Please sign in to use all the tools
Once you sign in, the reading tools are yours: larger or smaller text, typeface, line spacing, focus mode, highlighting and bold, and notes saved to your account and gathered in My study.
A people analytics team finds that branches offering more training have higher employee turnover. The quick conclusion, 'training makes people leave,' could cancel a useful programme. The same branches may have faced a difficult technology change that caused management to increase training while also raising pressure and departures. Or branches with existing retention trouble may have chosen to train more. Association describes variables moving together in observed data. Causation asks what would have happened to the same people or branches had training alone changed while other important conditions remained comparable. We cannot observe both states for the same individual at the same time, so we design a comparison that approximates the counterfactual. CIPD distinguishes correlation and causation in people analytics, but a real programme decision needs the relationships in this particular case, not a repeated slogan.
Draw a small causal map in words before fitting a model. Workload pressure may increase both training and departure: it is a common cause that confounds the relation. Previous experience may affect both the chance of being selected for training and the chance of leaving. Satisfaction measured after training may lie on training's causal pathway; adjusting for it as if it were a prior characteristic could remove part of the effect of interest. Requiring six-month survival for inclusion excludes early leavers and creates survivorship bias. The map is useful only if it yields a list of pre-intervention drivers of assignment and outcome, and a timeline showing that the proposed cause preceded the result. Ask which difference between groups could alter the interpretation, and which variable must not be controlled away because it is an effect of training or of the sampling rule.
An observational comparison is not automatically worthless, but it needs tests. Compare prior tenure, job type, manager, workload, and pre-training performance across groups. If those most at risk of leaving were selected for training, a raw difference may exaggerate harm or reverse the true direction. Matching or weighting on measured pre-exposure factors can help; neither removes unmeasured confounding. If one team received training at a known date, difference-in-differences may help when pre-intervention trends are comparable, provided composition and simultaneous changes are checked. A method name is not a guarantee. Every method carries assumptions that can be examined only partly and limits that belong in the decision brief. Match evidence strength to decision stakes: a small reversible pilot can use preliminary evidence, whereas a broad people-impacting change calls for stronger support and harm guardrails.
Keep three outputs separate. Description reports the observed difference, sample, and time. Explanation lists plausible mechanisms and competing accounts, such as workload or selection of people at higher departure risk. Recommendation specifies what to do despite uncertainty: preserve training needed for safe work, test one defined onboarding form, monitor benefit and harm, and revisit the choice at a named date. 'We do not know the effect' is not an instruction to suspend decisions forever; it is a refusal to credit an unproved cause with an observed difference. A positive result in one branch also does not automatically transport to every role. Transferability depends on similarity of tasks, context, and delivery mechanism.
Find a second story that could explain the result
A manager says mentoring reduced exits and cites participants against nonparticipants. Before accepting that explanation, map another route to the same outcome. Managers may have selected workers who were already more stable. People on easier shifts may have had more time to participate. In both stories, earlier conditions influence participation and departure. Start with what differed before the programme and seek suitable evidence. Request invitation dates, selection rules, and relevant pre-invitation work characteristics. Privacy and authorized use remain necessary while investigating, even when the purpose is better causal reasoning.
Write a short timeline: worker conditions before invitation, the offer, participation, and the outcome. This distinguishes prior characteristics from variables that may change because of treatment. Satisfaction measured after mentoring, for example, could lie on the programme's effect pathway. Comparing workers after fixing later satisfaction may change the question being estimated. A pre-invitation shift pattern describes a condition already present during selection. Place a variable in the story before adding it to a model. If records do not establish timing, disclose that rather than assuming a field is pre-treatment simply because it appears in the file.
Inspect who reaches final follow-up too. If the happiest participants answer a survey while leavers' data disappear, survey results do not represent the group that started. Keep counts of invited people, participants, and those with an observed outcome, with known missingness reasons. Compare follow-up completeness before interpreting satisfaction or retention. Where completeness differs, examine its consequences or improve source coverage. Respondent findings can be useful when explicitly labelled as such, but generalizing to all workers needs further evidence. Ask who is absent from the picture and how that absence could change the conclusion.
Improve design before rollout rather than attempting to repair comparison afterward. If resources cover half the eligible workers now and the rest later, discuss randomizing the first-wave offer. Establish eligibility and operating constraints beforehand and record who receives an offer. In the first analysis, compare outcomes by assigned offer, retaining follow-up for those who decline. That preserves the meaning of initial allocation. The effect of completing mentoring is a different question when completion varies. The randomized-evaluation reference explains comparison design; figures here are fictional teaching inputs, not published study findings.
Behaviour can change outside the offered group. A participant might share advice or a manager may reorganize work for the whole team. Consider this when designing comparison. If the waiting group also receives programme effects, the difference may no longer compare exposure with no exposure. Document how and when knowledge spreads where possible. Independent teams might support phased allocation if appropriate and operationally acceptable, but that choice needs methodological assessment of comparison size and design. Consult the specialist rather than claiming team separation solves everything. Clearly name the offer unit and the outcome-measurement unit.
Close the analysis page with an explicit claim ladder: describe the observed gap, state a possible explanation, and propose a design to test it. Use effect language when design and assumptions justify it, placing those conditions near the result. At tomorrow's meeting, present a serious competing explanation beside the preferred one and request evidence that could discriminate between them. This shifts disagreement from personal confidence to checkable evidence. A smaller pilot or delayed costly expansion may be the best provisional recommendation while comparison remains unresolved.
Lesson summary
A people analytics team finds that branches offering more training have higher employee turnover.