Mediation Analysis in WFM
Mediation Analysis in WFM is the decomposition of a causal effect into the part that flows through an intermediate variable (the indirect effect) and the part that does not (the direct effect). Where a causal diagram identifies a chain — X → M → Y — mediation analysis quantifies how much of the effect of X on Y is carried by the mediator M. For workforce management, the question is constant: a program "works," but through what mechanism, and how much of the result would survive if the mechanism were removed?[1]
Direct, indirect, and total effects
In the simplest model, three paths matter: the effect of X on the mediator (a), the effect of the mediator on Y (b), and the direct effect of X on Y that bypasses the mediator (c′). The indirect effect is the product a × b; the total effect is the sum of the direct and indirect effects, c′ + a × b.[2] A mediator may carry all of an effect (full mediation, c′ ≈ 0), part of it (partial mediation), or none. Modern treatments express these quantities as natural direct and natural indirect effects within a formal causal framework, which makes the assumptions explicit and extends the idea to nonlinear and interacting effects.[3]
WFM examples
- Coaching → adherence → CSAT. If a coaching program raises customer satisfaction, mediation asks how much of that gain runs through improved adherence versus other routes (morale, knowledge). If the effect is fully mediated by adherence, then adherence is the lever to manage; if much is direct, adherence is the wrong target.
- Occupancy → burnout → attrition. The effect of occupancy on attrition may run largely through burnout. Knowing the split tells whether to manage occupancy directly or to intervene on the burnout mechanism.
- Automation → volume mix → cost. The cost effect of automation is partly direct (deflected contacts) and partly indirect (a changed mix that shifts handle time on the residual work). Attributing the saving correctly requires separating the two.
The cautions
Mediation is powerful but rests on strong assumptions, and casual mediation analysis is a common source of wrong conclusions:
- The mediator is post-treatment. Controlling for a mediator to estimate a total effect is a mistake — it removes part of the very effect being measured, the chain-blocking result from d-separation. Mediation deliberately conditions on the mediator, so it answers a different question (the direct effect), not the total effect.
- Mediator–outcome confounding. Valid decomposition requires no unmeasured confounding of the mediator–outcome relationship. An unobserved factor that drives both adherence and CSAT biases the split — and unlike treatment confounding, randomizing X does not fix it.[3]
- It is not the front-door criterion. Both involve a mediator, but the front-door criterion uses the mediator to recover a total effect despite unmeasured treatment confounding; mediation analysis assumes the effect is identified and asks how it is split.
Maturity Model Position
In the WFM Labs Maturity Model™, mediation thinking marks the shift from "does it work?" to "why does it work, and which lever should be pulled?"
- Level 1–2 (Emerging / Foundational) — programs are judged only on total before-and-after movement; mechanisms are assumed, not measured.
- Level 3 (Progressive) — analysts ask which mechanism carries an effect, distinguish direct from indirect, and avoid controlling for mediators when they want the total effect.
- Level 4–5 (Advanced / Pioneering) — effect decomposition with explicit assumptions informs where to intervene, and mechanism estimates feed the design of automated and targeted programs.
See also
- Causal Diagrams (DAGs) in WFM
- d-separation
- Front-Door Criterion in WFM
- Causal Inference in Workforce Management
- Correlation and Causation in WFM
- A/B Testing for WFM Experiments
References
- ↑ MacKinnon, D. P. (2008). Introduction to Statistical Mediation Analysis. Lawrence Erlbaum Associates. ISBN 978-0-8058-3974-6.
- ↑ Baron, R. M., & Kenny, D. A. (1986). "The Moderator–Mediator Variable Distinction in Social Psychological Research". Journal of Personality and Social Psychology, 51(6), 1173–1182.
- ↑ 3.0 3.1 VanderWeele, T. J. (2015). Explanation in Causal Inference: Methods for Mediation and Interaction. Oxford University Press. ISBN 978-0-19-932587-4.
