Front-Door Criterion in WFM

From WFM Labs
The front-door criterion: the effect of Coaching on CSAT is identifiable through the mediator Adherence even though the confounder Motivation is unobserved.

Front-Door Criterion in WFM is a method from causal-diagram theory for identifying the causal effect of one variable on another even when the confounder between them cannot be measured, provided a suitable intermediate variable (a mediator) is available. It is the complement to backdoor adjustment: where the backdoor criterion works by measuring and adjusting for the confounders, the front-door criterion works by routing the analysis through the mechanism that transmits the effect. The criterion was formalized by Judea Pearl, who showed that an effect ordinarily considered unidentifiable could be recovered through a fully mediating variable.[1]

The problem it solves

In many WFM analyses the confounder is real but unmeasured. Suppose an operation wants to know whether a coaching program improves customer satisfaction. Agent motivation plausibly drives both who engages with coaching and their CSAT, so it confounds the relationship — but motivation is not captured in any system. Backdoor adjustment is impossible because the variable that would need to be controlled for does not exist in the data. The front-door criterion offers a way forward when there is a measurable variable through which the entire effect flows.

The three conditions

For a mediator M to license front-door identification of the effect of X on Y, three conditions must hold:[2]

  1. Full mediation: M intercepts every directed path from X to Y — there is no direct X → Y arrow that bypasses M.
  2. No confounding of X→M: there is no unblocked backdoor path from X to M (the unobserved confounder affects Y but not M directly).
  3. X blocks M→Y confounding: every backdoor path from M to Y is blocked by X.

When these hold, the effect of X on Y is computed in two stages: estimate the effect of X on M, estimate the effect of M on Y (adjusting for X), and combine them. The unmeasured confounder never needs to be observed because it does not touch the X → M link.

A worked WFM example

In the diagram, Coaching → Adherence → CSAT is the mechanism, while unobserved Motivation confounds coaching and CSAT. If coaching affects CSAT only by changing adherence (full mediation), if motivation does not directly change adherence, and if adherence's effect on CSAT is not separately confounded once coaching is accounted for, then the coaching→CSAT effect is recoverable: measure how much coaching shifts adherence, measure how much adherence shifts CSAT, and chain the two. This mirrors Pearl's canonical smoking → tar → cancer illustration, where the genotype confounder need not be observed.[3]

When it applies, and its limits

The front-door criterion is powerful but demanding, because its conditions are strong and often only approximately true:

  • Full mediation is rare. If coaching also affects CSAT directly — through morale, say — the no-direct-path condition fails and the estimate is biased. Most real WFM mechanisms have some direct component.
  • The mediator must itself be unconfounded with the outcome except through the cause. An unmeasured factor that affects both adherence and CSAT would break condition three.
  • It is a tool of last resort. Where a confounder can be measured, backdoor adjustment is simpler and more robust; where randomization is possible, an experiment is better still. The front-door criterion earns its place only when confounding is genuinely unobservable and a clean mediator exists.

Used carefully, it expands what can be estimated from observational WFM data; used loosely, it manufactures false confidence. As always, the assumptions live in the diagram, and the estimate is only as good as the diagram is true.

Maturity Model Position

In the WFM Labs Maturity Model™, front-door reasoning is an advanced identification capability.

  • Level 1–2 (Emerging / Foundational) — effects with unmeasured confounders are either ignored or estimated naively as correlations.
  • Level 3 (Progressive) — analysts recognize when a confounder is unobservable and know that a fully mediating, measurable mechanism can sometimes rescue identification.
  • Level 4–5 (Advanced / Pioneering) — front-door and backdoor identification are part of the standard toolkit, applied with explicit diagrams and sensitivity analysis, and embedded in how causal questions are posed to data-science partners.

See also

References

  1. ↑ Pearl, J. (1995). "Causal Diagrams for Empirical Research". Biometrika, 82(4), 669–688. doi:10.1093/biomet/82.4.669.
  2. ↑ Pearl, J. (2009). Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge University Press. ISBN 978-0-521-89560-6.
  3. ↑ Pearl, J., & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books. ISBN 978-0-465-09760-9.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Front-Door_Criterion_in_WFM&oldid=3331"