3. Identification strategies for observational process data

In a designed experiment, randomization severs the arrows from unobserved causes to the treatment, so a simple difference in means estimates a causal effect. On a production line, randomization is usually impossible: operators adjust setpoints in response to alarms, maintenance is scheduled when a machine looks degraded, and grade changes follow customer orders. The result is confounding, and the central question of this chapter is not "which estimator should I run?" but "is the causal effect identified at all, and by what argument?" Identification is a property of the causal graph plus the data you can realistically collect — not of the software.

We use the standard notation: do(X = x) denotes an intervention that sets the treatment X to value x, and the target quantity is the average treatment effect ATE = E[Y | do(X = 1)] − E[Y | do(X = 0)]. The four strategies below are ordered by how much they demand of your process knowledge and your data.

Backdoor adjustment: block the confounding paths

A backdoor path is any path from X to Y that begins with an arrow pointing into X. If a set Z blocks every backdoor path and contains no descendants of X, then

E[Y | do(X = x)] = Σ_z E[Y | X = x, Z = z] · P(Z = z)

Consider a heat-treatment cell where the furnace temperature X affects hardness Y, but ambient humidity Z raises both the operator's tendency to increase temperature and the oxidation rate that lowers hardness. Humidity is a confounder, so the crude regression of Y on X is biased. Conditioning on Z — by stratification, matching, or including it as a covariate — closes the backdoor path and recovers the causal effect.

Two practical cautions dominate industrial applications. First, do not condition on mediators. If temperature raises hardness partly by increasing the depth of the martensitic layer M, then adjusting for M removes part of the very effect you want and induces bias. Second, do not condition on colliders. If both temperature and humidity influence whether a part is scrapped, and you restrict the analysis to accepted parts, you open a spurious path between X and Z. A useful discipline is to draw the DAG with the process engineer before touching the data, mark which variables are measured, and only then choose the adjustment set.

Frontdoor adjustment: when the confounder is unmeasured

Sometimes the confounder is genuinely unobservable — for example, the general "tool wear state" that drives both the operator's feed-rate override X and surface finish Y. If, however, the effect of X on Y is fully transmitted through a measured intermediate M (say, cutting force), and no unmeasured variable confounds X–M or M–Y, the frontdoor formula applies:

E[Y | do(X = x)] = Σ_m E[Y | M = m] · P(M = m | X = x)

In practice this means estimating two regressions: the effect of feed-rate override on cutting force, and the effect of cutting force on surface finish, then composing them. The assumption is strong — no leakage path from X to Y that bypasses M — and it should be defended with process physics, not statistics. Frontdoor is powerful but fragile; use it only when the mediation is mechanistically complete.

Instrumental variables: exploiting exogenous variation

An instrument Z is valid if it (i) affects X, (ii) affects Y only through X (exclusion restriction), and (iii) shares no unmeasured cause with Y. In process settings, natural instruments include upstream raw-material lot variation, ambient conditions that shift a setpoint, or a scheduled maintenance event that changes a machine parameter. With a single binary instrument, the Wald estimator gives the local average treatment effect:

LATE = (E[Y | Z = 1] − E[Y | Z = 0]) / (E[X | Z = 1] − E[X | Z = 0])

Suppose a supplier change shifts average line speed by 4 m/min and the corresponding change in coating thickness is 0.9 µm. The Wald ratio is 0.9 / 4 = 0.225 µm per m/min — but only for the subpopulation whose speed actually responded to the supplier change (the compliers). Always report the first-stage strength; a weak instrument (small denominator) inflates variance and bias dramatically. The exclusion restriction is untestable and must be argued from engineering knowledge.

Do-calculus: a systematic toolkit

Do-calculus provides three rules for transforming expressions containing do() into ordinary conditional probabilities, using only the graph. Rule 1 lets you add or remove observations when they are irrelevant; Rule 2 converts interventions to observations when all backdoor paths are blocked; Rule 3 handles cases where an intervention on one variable does not affect another. Applied iteratively, these rules can identify effects that neither backdoor nor frontdoor handles directly — for instance, a graph with two unmeasured confounders where a measured mediator and a proxy variable together break the deadlock.

For day-to-day work, the practical value of do-calculus is diagnostic: it tells you whether an effect is identifiable before you invest in estimation, and it produces the exact adjustment formula. Software such as dowhy or causal-learn can enumerate identification strategies from a DAG, but the DAG itself must come from you.

Worked example: coating thickness on a roll-to-roll line

An engineer wants the causal effect of line speed X (m/min) on coating thickness Y (µm). The DAG includes a measured confounder — slurry viscosity Z — which affects both the operator's speed choice and the deposited thickness. A backdoor adjustment is therefore the first candidate.

Step 1 — Verify the adjustment set. Draw the DAG, confirm Z blocks all backdoor paths, and confirm Z is not a descendant of X. Viscosity is measured upstream of the speed decision, so it qualifies.

Step 2 — Estimate. Fit a linear model Y ~ X + Z on 240 production runs. Suppose the fitted coefficient on X is −0.42 µm per m/min with a 95% CI of [−0.55, −0.29]. The crude model Y ~ X gives −0.61, showing that viscosity confounding inflated the apparent effect by roughly 45%.

Step 3 — Check balance. Compare viscosity distributions across speed tertiles before and after weighting or matching. If the standardized mean difference drops below 0.1, the adjustment is behaving as intended.

Step 4 — Sensitivity. Ask how strong an unmeasured confounder would need to be to explain away the effect. If a plausible omitted variable (e.g., operator skill) would need an implausibly large association with both X and Y, the conclusion is robust.

Choosing among the four strategies

The recurring theme is that identification is an argument about the process, not a button in software. Before estimating anything, write down the DAG, name the unmeasured variables, and state which strategy you are invoking and why its assumptions hold on this line. If you cannot make that argument, no estimator will save the analysis — and the honest conclusion is that the effect is not identified from the data at hand.