Causal Inference

The fundamental logic of asking 'What would have happened if...?' In cold chain failure analysis, this means comparing the actual outcome of a shipment that experienced an excursion against the hypothetical outcome if the excursion had not occurred. Structural Causal Models (SCMs) encode these relationships as directed acyclic graphs (DAGs), allowing analysts to simulate interventions mathematically. For example, a counterfactual query might ask: 'Given that shipment S-451 arrived degraded, what is the probability it would have remained viable if the compressor had not failed at hour 12?' This framework isolates the specific contribution of a failure event from background degradation.

Visual and mathematical representations of causal assumptions. A DAG encodes expert knowledge about which variables influence others using nodes (variables like ambient temperature, door openings, compressor cycles) and directed edges (arrows indicating causal direction). The 'acyclic' constraint means no variable can cause itself, directly or indirectly. In cold chain monitoring, a DAG might model how ambient weather influences reefer unit load, which influences internal temperature, which influences product degradation rate. DAGs are essential for identifying confounders—variables that affect both the suspected cause and the outcome, creating spurious correlations that must be controlled for.

A formal mathematical framework developed by Judea Pearl for reasoning about interventions. The do-operator—written as P(Y | do(X=x))—represents the probability of outcome Y when we forcibly set variable X to value x, distinct from merely observing X=x. This distinction is critical: observing that shipments with frequent door openings have more excursions does not prove door openings cause excursions; a common cause like poor loading procedures could drive both. Do-calculus provides three rules for transforming observational data into interventional estimates when a valid DAG is available, enabling causal effect estimation without costly randomized experiments.

A technique for estimating causal effects when unobserved confounding is present. An instrument is a variable Z that: (1) affects the treatment X, (2) affects the outcome Y only through X, and (3) shares no common causes with Y. In cold chain logistics, a valid instrument might be distance from distribution center—it affects delivery time (treatment) but is unlikely to be correlated with unobserved factors like packaging quality that also affect product degradation (outcome). Two-stage least squares (2SLS) regression is the classic estimation method, though modern approaches use deep learning for non-linear instrumental variable estimation.

A quasi-experimental design that estimates causal effects by comparing the change in outcomes over time between a treatment group and a control group. The key assumption is parallel trends: in the absence of treatment, both groups would have followed the same trajectory. In cold chain monitoring, DiD could evaluate the impact of a new passive packaging solution by comparing temperature excursion rates before and after implementation across lanes that adopted the solution versus lanes that did not. Modern extensions like synthetic control methods construct a weighted combination of control units to create a more credible counterfactual when parallel trends are questionable.

Decomposes a total causal effect into direct and indirect pathways. In a cold chain context, a mediation analysis might reveal that the effect of ambient temperature on product degradation operates primarily through reefer compressor duty cycle (the mediator), rather than directly. This decomposition is crucial for CAPA design: if the effect is fully mediated, interventions should target the mediator (e.g., upgrading compressor capacity) rather than the distal cause. The product-of-coefficients method and causal mediation formulas based on counterfactuals provide formal estimation frameworks, with bootstrapping used for standard error calculation.