Causal Inference

Causal inference addresses the fundamental problem of causal inference: we can never observe both the treatment and the control state for the same unit simultaneously.

• Counterfactual: The unobserved outcome—what would have happened if we hadn't changed a machine parameter.
• Factual: The observed outcome after the intervention.
• Goal: Estimate the counterfactual using statistical methods to isolate the true treatment effect from confounding noise.

DAGs are the visual language of causal assumptions, encoding expert knowledge about the manufacturing process.

• Nodes: Represent variables like sensor readings, machine settings, or environmental conditions.
• Edges: Represent direct causal relationships, not mere correlations.
• d-separation: A graphical criterion for determining if a path between variables is blocked, enabling the identification of confounding variables that must be controlled for.
• DAGs make assumptions explicit and auditable before any statistical test is run.

Developed by Judea Pearl, do-calculus provides a formal symbolic language to reason about interventions.

• P(Y|X): The probability of Y given we observe X. This is passive observation.
• P(Y|do(X)): The probability of Y given we actively set X to a specific value. This is intervention.
• Do-calculus provides three rules for transforming expressions with the do operator into standard conditional probabilities, allowing causal effects to be estimated from observational data when a valid DAG exists.

An instrumental variable (IV) is a tool used to estimate causal effects when a randomized controlled trial is impossible and unobserved confounding exists.

• Relevance: The instrument must be strongly correlated with the treatment (e.g., a specific raw material batch that forces a machine setting).
• Exclusion: The instrument affects the outcome only through its effect on the treatment.
• Exogeneity: The instrument must be independent of the unobserved confounders.
• In manufacturing, a valid instrument might be a pseudo-random supplier allocation that influences a process parameter but is otherwise unrelated to final quality.

DiD is a quasi-experimental technique that estimates a causal effect by comparing the change in outcomes over time between a treatment group and a control group.

• Parallel Trends Assumption: The key assumption is that, in the absence of the treatment, the difference between the two groups would have remained constant over time.
• Application: Comparing a production line that received a new AI optimization algorithm (treatment) against a similar line that did not (control), both before and after the deployment.
• The causal effect is the divergence in their trajectories post-intervention, netting out pre-existing differences and secular time trends.

A statistical concept of causality specific to time-series data, such as high-frequency sensor telemetry.

• Principle: A variable X 'Granger-causes' Y if past values of X contain information that helps predict Y above and beyond the information contained in past values of Y alone.
• Critical Caveat: Granger causality is a test of predictive causality, not true structural causality. It can be fooled by a common latent driver affecting both X and Y with different time lags.
• It serves as a valuable exploratory tool for hypothesis generation in complex industrial systems before applying more rigorous structural methods.

The fundamental distinction that correlation measures mutual relationship while causation establishes directional influence. In manufacturing, vibration amplitude may correlate with defect rate, but only causal analysis reveals whether vibration causes defects or both stem from a third factor like tool wear. Key differentiators:

• Correlation: Symmetric, no directionality
• Causation: Asymmetric, requires temporal precedence
• Spurious correlation: Statistical coincidence without mechanistic link
• Confounding variable: Hidden factor driving both observed variables

DAGs are graphical models that encode causal assumptions using nodes (variables) and directed edges (causal relationships) with no cyclic paths. In root cause analysis, a DAG maps how equipment settings, ambient temperature, and raw material properties influence final product quality. Essential properties:

• Nodes: Represent observed and unobserved variables
• Edges: Direct causal influence from parent to child
• d-separation: Criterion for identifying conditional independence
• Backdoor paths: Non-causal associations requiring blocking via adjustment

Counterfactuals answer 'what would have happened if' questions by comparing observed outcomes to hypothetical alternatives. In manufacturing, this means asking: 'If we had used batch A instead of batch B, would the defect have occurred?' Core concepts:

• Potential outcomes framework: Each unit has multiple possible outcomes under different treatments
• Individual treatment effect: Difference between factual and counterfactual outcomes
• Fundamental problem: Only one outcome is ever observed per unit
• Average treatment effect (ATE): Population-level causal estimate

The gold standard for causal inference where treatment assignment is randomized to eliminate confounding. In industrial settings, this translates to A/B testing on production lines—randomly assigning machine parameters to isolate their effect on quality. Design elements:

• Randomization: Breaks all backdoor paths by design
• Control group: Baseline for comparison
• Blinding: Prevents operator bias in measurement
• Covariate balance: Ensures groups are comparable on pre-treatment characteristics

An instrumental variable (IV) is a tool for estimating causal effects when randomization is infeasible and unobserved confounding exists. The instrument must influence the treatment but affect the outcome only through the treatment. Manufacturing example: using supplier lead time as an instrument to estimate the causal effect of raw material quality on final yield. Requirements:

• Relevance: Instrument strongly predicts the treatment
• Exclusion restriction: No direct effect on outcome
• Monotonicity: Effect on treatment is uniformly directional

A statistical hypothesis test for determining whether one time series is useful in forecasting another. Granger causality does not imply true causation but identifies predictive temporal precedence—critical for industrial process control loops. Application: testing whether spindle load Granger-causes surface finish deviation in CNC machining. Key points:

• Temporal ordering: Cause must precede effect
• Incremental predictability: Past values of X improve forecast of Y
• Limitation: Cannot distinguish direct from indirect causal chains
• Stationarity requirement: Time series must have constant statistical properties