Find root causes with causal inference

Perform mathematical causal inference on incident data using Pearl's SCMs

Incidents with rich quantitative data where you need defensible, reproducible, mathematical causal claims (not just narrative RCA), especially when distinguishing multiple confounding factors.

• · Observational data (metrics, logs, timestamps) with quantifiable variables
• · Directed acyclic graph (DAG) representing suspected causal relationships
• · Conditional probability tables (CPTs) for Bayesian network
• · Counterfactual scenarios ('what if we had done X')

• · Structural causal model (U, V, F) encoding causal mechanisms
• · Causal effect estimates via do-calculus or adjustment formula
• · Posterior probabilities of fault hypotheses (Bayesian fault diagnosis)
• · Counterfactual predictions ('would Y=y' if X=x')

• · Python with causal inference libraries (causalml, DoWhy, pgmpy)
• · Statistical computing tools (R, Julia optional)
• · Bayesian network inference engines

• · Quantifiable observational data from incident (not purely qualitative)
• · Incident has multiple candidate causes (not single-component failure)
• · Time-series or temporal ordering of events is known
• · Willingness to specify DAG based on domain knowledge
• · Sufficient sample size for statistical validity (else go qualitative)

• · Misspecified DAG → all downstream inference is wrong
• · Unmeasured confounders that backdoor criterion cannot adjust for
• · Insufficient data → estimates have high variance
• · Conflating correlation with causation by omitting causal graph
• · Over-interpreting counterfactuals without verifying against reality
• · Pearl's ladder rungs confused (association ≠ intervention ≠ counterfactual)

• · Grounded in Judea Pearl's peer-reviewed causal inference framework (2018, The Book of Why)
• · Do-calculus provides formal identifiability conditions (not heuristic)
• · Three-rung ladder (association/intervention/counterfactual) prevents ladder confusion
• · Bayesian networks tested on six industrial case studies (85-94% accuracy with correct DAG)
• · Frontdoor criterion solves unmeasured confounder cases when applicable
• · Explicit acknowledgment of DAG sensitivity (garbage in → garbage out)