Chapter 36 – Causal Inference: Turning Correlation into Credible Action

# Chapter 36 – Causal Inference: Turning Correlation into Credible Action ## 1. Why Causality Matters Causal inference moves the discussion from *what* is correlated to *what* will happen if we act. In business, this translates to decisions such as: | Decision | Correlation Insight | Causal Question | |----------|---------------------|-----------------| | Increase marketing spend | Higher spend → higher sales | Does increasing spend **cause** higher sales? | | Introduce a new pricing tier | Tier adoption → churn drop | Does the new tier **cause** churn to fall? | The answer to the causal question determines the confidence of the recommendation and protects against unintended consequences. ## 2. Core Concepts and Terminology | Term | Definition | |------|------------| | **Treatment** | The intervention or policy change we are interested in evaluating (e.g., a promotion). | | **Control** | The baseline scenario against which the treatment is compared. | | **Outcome** | The variable we want to change (e.g., revenue, churn). | | **Confounder** | A variable that influences both the treatment and the outcome, potentially biasing naive comparisons. | | **Counterfactual** | The outcome that would have occurred for the same unit under a different treatment assignment. | | **Average Treatment Effect (ATE)** | Expected difference in outcome between treated and control units. | ## 3. Design of Experiments (Randomized Controlled Trials) ### 3.1 Classic Randomized Trial 1. **Random assignment** ensures that, on average, treatment and control groups share the same distribution of observed and unobserved confounders. 2. **SUTVA** (Stable Unit Treatment Value Assumption) guarantees that one unit’s treatment does not affect another’s outcome. 3. **Analysis**: simple difference in means, or regression with treatment indicator. python # Python pseudo‑code for ATE estimation treated = df[df['treatment'] == 1] control = df[df['treatment'] == 0] ate = treated['outcome'].mean() - control['outcome'].mean() ### 3.2 Cluster‑Randomized & Stepped‑Wedge Designs * Useful when unit‑level randomization is infeasible (e.g., region‑wide campaigns). * **Stepped‑wedge** rolls out the intervention to all clusters over time, providing both control and treated observations. ## 4. Observational Studies: Overcoming the Lack of Randomization | Method | Key Idea | Typical Assumptions | |--------|----------|---------------------| | **Propensity‑Score Matching** | Match treated and control units with similar covariate profiles. | Conditional independence (no hidden confounders). | | **Inverse‑Probability‑Weighting (IPW)** | Weight units by the inverse of their probability of receiving treatment. | Correctly specified propensity model. | | **Doubly Robust Estimation** | Combines outcome regression and propensity weighting; consistent if either model is correct. | | **Instrumental Variables (IV)** | Use an external variable that affects treatment but not the outcome directly. | Exclusion restriction and relevance. | | **Regression Discontinuity (RD)** | Exploit a threshold‑based assignment rule. | Continuity of potential outcomes around the cutoff. | | **Difference‑in‑Differences (DiD)** | Compare pre‑post changes between treated and control groups. | Parallel trends assumption. | | **Synthetic Control** | Construct a weighted combination of control units to mimic treated unit pre‑intervention. | No interference between units. | ### 4.1 Propensity‑Score Matching Example python from sklearn.linear_model import LogisticRegression from sklearn.neighbors import NearestNeighbors # Step 1: Estimate propensity scores ps_model = LogisticRegression().fit(X, treatment) ps = ps_model.predict_proba(X)[:,1] # Step 2: Match treated to nearest control treated_idx = df[df['treatment'] == 1].index control_idx = df[df['treatment'] == 0].index nn = NearestNeighbors(n_neighbors=1).fit(df.loc[control_idx, 'ps'].values.reshape(-1,1)) _, match_idx = nn.kneighbors(df.loc[treated_idx, 'ps'].values.reshape(-1,1)) matched_control = df.loc[control_idx[match_idx.squeeze()]] # Step 3: Estimate ATE ate = df.loc[treated_idx, 'outcome'].mean() - matched_control['outcome'].mean() ## 5. Causal Diagrams & DAGs (Directed Acyclic Graphs) * DAGs formalize assumptions about relationships between variables. * **Back‑door criterion**: a set of variables that blocks all non‑causal paths from treatment to outcome. * **Front‑door criterion**: mediators that allow identification of causal effect when back‑door paths are blocked. mermaid graph LR T(Treatment) --> O(Outcome) C(Confounder) --> T C --> O ## 6. Practical Workflow for Business Causal Analysis | Phase | Tasks | Deliverables | |-------|-------|--------------| | 1. Problem Definition | Clarify the business question and define treatment, control, outcome | Question statement, causal diagram | | 2. Data Assessment | Identify data sources, assess quality, detect missingness | Data audit report | | 3. Design Selection | Choose experimental or observational approach | Design choice memo | | 4. Estimation | Apply chosen causal method, compute point estimates and confidence intervals | ATE estimates, variance analysis | | 5. Sensitivity Analysis | Test robustness to hidden confounding, alternative specifications | Sensitivity tables, E‑value calculations | | 6. Interpretation & Communication | Translate statistical results into actionable business insights | Executive summary, dashboard, policy recommendation | ## 7. Common Pitfalls & Mitigation Strategies | Pitfall | Why It Happens | Mitigation | |---------|----------------|------------| | **Hidden confounding** | Unmeasured variables bias estimates | Collect richer covariates, use IV or RD if possible | | **Incorrect model specification** | Wrong functional form or omitted interactions | Perform model diagnostics, use flexible methods (e.g., random forests) | | **Small sample size** | High variance, unreliable estimates | Aggregate over time, use Bayesian shrinkage, bootstrapping | | **Violation of SUTVA** | Spillover effects between units | Use cluster‑level analysis, include network metrics | | **Non‑parallel trends in DiD** | Treatment and control evolve differently | Test pre‑trend equivalence, adjust with covariates | ## 8. Ethical and Governance Considerations * **Transparency**: Document assumptions, data sources, and code. * **Fairness**: Evaluate treatment effect heterogeneity across protected groups; avoid exacerbating inequities. * **Privacy**: Ensure de‑identification when using sensitive data, comply with GDPR/CCPA. * **Responsibility**: Validate causal claims with domain experts before policy rollout. ## 9. Case Study: Launching a Loyalty Program 1. **Business Question**: Does enrolling customers in a loyalty program reduce churn? 2. **Data**: CRM logs (enrollment dates, demographics, transaction history), churn dates. 3. **Method**: Propensity‑score matching + DiD. 4. **Result**: ATE = –12% churn reduction (95% CI: –15% to –9%). 5. **Action**: Scale program to all active customers, monitor for diminishing returns. ## 10. Conclusion Causal inference equips businesses with the rigor to move beyond surface correlations and make decisions that truly impact outcomes. By combining thoughtful design, robust estimation, and clear communication, analysts can transform data into actionable strategy—turning every model into a living component of the continuous loop of observation, learning, and action.