Why this section matters
Section 1.5 reframes a long-standing debate in statistics for marketing analytics: should we prioritize interpretable models or predictive performance? The answer, in practice, is neither extreme. For causal marketing decisions, we need both.
The two cultures, adapted to marketing
Leo Breiman described two statistical cultures:
- Data modelling: explicit probabilistic structure, parametric assumptions, interpretable parameters.
- Algorithmic modelling: flexible, data-adaptive prediction with minimal structural assumptions.
In marketing analytics, this maps to a familiar tension:
- Prediction-first systems (churn models, recommendation engines, bid optimization) are effective at forecasting within the current environment.
- Decision-first questions (launch, pricing, media mix, market entry) require causal inference—not just prediction.
The key insight from Section 1.5 is that the real tension is not “prediction vs interpretation,” but rigid assumptions vs flexible methods, and prediction vs causality. Neither culture alone guarantees causal validity.
Why prediction is insufficient
Prediction answers: “What happens under the current system?”
Causal inference answers: “What would happen if we intervene?”
Marketing strategy lives in the second space. A model that predicts sales from ad spend does not, by itself, tell us whether changing spend causes sales to move. To get causal answers, we need:
- Explicit estimands (ATE, ATT, cohort-time effects)
- Explicit identification assumptions (parallel trends, unconfoundedness, factor structure, SUTVA)
- Diagnostics and sensitivity analysis to validate those assumptions
A synthesis: modern panel causal methods
Section 1.5 argues that modern panel methods provide a practical synthesis:
- Algorithmic flexibility handles high-dimensional confounders and heterogeneity.
- Design-based identification makes causal structure explicit and testable.
A representative example is double machine learning: ML models estimate nuisance functions, while causal identification ensures the final effect estimate is valid.
This approach prioritizes causal validity, using machine learning as a tool—rather than treating prediction as the objective.
A pragmatic workflow (four steps)
The section proposes a pragmatic, application-first workflow:
Specify the estimand and assumptions
- Define the target quantity (ATE, ATT, elasticity, $\tau(g, t)$)
- State the assumptions needed for a causal interpretation
Choose an estimator that matches the design
- DiD, synthetic control, interactive fixed effects, doubly robust ML
- Estimator must align with the identification strategy
Run diagnostics
- Pre-trends, placebos, balance checks, robustness, specification curves
Conduct sensitivity analysis
- Quantify how strong violations must be to overturn conclusions
This workflow integrates the strengths of both cultures: flexible modelling inside a disciplined causal design.
Practical takeaway
Marketing analytics needs more than accurate forecasts—it needs credible counterfactuals. The two-cultures framing is useful, but the real goal is causal decision support. Modern panel methods help bridge that gap by blending:
- Flexible machine learning for high-dimensional data
- Explicit causal assumptions for interpretability and validity
This is the foundation for the methods that follow in the MMM series.
References
- Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 1.5.
- Breiman, L. (2001). Statistical Modeling: The Two Cultures. Statistical Science.