The core point
Marketing data is not just another application of standard econometrics. It has structural features that routinely violate assumptions behind off-the-shelf estimators. Section 1.8 highlights four differences that make marketing a distinct causal inference domain.
1) Targeting bias: endogeneity by design
Marketing treatments are strategic. Ads, coupons, and retention offers are aimed at people most likely to buy or churn. That makes treatment a function of predicted outcomes. In this setting, naive comparisons can be badly biased (even sign-flipped). Credible designs need explicit identification strategies and diagnostics, not just more covariates.
2) Interference is normal, not exceptional
Marketing actions spill across units: social referrals, geographic spillovers, and competitive reactions are the rule. SUTVA often fails. Designs must model spillovers directly (exposure mappings, cluster designs, spatial models) or risk misattributing effects.
3) Large-N, moderate-T, high-sparsity panels
Marketing panels often have millions of units with moderate time horizons and infrequent outcomes. This structure breaks many time-series tools and favors low-rank factor models and matrix completion approaches that borrow strength across units.
4) Complex dynamics: adstock and wear-out
Marketing effects are not instantaneous. Carryover, saturation, and habit formation create dynamic responses. Estimands must reflect lag structure and stock-and-flow behavior, and the identification strategy should be compatible with those dynamics.
Table 1.3: Structural differences across domains
| Feature | Clinical / Biostats | Financial Econometrics | Marketing Analytics |
|---|---|---|---|
| Primary goal | Efficacy / Safety | Risk / Forecasting | Incrementality / ROAS |
| Assignment | Randomised (RCT) | Systemic / Exogenous | Strategic (Targeting) |
| Interference | Rare (except vaccines) | Market-wide equilibrium | High (Social / Spatial) |
| Data shape | Small N, Short T | Small N, Long T | Large N, Moderate T |
| Sparsity | Low (complete records) | Low (continuous trading) | High (infrequent purchase) |
| Key method | Survival / Mixed Models | Time Series / GARCH | DiD/Event-Study, Synthetic Control, Matrix Completion |
Takeaway
Marketing analytics sits in a unique data regime where targeting, spillovers, sparsity, and dynamics are the norm. Panel causal methods are built to operate in this regime, but only when their assumptions are made explicit and tested with diagnostics.
References
- Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 1.8.
- Blake, T., Nosko, C., & Tadelis, S. (2015). Consumer heterogeneity and paid search effectiveness. Marketing Science.