From proxy metrics to causal questions

Marketing measurement has long relied on intermediate metrics like impressions, clicks, and page views. These are association-level signals: useful for optimization, but insufficient for answering “what if we change strategy?” A campaign can boost clicks without increasing incremental sales, and a loyalty program can shift timing rather than expand consumption. The causal revolution starts by insisting on incrementality.

Platform experiments help, but have limits

Large platforms now offer causal tools (e.g., conversion lift or geo experiments). These are valuable because they randomize eligibility or exposure and provide clean short-run estimates. But they also have limitations:

  • Proprietary implementations that are hard to audit
  • Aggregated outputs that obscure heterogeneity and dynamics
  • Platform-specific designs that limit cross-channel comparison
  • Narrow time windows that miss long-run effects

Panel methods extend experimental logic

Panel causal methods apply design-based reasoning to historical and cross-channel data. They can incorporate external signals (competitors, macro conditions, offline behavior), measure longer horizons, and model spillovers and dynamics explicitly. The key is transparency: assumptions are stated, diagnostics are explicit, and sensitivity checks are expected.

The new challenges for causal inference

Modern marketing data also creates new identification threats:

  • Algorithmic confounding: targeting rules use predicted outcomes, making treatment depend on past outcomes.
  • Privacy and measurement gaps: cookie loss, tracking limits, and aggregation increase missingness and bias.
  • Nonstationarity: preferences, platforms, and competitive landscapes change over time.
  • Interference at scale: spillovers and strategic reactions violate simple independence assumptions.

These problems raise the bar for credible inference and push methods toward robustness, sensitivity analysis, and triangulation.

Why this matters for MMM

MMM is most valuable when it moves beyond descriptive fit and toward causal interpretation. The causal revolution does not replace MMM; it re-anchors it in design-first logic. Panel methods and modern experiments make it possible to ask harder questions while staying honest about assumptions.

Table 1.2 at a glance

ExampleMain challengesMethods emphasizedTypical insights
Loyalty programme rolloutSelection bias, spillovers, dynamics, heterogeneityStaggered DiD, event studies, spillover models, causal forestsAverage lift that grows over time; modest spillovers; strong heterogeneity by store type
TV advertising carryoverEndogeneity, carryover, mediation, seasonalitySynthetic control, distributed lags, high-dimensional controls, robust inferenceImmediate lift with multi-week decay; mediated effects via search; competitive offsets
Platform entryUnique units, staggered timing, competition, general equilibriumSynthetic control, synthetic DiD, factor models, spillover modelsAverage gains with large heterogeneity; competitive response reduces impact; category expansion possible

Takeaway

The future of marketing measurement is causal, not just predictive. Platforms supply experiments, panels supply flexibility, and credible inference depends on explicit assumptions and diagnostics. That is the core shift this section highlights.

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 1.7.
  • Lewis, R., & Rao, J. (2015). The unfavorable economics of measuring the returns to advertising. Quarterly Journal of Economics.
  • Chernozhukov, V., et al. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal.