Why examples matter

Abstract methods become clearer when tied to concrete decisions. Section 1.4 introduces three stylised—but realistic—scenarios that map directly to panel causal tools and the identification issues they address.

Example 1: Loyalty programme with staggered rollout

A retailer phases a loyalty programme across stores over several years. Early adopters are typically large, high-income locations where uptake is expected to be strong, while later adopters include competitive markets that need a defensive boost. This creates selection bias: treated stores differ from controls even before launch, so raw comparisons confound baseline differences with true programme impact.

The estimands are richer than a single average effect. Teams often want a panel-level ATT, cohort–time effects $\tau(g, t)$, and event-time dynamics $\theta_k$ to see whether effects build slowly as customers accumulate points. Two complications are common: anticipation (customers delay purchases when they hear the programme is coming) and spillovers (nearby non-programme stores lose traffic to programme stores or benefit from word-of-mouth). A practical analysis therefore layers staggered DiD and event studies with diagnostics (placebos, pre-trends), then extends to spillover-aware models when interference is material.

Example 2: TV advertising carryover and mediation

A consumer brand varies TV spend across markets and weeks. Endogeneity is central: spend typically increases when managers anticipate demand or respond to competitor moves. Even with fixed effects, time-varying confounders (weather, promotions, local shocks) can drive both spend and sales.

In TV planning, exposure is often summarized by GRP (Gross Rating Points). GRP aggregates delivery as the product of reach and average frequency in the target audience. Formally:

$$ ext{GRP} = \text{Reach (\%)} \times \text{Average Frequency} $$

Equivalently, GRP is the sum of rating points across all spots in the plan. For example, if a campaign reaches 40% of the target audience at an average frequency of 3, then $\text{GRP} = 40 \times 3 = 120$. This matters for MMM because GRP is a common exposure input, and its construction clarifies what variation is being modeled (reach vs frequency vs total delivery).

Carryover matters just as much as confounding. TV effects often decay over multiple weeks, so the estimand is a dynamic dose–response with a long-run multiplier rather than a single-week lift. Choosing the lag structure (geometric decay vs flexible lags) is part of the identification story. Mediation complicates reporting: if TV drives search and search drives sales, controlling for search estimates a direct effect, not the total effect. Clear estimand definitions and sensitivity checks keep the analysis honest. In practice, teams combine distributed lag models with synthetic control or high-dimensional adjustment to guard against spurious trends.

Example 3: Platform market entry

A delivery platform enters cities over time, while incumbents respond strategically. Entry is targeted: larger or faster-growing markets are prioritized, which makes treated cities incomparable to untreated ones. Entry timing is staggered, and the competitive response (price cuts, promotion bursts, commission changes) can partially offset the platform’s impact.

The analysis needs to separate platform effects from broader market dynamics. Staggered DiD and synthetic control are natural starting points, especially when pre-treatment fit is strong. Factor models help absorb common shocks (e.g., national demand shifts). Spillover models are useful when entry in one city affects neighboring markets or when incumbent platforms adjust across multiple geographies. The result is rarely a single number: heterogeneous effects by city size, restaurant type, or competitive intensity are often the most actionable outputs.

Takeaway

These three challenges illustrate why panel methods are more than statistical machinery: they are a design-based response to selection bias, dynamics, spillovers, and heterogeneity. Each scenario points to a different set of tools, diagnostics, and assumptions—setting the stage for the method chapters that follow.

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 1.4.
  • Abadie, A., Diamond, A., & Hainmueller, J. (2010). Synthetic control methods for comparative case studies. Journal of the American Statistical Association.
  • Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing. Journal of Econometrics.