Why a method selection map

Section 3.10 emphasizes a simple rule: the design of the panel study should determine the method. The map links assignment mechanism, treatment timing, data structure, and interference to the estimators that are credible under those design features. It does not prescribe a single best method. It clarifies tradeoffs, assumptions, and which diagnostics matter.

Randomized block designs

When treatment is randomized once and then held fixed, standard difference-in-differences (DiD) applies. The estimand is often the ATT, and event-time effects $\theta_k$ are natural when dynamics matter. Parallel trends holds in expectation by design. Inference can use clustered standard errors, randomization inference, or wild cluster bootstrap depending on the number of clusters.

Staggered adoption without heterogeneity

If units adopt treatment at different times and effects are truly constant across cohorts and time, two-way fixed effects (TWFE) recovers ATT under parallel trends and no-anticipation. TWFE is computationally simple but relies on constant effects, an assumption that is typically too strong in marketing settings. When effects vary, TWFE can place negative weights on some cohort-time effects, biasing the estimate.

Staggered adoption with heterogeneity

With heterogeneous effects, use modern estimators (Callaway-Sant’Anna or Sun-Abraham) that target cohort-time effects $\tau(g,t)$. These methods estimate $\tau(g,t)$ using not-yet-treated controls and then aggregate with user-defined non-negative weights $\omega_{gt}$:

$$ \tau_{\text{summary}} = \sum_{g,t} \omega_{gt} \hat{\tau}(g,t). $$

Different weights answer different questions: cohort-size weights yield ATT, event-time aggregation yields $\theta_k$, and calendar-time weights give period-specific effects.

Single treated unit or few treated units

When only one or a few units are treated, synthetic control constructs a weighted donor pool to match pre-treatment outcomes and uses the post-treatment gap as the effect. Inference relies on placebo tests.

Hybrid methods such as synthetic difference-in-differences (SDID) weight both units and time periods, often improving pre-treatment fit while maintaining a parallel trends interpretation in the reweighted data.

If treated and control units are exposed to shared shocks that violate parallel trends, factor models and matrix completion replace parallel trends with a low-rank structure for untreated outcomes. These methods rely on stable latent factor loadings. They work best when there are enough units and pre-treatment periods to estimate the low-rank structure.

Dynamic effects and carryover

When treatment has lagged or cumulative effects, distributed lag models and other dynamic panel models parameterize the effect path. These methods require assumptions about lag length and functional form. Misspecification can bias long-run multipliers even if short-run averages look reasonable.

Spillovers and interference

When SUTVA fails, estimators must model exposure mappings $h_i(D_{-i,t})$ explicitly. Network, spatial, and exposure-map designs separate direct and spillover effects. Cluster-randomized designs can identify total effects within clusters, but not separate components without additional structure. Partial identification can provide bounds when spillovers are uncertain.

Machine learning for nuisance functions and heterogeneity

Machine learning is most useful for nuisance functions (propensity scores, outcome models) and heterogeneity, not as a replacement for design. Double machine learning, causal forests, and high-dimensional controls preserve identification assumptions while relaxing functional form restrictions. Post-selection inference is important when variable selection is data-driven.

Takeaway

Let the design choose the estimator. When multiple methods are plausible, compare them for robustness. When the design points clearly to one method, use it even if it is computationally heavier or less precise. Credibility beats convenience.

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 3.10.
  • Callaway, B., and Sant’Anna, P. H. C. (2021). Difference-in-differences with multiple time periods.
  • Sun, L., and Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous effects.
  • Abadie, A., Diamond, A., and Hainmueller, J. (2010). Synthetic control methods for comparative case studies.
  • Arkhangelsky, D., et al. (2021). Synthetic difference-in-differences.
  • Bai, J. (2009). Panel data models with interactive fixed effects.
  • Athey, S., et al. (2021). Matrix completion methods for causal inference.
  • Chernozhukov, V., et al. (2017). Double/debiased machine learning.
  • Athey, S., and Imbens, G. (2019). Causal forests.
  • Belloni, A., et al. (2014). Post-selection inference in high-dimensional models.