Why this interlude matters

Section 3.12 contrasts design-first causal inference with structural industrial organization (IO). It explains what each tradition targets, the assumptions that do the identifying work, and how to build a pragmatic hybrid workflow.

Two traditions, different targets

Design-first methods estimate credible causal effects of concrete interventions using transparent assumptions and diagnostics such as pre-trends, placebo tests, overlap checks, and design-faithful inference. The goal is a defensible local causal effect.

Structural IO targets behavioral primitives and market mechanisms. It starts from a model of consumer demand and firm conduct, then uses equilibrium conditions to recover objects like elasticities, markups, and marginal costs. A canonical example is the Berry-Levinsohn-Pakes (BLP) demand model, which recovers demand elasticities and supports counterfactual simulations of mergers, taxes, assortment changes, and pricing policies.

Key building blocks in structural IO:

  • Demand system with utility, heterogeneity, and outside options.
  • Supply side with pricing and conduct assumptions (price competition, collusion, or conduct parameters).
  • Equilibrium conditions that map primitives to observed prices and shares.
  • Instruments or exclusion restrictions that separate demand shocks from price variation.
  • Estimation framework (GMM or MLE) that fits the model to moments in the data.

Strengths and vulnerabilities

Design-first excels in transparency and internal validity. Diagnostics make assumptions visible, and estimates are decision-ready for the specific design and time window. The limitation is scope: effects are local and do not necessarily generalize to new policies or strategic environments.

Structural models excel at mechanism and counterfactual reach. They can answer “what if” questions far beyond observed support. The tradeoff is reliance on functional-form assumptions, exclusion restrictions, and equilibrium selection. Identification often hinges on instrument quality and the extent of variation they generate in prices or product attributes. Computational burdens can crowd out robustness checks, and mis-specified structure can lead to misleading counterfactuals.

A productive synthesis

The recommended workflow is sequential and pragmatic:

  1. Use design-first tools (DiD, event studies, synthetic control, SDID, factor models) to estimate credible local effects with full diagnostics.
  2. Use those reduced-form estimates to discipline structural models by calibrating elasticities, substitution patterns, or dynamics.
  3. Treat structural counterfactuals as sensitivity analysis when they cannot match credible reduced-form moments without implausible parameters.

This aligns evidence with decisions: design-first establishes what happened under a feasible design, structure explores why and what might happen under new policies.

BLP in one paragraph

BLP specifies indirect utility as $u_{njt} = x'_{jt}\beta_n - \alpha p_{jt} + \xi_{jt} + \varepsilon_{njt}$, where $x_{jt}$ are observed product characteristics, $p_{jt}$ is price, and $\xi_{jt}$ is an unobserved shock. Random coefficients capture taste heterogeneity. Market shares invert to mean utilities, and instruments address price endogeneity. With a supply side, the model delivers markups and conduct and enables counterfactual equilibrium simulations. Credibility hinges on the exclusion restriction and the support of the identified price variation.

Practical considerations for structural IO:

  • Data demands are heavy: product-level prices, shares, and rich characteristics.
  • Instruments must shift prices without directly shifting demand; weak instruments undermine identification.
  • Model fit should be checked against reduced-form evidence (elasticities, pass-through, substitution patterns).
  • Counterfactuals should report sensitivity to alternative conduct assumptions and instrument sets.

When to prefer which

Use design-first when the goal is to measure realized effects of specific rollouts, when diagnostics are paramount, or when interference and nonstationarity complicate modeling. Use structural approaches when the decision requires market-wide simulations, welfare analysis, or strategic counterfactuals far from observed support. Combine them when decisions require both credible local evidence and mechanism-based exploration.

Reporting and diagnostics across traditions

Structural work benefits from design-style diagnostics: pre-trend checks for reduced forms, placebo tests, instrument strength reporting, and sensitivity to equilibrium assumptions. Design-first work benefits from structural sanity checks: elasticities and substitution patterns should be economically plausible and consistent with prior evidence.

Takeaway

Design-first and structural IO are complements, not substitutes. Use design-first for credibility and structure for reach. The best practice is to link them: let design-based evidence discipline structural counterfactuals.

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 3.12.
  • Berry, S., Levinsohn, J., and Pakes, A. (1995). Automobile prices in market equilibrium.
  • Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry.
  • Berry, S., Levinsohn, J., and Pakes, A. (2014). Differentiated products demand systems.
  • Train, K. (2009). Discrete Choice Methods with Simulation.
  • Einav, L., and Levin, J. (2010). Empirical industrial organization.
  • Aguirregabiria, V., and Mira, P. (2010). Dynamic discrete choice structural models.