Identification Design

1. Problem Statement

The paper addresses a fundamental gap in the literature on strategic communication and causal inference. Standard models of strategic communication (e.g., Bayesian Persuasion) assume a decision maker (DM) with an exogenous prior who updates beliefs upon observing a specific signal realization. However, in empirical research (particularly microeconometrics and regulatory evaluations), researchers often disclose aggregate data distributions (e.g., joint distributions of outcomes, treatments, and covariates) rather than individual-level data.

The core problem is: How should an objective, prior-free decision maker respond to an information structure (a researcher's data disclosure) when they face ambiguity about the true state of the world? The DM does not know the true distribution of states ($\mu$) but observes the distribution of signals generated by the researcher. The DM must choose an action (e.g., a treatment policy) that maximizes their worst-case payoff over the set of all state distributions consistent with the observed data and structural assumptions.

The paper asks:

1. Under what conditions is an environment manipulable? (i.e., can a researcher always design an information structure to induce the DM to take any desired action, regardless of the true state?)
2. How do specific disclosure policies (e.g., revealing only marginal distributions of covariates) affect the ability to manipulate the DM?
3. What are the sharp bounds on payoffs when exact identification is impossible?

2. Methodology and Model Framework

The Environment

• States ($\Omega$): A finite set of states representing the true distribution of latent variables (e.g., potential outcomes, treatment assignment).
• Actions ($A$): A compact metric space of actions (e.g., choosing a treatment $a \in \{0, 1\}$).
• Admissible Priors ($P$): A set of probability distributions over $\Omega$ consistent with structural assumptions (e.g., unconfoundedness, strict overlap). The true distribution $\mu \in P$.
• Information Structure ($\Sigma, E$): A mapping where $E: \Omega \to \Delta(\Sigma)$ assigns a distribution of messages (signals) to each state. The DM observes the population distribution of signals $E_\mu$, not individual realizations.
• Identified Set: Given an observed signal distribution $E_\mu$, the set of plausible state distributions is $P_\mu(\Sigma, E) = \{ \nu \in P \mid E_\nu = E_\mu \}$.
• Decision Rule: The DM chooses action $\alpha$ to maximize the worst-case expected utility:
  $$\max_{\alpha} \inf_{\nu \in P_\mu(\Sigma, E)} U(\alpha, \nu)$$
  An information structure implements action $\alpha$ if $\alpha$ is the unique solution to this maximin problem.

Key Concepts

• Manipulability: An environment is manipulable if for every action $\alpha$ and every true state $\mu$, there exists an information structure that implements $\alpha$.
• Marginal Information Structures: A restriction reflecting real-world practice where the researcher discloses the joint distribution of a subset of variables (Outcome $Y$, Treatment $T$, and a subset of Covariates $Z$), while omitting others ($W$).

3. Key Contributions and Results

A. Characterization of Manipulability

Theorem 1 provides a necessary and sufficient condition for an environment to be manipulable:

• An environment is manipulable if and only if the worst-case payoff over the set of admissible priors is identical for all actions.
  $$\inf_{\nu \in P} U(\alpha, \nu) = \inf_{\nu \in P} U(\beta, \nu) \quad \forall \alpha, \beta \in A$$
• Implication: If actions have different worst-case payoffs, the DM will never choose the action with the strictly lower worst-case payoff, regardless of the information structure.

B. Application to Treatment-Effects Models

The paper applies this framework to causal inference models with binary treatments, outcomes, and covariates.

• Proposition 2: All treatment-effects models are manipulable.
  • Reasoning: In these models, the set of admissible payoff vectors forms a hypercube $[\min Y, \max Y]^A$. Consequently, the worst-case payoff for any action is the global minimum outcome ($\min Y$). Since all actions share the same worst-case payoff, the DM is indifferent in the worst case, allowing the researcher to manipulate the DM into choosing any specific action.
• Proposition 3 (Strong Manipulability): Treatment-effects models are strongly manipulable.
  • Manipulation is feasible via almost fully informative structures (disclosing $n-1$ linearly independent statistics in an $n$-state model).
  • These structures are robust to tie-breaking rules and do not "immiserate" the DM (the true payoff under the implemented action is exactly identified, preventing ex-post revelation of manipulation).

C. Marginal Information Structures and Identification

The paper restricts attention to marginal information structures (disclosing $Y, T, Z$ but omitting $W$).

• Theorem 2 (Identification Condition): A marginal information structure identifies the treatment-effects model (making it non-manipulable) if and only if the disclosed covariates $Z$ are rich enough to make the assignment mechanism measurable.
  • Specifically, $P(T|X)$ must depend only on $Z$. If the assignment depends on omitted variables $W$, the model is not fully identified.
• Implication for Experiments: In randomized experiments where assignment is independent of covariates ($P(T|X) = P(T)$), the model is identified regardless of omitted covariates.
• Implication for Observational Studies: These remain partially manipulable. Researchers can select which covariates to disclose to shift the DM's worst-case beliefs.

D. Sharp Payoff Bounds

For observational studies where exact identification fails (due to omitted variables), the paper derives tight bounds on the DM's payoff.

• Theorem 3: Defines lower ($L(a)$) and upper ($H(a)$) bounds for the expected payoff of action $a$.
  • The bounds depend on the observed marginal distribution of $(Y, T, Z)$ and the extreme possible values of the unobserved counterfactuals for the omitted groups.
  • Proposition 4: An action is implementable via a marginal structure if and only if it maximizes the lower bound $L(a)$ among all actions.
• Significance: This quantifies the "scope for manipulation." As the set of disclosed covariates grows, the gap between $L(a)$ and $H(a)$ shrinks, reducing the ability to manipulate the DM.

E. Non-Manipulable Environments

For environments that do not satisfy the manipulability condition (where worst-case payoffs differ), the paper characterizes implementable actions under regularity assumptions (Assumptions 2 & 3).

• Theorem 4: An action $\alpha$ is implementable under true state $\mu$ if and only if there exists an admissible prior $\nu$ such that:
  1. $\alpha$ is optimal under full information with prior $\nu$ ($\alpha \in \alpha^*(\nu)$).
  2. The payoff of $\alpha$ under $\nu$ is no higher than under $\mu$ ($U(\alpha, \nu) \leq U(\alpha, \mu)$).
• This result links implementation to the existence of a "saddle point" in the decision problem, achievable via almost fully informative structures.

4. Significance and Implications

1. Theoretical Contribution: The paper endogenizes the "identified set" in econometrics, treating the information structure as a choice variable. It bridges the gap between Blackwell's theory of experiments, Bayesian Persuasion, and Partial Identification in econometrics.
2. Robustness to Manipulation: It demonstrates that standard treatment-effect models are inherently vulnerable to manipulation because the worst-case scenario for any treatment is the same (the worst possible outcome). This suggests that standard "worst-case" robustness checks in econometrics may be insufficient to prevent strategic disclosure.
3. Policy Recommendations:
   • Experiments: Full identification (and thus immunity to manipulation) is achieved if the disclosed covariates are sufficient to verify the assignment mechanism.
   • Observational Studies: Partial identification is inevitable if covariates are omitted. However, the paper provides sharp bounds to quantify the risk.
   • Disclosure Standards: The results support strict pre-registration and full disclosure of joint distributions (Outcome, Treatment, and all relevant Covariates) to eliminate the "scope for manipulation" via covariate selection.
4. Methodological Shift: It moves the focus from "what is the true effect?" to "what information structure is required to guarantee a specific policy choice?" This is crucial for regulatory bodies and policymakers who must act on potentially biased or selectively disclosed data.

In summary, Rosenthal provides a rigorous game-theoretic framework showing that while causal inference models are theoretically manipulable, the extent of this manipulation is strictly bounded by the richness of the disclosed covariates. The paper offers a mathematical justification for rigorous disclosure protocols in scientific research and regulation.