Treatment Choice with Nonlinear Regret

Technical Summary: Treatment Choice with Nonlinear Regret

Problem and Motivation
The paper addresses the problem of statistical treatment choice, specifically how to use data from randomized controlled trials to inform welfare-optimal policy decisions. The existing literature, initiated by Manski (2004), predominantly focuses on minimizing the mean welfare regret (expected regret). Under this criterion, optimal decision rules are typically "singleton rules," which dictate treating either the entire population or no one based on the data. The authors argue that this approach is sensitive to sampling uncertainty. For instance, in small samples, a minor fluctuation in data can cause the optimal rule to switch abruptly from treating everyone to treating no one, leading to high variance in welfare and potential large welfare losses in bad states of the world. The mean regret paradigm fails to capture decision-makers' aversion to such volatility or large losses.

Methodology
To address these limitations, the authors propose a novel framework grounded in statistical decision theory that minimizes the mean of a nonlinear transformation of welfare regret. Let $g(\cdot)$ be a nonlinear, strictly increasing function with $g'(0)=0$. The performance of a treatment rule $\hat{\delta}$ is evaluated via the "nonlinear regret risk":
$$R_g(\hat{\delta}, P) := E_{P^n}[g(\text{Reg}(\hat{\delta}))]$$
where $\text{Reg}(\hat{\delta})$ is the welfare regret relative to the infeasible optimal policy. The paper focuses particularly on the case where $g(r) = r^2$, referred to as mean square regret. This criterion penalizes rules not only for high mean regret but also for high variance in regret (which is equivalent to high variance in welfare).

The analysis proceeds in two stages:

1. Finite Sample Analysis: The authors derive Bayes and minimax optimal rules for finite samples, specifically within a Gaussian experiment with known variance.
2. Asymptotic Analysis: Using the limit experiment framework introduced by Le Cam (1986) and Hirano and Porter (2009), the authors extend the results to large samples for regular parametric models.

Key Contributions and Results

1. Incompleteness and Inadmissibility of Singleton Rules:
   The paper proves that singleton rules are not essentially complete and are often inadmissible under nonlinear regret criteria.

   • Theorem 3.1: Shows that for a wide class of distributions and nonlinear transformations satisfying mild conditions, the class of singleton rules is not essentially complete. This implies that fractional rules (which assign a proportion of the population to treatment) cannot be eliminated from consideration.
   • Theorem 3.2: Demonstrates that for homogeneous nonlinear transformations (degree $\alpha > 1$) and data from a one-parameter exponential family, singleton threshold rules are inadmissible. This contrasts sharply with results for mean regret, where singleton rules are admissible.

2. Finite Sample Optimal Rules (Mean Square Regret):

   • Bayes Optimality: The Bayes optimal rule is generally fractional. It is characterized as a "tilted" posterior probability matching rule. Specifically, the treatment fraction is the posterior probability that the treatment effect is positive, weighted by the square of the treatment effect. If the prior is supported on two symmetric points, this reduces to the standard Thompson sampling (posterior probability matching) rule.
   • Minimax Optimality: The minimax optimal rule is derived by identifying a least favorable prior. The authors show that the least favorable prior is supported on two symmetric points $\{\tau^*, -\tau^*\}$ with $\tau^* \approx 1.23$. The resulting minimax rule is a logistic transformation of the sample mean (or t-statistic):
     $$\hat{\delta}^*(\bar{Y}_1) = \frac{\exp(2\tau^* \bar{Y}_1)}{\exp(2\tau^* \bar{Y}_1) + 1}$$
     This rule is tuning-parameter-free and easy to compute.

3. Asymptotic Optimality:
   In the limit experiment, the authors derive feasible and asymptotically optimal rules.

   • Minimax Rule: The asymptotically minimax optimal rule retains the logistic structure, applied to the t-statistic of the average treatment effect.
   • Bayes Rule: The asymptotically Bayes optimal rule is a tilted posterior probability matching rule. Unlike the linear regret case where Bayes and minimax rules converge to the same Empirical Success (ES) rule, under mean square regret, the Bayes and minimax rules remain distinct, though their performance in terms of mean square regret is quantitatively similar.

4. Applications:

   • Normal Regression Model: The framework is applied to a normal regression model with binary treatment. The authors provide closed-form expressions for the finite-sample minimax rule and the asymptotically minimax feasible rule, which practitioners can implement using OLS estimates.
   • Sample Size Calculations: The paper compares the efficiency of singleton rules (ES and Hypothesis Testing rules) against the proposed minimax rule. It shows that singleton rules are significantly less efficient in terms of mean square regret. For example, to achieve the same worst-case mean square regret, the ES rule requires nearly 40% more observations, and the Hypothesis Testing rule requires approximately 11 times more observations than the proposed minimax rule.

Significance and Interpretation
The authors claim several key significances for their approach:

• Justification for Fractional Rules: The paper provides a decision-theoretic justification for implementing fractional treatment assignment rules in standard settings without model ambiguity or nonlinear welfare functions.
• Regret Aversion: The approach aligns with the concept of "regret aversion" from decision theory (Hayashi, 2008), accommodating decision-makers' concern for large welfare losses in bad states.
• Summary Statistic: Because the optimal rules are fractional (values between 0 and 1), they can be interpreted as a summary statistic quantifying the strength of evidence in favor of the treatment. The authors suggest this metric could serve as an alternative to the p-value, offering a measure of confidence that is grounded in an optimality criterion (minimax mean square regret) rather than just hypothesis testing.
• Tractability: Despite the theoretical shift, the resulting optimal rules are simple, closed-form, and easy for practitioners to implement.