Technical Summary: MiniMax Learning of Interpretable Factored Stochastic Policies from Conjoint Data, with Uncertainty Quantification

Problem Statement

Conjoint analysis has become the standard methodology for studying multidimensional preferences, typically estimating the Average Marginal Component Effect (AMCE). The AMCE quantifies the average causal effect of a single attribute by marginalizing over other attributes, often assuming a uniform distribution of opponent profiles. However, this approach has two critical limitations for strategic decision-making:

1. Strategic Interdependence: Real-world scenarios (e.g., elections, hiring) involve strategic opponents who do not present uniform profiles. AMCEs ignore the interdependence between a candidate's profile and their opponent's.
2. Interaction Neglect: AMCEs focus on marginal effects, potentially missing critical interaction effects between attributes that determine optimal strategies.

The paper addresses the challenge of offline policy optimization over exponentially large factorial action spaces using randomized preference data. The goal is to move from estimating marginal effects to learning interpretable stochastic policies that maximize expected outcomes, accounting for both attribute interactions and strategic adversarial dynamics.

Methodology

1. Framework and Identification

The authors frame the problem as learning a policy $\pi$ (a distribution over full profiles) to maximize expected win probability in a forced-choice setting.

• Data: Randomized conjoint experiments where respondents choose between two profiles, $T^a$ and $T^b$.
• Outcome Model: A Bernoulli Generalized Linear Model (GLM) or a Bayesian Neural Network models the choice probability $P(C_i=1 | T^a, T^b) = \sigma(\eta_i)$, where $\eta_i$ includes main effects and two-way interactions.
• Policy Class: To ensure interpretability and tractability in high-dimensional spaces, the policy is restricted to a product-of-Categoricals distribution. This means the probability of a full profile is the product of independent probabilities for each attribute level: $Pr_\pi(t) = \prod_d \pi_d(t_d)$.

2. Average-Case Optimization

In the non-adversarial setting, one party optimizes their profile distribution against a fixed reference distribution (e.g., uniform) of the opponent.

• Objective: Maximize expected outcome $Q(\pi)$ subject to an $L_2$ (or KL) regularization term that shrinks the learned policy toward the experimental assignment distribution $p$. This controls off-policy variance.
• Closed-Form Solution: Under a linear probability approximation with two-way interactions, the authors derive a closed-form optimizer (Proposition 3.1). The optimal policy $\pi^*$ is the solution to a linear system $C\pi^* = B$, where the matrix $C$ and vector $B$ depend on the estimated interaction coefficients and the regularization strength.
• Uncertainty Quantification: Using the Delta method, the uncertainty from the outcome model parameters (estimated via GLM) is propagated to the optimal policy $\pi^*$ and its value $Q(\pi^*)$. This provides asymptotic standard errors for both the policy probabilities and the expected win rate.

3. Adversarial Minimax Extension

To capture strategic competition, the framework extends to a zero-sum game where two agents (e.g., political parties) simultaneously choose mixed profile distributions to maximize their own win probability.

• Objective: Solve the minimax problem $\max_{\pi^A} \min_{\pi^B} Q(\pi^A, \pi^B)$.
• Institutional Constraints: The model incorporates institutional details, such as a two-stage process (primaries followed by a general election). The payoff function $Q_{inst}$ integrates over primary electorates and the general electorate, allowing for heterogeneous voter participation.
• Optimization: Since the factored policy class is generally non-convex, the authors employ a gradient-based ascent-descent algorithm (Algorithm 1). They utilize implicit differentiation or unrolling to compute gradients through the optimization trace, enabling end-to-end uncertainty propagation for the equilibrium strategies.
• Strategic Divergence: A data-driven metric is introduced to measure the divergence between parties' optimal strategies, quantifying how distinct their preferred candidate profiles are.

Key Contributions

1. Shift from AMCE to Stochastic Policy: The paper proposes a new estimand: a factored stochastic policy over full profiles. Unlike AMCEs, this policy is optimized for expected performance and remains interpretable at the attribute level.
2. Closed-Form Average-Case Optimizer: For linear outcome models with two-way interactions, the authors provide an analytical solution for the regularized optimal policy, along with a formal Delta-method framework for uncertainty quantification of both the policy and its value.
3. General Differentiable Optimization: For richer models (GLMs, Neural Networks) and higher-order interactions, the paper provides a general gradient-based procedure. It leverages implicit differentiation to compute standard errors for the optimal policy without requiring manual derivation of complex gradients.
4. Adversarial Minimax with Institutional Structure: The framework extends to a two-player minimax setting that explicitly models institutional constraints (e.g., primaries). This yields "restricted-equilibrium" mixed strategies that reflect realistic strategic constraints.
5. Empirical Validation:
   • Simulations: Monte Carlo simulations demonstrate that the method achieves low bias and rapidly declining RMSE as sample size increases, with coverage rates approaching nominal levels. The adversarial method outperforms non-adversarial (averaging) baselines, particularly in high-dimensional settings with strong interactions.
   • Real-World Application: Applied to U.S. presidential conjoint data (Ono & Burden, 2019), the adversarial policies produce restricted-equilibrium vote shares that align with historical election ranges (1976–2016). In contrast, average-case optimizers produce vote shares outside historical norms. The method successfully recovers strategic divergence between parties and maps historical candidates to the conjoint feature space.

Significance and Claims

The paper claims to bridge the gap between causal inference in conjoint analysis and strategic decision support. By moving from marginal effect estimation to policy learning, the authors enable researchers to answer "what is the optimal strategy?" rather than just "what is the effect of this attribute?".

The significance lies in:

• Interpretability: Maintaining the ability to read off attribute weights (e.g., "economy" vs. "immigration") even in complex, high-dimensional strategic settings.
• Robustness: Providing rigorous uncertainty quantification for both the learned policies and their predicted values, addressing the "black box" nature of many machine learning approaches in social science.
• Strategic Realism: Capturing the reality that candidates do not optimize in isolation but in response to opponents and institutional rules (primaries), leading to more accurate predictions of equilibrium outcomes.

The authors emphasize that while the factorized policy class is an approximation of the full joint distribution, it offers a tractable and interpretable middle ground. They acknowledge limitations, such as the reliance on two-step estimators and the need for accurate institutional modeling, but argue that the resulting "restricted-equilibrium" strategies provide a valuable, data-driven lens for understanding strategic polarization and candidate positioning.