Triage Score: A Counterfactual Risk Assessment Instrument

Technical Summary: Triage Score: A Counterfactual Risk Assessment Instrument

Problem Statement
Risk assessment instruments (risk scores) are widely deployed in high-stakes domains such as criminal justice and medicine to predict the likelihood of an undesired outcome (e.g., rearrest or illness) under a baseline decision of no intervention. While high scores typically trigger interventions (e.g., cash bail or intensive care), this approach suffers from a fundamental misalignment with welfare maximization. Risk scores predict outcomes only under a single baseline (no intervention), failing to account for counterfactual outcomes under alternative decisions. Consequently, they cannot distinguish between individuals for whom an intervention would meaningfully alter outcomes and those for whom it would not. This limitation obscures ethical and practical tradeoffs, such as the regret associated with unnecessarily detaining an individual who would not have reoffended had they been released. Existing frameworks often rely on "selective labels" (outcomes observed only for those who received a specific decision), complicating the evaluation of alternative policies.

Methodology
The authors propose a framework based on triage scores, a class of counterfactual risk assessment instruments designed to align decision-making with the objective of maximizing expected utility. Unlike standard risk scores that depend solely on the baseline potential outcome $Y(d_0)$, triage scores are constructed from additive counterfactual utilities that depend on the joint distribution of potential outcomes under alternative decisions.

1. Counterfactual Utility Framework:

   • The framework utilizes principal stratification (Frangakis and Rubin, 2002), defining strata based on the joint potential outcomes $(Y(0), Y(1), \dots, Y(K_D-1))$. In a binary setting, these strata include: Safe (no outcome under any decision), Backlash (outcome only under intervention), Preventable (outcome only under no intervention), and Hopeless (outcome under any decision).
   • Utilities are defined additively: $u(d, y) = u^d_{y_d} + \sum_{d' \neq d} \tilde{u}^{d'}_{y_{d'}}$. The first term represents the standard utility of the realized outcome, while the second term captures the counterfactual utility (or "regret") associated with the outcome that would have occurred under an alternative decision $d'$.
   • This additive structure allows the framework to distinguish between cases where an intervention is genuinely preventive versus cases where it is unnecessary (e.g., imposing cash bail on a "Safe" defendant).

2. Identification and Estimation:

   • Under the assumption of unconfoundedness (conditional independence of potential outcomes and decisions given observed covariates), the authors establish that the expected counterfactual utility is point-identified, even without observing the full joint distribution of potential outcomes.
   • The paper employs Augmented Inverse Probability Weighting (AIPW) estimators to estimate expected utilities. This approach is doubly robust, requiring consistent estimation of both the decision model (propensity of the judge's decision) and the outcome model.
   • To handle high-dimensional, unstructured text data (specifically probable-cause affidavits) which confound judicial decisions, the authors integrate GenAI-powered inference (GPI). Using large language models (e.g., Llama3) and architectures like DragonNet, they extract low-dimensional deconfounders from the text to adjust for unobserved heterogeneity, satisfying the unconfoundedness assumption more plausibly than with structured covariates alone.

3. Policy Learning:

   • The framework enables the learning of optimal decision rules by solving an empirical utility maximization problem. For binary decisions, this is reformulated as a weighted classification problem; for multiclass decisions, it involves selecting the decision with the highest estimated expected utility for each covariate profile.

Empirical Application and Results
The methodology is applied to a randomized controlled trial (RCT) in Utah evaluating the Public Safety Assessment (PSA), a pretrial risk score. In this study, judges were randomly assigned to receive or not receive PSA recommendations. The dataset includes 9,855 first-arrest cases with outcomes measured as New Criminal Activity (NCA), New Violent Criminal Activity (NVCA), and Failure to Appear (FTA).

• Utility Specification: The authors define utility parameters incorporating the cost of decisions (e.g., cost of cash bail), the cost of undesirable outcomes, and regret parameters for counterfactual outcomes.
• Comparative Analysis: The study compares three systems: Human-alone, Human-with-PSA, and an Optimal Decision Tree policy derived from the triage score framework.
• Key Findings:
  • Sensitivity to Utility Parameters: The preferred decision rule (Human vs. Human+PSA) is highly sensitive to the specification of utility parameters, particularly the cost of cash bail and the regret associated with unnecessary detention.
  • Divergence from Risk Scores: Under the standard risk score framework (which collapses strata based on baseline risk), the optimal policy often differs significantly from that derived under the triage score framework. The triage score framework identifies that cash bail may be counterproductive for "Safe" defendants (those who would not reoffend regardless), a nuance lost in standard risk scoring.
  • Policy Implications: The analysis demonstrates that conclusions regarding the value of algorithmic recommendations are not absolute but depend on the ethical weight placed on counterfactual outcomes. For instance, when regret for unnecessary detention is high, the triage framework often favors policies that reduce cash bail usage compared to standard risk-based approaches.

Contributions and Significance
The paper makes three primary contributions:

1. Theoretical Generalization: It introduces the triage score as a generalization of standard risk scores, shifting the focus from predicting a single baseline outcome to evaluating decisions based on the joint distribution of potential outcomes. This allows for the explicit incorporation of ethical considerations like regret.
2. Methodological Advancement: It provides a statistical decision-theoretic framework for evaluating and learning optimal policies under counterfactual utilities, establishing identification results under unconfoundedness and developing semiparametric estimators (AIPW) that accommodate high-dimensional text confounders via GenAI.
3. Empirical Demonstration: Through the Utah RCT, the paper illustrates that conventional risk score evaluations may yield substantively distinct and potentially suboptimal policy recommendations compared to triage-based approaches. It highlights that the "best" decision rule is contingent on the specific utility structure (ethical and practical tradeoffs) chosen by the decision-maker.

Claims and Limitations
The authors position the triage score not as a specific tool for a single setting but as a general decision-support framework. They explicitly state that they do not develop a specific triage score for the Utah application but rather demonstrate the framework's ability to capture rich utility structures. The paper acknowledges that the validity of the results relies on the unconfoundedness assumption, which is rendered plausible in their application by the availability of full judicial information (including PC affidavits) and the use of GPI to adjust for text-based confounders. The authors suggest future work should focus on the real-world elicitation of utility parameters with policymakers and the extension of the framework to dynamic, sequential decision-making settings.