Abstract
<title>Abstract</title> <p>Attorney involvement in workers’ compensation claims is associated with longer resolution times, increased costs, and greater administrative complexity. Prior studies typically treat legal representation as a binary outcome and do not examine how litigation risk evolves across different levels of predicted probability. This study models attorney involvement as a probabilistic risk process using a CatBoost classifier applied to 48,130 claims. Model performance is evaluated using discrimination and reliability metrics, including ROC-AUC, PR-AUC, Brier score, and Expected Calibration Error, to ensure that predicted probabilities provide reliable estimates of escalation risk. To examine how litigation risk varies across the population of claims, predicted probabilities are used to construct rank-based risk regimes. Claims are grouped into low-, medium-, and high-risk tiers, and feature contributions are analyzed within each tier to identify how drivers of attorney involvement change across the risk spectrum. The analysis then focuses on claims located near the boundaries between regimes and compares them with structurally stable cases, revealing asymmetric patterns among the features. Variables capturing injury severity and jurisdiction strongly influence risk escalation, whereas de-escalating transitions are associated with several contributing predictors, such as injured body part type. By modeling litigation as a probabilistic risk process and examining structural differences across regimes and boundaries, the study provides both practical risk ranking and insight into escalation dynamics in workers’ compensation systems.</p>