From Approval to State Entry: Temporal Machine Learning and Observational Causal Analysis of Generic Drug Adoption in U.S. Medicaid Markets

Regulatory approval does not guarantee that a generic drug will immediately or uniformly appear across state Medicaid markets. This study develops a reproducible longitudinal framework for predicting first generic-drug entry into U.S. state Medicaid markets and evaluating the observational relationship between prior peer-state adoption exposure and next-quarter entry. FDA generic approval information was linked with Medicaid State Drug Utilization Data to construct a drug–state–quarter panel containing 261 eligible ANDA-level launches and 174,471 observations. Predictive models were evaluated using a temporal grouped design that separated launches into 147 training, 66 validation, and 48 locked-test launches. A validation-selected FastTree model achieved a locked-test ROC AUC of 0.8221 and PR AUC of 0.1112 for predicting first entry in the following quarter. The causal analysis estimated the average treatment effect on the treated of high prior-quarter peer-state exposure using augmented inverse probability weighting under stated measured-confounder assumptions. The governed .NET estimate was 0.00157 on the risk-difference scale, with a 95% confidence interval from −0.00377 to 0.00928. An independently fitted grouped cross-fitted Python analysis produced an estimate of −0.01382, indicating that the causal conclusion was not robust across specifications. The findings support the usefulness of temporal machine learning for state-entry prediction while providing no definitive evidence that peer-state exposure causes increased next-quarter entry. The study contributes an auditable SQL Server, ML.NET, C#, and Python framework for longitudinal pharmaceutical-access research.