Criterion validity and divergent risk profiles of long-term opioid therapy across medicare and medicaid

Importance: Accurate identification of patients receiving long-term opioid therapy (LTOT) remains essential for clinical care and population health surveillance. Two widely used methods: prescription-based definitions and diagnostic codes are often treated as interchangeable, yet the criterion validity of code-based identification against a prescription-based reference standard is unknown, and the comparative risk profiles associated with each method across diverse populations are incompletely understood. Objective: To assess the criterion validity of diagnostic codes for identifying long-term opioid therapy (LTOT) using prescription-based LTOT as the reference standard, and to compare how each identification method corresponds to the risk of opioid use disorder, opioid poisoning, and other adverse outcomes across Medicare and Medicaid populations. Design, setting, and participants: Retrospective cohort study using 100% Medicare and Medicaid claims data from 2016–2022. Criterion validity was assessed using complete beneficiary populations with a pain diagnosis (65.5 million Medicaid; 59.6 million Medicare). Risk-profile comparisons used incident LTOT cohorts identified by prescription-based 90-day LTOT, Z79.891 code, or both, excluding those with pre-existing opioid use disorder diagnoses. Main outcomes and measures: Criterion validity metrics included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Cohen’s kappa for Z79.891 against prescription-based LTOT. Within incident LTOT cohorts, primary outcomes within 2 years included documented OUD diagnosis, opioid poisoning, medication for OUD (MOUD) receipt, all-cause mortality, drug mortality, and emergency department visits for pain. Results: In the full Medicaid population (N = 65.5 million), Z79.891 demonstrated sensitivity of 30.5% and specificity of 97.3% against prescription-based LTOT, with PPV of 40.2% and Cohen’s κ = 0.313 (fair agreement). In Medicare (N = 59.6 million), sensitivity was 34.6%, specificity 91.4%, PPV 23.0%, and κ = 0.210 (fair agreement). Approximately two-thirds of prescription-based LTOT patients lacked Z79.891 coding in both populations. Among incident LTOT cohorts (771,581 Medicaid; 4,376,993 Medicare), the two identification methods demonstrated divergent risk profiles across insurance programs. In Medicaid, patients identified through prescription-based LTOT (n = 390,793) had significantly lower odds compared to code-based LTOT patients (n = 267,577) for OUD diagnosis (OR, 0.713; 95% CI, 0.696–0.730) and MOUD receipt (OR, 0.374; 95% CI, 0.364–0.385), but higher odds for all-cause mortality (OR, 2.277; 95% CI, 2.224–2.331) and opioid poisoning (OR, 1.117; 95% CI, 1.063–1.173). In Medicare, this pattern reversed: prescription-based LTOT patients (n = 1,045,466) demonstrated significantly higher odds than code-based LTOT patients (n = 2,845,109) for OUD diagnosis (OR, 3.009; 95% CI, 2.948–3.071), opioid poisoning (OR, 1.890; 95% CI, 1.802–1.981), and pain-related ED visits (OR, 1.204; 95% CI, 1.192–1.215), but lower odds for all-cause mortality (OR, 0.895; 95% CI, 0.886–0.904) and MOUD receipt (OR, 0.886; 95% CI, 0.841–0.932). Across programs, individuals with both types of identified LTOT (113,211 in Medicaid, 486,418 in Medicare) had the highest risk of OUD diagnosis and opioid poisoning. Conclusions and relevance: Diagnostic codes demonstrate low sensitivity (30.5–34.6%) and only fair agreement (κ = 0.210–0.313) with prescription-based LTOT, indicating these methods identify largely different populations. These distinct populations show divergent, and often reversed, risk profiles across Medicare and Medicaid. Reliance on a single LTOT method of identification risks systematic misclassification of opioid-related risk, and population-specific, multi-method approaches to LTOT surveillance are needed.