Presentation (CTI)

OBJECTIVES: The PROMIS-29+2 Profile (PROPr) is a multi-domain HRQoL instrument that offers greater precision and covers more health domains than competing alternatives. However, its relatively high item burden may increase item nonresponse, leading to missing HRQoL estimates in study data. Standard missing data methods do not leverage cross-domain dependence to impute item-level nonresponses. Copula modeling provides a flexible framework for characterizing joint dependence among PROPr domains, allowing observed domains to inform estimation of missing domains. Analyzing domain dependencies is particularly important for comparative effectiveness research in criminal-legal-system-involved populations with opioid use disorder (OUD), where enrollment and follow-up are limited. Thus, modest rates of item nonresponse complicate inference.
METHODS: Using pooled RCT data from the Justice Community Opioid Innovation Network (n=1,250), we estimated pairwise dependence among HRQoL domains and report parametric estimates that can be used to impute missing domains. Gaussian and t-copulas were fit with unstructured dependence, allowing distinct parameters for each domain pair, while Frank copulas were fit pairwise due to dimensional limitations. Pairwise dependence was summarized using Kendall’s τ with bootstrapped 95% confidence intervals, and model fit was compared using AIC.
RESULTS: Kendall's tau estimates were broadly consistent across copula families, with stronger dependence observed among conceptually related domains and weaker dependence between pain and physical function. The multivariate t-copula provided better fit than the Gaussian copula, suggesting tail dependence. Across all pairwise comparisons, t-copulas consistently achieved lower AIC than Frank copulas. Overall, Gaussian and t-copulas captured the dominant dependence structure across domains, whereas Frank copulas performed less well in this multi-domain setting.
CONCLUSIONS: Copula models provide a practical approach for quantifying inter-domain dependence in HRQoL measures and offer a flexible, semi-parametric alternative to regression-based imputation when domain-level data are partially missing. Borrowing information across correlated domains may support more robust estimation under item nonresponse, improving downstream comparative effectiveness and decision-analytic modeling.