Reference-based multiple imputation for longitudinal bounded discrete outcomes
Abstract
Reference-based multiple imputation has been developed for binary and ordinal responses. We develop a parsimonious formulation for longitudinal bounded scores and counts that preserves their finite integer support and gives partial and complete reference borrowing an explicit distributional interpretation. Discrete Kumaraswamy margins provide closed-form cumulative distribution and quantile functions, simplifying fitting and imputation under a Gaussian copula. Borrowing is defined through a regression model indexed by a smooth median anchor, allowing continuation under the assigned treatment, partial borrowing, or jump to reference after a specified trigger. With shapes shared across arms, complete borrowing replaces the active marginal distribution with the reference distribution at the same visit and covariate values, while conditional imputations remain informed by retained outcomes. Bayesian estimation uses latent variable augmentation, and all sensitivity rules share one fitted posterior. Completed datasets are analyzed using a beta-binomial random intercept model, with treatment comparisons expressed as standardized mean score differences. Simulations assess parameter recovery and imputation followed by analysis. An application to depression scores illustrates attenuation of the treatment benefit at the final visit as borrowing increases, with confidence intervals including zero under all rules.
Related
- Comparison of parametric versus machine-learning multiple imputation in clinical trials with missing continuous outcomes.
- Reference-based sensitivity analysis for bounded longitudinal outcomes
- Bayesian multiple imputation for functional data
- One-step Outcome Imputation: An Alternative to Multiple Imputation
- Robust variance estimation for multiple imputation using predictive mean matching or parametric imputation models
BibTeX
@misc{burger2026referencebased,
title = {{Reference-based multiple imputation for longitudinal bounded discrete outcomes}},
author = {Divan Aristo Burger},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.39834},
eprint = {2610.39834},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.39834},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Reference-based multiple imputation for longitudinal bounded discrete outcomes
AU - Divan Aristo Burger
PY - 2026
DA - 2026/10/29
PB - arXiv
DO - 10.48550/arxiv.2610.39834
UR - https://arxiv.org/abs/2610.39834
AB - Reference-based multiple imputation has been developed for binary and ordinal responses. We develop a parsimonious formulation for longitudinal bounded scores and counts that preserves their finite integer support and gives partial and complete reference borrowing an explicit distributional interpretation. Discrete Kumaraswamy margins provide closed-form cumulative distribution and quantile functions, simplifying fitting and imputation under a Gaussian copula. Borrowing is defined through a regression model indexed by a smooth median anchor, allowing continuation under the assigned treatment, partial borrowing, or jump to reference after a specified trigger. With shapes shared across arms, complete borrowing replaces the active marginal distribution with the reference distribution at the same visit and covariate values, while conditional imputations remain informed by retained outcomes. Bayesian estimation uses latent variable augmentation, and all sensitivity rules share one fitted posterior. Completed datasets are analyzed using a beta-binomial random intercept model, with treatment comparisons expressed as standardized mean score differences. Simulations assess parameter recovery and imputation followed by analysis. An application to depression scores illustrates attenuation of the treatment benefit at the final visit as borrowing increases, with confidence intervals including zero under all rules.
ER -