Scholarlink

Background <p>Dynamic treatment regimens (DTRs) guide personalised sequential treatment decisions for patients with a range of clinical or behavioural diseases. Sequential multiple assignment randomised trials (SMARTs) are designed to evaluate and optimise DTRs by randomising participants at multiple stages based on intermediate outcomes. To identify optimal DTRs in SMARTs, the mean outcome of each DTR is often estimated via inverse probability weighting (IPW), a statistical method that uses the inverse probability of treatment to address potential bias in the design. Like other randomised controlled trials, SMARTs are subject to missing data. Handling missing data in SMARTs is complicated by the sequential randomisation and dependence on intermediate outcomes. We evaluated the performance of complete case analysis (CCA) and multiple imputation (MI) for handling missing data when estimating the DTR mean outcomes using IPW in a two-stage SMART.</p> Methods <p>We simulated 1000 datasets of 400 participants, based on a prototypical SMART design with two stages where only non-responders are re-randomised at stage 2. The estimands of interest were the four DTR means of a continuous outcome and were estimated using IPW. We defined four plausible missing data scenarios using missing data directed acyclic graphs (m-DAGs) and then assessed how each missing data method (CCA and MI) performed under different proportions of missingness (20%, 40%) and strengths of associations with missingness in stage 1 intermediate outcome, stage 2 treatment, and the final outcome.</p> Results <p>Minimal bias was observed with MI when estimating the mean outcomes of the DTRs in most scenarios, except for when stage 1 intermediate outcome was missing dependent on baseline variables and stage 1 treatment. When data were missing dependent on other variables (for example, stage 2 treatment missing dependent on stage 1 intermediate outcome), CCA generally showed greater bias than MI when estimating the mean outcomes of the DTRs. Empirical standard errors were comparable across both missing data methods, with MI generally producing slightly lower values.</p> Conclusion <p>We found that for a prototypical SMART design, MI generally showed close to zero bias and slightly lower standard errors compared to CCA when IPW was used to estimate the mean outcomes of DTRs in the settings explored.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Inverse probability weighted estimation of dynamic treatment regimen means in sequential multiple assignment randomised trials with missing data: a simulation study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月30日

查看全文
  • Jessica Xu,
  • Robert K. Mahar,
  • Katherine J. Lee,
  • Anurika P. De Silva,
  • Julie A. Simpson

摘要

Background

Dynamic treatment regimens (DTRs) guide personalised sequential treatment decisions for patients with a range of clinical or behavioural diseases. Sequential multiple assignment randomised trials (SMARTs) are designed to evaluate and optimise DTRs by randomising participants at multiple stages based on intermediate outcomes. To identify optimal DTRs in SMARTs, the mean outcome of each DTR is often estimated via inverse probability weighting (IPW), a statistical method that uses the inverse probability of treatment to address potential bias in the design. Like other randomised controlled trials, SMARTs are subject to missing data. Handling missing data in SMARTs is complicated by the sequential randomisation and dependence on intermediate outcomes. We evaluated the performance of complete case analysis (CCA) and multiple imputation (MI) for handling missing data when estimating the DTR mean outcomes using IPW in a two-stage SMART.

Methods

We simulated 1000 datasets of 400 participants, based on a prototypical SMART design with two stages where only non-responders are re-randomised at stage 2. The estimands of interest were the four DTR means of a continuous outcome and were estimated using IPW. We defined four plausible missing data scenarios using missing data directed acyclic graphs (m-DAGs) and then assessed how each missing data method (CCA and MI) performed under different proportions of missingness (20%, 40%) and strengths of associations with missingness in stage 1 intermediate outcome, stage 2 treatment, and the final outcome.

Results

Minimal bias was observed with MI when estimating the mean outcomes of the DTRs in most scenarios, except for when stage 1 intermediate outcome was missing dependent on baseline variables and stage 1 treatment. When data were missing dependent on other variables (for example, stage 2 treatment missing dependent on stage 1 intermediate outcome), CCA generally showed greater bias than MI when estimating the mean outcomes of the DTRs. Empirical standard errors were comparable across both missing data methods, with MI generally producing slightly lower values.

Conclusion

We found that for a prototypical SMART design, MI generally showed close to zero bias and slightly lower standard errors compared to CCA when IPW was used to estimate the mean outcomes of DTRs in the settings explored.