Wen-Wei Liu Oral Dissertation Defense

Abstract: Meta-regression is used to assess whether participant or study characteristics explain variation in treatment effects across studies. Aggregate-data (AD) meta-regression is often accessible, but study-level summaries may not adequately represent participant-level covariate information. Individual participant data (IPD) meta-analysis can evaluate treatment effect modification more directly, although IPD are often difficult to obtain. This dissertation compares IPD and AD methods for estimating across-study treatment-covariate interactions and develops an approach to improve inference when only AD are available.

The first study compared one-stage IPD meta-analysis with traditional AD meta-regression for continuous outcomes. Across simulation scenarios, the IPD model generally showed lower bias and root mean squared error, confidence interval coverage closer to the nominal level, and greater power, especially when continuous covariates had substantial within-study variation.

The second study extended the comparison to binary outcomes. For continuous effect modifiers with moderate or high within-study dispersion, IPD generally provided more reliable inference, whereas differences were smaller when within-study variation was limited. For binary effect modifiers, the two approaches performed similarly overall, although IPD was modestly more powerful in several settings. The results also highlighted finite-sample logistic bias, sparse data, separation, and non-collapsibility of the odds ratio.

The third study developed a Bayesian hierarchical meta-regression model that treated each observed study-level covariate as an uncertain estimate of an underlying true covariate. With informative priors, the model reduced bias and improved interval coverage for continuous covariates. Improvements for binary covariates were greatest when covariate proportions varied sufficiently across studies. An application to pediatric asthma adherence interventions illustrated the importance of propagating covariate uncertainty and assessing prior sensitivity.

Overall, one-stage IPD meta-analysis provided the most reliable inference when participant-level data were available. When only AD are available, modeling uncertainty in study-level covariates can improve inference, although information lost through aggregation cannot be fully recovered.