Data-driven simulations: a new Quantitative Bias Analysis tool for real-world applications of survival analyses
Michael Abrahamowicz, PhD
Distinguished James 缅北强奸 Professor of Biostatistics
Department of Epidemiology, Biostatistics and Occupational Health | 缅北强奸
WHEN: Wednesday, January 15, 2025, from 3:30 to 4:30 p.m.
WHERE:听Hybrid | 2001 缅北强奸 College Avenue, Room 1140;
NOTE:听Michal Abrahamowicz will be presenting in-person
Abstract
This is a quite 鈥渁pplied鈥 talk that may be of interest for biostatisticians who collaborate on real-world clinical or epidemiological research projects. The goal is to promote use of targeted simulations to explore particular limitations of a specific real-world prognostic or epidemiological study. To this end, our data-driven simulations are designed so as to accurately reflect the salient characteristics of the real-world dataset being analyzed [1]. (In contrast to methods-driven simulations typically reported in statistical publications, which rely on data generated under arbitrary assumptions). After an overview of 7 generic steps proposed to implement our approach, I will illustrate it in two real-world time-to-event (survival) analyses, each dealing with a different data limitation The first illustration concerns assessing the impact of omitting an important prognostic factor on the adjusted Hazard Ratio (HR) for the exposure of interest, based on multivariable Cox proportional hazards (PH) analyses. Here, we show how data-driven simulations permit assessing the joint impact of (i) unmeasured confounding bias and (ii) non-collapsibility, while separating their effects. The second illustration focuses on the pharmaco-epidemiological study of the association between recent use of medication, modeled as a time-varying exposure, and the hazard of a transient cognitive impairment. The event is interval-censored as it can be detected only at discrete times of medical visits. Here we use the permutational algorithm, validated for simulating event times conditional on time-varying exposures and/or effects [2]. Here, our simulation results reveal how the strength of a systematic bias toward the null varies depending on the way event times are imputed, and help decide which of the divergent results of alternative imputation strategies may be closer to the (unknown) truth.
听
[1] Abrahamowicz M, Beauchamp ME, Boulesteix AL, Morris TP, Sauerbrei W, Kaufman JS. Data-driven simulations to assess the impact of study imperfections in time-to-event analyses. American Journal of Epidemiology. 2024 May:kwae058. doi: 10.1093/aje/kwae058 [Online ahead of print].
[2] Sylvestre MP & Abrahamowicz M. Comparison of algorithms to generate event times conditional on time-dependent covariates. Statistics in Medicine 2008; 27: 2618-2634.
Speaker bio
Michal Abrahamowicz is a Distinguished James 缅北强奸 Professor of Biostatistics at 缅北强奸. He develops new, flexible statistical methodology for survival analyses, with focus on time-varying exposures and effects. He also explores, and attempts to correct for, different biases in epidemiological studies and promotes creative applications of statistical simulations. He is a co-founder and the co-chair of the international STRATOS initiative for improving the analyses of observational studies. He is an Honorary Lifetime member of the International Society for Clinical Biostatistics.