Variance and Dissent: The Effect of Survivor Bias
Conclusion about the association between valve surgery and mortality in an infective endocarditis cohort changed after adjusting for survivor bias
Survivor bias commonly weakens observational studies, even those published in premier journals. It occurs because patients who live longer are more likely to receive treatment than those who die early. We sought to quantify the effect of survivor bias on the association between valve surgery and mortality in infective endocarditis (IE).Survivor treatment bias, treatment selection bias, and propensity scores in observational research
We would like to thank the editors for the invitation to comment on the article by Tleyjeh et al. published in this issue of the journal [1]. Tleyjeh et al. address the important issue of survivor treatment bias in observational studies and then propose two statistical methods for accounting for this bias. Studies with time-to-event outcomes in which the exposure of interest occurs during the same period during which outcomes occur can be susceptible to survivor treatment bias, also referred to as “immortal time bias” [2] or “time-dependent bias” [3–5].Author's response: the design of observational studies—defining baseline time
We would like to thank the editors both for the initial invitation to comment on the article by Tleyjeh et al. published in this issue of the journal and for the opportunity to respond to their reply [1–3]. In our response, we would like to expand upon an important issue raised by Tleyjeh et al.—that of the lack of a uniform and consistent definition of “baseline” time in some observational studies. This issue merits discussion as it potentially affects many observational studies.Propensity score analysis with a time-dependent intervention is an acceptable although not an optimal analytical approach when treatment selection bias and survivor bias coexist
We are grateful to Austin and Platt [1], who are foremost authorities in the field, for providing a detailed statistical evaluation of our work [2]. These publications provide a platform to launch a much needed discussion of the issues related to time-dependent intervention in propensity score analysis (PSA).
