Authors:
Sondhi A.
Databases derived from electronic health records (EHRs) are commonlysubject to left truncation, a type of selection bias that occurs when patientsneed to survive long enough to satisfy certain entry criteria. Standard methodsto adjust for left truncation bias rely on an assumption of marginal indepen-dence between entry and survival times, which may not always be satisfied inpractice. In this work, we examine how a weaker assumption of conditionalindependence can result in unbiased estimation of common statistical parame-ters. In particular, we show the estimability of conditional parameters in atruncated dataset, and of marginal parameters that leverage reference datacontaining non-truncated data on confounders. The latter is complementary toobservational causal inference methodology applied to real-world externalcomparators, which is a common use case for real-world databases. We imple-ment our proposed methods in simulation studies, demonstrating unbiasedestimation and valid statistical inference. We also illustrate estimation of a sur-vival distribution under conditionally independent left truncation in a real-world clinico-genomic database.
Sources:
Pharmaceutical Statistics