All Times EDT
Keywords: real world data, truncated data, survival analysis
Purpose: In real world datasets, such as those derived from electronic health records, patients often need to satisfy an entry criteria in order to be observed. When analyzing survival outcomes, this leads to a selection bias known as left truncation: patients who did not live long enough to enter the dataset are not observed and thus their outcomes are truncated. Standard methods can only suitably account for this process when survival and entry time are independent.
Methods: We conduct simulation studies of three common analytic settings given dependent left truncation in order to quantify the magnitude and direction of estimator bias. We also outline a procedure for conducting a simulation-based sensitivity analysis for an arbitrary dataset subject to dependent left truncation in order to better quantify uncertainty around specific analytic results.
Results: Our simulation results show that estimator bias varies substantially between analytic settings. In particular, estimation of absolute quantities (e.g. median survival time) is more subject to bias than that of relative quantities (e.g. hazard ratios). When comparing a truncated real world arm to a trial treatment arm, we observe that the estimated hazard ratio is biased towards the null, providing conservative inference. The most important data-generating aspect contributing to bias is the proportion of left truncated patients, given any level of dependence between survival and entry time.
Conclusions: Our simulation results show that the amount of bias due to dependent left truncation varies across analytic settings and truncation prevalence. However, there are settings where the bias may be low or predictable enough to provide useful inference, even without applying mitigations. For specific datasets and analyses that may differ from the examples provided, we recommend applying our simulation-based sensitivity analysis approach in order to determine how results would change given varying parameters.