Accounting for Delayed Entry in Analyses of Overall Survival in Clinico-Genomic Databases

Abstract

Background: Clinico-genomic databases favor inclusion of long-term survivors, leading to potentially biased overall survival (OS) analyses. Risk set adjustments relying on the independent delayed entry assumption may mitigate this bias. We aimed to determine whether this assumption is satisfied in a dataset of patients with advanced non-small cell lung cancer (aNSCLC), and to give guidance for clinico-genomic OS analyses when the assumption is not satisfied.

Methods: We analyzed the association of timing of next-generation sequencing (NGS) testing with real-world OS (rwOS) in patient data from a United States-based nationwide longitudinal deidentified electronic health records-derived database. Estimates of rwOS using risk set adjustment were compared with estimates computed with respect to all patients, regardless of NGS testing.

Results: The independent delayed entry assumption was not satisfied in this database, and later sequencing had a negative association with the hazard of death after sequencing. In a model adjusted for relevant characteristics, each month delay in sequencing was associated with a 2% increase in the hazard of death. However, until the median survival time, estimates of OS using risk set adjustment are similar to estimates computed for all patients, regardless of NGS testing.

Conclusions: rwOS analyses in clinico-genomic databases should assess the independent delayed entry assumption. Comparisons versus broader population may be useful to evaluate the rwOS differences between calculations using risk set adjustment and patient cohorts where the bias relates to overrepresentation of long survivors.

Impact: This study illustrates practices that can increase the interpretability of findings from OS analyses in clinico-genomic databases.

Figure 1.

Baseline characteristics of all patients in the EHR (left) and all patients in the EHR with an NGS test date before 60 days after the advanced diagnosis date (right), as a function of the year of advanced diagnosis. 95% confidence intervals are included (shaded areas). Dx, diagnosis.

Figure 2.

Survival functions for patients with nonsquamous histology without a record of EGFR⁺ or ALK⁺ alterations. Survival functions are included for (i) all patients in the EHR (all); (ii) NGS-tested patients in the EHR, without accounting for delayed entry (naïve); (iii) NGS-tested patients in the EHR whose entry date is on or prior to the start of first-line therapy (pre-L1); and (iv) NGS-tested patients in the EHR, using risk set adjustment (RSA). A, Patients who received carboplatin and paclitaxel or pemetrexed (potentially also with bevacizumab) as their first-line therapy starting pre-2017, survival functions from start of first-line therapy. B, Survival functions for patients who received carboplatin, pembrolizumab, and pemetrexed as their first-line therapy starting post-2016.

Figure 3.

Estimates of the survival function using risk set adjustment from the simulation study, in which samples of 1,000 patients with 25%, 50%, and 75% of patients entering prior to the start of treatment were created. One hundred samples were created for each of the different percentages of patients entering prior to the index date, and the survival functions shown are the median survival functions across each of those samples. The survival function for the source population (with 48% of patients entering prior to the start of treatment) is essentially the same as that illustrated here for 50% of patients entering prior to the start of treatment.