COMPARING ESTABLISHED AND NOVEL INDEX DATE SELECTION METHODS: A SIMULATION STUDY IN TWO CLINICAL CONTEXTS
Author(s)
Neisha Opper, PhD, MPH1, Shivani Aggarwal, PhD, MS2, David Goldfarb, PhD, MPH3, Hoa Le, MD, PhD4.
1Landmark Science, La Crescenta, CA, USA, 2Landmark Science, Inc, Los Angeles, CA, USA, 3Landmark Science, New York, NY, USA, 4University of North Carolina, Chapel Hill, NC, USA.
1Landmark Science, La Crescenta, CA, USA, 2Landmark Science, Inc, Los Angeles, CA, USA, 3Landmark Science, New York, NY, USA, 4University of North Carolina, Chapel Hill, NC, USA.
OBJECTIVES: Appropriate index date selection is fundamental to target trial emulation, especially in therapeutic areas such as hematologic and solid cancers where patients often have multiple eligible lines of therapy (LoTs). Established approaches include first eligible LoT (FEL), random eligible LoT (REL), or all eligible LoTs (AEL). A recent study empirically evaluated a novel methodology, stratified random LoT (SRL), for large B-cell lymphoma (LBCL). Here, we compare this novel method to common LoT selection approaches in two different clinical contexts using simulated data.
METHODS: We evaluated four LoT selection methods in a simulation study using synthetic cohorts emulating trial-like relapsed/refractory LBCL and EGFR-positive non-small cell lung cancer (NSCLC) patients and external controls for overall survival where prognosis deteriorated by LoT. Index methods were compared in their ability to estimate effectiveness in naïve and stabilized inverse probability of treatment weighting (sIPTW) analyses. Evaluation metrics included bias, coverage and RMSE. The impact of sample size, effect size, and overlap in prior LoTs was investigated.
RESULTS: In naïve analyses, AEL demonstrated minimal bias (LBCL=−0.03; NSCLC=−0.003) and moderate-to-high coverage (82-86%); weighting improved coverage for LBCL but reduced it for NSCLC. Across naïve and sIPTW analyses, SRL outperformed FEL, showing lower bias and higher coverage [sIPTW for LBCL: SRL (bias=0.001, coverage=96%) vs FEL (bias=0.05, coverage=68%); sIPTW for NSCLC: SRL (bias=−0.05, coverage=81%) vs FEL (bias=0.02, coverage=64%)] and performed comparably to or better than AEL with sIPTW. SRL performance increased as LoT overlap decreased. Naïve REL yielded substantial bias and near-zero coverage, not addressed by weighting.
CONCLUSIONS: SRL outperformed other methods where LoT is a bigger confounder either due to steeper health declines (LBCL) or less LoT overlap between arms. Optimal index date selection depends on clinical and study characteristics. Coverage diagnostics are essential, as low bias alone may mask poor inferential performance.
METHODS: We evaluated four LoT selection methods in a simulation study using synthetic cohorts emulating trial-like relapsed/refractory LBCL and EGFR-positive non-small cell lung cancer (NSCLC) patients and external controls for overall survival where prognosis deteriorated by LoT. Index methods were compared in their ability to estimate effectiveness in naïve and stabilized inverse probability of treatment weighting (sIPTW) analyses. Evaluation metrics included bias, coverage and RMSE. The impact of sample size, effect size, and overlap in prior LoTs was investigated.
RESULTS: In naïve analyses, AEL demonstrated minimal bias (LBCL=−0.03; NSCLC=−0.003) and moderate-to-high coverage (82-86%); weighting improved coverage for LBCL but reduced it for NSCLC. Across naïve and sIPTW analyses, SRL outperformed FEL, showing lower bias and higher coverage [sIPTW for LBCL: SRL (bias=0.001, coverage=96%) vs FEL (bias=0.05, coverage=68%); sIPTW for NSCLC: SRL (bias=−0.05, coverage=81%) vs FEL (bias=0.02, coverage=64%)] and performed comparably to or better than AEL with sIPTW. SRL performance increased as LoT overlap decreased. Naïve REL yielded substantial bias and near-zero coverage, not addressed by weighting.
CONCLUSIONS: SRL outperformed other methods where LoT is a bigger confounder either due to steeper health declines (LBCL) or less LoT overlap between arms. Optimal index date selection depends on clinical and study characteristics. Coverage diagnostics are essential, as low bias alone may mask poor inferential performance.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
SA6
Topic
Study Approaches
Topic Subcategory
Decision Modeling & Simulation
Disease
SDC: Oncology, SDC: Rare & Orphan Diseases, STA: Multiple/Other Specialized Treatments