A SEQUENTIAL MATCHING APPROACH TO ADDRESS IMMORTAL TIME BIAS: APPLICATION TO THE MGUS STUDY
Author(s)
Dazheng Zhang, PhD1, Yongmei Chen, MS, MD1, Yimeng Shang, PhD1, Yong Lin, PhD2;
1Eli Lilly, Indianapolis, IN, USA, 2Eli Lilly and company, Indianapolis, IN, USA
1Eli Lilly, Indianapolis, IN, USA, 2Eli Lilly and company, Indianapolis, IN, USA
OBJECTIVES: In targeted trial emulation, patients in the treated arm can have different waiting times from diagnosis to treatment compared to active controls, creating immortal time bias (ITB). Existing methods, including propensity-score matching, fail to balance time-to-treatment distributions and cannot fully eliminate ITB. We propose a sequential matching method that removes ITB by perfectly balancing time from diagnosis to treatment between arms.
METHODS: We developed a sequential matching approach that assigns control patients to treated patients within identical time-from-diagnosis intervals. Time from diagnosis to treatment is divided into intervals, creating a series of small trials where patients with the same interval are matched. Within each interval-matched trial, propensity score matching balances other covariates. We assessed balance using distribution plots of immortal time and standardized mean differences (SMD). We illustrated it using Market Clarity data for Monoclonal Gammopathy of Undetermined Significance (MGUS) patients comparing GLP-1 agonists versus non-GLP-1 antidiabetic drugs.
RESULTS: A real-world data example was presented to compare the effectiveness of traditional baseline matching versus sequential trial emulation. Among 11,150 MGUS patients, 1,694 received GLP-1 agonists and 9,456 received non-GLP-1 antidiabetic drugs. Traditional baseline matching failed to balance time-to-treatment distributions between groups. Sequential matching yielded 1,327 matched pairs. Time from MGUS diagnosis to treatment was identically distributed between groups: mean 1.4 years (SD 0.8) for both GLP-1 and control arms. All covariates including race, sex, MGUS diagnosis timing, and BMI achieved excellent balance (SMD <0.1). Mean SMD across all covariates with 95% CI was 0.052 (0.007-0.097).
CONCLUSIONS: Sequential matching successfully eliminated ITB by achieving perfect balance in time-to-treatment distributions while maintaining covariate balance. This method offers a robust solution for targeted trial emulation where immortal time bias threatens validity. Limitations include reduced sample size and potential loss of generalizability. Future research should explore its applicability across diverse therapeutic areas.
METHODS: We developed a sequential matching approach that assigns control patients to treated patients within identical time-from-diagnosis intervals. Time from diagnosis to treatment is divided into intervals, creating a series of small trials where patients with the same interval are matched. Within each interval-matched trial, propensity score matching balances other covariates. We assessed balance using distribution plots of immortal time and standardized mean differences (SMD). We illustrated it using Market Clarity data for Monoclonal Gammopathy of Undetermined Significance (MGUS) patients comparing GLP-1 agonists versus non-GLP-1 antidiabetic drugs.
RESULTS: A real-world data example was presented to compare the effectiveness of traditional baseline matching versus sequential trial emulation. Among 11,150 MGUS patients, 1,694 received GLP-1 agonists and 9,456 received non-GLP-1 antidiabetic drugs. Traditional baseline matching failed to balance time-to-treatment distributions between groups. Sequential matching yielded 1,327 matched pairs. Time from MGUS diagnosis to treatment was identically distributed between groups: mean 1.4 years (SD 0.8) for both GLP-1 and control arms. All covariates including race, sex, MGUS diagnosis timing, and BMI achieved excellent balance (SMD <0.1). Mean SMD across all covariates with 95% CI was 0.052 (0.007-0.097).
CONCLUSIONS: Sequential matching successfully eliminated ITB by achieving perfect balance in time-to-treatment distributions while maintaining covariate balance. This method offers a robust solution for targeted trial emulation where immortal time bias threatens validity. Limitations include reduced sample size and potential loss of generalizability. Future research should explore its applicability across diverse therapeutic areas.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR62
Topic
Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
SDC: Oncology