Retrospective survival analysis is commonly used to study comparative effectiveness of treatments (tx) in data derived from electronic health records (EHR). Unlike clinical trials, the index date for survival must be defined as part of the study design, and there may be multiple candidate index dates per patient in cases where subjects receive sequential lines of therapy. We present a multi-state semi-Markov model with cause-specific Weibull hazards for tx switching and mortality. To assess the performance of proportional hazards models comparing novel and standard-of-care (SOC) tx, simulations are modeled on a nationwide EHR-derived de-identified oncology database. Patient-specific tx effects represent a mixture of responders and non-responders. When use of novel tx increases in later lines of therapy, and SOC treatments may be repeated, using first or random eligible line of therapy as index date leads to biased estimates of tx effect, even with covariate adjusted or inverse probability of treatment weighted models. Inclusion of all eligible lines of therapy, with robust standard errors to account for clustering by subject, shows reduced bias and well-calibrated standard errors.