I am interested in conducting a meta-analysis of overall survival (OS) in two clinical trials using the restricted mean survival time (RMST) method, as well as performing subgroup meta-analyses using conventional methods.
To perform the RMST analysis, I need to digitize the survival curves (Kaplan-Meier) from research papers (WebPlotDigitizer - Copyright 2010-2022 Ankit Rohatgi) and calculate the RMST. While I have a tutorial on how to do this, I am unsure if I am doing it correctly, so I would greatly appreciate your expertise in confirming its accuracy.
The main reason for choosing the RMST approach is that the survival curves do not exhibit a clear proportional hazards (PH) behavior. The relevant papers with DOIs are: DOI: 10.1056/EVIDoa2200015 and DOI: 10.1016/S0140-6736(23)00727-4.
We sincerely appreciate your kind help and recommendations.
My focus on RMST revolves around the understanding of the impact of immunotherapy on HCC. Similar to other tumors, the trials showed very late splitting curves, with most of them reporting the hazard ratio (HR) and using benchmark analysis. Interpreting the curves and the overall effect is challenging. Therefore, I would like to know if you consider evaluating the RMST to better understand the benefit between arms, and if a meta-analysis of RMST could be valuable in gaining deeper insights.
We focus too much on a dichotomy of PH vs non-PH when we should be using Bayesian models containing non-PH parameters (e.g., interaction between treatment and time) with priors on those parameters
RMST must be covariate-adjusted to be interpretable so we should not be estimating RMST from Kaplan-Meier
Yes. Any time you are estimating an absolute quantity (e.g. risk or life expectancy) the quantity will vary with patient risk factors, and absolute risk reduction and increase in life expectancy will expand for higher risk subjects. So estimates will be functions of covariates. Failure to adjust to covariate settings will result in marginal averages that may not apply to anyone, and won’t apply to the population as a whole because RCTs do not randomly sample from populations. Hence covariate distributions in RCTs may be different from distributions in the population.
The discussion referenced by @Pavlos_Msaouel is incredible. Regarding whether randomization solves the problem that absolute quantities must be covariate-specific, it clearly does not, because randomization does not narrow the within-treatment distribution of prognostic factors. Stratification can help in the very special case where (1) you don’t provide an average (over strata) difference but only provide stratum-specific estimates and (2) there is only one prognostic factor and it is categorical, and the one used in the stratification.