Mathematics · Statistics and Probability
Modeling oncology surrogate markers
Karen Johnson, Barrie Stokes, Donald D. Anthony, Robert C. Ward, Nick Freemantle, Steven M. Hill
See on the mapAbstract
9639 Background The use of surrogates in clinical trials relies on evidence that treatment benefit on an unmeasured outcome can be reliably predicted from treatment effect on the surrogate. Surrogates for survival have the potential to improve the efficiency and economy of clinical trials. To explore this we model treatment effects on surrogates and outcomes across trials. Methods An exhaustive search was done for randomised trials of medical therapy in advanced colorectal cancer (CC) and non-small cell lung cancer (LC) using standard reporting of median survival and tumour response rate (RR) or median time-to-progression (TTP). Data were also extracted on treatments, quality, age, stage, performance status, rescue therapy, and intent-to-treat. Survival difference between arms was modelled as a function of RR or TTP difference between maximal 2-arms in first line trials weighted by size. Because biology suggests monotonicity without curvilinearity, a linear model was fitted. Adjusted R2 (aR2) and 95% confidence and prediction bands were estimated. Results 13938 search and 25 other citations, 1639 abstracts, and 544 articles yielded 144 (CC) and 190 (LC) qualifying trials of 34862 and 43534 patients. Quality was generally poor. Diagnostics were satisfactory. All models sloped upward to the right indicating more benefit with greater RR or TTP differences. For models RR-CC, RR-LC, TTP-CC, TTP-LC slopes were 7.6, 8.1, 0.87, and 0.85 (all p<0.001) and aR2 0.08, 0.12, 0.19, and 0.31. Modeling both RR and TTP increased aR2 for CC to 0.36 but not for LC. Prediction bands crossed y=0 at 35–50% for RR models and 3–4 months for TTP models. Modeling certain drug class subset increased aR2 and slope but no two slopes significantly differed; adding covariates or fitting fractional polynomials changed results little (data not shown). Conclusions By these data RR and TTP show validity only with large differences. Surrogate validity is graded, not binary, and it independently varies with slope steepness and data scatter. This approach deserves wider exploration. Limitations include incomplete reporting, unsuspected important covariates, and measurement, trial design, and statistical issues. Supported by the Department of Health and Ageing, Canberra Author Disclosure Employment or Leadership Consultant or Advisory Role Stock Ownership Honoraria Research Funding Expert Testimony Other Remuneration Pfizer, sanofi-aventis, Amgen Pfizer, sanofi-aventis, Amgen Pfizer, sanofi-aventis, Amgen Pfizer, sanofi-aventis, Amgen