OBJECTIVES: Lack of IPD poses challenges in estimating the correlation between two endpoints. We proposed a novel approach to parameterize an illness-death model with pseudo-IPD and applied it to estimate the correlation between PFS and OS in previously untreated metastatic melanoma.

METHODS: For each trial identified by a systematic literature review, pseudo-IPD were reconstructed from digitized PFS and OS Kaplan-Meier curves by applying the Guyot algorithm. In the model, time-to-progression and OS were assumed to be independent Weibull random variables with shared shape and independent scale parameters to invoke a closed-form equation for estimating Pearson’s correlation between PFS and OS. Transitions between the model states were simultaneously estimated by iteratively applying fractional polynomial Bayesian models assuming a fixed but different proportion of patients progressing before death in each iteration. Candidate models were ranked according to their deviance information criteria. Predicted OS curves from best-fitting models were compared with reported OS curves visually and via restricted mean survival time (RMST). An overall correlation estimate was obtained through a random-effects meta-analysis of the arm-specific estimates using Fisher’s Z transformation.

RESULTS: Analyses included 24 trials (published between 2000-2020) with a total of 49 arms. The estimated proportion of patients experiencing progression (mean=0.816, standard error=0.011) had a tendency to be higher for modern treatments (0.70-0.94 across 36 arms) and lower in control arms in general. In 31/49 trial-arms, predicted OS captured the observed OS trend adequately with a statistically insignificant RMST at the 95% confidence level. The estimated endpoint correlations were 0.987 (95% CI: 0.979-0.992) on the full evidence base and 0.986 (95% CI: 0.975-0.992) among the aforementioned 31 trial-arms with better fit.

CONCLUSIONS: Modelling PFS and OS dependently using pseudo-IPD can strengthen surrogacy analyses by enabling endpoint correlation assessment. Our case-study found strong correlation between PFS and OS in previously untreated metastatic melanoma using published data.