OBJECTIVES: Digitization is commonly used to create pseudo-patient data from published Kaplan-Meier plots, to inform comparative (cost-)effectiveness analyses for health technology appraisal (HTA) in the absence of patient-level effectiveness data for all relevant treatments. While event times are clear in Kaplan-Meier plots, censor times are not always identifiable. In these cases, assumptions are required about the distribution of censoring within identifiable intervals. This study aimed to illustrate the importance of such assumptions, for HTA-relevant outcomes.

METHODS: As a case study, pseudo-patient data were generated by applying the widely used Guyot et al (2012) algorithm, as implemented in the R survHE package, to overall survival Kaplan-Meier data from a randomized, controlled trial (RCT) which exhibited significant censoring (Baselga et al, 2012). Next, modified versions of this algorithm were run more comprehensively utilizing information from risk tables and applying alternative assumptions about the distribution of censor points within intervals: all earliest; all mid-point; all latest; uniformly distributed; exponential-uniform mixture distributed. Event times were unchanged. To illustrate the potential implications of censoring assumptions, the estimated RCT hazard ratio (HR) for each was compared. Restricted-mean survival time (RMST) was also estimated for each censoring approach, assuming a Gompertz model.

RESULTS: HR estimates spanned 0.653-0.672 (Guyot: 0.653) and Gompertz 10-year RMST spanned 2.92-3.57 (Pertuzumab; Guyot: 3.57) years and 2.61-2.98 (Control; Guyot: 2.98) years. Changes to the placement of censoring points had little impact on the estimated relative treatment effect, but relatively large effect on estimated absolute survival time, with estimates based on Guyot et al (2012) appearing optimistic.

CONCLUSIONS: The results of this study highlight the importance of reasoned specification of censoring assumptions in the generation of individual patient data from Kaplan-Meier estimates. Time-to-event estimates can be highly influential in cost-effectiveness models; modest improvements can affect HTA decision-making.