IMPROVING THE EXTRAPOLATION OF SURVIVAL DATA; HOW SHOULD WE EXPLOIT ADDITIONAL DATA?

Purpose:

   The choice of methods for the extrapolation of survival data can alter estimates of cost-effectiveness. Recent NICE guidelines recommend using additional data (typically observational and non-comparative) either to select the optimal survival function, or to estimate survival beyond study follow-up. This data may be available in the form of summary statistics (scenario A), results from a fitted survival regression model (scenario B), or individual patient data (IPD) (scenario C). In each case, we consider the optimal methods to incorporate additional data using a cost-utility analysis of cemented and cementless Total Hip Replacement (THR) as an exemplar.

Method:

   The primary comparative data on THR were obtained from the National Joint Register of England (NJR). Additional data were taken from Hospital Episode Statistics (HES), an administrative database of publicly funded surgery in England. In scenario A, survival functions applied to the NJR data were compared with summary statistics from HES, and alternative measures of predictive accuracy were evaluated. In scenario B, a regression model estimated from the HES data was used to estimate survival beyond NJR follow-up. In scenario C, IPD from both sources were combined to estimate survival, with adjustment for differences in date of operation. Estimates from each method were applied to a Markov Model of THR to determine cost-effectiveness in men and women aged 60.

Result:

   The alternative methods lead to different conclusions on the cost-effectiveness of cementless THR. The value of additional data in summary form (scenario A) in guiding selection of survival functions is limited. When additional data is reported as a survival model, calibration on the primary data can improve survival predictions. When additional data are available as IPD (scenario C), joint estimation of survival allows superior control for differences between datasets compared to calibration of a model derived only from additional data (scenario B).

Conclusion:

   Additional data are best exploited by combining with study data prior to estimating survival (scenario C). This approach avoids conflating differences in patient populations and changes over time. Additional data in the form of a survival regression model (scenario B) provide a reasonable substitute where IPD is not available. The value of additional data in summary form (scenario A) is limited. This analysis illustrates the potential value of increasing access to IPD.