ISPOR Europe 2018
Barcelona, Spain
November, 2018
PCN356
Cancer, Multiple Diseases/No Specific Disease
Patient-Reported Outcomes & Patient Preference Studies (PRO)
Quality-of-Life (hrQoL), Health States/Utilities (HS)
HOW SHOULD WE VALIDATE UTILITY MAPPING ALGORITHMS BEFORE USE? AN EXAMPLE IN NON-SMALL CELL LUNG CANCER
Gregory J1, Dyer M2, Hoyle C2, Mann H2, Hatswell A3
1BresMed Health Solutions, Sheffield, UK, 2AstraZeneca, Cambridge, UK, 3Delta Hat Limited and University College of London, Nottingham, UK
OBJECTIVES

:
Mapping algorithms can be used to generate health state utilities when a preference-based instrument is not included in a clinical study. Whilst many mapping algorithms are published (with guidance on their creation), the literature is scarce on how the external validity of algorithms should be tested. Our aim was to investigate the performance of published mapping algorithms in non-small cell lung cancer (NSCLC) between the EORTC QLQ-C30 and EQ-5D instruments.

METHODS

:
We conducted a targeted literature review to identify published mappings, then applied these to a dataset of patients with EGFR-mutation positive NSCLC which contained both scales. Performance of the algorithms was evaluated using the mean absolute error, root mean squared error, and graphical techniques for the observed versus predicted EQ-5D utilities. These statistics were also calculated across the range of utility values (as well as ordinary least squares and quantile regression), to investigate how the mappings fitted across all values, in addition to the mean utility.

RESULTS

:
Three algorithms developed in NSCLC were identified. The algorithm based on response mapping (Young et al., 2015) fitted the validation dataset across the range of observed values with similar goodness-of-fit statistics to the original publication (overall MAE of 0.087 vs 0.134). The two algorithms based on beta-binomial models presented a poor fit to both the mean and distribution of utility values (MAE 0.176, 0.178).

CONCLUSIONS

:
The validation of mapping algorithms is key to demonstrating their generalisability beyond the original dataset. This is particularly the case across the range of plausible utility values beyond the mean utility value – with a simple comparison of patient populations being insufficient. Without this validation process it is possible a mapping algorithm does not generalise well to other datasets, which could lead to biased results and consequently, incorrect decisions regarding the adoption of health technologies.