OBJECTIVES: In health technology assessments indirect comparisons are often required when head-to-head data are unavailable. In these instances a survival model is usually fitted to one treatment, with hazard ratios (HRs) from the indirect comparison then applied to produce survival curves for the comparators. Common survival models are proportional hazard models or accelerated failure time (AFT) models. Best practice suggests that the treatment effect is applied to a model on the same scale, such as applying a HR to a proportional hazard survival model. However, in health technology assessments it is common for HRs to be applied to survival models with an AFT distribution. This study explores the effect of this, sub-optimal, practice.

METHODS: We simulated survival data for two treatments using an AFT model, and then estimated the treatment effect in terms of a HR. We then fitted an AFT model to the control group and estimated survival in the experimental group using the estimated HR. We compared the restricted mean survival time estimated for the experimental group to the true restricted mean survival time. Mean percentage bias was estimated over 1,000 iterations. This study explored whether the scale of bias varied depending on the AFT model, size of the treatment effect or whether censoring was present.

RESULTS: The percentage bias in estimates varied significantly depending on the AFT model and size of the treatment effect. Estimates with a Weibull model were not biased, whereas the bias in a log-normal or log-logistic ranged from negligible with a treatment effect of one, to 87% with a treatment effect of 0.5.

CONCLUSIONS: The application of a HR to certain AFT models results in biased survival estimates. The bias is greater in instances where the treatment effect is large, or where survival estimates are already uncertain due to censoring.