OBJECTIVES: Health Technology Assessment (HTA) bodies increasingly recognize the importance of including caregiver spillovers in reimbursement decision-making. However, examples of HTAs incorporating within-trial caregiver data remain scarce. The EPISODE study assessed the impact of seizure dogs on individuals with severe refractory epilepsy and their primary caregivers. This paper explores various methods to link caregiver EQ-5D data with parameters commonly used in economic evaluations.

METHODS: 266 caregiver-patient dyad observations were collected over the three-year follow-up of the randomized controlled trial. Random-effects Tobit models explored the following predictor variables of caregiver EQ-5D-5L utilities separately: patient EQ-5D-5L utilities, patient EQ-5D-5L domain scores, seizure frequency (the primary outcome of the EPISODE trial), seizure-free days, treatment arm, time on treatment, and informal care hours. Utilities were calculated using the Dutch tariff. Predictive accuracy of the different models was evaluated using the mean absolute prediction error (MAE), and the root mean squared prediction error (RMSE).

RESULTS: The mean utility score was 0.832 (SD 0.147) for caregivers when patients received standard of care alone, and 0.861 (SD 0.110) when patients were partnered with a seizure dog condition. The models for predicting caregiver utility scores as a function of informal care hours (MAE 0.0878, RMSE 0.1136) or seizure-free days (MAE 0.0914, RMSE 0.1285) performed best. Among the other models analyzed, there was minimal variation in predictive accuracy (MAE ranges 0.934 – 0.957, RMSE ranges 0.1313 - 0.1347). Adding patient age and patient gender to the models slightly improved the predictions.

CONCLUSIONS: This analysis demonstrates the feasibility of estimating caregiver spillovers using within-trial data. Different methods to model these effects may lead to different results. Therefore, careful consideration is essential when determining how to integrate such data into economic evaluations, ensuring that clinical plausibility is considered alongside the predictive performance of the modelling approach.