Comparing Choice Share Prediction Methods for Discrete Choice Experiments: A Case Study With Hypomethylating Agents in Myelodysplastic Syndromes

OBJECTIVES:

Discrete choice experiments (DCEs) can be used to produce choice share predictions, translating patient preferences to product demand. In the healthcare literature, choice share predictions are conventionally calculated using the ‘share of preference’ method. An alternative, but debated, method is the first-choice method. There has been little investigation on differences in predictions between the methods in healthcare.

To compare accuracy of choice share predictions of an in-survey DCE hold-out task using different methods.

METHODS:

An online DCE survey was conducted to elicit preferences of patients with myelodysplastic syndromes comparing benefits, risks, and administration burden of hypomethylating agents (HMAs). Each respondent was shown 14 choice tasks (D-efficient design) with three options: two treatments and an opt-out. Respondents were then presented a hold-out task comprised of four available marketed real-life HMA products (an oral pill, two infusions, and one injection) and an opt-out.

A Multinomial Logit (MNL) and a Mixed Logit (MXL) model were fit, and predictions were made for the hold-out task using three methods: first-choice method per MXL model draw, share-of-preference MXL (incorporating heterogeneity), and share-of-preference MNL. Respondents’ actual choices in the hold-out task were compared to the choice share predictions and the mean absolute error (MAE) was calculated.

RESULTS:

184 respondents completed the survey. MAE was lowest for the first-choice MXL predictions with an MAE of 1.3% (hold-out task choices were oral: 77%, injection: 13%, IV-5: 5%, IV-7: 4%, opt-out: 0.7%; first-choice MXL predictions: 77%, 11%, 6%, 4%, 2%). The predictions were less accurate for the share-of-preference MXL (MAE of 9.5%) and the MNL (MAE of 13.4%).

CONCLUSIONS:

The first-share predictions closely resembled the hold-out task choices, whereas the conventional share-of-preference prediction method did not. Research is required to understand whether this result is generalizable (e.g., when actual choices are less extreme) and to provide researchers with guidance on the prediction method.