List of publications, posters and presentations

Orme ME,  Perard R. Double-hurdle and Heckman models for assessing patient preferences with zero responses: an example using HIV treatment adherence ratings (594)


Background: Patient preference data often contain responses indicating no preference. Different interpretations of this ‘zero’ observation require different estimation approaches. Objectives: To compare the two-stage Heckman and double-hurdle model. Methods: Data were collected from HIV patients participating in a prospective survey and discrete choice experiment (DCE). The DCE consisted of 12 hypothetical, unlabelled paired choices. For each scenario, patients used a sliding scale (0 (no preference) to 100 (strongest preference)) to rate the treatment option that they thought would maximise their adherence to treatment. The responses were analysed in STATA v13.1. In the Heckman model the probability of a non-zero observation is from a probit model fitted to the whole sample (stage 1: participation model). Stage 2 (adherence behaviour model) estimates preference weightings from the non-zero observations using the inverse Mills ratio as an explanatory variable (calculated in stage 1). A double hurdle model estimates both stages simultaneously using all observations. Results: The sample consisted of 278 UK respondents. The median age of the respondents is 44 years (range 21-66), 36% of the respondents are on a single tablet regimen, 9% take five tablets or more per day and 43% of patients reported that they never skip treatment (maximum self-reported adherence). Across all preference observations, 63% were rated 0. Both models provide similar treatment attribute rankings and preference weightings. Age, single tablet regimen and maximum self-reported adherence were found to be significant predictors of non-zero responses in the Heckman participation model but not in the double-hurdle model. Conclusion: For this example we infer that a zero response represents non-participation rather than a utility maximising solution and that the Heckman two-stage model provides the best estimate of likely adherence weighting in this case.

Presented at the 12th Annual Meeting of the HTAi, 15-17 June 2015, Oslo, Norway.

Citation: HTAI 2015

HTAI 1 2015