Seamus Kent1, Alastair Gray1, Iryna Schlackow1, Crispin Jenkinson2, Emma McIntosh3. 1. Health Economics Research Centre, Nuffield Department of Population Health, University of Oxford, Oxford, UK (SK, AG, IS) 2. Health Services Research Unit, Nuffield Department of Population Health, University of Oxford, Oxford, UK (CJ) 3. Health Economics and Health Technology Assessment, University of Glasgow, Glasgow, Scotland, UK (EM)
Abstract
OBJECTIVE: To compare a range of statistical models to enable the estimation of EQ-5D-3L utilities from responses to the Parkinson's Disease Questionnaire 39 (PDQ-39). METHODS: Linear regression, beta regression, mixtures of linear regressions and beta regressions, and multinomial logistic regression were compared in terms of their ability to accurately predict EQ-5D-3L utilities from responses to the PDQ-39 using mean error (ME), mean absolute error (MAE), and mean square error (MSE), overall and by Hoehn and Yahr stage. Models were estimated using data from the PD MED trial (n = 9123) and assessed on both the estimation data as well as external data from the PD SURG trial (n = 917). RESULTS: Overall, the differences in the metrics of fit between models were small in both data sets, with performance poorer for all models in PD SURG. The performance across Hoehn and Yahr stages 1 to 3 were also similar, but multinomial logistic regression was found to exhibit less bias and better individual-level predictive accuracy in PD MED for those in Hoehn and Yahr stages 4 or 5. Overall, the multinomial logistic regression reported an ME of 0.038 out of sample and MAEs of 0.128 and 0.164 and MSEs of 0.030 and 0.044 in the estimation and external data sets, respectively. Poorer levels of the mobility domain score of the PDQ-39 were associated with increased odds of reporting problems for all EQ-5D domains except anxiety/depression. CONCLUSIONS: Finite mixture models with only few components can approximate the distribution of EQ-5D-3L utilities well but did not demonstrate improvements in predictive accuracy compared with multinomial logistic regression in the present data set.
OBJECTIVE: To compare a range of statistical models to enable the estimation of EQ-5D-3L utilities from responses to the Parkinson's Disease Questionnaire 39 (PDQ-39). METHODS: Linear regression, beta regression, mixtures of linear regressions and beta regressions, and multinomial logistic regression were compared in terms of their ability to accurately predict EQ-5D-3L utilities from responses to the PDQ-39 using mean error (ME), mean absolute error (MAE), and mean square error (MSE), overall and by Hoehn and Yahr stage. Models were estimated using data from the PD MED trial (n = 9123) and assessed on both the estimation data as well as external data from the PD SURG trial (n = 917). RESULTS: Overall, the differences in the metrics of fit between models were small in both data sets, with performance poorer for all models in PD SURG. The performance across Hoehn and Yahr stages 1 to 3 were also similar, but multinomial logistic regression was found to exhibit less bias and better individual-level predictive accuracy in PD MED for those in Hoehn and Yahr stages 4 or 5. Overall, the multinomial logistic regression reported an ME of 0.038 out of sample and MAEs of 0.128 and 0.164 and MSEs of 0.030 and 0.044 in the estimation and external data sets, respectively. Poorer levels of the mobility domain score of the PDQ-39 were associated with increased odds of reporting problems for all EQ-5D domains except anxiety/depression. CONCLUSIONS: Finite mixture models with only few components can approximate the distribution of EQ-5D-3L utilities well but did not demonstrate improvements in predictive accuracy compared with multinomial logistic regression in the present data set.
Authors: Mónica Hernández Alava; Allan Wailoo; Stephen Pudney; Laura Gray; Andrea Manca Journal: Health Technol Assess Date: 2020-06 Impact factor: 4.014
Authors: Gang Chen; Miguel A Garcia-Gordillo; Daniel Collado-Mateo; Borja Del Pozo-Cruz; José C Adsuar; José Manuel Cordero-Ferrera; José María Abellán-Perpiñán; Fernando Ignacio Sánchez-Martínez Journal: Patient Date: 2018-12 Impact factor: 3.883
Authors: Tomasz Fundament; Paul R Eldridge; Alexander L Green; Alan L Whone; Rod S Taylor; Adrian C Williams; W M Michael Schuepbach Journal: PLoS One Date: 2016-07-21 Impact factor: 3.240