Quality of life and patient preferences: identification of subgroups of multiple sclerosis patients

Abstract

Purpose

The aim of this study was to estimate preferences related to quality of life attributes in people with multiple sclerosis, by kee** heterogeneity of patient preference in mind, using the latent class approach.

Methods

A discrete choice experiment survey was developed using the following attributes: activities of daily living, instrumental activities of daily living, pain/fatigue, anxiety/depression and attention/concentration. Choice sets were presented as pairs of hypothetical health status, based upon a fractional factorial design.

Results

The latent class logit model estimated on 152 patients identified three subpopulations, which, respectively, attached more importance to: (1) the physical dimension; (2) pain/fatigue and anxiety/depression; and (3) instrumental activities of daily living impairments, anxiety/depression and attention/concentration. A posterior analysis suggests that the latent class membership may be related to an individual’s age to some extent, or to diagnosis and treatment, while apart from energy dimension, no significant difference exists between latent groups, with regard to Multiple Sclerosis Quality of Life-54 scales.

Conclusions

A quality of life preference-based utility measure for people with multiple sclerosis was developed. These utility values allow identification of a hierarchic priority among different aspects of quality of life and may allow physicians to develop a care programme tailored to patient needs.

Appendix 1

Latent class model (finite mixture logit model) in discrete choice experiments.

Following the random utility framework, the utility that individual n assign to alternative j can be written as U _jn  = V _jn  + ɛ _jn , where $V_{jn} = \sum_{K = 1}^{K} \beta_{k} x_{jkn}$ is the deterministic part of utility (in which K is the number of attributes used in the experiment, k = 1,2,…,K, x _nk is an observed variable related to attribute k and β _k is the attribute coefficient, homogeneous across the population), while ɛ _jn is the stochastic part, also capturing the unobserved heterogeneity Given two alternatives i and j, an individual will choose alternative j if U _jn  > U _in .

Let P _nt (j|β) give the probability of respondent n choosing alternative j on an occasion (called here scenario) t, conditional on a vector of attributes coefficients (β), in a fixed logit model we have:

$$P_{nt} (j|\beta ) = \frac{{\exp \left( {x_{jnt} \beta } \right)}}{{\sum\nolimits_{j = 1}^{J} {\exp \left( {x_{jnt} \beta } \right)} }}$$

Preference variation among individuals that is unaccounted for in modelling can result in a biased estimate. Two types of approaches allow to take into account differences in preferences: the random coefficient and the latent class models.

In the random utility model (continuous mixture model), the vector β follows a random distribution with parameters Ω and the choice probabilities are given by:

$$P_{nt} (j|\varOmega ) = \int_{\beta } {P_{nt} (j|} \beta )f(\beta |\varOmega ){\text{d}}\beta$$

where P _nt (j|β) is the logit choice probability and f(β|Ω) is the density function for the vector of attributes coefficients β.

In a latent class model (finite mixture model), preference variation is accommodated by identifying C groups of respondents with different values for the vector of attributes coefficients (β _c). In this model, f(β) is a discrete distribution and the choice probability is the weighted sum of the choice probabilities across the C classes, with the class allocation probability π _nc has been used as weight:

$$P_{nt} (j|\beta_{c} ) = \sum\nolimits_{c = 1}^{C} { \, \pi_{nc} P_{nt} (j|\beta_{c} )}$$

This specification is useful if there are C segments in the population, each of which has its own choice preferences. The share of population in class c is π _nc ($\sum\nolimits_{c = 1}^{C} { \, \pi_{nc} = 1}$), and it is estimated along with the βs for each segment.