Tony O'Hagan (75KB)

Download Report

Transcript Tony O'Hagan (75KB)

Introduction
Tony O’Hagan
25 Sept 07
FF8 - Discrete Choice Data
The basic idea
• We have a finite collection of statuses: S1, S2,
…, SN
• We want to assign utility values to these: u(S1),
u(S2), …, u(SN)
• We ask people to compare two or more
statuses, and say which they prefer
• From many such expressed preferences, we aim
to make inference about the underlying utilities
25 Sept 07
FF8 - Discrete Choice Data
The basic problem
• Preferring Si to Sj basically means u(Si) > u(Sj)
• But if all we learn from the data is these
inequalities, then all we can infer is the ordering
of the utilities
• We cannot infer values for the utilities
• E.g. If S1 has highest utility, then
› I will always choose this in any comparison
› I can infer it has higher utility than any other status
› But not how much higher
25 Sept 07
FF8 - Discrete Choice Data
Just add noise
• In practice, people aren’t that consistent
• Model
› Person states preference for Si over Sj if v(Si) > v(Sj)
› Where v(S) = u(S) + e
› And e is an error term
» May be due to the person not evaluating their utilities accurately, or
because of variation between people
• Now if two statuses have similar utilities there is a
chance that the preference will be stated the other way
round
› We can infer differences in utilities from the frequency of
preferences
25 Sept 07
FF8 - Discrete Choice Data
But …
• We have to assume a distribution for e
› Including variance
» which will strongly affect frequency of preference reversal
› And it is usual to assume independence
» which I don’t believe
• I can’t think of any other problem where getting
better quality data would make the problem
worse
› Which suggests there is something unsatisfactory
about this solution
25 Sept 07
FF8 - Discrete Choice Data
Health states
• We will hear about various versions of this
approach
• In the context of health economics, the statuses
we wish to value are those describing health
• A health state descriptive system typically
represents health by a level in each of a number
of dimensions
› E.g. EQ-5D has 5 dimensions and 3 levels in each
dimension
› So this system has 35 = 243 distinct health states
25 Sept 07
FF8 - Discrete Choice Data
Alternative choice systems
• Consider the EQ-5D state 2 1 1 2 3
›
›
›
›
›
Mobility – some difficulties with mobility
Self-Care – no problems, can care for self
Usual Activities – can engage in all usual activities
Pain/Discomfort – some pain/discomfort
Anxiety/Depression – severe anxiety/depression
• Status
› The statuses are whole state descriptions like above (N = 243)
› The statuses are individual levels (N = 15)
» Utility for whole state is then obtained by assuming additivity
• Task
› Compare all in a set, to get a full ranking from best to worst
› Identify only the best and worst in a set (best-worst method)
25 Sept 07
FF8 - Discrete Choice Data