Estimating a preferencebased index from HRQoL measures: lessons from developing the SF6D

1
Estimating a preference-based index from HRQoL
measures lessons from developing the SF-6D

• John Brazier PhD
• Health Economics and Decision Science
• University of Sheffield,UK
• Prepared for a Workshop of the IOM
  Committee to Evaluate Measures of Health benefits
  for Environmental, Health and Safety Regulation,
  Washington, November 30 2004

2
Measuring health related quality of life SF-36

• A 36 item questionnaire designed for
  administration in a wide range of interfaces and
  languages
• Measures general health across 8 dimensions
• produced 8 dimension scores (summative
  scoring)
• and two summaries (scored by factor
  analysis)
• Most widely used measure of HRQOL in the world
• Validated across a wide range of medical
  conditions
• No preference-based single index

3
Approaches to deriving a single index from HRQoL
measures

• Apply arbitrary weights to dimension scores
• will not reflect peoples preferences
• Map or cross walk between HRQoL measure onto
  existing preference-based measure (e.g QWB or
  EQ-5D)
• Only ever second best since depends on overlap
  between descriptive systems
• Loose sensitivity and coverage of HRQoL
• Value HRQoL
• this has been the approach with SF-6D

4
Direct valuation of a HRQoL measure example the
SF-36
Aim was to obtain a preference-based measure of
health from the SF-36 which

• must be readily derived from SF-36 and SF-12
  questionnaire data
• based on a recognised valuation technique
• (e.g. standard gamble)

5
Stages of research

• Adapt the SF-36 into a simplified health state
  classification amenable to valuation (currently
  defines billions of possible states or
  combinations)
• Value a sample of states defined by the new
  classification
• Use multivariate statistical analysis to estimate
  an algorithm for scoring the new classification

6
Construction of a simplified descriptive system
the SF-6D

• Reduced from 8 to 6 dimensions by excluding
  general health and combing role limitation
  dimensions
• 11 items selected on basis of results from Rasch
  analysis (pattern of responses) and factor
  analysis (correlations)
• Result was a 6 dimensional multi-level (n)
  classification the 6D physical functioning (6),
  role limitation (4), social functioning (5), pain
  (6), mental health (5), vitality (5)
• 18,000 Health states defined by selecting one
  level from each dimension

7
Example dimension Social functioning

• Your health limits your social activities (like
  visiting friends, relatives etc.) none of the
  time
• Your health limits your social activities (like
  visiting friends, relatives etc.) a little of the
  time
• Your health limits your social activities (like
  visiting friends, relatives etc.) some of the
  time
• Your health limits your social activities (like
  visiting friends, relatives etc.) most of the
  time
• Your health limits your social activities (like
  visiting friends, relatives etc.) all of the time
  

8
Health state 424444

• Your health limits you a lot in moderate
  activities (such as moving a table, pushing a
  vacuum cleaner, bowling or playing golf)
• You are limited in the kind of work or other
  activities as a result of your physical health
• Your health limits your social activities (like
  visiting friends, relatives etc.) most of the
  time
• You have pain that interferes with your normal
  work (both outside the home and housework)
  moderately
• You feel tense or downhearted and low most of the
  time.
• You have a lot of energy a little of the time

9
2. Valuing SF-6D health states

• 249 SF-6D health states (out of a possible
  18,000) selected on basis of orthogonal design a
  supplemented by a random sample
• representative sample of 836 members of UK
  general population were interviewed (65 response
  rate)
• respondents asked to rank and value six SF-6D
  health states using the ping pong variant of
  standard gamble developed at McMaster

10
Results of survey

• 3518 valuations across 249 health states (14
  valuations per state)
• mean health state values ranged from 0.21 to 0.99
  (standard deviations 0.2 to 0.45)
• distribution of values was negatively skewed

11
3. Modelling preference data

• i 1, 2, , 5 health states (HS)
• j 1, 2, , 495 respondents
• yij SG score for HS i valued by respondent j
• x vector of main effects dummies
• r vector of interaction effects
• z vector of personal characteristics
• g link function specifying functional form

12
Modelling issues
Started with a simple additive OLS model with
dimension levels entered as a dummy variables

• Issues
• Individual (observations 3518) vs. mean
  (observations 249) level modelling
• Hierarchical data at individual level - used
  random effects model
• Skewness so examined alternative functional
  forms
• Interactions - used various summary extreme and
  count measures

13
Modelling results

• Estimates
• estimates of main effects robust across
  specifications
• broadly consistent with 6D (though some
  inconsistencies)
• pain, mental health and physical functioning had
  largest effects
• Fit
• predictions 79 within /- .1 and - 53 within
  /- 0.05
• explanatory power of 0.51 for mean models
• tendency to over predict at the lower end
  (predictions 0.301 to 1.0)

14
Preference-based scoring algorithm
15
Further work on SF-6D

• Completed
• tested psychometric performance of SF-6D compared
  to 8 dimension and 2 summary scores in 9 medical
  conditions and found index compared favourably
  in terms of effects sizes and standardised
  response means (paper in preparation)
• Estimated a model for the SF-12 (Med care
  200442851-859))
• Valuation work in other countries Japan
  completed.
• Comparisons with other preference-based measures
  (Health Economics, 2004, 13(90)873-884)
• Application of this method to condition specific
  measures (including atopic dermatitis,
  incontinence and menopausal symptoms)
• On-going
• Bayesian methods of estimation and design (using
  Gaussian processes)
• Valuation using Rank and pair wise data
• Valuation work in other countries Hong Kong
  on-going and Singapore in preparation)

16
Key lesson descriptive system

• Comparisons with other preference-based measures
  reveals a floor in the scale lowest score is 0.3
  compared to -0.59 EQ-5D (UK) and 0.36
• Floor effects evident for physical functioning
  and role limitation, but not other dimensions
• When designing a descriptive system need to
  ensure coverage (i.e. dimensions) and range of
  severity adequate
• Revise SF-6D by applying IRT to item pools to
  lower the floor

17
Key lesson design of valuation study and
modelling health state

• Gaussian process estimation produces better
  fitting models (Kharroubi et al, Applied
  Statistics)
• It overcame problem with over-predicting lower
  end (now 0.21) and reduced errors in predictions
• Provides a more informed basis for sampling
  states for future valuation work using prior
  knowledge (e.g. other valuation work)

18
Key lesson whose values

• General population give very low values serious
  question mark over negative values in public
  policy (particularly in EQ-5D and HUI3) and
  relationship to the views of people in those
  states
• Valuation methods based on cold calling, where
  the respondent has little time for reflection
  but should largely uninformed preferences be used
  in public policy?
• Argument for experiential values that is the
  values of those experiencing the state, but this
  raises concerns about the role of adaptation and
  context effects
• Third way a representative panel of the general
  population is better informed of what a state is
  like prior to valuation exercise

19
Conclusions from SF-6D work

• A preference-based measures has been successfully
  developed from the SF-36 based on 1)
  re-designing the descriptions 2) values obtained
  from a sample of the population 3) statistical
  inference to predict values
• SF-6D is being widely used in economic evaluation
  in health care
• Limited information loss of moving to index (and
  in any case retain the SF-36/SF-12)
• Main concern with the SF-6D is the floor in the
  descriptive system but this can be improved
  using results from item pools
• Modelling and design of valuation work can be
  improved using Bayesian methods
• Source of values an important question for this
  workshop