measuring well-being - psychometric perspectives · overview 1 background 2 problems with single...

Measuring Well-BeingPsychometric Perspectives

Sam Norton

[email protected]

18 June 2013

Overview

1 Background

2 Problems with single items and sum scores

3 Generalised latent variable modellingPositive and negative aspects of subjective well-being

4 Hierarchical modelsUsing bifactor models to parse common and specific variance

5 Conclusion

S. Norton Measuring Well-Being 18 June 2013 2 / 36

What is well-being?

I Broad constructincorporating psychological,physical, social andmaterial/economiccomponents

I Poorly defined and typicallyoperationalised in a varietyof ways

I Focus on subjective andpsychological well-being


Public policy

I GDP and life satisfaction


Subjective and psychological well-being

I Subjective well-beingI Hedonic well-being – focussed on pleasure and affectI Ed Diener – Incorporates positive and negative affect, and life

satisfactionI Psychological well-being

I Based on Aristotelian concept of eudaimonic well-being – ‘thegood life’

I Carol Ryff – self-acceptance, personal growth, purpose in life,positive relations with others, environmental mastery andautonomy

I Martin Seligman – Authentic happiness model emphasisingfinding meaning in life

I Most health research focuses on the negative end ofsubjective well-being → negative affect


Issues with measuring subjective well-being

I Not directly observableI Positive or negative emphasisI Scope → very general / very specificI State vs traitI Evaluative or affectiveI Cultural differences


Issues to focus on today

I Problems with single items and sum scoresI Generalised latent variable modellingI Hierarchical models


Latent variables

I Observed variables used tomeasure something not directlyobservable

I Recovered from the commonvariance (co-variance) betweenitems

I Ubiquitous in statistics, not justpsychometrics


Problems with single items

I Perceived issues, compared to multi-item scalesI Poor reliabilityI Less validity

I More important issuesI Problems with scaling → restricted response categories,

skewed response distributionI Can’t partition response variance into common variance,

unique variance and measurement error

I Single items fine in many circumstances where a generale.g. life satisfaction


Problems with sum scores

I Still common practice to sum (or average) responses onmulti-item scales

I Increased reliability and validity compared to single itemsI Reliability is an issue for short scales, leads to redundant or

similar items being includedI Restrictive assumptions needed to allow comparison across

samples (classical test theory)I Appropriate to sum scores of items with binary/ordinal

response formats and treat this total score as continuous?


Two alternative methods

I Factor analysisI Items measured on a continuous scale (multivariate normal)I Factor loadingsI Allows for a more precise estimate of reliability than

Chronbach’s αI Item response theory

I Binary or ordinal response scaleI Item difficult and discrimination parametersI Information function provides a more sophisticated insight into

reliability than Factor analysis

I Not quite so different!


Measurement models

Confirmatory factor analysis= linear regression

F

y1

y2

y3

e1

e2

e3

λ1

λ2

λ3

Item response theory= logistic regression

F

y∗1

y∗2

y∗3

y1

y2

y3

e1

e2

e3

λ1

λ2

λ3


Generalised latent variable modelling

I Standard CFA and IRT models subsumed within a family ofgeneralised latent variable models

I Increasing number of stats packages available to fit thesekinds of models

I Mplus, GLLAMM, R, Mx, Factor . . . Stata v13


Positive and negative aspects of subjectivewell-being

I Most health research focuses on the negative end ofsubjective well-being → negative affect

I Evidence of independence of positive and negative affectI Comparison of two scales, assuming unidimensionality

I Hospital Anxiety and Depression Scale – 7-item depressionsubscale with 4-point ordinal response format

I Cardiovascular disease, N = 893I Chronbach’s α = .83

I Short Depression Happiness Questionnaire – 6-item scale with4-point ordinal response format

I General population, N = 1262I Chronbach’s α = .88


The problem with focussing on the negative

Two parameter logistic model fitted for each scale using Mplus

Item LoadingHADS2 0.760HADS4 0.795HADS6 0.625HADS8 0.414HADS10 0.576HADS12 0.722HADS14 0.418

Item LoadingSDHS1 0.855SDHS2 -0.884SDHS3 0.769SDHS4 -0.824SDHS5 -0.933SDHS6 0.805


Item characteristic curves

I Hospital Anxiety and Depression Scale - Depression subscale


Item characteristic curves

I Short Depression Happiness Questionnaire


Test information function

I Hospital Anxiety and Depression Scale - Depression subscale


Test information function

I Short Depression Happiness Questionnaire


Independent constructs

I So far have assumed positive and negative affect two ends ofa continuum

I Evidence of independence of positive and negative affectI Analysis of the SDHS reveals two factor positive–negative

structure provides ‘better fit’, however correlation betweenfactor is .86

I CFA useful for testing latent structure against theory


Hierarchical models

I Most CFA studies focus on separate correlated factorsI Often theoretically salient higher-order constructs, or high

inter-factor correlations indicating factors are not entirelyindependent

I Two approaches to modelling hierarchical structureI Higher-orderI Bifactor


Higher-order vs bifactor models

F2

F1

G

F2

F1

G


HADS meta CFA

I Norton et al (2013). Journal of Psychosomatic ResearchI Hospital Anxiety & Depression Scale, 14-items split across two

subscalesI Twenty-one studies, provided data from 28 unique samples

(N = 21, 820)I Compared ‘best fit’ latent structures previously identified plus

bifactorI Re-analysis and multivariate meta-analysis


HADS Meta CFA model fit

RMSEA CFI TLI BIC1. Unidimensional 0.089 0.863 0.838 7827712. Zigmond & Snaith 0.058 0.943 0.932 7749458. Higher-order 0.053 0.953 0.943 7739639. Bifactor 0.043 0.974 0.963 771965


Implications for mood disorders

Watson, D. (2005). Journal of abnormal psychology, 114(4), 522


DSM-V Reliability

Am J Psychiatry. 2013;170(1):1-5


DSM-V Reliability


Well-being in general

I So far have shown bifactor model useful applied to symptomsof anxiety and depression

I Also evidence of its applicability more widely to well-beingI Chen et al (2012) evidence of a general

psychological/subjective well-being factorI 795 university studentsI Examined bifactor models for subjective well-being,

psychological well-being and a combined well-being model


Chen et al (2012) subjective well-beingRMSEA = .054 (CI: .047 to .062), SRMR = .026, CFI = .974


Chen et al (2012) psychological well-beingRMSEA = .075 (CI: .071 to .079), SRMR = .045, CFI = .908


Chen et al (2012) combined well-beingRMSEA = .061 (CI: .058 to .063), SRMR = .051, CFI = .900


Conclusions

I Well-being a broad concept incorporating psychological,physical, social and material

I Single items are fine in some circumstancesI Avoid use of sum scores where possible, instead use factor

scoresI Tests do not have to be long → Computer adaptive testingI Generalised latent variable modelling offers advantages over

over traditional CFA → best of both worldsI Think carefully about what you want to measure (and how to

measure it)!


Thank you, any questions?


measuring well-being - psychometric perspectives · overview 1 background 2 problems with single...

Documents