Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer
description
Transcript of Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer
![Page 1: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/1.jpg)
Multiple Indicator Growth Models aka, 2nd Order Growth
Daniel E BontempoScott M. Hofer
![Page 2: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/2.jpg)
Friday Harbor Psychometrics Workshop 2010
Acknowledgements
• Funded in part by Grant R13AG030995-01A1 from the National Institute on Aging
• The views expressed in written conference materials or publications and by speakers and moderators do not necessarily reflect the official policies of the Department of Health and Human Services; nor does mention by trade names, commercial practices, or organizations imply endorsement by the U.S. Government.
![Page 3: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/3.jpg)
Modeling Processes of Change
• As with Growth Modeling (GM) and Latent Growth Modeling (LGM), the goal remains to model overall pattern of within-person change, between person differences in change trajectories, and covariates that predict these fixed patterns and individual differences
• This goal is facilitated by:– Reliability of measurements– Magnitude of systematic within-person variation– Number of measurement occasions
![Page 4: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/4.jpg)
Multiple Indicator Growth Models• Reliability is greater when growth
modeling is based on an explicit measurement model that attenuated for measurement error
• Separation of reliable time-specific variance and measurement error variance
• Increased power for growth model parameters (i.e., smaller standard errors)
![Page 5: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/5.jpg)
Factor Models vs Latent Growth Modelsvs 2nd Order Latent Growth
• Each model has two latent factors
• Growth model has fewer indicators because the latent growth factors cross load
• In the factor model the loadings are estimated to get the best factor
• In the LGB the loadings ares specified to get a GF corresponding to a specific pattern f growth
![Page 6: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/6.jpg)
2nd Order Factor vs 2nd Order Growth
• The only difference is in the loadings at the 2nd level
![Page 7: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/7.jpg)
Curve-of-Factors
0
mem1
m1
apple1
0, u1
e1m2
address1
0, u2
e2m3
newold1
0, u3
e3
b1
1
b2
1
b3
1
0
mem2
m4
apple2
0, u4
e4m5
address2
0, u5
e5m6
newold2
0, u6
e6
b4
1
b5
1
b6
1
0
mem3
m7
apple3
0, u7
e7m8
address3
0, u8
e8m9
newold3
0, u9
e9
b7
1
b8
1
b9
1
0
L_m
0
S_m
mean_Lm, var_LmvLm
mean_Sm, var_Sm
vSm
1 1
0, v_mem1
vmem1
0, v_mem2
vmem2
0, v_mem3
vmem31 1 1
1 1 10 3.5 7
![Page 8: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/8.jpg)
But WAIT! - Is the indicator reflective?The Nature and Direction of Relationships Between Constructs and Measures: Reflective vs Formative
R
F
MEdwards, J. R., & Bagozzi, R. P. (2000). On the nature and direction of relationships between constructs and measures. Psychological Methods, 5(2), 155-174.
![Page 9: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/9.jpg)
R-F Importance
• One vs multiple error terms– Error in latent construct
• Misspecification can greatly bias estimation of structural parameters– MacKenzie, S. B., Podsakoff, P.
M., & Jarvis, C. B. (2005). The Problem of Measurement Model Misspecification in Behavioral and Organizational Research and Some Recommended Solutions. Journal of Applied Psychology, 90(4), 710-730.
![Page 10: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/10.jpg)
R-F in Practice• In 1990, a survey of 15 articles using latent
variables published in 1985-1986 in key social science journals, found constructs such as stressful life events, life change, illness, and instrumental activity of daily living inappropriately specified as reflective constructs.
– Cohen, P., Cohen, J., Teresi, J., Marchi, M. L., & et al. (1990). Problems in the measurement of latent variables in structural equations causal models. Applied Psychological Measurement, 14(2), 183-196.
• Still in 2005, another study finds important constructs in organizational and behavior research to be commonly mis-specified as reflective.
– MacKenzie, S. B., Podsakoff, P. M., & Jarvis, C. B. (2005). The Problem of Measurement Model Misspecification in Behavioral and Organizational Research and Some Recommended Solutions. Journal of Applied Psychology, 90(4), 710-730.
![Page 11: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/11.jpg)
Example: Activity-Limiting Illness
• Formative– Presence/severity of
cerebrovascular disease
– Presence/severity of cardiovasuclar disease
– Presence/severity of muscular-skeletal disease
– Presence/severity of cancer & immune system related problems
• Reflective (functional)– Pain severity/persistence– Energy level– Fatigue-proneness– Activity limitation
![Page 12: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/12.jpg)
Classical Test Theory
• T(rue score) is a hypothetical average of an infinite number of scores– Xi = T + ei
– Systematic errors in X become part of T
• ξ is not an average, but a latent variable that corresponds to a construct– X i = λiξ + δi
– Relation between ξ and X i can vary
– δi contains random error and systematic variance
![Page 13: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/13.jpg)
Items, Scores, Composites
![Page 14: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/14.jpg)
Items, Scores, Composites
![Page 15: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/15.jpg)
Optimizing Analysis of Change with Multiple Indicator Models
• Latent factor growth models will improve precision of growth model parameters by disattenuating for occasion-specific measurement error
• Requirements– Multiple indicators of a common factor at
each occasion• Item-level data• Alternative tests (i.e. scales)• Multiple methods/raters
– Evidence for Strong Factorial Invariance across occasions
– Other evidence of measurement equivalence for formative indicators
![Page 16: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/16.jpg)
Ordinal adds an “order” or two that may not count
• The manifest Xs have several latent y’s asociated with each X
• Not really a factor order
![Page 17: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/17.jpg)
Hierarchy of FI Tests
• Evidence for measurement invariance involves a nested sequence of increasingly stringent tests (Meredith,1993; Meredith & Horn, 2001).
• Configural (Baseline) – intercepts, loadings, & uniquenesses are freely estimated across groups, with minimal constraints only for identification
• Weak – constrain factor-variable regressions (i.e., loadings): factor variance in common metric
• Strong – constrain intercepts: factor means in common metric
• Strict – constrain uniquenesses (i.e., residuals): manifest scores comparable
![Page 18: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/18.jpg)
Factor Analytic Measurement Parameters
l
l
l
l
i
i
i
i
m, v
0, v
0, v
0, v
0, v
f1
l
l
l
l
i
i
i
i
m, v
0, v
0, v
0, v
0, v
f1
cov
l
l
l
l
i
i
i
i
m, v
0, v
0, v
0, v
0, v
f2
Group A
cov
l
l
l
l
i
i
i
i
m, v
0, v
0, v
0, v
0, v
f2
Group B
A regression like relationship with an intercept, slope, and residual.
InterceptsLoadingsUniquenesses
Invariance entails comparable parameters across groups.
![Page 19: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/19.jpg)
P5P4P3P2P1 P6 P8P7
Factor 1 Factor 2
1.0 a2 a3 a4 a6 a7 a81.0
Longitudinal Factorial Invariance
e8e7e5 e6e4e3e2e1
P5P4P3P2P1 P6 P8P7
Factor 1 Factor 2
1.0 a2 a3 a4 a6 a7 a81.0
e8e7e5 e6e4e3e2e1
Time 1 Time 2Mean
InterceptsMean
Intercepts
M1M2
0.00.0
![Page 20: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/20.jpg)
Weak Factorial Invariance
• Equality constraints on the factor-variable regressions (i.e., factor loadings) across groups while ensuring the factor variances and covariances are free to vary
• This condition is minimally necessary for establishing evidence for invariant measurement operations (Horn, 1991; Horn, McArdle, & Mason, 1983; Meredith, 1964; 1993)
![Page 21: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/21.jpg)
Strong Factorial Invariance
• Additional constraint of invariant mean intercepts across groups and fixed reference group factor mean
– Permits differences in means to be expressed at the factor level
– Factor mean differences across groups/time are expressed as differences relative a reference group or occasion
• Factor means for reference group/occasion are set to zero or some arbitrary value (Sorbom, 1974; see also Horn & McArdle, 1992)
![Page 22: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/22.jpg)
Strict Factorial Invariance
• Additional constraint of invariant unique variances as well as unique means and factor loadings
• Test of differences in magnitude of specific (i.e., unique systematic + error) variance– This model forces the combined specific and random error
components of corresponding variables to be equivalent across groups such that differences in variances across groups are permitted only at the latent variable level
![Page 23: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/23.jpg)
Evaluating Factorial Invariance• Configural
– Same pattern of factor loadings
• Weak (“Metric”) Factorial Invariance– Factor loadings equal
• Strong Factorial Invariance– Factor loadings equal– Mean intercepts equal
• Strict Factorial Invariance– Factor loadings equal– Mean intercepts equal– Unique variances equal
![Page 24: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/24.jpg)
LGC Factor-Level Specification
1. Strong or Strict Factorial Invariance established– Factor-variable regressions and variable intercepts constrained to
be equal, Reference occasion factor mean and variance fixed to 0.0 and 1.0, respectively
2. Factor-level covariances and means are fixed to zero
3. Metric for LGC parameters is identified by either of two approaches– Reference Occasion: LGC Level mean = 0; variance=1
• Corresponding manifest intercepts equal
– Manifest Variable: LGC Level mean and variance estimated• Fixed factor loading
• Reference variable intercept (all occasions) = 0 (provides factor metric)
• All other corresponding manifest intercepts estimated but equal
![Page 25: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/25.jpg)
Scaling of Growth Factors
• Ferrer, Balluerka, & Widaman (2009) Methodology
• Recommend an “approximately standardized metric” at first occasion of measurement– Change is expressed in Cohen’s d metric
• Standard deviation units based on T1 between-person differences
![Page 26: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/26.jpg)
Ferrer et al. Specification of the LGC model of factor outcomes
• Begin with the Strong or Strict FI model (standardized reference factor metric)
• Fix the mean of the LGC intercept to 0. • Fix the reference variable factor loading to the
value obtained in the strong or strict FI analysis. • Estimate the factor variances at each occasion. • Fix the factor means at each occasion to 0.
![Page 27: Multiple Indicator Growth Models aka, 2 nd Order Growth Daniel E Bontempo Scott M. Hofer](https://reader035.fdocuments.net/reader035/viewer/2022062517/568139c4550346895da16dae/html5/thumbnails/27.jpg)
Advantages of Modeling Factors in LGC
• Separates reliable common and specific variance from error variance
• Evaluation of measurement comparability over time • Reduces bias of time-specific systematic variance
and measurement error– Increases power for growth factor parameters (smaller
standard error)
– Disattenuates time-specific residuals for within-person analysis