Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
What Multi-Level Modeling Can Teach UsAbout Single-Level Modeling & Vice Versa:
The Case of Latent Transition Analysis
Bengt [email protected]
Mplus: www.statmodel.com
Mplus Web Talks: No. 1Recorded June 2020
This talk builds on the keynote address planned for the June 2020 ModernModeling Methods Conference at UCONN that was cancelled due to the coronavirus. I thank Tihomir Asparouhov and Ellen Hamaker for comments and Noah
Hastings for assistance.Bengt Muthen Mplus Web Talks: No. 1 1/ 80
Outline
Modeling:StatisticsPsychologyPsychometrics
Applications of LTA:Mood data (Eid & Langeheine, 2003)Dating and sexual risk behavior data (Lanza & Collins, 2008)Reading proficiency (Kaplan, 2008)Is LTA the right type of model?
Multi-level time series analysis (DSEM)
Appendix A: Sample size requirements
Appendix B: Computing time
Appendix C: New features in Mplus
Bengt Muthen Mplus Web Talks: No. 1 2/ 80
Hidden Markov - Latent Transition Analysis
u: observed categorical variable (latent class indicator)c: latent categorical variable (latent class variable)
u1 u3u2
c1 c2 c3
u4
c4
u5
c5
u11
c1 c2 c3
u12 u31 u32u21 u22
Bengt Muthen Mplus Web Talks: No. 1 3/ 80
LTA References
u11
c1 c2 c3
u12 u31 u32u21 u22
Langeheine (1988). New developments in latent class theory. InLangeheine & Rost (eds), Latent Trait and Latent Class Models
Graham et al. (1991). Modelling transitions in latent stage-sequentialprocesses: A substance use prevention example. Journal of Consultingand Clinical Psychology
Collins & Wugalter (1992). Latent class models for stage-sequentialdynamic latent variables. Multivariate Behavioral Research
Langeheine & van de Pol (2002). Latent Markov chains. In Hagenaars& McCutcheon (eds), Applied Latent Class Analysis
Collins & Lanza (2010). Latent Class and Latent Transition Analysis.
Bengt Muthen Mplus Web Talks: No. 1 4/ 80
An LTA Example: Dating Data
Lanza & Collins (2008). A new SAS procedure for latent transitionanalysis: Transitions in dating and sexual risk behavior.Developmental Psychology, 44, 446-456.
Data:
National Longitudinal Survey of Youth (NLSY97), N=29373 time points one year apart starting at ages 17-18
Items:
Past-year number of dating partners (0, 1, 2 or more)Past-year number of sexual partners (0, 1, 2 or more)Exposed to STD in past year (no, yes)
Bengt Muthen Mplus Web Talks: No. 1 5/ 80
Dating Data: Measurement Probabilities for 5 Classes
Past year Nondaters Daters Monogamous Multi-safe Multi-exposed
# dating partners0 0.789 0.008 0.096 0.031 0.0251 0.166 0.214 0.641 0.026 0.044
>=2 0.045 0.778 0.262 0.943 0.932# sex partners
0 0.975 0.948 0.000 0.104 0.0001 0.014 0.048 0.961 0.258 0.118
>=2 0.011 0.004 0.039 0.638 0.882Exposed to STD
No 1.000 1.000 0.385 1.000 0.187Yes 0.000 0.000 0.615 0.000 0.813
Bengt Muthen Mplus Web Talks: No. 1 6/ 80
Dating Data: Transition Probabilities
Past year Nondaters Daters Monogamous Multi-safe Multi-exposed
Nondaters 0.627 0.197 0.096 0.038 0.042Daters 0.023 0.626 0.173 0.086 0.091
Monogamous 0.034 0.032 0.679 0.056 0.199Multi-safe 0.036 0.000 0.177 0.584 0.203
Multi-exposed 0.021 0.033 0.201 0.055 0.690
Bengt Muthen Mplus Web Talks: No. 1 7/ 80
LTA Features
u11
c1 c2 c3
u12 u31 u32u21 u22
1 Measurement probabilities: P(Ut|Ct) - LCA for each time point,measurement invariance across time
2 Initial status probabilities: P(C1)
3 Transition probabilities: P(C2|C1), P(C3|C2)
Extensions:
Mover-Stayer modeling
Multiple-group analysis: Measurement invariance
Covariates: Influencing latent class probabilities and transitionprobabilities
Bengt Muthen Mplus Web Talks: No. 1 8/ 80
What’s Missing in These Models?
u1 u3u2
c1 c2 c3
u4
c4
u5
c5
u11
c1 c2 c3
u12 u31 u32u21 u22
Single indicator per time pointA statistical perspective
Multilevel modeling: Level 1 = time, level 2 = subjectRandom effects, especially random intercepts
A substantive perspectiveTrait theory in psychologyBetween-subject differences that are stable over time
Multiple indicators per time pointA psychometric perspective
Measurement non-invarianceMultilevel factor analysisMultilevel latent class analysis
Bengt Muthen Mplus Web Talks: No. 1 9/ 80
A Quick Reminder of Logistic Regression
1
0.5
0
P(Ui = 1|X1i,X2i) =1
1+ e−(β0+β1 X1i+β2 X2i). (1)
Logistic regression with a binary covariate X1 and a continuouscovariate X2. In factor analysis and IRT, X2 can be a factor and inLTA, X1 can be a latent class variable.
Bengt Muthen Mplus Web Talks: No. 1 10/ 80
Logistic Regression Continued
P(Ui = 1|X1i,X2i) =1
1+ e−(β0+β1 X1i+β2 X2i). (2)
Recalling that the odds is P(U = 1)/[P(U = 0)], this implies a linearequation for the log odds, also referred to as the logit:
logit[P(U = 1|X1,X2)] = β0 +β1 X1 +β2 X2. (3)
Interpretation of the influence of X on U in terms of odds ratios. Forinstance, the influence of X1 on U is eβ1 .
See also chapter 5 of the book:
Muthen, Muthen & Asparouhov (2016). Regression and MediationAnalysis Using Mplus.http://www.statmodel.com/Mplus_Book.shtml
Bengt Muthen Mplus Web Talks: No. 1 11/ 80
Measurement Probabilities as a Function of Latent Classes
Binary latent class indicator U, latent class variable C
P(U = 1|C = c) = 1/(1+ e−Lc), where L is the logit for class c
Lc = logit[P(U = 1|C = c)] = β0c =−τc (threshold)
High logit gives high probability, high τ gives low probability
L = 0 and τ = 0 give P = 0.5
L =−1 and τ = 1 give P = 0.27
L = 1 and τ =−1 give P = 0.73
How do you get L from P?
L = log odds = log[P(U = 1)/P(U = 0)]
Bengt Muthen Mplus Web Talks: No. 1 12/ 80
Longitudinal Data: Random Intercept Logistic Regression
Two representations of regression of a binary U on a covariate X attime t for subject j using a random intercept uj:
Logit [P(Utj = 1|Xtj)] = uj +β Xtj, (4)
uj = β0 + εj. V(εj) = σ2u (5)
Two-level, long format: Single-level, wide format:
x1 x2 x3
u
1 1 1
u1 u2 u3
β β β
u
Between (subject):Within (time):
xt
ut
β
Bengt Muthen Mplus Web Talks: No. 1 13/ 80
An Aside: Panel Data, Time-Invariant Omitted Variables,Fixed vs Random Effects
Logit [P(Utj = 1|Xtj)] = uj +β Xtj = β0 + εj (+γ Zj)+β Xtj.
x1 x2 x3
u
1 1 1
u1 u2 u3
β β β
Random effects modeling is criticizedfor not including the likely correlationbetween uj and Xtj. But this is resolvedby standard multilevel modeling:
Logit [P(Utj = 1|Xtj)] = uj +βw (Xtj−X.j),
uj = β0 +βb X.j +δj
Raudenbush & Bryk (2002). Hierarchical Linear Models. Sage. See,e.g., Table 5.11
Asparouhov & Muthen (2019). Latent variable centering of predictorsand mediators in multilevel and time-series models. StructuralEquation Modeling, 26, 119-142
Bengt Muthen Mplus Web Talks: No. 1 14/ 80
Panel Data Fixed vs Random Effects: References
Wooldridge (2002). Econometric Analysis of Cross Sectional andPanel Data. The MIT Press.
Allison (2009). Fixed effects regression models. Sage
Bollen & Brandt (2010). A general panel model with random and fixedeffects: A structural equations approach. Social Forces, 89, 1-34
Bell & Jones (2015). Explaining fixed effects: Random effectmodeling of time-series cross-sectional and panel data. PoliticalScience Research and Methods, 3 , 133-153
Bou & Satorra (2017). Univariate versus multivariate modeling ofpanel data: Model specification and goodness-of fit testing.Organizational Research Methods
Antonakis et al. (2019). On ignoring the random effects assumption inmultilevel models: Review, critique and recommendations.Organizational Research Methods
Hamaker & Muthen (2020). The fixed versus random effects debateand how it relates to centering in multilevel modeling. Psych Methods
Bengt Muthen Mplus Web Talks: No. 1 15/ 80
Single-Indicator LTA (Hidden Markov)
u
1 1 1
u1 u2 u3 u1 u2 u3
c1 c2 c3 c1 c2 c3
The Hidden Markov model on the left can use the fastbackward-forward Baum-Welch algorithm for ML estimation
The random intercept model on the right loses this simplicity - the U’sare no longer independent conditional on the Cs
ML estimation needs numerical integration with much heaviercomputationsBartolucci et al. (2012). Latent Markov Models for LongitudinalData
Bengt Muthen Mplus Web Talks: No. 1 16/ 80
Hidden Markov Modeling with a Random Intercept
u
1 1 1
u1 u2 u3
c1 c2 c3
Altman (2007). Mixed hidden Markov models. Journal of the AmericanStatistical Association, 102, 201-210.
Relapsing-remitting multiple sclerosis patients, N=39
Symptoms worsen and then improve in alternating periods ofrelapse and remission
Outcome: Number of lesions in the brain (count variable)
T=24 (monthly scans for two years)
2 latent classes (hidden states): Relapse versus remission
Bengt Muthen Mplus Web Talks: No. 1 17/ 80
What’s Missing?
Single indicator per time point
A statistical perspectiveMultilevel modeling: Level 1 = time, level 2 = subjectRandom effects, especially random intercepts
A substantive perspectiveTrait theory in psychologyBetween-subject differences that are stable over time
Multiple indicators per time point
A psychometric perspectiveMeasurement non-invarianceMultilevel factor analysisMultilevel latent class analysis
Bengt Muthen Mplus Web Talks: No. 1 18/ 80
Latent Trait-State Analysis
Kenny & Zautra (1995). The trait-state-error model for multiwavedata. Journal of Consulting and Clinical Psychology.
y1 y3y2
s1 s2 s3
y4
s4
y5
s5
t
t represents a continuous latent trait variable
y1-y5 represent continuous observed variables
s1-s5 represent continuous latent state variables
Bengt Muthen Mplus Web Talks: No. 1 19/ 80
Latent Trait-State Analysis Continued
Cole et al. (2005). Empirical and conceptual problems withlongitudinal trait-state models: Introducing a trait-state-occasionmodel. Psychological Methods
Kenny & Zautra (2001). Trait-state models for longitudinal data. In A.Sayer & L. M. Collins (Eds.), New methods for the analysis of change(pp. 243-263). Washington, DC: American Psychological Association.- STARTS model
Wagner at al. (2015). Self-esteem is mostly stable across youngadulthood: Evidence from latent STARTS models. Journal ofPersonality
Ludtke et al. (2018). More stable estimation of the STARTS model: ABayesian approach using Markov Chain Monte Carlo techniques.Psychological Methods
Bengt Muthen Mplus Web Talks: No. 1 20/ 80
Cross-Lagged Panel Modeling (CLPM)
z1
y1
z3
y3
z2
y2
Hamaker, Kuiper, Grasman (2015). A critique of the cross-laggedpanel model. Psychological Methods, 1, 102-116.
Bengt Muthen Mplus Web Talks: No. 1 21/ 80
Hamaker et al. (2015) Critique of CLPM
z1
y1
z3
y3
z2
y2
”if stability of constructs is to some extent of a trait-like, timeinvariantnature, the autoregressive relationships of the CLPM fail to adequatelyaccount for this”
”As a result, the lagged parameters that are obtained with the CLPM donot represent the actual within-person relationships over time”
”this may lead to erroneous conclusions regarding the presence,predominance, and sign of causal influences
an alternative model ... separates the within-person process fromstable between-person differences through the inclusion of randomintercepts”
Bengt Muthen Mplus Web Talks: No. 1 22/ 80
Random Intercept Cross-Lagged Panel Model (RI-CLPM)
z1
y1
wy1
wz1
y
z
z3
y3
wy3
wz3
z2
y2
wy2
wz2
Hamaker et al. (2015)Mplus scripts:http://www.statmodel.com/RI-CLPM.shtml
Bengt Muthen Mplus Web Talks: No. 1 23/ 80
What’s Missing? Multiple-Indicator Case
Single indicator per time point
A statistical perspectiveMultilevel modeling: Level 1 = time, level 2 = subjectRandom effects, especially random intercepts
A substantive perspectiveTrait theory in psychologyBetween-subject differences that are stable over time
Multiple indicators per time pointA psychometric perspective
Measurement non-invarianceMultilevel factor analysisMultilevel latent class analysis
Bengt Muthen Mplus Web Talks: No. 1 24/ 80
Multiple Indicators in Cross-Sectional Data:Capturing Response Bias in Self-Report Questionnaires
Individual differences in response styles e.g. due to acquiescence orsocial desirability:
Wolf (1940). Factor analysis to 1940. Psychometric Monograph No. 3
Individual differences in intercepts and loadings
Maydeu-Olivares & Coffman (2006). Random intercept factoranalysis. Psychological Methods
Individual differences captured by an added response style factorwith factor loadings fixed at 1, that is, a random intercept:Different individuals are allowed to have different intercept values
Schmalbach et al. (2020). From bi-dimensionality touni-dimensionality in self-report questionnaires. European Journal ofPsychological Assessment.
Applications to established questionnaires for depression,procrastination, life orientation, self-esteem, etc
Bengt Muthen Mplus Web Talks: No. 1 25/ 80
Random Intercept Factor Analysis
1 1 11 1
u
u uu3 4 5u2u1
f
λ λ λ λ λ54321
Random intercept factor analysis for a binary factor indicator Ur(r = 1, . . . ,p), individual i, and factor f :
logit[P(Uri = 1|fi)] =νr +ui +λr fi, (6)
ui ∼ N(0,σ2), (7)
where ui is a random intercept factor with loadings fixed at 1 for all thefactor indicators. A single added parameter: σ2.
Bengt Muthen Mplus Web Talks: No. 1 26/ 80
Random Intercept Factor Analysis Conclusions
Measurement non-invariance captured by a random intercept:Allows better estimation of the model of interestSuggests that measurement non-invariance is ok if modeled
But is the model realistic?Different indicators may have different random intercept variance
We need a design where each indicator is multiply measuredMultilevel data such as different students nested in differentschools so that an intercept is allowed to vary across schoolsLongitudinal data where time is nested within individual so thatan intercept for each indicator is allowed to vary acrossindividuals
Bengt Muthen Mplus Web Talks: No. 1 27/ 80
Multiple Indicators in Cross-Sectional Data:Multilevel Factor Analysis
Within
Between
fb
u1 u2 u3
fw
Students within schools:
Typically more within-level factors than between-level factorsfb interpretation different from fw interpretation
Bengt Muthen Mplus Web Talks: No. 1 28/ 80
Multilevel Factor Analysis Origins
Yij = ν +YBj +YWij (8)
= ν +ΛB fBj + εBj︸ ︷︷ ︸+ΛW fWij + εWij︸ ︷︷ ︸, (9)
with covariance structure V(Yij) = ΣB +ΣW where each covariancematrix has a factor model structure, Λ Ψ Λ′+Θ.
Harnqvist (1978). Primary mental abilities of collective and individuallevels. Journal of Educational Psychology
Goldstein & McDonald (1988). A general model for the analysis ofmultilevel data. Psychometrika
McDonald & Goldstein (1989). Balanced versus unbalanced designsfor linear structural relations in two-level data. British Journal ofMathematical and Statistical Psychology
Muthen (1994). Multilevel covariance structure analysis. SociologicalMethods & Research
Bengt Muthen Mplus Web Talks: No. 1 29/ 80
Multilevel Factor Analysis: The Random Intercepts Aspect
For each binary Uij factor indicator, the model
logit[P(Uij = 1|.)] = ν +λB fBj + εBj +λW fWij (10)
can be expressed by the two equations (level-1 and level-2):
logit[P(Uij = 1|.)] = uj +λW fWij , (11)
uj = ν +λB fBj + εBj, (12)
This is in line with two-level regression where the intercept uj israndom, varying across Level-2 units (schools)
Bengt Muthen Mplus Web Talks: No. 1 30/ 80
Random Intercept as Measurement Non-Invariance
logit[P(Uij = 1|.)] = uj +λW fWij , (13)
uj = ν +λB fBj + εBj, (14)
Can be interpreted as measurement non-invariance with respect to theintercept - expressed as a random effect:
Continuous outcomes: Jak et al. (2013; 2014) in SEM, Muthen &Asparouhov in SM&RCategorical outcomes: Fox (2005) BJMSP, Fox & Glas (2001)Psychometrika
Measurement non-invariance ok because you model it
Bengt Muthen Mplus Web Talks: No. 1 31/ 80
Random Intercept for the Factor: Multilevel IRT
Fox (2010). Bayesian Item Response Theory
Fox & Verhagen (2011). Random item effects modeling forcross-national survey data
logit[P(Uij = 1|.)] = uj +λW fWij , (15)
uj = ν +λB fBj + εBj, (16)
Multilevel IRT assumes λB = λW , εB = 0:
logit[P(Uij = 1|.)] = uj +λ fWij , (17)
uj = ν +λ fBj +0, (18)
I.e. a random intercept for the factor f , not a separate one for each U:
logit[P(Uij = 1|.)] = ν +λ fij, (19)
fij = fBj + fWij . (20)
Bengt Muthen Mplus Web Talks: No. 1 32/ 80
Two Types of Multilevel Factor Models
fw
u1 u2 u3 u4 u5
(a) (b)
Between
u2 u3u1 u4
fb
u5
f
u1 u2 u3 u4 u5
Within
Between
Within
u2 u3u1 u4
fb
u5
Bengt Muthen Mplus Web Talks: No. 1 33/ 80
Multilevel Factor Analysis versus Latent Class Analysis
c
u1 u2 u3 u4 u5
u2 u3u1 u4
f
u5
fw
u1 u2 u3 u4 u5
u2 u3u1 u4
fb
u5
(a)
Within
Between
c#1 c#2
u1
c
u2 u3 u4 u5
(b)
(c)
c#1 c#2
u1
c
u2 u3 u4 u5
(b)
Within
Between
Within
Between
Within
Between
Bengt Muthen Mplus Web Talks: No. 1 34/ 80
Two Approaches to Multilevel Latent Class Analysis
Vermunt (2003, 2008), Asparouhov-Muthen (2008), Henry-Muthen(2010)
c
u1 u2 u3 u4 u5
u2 u3u1 u4
f
u5
fw
u1 u2 u3 u4 u5
u2 u3u1 u4
fb
u5
(a)
Within
Between
c#1 c#2
u1
c
u2 u3 u4 u5
(b)
(c)
c#1 c#2
u1
c
u2 u3 u4 u5
(b)
Within
Between
Within
Between
Within
Between
Bengt Muthen Mplus Web Talks: No. 1 35/ 80
Multilevel Latent Class Analysis: An Application
Henry & Muthen (2010). Multilevel latent class analysis: Anapplication of adolescent smoking typologies with individual andcontextual predictors. Structural Equation Modeling
Smoking behavior for 10,772 9th grade females in 206 ruralcommunities
Six categorical indicators of three latent classes of student smokingbehavior
Random intercepts for the indicators
Across-community variation in the random intercepts found for severalindicators. Variation related to proportion of youth living in poverty
E.g. “Most friends are smokers” had a much larger probability of beingendorsed in communities with a large poverty proportion
Bengt Muthen Mplus Web Talks: No. 1 36/ 80
Going Deeper into Multilevel Factor Analysis
Within
Between
fb
u1 u2 u3
fw
Cross-sectional: Students within schools
But the case of time points within individuals is also multilevel - sowhy not apply this model to longitudinal data?
Bengt Muthen Mplus Web Talks: No. 1 37/ 80
What Multilevel Factor Modeling Can Teach Us:Longitudinal Factor Model for T=2
==
=u11 u12 u13 u u u21 22 23
fw1 fw2
fb
u1 u2 u3
Within
Between
fb
u1 u2 u3
fw
Bengt Muthen Mplus Web Talks: No. 1 38/ 80
Longitudinal Factor AnalysisBased on Multilevel Thinking
==
=u11 u12 u13 u u u21 22 23
fw1 fw2
fb
u1 u2 u3
Model extended to include auto-regression for the fw factors
Corresponds to a cross-sectional model where students within schoolsinfluence each other
This multilevel-based longitudinal model is quite different from the”naive” longitudinal model which uses only the bottom part
Bengt Muthen Mplus Web Talks: No. 1 39/ 80
Longitudinal Factor Analysis
Hamaker et al. (2016). Using a few snapshots to distinguish mountainsfrom waves: Weak factorial invariance in the context of trait-stateresearch. Multivariate Behavioral Research. - Based on multilevel FA
Argues that factors obtained from cross-sectional factor analysis arepartly determined by between- and partly by within-person covariancestructure (Cattell, 1978): “An uninterpretable blend” (R & B, 2002)
Multilevel-based longitudinal factor analysis allows differencesbetween between-person, trait-like factor structure and within-person,state-like factor structure
T=2 is sufficient to determine the within-person factor structure. Noneed for intensive longitudinal data to estimate a factor model for eachperson (N=1 analysis; Cattell’s P-technique)
Relates to work in Joreskog (1970; 1979), Raffalovich-Bohrnstedt(1987) in SM&R, Marsh-Grayson (1994) in SEM, Eid (1996) inMPRO, Dumenci-Windle (1996) in MBR, Geiser-Lockhart (2012) inPsychological Methods, Geiser et al. (2015) in Behav Res
Bengt Muthen Mplus Web Talks: No. 1 40/ 80
A Longitudinal Factor Analysis Application
Depression in adolescents (Dumenci & Windle, 1996); used inHamaker et al. (2016)
Four waves of data for 372 males and 433 females
Four subscales of the CES-D: Depressed affect, positive affect, somaticcomplaints, interpersonal. One factor on both “levels”
BIC for the naive model model versus the multilevel-based model, bothwith AR(1):
Males: 26123 versus 25465Females: 32216 versus 31423
The naive model gives inflated reliabilities (standardized loadings)relative to fw and inflated AR
Are we interested in the reliability wrt to fw or fb or both or thecombination?
Bengt Muthen Mplus Web Talks: No. 1 41/ 80
What Does Longitudinal Factor Analysis Imply forCross-Sectional Factor Analysis (and LCA)?
Generate data from a longitudinal factor analysis model
Analyze one time point using regular factor analysis
For example:Using data generated by 1 between factor with different loadingsacross indicators and 1 within factor, regular factor analysis maymistakenly settle on 2 factorsBut the model that the data were generated from doesn’t havevariation across individuals in 2 factors - but variation in 1 factoracross individuals and variation in 1 factor across time
Non-aggregatable model
Similar situation in cross-sectional multilevel settings:Muthen (1989). Latent variable modeling in heterogeneouspopulations. Psychometrika, 54:4, 557-585Muthen & Satorra (1995). Complex sample data in structuralequation modeling. Sociological Methodology, 25, 267-316
Bengt Muthen Mplus Web Talks: No. 1 42/ 80
Conclusions from the Modeling Part of the Talk
Modeling Lessons Learned
Statistics:Random intercept
Psychology:Latent trait
Psychometrics:Measurement Non-Invariance
Bengt Muthen Mplus Web Talks: No. 1 43/ 80
Back to Multiple-Indicator Latent Transition Analysis
u11
c1 c2 c3
u12 u31 u32u21 u22
1 Initial status probabilities: P(C1)
2 Transition probabilities: P(C2|C1), P(C3|C2)
3 Measurement probabilities: P(Ut|Ct) - LCA for each time point,measurement invariance across time
- Now it is clear what is missing
Bengt Muthen Mplus Web Talks: No. 1 44/ 80
Random Intercept LTA (RI-LTA)
u11
c1 c2 c3
u12 u31 u32u21 u22
f
λ1 λ1 λ1 λ2λ2λ2
Muthen & Asparouhov (2020). Random intercept latent transitionanalysis (RI-LTA). Under review in Psychological Methods.
Papers and Mplus scripts:http://www.statmodel.com/RI-LTA.shtml
Bengt Muthen Mplus Web Talks: No. 1 45/ 80
Features of RI-LTA
u11
c1 c2 c3
u12 u31 u32u21 u22
f
λ1 λ1 λ1 λ2λ2λ2
Fits the data much better than regular LTA which is unnecessarilyrestrictive and gives distorted results
Does not confound between- and within-subject sources of variation:Ct−1 −> Ct represents a within-subject process free of time-invariantbetween-subject differences (trait differences; meas. non-invariance)
Latent class indicators correlate over time beyond what is captured bythe latent class correlations over time:
Tends to reduce the probability of subjects staying in the sameclass as compared to regular LTAReduces the need for Mover-Stayer modeling because Moversand Stayers can be captured by different random intercept values
Bengt Muthen Mplus Web Talks: No. 1 46/ 80
Software Improvements for LTA and RI-LTA
RI-LTA can be time consuming due to numerical integration andneeding many random starts to find the global maximum.Mplus Version 8.4 released last November:
Significant speed improvements for computationally demandingmixture models such as with LTA and RI-LTA using a new three-stagerandom starts search and using specialized algorithms drawing onBaum-Welch ideas
Asparouhov & Muthen (2019). Random Starting Values andMultistage Optimization (Technical Report: http://www.statmodel.com/download/StartsUpdate.pdf)
”A 20 hours computation in Mplus 8.3 can be done in Mplus 8.4in less than 15 minutes, by utilizing the advantages of thethree-stage estimation, the Baum-Welch algorithm, as well asupdated hardware (i9-9900k Intel CPU)”
Substantially simplified output for mixture models with multiple latentclass variables
Bengt Muthen Mplus Web Talks: No. 1 47/ 80
An Aside: Different Types of Random Effects
Continuous - normal. Main model used in Muthen-Asparouhov (2020)
Discrete - non-parametric (one extra latent class variable):Latent trait related:
Eid & Langeheine (1999). The measurement of consistency andoccasion specificity with latent class models: A new model and itsapplicaton to the measurement of affect. Psychological MethodsEid & Langeheine (2003). Separating stable from variableindividuals in longitudinal studies by mixture distribution models.Measurement: Interdisciplinary Research and Perspectives
Random effects related for hidden Markov:Maruotti (2011). Mixed hidden Markov models for longitudinaldata: An overview. International Statistical ReviewBartolucci et al (2012). Latent Markov Models for LongitudinalData
The discrete alternative using a binary random effect variable ispresented in Muthen-Asparouhov (2020) as based on theEid-Langeheine parameterization. It is somewhat less time consumingbut gave a worse BIC than the continuous-normal model in allapplications. More than 2 categories was not needed
Bengt Muthen Mplus Web Talks: No. 1 48/ 80
Outline
Modeling:StatisticsPsychologyPsychometrics
Applications of LTA:Mood data (Eid & Langeheine, 2003)Dating and sexual risk behavior data (Lanza & Collins, 2008)Reading proficiency (Kaplan, 2008)Is LTA the right type of model?
Multi-level time series analysis (DSEM)
Appendix A: Sample size requirements
Appendix B: Computing time
Appendix C: New features in Mplus
Bengt Muthen Mplus Web Talks: No. 1 49/ 80
Regular LTA Fits Worse than RI-LTA Most of the Time
Analyses of 4 data sets from the LTA literature (data athttp://www.statmodel.com/RI-LTA.shtml):Table 1: Model fitting results
Life Satisfaction (non-stationary) Mood (stationary)N=5147, T=5, R=1, J=5 N=494, T=4, R=2, J=2
Model # par’s logL BIC Model # par’s logL BIC
Regular LTA 11 -15326 30745 Regular LTA 7 -2053 4150
RI-LTA 12 -15267 30637 RI-LTA 9 -2018 4093
Reading proficiency (non-stationary) Dating and sexual risk behavior (stationary)N=3574, T=4, R=5, J=3 N=2933, T=3, R=3, J=5
Regular LTA 35 -21793 43873 Regular LTA 49 -16202 32796
RI-LTA 40 -20329 40985 RI-LTA 52 -16043 32502
1
Bengt Muthen Mplus Web Talks: No. 1 50/ 80
Mood Data. Eid & Langeheine (2003)
German Multidimensional Mood Questionnaire data, N=494, ages17-77 (Steyer et al., 1997)
4 time points 3 weeks apart
2 binary items based on ratings of momentary sadness and unhappiness
Stationary model (equal transition probabilities over time)
Eid & Langeheine (2003). Separating stable from variable individualsin longitudinal studies by mixture distribution models. Measurement:Interdisciplinary Research and Perspectives
Bengt Muthen Mplus Web Talks: No. 1 51/ 80
Model Fit for Mood Data
Model # parameters log likelihood BIC
Standard
1 Regular LTA 7 -2053 41502 RI-LTA 9 -2018 4093
Bengt Muthen Mplus Web Talks: No. 1 52/ 80
Mood Data Measurement Probability Estimates
Classes
Regular LTA
High mood Low moodIndicator
Sad 0.089 0.902Unhappy 0.038 0.857
RI-LTA
High mood Low moodIndicator
Sad 0.286 0.748Unhappy 0.164 0.804
Bengt Muthen Mplus Web Talks: No. 1 53/ 80
Mood Data Transition Probability Estimates
2 latent classes: High mood and Low mood
Classes
Regular LTA
Classes High mood Low moodHigh mood 0.803 0.197Low mood 0.248 0.752
RI-LTA
Classes High mood Low moodHigh mood 0.691 0.309Low mood 0.486 0.514
Bengt Muthen Mplus Web Talks: No. 1 54/ 80
Model Fit for Mood Data: Mover-Stayer Modeling
Model # parameters log likelihood BIC
Standard
1 Regular LTA 7 -2053 41502 RI-LTA 9 -2018 4093
Mover-Stayer
3 Regular LTA 8 -2037 41234 RI-LTA 10 -2017 4096
Unlike regular LTA, RI-LTA does not need a Mover-Stayercomponent
Large negative and positive intercept factor scores captureStayers
Bengt Muthen Mplus Web Talks: No. 1 55/ 80
Transitions in Dating and Sexual Risk BehaviorLanza and Collins (2008) Developmental Psychology
Data:
National Longitudinal Survey of Youth (NLSY97), N=29373 time points one year apart starting at ages 17-18
Items:
Past-year number of dating partners (0, 1, 2 or more)Past-year number of sexual partners (0, 1, 2 or more)Exposed to STD in past year (no, yes)
Bengt Muthen Mplus Web Talks: No. 1 56/ 80
Model Results for Dating Data
Regular LTA (# pars = 49, logL = -16202, BIC = 32796):5 classes: Nondaters, Daters, Monogamous, Multipartner-Safe,Multipartner-Exposed
RI-LTA (# pars = 52, logL = -16043, BIC = 32502):5 classes but only Nondaters, Daters, and Multipartner-Exposedcorrespond to the solution obtained with regular LTATwo different classes: (1) Monogamous wrt sex partner but withmultiple dating partners, (2) Monogamous-ExposedTransition probabilities:
Past year Nondaters Daters Monogamous? Mono-exposed Multi-exposed
Nondaters 0.583 0.264 0.029 0.058 0.067Daters 0.014 0.660 0.183 0.013 0.130
Monogamous? 0.008 0.067 0.405 0.200 0.320Mono-exposed 0.071 0.007 0.040 0.639 0.243Multi-exposed 0.031 0.050 0.172 0.127 0.619
Bengt Muthen Mplus Web Talks: No. 1 57/ 80
Model Fit for Dating DataAdding a Mover-Stayer Component
Table 1: Model fitting results for Dating data
Model # par’s logL BIC
Standard
Regular LTA 49 -16202 32796RI-LTA 52 -16043 32502
Mover-Stayer
Regular LTA 50 -16194 32787RI-LTA 53 -16041 32506
1
Bengt Muthen Mplus Web Talks: No. 1 58/ 80
Dating Data Estimates
RI-LTA allows a continuum of Staying instead of a Stayer class:
RI-LTA model captures staying also via different intercept factor values
RI-LTA does not need a Stayer class or the Stayer class is smaller
Regular LTA Mover-Stayer model estimates 20% StayersRI-LTA Mover-Stayer model estimates 9% Stayers
RI-LTA makes it possible to learn more about the indicators:
Estimated RI-LTA intercept factor loadings (all significant):
# dating partners = 1.55# sex partners = 4.22 (strongest indicator of the random interceptfactor, that is, the time-invariant trait)exposed to STD = 1.51
Bengt Muthen Mplus Web Talks: No. 1 59/ 80
Reading Proficiency. Kaplan (2008)
Early Childhood Longitudinal Study (ECLS), N = 3574
4 time points: Fall and Spring of Kindergarten and Fall and Spring ofGrade 1
5 binary items representing a stage-sequential process:
Basic reading skills of letter recognitionBeginning soundsEnding letter soundsSight wordsWords in context
3 latent classes corresponding to 3 stages of learning
Kaplan (2008). An overview of Markov chain methods for the study ofstage-sequential developmental processes. Developmental Psychology
Bengt Muthen Mplus Web Talks: No. 1 60/ 80
Model Fit for Reading Data
Model # parameters log likelihood BIC
Standard
Regular LTA 35 -21793 43873RI-LTA 40 -20329 40984
Bengt Muthen Mplus Web Talks: No. 1 61/ 80
Reading Data Measurement Probability Estimates
Class 1 = low alphabet knowledge, Class 2 = early word reading,Class 3 = early reading comprehension
Regular LTA RI-LTA
Classes Classes1 2 3 1 2 3
Letrec 0.505 0.994 1.000 0.627 0.939 0.979Begin 0.066 0.917 0.984 0.303 0.806 0.941Ending 0.013 0.661 0.972 0.167 0.630 0.904Sight 0.000 0.051 0.985 0.020 0.208 0.808WIC 0.000 0.001 0.509 0.005 0.058 0.460
Bengt Muthen Mplus Web Talks: No. 1 62/ 80
Reading Data Latent Class Probabilities
Regular LTA
Fall K Spring K Fall 1st Spring 1st
ClassProbabilities
1 0.694 0.235 0.142 0.0412 0.284 0.635 0.627 0.1543 0.023 0.130 0.232 0.805
RI-LTA
Fall K Spring K Fall 1st Spring 1st
ClassProbabilities
1 0.948 0.161 0.040 0.0102 0.049 0.818 0.880 0.0173 0.003 0.022 0.080 0.973
Bengt Muthen Mplus Web Talks: No. 1 63/ 80
Reading Data Transition Probabilities
Regular LTA
1 2 3 1 2 3
TransitionProbabilities Fall K -– Spring K Spring K -– Fall 1st
1 0.338 0.649 0.012 0.596 0.401 0.0022 0.001 0.652 0.348 0.002 0.837 0.1613 0.000 0.000 1.000 0.002 0.003 0.994
RI-LTA
1 2 3 1 2 3
TransitionProbabilities Fall K -– Spring K Spring K -– Fall 1st
1 0.170 0.820 0.010 0.240 0.742 0.0182 0.000 0.819 0.181 0.001 0.931 0.0683 0.000 0.000 1.000 0.022 0.000 0.978
Bengt Muthen Mplus Web Talks: No. 1 64/ 80
The Meaning of the Random Intercept Factor:The Measurement Non-Invariance Perspective
u11
c1 c2 c3
u12 u31 u32u21 u22
f
λ1 λ1 λ1 λ2λ2λ2
For binary latent class indicator Ur, individual i, at time t, and class c,
logit[P(Urit = 1|Cit = c, fi) =−τrc +λr fi, (21)
for threshold τ and random intercept factor loading λ . A high logitvalue implies a high probability of U = 1, e.g. λ > 0, fi > 0.
In the presentation of measurement probabilities, the random interceptfactor f was integrated out, that is, considering the overall conditionalprobability P(Urit = 1|Cit = c)
Individuals with different fi values have different conditionalprobabilities P(Urit = 1|Cit = c, fi) as shown in (21)
This means that the random intercept factor f can be seen as capturingmeasurement non-invariance
Bengt Muthen Mplus Web Talks: No. 1 65/ 80
Measurement Probabilities for the Reading Datafrom a Measurement Non-Invariance Perspective
Regular LTA RI-LTA
Classes Classes1 2 3 1 2 3
Letrec 0.505 0.994 1.000 0.627 0.940 0.979Begin 0.066 0.917 0.984 0.303 0.806 0.941Ending 0.013 0.661 0.972 0.167 0.631 0.904Sight 0.000 0.051 0.985 0.021 0.209 0.808WIC 0.000 0.001 0.509 0.005 0.059 0.460
RI-LTA
f =−0.5 f =+0.5
Letrec 0.384 0.987 0.998 0.954 1.000 1.000Begin 0.045 0.809 0.977 0.425 0.985 0.999Ending 0.013 0.439 0.942 0.151 0.911 0.995Sight 0.000 0.005 0.850 0.001 0.185 0.996WIC 0.000 0.000 0.091 0.001 0.010 0.811
Bengt Muthen Mplus Web Talks: No. 1 66/ 80
Exploring Measurement Non-Invariance in the Reading Data
RI-LTA
f =−0.5 f =+0.5
Letrec 0.384 0.987 0.998 0.954 1.000 1.000Begin 0.045 0.809 0.977 0.425 0.985 0.999Ending 0.013 0.439 0.942 0.151 0.911 0.995Sight 0.000 0.005 0.850 0.001 0.185 0.996WIC 0.000 0.000 0.091 0.001 0.010 0.811
RI-LTA, allowing measurement non-invariance across poverty groups
Poverty (19%) Non-poverty (81%)
Letrec 0.371 0.880 0.977 0.692 0.969 0.990Begin 0.116 0.641 0.911 0.343 0.858 0.961Ending 0.052 0.421 0.840 0.187 0.687 0.932Sight 0.003 0.082 0.653 0.019 0.230 0.857WIC 0.000 0.011 0.237 0.003 0.062 0.512
The random intercept factor f can perhaps be viewed as a readingpreparedness dimension. It is strongly related to poverty
Bengt Muthen Mplus Web Talks: No. 1 67/ 80
RI-LTA Model Variations
Lag-2, lag-3 etc for the latent class variables
Distal outcomes
Time trend
Multiple processes (cross-lagged RI-LTA)
Multilevel RI-LTA
Bengt Muthen Mplus Web Talks: No. 1 68/ 80
Is LTA the Right Type of Model? Classes vs Factors
Latent classes or latent dimensions?
Latent transition analysis versus Longitudinal factor analysis (LFA)
LFA has many factors - high-dimensional integration for ML due tocategorical outcomes but easy for Bayes
LFA number of dimensions = T*(number of within factors) +number of of indicators/time point (one between factor)
Mood: 4*1 + 2 = 6, Dating: 3*1 + 3 = 6, Reading: 4*1 + 5 = 9
==
=u11 u12 u13 u u u21 22 23
fw1 fw2
fb
u1 u2 u3
Bengt Muthen Mplus Web Talks: No. 1 69/ 80
Is LTA the Right Type of Model: RI-LTA vs LFA Using ML
Mood data:
RI-LTA: 9 par’s, logL =-2018, BIC = 4093LFA: 16 par’s, logL = -2007, BIC = 4113
Dating data:
RI-LTA: 52 par’s, logL = -16043, BIC = 32502LFA: 20 par’s, logL = -17278, BIC = 34715
Reading data:
RI-LTA: 40 par’s, logL = -20329, BIC = 40984LFA: 29 par’s, logL = -19939, BIC = 40115LFA indicator variances (random intercepts; measurementnon-invariance): Letrec = 14.8, Begin = 6.9, Ending = 6.0, Sight= 6.0, WIC = 16.1LFA factor means increase significantly, except between Spring Kand Fall 1st
Can we use a model that fits much worse than another model?
Bengt Muthen Mplus Web Talks: No. 1 70/ 80
Outline
Modeling:StatisticsPsychologyPsychometrics
Applications of LTA:Mood data (Eid & Langeheine, 2003)Dating and sexual risk behavior data (Lanza & Collins, 2008)Reading proficiency (Kaplan, 2008)Is LTA the right type of model?
Multi-level time series analysis (DSEM)Appendix A: Sample size requirements
Appendix B: Computing time
Appendix C: New features in Mplus
Bengt Muthen Mplus Web Talks: No. 1 71/ 80
The Vice Versa: What Single-Level Modeling Can Teach UsAbout Multi-Level Modeling
u11
c1 c2 c3
u12 u31 u32u21 u22
f
λ1 λ1 λ1 λ2λ2λ2
How do you do analyze relations between variables at different timepoints using two-level modeling?
Multilevel time series analysis (time and individual)Other random effects can be added: Random slopes, randomvariances, random AR, random transition probabilitiesMany time points required - intensive longitudinal data (EMA,ESM, daily diary)
Bengt Muthen Mplus Web Talks: No. 1 72/ 80
Dynamic Structural Equation Modeling (DSEM)
Single-level, wide Two-level, long, time series (DSEM)
u11
c1 c2 c3
u12 u31 u32u21 u22
f
λ1 λ1 λ1 λ2λ2λ2
Within (time)
u1 u2
f
λ λ
ct-1 ct
u1t u2t
Between (subject)
1 2
DSEM data are in long format:2 outcomes per time point results in 2 columns of data (not 2*T)
Across-time effects specified by using lags: C ON C&1 (lag 1)Bengt Muthen Mplus Web Talks: No. 1 73/ 80
Dynamic Factor Analysis
fb φ
Between
φfwtfwt-1
Within
Bengt Muthen Mplus Web Talks: No. 1 74/ 80
Example: Affective InstabilityIn Ecological Momentary Assessment
Jahng S., Wood, P. K.,& Trull, T. J., (2008). Analysis of AffectiveInstability in Ecological Momentary Assessment: Indices UsingSuccessive Difference and Group Comparison via MultilevelModeling. Psychological Methods, 13, 354-375
An example of the growing amount of EMA data
84 outpatient subjects: 46 meeting borderline personality disorder(BPD) and 38 meeting MDD or DYS
Each individual is measured several times a day for 4 weeks for total ofabout 100 assessments
A mood factor for each individual is measured with 21 self-ratedcontinuous items
The research question is if the BPD group demonstrates more temporalnegative mood instability than the MDD/DYS group
Bengt Muthen Mplus Web Talks: No. 1 75/ 80
Dynamic Structural Equation Modeling
Two-level analysis: randomeffects varying across subjectsCross-classified analysis (oftime and subject): randomeffects varying across subjectsand time
fb φ
Between
φfwtfwt-1
Within
Dynamic structural equation modeling (DSEM):
Asparouhov, Hamaker & Muthen (2017). Dynamic Latent ClassAnalysis. Structural Equation Modeling: A MultidisciplinaryJournal, 24:2, 257-269Asparouhov, Hamaker & Muthen (2018). Dynamic structuralequation models. Structural Equation Modeling: AMultidisciplinary Journal, 25:3, 359-388
Papers at http://www.statmodel.com/TimeSeries.shtml
Bengt Muthen Mplus Web Talks: No. 1 76/ 80
Appendix A: Sample Size Requirements
LTA typically needs large samples (> 1,000)
RI-LTA (limited Monte Carlo simulation study in Muthen-Asparouhov,2020; binary latent class indicators):
T ≥ 3: Good performance at N ≥ 1,000T = 2: N > 4,000 needed
RI-LTA needs larger samples than regular LTA (assuming eachmodel is correct)But if data have been generated by RI-LTA, RI-LTA estimates arebetter than regular LTA at sample size as low as N=500
DSEM (continuous outcomes):
Schultzberg-Muthen (2018). Number of subjects and time pointsneeded for multilevel time series analysis: A simulation study ofdynamic structural equation modeling. SEM journalT ≥ 10 for minimal models, otherwise T ≥ 20-50
Bengt Muthen Mplus Web Talks: No. 1 77/ 80
Appendix B: RI-LTA Computing time
Timings using different Intel CPUs (hours:minutes:seconds)
i7-7700k i7-7700k i9-9900k i9-9900kReg LTA RI-LTA RI-LTA RI-LTA
Proc=8 Proc=12
Mood (N=494, 00:00:04 00:00:12 - -T=4, 2 classes)
Dating (N=2937, 00:00:41 00:06:00 00:03:31 00:03:25T=3, 5 classes)
Dating (N=2937, 00:05:00 00:40:00 00:21:00 00:19:00T=3, 5 classes,adding 4 covariates)
Bengt Muthen Mplus Web Talks: No. 1 78/ 80
Appendix C: New Mplus Featuresin Version 8.3 and Version 8.4
http://www.statmodel.com/verhistory.shtml
Mixture modeling:Significant speed improvements for computationally demandingmixture models, particularly with multiple latent class variablessuch as with Latent Transition AnalysisSubstantially simplified output for mixture models with multiplelatent class variables
Bayesian analysis:Significant speed improvements for Bayesian computations usinga new parallelized computing approach (use PROCESSORS = 8)Bayesian estimation of two-level models with latent variableinteractions using the XWITH option. This is especially helpfulfor models with moderation where ML has problemsExpanded Bayesian fit statistics in 3 areas: improved posteriorpredictive p-values (PPP) when there are missing data, BayesianCFI/TLI/RMSEA including confidence intervals, and a Bayesianversion of the Wald test of parameter restrictions using theMODEL TEST command
Bengt Muthen Mplus Web Talks: No. 1 79/ 80
Some Recent Mplus-Related Papers
Asparouhov & Muthen (2019). Random Starting Values and MultistageOptimization (Technical Report)
Asparouhov & Muthen (2019). Bayes Parallel Computation: Choosing thenumber of processors (Technical Report)
Asparouhov & Muthen (2019). Bayesian estimation of single and multilevelmodels with latent variable interactions. Forthcoming in Structural EquationModeling
Asparouhov & Muthen (2019). Advances in Bayesian Model Fit Evaluation forStructural Equation Models. Forthcoming in Structural Equation Modeling
Asparouhov & Muthen (2019). Latent variable centering of predictors andmediators in multilevel and time-series models. Structural Equation Modeling:A Multidisciplinary Journal, 26, 119-142.
Asparouhov & Muthen (2020). Comparison of models for the analysis ofintensive longitudinal data. Structural Equation Modeling: A MultidisciplinaryJournal, 27(2) 275-297
Muthen & Asparouhov (2020). Random intercept latent transition analysis(RI-LTA). Under review in Psychological Methods
Bengt Muthen Mplus Web Talks: No. 1 80/ 80