alexandre j.s. morin, simon a. houle, & david litalien...panel discussion: increasing our...
TRANSCRIPT
Panel Discussion: Increasing our Analytic Sophistication
Person-Centered AnalysesAlexandre J.S. Morin, Simon A. Houle, &
David Litalien
2017 Conference on Commitment
Focus: Mixture Modeling / Generalized Structural Equation Modeling.
Meyer, J.P., & Morin, A.J.S. (2016). A person-centered approach
to commitment research: Theory, research, and
methodology. Journal of Organizational Behavior, 37, 584-612.
Morin, A.J.S. (2016). Person-centered research strategies in
commitment research. In J.P. Meyer (Ed.), The Handbook of
Employee Commitment (p. 490-508). Cheltenham, UK: Edward
Elgar
Morin, A.J.S., & Wang, J.C.K. (2016). A gentle introduction to
mixture modeling using physical fitness data. In N. Ntoumanis, &
N. Myers (Eds.), An Introduction to Intermediate and Advanced
Statistical Analyses for Sport and Exercise Scientists (pp. 183-210).
London, UK: Wiley.
http://smslabstats.weebly.com/
https://www.statmodel.com/
https://www.statisticalinnovations.com/
https://www.r-project.org/
http://www.gllamm.org/
Psychological Constructs Can Be Categorical
The latent variable
is categorical
C
S1 S2 S3 SX…
S1 S2 S3 SX…
F
Factor Analyses
Vs. latent profile analyses
S1 S2 S3 SX…
F
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Person 1 Person 2 Person 3 Sample Mean
Factor (or group of competencies) 1
Factor (or group of competencies) 2
S1 S2 S3 SX…
F1 F2
The latent variable
is categorical C
S1 S2 S3 SX…
0
1
2
3
4
5
6
Profile 1 Profile 2 Profile 3 Profile 4
Competency 1
Competency 2
Competency 3
Competency 4
Factor Analyses
Vs. latent profile analyses
3 Characteristics of Person-Centered Models
(Mixture Models)
• Typological
• Prototypical
• Exploratory
Typological: Resulting in a classification system designed to help
categorize individuals more accurately into qualitatively and
quantitatively distinct subpopulations.
0
1
2
3
4
5
6
Profile 1 Profile 2 Profile 3 Profile 4
Competency 1
Competency 2
Competency 3
Competency 4
• Prototypical: With all participants having a probability of
membership in all profiles based on their degree of similarity
with the profiles’ specific configurations.
0
1
2
3
4
5
6
Profile 1 Profile 2 Profile 3 Profile 4
Competency 1
Competency 2
Competency 3
• Exploratory: Although exceptions exist. For this reason, a
class-enumeration process typically needs to be conducted –
and multiple random starts need to be integrated. Model LL #fp AIC CAIC BIC ABIC Entropy
Organizational Commitment
1 Profile -4704.605 6 9421.211 9457.135 9451.135 9432.078 Na
2 Profiles -4586.674 13 9199.347 9274.185 9264.185 9222.894 0.556
3 Profiles -4467.097 20 8974.193 9087.943 9073.943 9010.419 0.688
4 Profiles -4401.543 27 8857.086 9009.748 8991.748 8905.990 0.705
5 Profiles -4364.052 34 8796.105 8987.679 8965.679 8857.688 0.681
6 Profiles -4320.230 41 8722.460 8952.947 8926.947 8796.722 0.712
7 Profiles -4301.966 48 8699.932 8969.331 8939.331 8786.873 0.726
8 Profiles -4292.316 55 8694.632 9002.944 8968.944 8794.253 0.753
9 Profiles -4267.496 62 8658.992 9006.217 8968.217 8771.292 0.721
10 Profiles -4258.751 69 8655.501 9041.638 8999.638 8780.480 0.732
7000
7500
8000
8500
9000
9500
1 2 3 4 5 6 7 8 9 10
Number of Latent Profiles
AIC
CAIC
BIC
ABIC
Method Vs. Application
By exploratory, I am referring to the method itself which, like
EFA (or ESEM) is exploratory in nature (all items are associated
with all latent factors or all latent categories). In addition,
absolute fit indices are not available with mixture models.
However, the applications of both methods (e.g., using target
rotation or Bayes prior for EFA/ESEM) can be confirmatory and
theory-driven.
C
S1 S2 S3 SX…
X
XC
Y
YC
C
S1 S2 S3 SX…
C
S1 S2 S3 SX…
FC
Generalized SEM
Latent Profile Analysis
K
k
ykky
K
k
ykky1
22
1
2)(
C
S1 S2 S3 SX…
Mixture Indicators?• Anything goes, and any form of combination.
• Mixture models are not affected by the scale of the
indicators (contrary to cluster analyses).
• In theory, fully latent mixture models can be
estimated where items are used to estimate factors,
which are themselves used to estimate profiles.
• But mixture models are very complex, and
computer-intensive. So they are more typically
estimated using scale scores.
Scale Scores ? • What about measurement errors???
• Why not use factor scores:
o Provide a partial control for measurement error
o Forces you to demonstrate that the measurement model fits
the data well.
o Preserves the nature of the measurement model: invariance,
cross loadings, method factors, etc.
o Naturally standardized. Morin, A.J.S., Meyer, J.P., Creusier, J., & Biétry, F. (2016). Multiple-group analysis of similarity in latent profile
solutions. Organizational Research Methods, 19, 231-254.
Morin, A.J.S., & Marsh, H.W. (2015). Disentangling Shape from Levels Effects in Person-Centred Analyses:
An Illustration Based University Teacher Multidimensional Profiles of Effectiveness. Structural Equation
Modeling, 22, 39-59.
Morin, A.J.S., Boudrias, J.-S., Marsh, H.W., Madore, I., & Desrumaux, P. (2016). Further reflections on
disentangling shape and level effects in person-centered analyses: An illustration aimed at exploring the
dimensionality of psychological health. Structural Equation Modeling, 23, 438-454
Morin, A.J.S., Boudrias, J.-S., Marsh, H.W., McInerney, D.M., Dagenais-Desmarais, V., & Madore, I. (In Press).
Complementary variable- and person-centered approaches to exploring the dimensionality of
psychometric constructs: Application to psychological wellbeing at work. Journal of Business and
Psychology. EarlyView DOI 10.1007/s10869-016-9448-7
Basic Mplus input:DATA:
FILE IS Commitment.dat;
VARIABLE:
NAMES = ID Status unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1
NC1 CC1 AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1;
MISSING = all (999);
IDVARIABLE = ID;
CLASSES = c (3);
Analysis:
TYPE = MIXTURE;
ESTIMATOR = MLR;
PROCESS = 3;
STARTS = 3000 100;
STITERATIONS = 100;
CLASSES: serves to identify the number of latent profiles to be estimated.
STARTS: serves to identify the number of random starts (first number) and the number retained for final optimization (second number)
STITERATIONS : Identifies the number of iterations for each random starts
Basic Mplus input:DATA:
FILE IS Commitment.dat;
VARIABLE:
NAMES = ID Status unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1
NC1 CC1 AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1;
MISSING = all (999);
IDVARIABLE = ID;
CLASSES = c (3);
Analysis:
TYPE = MIXTURE;
ESTIMATOR = MLR;
PROCESS = 3;
STARTS = 3000 100;
STITERATIONS = 100;
PROCESS: Your best friend with mixture models… Allows you to use multiple processors in parallel.
RANDOM STARTS RESULTS RANKED FROM THE BEST TO THE WORST
LOGLIKELIHOOD VALUES
15 perturbed starting value run(s) did not converge.
Final stage loglikelihood values at local maxima, seeds, and initial stage start numbers:
-6095.887 991399 1433
-6095.887 165268 2436 The best loglikelihood needs to be
-6095.887 551639 55 replicated. This is the case here.
-6095.887 58353 1723
-6095.887 168648 1788
-6095.887 445692 2796
-6096.828 762461 425
-6096.828 62715 2599
-6104.647 264951 1130
-6104.647 930872 277
-6104.647 670998 1876
-6104.698 98068 998
-6105.491 361131 2563
-6105.491 293428 2473
-6105.491 18741 2970
…
Model section for LPAMODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%c#2%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%c#3%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%OVERALL%This section is used to describe parts of the model that are common to all profiles.
%C#1% to %C#k% This section is used to describe parts of the model that are specific to each profiles.
MODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%c#2%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%c#3%
[AC1 NC1 CC1];
AC1 NC1 CC1;
MODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1];
%c#2%
[AC1 NC1 CC1];
%c#3%
[AC1 NC1 CC1];
MODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1];
AC1 NC1 CC1 (v1-v3);
%c#2%
[AC1 NC1 CC1];
AC1 NC1 CC1 (v1-v3);
%c#3%
[AC1 NC1 CC1];
AC1 NC1 CC1 (v1-v3);
Peugh, J. & Fan, X. (2013). Modeling unobserved heterogeneity using latent profile analysis: A Monte Carlo simulation. Structural Equation Modeling, 20, 616-639.
To request the free estimation of the means and variances of the indicators in each profile
Mplus default: Variances are specified as equal across classes. Right: short-cut syntaxLeft: Complete syntax
Unequal variances Equal Variances
0
1
2
3
4
5
6
7
0
1
2
3
4
5
6
7
Suggested end of the inputOUTPUT:
SAMPSTAT CINTERVAL RESIDUAL SVALUES TECH1 TECH7 TECH11 TECH14;
! To save the probabilities of class membership for each participants according to the ID
variable in a new data set. Name this data set as you want but make sure to change
the name from one model to the next.
SAVEDATA:
FILE IS LPA_5_final.dat;
FORMAT IS FREE;
SAVE=CPROBABILITIES;
Morin, A.J.S., Morizot, J., Boudrias, J.-S., & Madore, I. (2011). A multifoci person-centered perspective on workplace affective commitment: A latent profile/factor mixture Analysis. OrganizationalResearch Methods, 14 (1), 58-90.
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
OrganizationSupervisorGroupJobCareerWorkCustomers
7%workplace-committed
31%supervisor-committed
17%career-committed
19%Uncommited
5%committed
Class Enumeration• Mixture models are essentially exploratory.
• The typical approach to determine whether profiles are
needed, and the number of profiles that is needed is called the
class enumeration process.
• Essentially, solutions varying from 1 to some number of
profiles are estimated, and contrasted.
• “Some” number: This is typically determined based on the
following considerations (with no absolute rules):
o What is the range of profiles that you can expect based on
the literature: Add at least 2-3 to this number.
o Sample size: smaller samples makes it possible to identify
fewer profiles.
Class Enumeration
Criteria to guide the selection of the optimal solution:
o Theoretical Conformity and Value
o Statistical adequacy
oDoes adding one profile adds something
meaningful to the solution ? OR does it simply
split an existing profiles into two with a similar
shape that is simply more or less extreme?
o The size of the added profiles: Tricky criteria.
• Some mention 1% or 5%.
• But this really depends on the sample size.
• You want a solution that is stable… Profiles smaller
than 25, or perhaps ideally 50, participants are
generally desirable.
Class Enumeration
o Statistical indicators
• Akaike Information Criteria (AIC)
• Consistent AIC (CAIC)
• Bayesian Information Criterion (BIC)
• Sample size adjusted BIC (ABIC)
• A lower value on these indicators suggest a better
fitting model. How much lower ? No clear guideline
exists here.
• Simulation studies shows that the AIC does not
work.
• The CAIC is calculated by hand: BIC + number of
free parameters.
Class Enumeration
o Statistical indicators
• The model log likelihood could be used to calculate a
chi-square difference test (as in SEM), to compare
nested models including the same number of profiles,
but not models including a different number of profiles.
• The Lo-Mendell-Rubin (LMR) and the Adjusted LMR
(aLMR) likelihood ratio test. Typically both give the
same consultation and only one is reported. TECH 11
• The Bootstrap Likelihood Ratio Rest (BLRT).TECH 14
• A non-significant LMR, aLMR, BLRT supports the
model with one fewer class.
• Simulation studies shows that the LMR-aLMR do not
work well.
Class Enumeration
o Statistical indicators
• Entropy: Varies between 0 and 1 and summarizes the
classification accuracy of participants into profiles.
• i.e. when most participants have a probability close
to 1 of membership into their most likely profile and
a probability close to 0 of membership into the other
profiles, the entropy will be close to 1.
• This is also an indicator of class separation.
• The entropy should never be used to select the final
solution. Mixture models can be fully proper even
when the classification is more “fuzzy”, it just means
that correspondences to prototypes are not as clear.
Class Enumeration
LMR, aLMR, BLRT• These tests compare a k-profile solution with a k-1-profile
solution.
• As such, the LogLikelihood associated with the k-1 profile
solution should correspond to the true k-1 profile solution,
and at least some of the BLRT (bootstrapped) replications
should converge properly.
• When this does not happen, you may want to increase the
number of random starts and iterations specifically used for
these tests.
Analysis:
TYPE = MIXTURE;
ESTIMATOR = MLR;
PROCESS = 3;
STARTS = 3000 100;
STITERATIONS = 100;
K-1STARTS = 3000 100;
LRTSTARTS = 3000 100 3000 100;
LRTBOOTSTRAP = 100;
OUTPUT:
SAMPSTAT CINTERVAL RESIDUAL SVALUES TECH1 TECH7 ;
TECH11 TECH14;
K-1STARTS : For LMR/ALMR and BLRT. Specify the number of random starts for the model with one less profile.
LRTSTARTS : For BLRT. Specify the number of random starts for the model itself (second) and the model with one less profile (first).
LRTBOOTSTRAP : For BLRT. Specify the number of bootstrap draws
Shortcut?Conduct the class enumeration process without requesting
LMR/aLMR and BLRT. Then pick the best final solution for each
number of profiles and use either the seed, or the starts values,
to re-estimate this solution while turning the random start
function to 0 and only focusing on the k-1 starts,
1. Using the seed
RANDOM STARTS RESULTS RANKED FROM THE BEST TO THE WORST
LOGLIKELIHOOD VALUES
15 perturbed starting value run(s) did not converge.
Final stage loglikelihood values at local maxima, seeds, and initial stage start numbers:
-6095.887 991399 1433
-6095.887 165268 2436
-6095.887 551639 55
-6095.887 58353 1723
-6095.887 168648 1788
-6095.887 445692 2796
-6096.828 762461 425
-6096.828 62715 2599
-6104.647 264951 1130
-6104.647 930872 277
-6104.647 670998 1876
-6104.698 98068 998
-6105.491 361131 2563
-6105.491 293428 2473
-6105.491 18741 2970
…
Analysis:
TYPE = MIXTURE;
ESTIMATOR = MLR;
PROCESS = 3;
STARTS = 0;
OPTSEED = 991399;
K-1STARTS = 3000 100;
2. Using the starts values
MODEL COMMAND WITH FINAL ESTIMATES USED AS STARTING VALUES
%OVERALL%
[ c#1*-0.93515 ];
[ c#2*0.49113 ];
%C#1%
[ CC1*0.39664 ];
[ NC1*-0.93155 ];
[ AC1*-1.59367 ];
CC1*1.59137;
NC1*2.12542;
AC1*0.53789;
%C#2%
[CC1*-0.15201 ];
[NC1*-0.29812 ];
[AC1*-0.18248 ];
CC1*0.72513;
NC1*0.43116;
AC1*0.42376;
%C#3%
[CC1*0.08333 ];
[NC1*0.86526 ];
[AC1*0.94111 ];
…
AND
Analysis:TYPE = MIXTURE; ESTIMATOR = MLR;PROCESS = 3; STARTS = 0; K-1STARTS = 3000 100;
Class Enumeration: In Practice• The various indicators will very seldom converge on the same
solution, hence the importance of considering all of them, as
well as the other criteria (theoretical and statistical adequacy,
added-value, size, etc.).
• Often, the indicators will even keep on decreasing (AIC,
CAIC, BIC, ABIC) or being significant (LMR, aLMR, BLRT)
when profiles are added, up until the maximum number of
profiles that you have decided to estimate, and even when
solutions stop being proper statistically (revealing even
“empty” profiles.
• The information criteria can then be represented by an elbow
plot. In these plots, the key is not to locate the first angle (as
in the scree plot for EFA), but the point after which a plateau
is reached (decreases becomes negligible).
Model LL #fp AIC CAIC BIC ABIC Entropy
Organizational Commitment 1 Profile -4704.605 6 9421.211 9457.135 9451.135 9432.078 Na
2 Profiles -4586.674 13 9199.347 9274.185 9264.185 9222.894 0.556
3 Profiles -4467.097 20 8974.193 9087.943 9073.943 9010.419 0.688 4 Profiles -4401.543 27 8857.086 9009.748 8991.748 8905.990 0.705
5 Profiles -4364.052 34 8796.105 8987.679 8965.679 8857.688 0.681
6 Profiles -4320.230 41 8722.460 8952.947 8926.947 8796.722 0.712 7 Profiles -4301.966 48 8699.932 8969.331 8939.331 8786.873 0.726
8 Profiles -4292.316 55 8694.632 9002.944 8968.944 8794.253 0.753
9 Profiles -4267.496 62 8658.992 9006.217 8968.217 8771.292 0.721 10 Profiles -4258.751 69 8655.501 9041.638 8999.638 8780.480 0.732
7000
7500
8000
8500
9000
9500
1 2 3 4 5 6 7 8 9 10
Number of Latent Profiles
AIC
CAIC
BIC
ABIC
Model LL #fp AIC CAIC BIC ABIC Entropy aLMR BLRT
1 -9582 12 19188 19262 19250 19212 Na Na Na
2 -8824 25 17827 17852 17827 17747 0.784 ≤ 0.001 ≤ 0.001
3 -8556 38 17189 17422 17383 17263 0.832 ≤ 0.001 ≤ 0.001
4 -8430 51 16964 17276 17225 17063 0.821 0.0237 ≤ 0.001
5 -8353 64 16834 17225 17161 16958 0.844 ≤ 0.001 ≤ 0.001
6 -8301 77 16757 17228 17151 16906 0.792 0.013 ≤ 0.001
7 -8277 90 16733 17283 17193 16908 0.784 0.024 ≤ 0.001
8 -8227 103 16660 17290 17187 16859 0.741 1.000 ≤ 0.001
Output1. Is the log likelihood replicated ?
Final stage loglikelihood values at local maxima, seeds, and initial stage start
numbers:
-2217.891 804821 3936
-2217.891 743830 1368
-2217.891 50983 834
-2217.891 816796 1728
-2217.891 619846 3528
-2217.891 554395 2195
-2217.891 822698 621
-2217.891 674419 2870
-2217.891 657909 2912
-2217.891 968846 970
-2217.891 714455 476
-2217.891 112441 3926
-2217.891 786248 2076
…
Output2. Model fit
THE MODEL ESTIMATION TERMINATED NORMALLY
MODEL FIT INFORMATION
Number of Free Parameters 20
Loglikelihood
H0 Value -2217.891
H0 Scaling Correction Factor 0.853
for MLR
Information Criteria
Akaike (AIC) 4475.783
Bayesian (BIC) 4559.752
Sample-Size Adjusted BIC 4496.272
(n* = (n + 2) / 24)
Output2. Model fit
GO TO THE END OF THE OUTPUT
TECHNICAL 11 OUTPUT
Random Starts Specifications for the k-1 Class Analysis Model
Number of initial stage random starts 5000
Number of final stage optimizations 200
VUONG-LO-MENDELL-RUBIN LIKELIHOOD RATIO TEST FOR 2 (H0) VERSUS 3 CLASSES
H0 LoglikelihoodValue -2717.569
2 Times the Loglikelihood Difference 999.355
Difference in the Number of Parameters 7
Mean 29.905
Standard Deviation 54.016
P-Value 0.0000
LO-MENDELL-RUBIN ADJUSTED LRT TEST
Value 976.842
P-Value 0.0000
Output2. Model fit
TECHNICAL 14 OUTPUT
Random Starts Specifications for the k-1 Class Analysis Model
Number of initial stage random starts 5000
…Number of bootstrap draws requested 100
PARAMETRIC BOOTSTRAPPED LIKELIHOOD RATIO TEST FOR 2 (H0) VERSUS 3 CLASSES
H0 LoglikelihoodValue -2717.569
2 Times the Loglikelihood Difference 999.355
Difference in the Number of Parameters 7
Approximate P-Value 0.0000
Successful Bootstrap Draws 0WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED IN 93 OUT OF 100
WARNING: 100 OUT OF 100 BOOTSTRAP DRAWS DID NOT CONVERGE.
THE P-VALUE MAY NOT BE TRUSTWORTHY.
INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.
Output3. Class Counts
FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASSES
BASED ON THE ESTIMATED MODEL
1 30.99528 0.06300
2 429.02931 0.87201
3 31.97540 0.06499
FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS
BASED ON ESTIMATED POSTERIOR PROBABILITIES
1 30.99528 0.06300
2 429.02931 0.87201
3 31.97540 0.06499
CLASSIFICATION QUALITY Entropy 1.000
CLASSIFICATION OF INDIVIDUALS BASED ON THEIR MOST LIKELY LATENT CLASS
MEMBERSHIP
1 31 0.06301
2 429 0.87195
3 32 0.06504
Output3. Class Counts
FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASSES
BASED ON THE ESTIMATED MODEL1 30.99528 0.06300
2 429.02931 0.87201
3 31.97540 0.06499
FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS
BASED ON ESTIMATED POSTERIOR PROBABILITIES1 30.99528 0.06300
2 429.02931 0.87201
3 31.97540 0.06499
CLASSIFICATION QUALITY Entropy 1.000
CLASSIFICATION BASED ON MOST LIKELY CLASS MEMBERSHIP1 31 0.06301
2 429 0.87195
3 32 0.06504
Fully model based estimate of the size of each profile (n = 30.999; 6.3%)
Based probability of class membership. Typically equal to model-based.
Based on most likely class assignment. Less accurate.
Output4. Classification Summary
Average Latent Class Probabilities for Most Likely Latent Class Membership (Row)
by Latent Class (Column)
1 2 3
1 1.000 0.000 0.000
2 0.000 1.000 0.000
3 0.000 0.001 0.999
Output5. Results
For LPA, stick with the unstandardized results.
For estimates of correlations or regression, then also look at the STDYX results.
MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value
Latent Class 1
Means
FAC 1.139 0.002 460.118 0.000
FCC 4.183 0.314 13.324 0.000
FNC 1.213 0.016 74.366 0.000
Variances
FAC 0.000 0.000 7.395 0.000
FCC 3.052 0.373 8.187 0.000
FNC 0.008 0.001 7.307 0.000
…
Reporting• Raw scores: Impossible to compare across scales or studies.
• Normed scores: Ideal, but seldom available.
• Standardized scores (no need to transform if using factor
scores): Easily interpretable, but also sample specific.
Recommendation:
• Histogram (do not use line graphs, these are impossible to
interpret) using standardized scores for the means
• Table for the exact mean and variance results, with confidence
intervals.
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5 Profile 6 Profile 6
Mean [CI] Mean [CI] Mean [CI] Mean [CI] Mean [CI] Mean [CI] Mean [CI]
Affective Commitment -.028
[-.284; .228]
-.792
[-1.065; -.519]
-.435
[-.632; -.239]
-1.623
[-1.753; -1.492]
.393
[.102; .684]
.632
[.525; .740]
.945
[.863; 1.027]
Normative Commitment .088
[-.093; .268]
-.854
[-1.083; -.625]
-.406
[-.526; -.286]
-1.445
[-1.587; -1.304]
-.010
[-.251; .232]
.555
[.424; .685]
1.019
[.855; 1.184]
Continuance Commitment:
High Sacrifice
1.050
[.898; 1.203]
-.960
[-1.053; -.866]
-.081
[-.183; .020]
-.702
[-.923; -.481]
-.604
[-.927; -.282]
.496
[.364; .627]
1.639
[1.533; 1.744] Continuance Commitment:
Low Alternatives
1.427
[1.277; 1.577]
-.364
[-.466; -.261]
.379
[.216; .543]
.777
[.535; 1.020]
-.811
[-1.094; -.529]
.172
[-.007; .350]
1.289
[1.076; 1.501]
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5 Profile 6 Profile 6 Variance
[CI]
Variance
[CI]
Variance
[CI]
Variance
[CI]
Variance
[CI]
Variance
[CI]
Variance
[CI]
Affective Commitment .277
[.058; .495]
.286
[.198; .374]
.245
[.159; .331]
.277
[.202; .351]
.281
[.126; .436]
.156
[.129; .182]
.065
[.036; .095]
Normative Commitment .317
[.201; .433]
.227
[.155; .300]
.205
[.168; .242]
.300
[.207; .394]
.393
[.243; .544]
.207
[.146; .267]
.216
[.126; .306]
Continuance Commitment: High Sacrifice
.033 [-.016; .082]
.028 [.004; .051]
.139 [.107; .171]
.598 [.353; .843]
.241 [.152; .330]
.140 [.099; .181]
.105 [.074; .135]
Continuance Commitment:
Low Alternatives
.082
[-.086; .251]
.060
[.009; .111]
.214
[.168; .259]
.833
[.577; 1.089]
.214
[.144; .284]
.274
[.198; .350]
.408
[.276; .540]
BUT
HOF
C
S1 S2 S3 SX…
Factor mixture analysis.
Shape Versus Level• Profiles should present qualitative (shape) differences.
• Ordered profiles, showing only quantitative level differences,
would be better represented by variable-centered methods.
• In some cases both level and shape effects can be strong,making the identification of qualitatively distinct profilesharder since strong level effects tend to create equally strongquantitative differences.
• For instance, some employees may be globally good, or badacross dimensions of effectiveness when compared to others.
• However these profiles may still present relative areas ofstrength and weaknesses worthy of consideration.
• Unfortunately, these specific areas will be harder to identifygiven that even these weaknesses may be relatively strongcompared to the levels observed in weaker profiles.
-2
-1.5
-1
-0.5
0
0.5
1
1.5
Class 1 Class 2 Class 3 Class 4
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Class 1 Class 2 Class 3 Class 4
Morin, A.J.S., & Marsh, H.W. (2015). Disentangling Shape from
Levels Effects in Person-Centred Analyses: An Illustration Based
University Teacher Multidimensional Profiles of Effectiveness.
Structural Equation Modeling, 22, 39-59.
Model 1C
S1 S2 S3 SX…
X
Model 1: Classical latent profile analysis.
MODEL:
%OVERALL%
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9];
VAR1-VAR9 ON CONT;
! Due to the control, conditional independence needs to be specified.
VAR1 WITH VAR2-VAR9@0; VAR2 WITH VAR3-VAR9@0;
VAR3 WITH VAR4-VAR9@0; VAR4 WITH VAR5-VAR9@0;
VAR5 WITH VAR6-VAR9@0; VAR6 WITH VAR7-VAR9@0;
VAR7 WITH VAR8-VAR9@0; VAR8 WITH VAR9@0;
%c#1%
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9];
%c#2%
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9];
[…]
Model 3
HOF
C
S1 S2 S3 SX…
X
Model 3: Factor mixture analysis with a class
invariant higher order latent factor.
MODEL:
%OVERALL%
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9];
VAR1-VAR9 ON CONT;
VAR1 WITH VAR2-VAR9@0; VAR2 WITH VAR3-VAR9@0;
VAR3 WITH VAR4-VAR9@0; VAR4 WITH VAR5-VAR9@0;
VAR5 WITH VAR6-VAR9@0; VAR6 WITH VAR7-VAR9@0;
VAR7 WITH VAR8-VAR9@0; VAR8 WITH VAR9@0;
HOF BY VAR1* VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[HOF@0]; HOF@1;
%c#1%
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9];
%c#2%
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9;
[VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9];
[…]
-2
-1.5
-1
-0.5
0
0.5
1
1.5
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
F1 = Learning/value
F2 = Enthusiasm
F3 = Group interaction
F4 = Organization/Clarity
F5 = Individual Rapport
F6 = Workload/difficulty
F7 = Breadth of Coverage
F8 = Exams/grading
F9 = Assignments/readings
Results from the latent profiles models based on 9 factors (Model 1)
-1.5
-1
-0.5
0
0.5
1
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
F1 = Learning/value
F2 = Enthusiasm
F3 = Group interaction
F4 = Organization/Clarity
F5 = Individual Rapport
F6 = Workload/difficulty
F7 = Breadth of Coverage
F8 = Exams/grading
F9 = Assignments/readings
Results from the factor mixture models based on 9 factors (Model 3)
Poor (11%)
Good (25%)Cognitive
Performance Oriented
(25%)
Relations and Performance
Oriented (20%)
Cognitive and Mastery
Oriented (19%)
BUTThe Factor Mixture approach also assumes that the mean level
on the global factor is equal across all profiles, and corresponds
to the sample average.
Morin, A.J.S., Boudrias, J.-S., Marsh, H.W., Madore, I., & Desrumaux, P. (2016). Further reflections on disentangling shape and level effects in person-centered analyses: An illustration aimed at exploring the dimensionality of psychological health. Structural Equation Modeling, 23, 438-454
Morin, A.J.S., Boudrias, J.-S., Marsh, H.W., McInerney, D.M., Dagenais-Desmarais, V., & Madore, I. (In Press). Complementary variable- and person-centered approaches to exploring the dimensionality of psychometric constructs: Application to psychological wellbeing at work. Journal of Business and Psychology. Early View DOI 10.1007/s10869-016-9448-7
The World Health Organization (2014) defines psychological
health as a state characterized not only by the absence of
psychological distress, but also by the presence of psychological
wellbeing.
Wellbeing
Health
+ -
Distress
Psychological Health at Work
Preliminary results showed that a ESEM model clearly fitted the
data better than a CFA model, and that a bifactor-ESEM model
was even better.
Profiles were calculated from factor scores saved from the ESEM
model, as well as from the Bifactor-ESEM model.
Psychological Wellbeing:• Harmony (BHAR)• Serenity (BSER)• Involvement (BIMP)
Psychological Distress:• Irritability/ aggression (DIRR)• Anxiety/depression (DANX)• Withdrawal (DDES)
Latent Profile Model
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
Profile 1 Profile 2 Profile 3 Profile 4
Harmony
Serenity
Involvement
Irritability
Anxiety/Depression
Distance
PURE LEVELArgues for a single dimension, and for the idea that wellbeing and distress are opposite ends
of the same continuum
Factor Mixture Model
Latent Profile Model (Bifactor)
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Class 1 Class 2 Class 3 Class 4 Class 5
G-Factor
BHAR
BSER
BIMP
DIRR
DANX
DDES
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
Global Psychological Health
Harmony
Serenity
Involvement
Irritability
Anxiety/Depression
Distance
Normative Profile: 61% of the sample for which health is best represented as a single dimension (with inter-individual variability)
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
Global Psychological Health
Harmony
Serenity
Involvement
Irritability
Anxiety/Depression
Distance
Adapted Profile: 11% of the sample who have high level of global health, do not feel any distress, but are also not really thriving
Flourishing Profile: Only 1.38% of the sample who really thrive. Suggesting that not feeling any distress is critical to being in good health, but that high level of wellbeing are necessary to thrive.
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
Global Psychological Health
Harmony
Serenity
Involvement
Irritability
Anxiety/Depression
Distance
Stressfully Involved Profile: 14% of the sample who appear to cope with work stress by increasing involvement
Harmoniously Distanced Profile: 12% of the sample who appear to cope with work stress by distancing themselves, thus restoring some level of harmony and serenity.
Factor Mixture Models
C
S1 S2 S3 SX…
F
HOF
C
S1 S2 S3 SX…
Factor Mixture Models• Can be used as a multiple-group CFA model to test for the
invariance of a measure across unobserved groups of
participants.
• I.e. the profiles are used to locate “response sets”.
• Carter, N.T., Dalal, D.K., Lake, C.J., Lin, B.C., & Zickar, M.J. (2011). Using
mixed-model item response theory to analyze organizational survey
responses: An illustration using the Job Descriptive Index. Organizational
Research Methods, 14, 116-146.
• Tay, L., Diener, E., Drasgow, F., & Vermunt, J.K. (2011). Multilevel mixed-
measurement IRT analysis: An explication and application to self-reported
emotions across the world. Organizational Research Methods, 14, 177-207.
• Tay, L., Drasgow, D.A., & Vermunt, J.K. (2011). Using mixed-measurement
item response theory with covariates (MM-IRT-C) to ascertain observed
and unobserved measurement equivalence. Organizational Research
Methods, 14, 147-176.
Factor Mixture Models• Can be used as an overarching framework to explore the
“true” underlying nature of psychological constructs as
continuous, ordinal, nominal, etc.
• Lubke, G.H., & Muthén, B.O (2005). Investigating population heterogeneity
with factor mixture models. Psychological Methods, 10, 21-39.
• Borsboom, D., Rhemtulla, M., Cramer, A.O.J., van der Maas, H.L.J., Scheffer,
M., & Dolan, C.V. (2016). Kinds versus continua: a review of psychometric
approaches to uncover the structure of psychiatric constructs. Psychological
Medicine, 46, 1567-1579.
• Clark, S.L., Muthén, B.O., Kaprio, J., D’Onofrio, B.M., Viken, R., & Rose, R.J.
(2013). Models and strategies for factor mixture analysis: An example
concerning the structure of underlying psychological disorders. Structural
Equation Modeling, 20, 681-703.
• Masyn, K., Henderson, C., & Greenbaum, P. (2010). Exploring the latent
structures of psychological constructs in social development using the
Dimensional-Categorical Spectrum. Social Development, 19, 470–493.
CFA Vs. latent profile analyses
Choosing between CFA and LPA is not easy since a
k-class LPA model has identical covariance
implications than a k-1 common factor model and
thus represents an equivalent model (Bauer &
Curran, 2004; Steinley & McDonald, 2007).
Construct Validation
Construct validation is the only way to demonstrate
the meaningfulness of latent profile solutions
oHeuristic Value
o Theoretical Conformity and Value
o Statistical adequacy
Construct Validation
Construct validation is the only way to demonstrate
the meaningfulness of latent profile solutions
oHeuristic Value
o Theoretical Conformity and Value
o Statistical adequacy
o Shows meaningful and well-differentiated
relations to key covariates:
• Predictors
• Correlates
• Outcomes.
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
Age
Tenure
Relations colleagues
Relations supervisor
Justice
Satisfaction
31%supervisor-committed
17%career-committed
7%workplace-committed
19%Uncommited
5%committed
-1.0
-0.5
0.0
0.5
1.0
Profile 1 Profile 2 Profile 3 Profile 4 Profile 5
In-role performance
OCB tasks
OCB group
OCB organization
OCB supervisor
OCB customers
Intent to quit
31%supervisor-committed
17%career-committed
7%workplace-committed
19%Uncommited
5%committed
Predictors and Outcomes
C
S1 S2 S3 SX…
X
XC
Y
YC
Indicator
Multinomial logistic
regressionThis is a key advantage of mixture models, relative to
cluster analyses.
Predictors and Outcomes
C
S1 S2 S3 SX…
X
XC
Y
YC
But these should not change the nature of the profiles.
Indicator
Multinomial logistic
regression
Always conduct the class enumeration process before adding covariates (Diallo et
al., 2016).
Solution 1. Use the start values from the retained solution and
turn off the random start function
Predictors and Outcomes
C
S1 S2 S3 SX…
X
XC
Y
YC
But these should not change the nature of the
profiles.
Solution 2: Auxiliary functions: (E) For simple correlates(DCON) (DCAT) for outcomes(R3STEP) for predictors(DU3STEP) (DE3STEP) for continuous outcomes(BCH) for continuous outcomes
Indicator
Multinomial logistic
regression
Integrating covariates to the final retained
solution
Using the starts values
MODEL COMMAND WITH FINAL ESTIMATES USED AS STARTING VALUES
%OVERALL%
[ c#1*-0.93515 ];
[ c#2*0.49113 ];
%C#1%
[ CC1*0.39664 ];
[ NC1*-0.93155 ];
[ AC1*-1.59367 ];
CC1*1.59137;
NC1*2.12542;
AC1*0.53789;
%C#2%
[CC1*-0.15201 ];
[NC1*-0.29812 ];
[AC1*-0.18248 ];
CC1*0.72513;
NC1*0.43116;
AC1*0.42376;
%C#3%
[CC1*0.08333 ];
[NC1*0.86526 ];
[AC1*0.94111 ];
…
AND
Analysis:TYPE = MIXTURE; ESTIMATOR = MLR;PROCESS = 3; STARTS = 0;
%OVERALL%
[ c#[email protected] ];
[ c#[email protected] ];
%C#1%
[ [email protected] ];
[ [email protected] ];
[ [email protected] ];
%C#2%
%C#3%
…
AND
Analysis:TYPE = MIXTURE; ESTIMATOR = MLR;PROCESS = 3; STARTS = 0;
DATA: FILE IS Commitment.dat;
VARIABLE:
NAMES = ID unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1NC1 CC1 AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1 Pred1 Pred2 ;
MISSING = all (999); IDVARIABLE = ID; CLASSES = c (3);
Analysis: […]MODEL:
%OVERALL%
c#1-c#2 ON Pred1 Pred2 ;
%c#1%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%c#2%
[AC1 NC1 CC1];
AC1 NC1 CC1;
%c#3%
[AC1 NC1 CC1];
AC1 NC1 CC1;
Always one less class statement than the total number of profiles (c#1-c#2 with 3 profiles, c#1-c#3 with 4 profiles, etc.)
Approaches 1-2: Predictors
Multinomial logistic regressionCategorical Latent Variables
C#1 ON
Pred1 1.122 0.409 2.745 0.006
Pred2 -0.012 0.076 -0.153 0.879
C#2 ON
Pred1 0.806 0.390 2.067 0.039
Pred2 0.133 0.082 1.620 0.105
LOGISTIC REGRESSION ODDS RATIO RESULTS
Categorical Latent Variables
C#1 ON
Pred1 3.072
Pred2 0.988
C#2 ON
Pred1 2.238
Pred2 1.142
• All of these coefficients refer to the likelihood of being
member of the target profile relative to the referent profile
(always the last one). Pairwise comparisons.
• These coefficients can be converted to easier to interpret
odds ratio by taking the antilogarithm of the unstandardized
regression coefficient OR = ln-1(coef) = ecoef. Easy to calculate
in Excel using the EXP function.
• An odds ratio indicates, for each unit increase in the
coefficient, how much the likelihood of membership into the
target group has increased (recovery) relative to the
comparison group.
• OR = 1.5: 1.5 times more likely (50% more likely)
• OR = -.5: 50% less likely; -.33 = 3 times less likely.
ALTERNATIVE PARAMETERIZATIONS FOR THE
CATEGORICAL LATENT VARIABLE REGRESSION
Parameterization using Reference Class 1
C#2 ON
Pred1 -0.317 0.393 -0.806 0.420
Pred2 0.145 0.059 2.449 0.014
C#3 ON
Pred1 -0.812 0.553 -1.469 0.142
Pred2 0.215 0.069 3.102 0.002
Parameterization using Reference Class 2
C#1 ON
Pred1 0.317 0.393 0.806 0.420
Pred2 -0.145 0.059 -2.449 0.014
C#3 ON
Pred1 -0.495 0.533 -0.928 0.353
Pred2 0.070 0.061 1.156 0.248
You just need to calculate the
OR
Approaches 1-2: OutcomesDATA: FILE IS Commitment.dat;
VARIABLE:
NAMES = ID unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1 NC1 CC1 AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1 Out1 Out2 ;
MISSING = all (999); IDVARIABLE = ID; CLASSES = c (3);
Analysis: […]
MODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1]; AC1 NC1 CC1;
[Out1 Out2] (o11-012);
%c#2%
[AC1 NC1 CC1]; AC1 NC1 CC1;
[Out1 Out2] (o21-022);
%c#3%
[AC1 NC1 CC1]; AC1 NC1 CC1;
[Out1 Out2] (o31-032);
MODEL CONSTRAINT:
NEW (ye12); ye12 = o11-o21;
NEW (ye13); ye13 = o11-o31;
NEW (ye23); ye23 = o21-o31;
NEW (yp12); yp12 = o12-o22;
NEW (yp13); yp13 = o12-o32;
NEW (yp23); yp23 = o22-o32;
MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value
Latent Class 1
Means
FAC 1.139 0.002 460.118 0.000
FCC 4.183 0.314 13.324 0.000
FNC 1.213 0.016 74.366 0.000
OUT1 -1.460 0.157 -9.293 0.000
OUT2 -1.438 0.122 -11.823 0.000
Variances
FAC 0.000 0.000 7.395 0.000
FCC 3.052 0.373 8.187 0.000
FNC 0.008 0.001 7.307 0.000
OUT1 0.226 0.013 18.047 0.000
OUT2 0.161 0.006 24.915 0.000
…
New/Additional Parameters
YE12 -1.287 0.091 -14.163 0.000
YE13 -2.533 0.168 -15.116 0.000
…
Integrating covariates to the final retained
solution
3. Auxiliary
Auxiliary: Predictors DATA: FILE IS Commitment.dat;
VARIABLE:
NAMES = ID unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1NC1 CC1
AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1;
MISSING = all (999); IDVARIABLE = ID; CLASSES = c (3);
AUXILIARY = Pred1 (R3STEP) Pred2 (R3STEP);
Analysis: […]
MODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1]; AC1 NC1 CC1;
%c#2%
[AC1 NC1 CC1]; AC1 NC1 CC1;
%c#3%
[AC1 NC1 CC1]; AC1 NC1 CC1;
TESTS OF CATEGORICAL LATENT VARIABLE MULTINOMIAL
LOGISTIC REGRESSIONS USING THE 3-STEP PROCEDURE
Two-Tailed
Estimate S.E. Est./S.E. P-Value
C#1 ON
Pred1 1.646 0.616 2.673 0.008
Pred1 0.221 0.231 0.958 0.338
C#2 ON
Pred1 1.562 0.606 2.575 0.010
Pred2 0.317 0.229 1.388 0.165
Multinomial logistic regression
Including all alternative referent profiles.
You only need to calculate the OR.
Warnings
Sometimes the nature of the profiles changes when the
R3STEP approach is applied. In this situation, you will
get a warning. This is the main limitation of this
approach.
Auxiliary: CorrelatesDATA: FILE IS Commitment.dat;
VARIABLE:
NAMES = ID unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1NC1 CC1
AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1;
MISSING = all (999); IDVARIABLE = ID; CLASSES = c (3);
AUXILIARY = Cor1 (e) Cor2 (e);
Analysis: […]
MODEL:
%OVERALL%
%c#1%
[AC1 NC1 CC1]; AC1 NC1 CC1;
%c#2%
[AC1 NC1 CC1]; AC1 NC1 CC1;
%c#3%
[AC1 NC1 CC1]; AC1 NC1 CC1;
EQUALITY TESTS OF MEANS ACROSS CLASSES USING POSTERIOR
PROBABILITY-BASED MULTIPLE IMPUTATIONS WITH 4 DEGREE(S) OF
FREEDOM FOR THE OVERALL TEST AND 1 DEGREE OF FREEDOM FOR
THE PAIRWISE TESTS
Cor1
Mean S.E. Mean S.E.
Class 1 0.744 0.037 Class 2 0.740 0.035
Class 3 0.598 0.064
Chi-Square P-Value Chi-Square P-Value
Overall test 4.473 0.346 Class 1 vs. 2 0.008 0.930
Class 1 vs. 3 3.783 0.052 Class 2 vs. 3 3.628 0.057
Cor2
Mean S.E. Mean S.E.
Class 1 29.797 0.267 Class 2 30.549 0.287
Class 3 32.491 0.661
Chi-Square P-Value Chi-Square P-Value
Overall test 20.743 0.000 Class 1 vs. 2 3.434 0.064
Class 1 vs. 3 14.094 0.000 Class 2 vs. 3 7.019 0.008
Auxiliary: OutcomesDATA: FILE IS Commitment.dat;
VARIABLE:
NAMES = ID unit Pred1 Pred2 Cor1 Cor2 Out1 Out2 AC1NC1 CC1
AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1;
MISSING = all (999); IDVARIABLE = ID; CLASSES = c (3);
AUXILIARY = Out1 (DCON) Out1 (DCAT);
AUXILIARY = Out1 (DU3STEP) Out1 (DE3STEP);
AUXILIARY = Out1 (BCH) Out1 (BCH);
DCON/DCAT: Continuous/Categorical outcomes. Generally works well. DU3STEP/ DE3STEP: Continuous outcomes specified as having either unequal or equal variance across profiles. Works well, but same instability than R3STEP. BCH: Works very well, for continuous outcomes. No clear winner. I use DCAT for categorical outcomes and BCH otherwise.
EQUALITY TESTS OF MEANS/PROBABILITIES ACROSS CLASSES
OUT1
Prob S.E. Odds Ratio S.E. 2.5% C.I. 97.5% C.I.
Class 1
Category 1 0.236 0.041 1.000 0.000 1.000 1.000
Category 2 0.764 0.041 4.834 2.838 0.000 10.397
Class 2
Category 1 0.254 0.036 1.000 0.000 1.000 1.000
Category 2 0.746 0.036 4.367 2.530 0.000 9.326
Class 3
Category 1 0.429 0.066 1.000 0.000 1.000 1.000
Category 2 0.571 0.066 1.984 1.210 0.000 4.354
Chi-Square P-Value Degrees of Freedom
Overall test 14.692 0.005 4
Class 1 vs. 2 0.106 0.745 1
Class 1 vs. 3 6.001 0.014 1
Class 2 vs. 3 5.072 0.024 1
DCAT
EQUALITY TESTS OF MEANS/PROBABILITIES ACROSS CLASSES
OUT2
Approximate Approximate
Mean S.E. Mean S.E.
Class 1 29.674 0.252 Class 2 30.539 0.272
Class 3 32.817 0.619
Approximate
Chi-Square P-Value Degrees of Freedom
Overall test 29.957 0.000 4
Class 1 vs. 2 5.457 0.019 1
Class 1 vs. 3 22.122 0.000 1
Class 2 vs. 3 11.350 0.001 1
DCON
EQUALITY TESTS OF MEANS ACROSS CLASSES USING THE 3-STEP
PROCEDURE WITH 4 DEGREE(S) OF FREEDOM FOR THE OVERALL TEST
OUT1
Mean S.E. Mean S.E.
Class 1 999.000 999.000 Class 2 999.000 999.000
Class 3 999.000 999.000
Chi-Square P-Value Chi-Square P-Value
Overall test 999.000 999.000 Class 1 vs. 2 999.000 999.000
Class 1 vs. 3 999.000 999.000 Class 2 vs. 3 999.000 999.000
Final Class Counts and Proportions for the Latent Class Patterns Based On
Estimated Posterior Probabilities for TURN: Step 1 vs. Step 3
Latent
Classes Step 1 Step 3
1 83.09193 0.31474 9.58211 0.03630
2 85.35850 0.32333 80.36276 0.30440
3 17.59572 0.06665 23.75866 0.08999
4 45.63659 0.17287 147.37306 0.55823
5 32.31726 0.12241 2.92341 0.01107
DU3STEP/DE3STEP
EQUALITY TESTS OF MEANS ACROSS CLASSES USING THE BCH
PROCEDURE WITH 4 DEGREE(S) OF FREEDOM FOR THE OVERALL TEST
OUT1
Mean S.E. Mean S.E.
Class 1 0.767 0.039 Class 2 0.739 0.036
Class 3 0.562 0.066
Chi-Square P-Value Chi-Square P-Value
Overall test 14.607 0.006 Class 1 vs. 2 0.248 0.618
Class 1 vs. 3 6.910 0.009 Class 2 vs. 3 5.283 0.022
OUT2
Mean S.E. Mean S.E.
Class 1 29.659 0.289 Class 2 30.530 0.293
Class 3 32.865 0.700
Chi-Square P-Value Chi-Square P-Value
Overall test 22.870 0.000 Class 1 vs. 2 3.925 0.048
Class 1 vs. 3 17.183 0.000 Class 2 vs. 3 9.046 0.003
BCH
Construct Validation
Construct validation is the only way to demonstrate
the meaningfulness of latent profile solutions
oHeuristic Value
o Theoretical Conformity and Value
o Statistical adequacy
o Shows meaningful and well-differentiated relations
to key covariates: Predictors – Correlates –
Outcomes.
oGeneralizes across samples and time points.
Construct Validation
Construct validation is the only way to demonstrate
the meaningfulness of latent profile solutions
oHeuristic Value
o Theoretical Conformity and Value
o Statistical adequacy
o Shows meaningful and well-differentiated relations
to key covariates: Predictors – Correlates –
Outcomes.
oGeneralizes across samples and time points.
Testing the Generalizability of LPA
Solutions
Construct Validation
Construct validation is the only way to demonstrate
the meaningfulness of latent profile solutions
oHeuristic Value
o Theoretical Conformity and Value
o Statistical adequacy
o Shows meaningful and well-differentiated relations
to key covariates: Predictors – Correlates –
Outcomes.
oGeneralizes across samples and time points.
Morin, A.J.S., Meyer, J.P., Creusier, J., & Biétry, F. (2016).
Multiple-group analysis of similarity in latent profile
solutions. Organizational Research Methods, 19, 231-254.
Stability and Change Across Samples and
Time
Within-Person Profile Similarity (Longitudinal)
Within-Sample (longitudinal) or Across Sample
Profile Similarity: o Configural
o Structural
o Dispersion
o Distributional
o Predictive
o Explanatory
Profile Similarity Across Samples
• Configural: o Tests if the same number of latent profiles can be identified across groups.
o Conduct the class enumeration process separately in each group.
• Structural
• Dispersion
• Distributional
• Predictive
• Explanatory
Group 1
Group 2
0
1
2
3
4
5
6
7
AC (Profile 1)NC (Profile 1)CC (Profile 1) AC (Profile 2)NC (Profile 2)CC (Profile 2)
0
1
2
3
4
5
6
7
8
AC (Profile 1)NC (Profile 1)CC (Profile 1) AC (Profile 2)NC (Profile 2)CC (Profile 2)
DATA:
FILE IS Commitment.dat;
VARIABLE:
NAMES = ID Country unit Pred1
Pred2 Cor1 Cor2 Out1 Out2 AC1
NC1 CC1 AC2 NC2 CC2;
USEVARIABLES = AC1 NC1 CC1;
MISSING = all (999);
IDVARIABLE = ID;
KNOWCLASS = cg (Status = 1 Status = 2);
CLASSES = cg (2) c (3);
Analysis:
TYPE = MIXTURE;
ESTIMATOR = MLR;
PROCESS = 3;
STARTS = 3000 100;
STITERATIONS = 100;
KNOWCLASS: identify a “known class” as one additional latent categorical variables in which the membership is “exact” rather than probabilistic. The model includes 2 latent categorical variables.
Model
%OVERALL%
c#1 on cg#1; c#2 on cg#1;
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
To indicate that class sizes differ across group (always one statement less than the number of groups).
Profile 1, Group 1
Label each parameter to indicate inequality across groups and profiles.
%cg#2.c#1%[AC1 NC1 CC1] (mm1-mm3); AC1 NC1 CC1 (vv1-vv3); %cg#2.c#2%[AC1 NC1 CC1] (mm4-mm6); AC1 NC1 CC1 (vv4-vv6); %cg#2.c#3%[AC1 NC1 CC1] (mm7-mm9); AC1 NC1 CC1 (vv7-vv9);
Profile Similarity Across Samples
• Configural
• Structural: o Prerequisite: Configural similarity.
o Tests whether the indicators’ within-profile levels are the same across groups.
• Dispersion
• Distributional
• Predictive
• Explanatory
Group 1
Group 2
0
1
2
3
4
5
6
7
AC (Profile 1)NC (Profile 1)CC (Profile 1) AC (Profile 2)NC (Profile 2)CC (Profile 2)
0
1
2
3
4
5
6
7
AC (Profile 1) NC (Profile 1) CC (Profile 1) AC (Profile 2) NC (Profile 2) CC (Profile 2)
Model
%OVERALL%
c#1 on cg#1; c#2 on cg#1;
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
Identify equal means across groups.
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (vv1-vv3);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (vv4-vv6);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (vv7-vv9);
Profile Similarity Across Samples
• Configural
• Structural
• Dispersion:o Prerequisite: Configural and structural similarity.
o Tests whether the indicators’ within-profile variability is the same across groups.
o Latent profiles do not assume that all members of a profile share the exact same configuration of indicators. This step tests whether within-profile inter-individual differences are stable across groups.
• Distributional
• Predictive
• Explanatory
Group 1
Group 2
0
1
2
3
4
5
6
7
AC (Profile 1) NC (Profile 1) CC (Profile 1) AC (Profile 2) NC (Profile 2) CC (Profile 2)
0
1
2
3
4
5
6
7
AC (Profile 1) NC (Profile 1) CC (Profile 1) AC (Profile 2) NC (Profile 2) CC (Profile 2)
Model
%OVERALL%
c#1 on cg#1; c#2 on cg#1;
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
Identify equal variances across groups.
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (v1-v3);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (v4-v6);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (v7-v9);
Profile Similarity Across Samples
• Configural
• Structural
• Dispersion
• Distributional:o Prerequisite: Configural and structural similarity.
o Tests whether the relative size of the profiles is similar across groups.
• Predictive
• Explanatory
Profile 1 Profile 2 Profile 3
Profile 1 Profile 2
Profile 3Profile 1 Profile 2 Profile 3
Group 1
Group 2
Model
%OVERALL%
c#1 on cg#1; c#2 on cg#1;
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
Take out the %overall% part
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (v1-v3);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (v4-v6);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (v7-v9);
Profile Similarity Across Samples
• Configural
• Structural
• Dispersion
• Distributional
• Predictive:o Prerequisite: Configural and structural similarity.
o This step includes predictors (direct inclusion) of profile membership to the most similar (2-4) model.
o Tests if the predictors-profiles relations are the same across groups.
o Use starts values from the retained (configural, structural, dispersion, distribution) solution to ensure stability
• Explanatory
C
S1 S2 S3 SX…
X
XC
Y
YC
Model
%OVERALL%
c#1-c#2 ON Pred1@0 Pred2@0;
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (v1-v3);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (v4-v6);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (v7-v9);
MODEL cg:%cg#1%c#1-c#2 ON Pred1 Pred2;%cg#2%c#1-c#2 ON Pred1 Pred2;
Constrain predictions to 0 in the %Overall% section to freely estimate them in both groups
FREE
Model
%OVERALL%
c#1-c#2 ON Pred1* Pred2*;
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (v1-v3);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (v4-v6);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (v7-v9);
MODEL cg:%cg#1%c#1-c#2 ON Pred1 Pred2;%cg#2%c#1-c#2 ON Pred1 Pred2;
SIMILAR
Profile Similarity Across Samples
• Configural
• Structural
• Dispersion
• Distributional
• Predictive
• Explanatory: o Prerequisite: Configural and structural similarity.
o This step includes outcomes (direct inclusion) of profile membership to the most similar (2-4) model.
o Tests if the profiles-outcomes relations are the same across groups or time points.
o Use starts values from the retained (configural, structural, dispersion, distribution) solution to ensure stability
C
S1 S2 S3 SX…
X
XC
Y
YC
Model
%OVERALL%
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
[Out1] (oa1);
[Out2] (ob1);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
[Out1] (oa2);
[Out2] (ob2);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
[Out1] (oa3);
[Out2] (ob3);
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (v1-v3); [Out1] (oaa1);[Out2] (obb1);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (v4-v6); [Out1] (oaa2);[Out2] (obb2);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (v7-v9); [Out1] (oaa3);[Out2] (obb3);
FREE
MODEL CONSTRAINT:
NEW (y12); y12 = oa1-oa2;
NEW (y13); y13 = oa1-oa3;
NEW (y23); y23 = oa2-oa3;
NEW (z12); z12 = ob1-ob2
NEW (z13); z13 = ob1-ob3;
NEW (z23); z23 = ob2-ob3;
NEW (yy12); yy12 = oaa1-oaa2;
NEW (yy13); yy13 = oaa1-oaa3;
NEW (yy23); yy23 = oaa2-oaa3;
NEW (zz12); zz12 = obb1-obb2
NEW (zz13); zz13 = obb1-obb3;
NEW (zz23); zz23 = obb2-obb3;
Model
%OVERALL%
%cg#1.c#1%
[AC1 NC1 CC1] (m1-m3);
AC1 NC1 CC1 (v1-v3);
[Out1] (oa1);
[Out2] (ob1);
%cg#1.c#2%
[AC1 NC1 CC1] (m4-m6);
AC1 NC1 CC1 (v4-v6);
[Out1] (oa2);
[Out2] (ob2);
%cg#1.c#3%
[AC1 NC1 CC1] (m7-m9);
AC1 NC1 CC1 (v7-v9);
[Out1] (oa3);
[Out2] (ob3);
%cg#2.c#1%[AC1 NC1 CC1] (m1-m3); AC1 NC1 CC1 (v1-v3); [Out1] (oa1);[Out2] (ob1);
%cg#2.c#2%[AC1 NC1 CC1] (m4-m6); AC1 NC1 CC1 (v4-v6); [Out1] (oa2);[Out2] (ob2);
%cg#2.c#3%[AC1 NC1 CC1] (m7-m9); AC1 NC1 CC1 (v7-v9); [Out1] (oa3);[Out2] (ob3);
SIMILAR
MODEL CONSTRAINT:
NEW (y12); y12 = oa1-oa2;
NEW (y13); y13 = oa1-oa3;
NEW (y23); y23 = oa2-oa3;
NEW (z12); z12 = ob1-ob2
NEW (z13); z13 = ob1-ob3;
NEW (z23); z23 = ob2-ob3;
France versus North America
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Profile 1 (CC-
Dominant)
Profile 2 (Moderately
Committed)
Profile 3 (AC/NC-
Dominant)
Profile 4 (NC-
Dominant)
Profile 5 (AC-
Dominant)
Affective Commitment (AC)
Normative Commitment (NC)
Continuance Commitment (CC)
Identical profiles were found in both cultures, with evidence of distributional differences
France versus North America
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Profile 1 (CC-
Dominant)
Profile 2 (Moderately
Committed)
Profile 3 (AC/NC-
Dominant)
Profile 4 (NC-
Dominant)
Profile 5 (AC-
Dominant)
Affective Commitment (AC)
Normative Commitment (NC)
Continuance Commitment (CC)
Confirmatory?Finch, W.H., & Bronk, K.C. (2011). Conducting confirmatory
latent class analysis using Mplus. Structural Equation Modeling, 18,
132-151.
• Using starts values and parameter constraints (equality
restrictions, deterministic restrictions [equality to a specific
value], and inequality restrictions) it is possible to rely on a
confirmatory approach to the estimation of LPA (as long as
prior research and theory is sufficiently advanced to provide
such hypotheses).
• The model constraint approaches even makes it possible to
implement proportionality constraints (e.g., twice as high), and
range restrictions constraints (e.g., between 1 and 3).
• The constrained model is then compared to the
unconstrained model in terms of fit.
Longitudinal Mixture Models
Latent Transition Analysis
C1
X1 X2 X3
C2
X1 X2 X3
Time t Time t + 1
Xi
… Xi
…
Kam, C., Morin, A.J.S., Meyer, J.P., & Topolnytsky, L. (2016).
Are commitment profiles stable and predictable? A latent
transition analysis. Journal of Management, 42 (6), 1462-1490.
Latent Transition Analysis• Latent transition analyses (LTA) are a very broad class of models
that can be used to assess the connections between any number
of latent categorical variables.
• Due to modern computer limitations, these models are typically
limited to 3, sometimes 4, latent categorical variables.
• These models can be used to assess the connections between
any type of latent categorical variable (LPA, MRA, growth
mixtures) based on the same, or different, sets of indicators
assessed at different, or similar, time points.
• The typical application of LTA is longitudinal, and assess the
connection between two LPA models based on the same set of
indicators measured at different time points.
• LTA can thus be used to assess the similarity of LPA solutions
over time.
CLASSES = c1 (4) c2 (4);
Analysis:
type = mixture;
estimator = MLR;
Process = 3;
STARTS = 10000 400;
STITERATIONs = 2000;
MODEL:
%OVERALL%
c2 on c1;
MODEL C1:
%c1#1%
[PAR8 TEA8 PEER8](ma1-ma3); PAR8 TEA8 PEER8 (va1-va3);
%c1#2%
[PAR8 TEA8 PEER8](ma4-ma6); PAR8 TEA8 PEER8 (va4-va6);
%c1#3%
[PAR8 TEA8 PEER8](ma7-ma9); PAR8 TEA8 PEER8 (va7-va9);
%c1#4%
[PAR8 TEA8 PEER8](ma10-ma12); PAR8 TEA8 PEER8 (va10-va12);
MODEL C2:
%c2#1%
[PAR11 TEA11 PEER11](mb1-mb3); PAR11 TEA11 PEER11 (vb1-vb3);
%c2#2%
[PAR11 TEA11 PEER11](mb4-mb6); PAR11 TEA11 PEER11(vb4-vb6);
%c2#3%
[PAR11 TEA11 PEER11](mb7-mb9); PAR11 TEA11 PEER11 (vb7-vb9);
%c2#4%
[PAR11 TEA11 PEER11](mb10-mb12); PAR11 TEA11 PEER11 (vb10-vb12);
FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS
PATTERNS BASED ON THE ESTIMATED MODEL
Latent Class
Pattern
1 1 5.07928 0.00202
1 2 464.55651 0.18508
1 3 0.00000 0.00000
1 4 0.00000 0.00000
2 1 105.76553 0.04214
2 2 89.60346 0.03570
2 3 623.25736 0.24831
2 4 0.00000 0.00000
3 1 8.21090 0.00327
3 2 433.19442 0.17259
3 3 57.47310 0.02290
3 4 495.70411 0.19749
4 1 43.44788 0.01731
4 2 27.33755 0.01089
4 3 156.36990 0.06230
4 4 0.00000 0.00000
FINAL CLASS COUNTS AND
PROPORTIONS FOR EACH LATENT CLASS
VARIABLE BASED ON THE ESTIMATED
MODEL
Latent Class
Variable Class
C1 1 469.63580 0.18711
2 818.62634 0.32615
3 994.58252 0.39625
4 227.15533 0.09050
C2 1 162.50359 0.06474
2 1014.69189 0.40426
3 837.10034 0.33351
4 495.70410 0.19749
• BASED ON ESTIMATED POSTERIOR PROBABILITIES• BASED ON THEIR MOST LIKELY LATENT CLASS
PATTERN
LATENT TRANSITION PROBABILITIES BASED ON THE ESTIMATED MODEL
C1 Classes (Rows) by C2 Classes (Columns)
1 2 3 4
1 0.011 0.989 0.000 0.000
2 0.129 0.109 0.761 0.000
3 0.008 0.436 0.058 0.498
4 0.191 0.120 0.688 0.000
MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value
Latent Class Pattern 1 1
Means
PAR8 -0.777 0.067 -11.608 0.000
TEA8 -0.956 0.128 -7.487 0.000
PEER8 -0.664 0.077 -8.612 0.000
PAR11 1.124 0.007 151.806 0.000
TEA11 0.785 0.126 6.215 0.000
PEER11 1.026 0.003 306.602 0.000
LTA with predictors• In tests of multi-group LPA similarity, we contrasted two models
for tests of predictive similarity:
1. One in which the effects of the predictors were freely
estimated in all samples.
2. One in which the effects of the predictors were constrained
to equality across samples.
• In LTA, we can contrast three models:
1. One in which the effects of the predictors were freely
estimated at all time points, and across profiles estimated at
the preceding time points (to predict specific profile-to-
profile transitions).
2. One in which the effects of the predictors were freely
estimated at all time points, but not initial profiles.
3. One in which the effects of the predictors were constrained
to equality across times points and initial profiles.
LTA with outcomes• Similar to tests of multi-group LPA similarity.
• Two models are contrasted for tests of explanatory similarity:
1. One in which the outcome levels are freely estimated in all
samples or time points.
2. One in which the outcome levels are constrained to equality
across samples or time points.
For applications, see:
Morin, A.J.S., & Litalien, D. (2017).Webnote: Longitudinal Tests of
Profile Similarity and Latent Transition Analyses. Montreal, QC:
Substantive Methodological Synergy Research Laboratory.
http://smslabstats.weebly.com/uploads/1/0/0/6/100647486/lta_dis
tributional_similarity_v02.pdf
Growth Mixture AnalysesGMA/GMM
Individual trajectories
Latent curve models allow one to synthesizeindividual trajectories with few latentparameters through a restricted factor model.
An average trajectory with
a random intercept and
slope is estimated
Y1 Y2 Y3 Y4
αy βy
1 1 1
1 1 2 3 0
Y5
1 4
Latent curve models allow one to synthesizeindividual trajectories with few latentparameters through a restricted factor model.
Random: Each individual has its own trajectory (intercept and
slope variance).Model-based
time-structured (trait) variability
Latent curve models allow one to synthesizeindividual trajectories with few latentparameters through a restricted factor model.
Residuals:Few persons
follow a perfectly linear
(curvilinear, etc.) trajectory.
Residual state-like variability.
Y1 Y2 Y3 Y4
αy βy
1 1 1
1 1 2 3 0
Y5
1 4
Y1 Y2 Y3 Y4
αy βy
1 1 1 1
1 0
*
Yt
1
itk itk itk itk itk
* *
C Predictors
Y1 Y2 Y3 Y4
αy βy
1 1 1 1
1 0
*
Yt
1
itk itk itk itk itk
* *
C Predictors
Growth Mixture Analyses combine Latent Profile Analyses and Latent Curve models by the addition of a latent class/profile variable to a latent curve model to allow for the identification of k subgroups following different trajectories.
In theory, all parameters can be freely estimated across classes.
Y1 Y2 Y3 Y4
αy βy
1 1 1 1
1 0
*
Yt
1
itk itk itk itk itk
* *
C Predictors
Time codes: Reflect the passage of time.
Linear: 0-1-2-3-4-5 (or some other reflecting the true passage of time, e.g. 0-1-3-4-6).
Quadratic: Addition of a second slope factor with squared time codes (0-1-4-9-16-25).
Piecewise: Two slopes with loadings reflecting the transition point: • 0-1-2-3-3-3-3• 0-0-0-0-1-2-3
Multibase: Free loadings reflecting the % of change between 0 and 1.
Morin, A.J.S., Maïano, C., Nagengast, B., Marsh, H.W., Morizot, J., &
Janosz, M. (2011). Growth Mixture Modeling of adolescents
trajectories of anxiety across adolescence: The impact of
untested invariance assumptions on substantive interpretations.
Structural Equation Modeling.
~ 1000 Montreal adolescents measured 5 times over a 4 year
period.
GMA Vs LCGALCGA (no inter-individual variability)
GMA (class specific variability that can either be invariant or freely estimated)
© Morin et al., 2010
Results: LCGA
LCGAMODEL:
%OVERALL%
i s q | ANXT1@-1 ANXT3@0 ANXT4@1 ANXT5@2 ANXT6@3;
i@0 s@0 q@0;
!No need for class-specific statements
NOTE: add the following for graphs of the trajectoriesPLOT:TYPE IS PLOT3;SERIES = ANXT1- ANXT6(*);
© Morin et al., 2010
Results: GMA, Mplus default
GMA-MDMODEL:
%OVERALL%
i s q | ANXT1@-1 ANXT3@0 ANXT4@1 ANXT5@2 ANXT6@3;
!No need for class-specific statements
© Morin et al., 2010
Results: GMA, free estimation
GGMA-LVMODEL:
%OVERALL%
i s q | ANXT1@-1 ANXT3@0 ANXT4@1 ANXT5@2 ANXT6@3;
! To specify the effects of predictors on class membership
c#1-c#4 ON SEX FAM SES FEELSE EB1 EB3 VICT1 WITN1;
! To specify the effects of predictors on the intercept factor (same thing for s and q)
I ON SEX FAM SES FEELSE EB1 EB3 VICT1 WITN1;
%c#1%
i s q; [I S Q];
i WITH s;
s WITH q;
i WITH q;
! To specify that the effects of predictors on the intercept factor are freely estimated in all classes (same thing for s and q)
I ON SEX FAM SES FEELSE EB1 EB3 VICT1 WITN1;
! Use same specifications for %c#2%, %c#3%, %c#4%, and %c#5%
Morin, A.J.S., Rodriguez, D., Fallu, J.-S., Maïano, C., & Janosz, M.
(2012). Academic Achievement and Adolescent Smoking: A
General Growth Mixture Model. Addiction.
741 non-smoking adolescents. Official GPA ratings obtained 8
times (beginning and end of school years).
And what about ?yitk
Latent curve models allow one to synthesizeindividual trajectories with few latentparameters through a restricted factor model.
Residuals:Few persons
follow a perfectly linear
(curvilinear, etc.) trajectory.
Residual state-like variability.
• Latent curve models consider these time-specific deviations from
the overall trajectories as random “errors” to be controlled
rather than substantively meaningful deviations from the generic
trajectory.
• Such deviations may indeed represent state-like “shocks” to the
overall trajectories. These “shocks” may result from meaningful
situation-specific perturbations (e.g. the death of a loved one) or
successes (e.g. admission into a highly competitive program). Such
time-specific, state-like, deviations may sometimes be quite strong
and/or vary across time and/or subpopulations,
• These deviations, when considered over time, provide a direct
estimate of state-like instability over time.
MODEL:%OVERALL%i s q | ngen1b1@0 ngen2b1@1 ngen1b2@2ngen2b2@3 ngen1b3@4 ngen2b3@5 ngen1b4@6ngen2b4@7 ;%c#1%i s q; [i s q];i WITH s q; s WITH q;ngen1b1 ngen2b1 ngen1b2 ngen2b2 ngen1b3ngen2b3 ngen1b4 ngen2b4;! Use same specifications for %c#2% and %c#3
Smoking onsets: 7.1%
Smoking onsets: 15.1%
Smoking onsets: 49.1%
A simulation study• Diallo, T.M.O, Morin, A.J.S. & Lu, H. (2016). Impact of
misspecifications of the latent variance-covariance and residual
matrices on the class enumeration accuracy of growth mixture
models. Structural Equation Modeling, 23 (4), 507-531. DOI:
10.1080/10705511.2016.1169188
• In a series of 4 studies, we contrasted population models
corresponding to the LCGA, Mplus default, or distinct latent
variance-covariance matrices, while also considering presence of
class-invariant or distinct residual structures.
• Overall, our results clearly show the advantages of relying on
freely estimated latent variance-covariance and residual
structures in each latent classes.
But this also depends on sample
size and convergence
Solinger, O.N., Van Olffen, W., Roe, R.A., & Hofmans, J. (2013) On Becoming (Un)Committed: A
Taxonomy and Test of Newcomer Onboarding Scenarios. Organization Science, 24, 1640–1661.
N = 72
Realism?• Start with a theoretically optimal model.
o The model converges
o The model converges but includes out-of-bound estimates: Constrain/fix them
o The model does not converge, or converge on a degenerate solution (many improper
estimates, empty profiles): Then you need to simplify the model.
• Possibly go with a model with equal residual and equal latent
variance-covariances to ensure that the best fitting solution
includes more than 1 profile.
• Then freely estimate either the residuals, or the latent
variance-covariances, depending on objectives.
• Models with homoscedastic residuals (equal over time) can
also be considered.
Mixture Regression / Mixture SEM
C
• Mixture regression analysis (MRA) identifies profiles of
participants differing from one another at the level of
estimated relations (regressions) between constructs.
• LPAs regroups participants based on their configuration on a
set of indicators
• MRAs identify subpopulations in which the relationships
among constructs differ.
Simply:
𝒀 = 𝒂𝒚𝒙 + 𝒃𝒚𝒙 ∗ 𝑿 + 𝜹
?
Mixture Regression
𝒀 = 𝒂𝒚𝒙 + 𝒃𝒚𝒙 ∗ 𝑿 + 𝜹
• Outcomes’ means and variance (representing the intercepts
and residuals of the regression) need to be freely estimated
across profiles to obtain profile-specific regression equations.
• More complex applications are possible, such as combining
LPAs and MRAs to create profiles based on both their
configuration of indicators, and relationships between
indicators.
• The estimation of this type of MRAs requires the free
estimation of the predictors’ means and variances in the
extracted profiles. Such models also reveal potential
interactions among predictors, resulting in profiles in which
the relations among constructs may differ as a function of
predictors’ levels.
%OVERALL%
AC2 ON AC1 NC1 CC1;
%c#1%
AC2 ON AC1 NC1 CC1;
AC2;
[AC2];
%c#2%
AC2 ON AC1 NC1 CC1;
AC2;
[AC2];
%c#3%
AC2 ON AC1 NC1 CC1;
AC2;
[AC2];
%OVERALL%
AC2 ON AC1 NC1 CC1;
%c#1%
AC2 ON AC1 NC1 CC1;
AC2 AC1 NC1 CC1;
[AC2 AC1 NC1 CC1];
%c#2%
AC2 ON AC1 NC1 CC1;
AC2 AC1 NC1 CC1;
[AC2 AC1 NC1 CC1];
%c#3%
AC2 ON AC1 NC1 CC1;
AC2 AC1 NC1 CC1;
[AC2 AC1 NC1 CC1];
Gillet, N., Morin, A.J.S., Sandrin, E., & Houle, S.A.
Upcoming !
Workaholism
Work Engagement
Sleep Difficulties
Work-Family Conflicts
Engaged
Workaholic
Engaged-Workaholic
Profile 1 Profile 2 Profile 3
Engagement –> Work-Family
Conflicts.041 (.086) .004 (.241) -.015 (.171)
Engagement –> Sleeping
Difficulties.023 (.095) .155 (.154) .417 (.316)
Workaholism –> Work-Family
Conflicts .550 (.059)** .440 (.153)* .582 (.097)**
Workaholism –> Sleeping
Difficulties.244 (.121)* .287 (.171) -.293 (.257)
Chénard-Poirier, L.-A., Morin, A.J.S., & Boudrias, J.-S. (In
Press). On the merits of coherent leadership
empowerment behaviors: A mixture regression
approach. Journal of Vocational Behavior.
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Profile 1 Profile 2 Profile 3 Profile 4
Delegation
Coaching
Recognition
Efficacy-Tasks
Improvement-Tasks
Collaboration-Group
Improvement-Group
Involvement-Organization
Efficacy-Tasks Improvement-Tasks Involvement-Org.
Predictor b (s.e) β b (s.e) β b (s.e) β
Profile 1
Delegation 0.05 (0.01)** .71 -0.08 (0.05) -.27 0.131 (0.24) .13 Coaching 0.01 (0.01) .15 0.15 (0.07)* .50 -0.057 (0.37) -.05
Recognition -0.02 (0.01)* -.28 -0.03 (0.03) -.11 0.100 (0.21) .10
Profile 2 Delegation 0.27 (0.13)* .51 0.25 (0.19) .34 0.09 (0.09) .32
Coaching -0.08 (0.12) -.15 0.22 (0.20) .31 0.06 (0.07) .21
Recognition -0.13 (0.13) -.24 -0.37 (0.15)* -.49 -0.14 (0.06)* -.48
Profile 3
Delegation 0.33 (0.11)** .51 0.13 (0.08) .28 0.19 (0.12) .31
Coaching -0.24 (0.09)** -.37 0.01 (0.08) .02 0.08 (0.11) .13 Recognition 0.12 (0.08) .20 0.02 (0.07) .04 -0.09 (0.12) -.15
Profile 4
Delegation 0.36 (0.17)* .31 -0.12 (0.18) -.12 0.01 (0.10) .02 Coaching 0.61 (0.18)** .46 0.54 (0.21)** .48 -0.12 (0.15) -.15
Recognition -0.47 (0.17)** -.39 -0.23 (0.22) -.22 0.22 (0.13)* .30
Efficacy-Tasks Improvement-Tasks Involvement-Org.
Predictor b (s.e) β b (s.e) β b (s.e) β
Profile 1
Delegation 0.05 (0.01)** .71 -0.08 (0.05) -.27 0.131 (0.24) .13 Coaching 0.01 (0.01) .15 0.15 (0.07)* .50 -0.057 (0.37) -.05
Recognition -0.02 (0.01)* -.28 -0.03 (0.03) -.11 0.100 (0.21) .10
Profile 2 Delegation 0.27 (0.13)* .51 0.25 (0.19) .34 0.09 (0.09) .32
Coaching -0.08 (0.12) -.15 0.22 (0.20) .31 0.06 (0.07) .21
Recognition -0.13 (0.13) -.24 -0.37 (0.15)* -.49 -0.14 (0.06)* -.48
Profile 3
Delegation 0.33 (0.11)** .51 0.13 (0.08) .28 0.19 (0.12) .31
Coaching -0.24 (0.09)** -.37 0.01 (0.08) .02 0.08 (0.11) .13 Recognition 0.12 (0.08) .20 0.02 (0.07) .04 -0.09 (0.12) -.15
Profile 4
Delegation 0.36 (0.17)* .31 -0.12 (0.18) -.12 0.01 (0.10) .02 Coaching 0.61 (0.18)** .46 0.54 (0.21)** .48 -0.12 (0.15) -.15
Recognition -0.47 (0.17)** -.39 -0.23 (0.22) -.22 0.22 (0.13)* .30
Efficacy-Tasks Improvement-Tasks Involvement-Org.
Predictor b (s.e) β b (s.e) β b (s.e) β
Profile 1
Delegation 0.05 (0.01)** .71 -0.08 (0.05) -.27 0.131 (0.24) .13 Coaching 0.01 (0.01) .15 0.15 (0.07)* .50 -0.057 (0.37) -.05
Recognition -0.02 (0.01)* -.28 -0.03 (0.03) -.11 0.100 (0.21) .10
Profile 2 Delegation 0.27 (0.13)* .51 0.25 (0.19) .34 0.09 (0.09) .32
Coaching -0.08 (0.12) -.15 0.22 (0.20) .31 0.06 (0.07) .21
Recognition -0.13 (0.13) -.24 -0.37 (0.15)* -.49 -0.14 (0.06)* -.48
Profile 3
Delegation 0.33 (0.11)** .51 0.13 (0.08) .28 0.19 (0.12) .31
Coaching -0.24 (0.09)** -.37 0.01 (0.08) .02 0.08 (0.11) .13 Recognition 0.12 (0.08) .20 0.02 (0.07) .04 -0.09 (0.12) -.15
Profile 4
Delegation 0.36 (0.17)* .31 -0.12 (0.18) -.12 0.01 (0.10) .02 Coaching 0.61 (0.18)** .46 0.54 (0.21)** .48 -0.12 (0.15) -.15
Recognition -0.47 (0.17)** -.39 -0.23 (0.22) -.22 0.22 (0.13)* .30
Efficacy-Tasks Improvement-Tasks Involvement-Org.
Predictor b (s.e) β b (s.e) β b (s.e) β
Profile 1
Delegation 0.05 (0.01)** .71 -0.08 (0.05) -.27 0.131 (0.24) .13 Coaching 0.01 (0.01) .15 0.15 (0.07)* .50 -0.057 (0.37) -.05
Recognition -0.02 (0.01)* -.28 -0.03 (0.03) -.11 0.100 (0.21) .10
Profile 2 Delegation 0.27 (0.13)* .51 0.25 (0.19) .34 0.09 (0.09) .32
Coaching -0.08 (0.12) -.15 0.22 (0.20) .31 0.06 (0.07) .21
Recognition -0.13 (0.13) -.24 -0.37 (0.15)* -.49 -0.14 (0.06)* -.48
Profile 3
Delegation 0.33 (0.11)** .51 0.13 (0.08) .28 0.19 (0.12) .31
Coaching -0.24 (0.09)** -.37 0.01 (0.08) .02 0.08 (0.11) .13 Recognition 0.12 (0.08) .20 0.02 (0.07) .04 -0.09 (0.12) -.15
Profile 4
Delegation 0.36 (0.17)* .31 -0.12 (0.18) -.12 0.01 (0.10) .02 Coaching 0.61 (0.18)** .46 0.54 (0.21)** .48 -0.12 (0.15) -.15
Recognition -0.47 (0.17)** -.39 -0.23 (0.22) -.22 0.22 (0.13)* .30
%OVERALL%
CONSC AMELT COLL AMELG IMPL ON DELEG COACH RECON ;
CONSC AMELT COLL AMELG IMPL DELEG COACH RECON;
[CONSC AMELT COLL AMELG IMPL DELEG COACH RECON];
%c#1%
CONSC AMELT COLL AMELG IMPL ON DELEG COACH RECON ;
CONSC AMELT COLL AMELG IMPL DELEG COACH RECON;
[CONSC AMELT COLL AMELG IMPL DELEG COACH RECON];
%c#2%
CONSC AMELT COLL AMELG IMPL ON DELEG COACH RECON ;
CONSC AMELT COLL AMELG IMPL DELEG COACH RECON;
[CONSC AMELT COLL AMELG IMPL DELEG COACH RECON];
%c#3%
CONSC AMELT COLL AMELG IMPL ON DELEG COACH RECON ;
CONSC AMELT COLL AMELG IMPL DELEG COACH RECON;
[CONSC AMELT COLL AMELG IMPL DELEG COACH RECON];
%c#4%
CONSC AMELT COLL AMELG IMPL ON DELEG COACH RECON ;
CONSC AMELT COLL AMELG IMPL DELEG COACH RECON;
[CONSC AMELT COLL AMELG IMPL DELEG COACH RECON];
Organizational Research Methods: Special
Issue on Person-Centered Analyses
• Confirmatory LPA
• Multilevel LPA
• Extended Capabilities through Manual Auxiliary Approaches
(mediation, moderation, etc.).
• K-Centres functional clustering for longitudinal data
• Among others