Advances in Mixture Modeling And More

Advances in Mixture Modeling And More Bengt Muthén & Tihomir Asparouhov Mplus www.statmodel.com bmuthen@statmodel.com Keynote address at IMPS 14, Madison, Wisconsin, July 22, 14 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 1/ 48

Overview of Latent Variable Models Continuous latent Categorical latent variables variables Hybrids Cross-sectional Factor analysis, SEM Regression mixture analysis, Factor mixture models Latent class analysis analysis Longitudinal Growth analysis Latent transition analysis, Growth mixture models (random effects) Latent class growth analysis analysis Source: Muthén, B. (08). Latent variable hybrids: Overview of old and new models. In Hancock, G. R., & Samuelsen, K. M. (Eds.), Advances in latent variable mixture models, pp. 1 24. Charlotte, NC: Information Age Publishing, Inc. Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 2/ 48

Characterizing Mixture Models Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 3/ 48

Characterizing Growth (Mixture) Models Growth Modeling Paradigms (Continued) HLM (Raudenbush) Outcome Outcome GMM (Muthén) Outcome LCGA (Nagin) Escalating Early Onset Escalating Early Onset Normative Normative 15 40 Age 15 40 Age 15 40 Age y t+ 1 y t+ 1 y t+ 1 y t y t y t Replace rhetoric with statistics let likelihood decide HLM and LCGA are special cases of GMM 126 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 4/ 48

Growth Mixture Modeling (GMM) References Muthén, B. & Shedden, K. (1999). Finite mixture modeling with mixture outcomes using the EM algorithm. Biometrics Muthén, B., Brown, C.H., Masyn, K., Jo, B., Khoo, S.T., Yang, C.C., Wang, C.P., Kellam, S., Carlin, J., & Liao, J. (02). General growth mixture modeling for randomized preventive interventions. Biostatistics Muthén, B. & Asparouhov, T. (09). Growth mixture modeling: Analysis with non-gaussian random effects. In Fitzmaurice, G., Davidian, M., Verbeke, G. & Molenberghs, G. (eds.), Longitudinal Data Analysis. Boca Raton: Chapman & Hall Muthén & Asparouhov (14). Growth mixture modeling with non-normal distributions. Submitted Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 5/ 48

2. Today s Focus: Skewed Distributions Examples: Body Mass Index (BMI) in obesity studies (long right tail) Mini Mental State Examination (MMSE) cognitive test in Alzheimer s studies (long left tail) PSA scores in prostate cancer studies (long right tail) CD4 cell counts in AIDS studies (long right tail) Ham-D score in antidepressant studies (long right tail) Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 6/ 48

Body Mass Index (BMI): kg/m 2 Normal 18 < BMI < 25, Overweight 25 < BMI < 30, Obese > 30 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 7/ 48

BMI at Age 15 in the NLSY (Males, n = 3194) Count 2 210 0 190 180 170 160 150 140 130 1 110 100 90 80 70 60 50 40 30 10 0 Descriptive statistics for BMI15_2: n = 3194 Mean: 23.104 Min: 13.394 Variance:.068 %-tile: 19.732 Std dev.: 4.480 40%-tile: 21.142 Skewness: 1.475 Median: 22.045 Kurtosis: 3.068 60%-tile: 22.955 % with Min: 0.03% 80%-tile: 25.840 % with Max: 0.03% Max: 49.868 14 14 15 16 16 17 18 19 19 21 22 22 23 24 25 25 26 27 27 28 29 30 30 31 32 33 33 34 35 35 36 37 38 38 39 40 41 41 42 43 43 44 45 46 46 47 48 49 49 BMI15_2 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 8/ 48

3. Mixtures for Male BMI at Age 15 in the NLSY Skewness = 1.5, kurtosis = 3.1 Mixtures of normals with 1-4 classes have BIC = 18,658, 17,697, 17,638, 17,637 (tiny class) 3-class mixture shown below Distribution 0.22 0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 11 12 13 14 15 16 17 18 19 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 Class 1, 35.8% Class 2, 53.2% Class 3, 11.0% Overall 41 42 43 44 45 46 47 48 49 50 51 52 BMI15_2 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 9/ 48

Several Classes or One Non-Normal Distribution? Pearson (1895) Hypertension debate: Platt (1963): Hypertension is a disease (separate class) Pickering (1968): Hypertension is merely the upper tail of a skewed distribution Schork & Schork (1988): Two-component mixture versus lognormal Bauer & Curran (03): Growth mixture modeling classes may merely reflect a non-normal distribution so that classes have no substantive meaning Muthén (03) comment on BC: Substantive checking of classes related to antecedents, concurrent events, consequences (distal outcomes), and usefulness Multivariate case more informative than univariate Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 10/ 48

What If We Could Instead Fit The Data With a Skewed Distribution? Then a mixture would not be necessitated by a non-normal distribution, but a single class may be sufficient A mixture of non-normal distributions is possible 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 11 12 13 14 15 16 17 18 19 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 11/ 48

4. Introducing Mixtures of Non-Normal Distributions in Mplus Version 7.2 In addition to a mixture of normal distributions, it is now possible to use T: Adding a degree of freedom (df) parameter (thicker tails) Skew-normal: Adding a skew parameter to each variable Skew-T: Adding skew and df parameters (stronger skew possible than skew-normal) References Azzalini (1985), Azzalini & Dalla Valle (1996): skew-normal Arellano-Valle & Genton (10): extended skew-t McNicholas, Murray, 13, 14: skew-t as a special case of the generalized hyperbolic distribution McLachlan, Lee, Lin, 13, 14: restricted and unrestricted skew-t Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 12/ 48

Skew T-Distribution Formulas Y can be seen as the sum of a mean, a part that produces skewness, and a part that adds a symmetric distribution: Y = µ + δ U 0 + U 1, where U 0 has a univariate t and U 1 a multivariate t distribution. Expectation, variance (δ is a skew vector, ν the df): ν 1 Γ( 2 E(Y) = µ + δ ) ν Γ( ν 2 ) π, ( Var(Y) = ν Γ( ν 1 ν 2 (Σ + δδt 2 ) ) ) 2 ν Γ( ν 2 ) π δδt Marginal and conditional distributions: Marginal is also a skew-t distribution Conditional is an extended skew-t distribution; conditional expectation function not always linear; conditional variance not always homoscedastic Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 13/ 48

Examples of Skew-T Distributions 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 Percentiles for I in Class 2: 5%-tile: 18.026 10%-tile: 18.656 25%-tile: 19.916 50%-tile: 21.717 75%-tile: 24.327 90%-tile: 27.658 95%-tile: 30.448 0.05 0.04 0.03 0.02 0.01 0 13.705 14.705 15.705 16.705 17.705 18.705 19.705.705 21.705 22.705 23.705 24.705 25.705 26.705 27.705 28.705 29.705 30.705 31.705 32.705 33.705 34.705 35.705 36.705 37.705 38.705 39.705 40.705 41.705 42.705 43.705 44.705 45.705 46.705 47.705 48.705 49.705 50.705 51.705 52.705 53.705 13.705 14.705 15.705 16.705 17.705 18.705 19.705.705 21.705 22.705 23.705 24.705 25.705 26.705 27.705 28.705 29.705 30.705 31.705 32.705 33.705 34.705 35.705 36.705 37.705 38.705 39.705 40.705 41.705 42.705 43.705 44.705 45.705 46.705 47.705 48.705 49.705 50.705 51.705 52.705 53.705 0.11 0.1 0.09 0.08 0.07 0.06 Percentiles for I in Class 1: 5%-tile: 22.169 10%-tile: 22.586 25%-tile: 23.835 50%-tile: 26.333 75%-tile: 30.2 90%-tile: 35.494 95%-tile: 39.936 0.05 0.04 0.03 0.02 0.01 0 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 14/ 48

BMI at Age 15 in the NLSY (Males, n = 3194) Skewness = 1.5, kurtosis = 3.1 Mixtures of normals with 1-4 classes: BIC = 18,658, 17,697, 17,638, 17,637 (tiny class). 3-class model uses 8 parameters 1-class Skew-T distribution: BIC = 17,623 (2-class BIC = 17,638). 1-class model uses 4 parameters 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 11 12 13 14 15 16 17 18 19 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 15/ 48

4.1 Cluster Analysis with Non-Normal Mixtures Australian Institute of Sports Data: BMI and BFAT (n = 2) AIS data often used in the statistics literature to illustrate quality of cluster analysis using mixtures, treating gender as unknown Murray, Brown & McNicholas forthcoming in Computational Statistics & Data Analysis: Mixtures of skew-t factor analyzers : How well can we identify cluster (latent class) membership based on BMI and BFAT? Compared to women, men have somewhat higher BMI and somewhat lower BFAT Non-normal mixture models with unrestricted means, variances, covariance Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 16/ 48

Australian Sports Institute Data on BMI and BFAT (n = 2) Table : Comparing classes with unknown gender Normal 2c: LL = -1098, # par. s = 11, BIC = 2254 Class 1 Class 2 Male 78 24 Female 0 100 Skew-Normal 2c: LL = -1069, # par. s = 15, BIC = 2218 Class 1 Class 2 Male 84 18 Female 1 99 Skew-T 2c: LL = -1068, # par. s = 17, BIC = 2227 Class 1 Class 2 Male 95 7 Female 2 98 Normal 3c: LL = -1072, # par. s = 17, BIC = 2234 Class 1 Class 2 Class 3 Male 85 1 16 Female 1 14 85 T 2c: LL = -1090, # par. s = 13, BIC = 2250 Class 1 Class 2 Male 95 7 Female 2 98 PMSTFA 2c: BIC = 2224 Class 1 Class 2 Male 97 5 Female 5 95 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 17/ 48

Cluster Analysis by Mixtures of Factor Analyzers (McLachlan) Reduces the number of µ c, Σ c parameters for c = 1,2,...C by applying the Σ c structure of an EFA with orthogonal factors: Σ c = Λ c Λ c + Θ c (1) This leads to 8 variations by letting Λ c and Θ c be invariant or not across classes and letting Θ c have equality across variables or not (McNicholas & Murphy, 08). Interest in clustering as opposed to the factors, e.g. for genetic applications. (EFA mixtures not yet available in Mplus for non-normal distributions, but can be done using EFA-in-CFA.) Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 18/ 48

4.2 Non-Normal Mixtures of Latent Variable Models Models: Choices: Mixtures of Exploratory Factor Models (McLachlan, Lee, Lin; McNicholas, Murray) Mixtures of Confirmatory Factor Models; FMM (Mplus) Mixtures of SEM (Mplus) Mixtures of Growth Models; GMM (Mplus) Intercepts, slopes (loadings), and residual variances invariant? Scalar invariance (intercepts, loadings) allows factor means to vary across classes instead of intercepts (not typically used in mixtures of EFA, but needed for GMM) Skew for the observed or latent variables? Implications for the observed means. Latent skew suitable for GMM - the observed variable means are governed by the growth factor means Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 19/ 48

4.3 Growth Mixture Modeling of NLSY BMI Age 12 to 23 for Black Females (n = 1160) Normal BIC: 31684 (2c), 31386 (3c), 31314 (4c), 31338 (5c) Skew-T BIC: 31411 (1c), 31225 (2c), 31270 (3c) 60 58 56 54 52 50 48 46 44 42 40 38 36 34 32 30 28 26 24 22 18 16 14 12 10 8 Class 2, 45.8% Class 1, 54.2% 6 4 2 0 12 13 14 15 16 17 18 19 21 22 23 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling / 48

2-Class Skew-T versus 4-Class Normal 60 58 56 54 52 50 48 46 44 42 40 38 36 34 32 30 28 26 24 22 18 16 14 12 10 8 6 4 2 0 Class 2, 45.8% Class 1, 54.2% 12 13 14 15 16 17 18 19 21 22 23 60 55 50 45 40 35 30 25 15 10 5 Class 1, 18.1% Class 2, 42.3% Class 3, 30.1% Class 4, 9.6% 0 12 13 14 15 16 17 18 19 21 22 23 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 21/ 48

2-Class Skew-T: Estimated Percentiles (Note: Not Growth Curves) 60 60 55 50 45 40 Percentiles shown: 5th, 10th, 25th, 50th, 75th, 90th, and 95th 55 50 45 40 Percentiles shown: 5th, 10th, 25th, 50th, 75th, 90th, and 95th 35 35 30 30 25 25 15 15 10 10 5 5 0 12 13 14 15 16 17 18 19 21 22 23 0 12 13 14 15 16 17 18 19 21 22 23 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 22/ 48

2-Class Skew-T: Intercept Growth Factor (Age 17) 0.21 0.2 0.19 0.18 0.17 0.16 0.15 Distribution 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 Percentiles for I in Class 1 (low class): 5%-tile: 18.026 10%-tile: 18.656 25%-tile: 19.916 50%-tile: 21.717 75%-tile: 24.327 90%-tile: 27.658 95%-tile: 30.448 0.01 0 10 12 14 16 18 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 I 0.21 0.2 0.19 0.18 0.17 0.16 Distribution 0.15 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 Percentiles for I in Class 2 (high class): 5%-tile: 22.169 10%-tile: 22.586 25%-tile: 23.835 50%-tile: 26.333 75%-tile: 30.2 90%-tile: 35.494 95%-tile: 39.936 0.02 0.01 0 10 12 14 16 18 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 I Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 23/ 48

Regressing Class on a MOMED Covariate ( c ON x ): 2-Class Skew-T versus 4-Class Normal 0.8 0.8 0.6 0.6 Class 1, 22.0% Class 2, 38.7% Class 3, 28.7% Class 4, 10.6% Probability 0.4 Probability 0.4 Class 1, 44.4% 0.2 Class 2, 55.6% 0.2 0-2.5-1.5-0.5 0.5 1.5 2.5 3.5 4.5 0-2.5-1.5-0.5 0.5 1.5 2.5 3.5 4.5 MOMED MOMEDU Recall the estimated trajectory means for skew-t versus normal: 60 60 58 56 54 55 52 50 50 48 46 44 45 42 40 40 38 36 34 35 32 30 28 26 24 30 25 22 18 16 14 12 10 8 6 4 Class 2, 45.8% Class 1, 54.2% 15 10 5 Class 1, 18.1% Class 2, 42.3% Class 3, 30.1% Class 4, 9.6% 2 0 12 13 14 15 16 17 18 19 21 22 23 0 12 13 14 15 16 17 18 19 21 22 23 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 24/ 48

4.4 GMM of BMI in the Framingham Data: Females Ages 25 to 65 Classic data set Different age range: 25 to 65 Individually-varying times of observations Quadratic growth mixture model Normal distribution BIC is not informative: 16557 (1c), 15995 (2c), 15871 (3c), 15730 (4c), 15674 (5c) Skew-T distribution BIC points to 3 classes: 15611 (1c), 15327 (2c), 15296 (3c), 15304 (4c) Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 25/ 48

Framingham BMI, Females Ages 25 to 65: 3-Class Skew-T 45 40 13% 33% 54% 35 30 25 15 30 40 50 60 70 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 26/ 48

BMI and Treatment for High Blood Pressure BMI = kg/m 2 with normal range 18.5 to 25, overweight 25 to 30, obese > 30 Risk of developing heart disease, high blood pressure, stroke, diabetes Framingham data contains data on blood pressure treatment at each measurement occasion Survival component for the first treatment can be added to the growth mixture model with survival as a function of trajectory class Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 27/ 48

Parallel Process Model with Growth Mixture for BMI and Survival to First Blood Pressure Treatment bmi25 bmi26 bmi65 i s q c f rx25 rx26 rx65 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 28/ 48

Parallel Process Model with Growth Mixture for BMI and Survival to First Blood Pressure Treatment: Survival Curves for the Three Trajectory Classes 1 0.8 Probability 0.6 0.4 0.2 35.0% 11.6% 53.3% 0 26 32 38 44 50 56 62 68 Age Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 29/ 48

4.5 Growth Mixture Modeling: Math Achievement Trajectory Classes and High School Dropout. An Example of Substantive Checking via Predictive Validity Poor Development: % Moderate Development: 28% Good Development: 52% Math Achievement 40 60 80 100 40 60 80 100 40 60 80 100 7 8 9 10 Grades 7-10 7 8 9 10 Grades 7-10 7 8 9 10 Grades 7-10 Dropout: 69% 8% 1% Source: Muthén (03). Statistical and substantive checking in growth mixture modeling. Psychological Methods. - Does the normal mixture solution hold up when checking with non-normal mixtures? Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 30/ 48

Growth Mixture Modeling: Math Achievement Trajectory Classes and High School Dropout math7 math8 math9 math10 i s High School Dropout c Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 31/ 48

Growth Mixture Modeling: Math Achievement Trajectory Classes and High School Dropout 60 Count 55 50 45 40 35 30 25 15 10 5 Descriptive statistics for MATH10: n = 40 Mean: 63.574 Min: 29.600 Variance: 186.295 %-tile: 51.430 Std dev.: 13.649 40%-tile: 61.810 Skewness: -0.317 Median: 65.305 Kurtosis: -0.467 60%-tile: 68.400 % with Min: 0.05% 80%-tile: 75.280 % with Max: 0.25% Max: 95.170 (1076 missing cases were not included.) 0 29.9279 31.2393 32.5507 33.8621 35.1735 36.4849 37.7963 39.1077 40.4191 41.7304 43.0418 44.3532 45.6646 46.976 48.2874 49.5988 50.9102 52.2216 53.533 54.8444 56.1558 57.4672 58.7786 60.09 61.4014 62.7128 64.0242 65.3356 66.647 67.9584 69.2698 70.5812 71.8926 73.4 74.5154 75.8268 77.1382 78.4496 79.761 81.0724 82.3838 83.6952 85.0066 86.318 87.6294 88.9408 90.2522 91.5636 92.875 94.1864 MATH10 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 32/ 48

Growth Mixture Modeling: Math Achievement Trajectory Classes and High School Dropout Best solutions, 3 classes (LL, no. par s, BIC): Normal distribution: -34459, 32, 69175 T distribution: -34453, 35, 69188 Skew-normal distribution: -34442, 38, 69191 Skew-t distribution: -34439, 42, 697 Percent in low, flat class and odds ratios for dropout vs not, comparing low, flat class with the best class: Normal distribution: 18 %, OR = 17.1 T distribution: 19 %, OR =.6 Skew-normal distribution: 26 %, OR = 23.8 Skew-t distribution: 26 %, OR = 37.3 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 33/ 48

4.7 Disadvantages of Non-Normal Mixture Modeling Much slower computations than normal mixtures, especially for large sample sizes Needs larger samples; small class sizes can create problems (but successful analyses can be done at n = 100-0) Needs more random starts than normal mixtures to replicate the best loglikelihood Lower entropy Needs continuous variables Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 34/ 48

Advantages of Non-Normal Mixture Modeling Non-normal mixtures Can fit the data considerably better than normal mixtures Can use a more parsimonious model Can reduce the risk of extracting latent classes that are merely due to non-normality of the outcomes Can check the stability/reproducibility of a normal mixture solution Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 35/ 48

5. SEM Allowing Non-Normal Distributions Non-Normal SEM with t, skew-normal, and skew-t distributions: Allowing a more general model, including non-linear conditional expectation functions and heteroscedasticity Chi-square test of model fit using information on skew and df Missing data handling avoiding the normality assumption of FIML Mediation modeling allowing general direct and indirect effects Percentile estimation of the skewed factor distributions Asparouhov & Muthén (14). Structural equation models and mixture models with continuous non-normal skewed distributions. Forthcoming in Structural Equation Modeling. Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 36/ 48

ML Estimation: Normality Violations in SEM ML estimate robustness to non-normality in SEM SEs and chi-square can be adjusted for non-normality (sandwich estimator, Satorra-Bentler ) MLE robustness doesn t hold if residuals and factors are not independent (Satorra, 02) as with residual skew 0 Asparouhov-Muthén (14): There is a preconceived notion that standard structural models are sufficient as long as the standard errors of the parameter estimates are adjusted for failure of the normality assumption, but this is not really correct. Even with robust estimation the data is reduced to means and covariances. Only the standard errors of the parameter estimates extract additional information from the data. The parameter estimates themselves remain the same, i.e., the structural model is still concerned with fitting only the means and the covariances and ignoring everything else. Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 37/ 48

Non-Normal SEM where and Y = ν + Λη + ε η = α + Bη + ΓX + ξ (ε,ξ ) rmst(0,σ 0,δ,DF) ( Θ 0 Σ 0 = 0 Ψ The vector of parameters δ is of size P + M and can be decomposed as δ = (δ Y,δ η ). From the above equations we obtain the conditional distributions η X rmst((i B) 1 (α +ΓX),(I B) 1 Ψ((I B) 1 ) T,(I B) 1 δ η,df) ). Y X rmst(µ,σ,δ 2,DF) Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 38/ 48

5.1 Path Analysis of Firefighter Data (n = 354) M X Y Sources: X represents randomized exposure to an intervention M is knowledge of the benefits of eating fruits and vegetables Y is reported eating of fruits and vegetables Yuan & MacKinnon (09). Bayesian mediation analysis. Psychological Methods Elliot et al. (07). The PHLAME (Promoting Healthy Lifestyles: Alternative Models Effects) firefighter study: outcomes of two models of behavior change. Journal of Occupational and Environental Medicine Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 39/ 48

Firefighter M and Y Distributions 100 Count 90 80 70 60 50 40 30 Descriptive statistics for M: n = 354 Mean: 0.000 Min: -5.166 Variance: 1.244 %-tile: -0.832 Std dev.: 1.115 40%-tile: -0.166 Skewness: -0.261 Median: -0.166 Kurtosis: 2.289 60%-tile: 0.168 % with Min: 0.28% 80%-tile: 0.834 % with Max: 0.28% Max: 3.834 10 0-4.94073-4.49073-4.04073-3.59073-3.14073-2.69073-2.24073-1.79073-1.34073-0.890725-0.440725 0.00927491 0.459275 0.909275 1.35927 1.80927 2.25927 2.70927 3.15927 3.60928 M 60 55 50 45 40 35 Descriptive statistics for Y: n = 354 Mean: 0.418 Min: -3.000 Variance: 1.159 %-tile: -0.333 Std dev.: 1.076 40%-tile: 0.000 Skewness: 0.233 Median: 0.333 Kurtosis: 0.692 60%-tile: 0.667 % with Min: 0.28% 80%-tile: 1.333 % with Max: 0.56% Max: 4.000 30 25 15 10 5 0-2.825-2.475-2.125-1.775-1.425-1.075-0.725-0.375-0.025 0.325 0.675 1.025 Count 1.375 1.725 2.075 2.425 2.775 3.125 3.475 3.825 Y Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 40/ 48

Summary Comparison of Models for the Firefighter Example Distribution LogL Number of parameters BIC Normal -1058 7 2157 t-dist -1045 8 (adding df) 2137 Skew-normal -1055 9 (adding 2 skew) 2162 Skew-t -1043 10 (adding df and 2 skew) 2144 The df parameter of the t-distribution is needed to capture the kurtosis, but skew parameters are not needed The t-distribution allows for heteroscedasticity in the Y residual as a function of M; the conditional expectation functions are linear; usual indirect effect valid Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 41/ 48

Comparing Normal and T-Distribution Estimates Normal distribution - regular SEM Two-Tailed Estimate S.E. Est./S.E. P-Value M ON X 0.397 0.119 3.346 0.001 Y ON M 0.142 0.051 2.755 0.006 X 0.108 0.116 0.926 0.354 Intercepts Y 0.418 0.056 7.417 0.000 M 0.000 0.058 0.000 1.000 Residual Variances Y 1.125 0.085 13.304 0.000 M 1.3 0.090 13.304 0.000 New/Additional Parameters INDIRECT 0.056 0.026 2.127 0.033 T-distribution M ON X 0.371 0.110 3.384 0.001 Y ON M 0.119 0.059 2.003 0.045 X 0.134 0.115 1.161 0.246 Intercepts Y 0.384 0.055 6.940 0.000 M 0.005 0.053 0.093 0.926 Residual Variances Y 0.872 0.088 9.963 0.000 M 0.829 0.092 9.006 0.000 Skew and Df Parameters DF 7.248 1.851 3.915 0.000 New/Additional Parameters INDIRECT 0.044 0.026 1.718 0.086 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 42/ 48

5.2 Path Analysis Mediation Model for BMI ( n = 3839) bmi12 momed bmi17 bmi12 skewness/kurtosis = 1.34/2.77 bmi17 skewness/kurtosis = 1.86/5.29 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 43/ 48

Summary Comparison of Models for the BMI Example Distribution LogL Number of parameters BIC Normal -117 7 22471 t-dist -10789 8 (adding df) 21642 Skew-normal -10611 9 (adding 2 skew) 21295 Skew-t -10423 10 (adding df and 2 skew) 928 Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 44/ 48

Regression of BMI17 on BMI12: Skew-T vs Normal Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 45/ 48

Skew-T versus Normal Results The skew-t model can be further improved by letting the BMI skew be a function of mother s education The modified skew-t model shows that the direct effect is 85% of the total effect while the indirect effect is only 15% of the total effect The normal model finds that the direct effect is ignorable relative to the indirect effect These drastically different results illustrate the modeling opportunities when we look beyond the mean and variance modeling used with standard SEM Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 46/ 48

5.3 ACE Twin Modeling using Skew-T Twin 1 Twin 2 y1 y2 a c e a c e A1 C1 E1 A2 C2 E2 How will heritability estimates be affected? For which variables should the skew parameters be applied? How should growth mixtures be incorporated: Different heritability for different trajectory classes; is there genetic influence on trajectory class membership? Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 47/ 48

Advances in Mixture Modeling And More Bengt Muthén & Tihomir Asparouhov Mplus www.statmodel.com bmuthen@statmodel.com Bengt Muthén & Tihomir Asparouhov Advances in Mixture Modeling 48/ 48