LMS and QML approaches
lms_qml.RmdThe Latent Moderated Structural Equations (LMS) and the Quasi Maximum Likelihood (QML) Approach
Both the LMS and QML approaches work on
most models, but interaction effects with endogenous variables can be
tricky to estimate (see the vignette).
Both approaches, particularly the LMS approach, are
computationally intensive and are partially implemented in C++ (using
Rcpp and RcppArmadillo). Additionally,
starting parameters are estimated using the double-centering approach,
and the means of the observed variables are used to generate good
starting parameters for faster convergence. If you want to monitor the
progress of the estimation process, you can use
verbose = TRUE.
A Simple Example
Here is an example of the LMS approach for a simple
model. By default, the summary() function calculates fit
measures compared to a null model (i.e., the same model without an
interaction term).
library(modsem)
m1 <- '
# Outer Model
X =~ x1 + x2 + x3
Z =~ z1 + z2 + z3
Y =~ y1 + y2 + y3
# Inner Model
Y ~ X + Z + X:Z
'
lms1 <- modsem(m1, oneInt, method = "lms")
# Standardized estimates
summary(lms1, standardized = TRUE)
#> Estimating baseline model (H0)
#>
#> modsem (version 1.0.9):
#> Estimator LMS
#> Optimization method EM-NLMINB
#> Number of observations 2000
#> Number of iterations 84
#> Loglikelihood -14687.86
#> Akaike (AIC) 29437.72
#> Bayesian (BIC) 29611.35
#>
#> Numerical Integration:
#> Points of integration (per dim) 24
#> Dimensions 1
#> Total points of integration 24
#>
#> Fit Measures for H0:
#> Loglikelihood -17832
#> Akaike (AIC) 35723.75
#> Bayesian (BIC) 35891.78
#> Chi-square 17.52
#> Degrees of Freedom (Chi-square) 24
#> P-value (Chi-square) 0.826
#> RMSEA 0.000
#>
#> Comparative fit to H0 (no interaction effect)
#> Loglikelihood change 3144.01
#> Difference test (D) 6288.02
#> Degrees of freedom (D) 1
#> P-value (D) 0.000
#>
#> R-Squared:
#> Y 0.596
#> R-Squared Null-Model (H0):
#> Y 0.395
#> R-Squared Change:
#> Y 0.201
#>
#> Parameter Estimates:
#> Coefficients standardized
#> Information expected
#> Standard errors standard
#>
#> Latent Variables:
#> Estimate Std.Error z.value P(>|z|)
#> X =~
#> x1 0.926 0.004 211.4 0.000
#> x2 0.891 0.006 159.4 0.000
#> x3 0.912 0.005 180.7 0.000
#> Z =~
#> z1 0.927 0.006 166.0 0.000
#> z2 0.898 0.005 178.4 0.000
#> z3 0.913 0.005 166.0 0.000
#> Y =~
#> y1 0.961 0.002 406.4 0.000
#> y2 0.942 0.003 279.9 0.000
#> y3 0.951 0.002 412.9 0.000
#>
#> Regressions:
#> Estimate Std.Error z.value P(>|z|)
#> Y ~
#> X 0.482 0.020 24.2 0.000
#> Z 0.417 0.021 20.1 0.000
#> X:Z 0.511 0.021 24.7 0.000
#>
#> Covariances:
#> Estimate Std.Error z.value P(>|z|)
#> X ~~
#> Z 0.199 0.028 7.1 0.000
#>
#> Variances:
#> Estimate Std.Error z.value P(>|z|)
#> x1 0.142 0.008 17.6 0.000
#> x2 0.206 0.010 20.7 0.000
#> x3 0.169 0.009 18.1 0.000
#> z1 0.141 0.010 13.6 0.000
#> z2 0.193 0.009 21.1 0.000
#> z3 0.167 0.010 16.7 0.000
#> y1 0.077 0.004 17.2 0.000
#> y2 0.112 0.006 17.5 0.000
#> y3 0.096 0.005 21.2 0.000
#> X 1.000
#> Z 1.000
#> Y 0.514 0.016 31.4 0.000Here is the same example using the QML approach:
qml1 <- modsem(m1, oneInt, method = "qml")
summary(qml1)
#> Estimating baseline model (H0)
#>
#> modsem (version 1.0.9):
#> Estimator QML
#> Optimization method NLMINB
#> Number of observations 2000
#> Number of iterations 110
#> Loglikelihood -17496.22
#> Akaike (AIC) 35054.43
#> Bayesian (BIC) 35228.06
#>
#> Fit Measures for H0:
#> Loglikelihood -17832
#> Akaike (AIC) 35723.75
#> Bayesian (BIC) 35891.78
#> Chi-square 17.52
#> Degrees of Freedom (Chi-square) 24
#> P-value (Chi-square) 0.826
#> RMSEA 0.000
#>
#> Comparative fit to H0 (no interaction effect)
#> Loglikelihood change 335.66
#> Difference test (D) 671.32
#> Degrees of freedom (D) 1
#> P-value (D) 0.000
#>
#> R-Squared:
#> Y 0.607
#> R-Squared Null-Model (H0):
#> Y 0.395
#> R-Squared Change:
#> Y 0.211
#>
#> Parameter Estimates:
#> Coefficients unstandardized
#> Information observed
#> Standard errors standard
#>
#> Latent Variables:
#> Estimate Std.Error z.value P(>|z|)
#> X =~
#> x1 1.000
#> x2 0.803 0.013 63.96 0.000
#> x3 0.914 0.013 67.80 0.000
#> Z =~
#> z1 1.000
#> z2 0.810 0.012 65.12 0.000
#> z3 0.881 0.013 67.62 0.000
#> Y =~
#> y1 1.000
#> y2 0.798 0.007 107.57 0.000
#> y3 0.899 0.008 112.55 0.000
#>
#> Regressions:
#> Estimate Std.Error z.value P(>|z|)
#> Y ~
#> X 0.674 0.032 20.94 0.000
#> Z 0.566 0.030 18.96 0.000
#> X:Z 0.712 0.028 25.46 0.000
#>
#> Intercepts:
#> Estimate Std.Error z.value P(>|z|)
#> x1 1.023 0.024 42.89 0.000
#> x2 1.215 0.020 60.99 0.000
#> x3 0.919 0.022 41.48 0.000
#> z1 1.012 0.024 41.57 0.000
#> z2 1.206 0.020 59.27 0.000
#> z3 0.916 0.022 42.06 0.000
#> y1 1.038 0.033 31.45 0.000
#> y2 1.221 0.027 45.49 0.000
#> y3 0.955 0.030 31.86 0.000
#> Y 0.000
#> X 0.000
#> Z 0.000
#>
#> Covariances:
#> Estimate Std.Error z.value P(>|z|)
#> X ~~
#> Z 0.200 0.024 8.24 0.000
#>
#> Variances:
#> Estimate Std.Error z.value P(>|z|)
#> x1 0.158 0.009 18.14 0.000
#> x2 0.162 0.007 23.19 0.000
#> x3 0.165 0.008 20.82 0.000
#> z1 0.166 0.009 18.34 0.000
#> z2 0.159 0.007 22.62 0.000
#> z3 0.158 0.008 20.71 0.000
#> y1 0.159 0.009 17.98 0.000
#> y2 0.154 0.007 22.67 0.000
#> y3 0.164 0.008 20.71 0.000
#> X 0.983 0.036 26.99 0.000
#> Z 1.019 0.038 26.95 0.000
#> Y 0.943 0.038 24.87 0.000A More Complicated Example
Below is an example of a more complex model based on the theory of
planned behavior (TPB), which includes two endogenous variables and an
interaction between an endogenous and exogenous variable. When
estimating more complex models with the LMS approach, it is
recommended to increase the number of nodes used for numerical
integration. By default, the number of nodes is set to 16, but this can
be increased using the nodes argument. The
nodes argument has no effect on the QML
approach.
When there is an interaction effect between an endogenous and
exogenous variable, it is recommended to use at least 32 nodes for the
LMS approach. You can also obtain robust standard errors by
setting robust.se = TRUE in the modsem()
function.
Note: If you want the LMS approach to
produce results as similar as possible to Mplus, you should increase the
number of nodes (e.g., nodes = 100).
# ATT = Attitude
# PBC = Perceived Behavioral Control
# INT = Intention
# SN = Subjective Norms
# BEH = Behavior
tpb <- '
# Outer Model (Based on Hagger et al., 2007)
ATT =~ att1 + att2 + att3 + att4 + att5
SN =~ sn1 + sn2
PBC =~ pbc1 + pbc2 + pbc3
INT =~ int1 + int2 + int3
BEH =~ b1 + b2
# Inner Model (Based on Steinmetz et al., 2011)
INT ~ ATT + SN + PBC
BEH ~ INT + PBC
BEH ~ INT:PBC
'
lms2 <- modsem(tpb, TPB, method = "lms", nodes = 32)
summary(lms2)
#> Estimating baseline model (H0)
#>
#> modsem (version 1.0.9):
#> Estimator LMS
#> Optimization method EM-NLMINB
#> Number of observations 2000
#> Number of iterations 64
#> Loglikelihood -23439.2
#> Akaike (AIC) 46986.41
#> Bayesian (BIC) 47288.85
#>
#> Numerical Integration:
#> Points of integration (per dim) 32
#> Dimensions 1
#> Total points of integration 32
#>
#> Fit Measures for H0:
#> Loglikelihood -26393
#> Akaike (AIC) 52892.45
#> Bayesian (BIC) 53189.29
#> Chi-square 66.27
#> Degrees of Freedom (Chi-square) 82
#> P-value (Chi-square) 0.897
#> RMSEA 0.000
#>
#> Comparative fit to H0 (no interaction effect)
#> Loglikelihood change 2954.02
#> Difference test (D) 5908.04
#> Degrees of freedom (D) 1
#> P-value (D) 0.000
#>
#> R-Squared:
#> INT 0.364
#> BEH 0.259
#> R-Squared Null-Model (H0):
#> INT 0.367
#> BEH 0.210
#> R-Squared Change:
#> INT -0.003
#> BEH 0.049
#>
#> Parameter Estimates:
#> Coefficients unstandardized
#> Information expected
#> Standard errors standard
#>
#> Latent Variables:
#> Estimate Std.Error z.value P(>|z|)
#> PBC =~
#> pbc1 1.000
#> pbc2 0.914 0.012 76.48 0.000
#> pbc3 0.802 0.012 65.22 0.000
#> ATT =~
#> att1 1.000
#> att2 0.878 0.012 70.62 0.000
#> att3 0.789 0.011 69.76 0.000
#> att4 0.695 0.012 55.81 0.000
#> att5 0.887 0.015 60.88 0.000
#> SN =~
#> sn1 1.000
#> sn2 0.889 0.018 50.72 0.000
#> INT =~
#> int1 1.000
#> int2 0.913 0.014 64.03 0.000
#> int3 0.807 0.014 56.08 0.000
#> BEH =~
#> b1 1.000
#> b2 0.959 0.035 27.38 0.000
#>
#> Regressions:
#> Estimate Std.Error z.value P(>|z|)
#> INT ~
#> PBC 0.217 0.031 6.94 0.000
#> ATT 0.214 0.030 7.06 0.000
#> SN 0.176 0.027 6.58 0.000
#> BEH ~
#> PBC 0.233 0.020 11.54 0.000
#> INT 0.188 0.024 7.78 0.000
#> PBC:INT 0.205 0.020 10.29 0.000
#>
#> Intercepts:
#> Estimate Std.Error z.value P(>|z|)
#> pbc1 0.991 0.019 52.90 0.000
#> pbc2 0.978 0.017 55.99 0.000
#> pbc3 0.986 0.017 59.16 0.000
#> att1 1.009 0.022 45.04 0.000
#> att2 1.003 0.020 50.14 0.000
#> att3 1.013 0.016 63.73 0.000
#> att4 0.996 0.016 61.52 0.000
#> att5 0.988 0.019 52.07 0.000
#> sn1 1.001 0.021 47.22 0.000
#> sn2 1.006 0.020 50.89 0.000
#> int1 1.011 0.019 51.86 0.000
#> int2 1.009 0.016 63.02 0.000
#> int3 1.003 0.020 50.67 0.000
#> b1 0.999 0.018 56.04 0.000
#> b2 1.017 0.016 64.27 0.000
#> INT 0.000
#> BEH 0.000
#> PBC 0.000
#> ATT 0.000
#> SN 0.000
#>
#> Covariances:
#> Estimate Std.Error z.value P(>|z|)
#> PBC ~~
#> ATT 0.668 0.026 26.10 0.000
#> SN 0.668 0.025 27.05 0.000
#> ATT ~~
#> SN 0.623 0.024 26.04 0.000
#>
#> Variances:
#> Estimate Std.Error z.value P(>|z|)
#> pbc1 0.148 0.007 20.14 0.000
#> pbc2 0.159 0.006 25.20 0.000
#> pbc3 0.155 0.007 22.52 0.000
#> att1 0.167 0.007 24.92 0.000
#> att2 0.150 0.005 27.43 0.000
#> att3 0.159 0.005 30.66 0.000
#> att4 0.162 0.006 25.93 0.000
#> att5 0.159 0.007 22.68 0.000
#> sn1 0.178 0.016 11.16 0.000
#> sn2 0.156 0.011 14.05 0.000
#> int1 0.157 0.011 13.67 0.000
#> int2 0.160 0.006 25.51 0.000
#> int3 0.168 0.007 22.57 0.000
#> b1 0.185 0.023 8.02 0.000
#> b2 0.136 0.020 6.70 0.000
#> PBC 0.947 0.035 27.27 0.000
#> ATT 0.992 0.032 31.43 0.000
#> SN 0.981 0.033 29.84 0.000
#> INT 0.491 0.020 25.07 0.000
#> BEH 0.456 0.023 20.14 0.000
qml2 <- modsem(tpb, TPB, method = "qml")
summary(qml2, standardized = TRUE) # Standardized estimates
#> Estimating baseline model (H0)
#>
#> modsem (version 1.0.9):
#> Estimator QML
#> Optimization method NLMINB
#> Number of observations 2000
#> Number of iterations 73
#> Loglikelihood -26326.25
#> Akaike (AIC) 52760.5
#> Bayesian (BIC) 53062.95
#>
#> Fit Measures for H0:
#> Loglikelihood -26393
#> Akaike (AIC) 52892.45
#> Bayesian (BIC) 53189.29
#> Chi-square 66.27
#> Degrees of Freedom (Chi-square) 82
#> P-value (Chi-square) 0.897
#> RMSEA 0.000
#>
#> Comparative fit to H0 (no interaction effect)
#> Loglikelihood change 66.97
#> Difference test (D) 133.95
#> Degrees of freedom (D) 1
#> P-value (D) 0.000
#>
#> R-Squared:
#> INT 0.366
#> BEH 0.263
#> R-Squared Null-Model (H0):
#> INT 0.367
#> BEH 0.210
#> R-Squared Change:
#> INT 0.000
#> BEH 0.053
#>
#> Parameter Estimates:
#> Coefficients standardized
#> Information observed
#> Standard errors standard
#>
#> Latent Variables:
#> Estimate Std.Error z.value P(>|z|)
#> PBC =~
#> pbc1 0.933 0.004 217.38 0.000
#> pbc2 0.913 0.005 189.94 0.000
#> pbc3 0.894 0.005 163.46 0.000
#> ATT =~
#> att1 0.925 0.004 240.19 0.000
#> att2 0.915 0.004 211.80 0.000
#> att3 0.892 0.005 173.63 0.000
#> att4 0.865 0.006 138.97 0.000
#> att5 0.912 0.004 207.47 0.000
#> SN =~
#> sn1 0.921 0.007 128.23 0.000
#> sn2 0.913 0.007 125.71 0.000
#> INT =~
#> int1 0.912 0.006 163.04 0.000
#> int2 0.895 0.006 146.66 0.000
#> int3 0.866 0.007 125.52 0.000
#> BEH =~
#> b1 0.870 0.014 62.27 0.000
#> b2 0.894 0.014 64.03 0.000
#>
#> Regressions:
#> Estimate Std.Error z.value P(>|z|)
#> INT ~
#> PBC 0.243 0.032 7.47 0.000
#> ATT 0.242 0.029 8.27 0.000
#> SN 0.199 0.031 6.43 0.000
#> BEH ~
#> PBC 0.299 0.027 10.99 0.000
#> INT 0.219 0.028 7.89 0.000
#> PBC:INT 0.235 0.021 11.01 0.000
#>
#> Covariances:
#> Estimate Std.Error z.value P(>|z|)
#> PBC ~~
#> ATT 0.692 0.013 53.44 0.000
#> SN 0.695 0.014 51.47 0.000
#> ATT ~~
#> SN 0.634 0.015 42.01 0.000
#>
#> Variances:
#> Estimate Std.Error z.value P(>|z|)
#> pbc1 0.130 0.008 16.25 0.000
#> pbc2 0.166 0.009 18.71 0.000
#> pbc3 0.201 0.010 20.49 0.000
#> att1 0.144 0.007 20.09 0.000
#> att2 0.164 0.008 20.87 0.000
#> att3 0.204 0.009 22.08 0.000
#> att4 0.252 0.011 23.48 0.000
#> att5 0.168 0.008 20.96 0.000
#> sn1 0.153 0.013 11.49 0.000
#> sn2 0.167 0.013 12.53 0.000
#> int1 0.168 0.010 16.45 0.000
#> int2 0.199 0.011 18.27 0.000
#> int3 0.249 0.012 20.82 0.000
#> b1 0.244 0.024 10.05 0.000
#> b2 0.202 0.025 8.03 0.000
#> PBC 1.000
#> ATT 1.000
#> SN 1.000
#> INT 0.634 0.019 33.92 0.000
#> BEH 0.790 0.019 41.41 0.000