Title: The analysis of complex nonlinear structural equation models: A comparison of LMS and QML
1The analysis of complex nonlinear structural
equation modelsA comparison of LMS and QML
- Polina Dimitruk, Helfried Moosbrugger,
- Augustin Kelava Karin Schermelleh-Engel
- J. W. Goethe University, Frankfurt am Main,
Germany
Supported by a grant of the German Research
Foundation (DFG), Nr.Mo 474/6-1
2nd Conference of the European Association of
Methodology (EAM) July 2 5, 2006, Budapest
2Outline
- Nonlinear Regression Models
- LMS method
- QML method
- Simulations Studies
- Design of Monte-Carlo Study
- Results of Monte-Carlo Study
- Conclusions
3Nonlinear Regression Models
- There are 3 common effects in nonlinear
regression models - Interaction effect The slope of the regression
of the criterion variable on a predictor variable
varies with the realizations of a second
predictor variable. - Y ß0ß1Xß2Z ß3XZe
- 2. Quadratic effectsThe slope of the regression
of the criterion variable on a predictor variable
varies with the realizations of the predictor
variable itself. - Y ß0ß1Xß2X2e
- 3. Mixed nonlinear effects The slope of the
regression of the criterion variable on a
predictor variable varies not only with the
realizations of the predictor variable itself,
but also with the realizations of a second
predictor variable. - Y ß0ß1Xß2Zß3XZß4X2e
4Nonlinear Models
- Nonlinear regression models can contain several
nonlinear effects, i.e., one or more interaction
effects and one or more qudratic effects. In a
simple case a full nonlinear model includes a
criterion variable h, two predictor variables x1
and x2, a product term x1x2 and two quadratic
terms x12 and x22 - h g1x1 g2x2 w12x1x2 w11x12 w22x22
z - Linear model results when all nonlinear effects
(w12, w11 and w22) are zero, interaction model
results when all quadratic effects (w11 and w22)
are zero, quadratic model results when the
interaction effect (w12) is zero, and finally
full model results when all nonlinear effects
(w12, w11 and w22) are unequal to zero.
5LMS method(Latent Moderated Structural
Equations, Klein, 2000 Klein Moosbrugger,
2000)
- LMS was developed especially for the estimate and
check of latent interaction effects in structural
equation models. - LMS is a maximum likelihood estimation technique
for latent nonlinear models with multiple latent
product terms. The LMS method assumes
normal-distributed predictor variables and error
variables, and this implies the normal
distribution of the X variables. - implemented in Mplus (Muthén Muthén, 2004).
6QML method(Quasi-Maximum Likelihood, Klein
Muthén, 2006)
- The QML approach has been developed for an
efficient and computationally feasible estimation
of multiple nonlinear effects in structural
equation models with quadratic forms. - The QML method has been specifically developed
to provide robust, efficient, and computationally
inexpensive analysis of structural equation
models with multiple nonlinear effects. - QML only assumes that the conditional
distribution of the latent criterion given the
X-variables can be approximated as a normal
distribution.
7Research questions
- Although simulation studies for small latent
interaction models have shown that LMS and QML
estimators are consistent, asymptotically
unbiased, asymptotically efficient, and
asymptotically normally distributed, it is not
known up to now, whether these methods perform
equally well when larger models are analyzed
including multiple nonlinear terms. - Up to now it has not been investigated whether
QML is as efficient as the ML estimator of LMS
when the distributional assumptions are met and a
complex model with several nonlinear effects is
analyzed. - Furthermore, it was not systematically
investigated how LMS and QML behave with more
complex models with several nonlinear effects
with smaller sample sizes (N200 and N400).
8Design of the Monte-Carlo Study
- An extensive simulation study was carried out to
the analysis of complex nonlinear structural
equation models with one interaction term and two
quadratic terms. - The structural equation models were analysed
using the LMS and QML methods. - In the simulation study a structural equation
model with two latent predictor variables x1 and
x2 and a single criterion variable h is used.
9Design of Monte-Carlo Study
- Data were generated from a population with known
parameter values. - 500 data sets per condition combination were
drawn. - We chose sample size of N200 and N400.
- Latent predictor variables were correlated (f12
.50)
10Design of Monte-Carlo Study The following
parameters have remained constant in our study
- - linear effects g1 g2 .316 (10 unique
variance of the criterion variable h ) - - predictor correlation f12 .50
- - reliability of indicators rel(X) rel(Y)
.80 - - 3 indicators per each latent variable
-
11 Design of Monte-Carlo Study Data Generation
In our models the size of the latent interaction
effects and the size of both quadratic effects
were varied across three levels The unique
variance of the criterion variable explained by
the interaction term and/or both quadratic terms
are 0 , 2 , or 5 Interaction effect w12
Quadratic effects w11 and w22
Each of these 9 conditions was examined by LMS
and QML using 500 data sets of sample sizes N200
and N400. It is a simulation design with
9x2x236 conditions
12Results of Monte-Carlo Study
- In the following, well present 12 conditions, 4
conditions for each model - Interaction model
- Quadratic model
- Full model with one interaction effect and two
quadratic effects
13Results of Monte-Carlo StudyInteraction model
with interaction effect w12.127 DR2 2 The
selection of ?12 .127 gives a model where the
percentage of variance of ? explained by the
interaction term is 2.
14Results of Monte-Carlo StudyQuadratic model with
two quadratic effects w11w22.071DR2 2The
selection of ?11 ?22 .071 gives a model
where the percentage of variance of ? explained
by both quadratic term is 2.
15Results of Monte-Carlo StudyFull model with
interaction effect w12.200DR2 5 andtwo
quadratic effects w11w22.112DR2 5
16Results of Monte-Carlo Study
- All linear parameters are unbiased.
- The nonlinear parameter estimates of QML for N
200 show also no substantial bias. - But the Monte-Carlo standard deviations (SD) are
somewhat larger for QML than they are for LMS.
And also the standard errors (SE) of the
parameter estimates for QML are a little larger
than for LMS. For N 400 both methods show
practically no difference. - The study reveals a small tendency to an
overestimation of the Monte-Carlo standard
deviations (SD) for QML for N200, for which the
SE/SD ratios lie between .979 and 1.483. Opposed
to this, for LMS SE/SD ratios lie between .914
and 1.009 for sample size 200. For sample size
400 SE/SD ratios lie for LMS between .936 and
1.022 and for QML between 0.960 and 1.031.
17 Conclusions
-
- This matches the expectation about the possibly
somewhat lower efficiency of QML under the normal
condition, because QML is only an approximate ML
estimator - Both methods LMS and QML have been developed for
an efficiency of multiple nonlinear effects in
structural equation models with quadratic forms
and the sample size from N 200 and N 400. - The results of the simulation study show that
parameters of the nonlinear effects of LMS and
QML for the sample sizes of N 200 and N 400
were estimated unbiased. - A small advantage for LMS for small sample sizes
exists compared to QML that the standard errors
are estimated more accurately.
18 Conclusions
- The QML and LMS estimation of standard errors
showed no substantial bias which supports precise
significance testing of multiple nonlinear
effects. They are very efficient, computationally
feasible, and practically adequate approaches,
which is of particular relevance when rather
complex structural equation models with 3
nonlinear effects are to be analyzed. - Further, simulation studies are needed in order
to analyze more complex models with one ore more
interaction and several quadratic terms.
Furthermore, the nonnormality of predictor
variables should also be systematically varied in
order to investigate whether methodological
problems may even be aggravated using nonnormally
variables compared to normally distributed
predictor variables.
19- Thank you for your attention
20 References
- Aiken, L. S., West, S. G. (1991). Multiple
Regression Testing and interpreting
interactions. Newbury Park, CA Sage. - Klein, A. Moosbrugger, H. (2000). Maximum
likelihood estimation of latent interaction
effects with the LMS method. Psychometrika, 65,
457-474. - Klein, A. G. Muthén, B. O. (2006). Quasi
Maximum Likelihood Estimation of Structural
Equation Models with Multiple Interaction and
Quadratic Effects. Journal of Multivariate and
Behavioral Research (in press). - Muthén, L. K. Muthén, B. O. (2004). Mplus
user's guide version 3. Los Angeles, CA Muthén.
Muthén.