Dutch Mplus Users Meeting – @ Utrecht University, The Netherlands.
Below you can watch the presentations given at the 4th Mplus Users Meeting on the 30th of August 2012 as part of our three week Mplus summer school event. Click on the lecture name to open the video and enjoy!
Peg Burchinal (FPG Child Development Institute, University of Noth Carolina)
Using Mplus to Analyze Data in Psychology Journals: An Associate Editor’s View - VIDEO
Jon Heron (School of Social and Community Medicine University of Bristol)
Adolescent alcohol use and depressive symptoms
Using data from ALSPAC, a UK-based birth cohort, I will apply a range of longitudinal models to assess the change in depressive symptoms and problem alcohol use from childhood through adolescence. I will seek to address the problem of missing data as well as the fact there has been a lack of consistency in the questions administered over time.
Tina Kretschmer (University of Groningen, Interuniversity Centre for Social Science Theory and Methodology)
Trajectory membership information in path models
Mixture models have informed about distinguishable developmental trajectories in a variety of constructs including conduct problems. This talk will show how using this trajectory membership information in path models can inform about differential associations between risk factors and outcomes as a function of previous developmental pathways.
Jan Boom (Developmental Psychology, Utrecht University)
Combining IRT and LGM to Model Strategy Development
The Overlapping Waves Model is a metaphor to illustrate a typical sequence of increasing and decreasing use of strategies during cognitive development. I have developed a new statistical model to analyze such categorical longitudinal data. Strategy use can be scored as an ordinal variable with few categories and longitudinal development as a vector of such scores. Item Response Theory provides the means to relate the use of such strategies to an underlying developmental dimension. Movement of individuals along this dimension can be modeled by means of Latent Growth Modeling using Mplus. Visualization with an interactive 3D demo is used to illustrate.
Vicky Strauss (Arthritis Research UK Primary Care Centre, Keele University)
Modelling longitudinal data with excess zeros: an application
It is often found that the distribution of a count variable in health care utilisation contains a large amount of zeros, and has a long right tail. Therefore, the use of Poisson distribution becomes inappropriate. In this specific application, health care utilisation is measured by the number of morbidities consulted for in the same time periods (i.e. multimorbidity) in a three-year primary care consultation data. We aim to investigate the use of several approaches to address this issue in growth mixture models in determining multimorbidity trajectories over time. The approaches discussed in this talk are: Negative Binomial, Zero-Inflated Poisson, Zero-Inflated Negative Binomial, and finally a two-part model. Advantages, disadvantages and problems of different approaches will be considered in this talk.
Talk 2: History and Future of Mplus
Talk 1: Causally-Defined Direct and Indirect Effects in Mediation Modeling
The paper that this talk is based on summarizes some of the literature on causal effects in mediation analysis. The paper presents causally-defined direct and indirect effects for continuous, binary, ordinal, nominal, and count variables. The expansion to non-continuous mediators and outcomes offers a broader array of causal mediation analyses than previously considered in structural equation modeling practice. A new result is the ability to handle mediation by a nominal variable. Examples with a binary outcome and a binary, ordinal or nominal mediator are given using Mplus to compute the effects. The causal effects require strong assumptions even in randomized designs, especially sequential ignorability, which is presumably often violated to some extent due to mediator-outcome confounding. To study the effects of violating this assumption, the paper shows how a sensitivity analysis can be carried out. This can be used both in planning a new study and in evaluating the results of an existing study.
Ian Rothman (Business Intelligence Manager: Afriforte (Pty) Ltd)
Ellen Hamaker (Department of Methods & Statistics, Utrecht University, The Netherlands)
Talk 1: Issues in cross-lagged panel models
Talk 2: Dynamic multilevel models
Dynamic multilevel models that are based on time series models have proven valuable in studying individual differences in affect regulation and other dynamic processes. In this talk two ways for specifying such models in Mplus will be presented and compared. In addition, the (dis)advantages and limitations of both approaches will be considered.
Tihomor Asparouhov (moderator of Mplus)
Estimating new structural equation models with the Bayesian methods
We present several structural models that can be efficiently estimated with the Bayesian methods but cannot be estimated with frequentist methods. We consider the consequences of adopting simpler, less flexible and misspecified models for the sole purpose of using a frequentist estimation method. Simulation examples and real data examples are presented.
Julia Dietrich (University of Helsinki, Finland, and University of Erfurt, Germany)
An application of the new Three-Step Approach
Using data on adolescents’ developmental regulation in the domain of career choice, this talk will demonstrate an application of the new 3-step approach to growth mixture modeling in Mplus 7. The aim of the study is to identify typical trajectories of career goal engagement. To identiy trajectory classes, we compare the (traditional) Pseudo class 3-step approach, where the most likely latent class membership is saved and used in further analysis, with the 1-step approach and the (new) Mplus 7 3-step approach.
Joop Hox (Department of Methods & Statistics, Utrecht University, The Netherlands)
Lessons learned from simulation studies where maximum likelihood and Bayes are being compared.
Bayesian estimation is attractive in settings where data are sparse and/or models are complex. A persistent problem in multilevel modeling is the requirement of having a sufficient sample size at the highest level. This requirement actually applies to maximum likelihood estimation, which is asymptotic. Bayesian analysis is expected to outperform ML estimation in such settings. This presentation discusses Bayesian estimation in multilevel models with small numbers of groups. It also discusses what happens when the sample size becomes very small, how well does Bayesian estimation work here, and how can one see when things go wrong?