Steven Boker
Introduction to OpenMx XSEM with Application to Dynamical Systems Analysis:
OpenMx is a free, open source Extended Structural Equation Modeling (XSEM) package that allows specification and estimation of a wide variety of models from within the R statistical system. The first half of the workshop will introduce OpenMx and the specification of simple latent variable models within R as well as provide an overview of some of its more advanced features such as full information, definition variables, parallelizing, ordinal thresholds, multiple groups, and mixture distributions. The second half of the workshop will present dynamical systems analysis as a continuous-time generalization of time series models using systems of differential equations. We will begin by introducing first order (exponential decay) and second order (cyclic) linear differential equations and two methods for fitting these to time series data: state space forward prediction and time delay embedding. We will then move to coupled differential equations for estimating parameters of dyadic relationships. Finally, we will introduce models for individual differences in parameters of regulation and equilibria as well as time-varying versions of these equations. In preparation, the workshop attendee should have R loaded on her or his laptop and have some familiarity with the use of R. The R packages “OpenMx”, “psych”, and “deSolve” should be installed prior to attending. R scripts and data will be provided so that attendees can gain hands-on practice with OpenMx and data handling techniques for time-delay embedding and state space modeling.
Biography:
Dr. Boker’s research interests include the application of dynamical systems analytic techniques to psychological and physiological data. His contributions include methods for examining change in multivariate mixed cross-sectional and longitudinal data include the Statistical Vector Field method, Differential Structural Equation Modeling using local linear approximation of derivatives, and the Latent Differential Equations method for fitting differential equations models to multivariate multiple occasion data. He is currently pursuing research into methods for estimating models for nonstationary data — data for which model parameters or model structures change over time. Dr. Boker’s current NSF sponsored project is through the Human and Social Dynamics program. In collaboration with Jeffrey Cohn at University of Pittsburgh and Simon Baker in the Robotics Institute of Carnegie Mellon University, he is studying the coordination of gestures and facial expressions during dyadic conversation over a video phone. Dr. Boker’s lab uses state of the art computerized technology to test cognitive theories of interpersonal coordination and perception-action coupling during conversation, dance, and imitation learning tasks. His awards include the Raymond B. Cattell Award for distinguished early career contributions to multivariate psychology and the Tanaka award from the Society of Multivariate Experimental Psychology. See website: http://people.virginia.edu/~smb3u/
For more information about OpenMx, go to http://www.openmx-square.org/
Kenneth Bollen
Model implied instrumental variables using MIIVsem:
An R Package for Structural Equation Models (SEMs)
If our models were perfectly valid and our variables only came from normal distributions, the maximum likelihood and related estimators that dominate SEM software would be hard to beat. In reality, however, such structural and distributional assumptions are rarely if ever satisfied. This workshop will discuss more robust estimators that better represent real world conditions. Model Implied Instrumental Variable (MIIV) estimators are more robust to the approximate nature of our models and are asymptotically distribution free. In addition, they can test equation level fit so as to better localize model misspecification. The workshop will give an overview of the free R package MIIVsem. We will introduce the key ideas behind MIIV estimation; we will show how MIIVsem automates the selection of MIIVs, the estimation of coefficients and standard errors, and provides overidentification tests for equations. These and other features will be introduced and illustrated with a variety of empirical examples. We will provide instructions on downloading R and MIIVsem and will ask participants to bring their computers to the workshop so that they can run these empirical examples on their own machines during the last part of the workshop. No prior knowledge of R or MIIVsem is assumed.
Biography:
Ken Bollen is the Henry Rudolph Immerwahr Distinguished Professor in the Department of Psychology and Neuroscience and Department of Sociology at UNC at Chapel Hill. He is a faculty member in the Quantitative Psychology Program in the Thurstone Psychometric Laboratory. He also is chair of the Methods Core and a Fellow of the Carolina Population Center and a Faculty member of the Center for Developmental Science. Since 1980 he has been an instructor in the ICPSR Summer Program in Quantitative Methods of Social Research. Bollen’s primary areas of statistical research are in structural equation models, longitudinal methods, and latent growth curve models. More details are available in his c.v.
A pdf of the Dr. Bollen’s keynote slides is available here: MIIVs A New Orientation to SEM.
For a copy of the MIIVsem post-conference workshop slides, go to github.com/zackfisher/M3
Craig Enders
Multiple Imputation for Multilevel Data:
Multiple imputation has been widely available for many years, but classic imputation routines are inappropriate for multilevel data because they ignore clustering and generate imputations from a model that assumes zero intraclass correlations. Joint modeling (JM) and fully conditional specification (FCS) are the principal imputation frameworks for single-level data, and both have multilevel counterparts. The multilevel versions of JM and FCS apply different underlying models, and software packages offer different functionality. The JM and FCS tools that researchers currently have at their disposal work well for very specific tasks, but these methods are generally not equipped to handle a wide range of complexities that are typical of behavioral science data. The Blimp application for Mac OS and Windows was developed to address a range of scenarios that are difficult or impossible to handle with existing software. In particular, the FCS machinery implemented in Blimp can accommodate random slopes, nominal and ordinal variables, unique within- and between-cluster covariance structures, heterogeneous within-cluster variance structures, and models with up to three levels, with missing data at any level of the data hierarchy. An imputation model that implements any combination of these features can be specified with a simple command language or through the graphical user interface. The session will provide attendees with a broad overview of multilevel imputation, followed by a specific discussion of the FCS methodologies implemented in Blimp. The presentation will include a mixture of theoretical and applied topics and will include a software demonstration that illustrates the use of Blimp with popular analysis programs such as Mplus, R, SAS, and SPSS. The Blimp software, documentation, and various analysis scripts are available for download at www.appliedmissingdata.com/multilevel-imputation.html.
A copy of Dr. Enders’ revised workshop slides is available here: Multilevel Imputation Workshop Slides
Biography:
“My methodological research program has focused on developing and applying modern statistical methods for addressing missing data, principally maximum likelihood estimation, multiple imputation, and Bayesian estimation. These techniques were initially developed in mathematical statistics, sometimes making strong assumptions about the nature of the data. Because these assumptions are often not met in psychology data, problems arise when behavioral scientists attempt to apply modern analytic approaches. My goal has been to identify the domains in which these procedures produce adequate results and to develop new methods in domains in which they are not adequate. My research develops new models for the treatment of missing data for common statistical analyses methods, among them structural equation modeling, growth curve modeling, and multilevel modeling.
Most recently, I have been developing imputation methods (procedures that fill in the missing scores with replacement values) for multilevel data structures where observations are nested within higher-order organizational units (e.g., repeated measures within individuals; children within families; students within schools). Multilevel data are ubiquitous throughout psychology and the social sciences, yet methods for dealing with missing data in this context have received relatively little attention in the methodological literature. I was recently award a three-year grant from the Institute of Educational Sciences to develop statistical software for dealing with missing data in multilevel analyses.
I am originally from the Midwest and received a B.A. in Psychology from University of Nebraska – Lincoln. During my undergraduate years, I developed an interest in psychometrics, and I subsequently did my Ph.D. work in psychometrics and quantitative methods at the University of Nebraska in Educational Psychology. Prior to joining the faculty at UCLA in 2015, I was a professor in the Department of Psychology at Arizona State University.”