Discussion of OpenMx beta testing progress

Avatars, Multivariate Analysis, and Metamers: A Theory of Facial Dynamics and Affect Expression
Steve Boker
When we speak to one another, we adapt and coordinate our facial expressions in response to our inner affective and cognitive states and in response to our perceptions of the person with whom we are speaking. We say we are speaking "with" someone, since this is a shared experience; one that creates an interpersonal coupled system with feedback between the dynamics of the two speakers' perceptions and actions.

We have developed a method that can track the non-rigid motion of a person's face from a single video camera and then reconstruct near-photorealistic video from between 8 to 15 principal components to account for the appearance and shape of a person's face in real time. During the reconstruction phase (about 33ms) we can alter the person's appearance and/or dynamics in order to make manipulated tests of hypotheses concerning interpersonal coordination during natural conversation. After running more than 120 participants in videoconference experiments using the reconstructed computer graphic avatar models, only four expressed doubts during debrief about whether the faces they spoke with were video.

It is surprising that such a low-dimensional representation of the face can be mistaken for video of a real person. This talk presents the argument that a low-dimensional coding of facial dynamics is performed as we speak with others. The argument proceeds from analogy to the system of color vision. It is further speculated that the low dimensional structure found using factor analysis of emotion adjectives has a mapping to the low-dimensional coding of facial dynamics. One mathematical consequence of these theoretic proposals is that there would exist "facial metamers" -- just as there are many combinations of spectral lines that produce the same perceived color, it is predicted that there are many configurations of facial musculature contraction that are perceived as the same emotional intent. This would have the benefit of perception of similar emotional intent given individual differences in facial musculature. If the face-emotion mapping is developmentally acquired, there may be cultural/linguistic differences in facial metamers. That is to say, growing up in culture A may leave one effectively blind to important differences in facial dynamics in culture B.

Inference with Heywood cases
Stas Kolenikov
I will review several hypotheses associated with variance estimates below zero, and discuss various ways to test them. The estimates and tests have different behavior depending on whether the true value of the variance is above zero (a regular case), at zero (a boundary case), or below zero (a grossly misspecified case). Since the testing situations are non-standard, the distributions of the test statistics are often non-standard and different from the traditional chi-squares. Some new forms of standard errors, as well as new test statistics, are proposed, and their properties are studied in a simulation project.

Running WinBUGS Through R
Mike Clark
Running WinBUGS through R has many potential benefits, including an easy solution for batching WinBUGS jobs and aggregating results. During this presentation, I will demonstrate the R2WinBUGS and CODA packages for R, and I will also share some preliminary results from my WinBUGS simulation research, which involves using WinBUGS to fit item response theory (IRT) models to small samples.

Obtaining Psychometric Report Cards using Diagnostic Classification Models
Emily Fall
Diagnostic classification modeling (DCM) can be used in a variety of settings from educational testing to clinical diagnosis to evaluate skills or traits of individuals. Classifications are made using both categorical response variables and categorical latent variables. Individuals are evaluated based on a set of skills rather than a continuous latent trait, so that markers of the trait can be identified and ‘treated.’ The theory and application of DCM is extensive and growing rapidly. I’ll outline some of these models, their applications, and DCM’s place in the broader world of psychometric testing.

Fall Break

Two Researchers and the Autoregressive Conundrum
Pascal Deboeck
Many methods are available for modeling intraindividual change. The modeling of intraindividual variability, however, requires significant advancement in the statistical models that are being used. One prevalent approach is to incorporate an autoregressive component to describe intraindividual variability; that is, a component where some observation is caused by a previous observation. The common discrete time approach differs from continuous time approaches that postulate that the observations are manifestations of an ongoing, underlying process. The results of two researchers will be presented, leading to puzzling conclusions about discrete-time autoregressive methods. Why and how a discrete time autoregressive model can be written as a continuous time model will be discussed.

Scale Setting and Model Specification in Second-order Latent Growth Curve Models
Ihno Lee
Second-order latent growth curve (LGC) models are a better alternative to first-order LGCs; growth is modeled from the error-free latent composites rather than from the variance-covariance matrix of the manifest indicators. Under conditions of factorial invariance, growth trajectories are reliable, and the method of scaling does not affect parameter estimates or model fit. However, some questions remain: 1) Can growth parameters be influenced by a non-arbitrary method of scale setting? 2) How does model specification, such as enforcing stationarity of residuals, influence parameter estimates and/or model fit? I will address these issues via a sequence of multivariate curve-of-factors models in a study of daily mood fluctuation.

Alternatives to Randomized Experiments: Designs and Strategies to Strengthen Inferences in Field Settings
Stephen West
Randomized experiments (REs) are preferred for causal inference. However, REs cannot be implemented to answer a number of questions in public health and in some areas of clinical, community, health, and prevention psychology (e.g., effects of Hurricane Katrina or secondhand tobacco smoke on health outcomes). Two approaches, Rubin's potential outcomes model from statistics and Campbell's pattern matching from the behavioral sciences, offer strong foundations for causal inference even in studies in which randomization is not possible or has been compromised. These approaches are applied to quantitative assignment designs (e.g., participants are assigned to treatment on the basis of need, merit, or risk) and designs in which the participants are assigned to treatment on the basis of unknown rules. These approaches can lead to relatively strong causal inferences, particularly if complementary design features are combined with the basic design. Empirical comparisons suggesting that alternative approaches produce similar effect size estimates to those of the RCT are described. Alternative designs permit a wider range of research questions to be answered and often permit more direct generalization of causal effects to practice; however, estimates of the magnitude of the causal effect may be more uncertain.
Presentation

TBA
Wei Wu & Kyle Lang

The Problem of Model Selection Uncertainty in Structural Equation Modeling
Kris Preacher
Model selection in structural equation modeling involves using selection criteria to select one model as superior and treating it as a best working hypothesis until a better model is proposed. A limitation of this approach is that sampling variability in selection criteria usually is not considered, leading to assertions of model superiority that may not withstand replication. I illustrate that selection decisions using information criteria (e.g., AIC, BIC) can be highly unstable over repeated sampling, and this uncertainty does not necessarily decrease with increases in sample size. I will discuss methods for addressing model selection uncertainty in SEM, as well as implications for practice.

TBA
David Johnson

Panel Discussion
Schoemann, Ihno Lee, & John Geldhof
A discussion of modern statistical techniques and the ethical obligation to use them, based on:

Erceg-Hurn, D. M., & Mirosevich, V. M. (2008). Modern robust statistical methods: An easy way to maximize the accuracy and power of your research. American Psychologist, 63(7), 591-601