Orginizational Meeting/Recruitment Day

Terri Pigott: 3:00 p.m., Watson 3 West Reading Room
Professor & Associate Dean, School of Education, Loyola University Chicago

This talk will introduce new advances in the methods for meta-analysis. As the use of meta-analysis increases, the complexity of the studies included in a research synthesis has spurred the development of new methods for synthesizing study results. These new methods include strategies for synthesizing the results of diagnostic tests, the inclusion of both aggregated data and individual participant data in a meta-analysis, and computing power for meta-analytic statistical tests.

IT: 2:00 p.m, Watson 503 A & B
Video to Web IT Presentation IT Client Consulting, University of Kansas

Neal Kingston: 3:00 p.m, Joseph R. Pearson Hall, Room 201
Professor, Psychology and Research in Education
Director, Center for Educational Testing and Evaluation, University of Kansas

Student test scores were first used to support test scores about students. Under No Child Left Behind they were used to hold schools accountable. Under Race to the Top they are being used to determine the effectiveness of individual teachers. Now some folks are arguing improvement in student scores should be used to determine the effectiveness of schools of education. Underlying all of these uses is the assumption that teachers have substantial impact on student test scores. Instructional sensitivity is the extent to which a test item can be influenced by good instruction. While there have been few studies of this phenomenon, those studies have identified few items that possess this characteristic. In this talk Neal Kingston will share information about a program underway at CETE and make a call for more methodological and substantive research.

David Flora, 3:00 p.m., Joseph R. Pearson Hall, Room 201
Associate Professor, Department of Psychology, York University

Researchers often hypothesize that the covariance structure for a set of psychological variables is organized according to a set of specific, narrow constructs along with general, more broad constructs. Such a hypothesis is typically examined using either a hierarchical factor model (of which the bifactor model is a special case) or higher order factor model. This talk will explain how these two types of models are distinct conceptually, though they do have a formal mathematical relationship. Issues of model identification, equivalence, and interpretation that are not well-recognized by researchers will be emphasized.

Bozenna Pasik-Duncan, Ph.D.
Professor, Department of Mathematics, University of Kansas, Courtesy Professor of EECS, IEEE Fellow & Distinguished Member of IEEE CSS

This talk focuses on controlled systems described by stochastic differential equations and adaptive control that includes self-optimizing controls for partially known continuous time stochastic systems in both finite and infinite dimensional spaces. For adaptive control of linear systems the weighted least squares estimators that always converge and are strongly consistent under weak assumptions and provide self-optimal adaptive controls under the natural assumptions of controllability and observability will be presented. The current research extends many of the results for control of stochastic systems with Brownian motion to systems with other more justifiable noise processes such as the family of fractional Brownian motions. In almost every application of control, the controlled system has unknown parameters so there are the fundamental problems of identification of unknown parameters and the simultaneous control of the stochastic system. The extension of optimal and adaptive control results to systems driven by processes other than Brownian motion is particularly important because empirical evidence from physical phenomena demonstrate the necessity of other noise processes such as the family of fractional Brownian motions in the mathematical models.

Victoria Savalei: 3:00 p.m., Watson 3 West Reading Room
Assistant Professor, Department of Psychology University of British Columbia(UBC)

There are two geometric representations of multivariate data: the plotting of subjects in the variable space and the plotting of variables in the subject space. While the former is well known and often used, the latter is rarely taught, yet can yield fascinating insights into many multivariate statistical procedures. In this teaching talk, I will review the subject space representation of multivariate data (i.e., viewing variables as vectors). I will show, by reviewing a bit of elementary vector geometry, that most statistical formulas are absolutely equivalent to a definition or a theorem in geometry. Time permitting, various multivariate techniques will then be illustrated from this perspective, most importantly multiple regression (with ANOVA as a special case) and principal component analysis. Even though these will be reviewed during the talk, reviewing concepts such as the definition of a vector, how to compute the length of a vector, how to add and subtract two vectors, how to compute the angle between two vectors, the definition of an inner product, and Pythagorean theorem, will help the audience get more out of the talk!

Neil Heffernan: 3:00 p.m, Joseph R. Pearson Hall, Room 150
Associate Professor, Computer Science Department, Worcester Polytechnic Institute
Co-Director of the Learning Science and Technologies PhD Program and Director of ASSISTments at WPI

In this talk I will describe a shared scientific instrument being used by multiple universities to do cognitive science research on human learning in K-12 schools. The web-based platform is called "ASSISTments" and is used by teachers and their students for 1) nightly homework support and 2) in-class formative assessment and differentiated instruction. Our schools think of ASSISTments as a valuable free public service, for this talk, I will address the different types of ways we use ASSISTments to measure knowledge, and the different randomized controlled experiment being conducted with ASSISTments. I will give multiple examples of the different kinds of ways researchers can use this system. Currently, we have 147 randomized controlled experimenters that conduct very short (minutes-long) experiments comparing different types of feedback. I will discuss an NSF funded 8 year-long longitudinal tracking study we are doing in cooperation with Professor Ryan Baker at Teachers College. Another example I will give will be a US Dept of Ed 10 million dollar math study where, in cooperation with Jim Pellegrino and Susan Goldman at UIC, we are exploring applying cognitive science principles to improve a commonly used middle school math textbook (Connected Math). I will talk about data mining work done with ASSISTments data showing how data collected with the system yields more accurate predictions of student knowledge. I will talk about the Efficacy Trail SRI is doing to see if ASSISTments can be used to raise Smarter Balance test scores. I will also talk about joint work with Zach Pardos in which Zach won an award in the KDD Cup challenge in predicting student performance. Zach’s use of Bayes Nets is to track knowledge is maybe of interest to Dynamic Maps. Finally, I will talk about the online professional development work that Gates Foundation is funding us to do as part of our plan to scale-up to 1 million children and the PD work we are doing with the ~80 teachers a week asking for accounts.

Fei Gu: 2:30 p.m. - 3:30 p.m., Watson 455
Graduate Student, University of Kansas

Estimating multilevel regression models as structural equation models was thoroughly discussed by Bauer (2003) and Curran (2003). Based on the equivalence between structural equation models and state space models (e.g., Chow, Ho, Hamaker, & Dolan, 2010), the state space formulation for the multilevel regression models can be derived by a direct translation of the corresponding structural equation formulation. In this paper, instead of translating the existing structural equation formulation, we introduce a more efficient state space approach to estimating multilevel regression models. Though the state space approach has been well established for decades in the time series literature, it does not receive much attention from educational and psychological researchers. To the best of our knowledge, the state space approach to estimating multilevel regression models is barely known, and (almost) never implemented, by multilevel modelers in education and psychology. We first provide a brief outline of the state space formulation. Then, state space forms for univariate and multivariate multilevel regression models are illustrated, and the utility of the state space approach is demonstrated with both simulated and real examples. It is concluded that the results from the state space approach are essentially identical to those from specialized multilevel regression modeling and structural equation modeling software. More importantly, the state space approach is a much efficient treatment for multilevel regression models.

Kyle Lang: 3:30 p.m. - 4:30 p.m., Watson 455
Graduate Student, University of Kansas

This talk will present the initial findings of a line of research aimed at streamlining missing data analysis by simplifying the use of multiple imputation. A novel analysis strategy was developed (i.e., the SuperMatrix Approach), and a Monte Carlo Simulation Study was conducted to assess the tenability of the proposed technique. Through aggregating multiply-imputed data sets prior to model estimation, the SuperMatrix (SM) Approach was envisioned as a way for researchers to reap the benefits of a principled missing data tool (i.e., multiple imputation), while maintaining the simplicity of complete case analysis. The ability of the SM Approach to produce accurate estimates of model fit, and related quantities, will be discussed. These SM-based estimates of model fit will be judged against estimates derived from two comparison conditions. The first comparison condition was based on a simple, naïve average the multiple estimates of model fit, and the second comparison conditions was based on FIML estimation. Specifically, empirical convergence rates, assessment of direct model fit, and the accuracy of Change in Chi-Squared values derived from each of the three techniques will be scrutinized. Finally, implications, limitations and future directions of the current work will be discussed.

Sunthud Pornprasertmanit: 3:00 p.m., Watson 455
Research Assistant, University of Kansas

This presentation discusses two popular frameworks in statistical influences: the Fisher’s model-based inference and the Neyman’s design-based inference. We show that both frameworks have restricted requirements which are not often satisfied with the current practice of collecting sample in psychological studies. As a result, accurate statistical inferences cannot be made. To mitigate the problem, we propose a practical approach that researchers may use for valid statistical inference with less-restricted requirements. A simulation study is conducted to evaluate the proposed approach. The importance of having a well-defined population is also emphasized in the presentation.

Ed Merkle: 3:00 p.m., Watson 3 West Reading Room
Assistant Professor, University of Missouri

The issue of measurement invariance commonly arises in factor-analytic contexts, with methods for assessment including likelihood ratio tests, Lagrange multiplier tests, and Wald tests. These tests typically require advance definition of group membership and number of groups (via an auxiliary variable), along with specification of the model parameters that potentially violate measurement invariance. In this talk, I study tests of measurement invariance that use individuals' scores (i.e., casewise gradients of the likelihood function) from the estimated factor analysis model. These tests can be viewed as generalizations of the Lagrange multiplier test, and they are especially useful for (1) isolating specific parameters affected by measurement invariance violations, (2) identifying subgroups of individuals that violated measurement invariance, and (3) developing novel statistics geared towards ordinal auxiliary variables. The tests are described in detail and illustrated via both simulation and application.
Tests of Measurement Invariance Without Subgroups: A Generalization of Classical Methods

Roger E. Millsap: 3:00 p.m., Watson 3 West Reading Room
Professor, Psychology, Arizona State University

Latent variable models represent the statistical relationships between measured variables (items, subtests, parcels) and the latent variables assumed to underlie those measures. Suppose that a single latent variable W underlies a set of measures X, and that we create a test score Y that is the sum of the measures. At minimum, we would like it to be true that when we rank order people on Y, that ranking induces a rank-ordering on W that is proper in some sense. Under what conditions will this be true? This is a question of stochastic ordering, and conditions leading to various types of stochastic ordering have been studied previously. We know, for example, the conditions under which certain ordering properties will hold when the measures X are binary items. Little is known however for the case in which the measures X are not binary, and the latent variable model is the common factor model, although this case is widely used in psychology. This case will be addressed here, and the conditions leading to a useful ordering property will be described. It is shown that the common factor model need not imply any useful ordering properties generally, but it can do so under some conditions that have not received attention previously in this context.

Jared Harpole: 3:00 p.m., Holiday Inn Holidome, Brazilian Room
Graduate Student, University of Kansas

The practice of thinning MCMC chains for item response theory (IRT) models has mixed reviews in the literature. Recently, Link and Eaton (2012) found that thinning MCMC chains from a t-distribution produced more biased estimates versus not thinning. The purpose of the present talk is to discuss the results of a simulation study extending the work of Link and Eaton (2012) to include the impact of thinning versus not thinning on parameter estimation when fitting a 2PL IRT model.
This talk will include a brief introduction to Bayesian estimation and Gibbs Sampling, provide background on thinning, and discuss findings and implications from the simulation study.