UBC SEM LAB in the UBC Department of Psychology conducts cutting-edge research in the area of structural equation modeling (SEM) for psychological and behavioral science data. SEM is a powerful statistical modeling tool that allows researchers to test complex theories involving multiple observed and latent variables. SEM can be thought of as a set of simultaneous regression analyses in which observed variables (e.g., scores on several aptitude tests) can be predicted from latent constructs that are not directly observable (e.g., general intelligence). These latent constructs can in turn predict each other, or be predicted by other observed (exogenous) variables. SEM subsumes many other multivariate statistical techniques, such as regression, path analysis, and confirmatory factor analysis. One benefit of SEM is that it allows researchers to estimate the relationships between latent variables controlling for measurement error and for item specificities. Another benefit of SEM is that it produces a test statistic for the overall model, allowing researchers to assess how consistent the observed data are with the proposed model. With increasing availability of software packages (we prefer the lavaan package in R), SEM has become very popular in psychology and other behavioural sciences. For an introduction, see Savalei and Bentler (2005).
The following are some research questions that are investigated in our lab:
- How can SEMs be estimated and evaluated more reliably with difficult kinds of data, such as incomplete data, nonnormal data, categorical data, data with a small number of observations, or some combination of these? We develop new methods for estimating and testing SEMs under these difficult conditions. We also evaluate existing methods relative to each other in large simulation studies and provide guidance to applied researchers based on the results. Many diverse ongoing projects fall into this category of research.
- What are the properties of popular approximate indices of fit used in SEM? Indices such as CFI and RMSEA are widely used by applied researchers, but their meaning remains poorly understood.
- How can SEMs best be used to model response biases in personality data? In particular, we are interested in the factor structure of those psychological scales that have both positively worded and reverse worded Likert items. Many different types of SEMs have been proposed to describe the structure of such scales, but little research has been done into whether these models are distinguishable and if so, which ones best describe real data. We are also interested in whether alternative scale formats produce scales with a simpler factor structure and eliminate some of the response biases created by the Likert items.