Research

Our software products are structural equation modeling (SEM), hierarchical linear and nonlinear modeling (HLM) and item response theory (IRT). The software packages are authored by respected figures in the academic world, including Karl Jöreskog and Dag Sörbom, Steve Raudenbush and Tony Bryk, Darrell Bock, David Thissen, Don Hedeker and Robert Gibbons. SSI is continuously focused on improving and extending the capabilities of our programs. In addition to an on-going implementation of new research tools that result from the research of individual authors, we are also involved in specific research and development funded by SBIR grants. SBIR is a highly competitive program that encourages small business to explore their technological potential while also allowing to benefit from the commercialization of the resulting products. Details on the aims of the individual projects we are currently involved in, followed in each case by a short motivation for the research, can be found below.

Structural equation and multilevel modeling projects:

  • The first phase of a project, titled "Longitudinal Analysis of Complex Survey Data with LISREL," was recently completed. Funded by NIAAA grant 1 R43 AA014999-01, the focus of this project was on the fitting of models collected from longitudinal surveys using complex sample designs.
  • This interest reflects the needs of policy makers and researchers for in-depth studies of social processes over time. Many of the longitudinal studies on alcohol abuse, a main interest of the National Institute on Alcohol Abuse and Alcoholism, are characterized by the fact that each individual longitudinal record is short, while the number of records is large. Consequently, software for structural equation modeling (SEM) and multilevel modeling is used to analyze these longitudinal data sets. Under this grant, SSI staff investigated and reported on ways that the existing LISREL program may be further developed to provide researchers with an integrated platform for the analysis of longitudinal data from complex sampling designs. Phase 1 results, which include the ability to handle structural equation models, multilevel models and generalized linear models with continuous, ordinal and nominal outcome variables, have been implemented in LISREL 8.72, which was released in April 2005. In addition to new program features, an additional draft manual, documenting new features and containing annotated examples, has been made available to users in PDF format. SSI has recently applied for a Phase II grant from NIAAA to extend this research over the next 2 years.
  • After successfully completing a SBIR Phase I project titled "SUPERMIX: Analysis of Clustered and Longitudinal Data", SSI is currently busy with research on a Phase II contract with the National Institute of Mental Health (Phase II contract no N44MH32056) to develop an integrated program called SUPERMIX which will accommodate two- and three-level data for continuous, binary, ordinal, nominal, time to event, and count data. The SUPERMIX program will provide these models in a user-friendly format so that all mental health researchers can have easy access to these new technologies, and is scheduled for release at the end of 2006.

    Historically, mental health research has been constrained by a lack of adequate statistical methods. With the advent of mixed-effects regression models, the complex multi-level sampling nature of these data (i.e., longitudinal and/or clustered sampling designs) can be accommodated in the statistical analysis. Nevertheless, traditional mixed-effects regression models for normally distributed outcomes often fail to accommodate the complexities of mental health research. Data collected in mental health research are often rated on binary, ordinal, or nominal scales. In mental health services research, service utilization expressed as a count is often the primary outcome of interest. The focus of this application is on the development of integrated statistical software that will provide a single platform upon which mental health researchers can analyze their data using methods that are appropriate for both the multilevel nature of their designs and distributional form of their outcomes.

Item Response Theory:

  • A Phase I proposal by SSI currently under review, titled "MEDPRO: An IRT-based Outcomes Research System" is ultimately aimed at producing a system for outcomes research, provisionally called MEDPRO, based on two intercommunicating computer applications that will be of particular interest to medical outcomes researchers and researchers active in other areas of behavioral measurement. One of these is a comprehensive program for item calibration and patient scoring, and the other program has facilities for the creation of computer displayed items for PROs instruments, for adaptively presenting these items to patients, and for capturing item responses and other patient information.

    Although health status following medical treatment has traditionally been measured by objective outcomes, there is now recognition of the importance of subjective reports of outcomes by the patient, by proxies interviewing the patient, such as psychologists, research nurses, or medical caseworkers, and by providers of care for the patient. Up to now, medical outcomes research has relied upon simple methods of classical test theory for combining responses to such instruments into measures of health status. Although the limitations of classical test theory had long been apparent in psychological and educational testing, it was not until the 1970s and '80s that modern statistical methods were successfully brought to bear on all these problems. The result of this work has been the development of a wide-ranging methodology for analysis of item response data in their many forms. The mathematical and statistical results in this field, and the principles by which they are implemented in practice, are now called item response theory (IRT). In addition to the available software packages for IRT available from SSI and other publishers, it is also possible to fit subsets of the models with existing software not primarily designed for IRT analysis. Examples are the structural equation modeling program LISREL and the hierarchical linear and nonlinear program HLM. Since none of the existing programs was designed for or is specifically geared towards mental health research, the ultimate objective of this proposal is to implement these two closely articulated programs in a system that will be particularly useful to workers in the outcomes research fields and other areas of behavioral measurement.