Bridging Disciplinary Boundaries (January 11 - 14, 2007)
|Saturday, January 13, 2007: 10:00 AM-11:45 AM|
|Seacliff C (Hyatt Regency San Francisco)|
|Applied Categorical Data Analyses: Binary Logistic, Ordered and Unordered Multinomial Logistic Regression Models to Illuminate What Works Best for Whom|
|Speakers/Presenters:||Mansoor AF Kazi, PhD, University at Buffalo (The State University of New York)|
Tom Nochajski, PHD, University at Buffalo (The State University of New York)
Carrie J. Petrucci, PhD, California State University, Los Angeles
This workshop will present an introduction to three different applied categorical data analyses: when the outcome is dichotomous (logistic regression), when the outcome has three or more unordered categories (nominal multinomial logistic regression), and when the outcome has three or more ordered categories (ordinal multinomial logistic regression) (Hosmer & Lemeshow, 2000; Kazi, 2003; Jaccard & Dodge, 2004). Each presenter will use datasets from their completed evaluations from California, New York and United Kingdom, and discuss real-world applications of the analyses. The didactic approach will be interactive, guiding the workshop participants through the requirements and limitations of each method. Appropriate time for a question-and-answer period will also be included.
With regard to dichotomous outcome variables (e.g. those that improved and those that did not), binary logistic regression will be used to investigate what interventions work and in what circumstances. In each example, the variables that may be influencing the outcome will be identified through bivariate analysis, other research findings, and/or practice wisdom, and then entered in a forward-conditional model. The variables that are actually influencing the outcome are retained in the equation, and those that are significant provide an exponential beta which is interpreted as an odds ratio, indicating the odds of the intervention achieving the outcome where the significant factor(s) may be present.
Nominal or unordered multinomial logistic regression will be explained using datasets that explored the outcome variable as clients who improved, stayed the same, or got worse from pre- to post-test. Ordinal multinomial logistic regression will be explained using driving while intoxicated datasets. Offender status in terms of 1, 2, 3 or more arrests, and prior treatment for alcohol or other drug problems, of none, one, two, three or more will be explored. Variable selection, model fit strategies, and interpretation of the coefficients and odds ratios will be illustrated with both types of multinomial logistic regression.
A recent search in Social Work Abstracts revealed the following number of articles when searching on type of analysis: regression (not including logistic regression) – 1,024 articles; logistic regression – 290 articles; and multinomial logistic regression – 15 articles. This bias towards the use of parametric statistics is also evidenced in social work research texts that seldom include in-depth discussions of non-parametric statistical tests or categorical data analyses beyond the chi-square test of independence. Yet, we know that in many circumstances, the data we are analyzing is not normally distributed, and the variables we are examining are commonly categorical. Thus, the use of parametric analyses often pushes the boundaries of acceptable usage of these statistics.
Regression analysis of categorical data can help researchers and practitioners understand outcomes, what predicts those outcomes, and characteristics of clients who are achieving those outcomes verses clients who are not; but if the search in Social Work Abstracts is any indication, technical assistance is needed to provide social work researchers with the requisite skills and knowledge to carry out these often useful but underutilized categorical analyses.
See more of Workshop
See more of Bridging Disciplinary Boundaries (January 11 - 14, 2007)