1 A Proportional Odds/Hazards Approach to Analyzing Likert Scale Data

Linear scoring coupled with parametric regression or ANOVA often is used to analyze responses measured on a Likert scale. Alternatively, we have had success in numerous cases by modeling the same data under relaxed assumptions of the proportional odds model. This is implemented conveniently by procedure LOGISTIC, where a test of the proportional odds parallelism assumption is displayed automatically. When this test fails, a satisfactory model sometimes results by switching from the default cumulative logit link to use of the complementary log-log link, implemented by option LINK=CLOGLOG. Model fit by means of the log-log link also can be assessed using the CLOGLOG option, after minimal data manipulation. Either link induces the equivalent of a proportional hazards model. Interpretation of proportional odds and hazards models in the context of a Likert scale response is illustrated using data obtained from a survey of 849 teachers and administrators involved in implementing a new teaching program in their schools. Of 8 items relating to the perceived degree of implementation of the program, the proportional hazards model represents responses for seven items very well. This paper is directed at survey specialists who employ Likert scale instruments and who are familiar with logistic regression