ASA Connect

Hello Faisal, My humble answer to your question would be "it depends". I say this because I realize that for various reasons one sometimes has no choice but to include an independent variable generated from a Likert scale question into the regression model. First, of course, if the response variable has been measured at the ordinal level one should use an Ordinal Logistic Regression model. Second, if instead one is including a predictor variable into the regression model that has been measured using a Likert scale, then 3 options come to my mind. 1). If it was a true Likert scale measuring a feeling or attitude, such as pain level on a 5 or 7 or 10 point scale, then the data generated from this should probably be entered as dummy variables. That is because the assumption that this scale has equal intervals between its values is rarely valid. I speak from personal experience on this one; the difference between my 4 and 5 in pain (on a 5 point scale) is very different from my sister's--it is subjective. A second common issue with a Likert scale used for this type of measurement is that some personality types are loathe to ever check the extreme values, while others tend to most often check the extreme values. Again, a challenge for the validity of such measurements. Of course, using dummy variables for this purpose is easier when the scale consists of only a few vales. 2). Add the ordinal predictor to the regression model, assuming it is an interval scale. Then remember to use caution when evaluating the relational between it and the response variable (assuming its coefficient is significant). Also, under this set of conditions, use caution when interpreting model predictions. 3). Use a different type of regression model altogether. Use a generalized linear model, such as one described in "Regression Models with Ordinal Variables," by Christopher Winship and Robert D. Mare (American Sociological Review, 1984). An old, but good, paper. Generalized Linear Models is an entire category of regression models, some of which make none, or very few, assumptions about the types of distributions involved. Your question continues to be intriguing and complicated and challenging to most of us, I am sure. ------------------------------ Gretchen Donahue ------------------------------