On 14-Apr-2004, [EMAIL PROTECTED] (Douglas Rugh , D.Rugh) wrote: > > I have a question about the levels of measurement in a multi-level > > regression model. The sampling frame is a representative cross > > section of students throughout the United States at each grade level. > > The three selection stages are: (1) geographic areas or primary > > sampling units (PSUs), (2) schools within PSUs, and (3) students > > within sampled schools. > > > > As for my model, I have an individual level measure of alcohol > > consumption which is ordinal. I also have a school level measure of > > alcohol consumption, which is an aggregate of the individuals within a > > school--also ordinal. Then I have a Primary Sampling Unit variable on > > a scale from 1 to 72. This third level measure is the one I am > > confused about. It is not an ordinal measure, it is a nominal > > measure. How do I handle that in the regression? It is unreasonable > > to dummy code this variable.
I recommend that you fit a regression decision tree to your data. Unlike normal function regressions, decision trees are insensitive to the scaling of variables (since it just divides the rows into groups), and categorical variables such as your PSU are handled naturally without having to generate dummy variables. Decision trees also do a very good job of handling variable interactions, and they produce models that are easily (visually) understood. If you are interested, I will be happy to run your data through my DTREG program and send you the results. -- Phil Sherrod (phil.sherrod 'at' sandh.com) . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
