"Li,Tom" <Li.Tom@ENDO.COM> replied:

>> To answer your question, I am hypothesizing that Variable A (measured >> from 1-5) will have a greater prevalence than variable B(measured from >> 1-7). I agree that the ordinal nature of the anchors is problematic, >> but only if the anchors are not continuous. >> The >> responses are behaviors, so, e.g. >> >> 1- don't do at all >> 2 - do a little bit >> 3 - do an average amount >> 4- do a lot >> 5- do all the time >> >> So basically I am saying that people engage in behavior A more than >> behavior B, except that behavior B is measured on a 7 point scale. If I >> assume that these are continuous, then can I not simply stretch the >> scale to 1-7, such that, for example, 1 stays a 1, 2 becomes 2.5, 3 >> becomes 4, 4 becomes 5.5, and 5 becomes 7? I am not sure about the >> implications of this, though, so I had posted to see if there are other >> options or if this would be problematic.

> Can you fit a regression model between the two scales and come up some > kind of mapping schedule?

I hate to be rude (okay, I am rude, it's just that I hate it), but I think that's a really bad idea. A regression model will totally ignore the Likert nature of the scale, and end up sticking inappropriate levels together.

The real problem is the desire to compare two scales which were set up so that a comparison is particularly awkward. The survey design is making the analysis so much harder.

Rather than try to stretch a 5-part scale up to 7, or vice versa, I would simplify both. Perhaps a simple yes/no, based on one of the following, depending on the goals of the project:

don't do at all vs. do at least some do seldom or not at all vs. do average or a lot do average or less vs. do more than average

Then the original poster would have a 2x2 grid that would be straightforward to analyze in PROC FREQ.

David -- David Cassell, CSC Cassell.David@epa.gov Senior computing specialist mathematical statistician