>Charlotte Manly wrote: (in response to my query...) > >> If that scale is well-constructed, it _is_ an interval scale >> (or awfully close to one). Measurement issues are beyond my >> area of expertise, but as I understand it, if you only label >> the endpoints, subjects tend to treat it as an interval >> scale. If you carefully label each code number (and can show >> that subjects perceive each label as about equally far from >> its neighbors) then it's also an interval scale. On the >> other hand, if your labels are something like, "never," >> "almost never," "pretty often," and "always," you'll have a >> hard time arguing that that's an interval scale. > > Thanks for the comments. I should have been more clear. I'm aware >that if the scale is the result of careful scale construction, the result is >a genuinely interval scale (or close enough for jazz...). I meant to direct >my question to "scales" (I hesitate to use the term for these as it begs the >question to some degree) for which that scale construction was NOT done. >That is, I meant to ask about the kind of thing that you referred to in the >last sentence of your reply. > > I have _always_ believed that one can treat carefully constructed >scales (e.g., the product of the Thurstone technique) as interval. Is it >also true that we can do so for the kind of "thrown-together scales" we (or >our students) so often use? > >Paul Smith >Alverno College >Milwaukee
If by "thrown together" you mean my third example, no, a t test wouldn't be appropriate. That is clearly only an ordinal scale, not interval. (The difference between never and almost never is a lot smaller than the difference between almost never and pretty often.) If you mean my first example, where I only labelled the endpoints, I would hazard to call that an interval scale. Rick Froman said it well: the mean of an ordinal scale isn't meaningful. If you can't assume equal intervals between data values, then the mean will be misleading because it weights outliers more heavily based on their distance from the center. The median just identifies the point at which have the data is larger and half smaller. So with ordinal data, better to use the median to report summary stats in this case, and the corresponding nonparametric statistical test (Mann-Whitney or Wilcoxon rank sums for independent samples t, Wilcoxon matched-ranks for matched samples t). Charlotte -- ====================================================== Charlotte F. Manly, Ph.D. | Psychological & Brain Sciences Assistant Professor | 317 Life Sciences Bldg ph: (502) 852-8162 | University of Louisville fax: (502) 852-8904 | Louisville, KY 40292 [EMAIL PROTECTED] http://www.louisville.edu/a-s/psychology/ http://www.louisville.edu/~cfmanl01 --- You are currently subscribed to tips as: [EMAIL PROTECTED] To unsubscribe send a blank email to [EMAIL PROTECTED]
