In article <firstname.lastname@example.org> email@example.com writes:
After running several experiments, I realized that I needed to classify my outcomes SUBJECTIVELY as unfavorable, slightly unfavorable, neutral, slightly favorable or favorable. Although using these results isn't very ideal, time constraints leave very little alternatives. What I need to do is to determine the expected outcome of an event based only on the number of experiment I ran and the subjective rating I gave it.
I was thinking of converting the ratings to a number scale (i.e. unfavorabe=0, slightly unfav=1, ..., favorable=4) and running a T-test at a 95% confidence level. I don't think the T-test is appropriate for subjective "measurements" though.
As you note, the choice of scores is arbitrary and hence a change in mean score is difficult to interpret. On the other hand if the general differences you wish to find are tendencies for the distribution to move in a particular direction, the means *should* differ. The t-test is also fairly robust to deviations from the assumptions of normal distribution. So the t-test is slightly wrong, but not likely to yield wildly misleading conclusions.
Somewhat better alternatives are offered by nonparametric tests such as Wilcoxon two-sample. These have the feature of being independent of the assigned scores. Also, the rank sums that enter have a rather nice interpretation, being related to the number of times an observation from group 1 is bigger that one from group 2 (sum over all pairs).
Make sure you use the Wilcoxon test corrected for *ties* (with only 5 categories, you are bound to get a lot of observations with the same value and you have to use average ranks).
-- O__ ---- Peter Dalgaard Blegdamsvej 3 c/ /'_ --- Dept. of Biostatistics 2200 Cph. N (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918 ~~~~~~~~~~ - (firstname.lastname@example.org) FAX: (+45) 35327907