[Reply to OP and to list. -- DFB.] >From your description, I'd guess you were overinterpreting the idea of "ordinal" in a sense like "ONLY ordinal". I'd be very surprised if your 7 scale points were not approximately equally spaced in some useful frame of reference. (By "approximately equally" I mean something like "ratio of <maximum interval between adjacent points> to <minimum interval ...> is less than, say, 2.)
In any case, you have 72 values for each subject, which means you can legitimately [!] do some averaging. So do a two-way ANOVA: 4 groups (between Ss) by 72 ratings (within Ss). Now I'm not altogether clear on what your data actually are. Perhaps (A) you have one rating for each S for each day; but your language is also consistent with (B) scores being taken several times a day and the data you've tabulated representing the maximum score that day. If (A) is the case, what you'll get is an analysis of means rather than of maxima; if (B) is the case, you'll have an analysis of daily maxima. If (A) is the case, then what you have in your cross-tabulation is only one score: the maximum observed among the 72 scores. It strikes me that this would be throwing away a lot of information; since surely you would like to be able to distinguish between (1) a person whose maximum score is "7" and all the other 71 scores are less than "4"; and (2) a person whose maximum score is "7" and 58 of the 71 scores are equal to "7" while the rest do not descend below "6". Your first sentence refers to "differences in distributions", not in averages or in maxima; and perhaps that's what you may want to attempt (also?). But tests for differences among distributions are notably insensitive to real differences unless one has a lot of data, so for such a purpose you should probably be using ALL the data in constructing distributions, not just the maxima of certain chunks of the data set (however well-defined the chunks may be!). On Wed, 18 Feb 2004, MMMM wrote: > I'm looking for methods of testing differences in distributions > between groups scored on an ordinal scale. > > I have 4 groups of subjects scored daily for 72 days on a 1-7 scale. > I would like to see whether there is a difference between groups in > terms of the maximum score achieved over these 72 days. I have taken > the maximum score for each subject and created a crosstab table. > (Counts of max. score by group.) > > Since the data are ordinal, I used Mantel-Haenszel measurements for > general association as opposed to simply the Chi-Square value. I don't understand this statement. If you're referring to the usual chi-square test for independence in a two-way table, the classifications are nominal: don't even have to be ordinal. Unless the M-H measurement to which you refer are more sensitive than chi-square. (I'm not familiar with that analysis.) > I found that there is an association between maximum score achieved > and group membership. > > I would now like to determine which groups are statistically > different, and which are not. I'm looking for a way to compare these > groups in a method similar to a multiple comparisons method that you > would do for an ANOVA. E.G. I found they're different, now which ones > are different? However, I'm having difficulty finding any way to do > this. I'll just point out that in MINITAB, a one-way ANOVA will produce, in character graphics, a nice display of 95% confidence intervals, in which it's quite easy to _see_ (as distinct from having to work out the meanings of numerical values reported only digitally) differences between the several groups. Good luck! -- DFB. ------------------------------------------------------------ Donald F. Burrill [EMAIL PROTECTED] 56 Sebbins Pond Drive, Bedford, NH 03110 (603) 626-0816 . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
