Rankings and partial rankings, with or without ties, can be conveniently analyzed using ISGROUPS=0 and STKEEP=No

 

Each row is an element or object to be ranked.

 

Each column is a set of rankings by a respondent. In the item label, place any interesting demographics about the respondent doing the ranking.

 

ISGROUPS=0  - Each respondent (column) has their own "ranking scale". This is equivalent to the Partial Credit model.

 

STKEEP=No - tied ranks are OK, but if the ranking goes: 1, 2, 2, 4, 5, .... in the data then, from a Rasch perspective this should be: 1, 2, 2, 3, 4, ....  which is what STKEEP=NO does for us.

 

The elements (rows) will have measures and fit statistics indicating respondent preference. The fit statistics for the respondents (columns) will indicate the extent of agreement of each respondent with the consensus. The measures for the respondents are usually meaningless (or the same) and can be ignored.

 

Fit statistics, DIF and DPF analysis, and contrast analysis of residuals are all highly informative.

 

Elements rated by only one respondent will be correctly measured (with large standard errors), except if that element is ranked top or bottom by that respondent. When that happens, construct a dummy ranking column in which the extreme elements are not-extreme. Give that ranking a low weight with IWEIGHT=.

 

If every ranking set includes every element, and ties are not allowed, then the elements can be columns and the respondents can be rows. ISGROUPS=0 is not required.

 

Example:

In the data, as collect, the ranked objects are columns, the respondents are rows, and the rankings as numbers.

 

Set up a standard rating-scale analysis of these data.

 

Run an analysis in Winsteps. Ignore the results except to verify that the data have been input correctly.

 

Now we want the respondents as columns and the objects as columns:

 

Winsteps "output files", "transpose". TRPOFILE=

 

Edit the transposed control file with ISGROUPS=0 and STKEEP=NO

 

Now run the Winsteps analysis. The results should make sense. The objects (rows) should have measures and fit statistics indicating respondent preference. The fit statistics for the respondents (columns) will indicate the extent of agreement of each respondent with the consensus.