Mike,
I have analysed a scale with PCM.  There are 30 items. Should I report the threshold estimates for all the items?
Only one item has disorderd thresholds. Should I recategorize the response options for all the scale now? Or should I recategorize the options for only that item?
Does the distance of 1.4 to 5 logit between thresholds also holds for individual items in PCM analysis?
The disorder is vey slight: -1.31, .69, .62. Is it very serious? It�s only for one item.

Thanks

Raschmad

Raschmad, if the items all have the same rating-scale category definitions, then please group the items together as much as possible (ISGROUPS=) to share the same rating-scale category definition. This simplifies communication with your audience.

The very slight disordering of thresholds is probably due to chance, but, even if it is not, your audience will be confused if you recode the rating scale for one item but not for the others. How will they (and you) compare outcomes between the recoded item and the other items?

Thanks Mike,
I meant partial-credit model. the reason why I used PCM instead of rating scale model is that it fits much better. Of course, communication of results to audience from RSM is easier. But in RSM several items misfit or overfit which I can't explain from a substantive point of veiw. But PCM with the same data provides excellent item fit values. the only problem is that each item has unique threshold estimates.

1. Should all threshold estimates be reported for every single item?
2. The distance between some thresholds in some items is .40. They are close but they are all ordered. this means that respondenets could distniguish among the category choices. Is the close distance a problem?

Thanks
Raschmad

Raschmad, Fit is a secondary consideration in the choice between PCM and RSM, particularly if items with similar threshold patterns are grouped together using ISGROUPS=. This is because differences in fit tend to be dominated by accidents in the data. Please see http://www.rasch.org/rmt/rmt143k.htm

Suggestion: model the data with PCM, output the person measures (PFILE=)
model the data with RSM
cross-plot the two sets of person measures
What difference has the choice made to the person measures?

If the difference makes no substantive difference, then RSM

1. If you decide on PCM, then communicate to your audience the substantive implications of the differences in the thresholds.

2.Are you sure that "threshold ordering = ability to distinguish between categories"?
Surely: "fit of responses within category = ability to distinguish between categories"
and
"threshold ordering -> width of category on the latent variable"
Where strong disordering = narrow category, and strong ordering = wide category.
(Remember that thresholds are pairwise between adjacent categories, not global across all categories. If you want to think about global thresholds then please use Rasch-Thurstone thresholds, instead of Rasch-Andrich thresholds.)

Thanks Mike,
I shifted to RSM. The threshold estimates are -1.46, .37, 1.09.
They are not within your guidelines of 1.4 to 5 logits. Shoudl I recategorize?

Raschmad

Raschmad, these are wonderful thresholds.
Please look at my guidelines again. 1.4 logits is only if you want your rating scale to act like a set of dichotomous items. Do you really? That is an unusual requirement for a rating scale.