Dear Mike,
In WINSTEPS manual for interpreting infit and outfit mean square stats. you have written that:
>2.0     Distorts or degrades the measurement system.
1.5 - 2.0     Unproductive for construction of measurement, but not degrading.
0.5 - 1.5     Productive for measurement.
<0.5     Less productive for measurement, but not degrading. May produce misleadingly good reliabilities and separations.

However, this is very different from what is practiced. Many researchers say, .70 to 1.3 for binary items and .60 to 1.4 for Likert items. And they remove items on the basis of these criteria.
Your guidelines are different. You say that values above 2 should be deleted.
Can you please explain?

Thanx
Raschmad

Raschmad, please see http://www.rasch.org/rmt/rmt83b.htm

We must choose the fit criteria to match the situation. Let's use the analogy of a road. When is it smooth enough to drive along? If we are racing sports cars, then even undulations may make the cars "bottom out". If we are driving a bull-dozer, then big pot-holes are no problem. Only crevasses and swamps must be avoided.

So, when we have the luxury of being able to construct a well-formulated test-instrument and administer it under well-controlled conditions, then we can impose tight fit criteria. If we are trying to extract useful information from a poorly-controlled dataset, then our fit criteria will aim at removing the worst of the garbage from the dataset.