5 - Best practices with Rasch Analysis
Here we will discuss some general principles about how to implement
the Rasch Model for your data.
- It is easiest to run the model with a large sample and few
items. Samples of less than 200 people can be problematic,
especially if there are a large number of items (20+) and responses to
the items are not distributed evenly.
- In general, minimal data adjustment is best. Always
try to see if you can make as few adjustments to your data as possible.
This means the outcome of your analysis will better support your
original survey instrument.
- Try to make testlets only among items that are conceptually
similar. It is easier to justify combining items that are very
similar (for instance, “feeling depressed” and “feeling anxious”) than
items that are extremely different (for instance, “walking 100m” and
“remembering important things”). If you have high correlation among
items that are very conceptually different, this may point to other
problems with the survey instrument that should likely be
- When recoding, try to collapse only adjacent
thresholds For instance, if you see that thresholds are
disordered in the pattern 2, 1, 3, 4, it is natural to try to collapse
thresholds 2 and 1 because they are adjacent. It would not make sense to
collapse thresholds 2 and 4 because they are not adjacent.
- When recoding, leave the first response option alone and do
not recode it. This first response option represents an answer
in the MDS of “no problems” or “no difficulty”, and it is best to leave
this response option alone as a baseline. There is a larger conceptual
difference between having “no problems” and “few problems” than there is
between having “few problems” and “some problems”, so it makes less
sense to collapse the first two response options than it does to
collapse the 2nd and 3rd response options.
- It is normal to have to run many iterations of the
model in order to find a solution that works best, especially
if you have many items in your instrument.