Title: Introduction to IRTRasch Measurement with Winsteps Ken Conrad, University of Illinois at Chicago Bar
1Introduction to IRT/Rasch Measurement with
WinstepsKen Conrad, University of Illinois at
Chicago Barth Riley and Michael Dennis,Chestnut
Health Systems
2Agenda
- 1230. Ken Conrad Power-point presentation on
classical - test theory compared to Rasch, includes history
and introduction to the Rasch model. - 215. Break
- 230. Discussion of an application of Rasch
analysis in the measurement of posttraumatic
stress disorder with interpretation of
Rasch/Winsteps output. - 315. Barth Riley Implications and Extensions of
Rasch Measurement. - 415. Break.
- 430. Mike Dennis Practical applications of
IRT/Rasch in SUD screening and outcome assessment - 515. Open discussion and Q A.
- 530. End of workshop.
3The Dream of Rulers of Human Functioning
- Beyond organ function to human functionWHO, 1947
- E.g., quality of life, need to ask person
- 1970s--Physical, social, and mental health
issues - Measuring many constructs requires many
itemstime, , burden - Todayneed for psychometric efficiency w/o loss
of reliability and construct validity
4Prevailing Paradigm, Classical Test Theory
- CTTmore items for more reliability
- Since we seek efficiency (fewer items), items
tend to be where most of the people arearound
the mean. - Resultredundancy at mid-range, few items at
extremes, ceiling and floor effects - Impossible to measure improvement of those in
ceiling and decline of those in floor.
5How children measure wooden rods (from Piaget)
- Classificationseparate the rods from the cups,
the balls, etc. (nominal) - Seriationline them up by size (ordinal)
- Iterationdevelop a unit to know how much bigger
(interval) - Standardizationmake a rule(r) and a process for
determining how many units each rod has - Children know that classification and seriation
are not measurement, Stevens did not nominal,
ordinal, interval, ratio
6Improvement IRT/Rasch measurement and computers
- Rasch measurement model enables construction of a
ruler with as many items as we want at any level
of the construct - The computer enables choice of items based on
each persons pattern of responses. - Each test is tailored to the individual, and not
all of the items are needed.
7Classical Test Theory
- A measure is a sample of items from an infinite
domain of items that represent the attribute of
interest. - Items are treated as replicates of one another in
the sense that differences among the items are
ignored in scaling. - More itemsmore reliability
- Everyone gets the same items
- Answers needed to all items
8- Ranking is sample dependent
- E.g., NBA players, jockeys. Height could be in
the same 1-5 ordinal metric where both a jockey
and NBA player could be rated 5, but this could
only be interpreted with reference to a
particular sample. The sample defines height. - With interval scaling, height defines the
sample. Over 6NBA, under 6jockey.
9Classical Test Theory
- Uses ordinal data as interval.
- Using presumably impermissible transformations,
i.e. using ordinal as interval, usually makes
little, if any, difference to results of most
analyses. - Thus, if it behaves like an interval scale, it
can be treated as one. - Just use the raw scores. Add em up.
- Clean and easy
10Assumption all items are created equal But we
know that is not true. Is that how we measure
potatoes? How about spelling? Items actually
range from Easy-gthard Like addition
-gt division E.g., Guttman 1111100000 Lack
of recent practice on item 5 1111011000 Educated
guess on item 8 1111100100 Slow, nervous
start 0111111000
11No Difficulty Parameter in CTT. What if two
students both got 5 out of 10 correct, but one
got the 5 easiest right and the other the 5
hardest? Easy-gthard Peter 1111100000
Paul 0000011111Do they have the same ability?
Wouldnt you like to get a better idea of what
happened on Pauls test? Did he arrive late?
Were test pages missing? Maybe they were word
problems, and Paul is a foreign student.
12- With CTT, extremely difficult to compare a
persons scores on two or more different
testsusually compare z-scores. - Assumes that samples of both tests center on the
same mean. - Assumes that all of the tests are normally
distributed, which is rarely the case.
13Assumptions of CTT
- CTT take the test, e.g., SD, D, A, or SA on 50
items. What if there is missing data? - CTT uses ordinal scaling, but assumes equal
intervals in the rating scale. However, we know
that distances between scale points usually are
not equal, e.g., The President is doing a good
job. - SD D A
SA - To WWII veterans Do you wear fashionable shoes?
- N SD D A
SA - CTT gives us very limited ability to examine the
performance of our rating scales. Do they really
work the way we want them to?
14Cronbachs Alpha
- Adding items improves alpha, but are they good
items? - Ceiling and floor effects improve alpha.
- CTT assumes homoscedasticitythat the error of
measurement is the same at the high end of the
scale as in the middle or at the low end. - However, ordinal measures are biased, especially
at the extremes where there is much more error.
15To Count gt To Measure
E.G., From counting potatoes to measuring their
quality. From counting number of drinks to
measuring substance use disorders. From summing
Likert ratings to linear, interval measurement.