itf {irtoys} | R Documentation |
Returns a statistic of item fit together with its degrees of freedom and p-value. Optionally produces a plot.
itf(resp, ip, item, stat = "lr", theta, groups, do.plot = TRUE, main = "Item fit")
resp |
A matrix of responses: persons as rows, items as columns, entries are either 0 or 1, no missing data |
ip |
Item parameters: a matrix with one row per item, and three columns: [,1] item discrimination a, [,2] item difficulty b, and [,3] asymptote c. |
item |
A single number pointing to the item
(column of resp , row of ip ),
for which fit is to be tested |
stat |
The statistic to be computed, either "chi"
or "lr" . Default is "lr" . See details below. |
theta |
A vector containing some viable estimate of the
latent variable for the same persons whose responses are given
in resp . If not given (and group is also missing),
EAP estimates will be computed from resp and ip . |
groups |
An object produced by function grp .
If not given, grp will be applied on theta with its
default values. |
do.plot |
Whether to do a plot |
main |
The title of the plot if one is desired |
Given a long test, say 20 items or more, a large-test statistic of item fit could be constructed by dividing examinees into groups of similar ability, and comparing the observed proportion of correct answers in each group with the expected proportion under the proposed model. Different statistics have been proposed for this purpose.
The chi-squared statistic
X^2=sum_g(N_gfrac{(p_g-π_g)^2}{π_g(1-π_g)},
where N_g is the number of examinees in group g, p_g=r_g/N_g, r_g is the number of correct responses to the item in group g, and π_g is the IRF of the proposed model for the median ability in group g, is attributed by Embretson & Reise to R. D. Bock, although the article they cite does not actually mention it. The statistic is the sum of the squares of quantities that are often called "Pearson residuals" in the literature on categorical data analysis.
BILOG uses the likelihood-ratio statistic
X^2=2sum_g<=ft[r_glogfrac{p_g}{π_g} + (N_g-r_g)logfrac{(1-p_g)}{(1-π_g)}right],
where π_g is now the IRF for the mean ability in group g, and all other symbols are as above.
Both statistics are assumed to follow the chi-squared distribution with
degrees of freedom equal to the number of groups minus the number of
parameters of the model (eg 2 in the case of the 2PL model). The first
statistic is obtained in itf
with stat="chi"
,
and the second with stat="lr"
(or not specifying stat
at all).
In the real world we can only work with estimates of ability, not with ability
itself, so the approach is a bit circular in defining the groups.
I have tried to offer some
extra flexibility with the arguments theta
nor group
:
theta
nor group
is specified, item.test
will compute EAP estimates of ability for the proposed model, group them,
and use medians for "chi"
or means for "lr"
. This is the
approximate behaviour of BILOG (assuming stat="lr"
).
qrs
and passing
them to item.test
as theta
.
"chi"
, means for "lr"
)
can be overriden by preparing the groups with grp
and passing
them to item.test
as group
. In that case, theta
is not
needed.
If the test has less than 20 items, item.test
will issue a warning.
For tests of 10 items or less, BILOG has a special statistic of fit, which
can be found in the BILOG output. Also of interest is the fit in 2- and
3-way marginal tables in package ltm
.
A list of:
statistic |
The value of the statistic of item fit |
dfr |
The degrees of freedom |
pvalue |
The p-value |
Ivailo Partchev
S. E. Embretson and S. P. Reise (2000), Item Response Theory for Psychologists, Lawrence Erlbaum Associates, Mahwah, NJ
M. F. Zimowski, E. Muraki, R. J. Mislevy and R. D. Bock (1996), BILOG–MG. Multiple-Group IRT Analysis and Test Maintenance for Binary Items, SSI Scientific Software International, Chicago, IL
data(Scored) p.2pl <- est(Scored, model="2PL", engine="ltm") fit <- itf(resp=Scored, ip=p.2pl, item=7)