User Guide for the TIMSS International Database.pdf - TIMSS and ...
User Guide for the TIMSS International Database.pdf - TIMSS and ...
User Guide for the TIMSS International Database.pdf - TIMSS and ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
S C A L I N G C H A P T E R 5<br />
of <strong>the</strong> coding was used in <strong>the</strong> scaling, so this facility in <strong>the</strong> model <strong>and</strong> scaling software was not<br />
used in <strong>TIMSS</strong>.<br />
Letting q be <strong>the</strong> latent variable <strong>the</strong> item response probability model is written as:<br />
T<br />
exp bij<br />
aij<br />
Pr ( Xij = ; A,b,<br />
| )<br />
( q + x)<br />
1 x q = K<br />
,<br />
i<br />
(2)<br />
T<br />
exp b q + a x<br />
<strong>and</strong> a response vector probability model as<br />
f<br />
å<br />
k = 1<br />
( ik ij )<br />
[ ]<br />
T<br />
( x q)= ( q ) x ( bq + A )<br />
; x| Y , x exp x ,<br />
(3)<br />
ì<br />
T<br />
with Y(<br />
q, x)= íå<br />
exp z bq + Ax<br />
îzÎW<br />
[ ( ) ]<br />
where W is <strong>the</strong> set of all possible response vectors.<br />
T I M S S D A T A B A S E U S E R G U I D E 5 - 3<br />
ü<br />
ý<br />
þ<br />
-<br />
1<br />
, (4)<br />
5.3 The Multidimensional R<strong>and</strong>om Coefficients Multinomial<br />
Logit Model<br />
The multidimensional <strong>for</strong>m of <strong>the</strong> model is a straight<strong>for</strong>ward extension of <strong>the</strong> model that<br />
assumes that a set of D traits underlie <strong>the</strong> individuals’ responses. The D latent traits define a<br />
D-dimensional latent space, <strong>and</strong> <strong>the</strong> individuals’ positions in <strong>the</strong> D-dimensional latent space<br />
are represented by <strong>the</strong> vector q=( q1, q2, K, qD<br />
) . The scoring function of response category<br />
k in item i now corresponds to a D by 1 column vector ra<strong>the</strong>r than a scalar as in <strong>the</strong><br />
unidimensional model. A response in category k in dimension d of item i is scored bikd. The<br />
T<br />
scores across D dimensions can be collected into a column vector bik = ( bik1, bik2, K , bikD)<br />
,<br />
T<br />
again be collected into <strong>the</strong> scoring sub-matrix <strong>for</strong> item i, Bi = ( bi1, bi2, K , biD)<br />
, <strong>and</strong> <strong>the</strong>n<br />
T T T collected into a scoring matrix B= ( B1 B2 BI)<br />
T<br />
, , K , <strong>for</strong> <strong>the</strong> whole test. If <strong>the</strong> item parameter<br />
vector, x, <strong>and</strong> <strong>the</strong> design matrix, A, are defined as <strong>the</strong>y were in <strong>the</strong> unidimensional model <strong>the</strong><br />
probability of a response in category k of item i is modeled as<br />
( ij<br />
) =<br />
Pr X = 1;<br />
A, B,<br />
x| q<br />
And <strong>for</strong> a response vector we have:<br />
exp(<br />
bijq+ a¢<br />
ijx)<br />
i<br />
exp(<br />
b q+ a¢<br />
x)<br />
K<br />
å<br />
k = 1<br />
ik ik<br />
[ ]<br />
( ) = ( ) ¢ ( + )<br />
f x; x| q Y q, x exp x Bq Ax<br />
,<br />
ì<br />
T<br />
with Y(<br />
qx , ) = íå<br />
exp z Bq+ Ax<br />
îzÎW<br />
[ ( ) ]<br />
The difference between <strong>the</strong> unidimensional model <strong>and</strong> <strong>the</strong> multidimensional model is that <strong>the</strong><br />
ability parameter is a scalar, q, in <strong>the</strong> <strong>for</strong>mer, <strong>and</strong> a D by one-column vector, q, in <strong>the</strong> latter.<br />
ü<br />
ý<br />
þ<br />
-1<br />
.<br />
(5)<br />
(6)<br />
(7)