Can Stata estimate a Rasch model?
|
Title
|
|
The Rasch model in Stata
|
|
Author
|
Jeroen Weesie, Department of Sociology/ICS, Utrecht University
|
|
Date
|
January 1999; minor revisions July 2011
|
Background
The Rasch model is one of the dominant models for binary items (e.g.,
success/failure on math problems) in psychometrics. Let y(ij) be the binary
response, where i=1,...,n, n = #subjects, j=1, ..., m, m = #items. (Sometimes
“unbalanced designs” are used in which subjects respond to only a subset of
the items. The description below applies without modification.) The Rasch
model can be written as a logit-linear model:
logit Pr(y(ij)=1 | eta(i)) = eta(i) − theta(j)
Here eta(i) can be interpreted as a person’s ability parameter and theta(j)
as an item-difficulty parameter. Moreover, it is assumed that, conditional on
eta(i), the y(i*) are independent (“local independence”). Actually, Rasch, a
Danish statistician, gave an axiomatic derivation of the model in the
1960s, in which, next to local independence, the main characterizing
properties were that the { y(+j) , y(i+) } form a sufficient statistic for {
eta, theta }.
|
Technical note:
When treating the etas and thetas as parameters (fixed effects), it has
long been known that the maximum likelihood estimators are
inconsistent in the standard asymptotic setup (n→infinity, m fixed).
This may become somewhat intuitive if we observe that with each
extra subject we have m extra observations but also one extra eta
parameter to estimate. Thus, intuitively, the number of observations
and hence, the amount of information per parameter, stay almost
fixed with increasing n.
A famous example of inconsistent maximum likelihood estimators with a
similar structure is the following. Let y(ij) ~
Normal(mu(i),v2), i=1, ..., n, j=1, 2. The mu(i) parameters
and v2 are to be estimated. The maximum likelihood
estimator for v2 is
sum{i,j} (y(ij) − y(i.))2 / (2n)
with expected value and also with probability limit
(1/2)v2, not v2.
|
In the 1980s, Andersen showed the conditional maximum likelihood (CML)
estimator of theta(j), where conditioning occurs on the subject’s
scores y(i+), is efficient and is actually asymptotically normal distributed,
with all the nice properties of ml-estimators in regular situations. For
instance, a conditional likelihood-ratio test has properties analogous to
those of the standard likelihood-ratio test. While other estimators have been
investigated, the CML estimator is the one most widely used for the
fixed-effects case—the etas are treated as parameters (Fischer and
Molenaar 1995).
Fitting the Rasch model in Stata
Consider the most probable case; you have a dataset in which the cases
refer to subjects, and the responses on m items are stored in m dichotomous
variables item_1, ..., item_m. You can obtain CLM estimates of the theta
parameters of the Rasch model using the conditional logit fixed-effects
estimator in
clogit
(xtlogit, fe
is equivalent to typing clogit).
This command requires all responses be stored in separate observations, while a
“group” variable is used to identify the observations that belong to the same
subject. This can be accomplished with the
reshape command.
Finally, you can describe the Rasch model as a “clogit model”
with m covariates x(ijk), k=1, ..., m, so that for all i, x(ijk) = −1
if j=k, and 0 otherwise. The regression coefficient of x(..k) is just
theta(k).
In a conditional logit model, effects of variables that are constant within
groups (subjects) are not identified. Since the x(..k) variables are a
complete classification, i.e., they sum to −1 in each record, the
thetas are not identified. If all x’s are included in the model, Stata
deals with this problem automatically by removing one of the x variables. You
can also withdraw one of the x variables from the set of predictor variables.
The associated theta parameter could be interpreted as being fixed at 0.
The outline for estimating the CML estimates of the theta parameters of
the Rasch model follows.
(1) If necessary, transform the data into long format.
(2) Create explanatory variables for the thetas.
(3) Estimate the theta parameters.
The “Bayesian” approach: random effects
It may come as only a little surprise that psychometricians also studied a
Gaussian random-effects estimator for the thetas. In Stata, this is obtained
by the random-effects panel estimator, as computed by the
xtlogit, re
command. This command requires the same data
organization as clogit.
Example
Consider data on 120 subjects who responded to 10 math problems, coded as 1
(correct) and 0 (incorrect). We want to know whether the 10 math problems
involve a one-dimensional scale so that the items can be ordered with
respect to difficulty, irrespective of the subjects’ abilities.
. use http://www.stata.com/support/faqs/dta/raschfaq, clear
. describe
Contains data from http://www.stata.com/support/faqs/dta/raschfaq.dta
obs: 120
vars: 11 19 May 2005 07:47
size: 1,920 (99.9% of memory free)
-------------------------------------------------------------------------------
storage display value
variable name type format label variable label
-------------------------------------------------------------------------------
math1 byte %9.0g Correct math item 1
math2 byte %9.0g Correct math item 2
math3 byte %9.0g Correct math item 3
math4 byte %9.0g Correct math item 4
math5 byte %9.0g Correct math item 5
math6 byte %9.0g Correct math item 6
math7 byte %9.0g Correct math item 7
math8 byte %9.0g Correct math item 8
math9 byte %9.0g Correct math item 9
math10 byte %9.0g Correct math item 10
subj_id int %9.0g
-------------------------------------------------------------------------------
Sorted by:
First, to get some feeling for the data, we describe the items
. summarize math*
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
math1 | 120 .8083333 .3952626 0 1
math2 | 120 .775 .4193332 0 1
math3 | 120 .6833333 .4671266 0 1
math4 | 120 .55 .4995797 0 1
math5 | 120 .5333333 .5009794 0 1
-------------+--------------------------------------------------------
math6 | 120 .45 .4995797 0 1
math7 | 120 .425 .4964157 0 1
math8 | 120 .3 .460179 0 1
math9 | 120 .2583333 .4395535 0 1
math10 | 120 .1666667 .3742406 0 1
They are already labeled so that the proportion of students that
answer the item correctly decreases. The subjects are aptly described via the
total number of correct answers.
. egen score=rowtotal(math*)
. tabulate score
score | Freq. Percent Cum.
------------+-----------------------------------
0 | 3 2.50 2.50
1 | 5 4.17 6.67
2 | 4 3.33 10.00
3 | 16 13.33 23.33
4 | 21 17.50 40.83
5 | 24 20.00 60.83
6 | 22 18.33 79.17
7 | 13 10.83 90.00
8 | 4 3.33 93.33
9 | 6 5.00 98.33
10 | 2 1.67 100.00
------------+-----------------------------------
Total | 120 100.00
To fit the Rasch model, we first have to
reshape the data.
. reshape long math, i(subj_id) j(item)
(note: j = 1 2 3 4 5 6 7 8 9 10)
Data wide -> long
-----------------------------------------------------------------------------
Number of obs. 120 -> 1200
Number of variables 12 -> 4
j variable (10 values) -> item
xij variables:
math1 math2 ... math10 -> math
-----------------------------------------------------------------------------
We now have to generate the predictor variables for the thetas:
. forvalues num =1/10{
gen Th`num' = -(item==`num')
}
To illustrate what the data now look like, we
list the cases.
* set format to compress the output of list
. format math item Th* %4.0f
* sort within subj_id on the identifier item of the math problem.
. sort subj_id item
* invoke list with options that improve readability.
. list math item Th* if sub==1, nodisplay noobs nolabel
+--------------------------------------------------------------------------+
| math item Th1 Th2 Th3 Th4 Th5 Th6 Th7 Th8 Th9 Th10 |
|--------------------------------------------------------------------------|
| 1 1 -1 0 0 0 0 0 0 0 0 0 |
| 1 2 0 -1 0 0 0 0 0 0 0 0 |
| 1 3 0 0 -1 0 0 0 0 0 0 0 |
| 0 4 0 0 0 -1 0 0 0 0 0 0 |
| 0 5 0 0 0 0 -1 0 0 0 0 0 |
|--------------------------------------------------------------------------|
| 0 6 0 0 0 0 0 -1 0 0 0 0 |
| 0 7 0 0 0 0 0 0 -1 0 0 0 |
| 0 8 0 0 0 0 0 0 0 -1 0 0 |
| 0 9 0 0 0 0 0 0 0 0 -1 0 |
| 0 10 0 0 0 0 0 0 0 0 0 -1 |
+--------------------------------------------------------------------------+
. list math item Th* if sub==2, nodisplay noobs nolabel
+--------------------------------------------------------------------------+
| math item Th1 Th2 Th3 Th4 Th5 Th6 Th7 Th8 Th9 Th10 |
|--------------------------------------------------------------------------|
| 1 1 -1 0 0 0 0 0 0 0 0 0 |
| 1 2 0 -1 0 0 0 0 0 0 0 0 |
| 0 3 0 0 -1 0 0 0 0 0 0 0 |
| 0 4 0 0 0 -1 0 0 0 0 0 0 |
| 1 5 0 0 0 0 -1 0 0 0 0 0 |
|--------------------------------------------------------------------------|
| 1 6 0 0 0 0 0 -1 0 0 0 0 |
| 1 7 0 0 0 0 0 0 -1 0 0 0 |
| 0 8 0 0 0 0 0 0 0 -1 0 0 |
| 0 9 0 0 0 0 0 0 0 0 -1 0 |
| 0 10 0 0 0 0 0 0 0 0 0 -1 |
+--------------------------------------------------------------------------+
Thus, we see that the subject with subj_id=1 has 10 records
associated with his response: the score for each item is in variable
math. The remaining 10 variables, Th1–Th10,
which are included in the output, are the independent variables. The
Th#s are nearly everywhere 0. We see that the response on the
first item (item>==1) is modeled only with the predictor variable
Th1, the response on the second item (item>==2) with the
second predictor Th2, etc.
We are now ready to request the CML estimates of the theta parameters of the
Rasch model using the
clogit command:
. * clm-estimator for Rasch model (etas are fixed effects)
. clogit math Th2-Th10, group(subj_id)
note: multiple positive outcomes within groups encountered.
note: 5 groups (50 obs) dropped due to all positive or
all negative outcomes.
Iteration 0: log likelihood = -436.11778
Iteration 1: log likelihood = -435.352
Iteration 2: log likelihood = -435.35069
Iteration 3: log likelihood = -435.35069
Conditional (fixed-effects) logistic regression Number of obs = 1150
LR chi2(9) = 243.80
Prob > chi2 = 0.0000
Log likelihood = -435.35069 Pseudo R2 = 0.2188
------------------------------------------------------------------------------
math | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Th2 | .241266 .3481052 0.69 0.488 -.4410077 .9235397
Th3 | .7921615 .3318989 2.39 0.017 .1416515 1.442672
Th4 | 1.450772 .3241392 4.48 0.000 .8154703 2.086073
Th5 | 1.528206 .3239802 4.72 0.000 .8932166 2.163195
Th6 | 1.913438 .3253857 5.88 0.000 1.275694 2.551183
Th7 | 2.030632 .3265377 6.22 0.000 1.390629 2.670634
Th8 | 2.662249 .3389274 7.85 0.000 1.997964 3.326535
Th9 | 2.904665 .3467128 8.38 0.000 2.225121 3.58421
Th10 | 3.55384 .3771324 9.42 0.000 2.814674 4.293006
------------------------------------------------------------------------------
We conclude that the first item (Th1) with associated Theta(1)=0 is
the easiest item, to which the most subjects responded correctly. Item 10 is
the most difficult item, to which the fewest subjects responded correctly.
clogit has dropped five subjects from the analysis. These subjects
responded either to all items correctly or to all items incorrectly; in a
conditional likelihood, these subjects carry no information about the
difficulty of the items.
Andersen has suggested a specification test, namely that the difficulty
parameters are the same for the “good” subjects and the
“poor” subjects, distinguished via their total score. While
Andersen subsequently employed a (conditional) likelihood-ratio test, we
will proceed by illustrating the new
hausman command. The
hausman test can be used to compare the theta estimates obtained from
the full sample with the theta estimates obtained from the
“good” or “bad” students.
. estimates store RASCH
. clogit math Th2-Th10 if score<=5, group(subj_id)
(output omitted)
. estimates store LESS
. hausman LESS RASCH
---- Coefficients ----
| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| LESS RASCH Difference S.E.
-------------+----------------------------------------------------------------
Th2 | .1507434 .241266 -.0905226 .172672
Th3 | 1.024031 .7921615 .2318693 .1606061
Th4 | 1.358351 1.450772 -.0924202 .1760475
Th5 | 1.414312 1.528206 -.1138945 .1773457
Th6 | 1.822396 1.913438 -.0910427 .1907797
Th7 | 2.014378 2.030632 -.016254 .2016512
Th8 | 3.414959 2.662249 .7527093 .3793093
Th9 | 3.102924 2.904665 .1982584 .3122054
Th10 | 3.414959 3.55384 -.1388815 .341348
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from clogit
B = inconsistent under Ha, efficient under Ho; obtained from clogit
Test: Ho: difference in coefficients not systematic
chi2(9) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 11.72
Prob>chi2 = 0.2293
We conclude that the difficulty of items does not appear to be different for
“poor” subjects. Similarly,
. clogit math Th2-Th10 if score>=5, group(subj_id)
(output omitted)
. estimates store MORE
. hausman MORE RASCH
---- Coefficients ----
| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| MORE RASCH Difference S.E.
-------------+----------------------------------------------------------------
Th2 | .146233 .241266 -.0950331 .4148311
Th3 | .5045323 .7921615 -.2876292 .3871818
Th4 | 1.307562 1.450772 -.1432097 .3401682
Th5 | 1.241154 1.528206 -.2870522 .343086
Th6 | 1.907459 1.913438 -.0059791 .3258093
Th7 | 1.73504 2.030632 -.2955911 .3263267
Th8 | 2.250694 2.662249 -.4115551 .3131327
Th9 | 2.610758 2.904665 -.2939078 .3127668
Th10 | 3.488208 3.55384 -.0656323 .3271359
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from clogit
B = inconsistent under Ha, efficient under Ho; obtained from clogit
Test: Ho: difference in coefficients not systematic
chi2(9) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 8.88
Prob>chi2 = 0.4483
and again we find no significant differences in the item difficulties.
Finally, we can also obtain the estimates for the thetas if we assume that
the etas are random effects.
. * xtlogit estimator for Rasch (etas are Gaussian random effects)
. xtset subj_id
. xtlogit math Th1-Th10, noconstant re
(iteration log omitted)
Random-effects logistic regression Number of obs = 1200
Group variable (i): subj_id Number of groups = 120
Random effects u_i ~ Gaussian Obs per group: min = 10
avg = 10.0
max = 10
Wald chi2(10) = 183.95
Log likelihood = -695.52405 Prob > chi2 = 0.0000
------------------------------------------------------------------------------
math | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Th1 | -1.669049 .2640865 -6.32 0.000 -2.186649 -1.151449
Th2 | -1.439218 .2510217 -5.73 0.000 -1.931211 -.9472242
Th3 | -.9006992 .2289021 -3.93 0.000 -1.349339 -.4520594
Th4 | -.2355981 .2161706 -1.09 0.276 -.6592846 .1880884
Th5 | -.1566347 .2156563 -0.73 0.468 -.5793133 .2660439
Th6 | .2365609 .2161699 1.09 0.274 -.1871243 .6602462
Th7 | .3559474 .2173414 1.64 0.101 -.0700339 .7819286
Th8 | .9920522 .2318226 4.28 0.000 .5376883 1.446416
Th9 | 1.231196 .2410622 5.11 0.000 .7587231 1.703669
Th10 | 1.860508 .2766864 6.72 0.000 1.318212 2.402803
-------------+----------------------------------------------------------------
/lnsig2u | -.1718778 .2528068 -.6673701 .3236145
-------------+----------------------------------------------------------------
sigma_u | .9176503 .1159941 .7162793 1.175634
rho | .2037977 .0410215 .1349109 .2958304
------------------------------------------------------------------------------
Likelihood-ratio test of rho=0: chibar2(01) = 55.65 Prob >= chibar2 = 0.000
Here the order of difficulties of the items is not affected by moving from a
fixed-effects to a random-effects estimator for the thetas.
A command for Rasch
It is important to carry out specification tests of the model to analyze
whether the model fits the data. In the literature, such tests have
been suggested and are indeed commonly applied by psychometricians. Some of
these tests, as proposed by Andersen, are (conditional) likelihood-ratio
tests for the hypotheses that the thetas are the same for groups of subjects
formed by an exogenous variable (Do men and women behave the same way?) and
by distinguishing groups via the total score: are the difficulty parameters
the same for the “smart” subjects (those for whom y(i+) is
relatively high) as those for the “dumb” subjects (those with a
low total score for y(i+))? CLM’s condition is on y(i+), so it is OK
to distinguish cases using y(i+). Above we have applied Hausman tests for
this purpose.
Another purpose of a Rasch analysis is to estimate the subject parameter
eta. In the fixed effects approach, the etas are commonly estimated by
maximum likelihood conditional on the CLM theta-estimates. For the
random-effects case, the etas are commonly estimated by posterior means.
Reference
- Fischer, H. G. and I. W. Molenaar. 1995.
- Rasch Models. Foundations, Recent
Developments and Applications. New York: Springer.
|
