Raschpackage eRM

Models
and the R
Reinhold Hatzinger
Institute for Statistics and Mathematics
WU Vienna

Munich 2010

Intro

What is eRm?
eRm short for extended
is an R package

Rasch modelling

is open source: no license fees, source code available,

GPL: share, change, and redistribute under certain conditions
for Rasch family models:
utilities for tting, testing, and displaying results
currently implemented models:
LPCM, PCM, LRSM, RSM, LLTM, RM, (LLRA)
uses CML estimation

Munich 2010

Intro

What is Item Response Theory (IRT)?

IRT is built around the central idea:
probability of a subject's certain reaction to a stimulus

described as a
a latent trait

can be

characterising the subject's location on

plus one or more parameters characterising the
function

0.6
0.4
0.0

0.2

Probability

0.8

1.0

stimulus

Latent Dimension

Munich 2010

The Rasch Model

The Rasch Model (RM) (Rasch, 1960)

exp(v i)
P (Xvi = 1 v , i) =
1 + exp(v i)
Xvi . . . person v gives correct
v . . . `ability' of person v
i . . . `diculty' of item i

I1
P1
P2
P3
P4
si

1
1
1
0
3

I2

0
0
1
1
2

Munich 2010

I3

0
1
0
1
2

I4

0
0
0
1
1

rv

1
2
2
3


answer to item

Raw Scores:
i xvi = rv
v xvi = si

The Rasch Model

0.6
0.4

Item 1
Item 2
Item 3

0.0

0.2

Probability to Solve

0.8

1.0

Several ICCs

Latent Dimension

Munich 2010

The Rasch Model

Rasch Model Assumptions / Properties

unidimensionality

P (Xvi = 1 v , i, ) = P (Xvi = 1 v , i)

response probability does not depend on other variables

suciency

f (xvi, ..., xvk v ) = g(rv v )h(xvi, ..., xvk )

score rv = i xvi (sum of responses) contains

all informaraw
tion on ability, regardless which items have been solved

conditional independence
for xed

Xvi Xvj v , i, j

there is no correlation between any two items

monotonicity for
response

Munich 2010

v > w f (xvi v , i) > f (xwi w , i), v , w

probability increases with higher values of

The Rasch Model

Parameter Estimation
Item Parameter Estimation

likelihood based methods:

dier in their treatment of person parameters
joint ML estimation (JML)
conditional ML estimation (CML)
marginal ML estimation(MML)

other methods available:

less often used
not covered here

Person Parameter Estimation

ML and weighted ML estimation
Bayes approaches
Munich 2010

The Rasch Model

Joint Maximum Likelihood (JML)

Lu =

exp(v v rv ) exp( i isi)

v i(1 + exp(v i))

sucient statistics are:

rv = i xvi for v
si = v xvi for i

problem: item parameter estimates inconsistent as

biased in nite samples with k(k 1)

Marginal Maximum Likelihood (MML)

integrate out the person parameter

nr

exp(r)
exp( isi)
Lm =
dG()

r
k (1 + exp( i))
i
i=1

distribution for must be specied, usually N (0, 1)

can be estimated in R using the ltm package (Rizopoulos,2009)
Munich 2010

The Rasch Model

Conditional Maximum Likelihood (CML)

condition on

rv

Lc = exp( isi)/ exp( xii)nr

i







r xr

person parameters do not occur in the conditional likelihood

items can be compared independent of persons (separation)
leads to specic objectivity
person free item calibration
`sample-independence':
actual sample not of relevance for inference on item parameters

CML estimates are unbiased and consistent as n

for estimability set 1 = 0 or i = 0
items with score si = 0 or n and person with rv = 0 or
removed prior to estimation
Munich 2010

are

The Rasch Model

MML vs CML
MML Advantages:
gives also estimates for persons with rv = 0 or
when research aims at person distribution
allows estimation of additional parameters
(2PL, 3PL models)
faster with large k

rv = k

CML Advantages:
when RM is used as measurement model (scale construction)
MML parameters can be biased if G() incorrectly specied
CML closer to concept of person-free assessment
allows for specic objectivity
several goodness-of-t tests not available with MML
distributional properties of CML and MML estimates are the
same asymtotically
Munich 2010

10

The Rasch Model

Person Parameter Estimation

using the unconditional likelihood
Lu =

exp(v v rv ) exp( i isi)

v i(1 + exp(v i))

and assuming the s to be known (from prior estimation)

slightly biased (bias smaller than s.e.'s of estimates)
no estimates for rv = 0 and rv = k
can be approximated using, e.g., spline interpolation
weighted ML estimation:
likelihood function is skewed, additional source of estimation bias
Warm (1989) suggests unbiasing correction, computationally unfeasible
Munich 2010

11

eRm

The R package eRm (extended Rasch modelling)

> library(eRm)

main functions concerning t of the RM:

RM(data) ts the RM and generates object of class dRm
[Link](drmobj) generates object of class ppar
plots from drm object:
 plotPImap(), plotICC(), plotjointICC()
plots from ppar object:
 plot()
extract information from drm object:
 coef(), vcov(), confint(), logLik(), [Link]()
extract information from ppar object:
 confInt(), logLik()

Munich 2010

12

eRm

Fitting the RM
> [Link] <- RM(data)
> [Link]
Results of RM estimation:
Call: RM(X = data)
Conditional log-likelihood: -156.3100
Number of iterations: 12
Number of parameters: 4
Basic Parameters eta:
eta 1
eta 2
eta 3
eta 4
Estimate 0.4292685 -1.1743542 -0.1496732 0.02667262
[Link] 0.1945618 0.2243309 0.1918824 0.19118379

 default is: RM(datamatrix, sum0 = TRUE, other options)

 sum0 denes constraints (for estimability):
TRUE . . . sum zero, FALSE . . . rst item set to 0
 the output gives easiness (not diculty) parameters!
Munich 2010

13

eRm
> summary([Link])
Results of RM estimation:
Call: RM(X = data)
Conditional log-likelihood: -156.3100
Number of iterations: 12
Number of parameters: 4
Basic Parameters (eta) with 0.95 CI:
Estimate Std. Error lower CI upper CI
eta 1
0.429
0.195
0.048
0.811
eta 2
-1.174
0.224
-1.614
-0.735
eta 3
-0.150
0.192
-0.526
0.226
eta 4
0.027
0.191
-0.348
0.401
Item Easiness Parameters (beta) with 0.95 CI:
Estimate Std. Error lower CI upper CI
beta I1
0.868
0.206
0.464
1.272
beta I2
0.429
0.195
0.048
0.811
beta I3
-1.174
0.224
-1.614 -0.735
beta I4
-0.150
0.192
-0.526
0.226
beta I5
0.027
0.191
-0.348
0.401

Munich 2010

14

eRm

Extracting Information
the item parameter estimates
> coef([Link])
eta 1
eta 2
eta 3
eta 4
0.42926853 -1.17435425 -0.14967319 0.02667262

the variance-covariance matrix of item parameter estimates

> vcov([Link])
[,1]
[1,] 0.037854306
[2,] -0.012554175
[3,] -0.008073628
[4,] -0.007959444

Munich 2010

[,2]
-0.01255417
0.05032436
-0.01171606
-0.01178009

[,3]
-0.008073628
-0.011716057
0.036818867
-0.007484464

[,4]
-0.007959444
-0.011780088
-0.007484464
0.036551241

15

eRm

Extracting Information (cont'd)

conndence intervals for the item parameter estimates
> confint([Link], "beta")
2.5 %
97.5 %
beta I1 0.46444284 1.2717297
beta I2 0.04793435 0.8106027
beta I3 -1.61403476 -0.7346737
beta I4 -0.52575584 0.2264095
beta I5 -0.34804072 0.4013860

the conditional log likelihood

> logLik([Link])
'Conditional log Lik.' -156.3100 (df=4)

Munich 2010

16

eRm

Plot ICCs

0.6
0.4

Item 1
Item 2
Item 3
Item 4
Item 5

0.0

0.2

Probability to Solve

0.8

1.0

> plotjointI([Link], xlim = c(-5, 5))

Latent Dimension

Munich 2010

17

eRm

Plot single ICC

0.6
0.4
0.0

0.2

Probability to Solve

0.8

1.0

> plotI([Link], i = 3)

Latent Dimension

Munich 2010

18

eRm

Plot ICCs
> plotICC([Link], [Link] = 1:4, ask = F, empICC = list("raw"),
+
empCI = list(lty = "solid"))

0.8

q
q
q

0.4

0.8
0.4

0.0

Probability to Solve

ICC plot for item I2

q
q

0.0

Probability to Solve

ICC plot for item I1

ICC plot for item I4

0.0

Latent Dimension

0.8
0.4

q
q

q
q

q
q

0.0

0.4

0.8

Munich 2010

Probability to Solve

Latent Dimension

ICC plot for item I3

Probability to Solve

Latent Dimension

Latent Dimension

19

eRm

Plot Person-Item Map

> plotPImap([Link])
PersonItem Map
Person
Parameter
Distribution

I1
q

I1

I2

ttx

I2

I3
q

I3
I4
q

I4
I5
q

I5

Latent Dimension

Munich 2010

20

eRm

Person Parameter Estimation

> pp <- [Link]([Link])
> pp
Person Parameters:
Raw Score
0
1
2
3
4
5

Estimate
-2.6310979
-1.5189967
-0.4615091
0.4374933
1.5217580
2.6659917

[Link]
NA
1.1498599
0.9565426
0.9636302
1.1669396
NA

if NAs in the data, dierent person parameters are estimated for

every NA-pattern group

Munich 2010

21

eRm

Methods for Person Parameter Estimation Results

> logLik(pp)
'Unconditional (joint) log Lik.' -10.85398 (df=4)
> confint(pp)
2.5 %
P1 -3.772681
P2 -1.451187
P3 -1.451187
P5 -2.336298
P6 -1.451187
P7 -2.336298
...

97.5 %
0.7346872
2.3261739
2.3261739
1.4132799
2.3261739
1.4132799

attention: confint(pp) gives values for all subjects

if there are Nas in the data, condence intervals are printed for
each NA group

Munich 2010

22

eRm

Plot of Person Parameter Estimates

> plot(pp)

Plot of the Person Parameters

1
0

Person Parameters (Theta)

Person Raw Scores

Munich 2010

23

Testing the Rasch Model

Testing the RM  Overview

RM allows to evaluate the quality of measurement
crucial assumptions empirically testable
aim: nd set of items that conform to the RM (`data t model')
various tests/diagnostics have been proposed
some implemented in eRm:






Andersen LR test
Wald-type test
nonparametric tests
item/person t indices
graphical procedures

Munich 2010

24

Testing the Rasch Model

Andersen's Likelihood Ratio Test (Andersen, 1973)

 `global' test (all items investigated simultaneously)
 powerful against violations of suciency and monotonicity
 can detect DIF (dierential item functioning or item bias ):
basic idea:
consistent item parameter estimates (`invariance') obtained from
any subgroup where the model holds
divide the sample according to score r,
obtain J 1 likelihoods of the form
(r)

Lc

r = 1, . . . , J 1

= exp( j sj )/(r; 1, . . . , J )nr

(r)

the total likelihood is

Munich 2010

(r)

Lc = Lc
r

25

Testing the Rasch Model

Andersen's Likelihood Ratio Test (cont'd)

then
=

Lc
(r)
r Lc

Z = 2 ln

= 1,

only if the RM holds

is asymptotically

2-distributed

with

df = (J 2)(J 1)

test can be used for any partition of the sample according to

extraneous variables (e.g., gender, age, ...)
Wald Test
allows for testing single items idea is again: sample into subgroups (usually 2)
(1)
(2)
(1)
(2)
using separate estimates j and j (and j , j ),

(1)
(2)
(1) (2)
Sj = (j j )/ + N (0, 1)

Munich 2010

26

Testing the Rasch Model

Nonparametric (`exact') Tests

Idea:
Parameter estimates depend only on marginals r and s
for any statistic of the data matrix, one can approximate the
null distribution
take random sample from the collection of equally likely data
matrices, compute null distribution of statistic
valid and powerful, even in small samples

Person/Item Fit

objective is to detect noticeable patterns

Expected response: vi = exp(v i)/(1 + exp(v i))
Residuals: evi = xvi vi

e2
1
vi
Example: Outt MSQ for items: ui =
n v vi(1 vi)
test statistics, e.g., nu2, are 2 with corresponding df
i

Munich 2010

27

Testing the Rasch Model

Graphical Procedure
underlying idea again subgroup homogeneity, plot

(1)

vs

(2)

I3

I4

I1

I5

I2

Beta for Group: Raw Scores > Median

Graphical Model Check

Beta for Group: Raw Scores <= Median

Munich 2010

28

eRm

Likelihoodratio- and Wald Tests

LR Test:
> lrt <- LRtest([Link], se = TRUE)
> lrt
Andersen LR-test:
LR-value: 2.407
Chi-square df: 4
p-value: 0.661

Wald Test:
> Waldtest([Link])
Wald test on item level (z-values):
beta
beta
beta
beta
beta

I1
I2
I3
I4
I5

z-statistic p-value
-0.832
0.405
-0.352
0.725
0.428
0.668
1.300
0.194
-0.411
0.681

Munich 2010

29

eRm

Item Fit Statistics

> itemfit(pp)
Itemfit Statistics:
Chisq df p-value Outfit MSQ Infit MSQ
I1 80.938 84
0.574
0.952
0.966
I2 78.491 84
0.649
0.923
0.934
I3 82.480 84
0.526
0.970
0.961
I4 85.144 84
0.445
1.002
1.024
I5 74.275 84
0.767
0.874
0.908

Nonparametric Tests
> t11 <- NPtest(data, method = "T11")
> t11
Nonparametric RM model test: T11 (global test - local dependence)
(sum of deviations between observed and expected inter-item correlations)
Number of sampled matrices: 500
one-sided p-value: 0.934

Munich 2010

30

eRm

Graphical Procedure
> plotGOF(lrt, conf = list())

I3

q
q

I4
q I5
I2

q
q

q
q

I1

q
q

Beta for Group: Raw Scores > Median

Graphical Model Check

Beta for Group: Raw Scores <= Median

Munich 2010

31

eRm

Polytomous Models
Partial Credit Modell (PCM)
exp[h(v + i) + hi]
P (Xvi = h) = mi
l=0 exp[l(v + i) + li]
h . . . response categories (h = 0, . . . , mi)
mi . . . number of response categories may
hi . . . category parameter

dier across items

Rating Scale Modell (RSM)

simplication:
mi = m . . . distances between categories are equal across all item
hi = h . . . `equistant scoring'

Munich 2010

32

eRm

ICCs for the PCM

0.8
0.6
0.2

0.4

Category 0
Category 1
Category 2
Category 3

0.0

probability for responding in category

1.0

ICC plot for item I2

Latent Dimension

Munich 2010

33

eRm

Comparison RSM vs PCM

RSM

PCM

Person
Parameter
Distribution

Person
Parameter
Distribution

QA20_1

q
q

ttx

q
q

q
q

QA20_3

q
q

q
q

q
q

Latent Dimension

Munich 2010

q q

q
q

qq qq

QA20_5

QA20_4

QA20_6

qq

2 34

qq

QA20_6

QA20_5

2 3

QA20_2

QA20_4

q q
q

ttx

QA20_3

QA20_2

QA20_1

Latent Dimension

34

eRm

R commands
main functions concerning t of polytomous models:
PCM(data) ts the PCM and generates object of class Rm
RSM(data) ts the RSM and generates object of class Rm
thresholds(rmobj) displays the itemparameter estimates as

thresholds
all other functions are the same as previously presented
(except for plotjointICC())

Munich 2010

35

eRm Summary

eRm Summary
core of eRm is the Linear Partial Credit Model (LPCM):
P (Xvi = h) =

where the

ih's

exp(hv + ih)
mi
l=0 exp(lv + ih)

are linearly reparameterised

mi

ih = wihpp
p

allows for a general algorithm:

according to the specication of the design matrix W = ((wih,p))
various models can be estimated
currently functions for:
LPCM, PCM, LRSM, RSM, LLTM, RM
Munich 2010

36

eRm Summary

The model hierarchy in eRm

the LPCM is the most general unidimensional model in this family
all other models are submodels
they are obtained by appropriately dening the design matrix W
LPCM
PCM
LRSM
RSM
LLTM
RM

Munich 2010

37

eRm Summary

eRm Features
Scope:
Scale Analysis (measurement models)
Modelling latent change (statistical models)
uni- and multidimensional (LLRA)
Models:
RM, RSM, PCM, LLTM, LRSM, LPCM, (LLRA)
Treatment of missing values (MCAR)
Dierent constraints for parameter estimation
Design matrix (default / user dened)
Estimation:
Itemparameters, `basic'- and eect parameters,
threshold parameters (all using CML)
Personparameters (JML)
Covariance matrices (condence intervals)
Support for stepwise item selection
Munich 2010

38

eRm Summary

eRm Features (cont'd)

Diagnostics, Model Tests, and Fit Statistics:
Andersen LR-test, Wald Test for single items
Global and item level nonparametric tests (for RM)
Itemt, Persont (using Pearson residuals)
Information criteria (AIC, BIC, cAIC)
Check for existence of ML estimates 
`well-conditioned datamatrix' (for RM)
some (nonpsychometric) logistic regression diagnostics
Plots:
Goodness-of-Fit Plots
ICC-Plots for single items (with optional empirical ICCs)
Joint ICC-Plot (for RM)
Person-Item Map
Miscellaneous :
Simulation of data matrices according to RM violations
...
Munich 2010

39

eRm Summary

Further Infos:
R Forge:

[Link]

Development platform
latest releases downloads
Discusssion and help forum
Project homepage [Link]

Publications:

Mair & Hatzinger (2007). Journal Statistical Software

Mair & and Hatzinger (2007). Psychology Science
Hatzinger & Rusch (2009). Psychology Science Quarterley

Munich 2010

40