Volumen 27 No 1. P´
ags. 27 a 41. Junio 2004

Modeling Abilities in 3-IRT Models
Edilberto Cepeda Cuervo*
´ Manuel Pela
Jose

Abstract
This paper considers situations where regression models are proposed
to model abilities on three parameter logistic models. After a summary of
the classical approach to estimate the item parameters and the abilities,
we provide an exposition of the maximum likelihood method to estimate
the regression parameters. We analyze how good these estimations are,
using a simulated study, and we include an application.
Key words: Logistic models, Ability modeling, Item Response Theory

1.

Introduction

In Item Response Theory (IRT) we have an estimation problem including
two types of parameters: the item parameters and the subject’s ability parameters. The difficulty in the general model is that the subject’s ability, which
appears as a nuisance parameter, cannot be eliminated from the likelihood
function by conditioning a sufficient statistic in the way proposed for the one
parameter logistic model by Rasch (see Andersen, 1980, chap 6). Moreover,
joint maximum likelihood estimation of the subject’s abilities and item parameters is not generally possible because the number of parameters increases
with the number of subjects and then, standard limit theorems do not apply
(Bock & Aitken, 1981).
A better approach to estimation in the presence of a random nuisance parameter is to integrate over the parameter distribution and to estimate the
* E-mail:
** E-mail:

ecepeda@unal.edu.co. Departamento de Matem´
jos-pela@uniandes.edu.co. Departamento de Matem´

Andes.

27

Instead. 2. ξi ) = ci + (1 − ci ) 1 1+ e−Dai (θj −bi ) . is a vector of ability explanatory variables.. 2. . . Verhelst & Eggen (1989) and Zwinderman (1991) formulated a structural model for the Rasch model assuming that the item parameters are known. Section 2 presents some topics on three parameter logistic models. The probability that a subject j with ability parameter θj solves item i with difficulty parameter bi is given by: p(uij = 1|θj .28 Edilberto Cepeda C. 1) distribution of the ability. j = 1. Section 7 presents an application and section 8 draws concluding remarks. by a simple reformulation of the Bock– Lieberman likelihood equation. & Jos´e Pel´ aez A.. xRj ). .. i = 1.. . This reformulation of the likelihood equation clearly states that the form of the ability distribution does not need to be known.. (1) . be I × n binary random variables. we can model the abilities using θj = xtj β + ²j . Three Parameter Logistic Models Let uij . Section 6 includes a simulated study....e. n.. a histogram). thus freeing the method from arbitrary assumptions about the ability distribution in the population effectively sampled. gender. where i indicates an item and j a subject. Section 4 reviews the maximum likelihood estimate of the abilities. small and large number of items. among others. a computational solution is possible to find for both. uij = 1 if subject j solves item i.... where β t = (β0 . n. number of hours of practice. Bock & Aitken (1981) show that. . In this paper we model the abilities as a regression model on a 3-IRT model. otherwise uij = 0... Section 3 reviews the marginal maximum likelihood estimate of item parameters. . This is not a new idea. If xtj = (1. Subject’s abilities can be explained by associated factors such as habits. socioeconomic status and parents’ education level.. I and j = 1.. j = 1.. i = 1. The item parameters can be estimated by integrating over the empirical distribution. .. structural parameters by maximum likelihood in the marginal distribution. I. Working with two parameter normal ogive model. β1 . n. σ 2 ). Bock & Lieberman (1970) have already taken this approach and have estimated item stress holds and factor loadings based on the assumption that subjects are random samples from a N (0... xij . it can be estimated as a discrete distribution on a finite number of points (i.. βR ) is a regression parameters vector and ²j ∼ N (0. Section 5 presents the ability regression models and the ML method for the parameter estimation.

If for items i and j. then there is no possibility of having random correct answers and the resulting model is called the unidimensional two parameter logistic model or Birnbaum’s model. f 0 (bi ) = 21 (1 − ci )Dai . the three parameters are independent from the sample location and dispersion of the ability. equation (1) reduces to: p(uij = 1|θj . ci = cj and ai < aj . θj and bi assume values between −∞ and ∞. 3. Given an item with known parameters. As the ability scale does not have practical interpretation in pedagogical terms. higher values of bi correability of solving the item equals 2 spond to higher difficulty levels for the item being considered. Consequently. bi . also called the Rasch Model (1961). 2 then bi may be interpreted as the subject’s ability necessary for the prob1 + ci . If θj = bi . Viewing p(uij = 1|θj . The Rasch model describes how the probability of correct answers depends on the subject’s overall ability and on the level of difficulty of the questions. One of the most important theoretical merits of the Rasch models is called by Fisher (1995) the “specific objectivity”. This fact justifies to take ai as a discrimination measure. If in equation (1) ci = 0. consisting in that item parameters do not depend on the characteristics of the subjects answering the test. we obtain the unidimensional one parameter logistic model. ξi ). ξi ) as a function f of θj . ξi ) = 1 + ci . thus. and subject’s parameters do not depend on the items chosen from a given set. then the item with discrimination measure aj discriminates more than the item with discrimination measure ai . 1968). and θj is the ability of j-th subject.29 Modeling Abilities in 3-IRT Models where ξi = (ai . ci ) are the parameters of item i. ci can be interpreted as the probability θj →−∞ of random correct answers. The following considerations provide a basis for the interpretation of the parameters involved in the IRT model: 1. Setting ai = a and c = 0 in equation (1). 0 ≤ ci < 1 and ai > 0 (Birnbaun. The latent continuous θ is called the ability. the curve which describes the relation between subject’s ability parameter and agreement probability is called Item Characteristic Curve (ICC) (See figure 1). it is necessary to define an ability scale characterized by sets of items . 2. Since ci = lim p(ui = 1|θj .

. < θP . θp . a = 1.0 0..1.8 1.2. θ0 < θ1 < .6 0.65.30 (see Andrade. a = 2 and b = 0.4 0. -2 -1 0 1 2 ability Figure 1: Item Characteristic Curve: Probability of agreement with (a) c = 0.2 Probability 0. depend of the conditional probabilities of correct answer p(u = 1|θ = θp ) and p(u = 1|θ = θp−1 ). Pertinent anchor levels. and p(u = 1|θ = θp ) − p(u = 1|θ = θp−1 ) ≥ 0.30 0. selected by the analyst. & Jos´e Pel´ aez A.5. Thus. the scale is defined only at the end of the statistical analysis of the data resulting from the test application. The ability scale is defined by a set of values. Specifically.. 2000).59 (dashed line). an item u is one anchor item of a level θp if and only if p(u = 1|θ = θp ) ≥ 0. (b) c = 0. p(u = 1|θ = θp−1 ) < 0. called “anchor levels”.0 Edilberto Cepeda C.5 and b = −0.5 (full line). with pedagogical interpretation in the test theoretical frame. The pedagogical interpretation of the scale is possible through the pedagogical interpretation of the set of items associated with each level.

To gain computational advantage it is convenient to work with different response patterns (Bock & Lieberman. ξn ) and ξi = (ai . .. i = 1.. ξi3 ). ci ) = (ξi1 . then the likelihood function is: L(ξ. l=1 and s X ∂` γl ∂ = p(˜ ul |ξ.. (2) R where ξ = (ξi . based on Andrade (2000). given their ability. I. 2. Different items are solved independently by each subject. Answers of different subjects are independent from each other. ξ)g(θ|η)dθ. . Z p(˜ ul |ξ. . η).31 Modeling Abilities in 3-IRT Models 3. the log likelihood function is: ³ n! ` = log Qs l=1 ´ γl ! + s X h i γl log p(˜ ul |ξ. 3. if γl . l=1 Thus. Assuming the ability having a known distribution g(θ|η).. if u ˜l is a response pattern. Item parameter estimation In this section we present the marginal maximum likelihood method to estimate item parameters. η) .. η)γl . r = 1. ξi2 . is the number of occurrences of response pattern u ˜l . where s is the number of different response patterns with γl > 0. To this end. s.. η) = Qs s Y n! l=1 γl ! p(˜ ul |ξ.. ∂ξir p(˜ ul |ξ. bi . η) ∂ξir l=1 (3) . Given the independence between different subject’s answers (hypothesis 1).. 1970). 2.. we consider two hypothesis: 1. l = 1. η) = p(˜ ul |θ.

p(˜ ul |ξ. I. η. . equation (3) can be written as: Z h s X ∂` ∂pi Wi i ∗ g (θ)dθ. s.. η) ∂ξir j=1 = I Z ∂ Y p(ujl |θ. we have: ∂p(˜ ul |ξ.. ξ)g(θ|η)dθ. .32 Edilberto Cepeda C. (4) = pi qi ∂ξir R Given (4). where pi = ci + (1 − ci ) . Wi = gl∗ (θ) = g(θ|ul . ξj )g(θ|η)dθ ∂ξir j=1 R = Z hY I R p(ujl |θ. ξj )g(θ|η) dθ. 2. ξj ) j6=i i£ ∂ i p(uil |θ. From equation (5) we obtain: Z h s i X ∂` = D(1 − ci ) γl (uil − pi )(θ − bi )Wi gl∗ (θ)dθ. η) The distribution gl∗ (θ) is the conditional distribution of θl given η. (5) .. η) ∂ξir = I ∂ Y p(ujl |ξj .. qi∗ = 1 − p∗i . item j is answered. η) (−1)uil +1 ih ∂ = p(˜ u |θ. ∂ξir ∂ξir ∂ξir 1 l = 1. ∂ai R l=1 s Z h l=1 R X ∂` = −Dai (1 − ci ) γl ∂bi i (uil − pi )Wi gl∗ (θ)dθ.. ξi ) = pi qi . ξ)g(θ|η) dθ l u 1−u il il ∂ξir ∂ξir R p i qi Z h³ uil − pi ´³ ∂pi ´i p(˜ ul |θ. & Jos´e Pel´ aez A. we have: 1 + e−Dai (θ−bi ) Z h i ∂p(˜ ul |ξ. ∂ξir where ujl = 1 if in the l-th pattern of response. i = 1. = γl (uil − pi ) ∂ξir ∂ξir p∗i qi∗ l R l=1 where p∗i = {1 + e−Dai (θ−bi ) }−1 . ξ) = p∗i qi∗ and pi qi p(˜ ul |θ. ∂ ∂ h uil 1−uil i ∂pi = (−1)uil +1 Since p(uil |θ. ξ)g(θ|η) . By the independence between responses.

∂ξi ∂ξi Given the independence between items. item parameters can be estimated individ∂ 2 log ` ually. ∂bi ∂ci ∂` = 0. η)) ∂ξi l=1 ( s X ∂ 2 p(˜ ul |ξ. since from the hypothesis. r. the computational process is simple. where qˆ(k) = q(ξˆ(k) ).. µ ¶t · ¸t ∂ ∂ log L(ξ. k = 1. q7 ) with qi = (k) (k) H (k) = diag(Hi ) and Hi (7) ∂ log L(ξ. 3. In this case. η). η) l=1 µ ¶µ ¶t ) ∂p(˜ ul |ξ.I. for i 6= j. = 0. η)/ξi ∂p(˜ ul |ξ. ∂ξi ˆ (k) ).. . pi R Then the equations for the estimation of item parameters are: ∂` ∂` = 0. we need the second derivative of `(ξ. 1981). η) . = 0. i=1. η) 0 . There will be groups formed by independent items and groups with non independent items.I and j = i. this can induce a linked structure in the items set... q = (q1 . ∂ai In order to apply the Newton-Raphson algorithm or Fisher scoring algorithm. when a theoretical structure for test elaboration exists... However. q2 . (6) p(˜ uj |ξ. 2. For ∂ξir ∂ξjk i = 1. η) p(˜ ul |ξ.. 2. η) ∂2` ∂` = = ∂ξi ∂ξit ∂ξi ∂ξi ∂ξi ∂ξi ·X ¸t s ∂ 1 p(˜ ul |ξ. η)/∂ξi − . Given the independence between items (Bock & Aitkin. η) = γj ∂ξi p(˜ ul |ξ. where =H i Hi = ∂ 2 `(ξ.33 Modeling Abilities in 3-IRT Models s X ∂` γl = ∂ci l=1 Z h Wi i (uil − pi ) ∗ gl∗ (θ)dθ. η)/∂ξi ∂ξit = γj p(˜ ul |ξ. η) Thus the Newton-Raphson algorithm equation is: ξˆ(k+1) = ξ (k) − (H (k) )−1 q (k) . . the ....

pij qij © ª−1 t ∗ p∗ij = 1 + e−Dai (xj β−bi ) and qij = 1 − p∗ij . & Jos´e Pel´ aez A. j=1 i=1 where pij = p(uij = 1|θj .34 Edilberto Cepeda C.. . Ability estimation Now we know the item parameters. ∂`(θ) ∂θj = = ¾· ¸ I ½ X uij − pij ∂pij pij qij ∂θj i=1 D I X ai (1 − ci )(uij − pij )Wij . probability of some response vector is given by the product of the group probabilities instead of the product of probabilities of individual item responses. ξi ) and qij = 1 − pij . 2. 4. = 0 for k 6= j.. i=1 where Wij = ∗ p∗ij qij . . Then for j = 1. n. By hypothesis (1) and (2). the logarithm of likelihood function can be written as: `(θ) = log L(θ) = n X I X © ª uij log(pij ) + (1 − uij ) log(qij ) . ∂`2 (θ) we need to calculate the second derivatives. ∂θj ∂θk ∂ 2 `(θ) ∂θj2 = ¾ I ½³ X uij − pij ´³ ∂ 2 pij ´ ³ uij − pij ´2 ³ ∂pij ´2 − pij qij ∂θj2 pij qij ∂θj i=1 (8) .. since by hypothesis (2). ∂θj2 ∂ 2 `(θ) section (3). In order to apply Newton-Raphson algorithm or Fisher scoring algorithm. For detail and comments about estimation methods see Mark & Raymond(1995) and Adams & Wilson(1992).

ξi2 . xR ).. . j = 1.. ²j ) = ci + (1 − ci ) 1 1+e −Dai (xtj β+²j −bi ) := pij . Assuming that the distribution of the random variable ²j . ξ = (ξ1 . ²j )g(²j |η)d²j ∂β Z r R i £ ∂ p(uij |β. βR ) is a parameter vector and ² is a random variable associated with error.. η) = p(uij = 1|β... (10) R t where β = (β0 . model (1) can be written as: p(uij = 1|β. (9) for i = 1. ξi .. Modeling the abilities Subject’s abilities may be explained by associated factors such as hours of practice. . ξi . j=1 i=1 Given that ∂P (uij |β. ..... ²j )g(²j |η) d²j . bi ... ∂ 2 pij ∂θj2 = ∗ 2 D2 a2i (1 − ci )(1 − 2p∗ij )(p∗ij qij ) . x1 .35 Modeling Abilities in 3-IRT Models where ∂pij ∂θj = ∗ Dai (1 − ci )p∗ij qij . In this model xt = (1. β1 .. . the logarithm of the likelihood function is: `= n X I n o X uij log(Pij ) + (1 − uij ) log(Qij ) . Given the independence between answers from different subjects (hypothesis 1) and that different items are solved independently by each subject (hypothesis 2). 5. socioeconomic status or education level. ξi . is known.. . ξi3 )t and g(²j |η) = N (0. is a explanatory variables vector. β1 . I. R ∂βr . ξi . η) ∂βr = = Z ∂ p(uij |β. ²j )g(²j |η)d²j := Pij . ξn )t . As in (7) we can use the Newton-Raphson algorithm to estimate subject’s abilities. with the model θ = xt β + ². ci )t = (ξi1 .. Thus. σ 2 ). β = (β0 . ξi . n. ξi . Z P (uij = 1|β... g(²j |η). .. ξi = (ai . βR ) ..

. ∂βr R j=1 i=1 where Wij = j = 1.. ∂` = 0... ∂βi ∂βr We may also consider models with two levels of explanatory variables. for r = 1. 2. applying the Fisher scoring algorithm. R. ∂`(θ) q = (q1 . school variables and student’s variables. ∂βs Thus. ²j ) as in (9). R.. q2 . with hij = −E . Then in the k-th school . ∗ ¡ ¢−1 p∗ij ∗ qij t ∗ . η). the Fisher scoring algorithm is given by: £ ¤−1 βˆk+1 = βˆk + I(β k ) qˆk .36 Edilberto Cepeda C. . requires the second derivative matrix. ξi . the estimates are obtained by solving the equations 1. The second informa∂βi h ∂ 2 `(θ) i tion matrix is I k = (hij ). Pij Qij ∂βs ∂βr (12) ∂ 2 pij ∗ = (1 − ci )D2 a2i xjs xjr (1 − 2p∗ij )p∗ij ∗ qij ∂βr βs R ∂Pij ∗ = (1 − ci )Dai xjr R p∗ij qij g(²j |η).. . Pij ∗ Qij Finally. ξi .. qr ) with qi = .. r = ∂βr The marginal maximum likelihood estimation of regression parameters. p∗ij = 1 + e−Dai (xi β+²j −bi ) and qij = 1 − pij . η) as in (10) and Qij = 1 − Pij = p(uij = 0|β.. Pij Qij ∂βr j=1 i=1 R (11) where pij = p(uij = 1|β.. So equation (11) can be written as: Z n X I X ∂`(θ) =D (1 − ci )ai xjr (uij − Pij )Wij g(²j |η)d²j . 2. R. .. given by: ∂ 2 `(θ) ∂βr βs where and J X I Z ½³ X uij − Pij ´³ ∂ 2 pij ´ = − Pij Qij ∂βs βr j=1 i=1 R ³ u − P ´2 ³ ∂P ´³ ∂p ´¾ ij ij ij ij g(²j |η)d²j . Pij = p(uij = 1|β. & Jos´e Pel´ aez A. For example. ξi . . we obtain: ∂` ∂βr = ¸ n X I Z ·³ X uij − Pij ´³ ∂pij ´ g(²j |η)d²j . ¡ ¢ where qˆk = qˆ β (k) .. η.

1749 × 10 2.220 × 10 −2 −5.02x2i −bi ) for i = 1. pij ).292 × 10−1 −4. For the variables X1 .. Deviation 4. X2 and X3 were simulated n=100 (500) values: x1i = 1 to obtain an intercept model.817 × 10−2 5.9945 × 10−3 Estimates 9. 6..1. 1. n. given by β0k = α0 + α1 zk and β1k = γ0 + γ1 zk for k = 1. If the parameters β0k and β1k are not the same between schools. Deviation 2.. (13) 1 + e−Dai (1−0. we considered twenty five items.5479 × 10−2 Estimates 4. better estimate values will be .038 × 10−3 The simulation shows that using more data. ξi ) = ci + (1 − ci ) .. x3i from an uniform distribution on the interval (−10. The estimates of regression parameters (with standard deviations) are given in table 1. 25 and j = 1.177 × 10−1 2.487 × 10−2 6. . Simulated study A simulated study was conducted to examine how the estimates are similar to the original parameters when we model the ability in the IRT models as a function of some explanatory variables (Model 13).37 Modeling Abilities in 3-IRT Models we can model the abilities as θj = β0k + β1k xj + ²j . So. 10)..441 × 10−2 S.837 × 10−1 1. I 25 25 10 Table 1: Ability model β0 β1 n 100 500 500 −1 β2 −1 9..926 × 10−2 S.6.. where xj is the ability explanatory variable corresponding to the j-th student...592 × 10−3 Estimates 9.955 × 10 S.357 × 10−2 2. it can be modeled as a function of school variables zk .5).5x1j +0. and on 0.980 × 10−2 2. The ci values were simulated from a discrete uniform distribution with mass value 0. θj = α0 + α1 zk + (γ0 + γ1 zk )xj + ²j .5) and U (−1. The values yi of interest variables Y were generated from a Bernoulli distribution with parameters (1. .955 × 10 −2 4.2.5 on 0. respectively. where pij is given by the deterministic model: 1 p(uij = 1|β. K. 1.4 and P (X2 = 1) = 0. Deviation 5. Initially.5. x2i from a discrete distribution with P (X2 = 0) = 0. . with a0i s and b0i s values simulated from uniform distributions U (1.819 × 10−1 −4.

0. 0.21.0.28. 0. 1. 0. 0.. specifications of communicative phenomenon. but with sets with fewer data.18..889. 0.42. 0. 1.96.-0. with one hundred items.74.18. candidates applying for spanish teachers are evaluated by the Ministry of Education.83. 0. 1=female) and X2 =age. From the marginal likelihood function we found the item parameters. We selected twenty one items with discrimination parameters: ai = 0.18. 0.68.12. 0.55.98. The academic and pedagogical abilities in spanish teaching are measured in the effectiveness of approaching capacity to language analysis. i=1. 7.179. obtained. As predictor variables of academic abilities we used X1 =gender (0=male.8. (β. 0.18.18. 0. -0.204.12.40. 0. 2. -0.54.549. 0.93. we need larger samples to obtain better estimates.978.2. The conclusions are the same for non deterministic models.29. language. universal and local literature.59. communicative and didactic abilities. as well as about the student’s capacity of developing strategies to teach these concepts.833.47. 0.09.67. we determine the likelihood value L(β. precision decreases quickly. We can see better estimates of parameters and smaller standard deviations in the study with I = 25 and n = 500. a sample size of thirty or fifty students may be very small. 0.16. with twenty five items. Application In the universities at El Salvador.32.25. ˆ 2 ) can be estimated using different algorithms. 0. ˆ σ In this application σ 2 is unknown.18. 0. for each value of σ 2 the regression model parameters are estimated as in secˆ σ 2 ). 0.17. 0. 0. 0. 0. For example. Finally. we compare tion 4. 1. 0.2.905..917. and probability of random correct answers: ci =0. With an evaluation of academical and pedagogical abilities of the candidates. 0.19.17. -0. 0.18.. as expected from the increasing likelihood information. 0.61. 1.14. 0.64. closer to true values and with small variance. i the likelihood values. 0.06.44.26. 0. Next.33. 0.24. includes the following topics: literature. to get the estimates as the values corresponding to the . 0. where θj = µj + ²j . First.0. 0.19. between 0 and 1. In general. -0. 0. and pedagogical techniques which guarantee an efficient practice of the teacher’s job in these topics in the first levels of the educational process.24.19.77. & Jos´e Pel´ aez A. 0.28.38 Edilberto Cepeda C.33. the Ministry intends to improve the education quality. 0.11.16.18. 0.526. -1. 1.16. The test to explore academic abilities included topics knowledge and disciplinary theories. -0. The test.025. 1. 0.83. 0. 0. 0.052. -1. difficulty parameters: bi =0.0. It was answered by 230 students. (some of them will be presented in a future paper) but here we estimate the regression parameters β for fixed values of σ 2 . 1.

367 × 10−3 ). An interpretation is that female students have more developed language abilities than male students.39 Modeling Abilities in 3-IRT Models -1190 -1195 -1205 -1200 Loglikelihood -1185 -1180 -1175 maximum of the likelihood function.083 × 10−2 (4. have contribution in the statistical explanation of differences in the abilities of university students.2 0.4 0. In general.35 the maximum likelihood estimates (and standard deviation) are βˆ0 = 0. close to true . βˆ1 = −2. Figure 2 is a plot of log likelihood versus variance.8 1. and βˆ2 = 0. it is necessary to consider large samples in order to obtain good estimates. although latent regression analysis with subject-level predictors would eliminate such problems.0 Variance Figure 2: Log likelihood function of variance. 8.740(0.047). while older students have less developed language abilities than younger students.526(0. For σ 2 = 0. It shows an increasing behavior before 0. 1991). From these estimates we can see that gender and age 0.35 (±0.01) and decreasing behavior afterwards. Conclusions The subject’s parameters estimates may be subject to considerable error and consequently these should not be considered as dependent variables in regression models (Zwindeman.6 0.092).

San Francisco. in ‘Statistical Theories Of Mental Tests Score’. Psychometrika 35. Samples of size fifty or less should be considered. Discrete statistical models with social science applications. D. & Wilson. AddisonWesley. pp. too small to obtain acceptable estimates. H. A random coefficients multinomial logit: Generalizing Rasch models. in the most of the cases. ‘Marginal maximun likelihood estimation of item parameters: An application of a EM algorithm’. (1968). M. References Adams. B. Technical report. Other extensions are possible. (1995). Anddersen.. 449–487. Latent regression can be used to determine the “mean ability” of groups of students. Some latent trait models and their use in inferring an examinee’s ability. values. Tavares. Acknowledgements This research was supported by research grants from Facultad de Ciencias de la Universidad de los Andes. Paper presented at the annual meeting of the American Educational Research Association. Andrade.40 Edilberto Cepeda C. A. (1992). Reading. Teoria da resposta ao item: conceitos e aplicac˜ oes. E. 179–197. Publishing Company. Psychometrika 60. R. including indicator functions as explanatory variables. (1980). Bock. explore classical and bayesian metodologies to model mean and variance parameters. for example to compare schools. Bock. D. & Cunha. M. A. ‘Fitting a response model for n dichotomously scored items’. 395–479. R. Birnbaun. G. M. R. D. (1981). (1970). ABE–Associac˜ao Brasileira de Estat´ıstica: S˜ao Pablo. Fisher. F. althought a rule establishing the minimum number of items in an test does not exist. (2000). Psychometrika 46. with small variance. H. North-Holand. & Jos´e Pel´ aez A. MA. & Aitkin. 179–197. & Lieberman. . for example. R. ‘Some neglected problems in IRT’.

Psychometrika 56. 271–288. Zwinderman. W.Modeling Abilities in 3-IRT Models 41 Mark. R. J. (1991). Probabilistic Models for Some Intelligence and Attainment Test. A. ‘Rasch models for item bundles’. (1995). ‘A generalized Rasch model for manifest predictors’. & Raymond. (1960). Mislevi. Psychometrika 60(2). 589–600. . J. A. (1983). G. ‘Item response model for grouped data’. Danish Institute for Educational Research. Rasch. Journal of Educational Statistics 8(4). H. 181–198.