Journal of Educational Measurement

Fall 2009, Vol. 46, No. 3, pp. 273–292

Automated Test Assembly for Cognitive Diagnosis

Models Using a Genetic Algorithm
Matthew Finkelman
Tufts University School of Dental Medicine
Wonsuk Kim and Louis A. Roussos
Measured Progress

Much recent psychometric literature has focused on cognitive diagnosis models

(CDMs), a promising class of instruments used to measure the strengths and weak-
nesses of examinees. This article introduces a genetic algorithm to perform auto-
mated test assembly alongside CDMs. The algorithm is flexible in that it can be
applied whether the goal is to minimize the average number of classification errors,
minimize the maximum error rate across all attributes being measured, hit a tar-
get set of error rates, or optimize any other prescribed objective function. Under
multiple simulation conditions, the algorithm compared favorably with a standard
method of automated test assembly, successfully finding solutions that were appro-
priate for each stated goal.

In recent years, much work in educational assessment has focused on providing

detailed diagnostic information to examinees, going beyond the reporting of a single
test score. For example, it might be desired that a mathematics test measure not only
an examinee’s general proficiency in the subject, but also specific abilities like factor-
ing polynomials, solving quadratic equations, using laws of exponents, and manipu-
lating fractions. To meet this need for diagnostic information, psychometricians have
developed new models often referred to as “cognitive diagnosis models” (CDMs) be-
cause they are intended to provide more sophisticated parameterizations that better
handle the complexities of cognitive psychological theories (Nichols, Chipman, &
Brennan, 1995).
In CDMs, the ability of examinee j is characterized not by a single latent trait, θj ,
but by a vector α j = (α1 j , . . . , α K j ) of K values, which are usually referred to as
“attributes” or “skills.” Of these terms, the former is more accurate because CDMs
are applied to measuring a wide variety of examinee latent characteristics, from fine-
grained cognitive skills to knowledge in broad content areas as well as personality
traits (Roussos, Templin, & Henson, 2007). In fact, they are now considered relevant
in any situation that involves obtaining information that is richer and more useful than
a single score (Roussos et al., 2007). The goal of CDMs is to effectively measure and
provide feedback on each attribute; parents, teachers, and the examinees themselves
can then refocus instruction on the basis of this feedback.
Because the quality of a test’s diagnostic information depends directly on the qual-
ity of its items, test assembly is a critical step in developing assessments. Indeed,
many examinations are constructed by first developing a large pool of valid, reli-
able items, and then carefully selecting items from this pool to meet the substantive
and statistical design constraints of the test. The goal of test development in this

Copyright 
c 2009 by the National Council on Measurement in Education 273
Finkelman, Kim, and Roussos

context is the construction of a test of sufficient statistical performance (such as re-

liability, psychometric information, or measurement precision) yet also meeting a
variety of substantive constraints (such as sufficient numbers of items measuring
the skill and knowledge areas of concern, variety of types of items, key balancing
for multiple-choice items, etc.). Within the context of using CDMs, statistical per-
formance constraints are applied not simply to the test as a whole, but also to the
individual attributes, and substantive constraints may also apply within as well as
between attributes.
To streamline the process of selecting items from a pool, many researchers have
studied automated test assembly (ATA), where a computer is used to select an appro-
priate set of items without human intervention (Theunissen, 1985; van der Linden,
2005; van der Linden & Boekkooi-Timminga, 1989). Henson and Douglas (2005)
combined CDMs with ATA by developing a cognitive diagnostic index (CDI) for
item discrimination and proposing the selection of items that maximize the index.
Although this heuristic method was shown to substantially improve test precision
compared to randomly determined tests, it has two potential shortcomings. First, as
pointed out by Henson, Roussos, Douglas, and He (2008), the CDI is an overall mea-
sure of an item’s discrimination; therefore, it does not give psychometric information
at the attribute level. In many applications of test assembly, it is desired that a diag-
nostic test not only provide overall precision, but also accurate measurement of each
attribute. Second, because Henson and Douglas recommended the selection of items
with largest CDI values, the resulting test will always be maximally informative with
respect to that index. This property of an ATA method is appropriate when the practi-
tioner seeks to achieve as accurate a test as possible. However, in many applications
of test assembly, the goal is to attain a specified (rather than maximized) level of pre-
cision; Birnbaum (1968), Lord (1980), and Theunissen (1985) discussed this target
function approach when conducting ATA alongside a unidimensional item response
theory (IRT) model. The application of target functions is common when the anal-
ysis of results includes year-to-year comparisons. In such cases, the current form
is often designed to have psychometric properties similar to those of the previous
year’s form, promoting between-year comparability of results. To date, however, no
CDI-based method exists to select items that meet a target level of overall precision.
Achieving a target level of precision at each attribute would be even more difficult, if
not impossible, for a CDI-based method in light of the fact that it does not measure
discrimination at the attribute level.
In this article, we introduce a genetic algorithm (GA) method that overcomes the
above difficulties. It is an ATA procedure that chooses item sets holistically rather
than item by item, ensuring that the selected items complement one another. The
method takes initial solutions as starting points, then assembles a test that always
performs as well as or better than these starting points. This new method can optimize
item selection either at the attribute level or based on overall precision, in accord with
the practitioner’s specifications. It is equally applicable whether minimizing the error
rates of a test, hitting a set of target error rates, or optimizing any other prescribed
objective function.
After introducing the GA, we turn our focus to comparing this method to the CDI
in simulation. Our research design includes the evaluation of these two procedures in

274
 Automated Test Assembly for CDMs Using a Genetic Algorithm

terms of three fitness functions: average number of classification errors, maximum

error rate, and ability to hit attribute-level target error rates. Multiple prior distribu-
tions, item pools, and levels of constraints are considered. The relative advantages
and disadvantages of the GA and CDI are discussed.

Cognitive Diagnosis Models

CDMs are a class of statistical models that are capable of measuring multiple traits
simultaneously. They are distinguished from multidimensional item response theory
(MIRT) models (Reckase, 1985) in that CDMs classify examinees into a discrete
number of levels (usually two) on each of the K attributes, whereas MIRT models
typically measure the ability vector of examinee j, θ j = (θ 1 j , . . . , θ Kj ), along a con-
tinuous K-dimensional spectrum of values. Hence, item and ability parameters of
CDMs are especially tailored to classification, facilitating the interpretation of re-
sults when the goal is to categorize examinees into discrete levels. Without loss of
generality, this article assumes that two levels (mastery and nonmastery) are speci-
fied for each attribute. An examinee’s ability is defined by his or her α j vector: αkj
= 1 if examinee j has achieved mastery on attribute k, and αkj = 0 otherwise (k =
1, . . . , K).
A fundamental component of CDMs is the so-called Q-matrix (Tatsuoka, 1985),
which indicates the attributes that are measured by each item. If there are I items
(indexed by i), then the Q-matrix is an I × K matrix. A given element qik is 1 if item
i measures attribute k, and 0 otherwise.
Many candidate CDMs have been proposed in the psychometric literature. These
include the restricted latent class model (Haertel, 1984; also called DINA, Junker
& Sijtsma, 2001), NIDA (Junker & Sijtsma, 2001), compensatory MCLCM (Maris,
1999; von Davier, 2005); and the reparameterized unified model (RUM; Roussos
et al., 2007). We only provide a formal definition of the RUM here because this was
the model utilized in simulation. However, we emphasize that both ATA procedures
discussed in this article (CDI and GA) are general; they can be applied alongside any
CDM. Thus, although the RUM was used for illustrative purposes, the comparison
between the CDI and GA could have been made with any other CDM. It is also
noteworthy that CDMs, like all other psychometric models, should always be tested
for misfit when applying them to real data. See de la Torre and Douglas (2004),
Roussos et al. (2007), and Sinharay (2006) for information on assessing the fit of
CDMs.
To begin our review of the RUM, let Yijk be a latent variable representing the event
that attribute k is correctly applied to item i by examinee j. Let rik = P(Yik j =
1 | αk j = 0) and πik = P(Yik j = 1 | αk j = 1), and define rik∗ = rik /πik . In words, rik
is the probability that attribute k is correctly applied to item i by an examinee who
has not mastered this attribute, πik is the corresponding probability for an exami-
nee who has mastered the attribute, and rik∗ is the ratio between the two. Smaller rik∗
values indicate better discernment between masters and nonmasters. Thus, rik∗ essen-
tially acts as a “reverse discrimination” parameter: the smaller it is, the higher the
discrimination. Related to πik is another variable π i∗ , which denotes the probability
of correctly applying all attributes that are required for item i, given that an examinee

275
Finkelman, Kim, and Roussos

has mastered all of these attributes. Note that if the application of attributes is locally
independent, then


K
πi∗ =
q
πikik
k=1

(Roussos et al., 2007). Now let η j denote the “supplemental ability” of examinee
j; this variable represents a composite of residual abilities for attributes α b that are
not included in the Q-matrix, but potentially have an effect on item responses. An
associated quantity is the probability of correctly applying attributes α b to item i,
given η j ; this value is denoted Pci (η j ). Finally, let X i j be an indicator variable equal
to 1 if examinee j answers item i correctly, and 0 otherwise. Under the RUM, the
probability that examinee j answers item i correctly is


K
P(X i j = 1 | α j , η j ) = πi∗ rik∗ (1−α jk )qik Pci (η j ) (1)
k=1

(Roussos et al., 2007). Our simulation study employed the “reduced” RUM, which
is given by Equation 1 with the additional assumption that Pci (η j ) = 1 for all i and
j. Note that parameters are constrained to satisfy 0 < π i∗ < 1 and 0 ≤ r ik
∗
≤ 1. These
conditions place probabilities in the appropriate range and guarantee that they are
monotonic, that is, that a master of attribute k always has at least as high a chance of
applying this attribute correctly as a nonmaster of the attribute.

Review of the CDI and ATA

When a unidimensional IRT model is employed, one common approach to ATA
is to optimize selected items in terms of their combined Fisher information, or test
information function (Birnbaum, 1968; van der Linden, 2005; van der Linden &
Boekkooi-Timminga, 1989). A necessary assumption for computing Fisher informa-
tion is that the log-likelihood function has a second derivative (Lord, 1980). This
assumption does not hold for CDMs, which are formulated using discrete levels of
ability rather than continuous distributions. Therefore, ATA based on Fisher infor-
mation is inapplicable to tests employing a CDM (Henson & Douglas, 2005).
As an alternative to Fisher information, Henson and Douglas (2005) proposed
a discrimination index for CDMs utilizing the concept of Kullback-Leibler infor-
mation (Chang & Ying, 1996). Kullback-Leibler information is easily applicable to
classification problems and does not require differentiability of the log-likelihood
function. It is a measure of the distance between two distributions; here, the distri-
butions in question are those induced by two different attribute vectors (say, α and
α  ). An item that produces a large distance between the distributions induced by α
and α  will in turn provide high discriminatory power for discerning these vectors.
Therefore, the Kullback-Leibler information of item i between α and α  , defined as
 
 L(α; X i )
K i (α, α ) = E α log , (2)
L(α  ; X i )

276
 Automated Test Assembly for CDMs Using a Genetic Algorithm

may be used as an index of the item’s ability to distinguish α from α  . Note that in
Equation 2, the subscript j has been dropped from α, α  , and X for the purpose of
simplicity (X i still denotes an indicator variable for a correct response to item i).
L(α; X i ) is the likelihood of α for the datum X i . Intuitively, L(α; X i ) indicates the
degree to which the result of X i is consistent with the hypothesis that α is the true
state of nature. The expectation is over the distribution of X i given α; thus, for the
dichotomous X i assumed in this article, K i (α, α  ) can be written more explicitly as
   
 Pα (X i = 1) Pα (X i = 0)
K i (α, α ) = Pα (X i = 1)log + Pα (X i = 0)log
Pα (X i = 1) Pα (X i = 0)

(Henson & Douglas, 2005).

One property of K i (α, α  ) is that it only measures the extent to which an item
discerns between the two specific attribute patterns, α and α  . The CDI takes into
account all possible attribute pattern pairs; this index is a weighted average of the
2 K (2 K − 1) pairwise K i (α, α  ) values. Because it is more difficult to disentangle
patterns with many identical elements than those with many nonidentical elements,
the CDI gives more weight to pairs of patterns with more identical elements. In
particular, the weight, which we denote ξ (α, α  ), is the reciprocal of the number of
nonidentical elements:

1
ξ (α, α  ) = .

K
(αk − αk )2
k=1

The CDI for item i, CDIi , is then defined by the following weighted average of
K i (α, α  ) values (Henson & Douglas, 2005):

ξ (α, α  )K i (α, α  )
α=α 
C D Ii =  .
ξ (α, α  )
α=α 

Henson and Douglas (2005) also provided an algorithm for using the CDI to con-
duct ATA, incorporating practical constraints such as content balance if necessary. In
their algorithm, items are picked one at a time. Suppose that the test design specifies
N items to be chosen out of the pool of I items, subject to a set of general constraints
set by the practitioner (examples of constraints include balance across attributes and
balance across the answer key). At each stage of selection, n = 1, . . . , N, every can-
didate item is checked to determine whether all constraints can be met, given that the
item is chosen. Items for which this statement is not true cannot be selected. Among
the remaining items, the one with the largest CDIi is added to the test. In the special
case where there are no constraints, this reduces to selection of the N items with the
highest CDIi values.

277
 The GA for CDMs
The CDI is an analytic procedure for finding a solution to the test assembly prob-
lem. In this section, a different approach based on GAs is undertaken. The GA and
CDI are both heuristic methods, but GA uses a local search heuristic instead of an
explicit formula to find its solution. As will be seen, a benefit of the GA is that it can
be tailored to the outcome of interest (namely, the error rates of the test), rather than
through an intermediate function based on Kullback-Leibler information. Thus, al-
though the CDI is a reasonable index and performs significantly better than randomly
determined tests (Henson & Douglas, 2005), the GA’s direct attention to error rates
carries the potential to enhance this improvement. Additionally, unlike the CDI, the
GA offers the ability to achieve a target error rate for each attribute.

Genetic Algorithms
GAs are a class of techniques to find or approximate the optimal solution to a prob-
lem that is difficult to solve analytically. They were introduced by Holland (1973)
in the computer science literature and can be applied to a diverse set of problems.
GAs are a type of local search algorithm; they gradually converge to a final solution
through a series of slight adjustments. The term “genetic” derives from an analogy to
evolution: Candidate solutions compete with one another, with “stronger” solutions
more likely to survive and produce other potential solutions known as “children.”
The system of candidates thus evolves and improves until it stabilizes. The GA then
selects the best candidate in the system as its official solution.
In the case of ATA, the  goal is to pick a set of N items out of a larger pool of
I items. There are thus NI candidate sets of items from which to select. For most
realistically sized item pools, there are far too many candidates to examine each
one and choose the best. An analytic method like CDI may be used to make the
selection; in this section, we instead study the potential of GAs to propose a suitable
set.
The use of GAs and other local search algorithms is not new to the measure-
ment field. In the domain of dimensionality analysis, Zhang and Stout (1999) uti-
lized a GA to optimize the DETECT index. In the domain of ATA, Veldkamp
(1999) used a local search algorithm called simulated annealing to assemble tests
in the face of multiple objectives. van der Linden (2005) gave a general discus-
sion of GAs in the ATA framework; Verschoor (2007) considered the topic in
great depth, providing specific details on applying GAs to ATA with a unidimen-
sional IRT model. Sun, Chen, Tsai, and Cheng (2008) also applied GAs alongside
a unidimensional model, using them to assemble parallel test forms. Such works
all illustrate the utility of local search algorithms as tools to solve measurement
problems.
A GA begins with S initial candidate solutions that are selected to meet all content
constraints; the S solutions may be based on statistical criteria or simply chosen at
random from I ∗N , the set of solutions satisfying all constraints. For instance, S identi-
cal copies of the solution based on maximum CDI may be used. Alternatively, initial
sets may be randomly generated and evaluated; the first S adhering to all constraints

278
 Automated Test Assembly for CDMs Using a Genetic Algorithm

are chosen. The initial S solutions are known as the “parents” of the first stage; each
parent is composed of N items. The parents are then blended together or altered to
form M children, with M > S; typically, each child is constrained to be a member
of I ∗N . From this set of M children, S are chosen to “survive” to the next stage, with
more desirable children more likely to survive. Desirability is measured by a fitness
function (Verschoor, 2007)—a statistical criterion specifically constructed to rate the
children. Once the S children have been selected for survival, those children become
the parents for the next stage, creating their own children that will in turn be rated
by the fitness function for survival. The process continues until either the fitness
function ceases to improve from stage to stage (that is, the system converges), or a
prescribed number of stages has passed. Once the stopping rule has been reached,
the best solution remaining in the system is chosen.
Given a set of parent solutions, there are several ways to create children for the next
stage. One method is recombination, also called crossover, by which two parents are
blended into a child (van der Linden, 2005). The GA of this section uses mutation,
rather than crossover; mutation alters the features of one parent to produce a child.
As in previous sections, let i index the items in the pool. We assume that a parent
is written as a vector of length N , where the vector components are the indexes
of items in that parent. In mutation, a subset of a parent’s items is removed and
replaced by other items, creating a child. That is, l of the parent’s items are exchanged
with another l items from the pool (the value of l can be either deterministic or
random). As always, the exchange is made judiciously so that the child satisfies all
constraints.

Special Considerations for CDMs

A main goal of this article is to develop a GA that is specifically tailored to CDMs.
As stated earlier, the combination of GAs and ATA is not novel; what is new here
is their joint application to CDMs. Previous ATA methods using GAs do not “pour
over” to CDMs because the usual fitness function is inapplicable. Specifically, the
test information function (TIF) was instrumental to the fitness function of Verschoor
(2007) and is a standard ATA criterion in unidimensional modeling; however, the TIF
is generally undefined in CDMs due to these models’ discretization of the ability
space (Henson & Douglas, 2005). Therefore, a new fitness function is needed in
order to use the GA alongside popular CDMs such as DINA, NIDA, compensatory
MCLCM, and RUM. This section will introduce three fitness functions, denoted F 1 ,
F 2 , and F 3 , that are tailored to CDMs.
To determine a suitable fitness function, recall that CDMs operate under a classi-
fication framework. Every examinee is categorized into one of multiple levels (here,
two) along each of the K attributes. The error rates of the respective attributes are
thus a fundamental measure of test accuracy. One natural desire would be to mini-
mize the average number of classification errors that are produced in examination.
For an individual K examinee with a true attribute vector of α, the number of classifica-
tion errors is k=1 |αk − α̂k |, where α k is the true component of α along dimension
k and α̂k is the corresponding estimate based on the test. The expected number of

279
Finkelman, Kim, and Roussos

errors with respect to a prior distribution on α, P(α), is then given by

K

 
F1 ≡ P(α)E α |αk − α̂k | , (3)
α k=1

where E α (.) denotes expected value under α. F 1 is our first fitness function, with
lower values considered better results. Note that this definition does not conform to
the common convention that higher values of the fitness function connote “better”
results, but this fact does not hinder its efficacy in any way. The negative of F 1 may
be taken rather than F 1 itself, if it is desired that this convention be followed.
Equation 3 combines the error rates of the different attributes into a single index.
However, as explained in the introduction, a practitioner may seek adequate precision
along each attribute, rather than reducing the problem to an overall index. In this
case, the error rates of the individual attributes may be considered. The error rate for
attribute k is

ek ≡ P(α)E α (|α k − α̂k |).
α

To ensure that all attributes are measured precisely, a minimax approach may be
taken, whereby we seek to minimize the maximum of the K error rates. That is, we
minimize
K 
F2 ≡ max P(α)E α (|αk − α̂k |).
k=1
α

F 2 may be used as a fitness function, with lower values considered better.

Either of the previous fitness functions is appropriate if the practitioner’s goal is
to achieve maximal precision. However, some applications of testing may require
that a test meets a target level of precision. As stated in the introduction, one notable
example is the case where a test is given annually, and it is desired that the error rates
of the current year match those of the previous year as closely as possible. In this
situation, the estimated error rates of the previous test become the target.
Let ε k denote the target error rate for attribute k; these values can be set a priori or
to match the error rates of a previous form. A measure of the difference between the
true error rates ek and their targets ε k is the sum of all absolute distances:


K
F3 ≡ |ek −εk |.
k=1

As with F 1 and F 2 , lower values of F 3 values are considered better.

One complicating factor in applying the above fitness functions to CDMs is that
they are very difficult to compute analytically. Therefore, we propose that they be
evaluated via simulation. In particular, B simulees are produced from the prior

280
 Automated Test Assembly for CDMs Using a Genetic Algorithm

distribution P(α), with B ∗ P(α) simulees exhibiting a true state of α. Each simulee
is administered all I items in the pool; that is, there is a single run in which a response
for every simulee and every item is simulated. Based on this simulation run, the fit-
ness functions may be evaluated for any candidate set of items. First, for simulee j
and candidate set of items u i , the posterior distribution of α j is computed based only
on the set of responses, X i j , that correspond to u i :

P(α j | X i j )L(α j ; X i j )
P(α j | X i j ) =  .
P(α j | X i j )L(α j ; X i j )
α

From this posterior distribution, a classification of simulee j is made with respect

to the candidate set u i . In particular, the simulee’s ability estimate α̂ j is the vector
that minimizes the expected number of errors with respect to the posterior distribu-
tion. That is, α̂ j is the vector minimizing
 K

 
P(α j | X i j ) |αk − α̂k | . (4)
α k=1

Once α̂ j has been determined, the expected posterior error rate for attribute k can
be determined for simulee j. Let α̂k j represent the observed classification for simulee
j along attribute k. Let P(αk = 1 | X i j ) denote the posterior probability of mastery
along attribute k for simulee j:

P(αk = 1 | X i j ) = [P(α j | X i j )].
α j :αk j =1

Then the expected posterior error rate, E(ek j | X i j ), for simulee j and attribute k is
equal to P(αk = 1 | X i j ) if α̂k j = 0; it is equal to 1 − P(αk = 1 | X i j ) if α̂k j = 1.
The overall expected posterior error rate for attribute k is then obtained by averaging
across all simulees:


B
E(ek j | X i j )
j=1
ēk = .
B
The values ēk are then used as a building block to evaluate any of the fitness functions
above—that is, F  K set u i . These three fitness functions
1 , F 2 , or F 3 —for the candidate
K
are evaluated by k=1 ēk , maxk=1
K
ēk , and k=1 |ēk −εk |, respectively.
In light of the somewhat complicated simulation procedure described above, two
observations are worth making. The first of these observations is that the use of
simulation is not unprecedented in item selection. Most notably, simulation is a fun-
damental part of several popular computerized adaptive testing (CAT) methods to

281
Finkelman, Kim, and Roussos

curb the exposure rates of items (Stocking & Lewis, 1998; Sympson & Hetter, 1985).
Although such simulation adds more time to the process, it has not prevented these
CAT methods from being used operationally. The second observation is that due to
the complexity of the above evaluation criteria, linear programming methods (van
der Linden, 2005) for direct optimization are not as easily applicable as GAs. The
three fitness functions introduced above define optimality with respect to a Bayesian
prior distribution, whether that optimality is with respect to the minimization of the
average number of errors, the minimax criterion, or the distance-to-target error rates.
The relation between error rates and item parameters cannot be written as a linear
function, and therefore linear programming is unable to directly optimize these error
rates. However, the GA proposed in the next subsection can be used as a tool for
optimizing (or approximating the optimum of) any of the three functions. This ap-
plication to CDMs is thus an illustration of the added flexibility of GAs over linear
programming techniques; such flexibility was previously listed as a benefit of GAs
and other local search heuristics by van der Linden (2005) and Verschoor (2007).

The Specific GA for CDMs

As a general rule, there is a tradeoff between the quality of a GA’s solution and its
running time (van der Linden, 2005). After all, by allowing a GA to run longer, we
enhance the chances of finding the optimal value of our selected fitness function. Due
to the many calculations performed by the GA (described in the previous subsection),
maintaining a reasonable running time is an important consideration in the current
application to CDMs. To achieve the correct balance between running time and the
solution’s fitness, the GA proposed in this subsection is relatively simple. Although
this simple GA may be more likely to find a local (rather than global) optimum, it is
more practicable in terms of computational expense.
The first manner in which our chosen GA is simplified regards the creation of
children from parents. We have used a mutation scheme in which items of the parent
vector are perturbed to new values, one item at a time (mathematically, this amounts
to setting l = 1). As in previous sections, let N denote the prescribed number of
items for the test, and let s = 1, . . . , S index the S parent vectors at the current itera-
tion of the GA. Now let (i s1 , . . . , isN ) represent the items comprising parent s. For the
current iteration, N children are created for each parent. The first child is created by
removing item i s1 and replacing it with another item, i s1 , so that the child satisfies
all content constraints. Following the notation above, the first child contains items
(i s1 , . . . , isN ). Similarly, the second child is composed of items (i s1 , i s2 , . . . , isN ),
and so forth until child N , which is composed of items (i s1 , . . . , i sN ). A clone of
each parent is also retained as an additional child; that is, a parent at the current
iteration is eligible to be a parent at the next iteration, if it is superior to the chil-
dren. Thus, there are N + 1 children propagated by each parent, for a total of M =
S(N + 1) children.
The above procedure is a simplified GA in that it uses only mutation, not crossover.
By only mutating one item at a time, we facilitate the search for children that meet
all constraints. After all, for each item in parent s (which itself has been chosen to

282
 Automated Test Assembly for CDMs Using a Genetic Algorithm

meet each constraint), it is trivial to identify and restrict attention to potential re-
placement items whose addition would not violate any constraints. Note also that we
have proposed each item to be replaced in exactly one child; this rule is more sys-
tematic than determining the items to be replaced at random. In particular, it allows
for the potential replacement of each item at every stage, facilitating the identifi-
cation and removal of less informative items. It was believed that this systematic
approach would remove such items more quickly and consistently than random re-
placement: By chance alone, a random replacement might skip the potential removal
of a problematic item for several iterations. Though not yet verified empirically, the
systematic method was thus expected to result in faster convergence to an adequate
solution than a random replacement of items.
The second manner in which our GA is simplified is in the selection of children
for survival. As stated earlier, the children with more desirable values of the fitness
function should be more likely to survive. If there is an element of randomness to
the choice of children for survival, then the selection is said to be stochastic; if, on
the other hand, only the strongest children are chosen, then the selection is deter-
ministic (Verschoor, 2007). Although stochastic selection is able to backtrack from
local optima, its adoption of interim solutions with suboptimal fitness causes a more
gradual progression toward convergence. On the other hand, deterministic selection
does not backtrack; because interim solutions always move forward in the direction
of better fitness, a more direct path is taken to the ultimate solution. We therefore
expect that a deterministic rule will result in faster convergence to an optimum, al-
beit perhaps a local one. To reduce computation time, as well as to eliminate the
need for fine tuning associated with stochastic selection, we utilize deterministic
selection in the GA of this article: the S children with the best fitness values are
chosen.
Overall, our GA selects a test form for a CDM using the following steps:
1. Generate responses for B simulees and all I items in the pool.
2. From the given pool of items, select S initial “parent” solutions.
3. Perform the mutation scheme described above for each parent, resulting in
S(N + 1) children.
4. Evaluate the fitness function for all children. The S children exhibiting the best
values of the fitness function are selected as parents for the next iteration.
5. Repeat steps 3 and 4 until the stopping rule is invoked.
6. Choose the solution in the system that exhibits the best value of the fitness
function. This becomes the official solution of the GA.
We emphasize that more complicated variations of the GA could also be em-
ployed; such variations include the addition of crossover, the random replacement
of items (rather than the replacement of each item in exactly one child), the use of
stochastic selection, and the allowance of interim solutions to venture into infeasible
space (rather than requiring them to satisfy all constraints). Provided that they are
well implemented, these complex variations enhance the GA’s ability to avoid local
optima. The extent to which such complications would change the results obtained
from the GA outlined above, as well as the extent to which they would increase

283
Finkelman, Kim, and Roussos

computation time, are questions that will be answered in future work. The goals
of this article, however, are to introduce the GA method to CDMs and investigate
whether it offers improvement over the CDI; therefore, our current attention will be
restricted to these purposes.
Simulations
Background and Design
The previous two sections described algorithms for combining ATA with CDMs,
including one previously proposed method (CDI) and one new method (GA). The
goal of this section is to compare these methods in simulation. Specifically, the two
approaches were evaluated based on their abilities to minimize the fitness functions
F 1 , F 2 , and F 3 . Results were analyzed separately for each of these three functions.
The simulation design used two item pools, two prior distributions on α, and two
levels of item constraints; these factors were completely crossed for a total of eight
conditions.
Item pools. Each item pool consisted of 300 items; the two pools were based on the
same Q-matrix, which contained five attributes. The choice of 300 items and five
attributes was similar to the design of Henson and Douglas (2005), whose method
involved simulations with a pool of 300 items and either four or eight attributes.
Because items typically do not measure all attributes, this study’s Q-matrix was
constrained to have 80 items measuring one attribute, 140 items measuring two at-
tributes, and 80 items measuring three attributes, for an average of two attributes per
item. Within each item, the attribute or attributes being measured were randomly de-
termined. This procedure resulted in 135 items measuring attribute 1, 115 measuring
attribute 2, 108 measuring attribute 3, 115 measuring attribute 4, and 127 measuring
attribute 5 across the two pools.
The two item pools were different in that one consisted of more highly discrimi-
nating items than the other. In both cases, items were assumed to follow the reduced
RUM, with specific item parameters generated at random. For both pools, every π i∗
was randomly generated from the U [0.75, 0.95] distribution. All rik∗ values were
randomly generated from U [0.40, 0.85] for the pool with high discrimination and
U [0.65, 0.92] for the pool with low discrimination. These ranges of rik∗ are similar
to those obtained in analyses by Jang (2005, 2006) and Roussos, Hartz, and Stout
(2003).
Prior distributions on α. First, we investigated the situation where the Bayesian prior
on α was uninformative, that is, where each of the 25 = 32 possible α vectors had
P(α) = 1/32. The second prior distribution corresponded to the case where exam-
inees’ latent abilities followed the multivariate standard normal distribution; these
latent abilities were then compared to specified cut points in order to determine mas-
tery or nonmastery on each attribute. The tetrachoric correlation of each pair of at-
tributes was set to .5, and the cut points were set so that the percentage of masters on
each attribute was 45, 50, 55, 60, and 65 for attributes 1, 2, 3, 4, and 5, respectively.
This prior distribution was considered more realistic than the uninformative prior:
in educational measurement, latent abilities are usually considered to be positively

284
 Automated Test Assembly for CDMs Using a Genetic Algorithm

correlated (Henson & Douglas, 2005), and the percentage of masters is seldom ex-
actly 50% for all attributes.
Constraints. Two different levels of constraints were specified. The first level im-
posed no constraints and was viewed as a baseline condition. The second level was
more realistic, imposing two types of constraints: a lower bound on the number of
times that each attribute must be measured, and a proper balance across the answer
key. Specifically, it was required that the 40 selected items measured each attribute
at least 15 times; a similar constraint on content had been studied by Henson and
Douglas (2005). This constraint ensured that all content areas would be adequately
represented, and thus the reporting of subscores would be defensible. Each item was
also randomly assigned a correct answer of A, B, C, or D; candidate item sets were
considered feasible only if each answer choice was represented between 8 and 12
times. Such balance ensures that capable examinees will not be distracted by a per-
ceived bias toward one answer choice over another.
It should be noted at this juncture that the use of more complicated substantive
constraints can be easily incorporated into the GA. For example, constraints on more
fine-grained content areas could be introduced within the measurement of each at-
tribute, as well as more sophisticated constraints having to do with student cognitive
modeling (such as different strategies leading to the use of different Q-matrices).
Other simulation details. One of the GA’s benefits is that it is flexible to the goal:
Unlike the CDI, it tailors itself to the choice of fitness function. To test this character-
istic, the GA was run for each of the fitness functions F 1 , F 2 , and F 3 and compared
with the CDI in simulation results. That is, the fitness function was considered an in-
put to the GA, so that the GA was tailored to the prescribed function when selecting
its items. Each GA solution was only evaluated with respect to the fitness function to
which it was tailored. In all cases, S = 3 copies of the CDI solution were chosen as
initial parents. Setting S = 3 had been successful in a previous psychometric applica-
tion of GAs (Zhang & Stout, 1999) and was expected to provide adequate variation
in the children without creating an intractable computational burden.
As described in the subsection entitled “Special Considerations for CDMs,”
the GA utilizes preliminary simulations in order to conduct item selection. For
each prior distribution and item pool, 20,000 “training” simulees were tested, with
20,000 P(α) simulees exhibiting a true ability vector of α. There were thus four
training data sets, from which both the constrained and unconstrained item selec-
tions were made. The GA was stopped if the fittest solution in the system remained
the same for 50 iterations, or if 500 total iterations were run, whichever came first. A
FORTRAN 6.1.0 program was specially written to carry out all steps of the analysis.
To compare the methods, a common set of simulees was needed. The training data
sets were not ideal for this purpose because the GA had selected its items based on
these data sets; hence, the GA would be likely to “capitalize on chance” if the training
sets were also used in evaluation. In other words, comparison of the methods using
training data would constitute an unfair advantage for the GA. To avoid this problem,
four new sets of 20,000 simulees were produced, again with true α values generated
proportional to their prior probabilities, and these test sets were used to evaluate the

285
TABLE 1
Average Number of Classification Errors, by Condition and Method

Constraints Prior Item Discrimination CDI GA GA% Gain

Yes Uniform High .402 .342 15.0%
Yes Uniform Low .879 .836 4.9%
Yes Corr = .5 High .286 .256 10.5%
Yes Corr = .5 Low .617 .608 1.4%
No Uniform High .418 .335 19.9%
No Uniform Low .879 .837 4.8%
No Corr = .5 High .292 .258 11.6%
No Corr = .5 Low .620 .611 1.3%

TABLE 2
Maximum Attribute Error Rate, by Condition and Method

Constraints Prior Item Discrimination CDI GA GA% Gain

Yes Uniform High .118 .072 39.0%
Yes Uniform Low .226 .182 19.6%
Yes Corr = .5 High .084 .057 32.4%
Yes Corr = .5 Low .159 .125 21.1%
No Uniform High .139 .071 48.5%
No Uniform Low .241 .176 27.2%
No Corr = .5 High .096 .056 41.9%
No Corr = .5 Low .166 .132 20.7%

methods. Simulees were classified based on the items selected by each ATA method,
using the α̂ estimates minimizing expression 4. The true α values were compared to
these α̂ estimates, yielding observed values of F 1 , F 2 , and F 3 .

Results Comparing the Methods

Table 1 gives the average number of classification errors for each condition and
ATA procedure, based on the test sets. This outcome measure corresponds to the
observed values of F 1 . Table 1 indicates that the GA resulted in an improvement
over CDI on all test sets. Over the eight conditions studied, the GA’s percentage
reduction of average errors ranged from 1.3% to 19.9% as compared to the CDI;
the greatest reductions were found in the highly discriminating item pool. As in
Henson and Douglas (2005), the CDI’s average number of errors actually decreased
under conditions with constraints, due to a greater balance of attributes measured. In
training data, the GA always exhibited better performance without constraints than
with them (results not shown); as Table 1 indicates, this trend sometimes reversed in
the test data.
Next, we compared the methods based on their observed values of F 2 . Here, the
GA always exhibited a substantial advantage over the CDI, with percentage reduction

286
TABLE 3
Sum of the Absolute Distances Between Realized and Target Error Rates, by Condition and
Method
Item Target Error Rate GA%
Constraints Prior Discrimination Per Attribute CDI GA Gain
Yes Uniform High .10 .135 .009 93.2%
Yes Uniform Low .20 .173 .015 91.4%
Yes Corr = .5 High .10 .214 .011 95.1%
Yes Corr = .5 Low .15 .151 .015 89.8%
No Uniform High .10 .179 .010 94.2%
No Uniform Low .20 .230 .005 97.7%
No Corr = .5 High .10 .208 .011 94.8%
No Corr = .5 Low .15 .175 .004 97.9%

of maximum error rate ranging from 19.6% to 48.5%. Again, the observed reductions
tended to be larger for the highly discriminating item pool. See Table 2 for a complete
list of maximum error rates.
Table 3 gives the sum of the absolute distances between realized and target er-
ror rates (i.e., observed values of F 3 ) for each condition and method. Not sur-
prisingly, the GA performed much better than selection based on maximum CDI,
which was not tailored to the targets. The GA’s percentage reduction, compared
to CDI, was always at least 89.8% over all eight conditions; the highest reduction
was 97.9%.
The above simulations assume that the P(α) values are known. When an informa-
tive prior is utilized, this assumption should generally benefit methods, such as the
GA, that incorporate the P(α) into the item selection process. Thus, one limitation
of the simulation study is that it does not investigate the GA’s robustness to a mis-
specified prior. Nevertheless, the conditions involving a uniform prior can be utilized
to make a “fair” comparison between the CDI and GA. In these cases the GA did not
benefit from informative P(α) values, but it achieved better results than the CDI for
all three fitness functions considered.

How Good Is the GA’s Solution?

The previous subsection assessed the GA’s performance on test data. Attention
now turns to its fundamental ability to locate a high-quality solution from the training
data, over which the search is actually conducted. The following analysis is designed
to evaluate the extent to which the GA is adept at finding good approximations to the
optimal solution, as opposed to a local optimum far from the global optimum.
An ideal evaluation of the GA
 solution would involve systematically calculating
the fitness of all possible 300
40
item sets for each condition, then comparing the GA
solution’s fitness compared to that of the optimum. This method is computationally
intractable, however, due to the fact that this is an astronomically high number of
item sets. Therefore, three other approaches were taken to analyze the GA’s ability
to find a high-quality solution within the training data. The three approaches were:

287
Finkelman, Kim, and Roussos

(a) comparison of the GA solution with the distribution of values observed in 4,000
randomly determined item sets, (b) comparison of GA results when varying  the ini-
tial parents, and (c) comparison of the GA solution with all possible NI item sets,
with I and N chosen to be computationally tractable. None of these short investiga-
tions represents an irrefutable confirmation that the GA solution is always close to
optimal; nevertheless, they give some evidence of the GA’s performance with respect
to optimality.
For brevity, F 1 was considered to be the only fitness function of interest in all in-
vestigations. Also, only one condition was analyzed in each investigation, namely the
condition with highly discriminating items, a uniform prior on α, and no constraints.
The choice of a uniform prior and no constraints was made because of the simplicity
and interpretability of these features; the choice of the highly discriminating item
pool was arbitrary. Due to the complex relationship between item parameters and
error rates, as well as the complexity of GA methodology itself, it is difficult to pre-
dict which conditions will yield similar results to those obtained here. In general,
the relative quality of a GA’s solution may be expected to depend upon the size
of the item pool, the number of items to be selected, the degree to which the opti-
mal solution is superior to other candidate solutions, and the proximity of the initial
solutions to the optimal solution (as well as their proximity to local optima). There-
fore, caution should be observed when generalizing the following results to other
conditions.
Comparison with randomly determined item sets. In this investigation, 4,000 item
sets out of the possible number were chosen at random and compared to the GA so-
lution (as well as the CDI) based on performance over the training data. Of the 4,000
randomly determined item sets, the best set had an average number of classification
errors of .539. The mean of the distribution for the 4,000 values was .675, and the
standard deviation was .043. Both ATA methods resulted in values (.411 for CDI,
.329 for GA) that were substantially better than any of the randomly selected sets.
Note that the values presented here are different from their counterparts in Table 1,
which were the results of test data rather than training data.
Comparison of GA results when varying the initial parents. In all previous applica-
tions of the GA considered herein, the same initial parents were used: three copies of
the CDI solution. A natural question is whether the GA’s solution would converge to
the same fitness value if initial parents other than these were input, and if not, how
close the two values would be. Therefore, the GA was rerun using three randomly
chosen item sets as initial parents. If varying the initial parents resulted in markedly
different fitness values, it would cast doubt upon the GA’s ability to consistently find
high-quality solutions. Such a result did not occur in this investigation, however;
comparing the results for the two sets of initial parents, the associated difference in
fitness was in the third decimal place (.329 using the CDI as initial parents, .331
using random initial parents). The fitness value of the CDI was .411; those of the
random initial parents were .612, .690, and .727. Thus, the fitness values of the GA
solutions were much closer to each other than to their starting values, suggesting that
the difference in quality due to varying the initial parents is negligible.

288
 Automated Test Assembly for CDMs Using a Genetic Algorithm

Comparison of the GA solution with all possible item sets. As explained above, com-
puting the fitness of all possible item sets is intractable. However, it is possible to
compute the fitness of all possible item sets in a smaller item pool, and then deter-
mine the GA solution’s rank with respect to this smaller pool. For this subsection, a
“subpool” of 20 items was selected at random from the highly discriminating pool.
The GA was then run on this subpool; its task was to select the 10 subpool items
yielding the smallest average number of classification errors over the training data.
Three copies of the CDI solution served as initial parents for the GA.
Both the CDI and GA found top-tier solutions within the subpool. The CDI solu-
tion had anaverage
 of 1.35 classification errors per simulee; this value ranked 11th
out of the 2010
= 184,756 possible item sets. The GA solution had an average of
1.30 classification errors per simulee, ranking first out of all possible sets. Although
optimal selection within this small subpool does not guarantee optimal selection for
the complete 300-item pool, it is an encouraging result nonetheless.

Summary and Discussion

Local search algorithms, though relatively new to the psychometric literature, have
been successfully used in several ATA applications (Veldkamp, 1999; Verschoor,
2007). In this article, a new GA was introduced to select items for CDM. This GA
circumvents potential problems arising from nonexistence of Fisher information in
CDMs by utilizing fitness functions that directly relate to classification accuracy.
The three particular functions defined were average number of classification errors,
maximum error rate, and distance to a set of target error rates.
Several theoretical and practical benefits of using the GA are as follows:

1. In contrast to methods that select items based on individual merit, such as the
CDI, the GA evaluates each item set as a unit. Such holistic selection guarantees
that the chosen items complement one another. Although linear programming
methods also treat item sets holistically, they are unable to directly optimize
Bayesian error rates alongside CDMs, as does the GA considered in this article.
2. The GA is flexible to the practitioner’s goal: it can be applied whether the fitness
function is the average number of classification errors, the maximum error rate,
proximity to target error rates, or any other numerical criterion.
3. The GA takes initial solutions as parents, then finds a solution that performs
as well as or better than those parents. Thus, any solution can potentially be
improved upon by inputting it to the GA as a parent.

We have given evidence that the GA is not sensitive to its initial parents. However,
using the CDI as an initial parent carries two advantages over random starting points.
First, because the GA’s solution cannot be worse than its initial parents, using the
CDI guarantees that the GA will never be outperformed by it. Second, because the
CDI solution is generally closer to optimal than randomly determined tests (Henson
& Douglas, 2005), its input as an initial parent tends to reduce the GA’s computa-
tional time. Thus, our proposed procedure may be considered a sequential method
where the practitioner first finds the CDI solution, then refines this solution via the
GA.

289
Finkelman, Kim, and Roussos

The “Simulation Results” section compared this sequential GA to the CDI using
simulation. The conclusions to be drawn from such simulations are specific to the
fitness function studied. When the defined fitness function was average number of
classification errors, the GA achieved a modest-to-significant gain in accuracy as
compared to the CDI. For this fitness function, the decision of whether to use the
GA or CDI depends on the cost of computation relative to the cost of classification
error. The CDI may be adequate if computation is expensive; on the other hand, if
classification error is considered more expensive, the GA can help reduce this cost.
Turning to the results for maximum error rate, the GA achieved a substantial gain
relative to the CDI under all eight conditions. Therefore, if a practitioner seeks to
avoid any high attribute-level error rates, the simulations herein recommend the GA
as the superior method. Finally, if the goal is to hit attribute-level target error rates,
then no method based on CDI is adequate, as there is no known function relating the
CDI to error rates. The GA, on the other hand, was able to hit prescribed targets with
high precision. Thus, whenever target error rates are sought, GA adds an important
dimension beyond what can be provided using the CDI.
Clearly, the widespread use of the GA is dependent on the ability of comput-
ers to perform the requisite tasks in an adequate amount of time. Because the GA
requires the use of simulated training data to assemble tests, its computational in-
tensity is considerably greater than that of analytic procedures like the CDI. We
note, however, van der Linden’s (2005) statement that “the current enormous power
of our computers has stimulated the interest in local search heuristics” (p. 94).
As this power continues to grow, the computational expense of the GA may be
expected to shrink in comparison to its contribution to enhancing classification
accuracy.
Further study includes simulations using CDMs other than the reduced RUM,
as well as varying the number of attributes, discrimination of item parameters, and
complexity of constraints. Robustness of the GA to a misspecified prior distribution
should also be examined. Furthermore, although we have provided a preliminary
analysis of the GA’s proximity to the optimal solution, this analysis was limited to
the condition with highly discriminating items, a uniform prior on α, and no con-
straints; moreover, the comparison of the GA with all possible item sets was limited
to a small subpool. Additional investigation is required to determine general condi-
tions when the procedure finds the optimal solution and when it does not. Finally,
it is important that the CDM ATA methods developed in this paper be applied in a
wide variety of real-data settings. The comparison of tests assembled by computer
and by test developers will likely produce important insights that will improve the
CDM ATA methods. These and other extensions of this article will be addressed in
future work.

Acknowledgments
The authors are indebted to Jon-Michael Brasfield and Angela Verschoor for help-
ful discussions of GAs.

290
 References
Birnbaum, A. (1968). Some latent trait models and their use in inferring an examinee’s ability.
In F. M. Lord & M. R. Novick (Eds.), Statistical theories of mental test scores (pp. 395–
479). Reading, MA: Addison-Wesley.
Chang, H., & Ying, Z. (1996). A global information approach to computerized adaptive test-
ing. Applied Psychological Measurement, 20, 213–229.
de la Torre, J., & Douglas, J. A. (2004). Higher-order latent trait models for cognitive diagno-
sis. Psychometrika, 69, 333–353.
Haertel, E. H. (1984). An application of latent class models to assessment data. Applied Psy-
chological Measurement, 8, 333–346.
Henson, R., & Douglas, J. (2005). Test construction for cognitive diagnosis. Applied Psycho-
logical Measurement, 29, 262–277.
Henson, R. A., Roussos, L. A., Douglas, J. A., & He, X. (2008). Cognitive diagnos-
tic attribute-level discrimination indices. Applied Psychological Measurement, 32, 275–
288.
Holland, J. (1973). Genetic algorithms and the optimal allocation of trials. SIAM Journal of
Computing, 2, 88–105.
Jang, E. E. (2005). A validity narrative: Effects of reading skills diagnosis on teaching and
learning in the context of NG TOEFL. Unpublished doctoral dissertation, University of
Illinois, Urbana-Champaign.
Jang, E. E. (2006). Pedagogical implications of cognitive skills diagnostic assessment for
teaching and learning. Paper presented at the annual meeting of the American Educational
Research Association, San Francisco.
Junker, B. W., & Sijtsma, K. (2001). Cognitive assessment models with few assumptions, and
connections with nonparametric item response theory. Applied Psychological Measurement,
25, 258–272.
Lord, F. M. (1980). Applications of item response theory to practical testing problems. Hills-
dale, NJ: Erlbaum.
Maris, E. (1999). Estimating multiple classification latent class models. Psychometrika, 64,
187–212.
Nichols, P. D., Chipman, S. F., & Brennan, R. L. (1995). Cognitively diagnostic assessment.
Hillsdale, NJ: Erlbaum.
Reckase, M. D. (1985). The difficulty of test items that measure more than one ability. Applied
Psychological Measurement, 9, 401–412.
Roussos, L. A., DiBello, L. V., Stout, W., Hartz, S. M., Henson, R. A., & Templin, J. L. (2007).
The fusion model skills diagnosis system. In J. P. Leighton & M. J. Gierl (Eds.), Cognitive
diagnostic assessment for education: Theory and applications (pp. 275–318). Cambridge,
UK: Cambridge University Press.
Roussos, L. A., Hartz, S. M., & Stout, W. M. (2003). Real data applications of the fusion model
skills diagnostic system. Paper presented at the annual meeting of the National Council on
Measurement in Education, Chicago.
Roussos, L. A., Templin, J. L., & Henson, R. A. (2007). Skills diagnosis using IRT-based
latent class models. Journal of Educational Measurement, 44, 293–311.
Sinharay, S. (2006). Model diagnostics for Bayesian networks. Journal of Educational and
Behavioral Statistics, 31, 1–33.
Stocking, M. L., & Lewis, C. (1998). Controlling item exposure conditional on ability in
computerized adaptive testing. Journal of Educational and Behavioral Statistics, 23, 57–
75.

291
Finkelman, Kim, and Roussos

Sun, K.-T., Chen, Y.-J., Tsai, S.-Y., & Cheng, C.-F. (2008). Creating IRT-based parallel test
forms using the genetic algorithm method. Applied Measurement in Education, 21, 141–
161.
Sympson, J. B., & Hetter, R. D. (1985). Controlling item-exposure rates in computerized adap-
tive testing. In Proceedings of the 27th annual meeting of the Military Testing Association
(pp. 973–977). San Diego, CA: Navy Personnel Research and Development Center.
Tatsuoka, K. K. (1985). A probabilistic model for diagnosing misconceptions in the pattern
classification approach. Journal of Educational Statistics, 12, 55–73.
Theunissen, T. J. J. M. (1985). Binary programming and test design. Psychometrika, 50, 411–
420.
van der Linden, W. J. (2005). Linear models for optimal test design. New York: Springer.
van der Linden, W. J., & Boekkooi-Timminga, E. (1989). A maximin model for test design
with practical constraints. Psychometrika, 54, 237–247.
Veldkamp, B. P. (1999). Multiple objective test assembly problems. Journal of Educational
Measurement, 36, 253–266.
Verschoor, A. J. (2007). Genetic algorithms for automated test assembly. Unpublished doc-
toral dissertation, University of Twente, Enschede, The Netherlands.
von Davier, M. (2005). A general diagnostic model applied to language testing data (ETS
Research Report 05-16). Princeton, NJ: Educational Testing Service.
Zhang, J., & Stout, W. (1999). The theoretical DETECT index of dimensionality and its ap-
plication to approximate simple structure. Psychometrika, 64, 213–249.

Authors
MATTHEW FINKELMAN is an Assistant Professor, Tufts University School of
Dental Medicine, 75 Kneeland Street, Boston, MA 02111; matthew.finkelman@
tufts.edu. His primary research interests include computerized adaptive testing, test assem-
bly, and biostatistics.
WONSUK KIM is a Psychometrician I, Measured Progress, 100 Education Way, Dover, NH
03820; [email protected]. His primary research interests include equat-
ing and test assembly.
LOUIS A. ROUSSOS is a Psychometrician II, Measured Progress, 100 Education Way, Dover,
NH 03820; [email protected]. His primary research interests include
cognitive diagnosis models, dimensionality analysis, and differential item functioning.

292