Sie sind auf Seite 1von 9

Journal of Applied ftychology Copyright 1982 by the American Psychological Association, Inc

1982, Vol 67, No. 6, 826-834 0021 -9O1O/82/67O6-O826S00.75

An Empirical Comparison of Item Response Theory and


Hierarchical Factor Analysis in Applications to the
Measurement of Job Satisfaction
Charles K. Parsons Charles L. Hulin
College of Management University of Illinois at Urbana-Champaign
Georgia Institute of Technology

Responses to the Job Descriptive Index (JDI; N = 1,349) are used to empirically
compare a two-parameter logistic item response theory (IRT) model to a hier-
archical factor analytic model suggested by Humphreys (1962). A hierarchical
model was chosen because the JDI is typically conceptualized as having five or
more dimensions, whereas current IRT estimation procedures are based on the
assumption of a unidimensional trait. The hierarchical analysis allows the in-
vestigator to create a general factor based on the common variance among the
usual common factors. IRT item parameters estimated by a maximum likelihood
algorithm agree very well with the item loadings on the first principal factor and
a general satisfaction factor from a matrix of tetrachoric correlations. These results
are consistent with a hierarchical job satisfaction model that has one general
factor and multiple group factors, as well as a logistic IRT model with two
parameters. The authors conclude that IRT can be applied in the job satisfaction
domain, where data are typically multidimensional, to provide evidence about
the general satisfaction factor. Implications of this research for applications of
IRT are discussed.

Item response theory (IRT) has been pro- Although the IRT models are typically as-
posed as a solution to many measurement sociated with aptitude measurement, Schmitt
problems (Lord, 1980). The advantages of (1981) noted that various IRT theorists
IRT models are derived from the following (Hambleton & Cook, 1977; Lord, 1977;
properties. First, they provide a scale of mea- Wright, 1977) have suggested that the appli-
surement (called the 6 scale) that is indepen- cability of the models to attitude measure-
dent of the particular items included in a ment is also worth investigating.
measurement instrument. Second, they pro- Because such applications of the IRT
vide item parameters that are independent methods to attitude measurement are rare
of the sample of persons being measured. and the methods themselves are probably not
Some of the applications of such measure- well understood by most researchers, it be-
ment models were discussed in Hulin, Dras- hooves us to examine the extent to which the
gow and Parsons (1982) and Lord (1980), results of the newer methods converge with
and an introduction of IRT and methods was results from the more traditional methods
presented in Hambleton and Cook (1977). when examining issues of interest in attitude
measurement. For example, it is useful to
This research was supported in part by Contract compare both theoretically and empirically
N00O-14-C-O9O4 from the Office of Naval Research, the more traditional factor analysis methods
Charles L. Hulin, principal investigator; and in part by to the newer IRT methods for describing the
the Department of Psychology, University of Illinois.
The authors thank Michael Levine for his assistance in structure of attitude items. This article briefly
the research as well as several anonymous reviewers who describes the similarity and dissimilarity of
provided many helpful comments on earlier drafts of the factor analytic and IRT methods and
this article. then applies both to job satisfaction data to
Requests for reprints should be sent to Charles K.
Parsons, College of Management, Georgia Institute of answer the question concerning the trait that
Technology, Atlanta, Georgia 30332. is basic to a set of observations.

826
ITEM RESPONSE THEORY IN JOB SATISFACTION DATA 827

The Factor Analytic Model For the model discussed in this article,
Most factor analysis methods are based on P,(0) is given by the formula
a model that can be described as "a formal 1
model about hypothetical components which P,(6) = (1)
1 +e~
account for linear relationships that exist be-
tween observed variables" (Mulaik, 1972, p. The terms a, and b, are called item param-
96; emphasis added).1 In studies of the mea- eters, with a, reflecting item discrimination
surement of job satisfaction, factor analytic and b, reflecting item difficulty for the Ith
methods are used to determine the number item. D is a scaling factor usually set to 1.702
of hypothetical factors and the relations of so that this logistic function closely follows
items to these factors. Because factor analytic a normal ogive curve.
methods are linear, they are theoretically in- A graphic representation of this model is
correct for discrete variables, such as items represented by an S-shaped curve. The b pa-
with limited scored response options. For rameter reflects the position of the curve on
example, the frequently used job satisfaction the horizontal axis (0), and the a parameter
measure, the Job Descriptive Index (JDI; is related to the slope of the curve at its steep-
Smith, Kendall, & Hulin, 1969) with three est point. Lower b values shift the curve to
item-response alternatives, has frequently the left and lower a values flatten the curve.
been factor analyzed in order to answer ques- A third parameter is frequently added to
tions about convergent and discriminant va- the model when individuals who do not rec-
lidity (Dunham, Smith, & Blackburn, 1977), ognize the correct response are able and have
similarity of factor structure across subpop- a tendency to give a correct response to an
ulations (Golembiewski & Yeager, 1978; item with probability greater than zero. The
Smith, Smith, & Rollo, 1974), and the num- effect of nonzero value for this parameter on
ber of meaningful scales that can be derived an ICC is most obvious at the low level of
from the JDI (Yeager, 1981). 6 where the probability of a correct response
does not approach zero. (The ICC has a non-
zero lower asymptote.) For a job satisfaction
Item-Response Models questionnaire, it can be argued that as the
level of 0 approaches very low values
A formal item-response model states a re- (-•x), the conditional probability of re-
lation between the probability of some type sponding positively to an item goes to zero
of response by an individual and some char- because guessing does not occur. Parsons
acteristic of the individual. A factor analytic (Note 1) has determined empirically that this
model is one type of item-response model; is a tolerably good assumption for the Job
it describes the linear relation between item Descriptive Index. Therefore, the third pa-
responses and constructs. Recently, in the rameter will be set to zero (identical to not
area of psychometrics, there has been an in- including it in the model) for the present ar-
creasing emphasis on ogive or logistic item- ticle (see Hulin. Drasgow, & Parsons, 1982.
response models (Birnbaum, 1968). In most or Lord, 1980, for further discussion on the
applications the probability of an item re- various item response models).
sponse is a function of a unidimensional at- As afinalnote on Equation 1, it is obvious
tribute of the respondent and certain char- that the slope of a linear item response model
acteristics of the item called item parameters.
When unidimensionality is assumed, the
function is called an item characteristic curve 1
This paper will use the term factor rather than com-
(ICC) and gives the conditional probability ponent to denote the hypothetical element that is com-
that a randomly chosen person from the pop- mon to a set of observed variables. The distinction is
ulation of all people at a given value of the important when applying factor analysis because com-
trait gives a particular response. The notation ponents are usually derived to account for the variance
of observed variables, whereas factors are derived to ac-
commonly used denotes the attribute by (0) count for covariance among observed variables (see
and the ICC for the /* item by P, (0). Mulaik, 1972, for further details).
828 CHARLES K. PARSONS AND CHARLES L. HULIN

will be related to the a parameter from the The theoretical work by Lord and Novick
logistic model. In fact, for the two-parameter (1968) shows us the relation between the
normal ogive model (similar to the logistic models for the one-factor case, and Reckase's
model), (1979) Monte Carlo results show us the re-
sults for multiple orthogonal factors. The
a, = 7 7 = = , (2) correspondence of the IRT analysis of job
satisfaction items to one extreme or the other
where for the Ith item, a, is item discrimi- will reveal the trait underlying the set of re-
nation and rbi is the biserial correlation be- sponses.
tween item and unidimensional 0 (Lord & However, there are a multitude of possible
Novick, 1968). For the same model, these factor analytic solutions in a set of data. Fac-
authors have also specified the relationship tor analysis investigations of job satisfaction
between a, and the factor loading of item i are usually based on the common factor
on the first principal factor (a,) from the model and use factor rotation techniques
matrix of tetrachoric correlations. Again, as- based on some mathematical criterion of
suming unidimensional 0, this relation is Thurstone's (1947) principles of simple
structure. Such rotations explicitly empha-
(3) size the differences between sets of items that
define each factor.
This relation clearly shows that when the An alternate approach is to examine a fac-
assumptions of the normal ogive model hold tor that is common to all items. The first
(including unidimensionality), the relation principal factor is an example of such a factor
between the IRT model and the factor ana- but suffers from lack of psychological inter-
lytic model is simple. Unfortunately, most pretability when multiple factors exist in the
data sets do not meet such an assumption. data. Humphreys (1962) has proposed the
use of hierarchical factor models that de-
scribe observed variables as a function of a
Unidimensional IRT Models in general factor and multiple, more specific,
Multidimensional Data common factors. The specific method will
Applications of the IRT model in real data be described in the Methods section of this
require understanding the consequences of paper.
violating the model's assumptions. In partic- Regarding general factors, Humphreys
ular, what does it mean to apply this unidi- (1962) and Humphreys and Hulin (Note 2)
mensional model to multidimensional data? have noted that the most striking character-
Reckase (1979) has presented some results istic of published matrices displaying inter-
that compare such an application of a three- correlations among very diverse measures of
parameter logistic IRT model in data sets of cognitive ability, based on large samples and
varying factorial complexity. As more or- reliable measures, are the sizes of the smallest
thogonal factors of equal size are present in correlations. They are typically positive and
the item set, the relation denoted in Equation suggest the presence of a general factor of
3 becomes weaker. On the other hand, the intelligence that is obscured by most com-
estimated a parameters do relate highly to mon factor analytic procedures. Similar ob-
the loading of one of the orthogonally rotated servations can be made concerning factor
factors. This is quite interesting, because the analytic studies of job satisfaction. Positive
convergence of a model that assumes uni- correlations among the commonly reported
dimensionality with one factor of a model factorially derived scale scores are the most
that assumes multidimensionality tells us obvious and general outcome of any such
what component is being measured by the study. Based on these arguments, the general
former. Based on his results, Reckase also factor, as well as the more specific common
suggested that stable parameter estimates for factors should be considered when compar-
the IRT model require a first principal com- ing factor analytic results to IRT results.
ponent that accounts for at least 20% of the As described in the opening paragraph, the
test variance. advantages of IRT to many measurement
ITEM RESPONSE THEORY IN JOB SATISFACTION DATA 829

problems has been thoroughly described else- the work itself, pay, promotional opportunities, super-
where. Strict adherence to the assumptions visor, and co-workers. Four scales are included in this
study (the pay scale was omitted). Also, three adjectives
of the model in the available computational from the co-workers scale ("easy to make enemies," "no
procedures would appear to make applica- privacy," and "hard to meet") were not included because
tions to many research settings impossible. they were not common to several other samples that
Longer scales are certainly one alternative, were originally under study by Parsons (1980). Therefore
60 items from the JDI were used to index job satisfac-
but they are not practical in many organi- tion.
zational contexts where research time and Only subject records with no missing data on the 60
questionnaire space are extremely precious JDI items were included. Although Lord (1974) has pre-
commodities. Application of the model and sented an acceptable solution for estimating both item
procedures to a multidimensional item set and 6 parameters for aptitude tests with omitted re-
sponses, it is based on assumptions that clearly are not
is useful to the extent that we obtain useful tenable for responses to the JDI. For instance. Lord
results, understand the trait being assessed by (1974. p. 250) stated the assumption that "examinees
the total item set, and are aware that we are wish to maximize their expected scores and that they
violating an assumption of the model. are fully informed about their best strategy for doing
this." After eliminating records with omitted responses,
With this caveat, the current study at- the sample consisted of 1,349 response records.
tempted to shed light on what trait is being
assessed by the unidimensional IRT model
in multidimensional job satisfaction item re- Parameter Estimation
sponse data. Specifically, two questions were Parameters for the IRT model were estimated from
asked: the maximum likelihood algorithm, LOGIST (Wood,
Wingersky, & Lord, Note 3). LOGIST requires dichoto-
1. Is there evidence to support the pres- mous scoring of items, with 1 indicating satisfaction and
ence of the general factor? 0 indicating no satisfaction. The responses scored 0 or
2. Does an IRT procedure that assumes 1 by the Smith et al. (1969) procedure were scored as
unidimensionality provide parameter esti- 0. Responses that would have received a 3 were scored
mates that converge with the general factor as 1. The justification for this adjustment comes from
Smith et al.'s results demonstrating that question-mark
or one of the more commonly cited facet fac- responses (scored 1) were more frequently given by in-
tors in the Job Descriptive Index? dividuals with low satisfaction.

Method Factor Extraction and Rotation


Subjects Previous factor analysis studies of the JDI have been
based on Pearson product-moment correlations and
Data consisted of responses to questionnaires by non- have not considered that the observed variables are bet-
managenal personnel performing a variety of jobs in a ter described as discrete than continuous. Tetrachoric
large international merchandising company. Usable correlations are another index of association sometimes
questionnaires were received from 1,632 employees dis- used in test development and can be used if two observed
tributed among 41 units around the United States. Re- variables, such as items, are scored dichotomously.
turn rates for questionnaires were approximately 95%. Then, if it is reasonable to assume that the trait under-
The surveys were administered by organizational staff lying each item response is continuous, "it is possible
members. Participation was voluntary and participants to secure an estimate of what the correlation could be
completed the surveys on company time. They were if the underlying traits were continuous and normally
collected and mailed to university researchers. Confi- distributed or if they were so measured as to give normal
dentiality of responses was assured although identifying distributions" (McNemar, 1969, p. 221). Because Lord
information was requested in order to complete a follow- and Novick (1968) have shown the relation between fac-
up study not relevant to the present one. In this sample, tor analysis of tetrachoric correlations and the normal
59% of the respondents considered themselves full-time ogive item response model, tetrachorics were used in this
workers and the other 41% were part time. Thirty per- study. This latter requirement excluded the use of con-
cent of the sample was male, the average age was 36.5 firmatory factor analysis (such as Joreskog's, 1970, max-
years, and average tenure was 6.6 years. Further descrip- imum likelihood procedure) to test the fit of the hier-
tion of the subjects, questionnaire, and results can be archical factor model (Lord. 1980, p. 21).
found in Miller (1979). All factoring is based on the 60-item tetrachoric cor-
relation matrix with the maximum correlation of each
Variables and Selection of Data item with the other items in the diagonal as communal-
lty estimates (referred to as the reduced matrix). The
The Job Descriptive Index (Smith et al.. 1969) was principal axes method (Harman. 1967) was used to ex-
used as the measure of job satisfaction. The JDI is a tract principal factors.
series of adjective checklists that assess satisfaction with The direct oblimin routine (Jennnch & Sampson,
830 CHARLES K. PARSONS AND CHARLES L. HULIN

Table 1 Table 3
First 12 Eigenvalues From Matrix of Schmid-Leiman Transformation Matrix
Tetrachoric Correlations
Factor h, VT-hi VT-H
Factor Eigenvalue
1 .518 .855 0 0 0
1 16.76 2 .507 0 .502 0 0
2 6.00 3 .425 0 0 .905 0
3 5.31 4 .640 0 0 0 .768
4 3.60
5 2.32 Note, h, is the loading of the ith first order factor on the
6 1.93 second order principal factor.
7 1.81
g 1.48 ment suggested by Reckase (1979) to obtain
9 1.31
10 1.16 stable parameter estimation results. Four fac-
11 1.07 tors were retained in order to represent the
12 .97 four satisfaction facets on which the scales
are based. An eigenvalue-greater-than-one
1966) was used to rotate the factors obliquely. The cor- criterion (Kaiser, 1961) is inappropriate for
relations among the oblique factors were then used as the tetrachoric correlation matrix. Therefore,
a basis for second-order factoring. One principal factor four factors were rotated obliquely; and their
was extracted from these correlations by using an iter-
ative extraction procedure that first uses squared mul- intercorrelations appear in Table 2. The sec-
tiple correlations as communality estimates, extracts the ond-order factor was extracted from these
desired number of factors, reestimates communality correlations and is defined by loadings rang-
from the factor loadings, refactofs the new reduced cor- ing from .425 to .641. The eigenvalue of the
relation matrix, and so on until the difference between first principal component (1.82) was sub-
two successive communality estimates becomes negli-
gible. The loadings of the first-order factors on the sec- stantially larger than that of the second com-
ond-order factor were used to construct a matrix for ponent (.84), thus justifying the choice of the
transforming the four oblique first-order factors into a model that contains one general factor. The
hierarchical configuration with one general factor and Schmid-Leiman (1957) transformation ma-
four orthogonal facet factors. This matrix was con- trix is presented in Table 3. The values in the
structed in the following manner. The loadings of the
four first-order factors on the one second-order factor first column (the loadings on the second-or-
(h,) are the first column of the transformation matrix. der factor) are smaller than those in the other
Theremainingfour columns represent a diagonal matrix four columns, indicating that the loadings of
with v l ~ / j f as the diagonal entries. This matrix was the items on the general factor will be smaller
then premultiplied by the factor pattern matrix from the
direct oblimin rotation. (The reader is directed to than the loadings on the facet factors.
Schmid & Leiman. 1957, for the specific procedures and The oblique factor pattern matrix was
rationale.) transformed to the hierarchical factor struc-
ture and appears in Table 4 along with the
Results loadings on the first principal factor and the
a values from LOGIST.
The eigenvalues of the first 12 factors from The factor loadings on all factors were
the 60-item tetrachoric matrix appear in Ta- transformed according to Equation 3 and
ble 1. First, note that the first eigenvalue is compared to the LOGIST a using correlations
16.76 which, in terms of proportion of total that appear in Table 5. First observe that the
variance (27.8%), exceeds the 20% require- correlation between d and the principal fac-
tor loadings (PF) is .966. This is very con-
Table2 sistent with the theoretical relation specified
Correlations of Oblique First-Order Factors by the normal ogive model. As mentioned
earlier, the first principal factor lacks psy-
Factor 1 2 3 4
chological interpretability when multiple fac-
1 1.00 tors exist in this data. Therefore, the general
2 .30 1.00 factor (G) and facet factors (Fl to F4) are
3 .20 .19 1.00 examined next. The correlation for G is .853,
4 .31 .31 .31 1.00 which is clearly higher than for any of the
ITEM RESPONSE THEORY IN JOB SATISFACTION DATA 831

Table 4
LOGJST a, Principal Factor Loadings, and Hierarchical Factor Matrix for 60 Job
Descriptive Index Items
Item a, PF G Fl F2 F3 F4
1 .517 .484 .459 -.041 -.027 .100 .546
2 .342 .381 .371 -.045 -.020 .024 .478
3 .850 .625 .574 .036 .065 -.001 .617
4 .677 .530 .509 .035 .043 -.032 .573
5 .853 .614 .529 .108 .160 .029 .427
6 .432 .427 .407 -.065 -.025 .124 .483
7 .768 .594 .510 .100 .097 .117 .404
8 .179 .215 .148 .167 .132 -.024 -.023
9 .796 .609 .502 .198 .149 .057 .322
10 .602 .473 .428 .044 .106 -.049 .434
11 .503 .493 .428 .087 .122 .002 .362
12 .280 .300 .235 .072 .085 .154 .083
13 .651 .532 .508 -.010 -.068 .058 .632
14 .032 .048 .030 .078 .053 -.052 -.013
15 .321 .353 .271 .210 .057 .052 .103
16 .171 .204 .218 -.082 .026 -.089 .353
17 .244 .290 .229 .117 .060 .079 .102
18 .843 .613 .568 .017 .037 .020 .632
19 1.053 .576 .432 .075 -.068 .846 .035
20 .724 .472 .345 .029 -.012 .768 -.031
21 .907 .589 .447 .094 .040 .662 .067
22 .903 .599 .466 .080 .016 .633 .133
23 1.071 .584 .444 .046 -.045 .831 .062
24 .844 .597 .451 .149 .100 .515 .072
25 .657 .486 .351 .081 .056 .655 -.046
26 .873 .542 .403 .046 -.033 .749 .005
27 1.109 .615 .463 .079 .017 .785 .044
28 .532 .465 .365 .352 -.121 .109 .205
29 .943 .634 .445 .706 -.031 -.011 .049
30 .888 .584 .399 .726 -.047 -.038 .007
31 .787 .590 .423 .542 -.040 .141 .062
32 .743 .589 .405 .654 -.026 .062 .006
34 1.087 .679 .483 .596 .070 .112 .033
35 .616 .523 .357 .430 .151 .141 -.071
36 .559 .470 .291 .673 .009 .011 .140
37 .515 .470 .346 .386 -.056 .154 .088
38 1.073 .650 .446 .785 -.057 -.004 .007
39 .689 .539 .358 .699 -.035 -.015 -.045
40 .878 .605 .418 .575 .103 .068 -.027
41 1.209 .666 .455 .728 .044 .021 -.026
42 1.127 .682 .489 .519 .187 .074 .036
43 .020 .057 .047 .180 -.079 -.178 .082
44 .712 .566 .404 .531 .054 .024 .048
45 .885 .555 .381 .667 -.023 -.026 .003
46 .512 .494 .400 -.097 .497 .119 .133
47 .724 .563 .433 .050 .610 -.025 .067
48 .438 .430 .318 -.047 .711 -.054 -.055
49 .498 .479 .379 -.024 .509 .024 .100
50 .677 .513 .390 .019 .665 -.076 .028
51 .682 .544 .420 -.024 .668 .006 .046
52 .440 .444 .338 -.077 .607 .065 -.003
53 .817 .621 .478 .049 .627 .027 .081
54 .385 .389 .282 .035 .487 .040 -.054
55 .659 .570 .448 .008 .575 .028 .110
56 .539 .476 .352 .010 .693 -.066 -.037
57 580 .457 .339 .065 .585 -.070 -.014
58 .690 .558 .424 -.021 .666 .065 .018
59 559 .526 .407 -.012 .590 .043 .056
60 .559 .513 .386 .065 .542 .029 .017
Note. PF = first principal factor; G = general factor, Fl = supervision satisfaction facet factor, F2 = co-workers
satisfaction facet factor; F3 = promotions satisfaction facet factor; F4 = work satisfaction facet factor.
g32 CHARLES K. PARSONS AND CHARLES L. HULIN

Table 5 with this number of parameters. Another


Correlations O/LOGIST a With Transformed criticism could be based on using second-or-
Factor Loadings der factoring rather than simply leaving the
Factor a factor intercorrelations to be interpreted. In
this case, it can be argued that the higher
First principal factor .966 order factoring is both a more meaningful
Satisfaction and parsimonious solution because it in-
General .853 volves a reduction of the factor intercorre-
Supervision .374
Co-workers -.081
lation matrix. The Schmid-Leiman trans-
Promotion -.402 formation simply used this reduction to de-
Work .036 fine a general factor, to orthogonalize the
facet factors, and to define all factors in terms
of observed variables. The main advantage
facet factors. Correlations for the loadings on of the hierarchical factor solution is that it
the oblique facet factors would be identical illustrates that items on different scales do
to those for the facet factors in the hierar- share common variance and this common
chical structure. variance can be effectively summarized by
Clearly, the general factor is the psycho- one second-order factor. This latter point
logically interpretable factor that is most rep- would not hold if the oblique rotation had
resentative of 8 from the IRT analysis. The resulted in factor intercorrelations that were
correlations for the first principal factor also not described by a single factor. That is, a
demonstrate that the theoretical relation for single general factor is not a necessary out-
unidimensional items also holds very well for come from the hierarchical transformation
the multidimensional items of the JDI. Be- process.
cause these correlations only reflect the sim- Because of the clear results from the hi-
ilarity of patterns of loadings and a, the root erarchical solution, the interpretation of the
mean squared difference for these two item IRT analysis is clear. The item discrimina-
indexes were also computed and are .057 for tion parameters describe the relation of the
PF and. 167 for G. Finally, the means for the general factor, however, we choose to extract
various indexes are .528 for a, .507 for PF, or represent it, to the probability of endorsing
and .392 for G. As can be seen, there is a the items. On the other hand, from the per-
tendency for the a to be larger than what spective of IRT, the dimensions of the latent
would be expected from the transformed space represented by the four facet factors
loadings. are also related to the probability of item en-
dorsement. When and if satisfactory esti-
Discussion mation procedures become available, these
dimensions can be incorporated into a
The hierarchical factor matrix provides a multidimensional item response model for
pattern of factor loadings that is psycholog- the JDI.
ically interpretable and can be compared to The current findings are important be-
the IRT results. Regarding its interpretabil- cause they suggest that IRT can be applied
ity, the addition of the general factor to the as a reasonably well fitting measurement
already moderately well fitting four-orthog- model for the JDI and probably other ques-
onal-factor solution could be criticized for tionnaire measures of job satisfaction that
making the matrix less parsimonious (more can be scored validly in a dichotomous man-
parameters). The obvious response to the ner. This conclusion does not exclude the
objection is that the orthogonal facets have development of item sets that better fit the
little substance in empirical observation. In unidimensional IRT model or the experi-
addition, parsimony is not the sole or even mentation with the shorter scales of facet sat-
the overriding goal of science. If this were the isfaction to determine whether or not further
case, then four-principal factors (without ro- improvements can be made.
tation) would be desirable because they ac- At the same time, these data strongly sug-
count for the maximum possible variance gest the necessity for developing estimation
ITEM RESPONSE THEORY IN JOB SATISFACTION DATA 833

procedures for multidimensional item re- els are required to provide evidence about
sponse theories that will describe responses which specific factors in the work situation
to each item of both a general 0 and a facet should be changed. We can operate as re-
0. It must be emphasized that the appeal of searchers or practitioners with either model
a multidimensional model over one that em- depending on our aims without making as-
phasizes and uses only the general factor sumptions that we have learned much about
from the hierarchical solution will depend on specific causes of job satisfaction when we
the goals of the researcher and the uses to use a general factor approach or that we
which the resulting scales are to be put. If the know much about the antecedents of behav-
aim is the prediction of behavioral responses iors reflecting general acceptance/rejection of
reflecting general acceptance or rejection of a job when we use multidimensional models.
a work situation, such as turnover or absen- Perhaps most importantly, this study has
teeism, then the use of job satisfaction scores demonstrated the convergence of evidence
reflecting the general factor will probably from three quite different approaches to the
provide predictive power equal to that gen- study of the meaning of different item re-
erated by a multidimensional approach. On sponses on job satisfaction questionnaires.
this point, we could not expect a general fac- Convergence among measures based on the
tor to outperform a multiple regression com- first principal factor, on the general factor
posite of JDI scales in accounting for vari- from a hierarchical factor model, and from
ance in a specific criterion. However, the gen- a unidimensional latent trait model are en-
eral factor may account for more variance couraging. The results of this study provide
in a wider range of behavioral responses than some evidence for interpreting what is being
any one weighted composite of JDI scales. estimated by 0s from the JDI. Both the ne-
In addition, in small samples, the general fac- cessity and limitation of future developments
tor would not be subject to the sampling error stressing multidimensional IRT models in
that severely limits cross validation of least job satisfaction have been pointed out. Re-
squares regression equations. Humphreys finements of the model will generate more
and Hulin (Note 2) have commented on this research aimed at specifying the usefulness
same point in the domain of ability mea- of general and specific job satisfaction mea-
surement and job performance prediction. sures.
The fit of the IRT discrimination indexes,
derived assuming local independence and Reference Notes
unidimensionality of 0, to the loadings of the 1. Parsons, C. K. Empirically investigating the two pa-
items on the general factor from the hierar- rameter logistic model for a job satisfaction question-
chical factoring suggests minimal violence naire. Paper presented at the meeting of the Ameri-
may be done to our data by fitting it to a can Institute for Decision Sciences. Boston, Novem-
ber 1981.
general unidimensional model. So long as we 2. Humphreys, L. G., & Hulin. C. L. The construct of
are aware that assumptions are being made intelligence in the historical perspective of classical
in this approach that are not entirely correct, test theory. Paper presented at the Educational Test-
our informed violation of these assumptions ing Service Symposium on Construct Validity,
Princeton. N.J., October, 1979.
should not mislead us. 3. Wood, R. L., Wingersky, M. S., & Lord, F. M. A
However, if the aims of the researcher are computer program for estimating ability and item
more specific, such as to test specific hy- characteristic curve parameters (ETS RM 76-6).
potheses about attitudinal or affective cor- Princeton. N.J.: Educational Testing Service, 1976.
relates of specific behaviors—voting for union
representation in National Labor Relations References
Board elections, being absent on specific Birnbaum, A. Some latent trait models and their use in
days, or volunteering to work overtime— inferring an examinee's ability. In F. M. Lord &
then more complex multidimensional mod- M. R. No vick (Eds.), Statistical theories of mental test
scores. Reading. Mass.: Addison-Wesley, 1968.
els are required. Similarly, if the aims of an Dunham, R. B.. Smith, F. J., & Blackburn, R. S. Val-
investigator are interventions designed to in- idation of the Index of Organizational Reactions with
crease levels of job satisfaction in an orga- the JDI, MSQ, and Faces scales. Academy of Man-
nization, then again, multidimensional mod- agement Journal, 1977, 20, 420-432.
834 CHARLES K. PARSONS AND CHARLES L. HULIN

Golembiewski, R. T., & Yeager, S. Testing the appli- thesis, University of Illinois at Urbana-Champaign,
cability of the JDI to various demographic groupings. 1979.
Academy ofManagement Journal, 1978,21,514-519. Mulaik, S. A. The foundations offactor analysis. New
Hambleton, R. K., & Cook, L. Latent trait models and York: McGraw-Hill, 1972.
their use in the analysis of educational test data. Jour- Parsons, C. K. Measuring appropriateness in the assess-
nal of Educational Measurement, 1977, 14, 75-96. ment of job satisfaction (Doctoral dissertation. Uni-
Harman, H. H. Modem factor analysis (2nd ed.). Chi- versity of Illinois at Urbana-Champaign, 1980). Dis-
cago: University of Chicago Press. 1967. sertation Abstracts International, 1980, 4IB, 725B-
Hulin, C. L., Drasgow, F., & Parsons, C. K. Item re- 726B. (University Microfilms No. 8018196)
sponse theory and its application to measurement in Reckase. M. D. Unifactor latent trait models applied to
social science. Homewood, 111.: Dow Jones-Irwin, multifactor tests: Results and implications. Journal of
1982. Educational Statistics, 1979, 4, 207-230.
Humphreys, L. G. The organization of human abilities. Schmid, J., & Leiman. J. The development of hierar-
American Psychologist, 1962, 17, 475-483. chical factor solutions. Psvchomelrika, 1957, 22, 53-
Jennrich, R. I.. & Sampson, P. F. Rotation for simple 61.
loadings. Psychometrika, 1966, 31, 313-323. Schmitt, N. Rasch analysis of the Central Life Interest
Jbreskog, K. G. A general method for the analysis of measure. Applied Psychological Measurement, 1981.
covariance structures. Biometrika, 1970,5 7,239-251. 5. 3-10.
Kaiser, H. A note on Guttman's lower bound for the Smith, P. C, Kendall, L. M., & Hulin. C. L. The mea-
number of common factors. British Journal of Sta- surement of satisfaction in work and retirement. Chi-
tistical Psychology, 1961, 14, 1-2. cago: Rand McNally, 1969.
Lord, F. M. Estimation of latent ability and item pa- Smith, P. C, Smith, O. W.. & Rollo, J. Factor structure
rameters when there are omitted responses. Psycho- for blacks and whites of the Job Descriptive Index and
metrika, 1974. 39, 247-264. its discrimination of job satisfaction. Journal of Ap-
Lord, F. M. Practical applications of item characteristic plied Psychology, 1974, 59, 99-100.
curve theory. Journal of Educational Measurement, Thurstone, L. L. Multiplefactor analysis. Chicago: Uni-
1977, 14, 117-138. versity of Chicago Press. 1947.
Lord, F. M. Applications of item response theory to prac- Wright, B. D. Solving measurement problems with the
tical testing problems. Hillsdale, N.J.: Erlbaum, 1980. Rasch model. Journal of Educational Measurement.
Lord, F. M, & Novick, M. R. Statistical theories of 1977. 74,97-116.
mental test scores. Reading, Mass.: Addison-Wesley,
1968. Yeager, S. J. Dimensionality of the Job Descriptive In-
McNemar. Q. Psychological statistics. New York: Wiley, dex. Academy of Management Journal, 1981, 24,
1969. 205-212.
Miller, H. Social influences on work attitudes of part-
time and full-time employees. Unpublished master's Received March 5, 1982 •

Das könnte Ihnen auch gefallen