Professional Documents
Culture Documents
net/publication/51189947
CITATIONS READS
40 2,436
3 authors:
James Rounds
University of Illinois, Urbana-Champaign
139 PUBLICATIONS 6,657 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
ExpiWell: Experience Sampling Platform and App for Social Scientists View project
All content following this page was uploaded by James Rounds on 07 November 2014.
Louis Tay
Rong Su
James Rounds
Author Note
Abstract
Ideas are bipolar dimensions. Two minimal criteria for bipolarity are proposed and examined
across 3 studies: (a) the correlation between opposite interest types should be negative; (b) after
correcting for systematic responding, the correlation should be greater than -.40. In study 1, a
meta-analysis using 26 interest inventories with a sample size of 1,008,253 participants showed
that correlations between opposite RIASEC types ranged from 0.08 to 0.16 (corrected
correlations ranged from -0.08 to -0.16). In study 2, structural equation models (SEM) were fit to
the Interest Finder (IF; Wall, Wise, & Baker, 1996) and the Interest Profiler (IP; Rounds et al.,
1999) with sample sizes of 13,939 and 1,061, respectively. The correlations of opposite RIASEC
types were positive, ranging from 0.17 to 0.53. None of the corrected correlations met the
direct estimate of the correlation between Data-Ideas end poles using targeted factor rotation did
not reveal bipolarity. Further, bipolar SEMs fit substantially worse than a multiple factor
representation of vocational interests. In study 3, a two-way clustering solution on the IF and the
IP respondents and items revealed a substantial number of individuals with interests in both
People and Things. We discuss key theoretical, methodological, and practical implications such
as the structure of vocational interests, the interpretation and scoring of interest measures for
Holland‟s model (1959; 1997) underscores the structure of vocational interests, in which
six distinct vocational interest types -- Realistic, Investigative, Artistic, Social, Enterprising, and
proposition is that bipolar People-Things and Data-Ideas dimensions underlie the RIASEC
hexagon (Prediger, 1982; Prediger & Swaney, 2004). Compared to Holland‟s (1997) spatial
model that uses relative distances to explain the relations among the interest types, Prediger‟s
(1982) model conceptualizes the relations among interest types with two bipolar factors (see
review by Rounds, 1995). Bipolarity asserts mutual exclusion, and by implication, a strong
negative correlation between opposite interest types. According to Prediger (1982; see Figure 1),
Social interests are directly opposed to Realistic interests. Similarly, interest types near the Data-
Ideas poles are bipolar with the Enterprising interests opposed to Investigative interests; and
interest literature shows that reported correlations among RIASEC interest types are positive
(e.g., Mount, Barrick, Scullen, & Rounds, 2005). Because the bipolarity assumption of People-
Things and Data-Ideas undergirds key theoretical, methodological, and practical issues within
(1982) has been very influential among vocational researchers. These dimensions serve as the
theoretical basis for spherical models of interest in which a third dimension – Prestige or Sex-
Type – was proposed in addition to People-Things and Data-Ideas (Deng, Armstrong, & Rounds,
2007; also see Tracey & Rounds, 1996). Furthermore, bipolarity flavors the discussion of
Bipolar Dimensions 3
between vocational interests and personality types, Goh and Leong (1993) interpreted the
Realistic-type as a tendency to “move away from people”, “asocial, not caring for people,
lacking of feeling, and lacking interpersonal skills,” posing them as the opposite of Social-type
individuals who “have the benefits and demands of rich interpersonal lives.”
Researchers have also used the bipolar People-Things and Data-Ideas dimensions to
organize and understand individual differences. Career aspirations and goals (Hirschi, 2010;
Junk & Armstrong, 2010), along with work relevant abilities (Tracey & Hopkins, 2001), have
been conceptualized using the bipolar nomenclature. Several researchers have linked a gender-
dimension and suggested a parallel between them for understanding occupational interests (e.g.,
Lippa, 2005). Similarly, the bipolar People-Things dimension has been used for explaining sex
Aside from its theoretical impact, the bipolarity assumption also underpins the
development of interest measures and scoring of vocational interest scales in practice. The
Jackson Vocational Interest Survey (Jackson, 1977, 2000) was design to include the People-
Things dimension. Recent research explicitly uses bipolar People-Things and Data-Ideas rating
scales for occupations (Primavera et al., 2010). Further, because it is assumed that the RIASEC
space is fundamentally bipolar, there are scoring procedures for counter weighting opposite
interest types (see p.54, ACT Interest Inventory Technical Manual; ACT, 2009).
Traditionally, evidence for bipolarity comes primarily from three methods: targeted
factor extraction (Overall, 1962), multidimensional scaling (MDS), and principle component
analysis (PCA) solution with a general interest factor. Prediger (1982, 2004) utilized targeted
underlying Holland‟s hexagon. To elaborate, targeted factor extraction ensures that the two
dimensions uncovered are positioned in a manner consistent to Figure 1. The primary criterion
used to assess the adequacy of this bipolar model was the amount of variance accounted for by
the targeted factors. As an indicator of absolute fit, it has been shown that the targeted factors
relative fit, the proportion of variance accounted for by the two targeted factors compared to the
two untargeted factors is generally higher than .95. Although these criteria show how well two
factors account for variability in the data, they do not provide direct evidence for bipolarity. To
do so, it is necessary to evaluate whether the dimensional poles are indexed by the relevant
interest types in an expected manner. For instance, bipolarity can be demonstrated if the
indicators of People and Things have high positive and negative loadings on the People-Things
dimension, respectively. However, evidence for high bipolar loadings has yet to be presented.
provides insufficient evidence to claim bipolarity. This is because MDS uses information about
the relative size of the intercorrelations between RIASEC interest scales. Two correlation
matrices based on different theoretical structures – bivariate and bipolar – as illustrated in Table
Importantly, a bivariate pattern – wherein People and Things are uncorrelated – can still produce
an ostensible bipolar dimension in the MDS space. Therefore, it is not appropriate to infer
Bipolar Dimensions 5
bipolarity simply because vocational interest types lie in opposition in the MDS space. As noted
dimensions with opposite interests lying on positive and negative poles for any MDS spatial
Another method for showing bipolarity (Rounds & Tracey, 1993) has been through the
use of targeted factor extraction where the first dimension is taken as a general interest factor and
the subsequent dimensions as People-Things and Data-Ideas. This dimensional structure was
considered because when a general interest factor was modeled in Prediger‟s (1982) seminal
work, it was found to account for a substantial amount of variance (around 40 percent), even
more than the People-Things and Data-Ideas dimensions combined (about 35 percent). The
substantive interpretation of this factor has been unclear and has been suggested as (a) systematic
error or response bias (Jackson, 1977; Prediger, 1982), or (b) a general interest level (Rounds &
Tracey, 1993). Through the use of structural equation modeling (SEM), this three-factor
representation has been shown to fit relatively better than a two-factor structure and was claimed
to be adequate (although the model fit values NFI = .79; TLI = .73 do not reach current standards
for reasonable fit) across a dataset compiled from 10 representative RIASEC measures (Rounds
& Tracey, 1993). Nevertheless, a factor representation alone is insufficient evidence for
comes from an application of a latent variable measurement model that enables bipolar scoring.
Tay, Drasgow, Rounds, and Williams (2009) found that many individuals were placed in the
middle of the continuum because of their dual interests in People and Things. The same trend
was found with Data-Ideas as many individuals had both Artistic and Conventional interests.
Bipolar Dimensions 6
This pattern suggests that People-Things and Data-Ideas may not be bipolar, but could be better
Lippa‟s (2005) data that the claims for a bipolar masculinity-femininity dimension in
occupational interest surveys was overstated and the strong negative correlations between
masculine and feminine interests were shown to be a statistical artifact, casting doubt on the
In view of the counterevidence for bipolarity, along with the insufficiency of inferring
bipolarity using previous methods (i.e., variance accounted for by targeted factor extraction and
MDS), we examined bipolarity using more innovative and appropriate methods, with specific
Bipolarity asserts that an endorsement of one interest type is incompatible with the other.
Nevertheless, a strict pattern of bipolar responses is usually not observed in empirical data;
rather, higher values on one preference tend to be tied to a lower preference for another. To
quantify whether bipolarity holds across individuals, the zero-order correlation is computed.
Theoretically, a large negative correlation close to -1.00 would indicate strong evidence for
bipolarity, whereas a small negative correlation provides weak evidence for bipolarity.
vocational interest structure. In the study of emotions, there has been considerable debate on
whether positive and negative emotions constitute bipolar opposites or whether they are bivariate
dimensions. Advocates for the independence of positive and negative emotions present
nonsignificant to low negative correlations between the two (Bradburn, 1969; Diener &
Emmons, 1984; Warr, Barter, & Brownbridge, 1983). On the other hand, low negative
Bipolar Dimensions 7
correlations may occur because specific measurement factors (e.g., item selection, measurement
error, systematic error) are not accounted for (Green, Goldman, & Salovey, 1993; Russell, 1979).
Importantly, there is generally an inverse relationship between positive and negative emotions,
and arguments surround why larger negative zero-order correlations close to -1.00 are not found
(Russell & Carroll, 1999). Minimally, a small negative zero-order correlation is an indicator that
there exists some degree of bipolarity. Similarly, a negative correlation, even a small one,
between opposite interest types can be one criterion for ascertaining bipolarity.
systematic error (e.g., general interest responding). Individuals who have a strong general
interest tend to endorse both People and Things, or both Data and Ideas, whereas individuals
with more specific interests tend to reject many interest items. Thus, general interest responding
may interfere with observing a clear pattern of bipolar responses and it is important to correct for
this. Guideposts for evaluating the degree of bipolarity in correlations after correcting for
systematic error can be obtained from previous studies on emotions. Foremost, a value of around
-0.84 was obtained when SEM was applied to semantic opposites happy and sad (Green, et al.,
1993). Using a broader set of semantically bipolar terms, partial correlations in the range of -0.70
to -0.77 were obtained but were described to be only “approximately” bipolar, rather than strictly
bipolar (Bentler, 1969). Finally, Tellegen, Watson, and Clark (1999) reported positive and
negative emotions yielded a corrected correlation of -0.43 that was judged to be moderately
measurement error, the types of scales, and the scale formats used has led to a proposal that a
value of -.467 can be indicative of bipolarity (see Russell & Carroll, 1999).
Bipolar Dimensions 8
From these studies, we conclude that a corrected correlation more negative than -.40
would provide preliminary evidence for bipolar dimensions, whereas any value less negative
than -.40 would be evidence against bipolarity. In our view, a value of -.40 does not necessarily
warrant the conclusion of bipolarity, but should be a minimal requirement to claim that bipolarity
is plausible. Thus, we use the rule-of-thumb of -0.40 as a second minimal criterion. This value is
also in the ballpark between the recommended values for medium to large effect sizes (Cohen,
1992). If bipolarity exists, a moderate to large effect size should be observed. To summarize, two
conditions need to be fulfilled in order for bipolarity in the People-Things and Data-Ideas
dimensions to be plausible: (a) there must be a negative zero-order correlation between the two
poles; and (b) the corrected zero-order correlation between two poles should be larger than -
0.40.1
does not directly convey information on the proportion of individuals who express their interests
in a bipolar fashion. Examining this information is important because we can ascertain the
individuals exhibit bipolar responding? Or does bipolarity describe only a small subpopulation of
individuals? In doing so, we can uncover the generalizability of bipolarity across individuals. To
achieve this, we used graphical clustering methods to examine whether a majority of individual
responses follow a pattern of mutual exclusivity. Further, graphical clustering evaluates the
usefulness of the bipolarity concept from a person-centered approach rather than a variable-
Slightly different methodology was used to evaluate the bipolarity in People-Things and
Data-Ideas because the People-Things dimension correspond directly to two RIASEC types (i.e.,
Social and Realistic interests) whereas the Data-Ideas dimension do not have corresponding
interest types at the end poles. In Study 1, we used meta-analysis to examine whether the
correlations from a diverse set of interest measures are negative – a necessary condition for
bipolarity. For People-Things, this can be directly examined through the Social-Realistic
correlation; for Data-Ideas, this is indirectly examined with the Investigative-Enterprising and
account for the general factor and then reexamine the meta-analytic correlations. Residual
In Study 2, we used two different large samples of individuals who responded to interest
inventories. The response data allow us to estimate the opposite RIASEC type correlations.
Negative correlations would indicate a possible bipolar relationship. With primary data, we can
estimate the Data-Ideas correlation by conducting targeted factor extraction (Overall, 1962)
following Prediger‟s procedures (1982; Prediger & Swaney, 2004). We also examined the
corrected correlations to determine if the values are larger than -0.40. By controlling for the
general interest factor and measurement error through the use of SEM (cf. Green, et al., 1993),
from the two different samples. Because People-Things directly correspond to Social and
Realistic interests, we only examine the extent to which responses are bipolar on the People-
Things dimension.
Study 1
Bipolar Dimensions 10
Meta-analysis
Technical manuals of vocational interest inventories constituted the data source for this
analysis. We used technical manuals because manuals generally include results from large
samples that are representative of different ages and ethnic groups. Hedges and Nowell (1995)
recommended the use of large well-sampled data sets as an alternative to traditional meta-
analytic samples. Additionally, these technical manuals represent the empirical foundation of the
interest measures that are used most frequently in applied settings when working with
individuals who are making career-related decisions. Therefore, in addition to providing insight
into the relationship between opposite interest types, meta-analytic results from interest
inventory manuals also have important practical implications for career guidance and interest
inventory development.
Multiple literature search strategies were used to identify available interest inventory
manuals, following the procedures laid out in Su, Rounds, and Armstrong (2009). We went
through the most recent edition of the Mental Measurements Yearbook (Buros, 1938, 1941,
1949, 1953, 1959, 1965, 1972, 1978; Conoley & Impara, 1995; Conoley & Kramer, 1989;
Geisinger, Spies, Carlson, & Plake, 2007; Impara & Plake, 1998; Kramer & Conoley, 1992;
Mitchell, 1985; Plake & Impara, 2001; Plake, Impara, & Spies, 2003; Spies & Plake, 2005) to
identify interest inventories published through 2007. We then searched the publishers‟ websites
or contacted the publishers directly to determine whether any new edition of a test had been
published or any new norm sample had been used since 2007. The third author identified interest
publishers, and the archival collection of the Buros Institute of Mental Measurements. The
Inclusion Criteria
Inventories were included in the current meta-analysis if they satisfy the following
criteria: First, the inventories were published in English with norm samples from the United
States or combined norm samples from both the United States and Canada. Second, the
inventories were intended to measure vocational interests and were constructed under Holland‟s
(1959, 1997) RIASEC model or have scales corresponding to the RIASEC interest types. Third,
correlation matrices for the RIASEC scales were reported in the technical manuals. Different
editions of an inventory were counted as separate studies if item changes had occurred between
between 1965 and 2009, and a total of 49 samples consisting of 1,008,253 participants.
Analytical Procedures
scales was calculated according to the random-effects model of error estimation. The analysis
was performed in SPSS 16.0 using the modules given by Lipsey and Wilson (2001, pp. 208–
220). The aim was to examine whether opposite RIASEC types generally have negative
correlations.
the general interest factor (factor loadings are freely estimated), was fit to each of the correlation
matrices in Table 2. To ascertain the degree to which opposite RIASEC types correlate after
partitioning out the general interest factor, we examined the residual correlations. A residual
Bipolar Dimensions 12
correlation greater than -0.40 would indicate that the bipolar assumption between opposite
To further illustrate the correlation pattern for the RIASEC scales, we included
correlation matrices from a number of interest inventories held in high regard by the professional
testing community of test developers, teachers, researchers, and practitioners. To identify these
interest inventories, we perused professional test and measurement books and selected four
inventories that were most endorsed based on a broad set of criteria, such as rigor in inventory
The Strong Interest Inventory (SII; Donnay, Morris, Schaubhut, & Thompson, 2005) is
one of the earliest interest inventories developed by E. K. Strong, Jr. in 1927 and has been
widely used ever since. In 1974, RIASEC scales called the General Occupational Theme (GOT)
scales were added to the SII profile. The internal consistencies for the GOT scales range from
.90-.95. The Self-Directed Search (SDS; Holland, Fritzsche, & Powell, 1994) was developed by
Holland as a self-administered, self-scored, and self-interpreted interest assessment and has had
since its inception a wide impact on career decision-making and counseling. The internal
consistency coefficients for its RIASEC scales range from .90-.94. The Career Assessment
Inventory – Vocational Version (CAI-V; Johansson, 1984) closely resembles the Strong Interest
Inventory but was developed to target those careers requiring no postsecondary training up
through four years of college. In spite of its relatively brief history, CAI-V has been widely
endorsed among practitioners and has received positive feedback from the testing community.
The reliabilities for its RIASEC scales are also around .90s. Lastly, the Unisex Edition of ACT
Bipolar Dimensions 13
Interest Inventory (UNIACT; ACT, 1995) is an interest assessment administered along with the
ACT test to tens of thousands of Grade 8, Grade 10, and Grade 12 students every year. Aside
from its wide practical impact, it is known for showing relatively small gender different at both
item and scale level. The reliabilities for its RIASEC scales range from .87-.92. All four
inventories have excellent professional reputations and have received abundant support for their
construct and criterion-related validity (see Savickas & Spokane, 1999; Walsh & Betz, 2000 for
reviews). These inventories are widely used in applied settings and have had a large impact on
the career choices of individuals, and therefore, provide important information to corroborate the
meta-analytic results.
Results
Meta-analytic correlations. A weighted overall correlation matrix was obtained from all
49 correlation matrices with a sample size of 1,008,253, as shown in Table 2. Table 2 also shows
correlation matrices from the four highly regarded interest inventories. Although past empirical
studies on the structure of vocational interests have generally supported very similar correlation
pattern of males and females and the equivalence of their RIASEC interest structure (e.g., ACT,
2009; Darcy & Tracey, 2007; Holland et al., 1997; Rounds & Day, 1999), we also examined the
meta-analytic correlations by gender when separate correlation matrices were available. Bolded
values denote the correlations between opposite interest types: Realistic-Social (Things-People),
As shown in Table 2, the overall meta-analytic correlations were all positive, contrary to
the bipolar assumption: Realistic and Social interests (i.e., Things and People) were correlated at
interests were .16 and 08, respectively. When the overall meta-analytic correlation was broken
Bipolar Dimensions 14
down by gender, only females exhibited a small negative correlation of -.03 on Artistic-
Conventional; all the other opposite RIASEC types had positive correlations ranging from .08 to
.18.
Examining the correlations by interest inventories and gender, we found that the largest
inverse correlation was -0.05 in the CAI-V (2005) for females. Nevertheless, the trend of
correlations depicts a picture of co-endorsement of opposite RIASEC interests rather than mutual
exclusion. Among the correlations for opposite interests from the four highly regarded interest
inventories, only three were negative (ranging from -.02 to -.05) whereas the other 21 were
positive (ranging from .04 to .39).These positive correlations cast doubt on the assumption that
literature, there has been concern that a general interest factor (Jackson, 1977) or acquiescence
would attenuate the negative correlation (Bentler, 1969). In view of this, we used SEM to control
for the general interest factor in the meta-analytic correlations; specifically, we examined the
residual meta-analytic correlations between all three pairs of opposite interests after partialling
Table 3 shows the meta-analytic correlations after correcting for the general interest
factor. The overall residual correlations for People versus Things was only slightly negative (r =
-.08). Similarly, there did not seem to be strong evidence for the bipolarity of Data-Ideas given
were -.13 and -.18, respectively. By gender, the overall residual correlations ranged from -.06 to
-.23 and averaged -.15. These coefficients were substantially smaller than the minimal criterion
of -.40.
Bipolar Dimensions 15
Analysis by interest inventories and gender showed that the largest negative correlation
between People and Things was found for the Strong Interest Inventory (2005) for female
respondents (r = -0.12); yet, in several interest inventories, the corrected correlation between
People and Things was zero. Similarly, there was little evidence for bipolarity of the Data-Ideas
dimension. The largest negative corrected correlation for Investigative-Enterprising was -.23 for
males in the Strong Interest Inventory (2005). For Artistic-Conventional the largest negative
value was -.24 for females in the CAI-V (1984). Therefore, correcting the meta-analytic
correlations for the general interest factor did not reveal bipolarity in the People-Things and
Data-Ideas dimension.
Summary. Meta-analytic results did not meet the two necessary criteria for bipolarity.
Positive correlations were found for all opposite RIASEC types in the overall meta-analytic
correlations. When a correction was applied to these positive correlations, only small negative
Study 2
Primary Studies
vocational interest inventories. In this study, we used data from two primary studies to specify
more intricate structural models that allow us to control for measurement error and the general
interest factor at the level of the individual. Following the two bipolarity criteria, we first
estimate the correlations from opposite RIASEC types to determine if the values are negative.
Then, we control for the general interest factor to determine the extent to which corrected
Conventional) provides only indirect evidence for bipolarity. With primary data, we can use
Prediger‟s (1982) original analytic procedure – targeted factor extraction – to obtain the People-
Things and Data-Ideas dimension. If end poles are bipolar, Realistic and Social interests are
expected to have low and high loadings on the People-Things dimension, respectively. In the
same way, Artistic and Conventional interests, along with Investigative and Enterprising
interests, should have low and high loadings on the Data-Ideas dimension, respectively. With the
factor loadings, we can directly estimate the correlations between the end poles.
Interest Finder. The Interest Finder (IF; Wall, Wise, & Baker, 1996) is a measure of
vocational interests designed for the Armed Services Vocational Aptitude Battery (ASVAB)
Career Exploration Program. Responses to the IF were obtained as part of the Freedom of
Information Act (FOIA) from the Defense Manpower Data Center (DMDC). No demographic
data was provided and we were unable to fit models to different subpopulations (e.g., gender).
The IF was administered to a sample of 14,297 individuals in 1997 as part of the IF data norming
based on national probability sampling so that the sample conforms to the US Census data. In
our analysis, listwise deletion was implemented: participants with less than 3 percent missing
responses were used, resulting in a sample size of 13,939. Respondents indicated „dislike‟ („0‟)
or „like‟ („1‟) to 40 interest statements per RIASEC type. Each 40 item scale is further divided
into three areas: 12 items assess training, 14 assess activities, and 14 assess occupations. We
present the means, standard deviations, reliabilities, and the correlations of the RIASEC types in
Table 4. Importantly, there are no floor or ceiling effects as evident from the mean levels, and all
Bipolar Dimensions 17
of the RIASEC scales exhibit excellent reliabilities (see Wall & Baker, 1997 for validity of the
IF).
O*NET Interest Profiler. The O*NET Interest Profiler (IP; Rounds, Smith, Hubert,
Lewis, & Rivkin, 1999) is a vocational interest measure used as part of the U.S. Department of
Labor O*NET career exploration program. The IP consists of 180 items, with 30 items per
interest type. All responses were dichotomized so that „not sure‟ and „dislike‟ were coded as „0‟
and „like‟ was coded as „1.‟ This data set was collected to evaluate the reliability and validity of
the IP (Rounds, Walker et al., 1999) with 1,061 individuals (437 males and 634 females) from
four regions across the United States: Michigan, New York, North Carolina and Utah. Data
collection sites included employment service offices, high schools, junior colleges, technical-
trade schools, universities and government agencies. Participants were heterogeneous in terms of
ethnicity, age, and education. The sample was 25% African American, 1.5% Asian or Pacific
Islanders, 10% Hispanic, 2.5% Native American, 59% White non-Hispanics, and 1.6% members
of other racial/ethnic groups. The mean age of the participants was 32.62 (SD = 12.05) and
ranged from younger than 18 years old to above 50 years old. Educational background of the
participants was 38.2% high school degree, 36.4% with some college including a college degree,
20.4% less than high school degree, and 4.1% education beyond college. Table 4 presents the
means, standard deviations, reliabilities and correlations of the IP RIASEC types from this
sample.
Analysis
All analyses in this study were conducted using the statistical software Mplus 5.2
(Muthén & Muthén, 2007). Maximum likelihood (ML) estimation was used for all the models
Bipolar Dimensions 18
with the exception of the O*NET Interest Profiler data in which weighted least squares estimator
Interest Finder. A series of factor models were employed to examine the correlations
between the opposite RIASEC types so that the bipolarity of People-Things and Data-Ideas
could be assessed. The factor indicators for each of the RIASEC types were based on the mean
scores of activities, training, and occupations. Using a multiple factor model, we examined the
factor correlations without controlling for the general interest factor as illustrated on the left side
of Figure 2. Thus, measurement error is taken into account, but not the general interest factor.
Next, we fitted a bifactor model, and estimated the factor correlations of opposite RIASEC types
controlling for the general interest factor as shown on the right side of Figure 2. Importantly, the
bifactor model allowed us to examine – controlling for both the general interest factor and
measurement error – whether the factor correlations of interests are higher than -.40. In this
model, we only estimated three factor correlations because these correlations are of theoretical
interest. Further, unlike the multiple-factor model where all factor correlations are estimated,
Because the Data-Ideas dimension does not lie on specific RIASEC types, an
examination of the opposite RIASEC types does not directly address whether Data-Ideas is
bipolar. We used the targeted factor extraction procedure to obtain two uncorrelated factors. The
People-Things and Data-Ideas dimensions are extracted using on a target loading matrix based
on the Cartesian coordinates of the RIASEC types on Holland‟s hexagon. In this manner, we can
estimate a correlation between the end poles for Data-Ideas along with People-Things. Negative
Targeted factor extraction does not account for general interest factor. Following the
procedures by Rounds and Tracey (1993), we estimated a model in which we also extracted the
general interest factor through the use of a target vector of ones, in addition to targeted factor
extraction of the People-Things and Data-Ideas dimensions. In doing so, we can determine
whether the corrected correlations reach the minimal -.40 criteria for bipolarity.
O*NET Interest Profiler. The analysis was conducted using the same stepwise procedure
used with the IF. Factor analysis was conducted at the item-level because there were no
subscales across the RIASEC interest types. To maximize the possible negative correlation
between the two factors, factor analysis was initially conducted for each RIASEC type with
itemsthat had high factor loadings (> .8). Item selection was also necessary for practical purposes
because a multiple-factor and bifactor model with a total number of 180 dichotomous indicators
is difficult to estimate. A smaller set of indicators representative of each RIASEC factor was
Model-data fit. To determine whether the specified model fits the data, it has been
recommended that a range of model fit indices be utilized and that no single index be relied upon
exclusively (Hu & Bentler, 1999; Martens, 2005). Model-data fit was evaluated with three
categories of indices which include: (a) comparative fit indices, which compare the hypothesized
model to the independence model, including the Tucker-Lewis index (TLI; also known as the
nonnormed fit index) and the comparative fit index (CFI) (Bentler, 1990; Bentler & Bonett,
1980). The closer they are to 1.00, the better the fit. It has been recommended that values around
0.95 or higher show good model fit (Hu & Bentler, 1999); (b) the root mean squared error of
approximation (RMSEA) was also examined. A RMSEA value of zero indicates that the model
fits the data exactly. The criteria proposed for RMSEA are values smaller than 0.06 (Hu &
Bipolar Dimensions 20
Bentler, 1999), indicating reasonable model-data fit; (c) the standardized root mean square
residual (SRMR) is a measure of the standardized difference between the observed and model-
implied covariance. Similar to the RMSEA, a small value indicates a good model-data fit.
Results
In the following, we provide the necessary information for evaluating bipolarity and the
fit of the SEMs. The full set of SEM results may be obtained from the first author.
Interest Finder. Table 5 summarizes the fit statistics for the models and the key
correlations of interest. The model fit for the multiple-factor model was reasonable (χ2(120)=
8892.24, CFI = .96, TLI = .94, RMSEA = .072, SRMR = .039). All the factor loadings in the
model were large and ranged from 0.78 to 0.90 (M = 0.86, SD = 0.03), indicating that the scales
were representative of the RIASEC factors. As shown in Table 5, the People and Things
correlation was 0.35. The correlation for Data-Ideas was tested indirectly with Investigative-
Enterprising and Artistic-Conventional, which were 0.53 and 0.43, respectively. The positive
values were expected as Study 1 showed the overall observed meta-analytic correlation was
positive rather than negative. In addition, the zero-order correlations in Table 4 were positive as
well. Therefore, none of the results meet the first criterion for bipolarity.
The bifactor model yielded a reasonable model fit, comparable to that of the multiple-
factor model (χ2(114)= 8013.38, CFI = .96, TLI = .95, RMSEA = .071, SRMR = .045). After the
general interest factor is taken into account the RIASEC factor loadings ranged from 0.28 to 0.75
(M = 0.57, SD = 0.12), which was substantially smaller than that of the multiple-factor model.
The factor loadings on the general interest factor were slightly larger and ranged from 0.43 to
0.77 (M = 0.64, SD = 0.09). This pattern of factor loadings indicates that individuals‟ general
interest levels contribute to endorsements on the RIASEC scales. By modeling the general
Bipolar Dimensions 21
interest factor, we statistically control for this effect. Nevertheless, the opposite RIASEC
correlations did not reach the second necessary criterion of -0.40 for bipolarity (People-Things =
In Table 5, we also present the results from targeted factor extraction of People-Things
and Data-Ideas. The total variance accounted for by the People-Things and Data-Ideas factors
was 53.8 percent, close to the average of 60 percent found in previous studies by Prediger (1982;
Prediger & Swaney, 2004). Furthermore, the Fit Index typically used to evaluate model fit was
1.00, indicating that the targeted factor rotation accounted for the same amount of variance as the
freely estimated two-factor solution. On the People-Things dimension, the average loadings for
People and Things were 0.23 (SD = .06) and -0.67 (SD = .04), respectively. In this case, because
the indicators lie directly on the rotated dimension, a correlation between the end poles can be
computed by multiplying the average loadings 0.23 × -0.67 = -0.15. Surprisingly, for the Data-
Ideas dimension, the loadings for the opposite poles were both positive. As depicted in Figure 1,
Conventional and Enterprising types on positive end of the dimension had average loadings of
0.66 (SD = .01) and 0.71 (SD = .03), respectively; Investigative and Artistic types on the
negative end of the dimension had average loadings of 0.63 (SD = .03) and 0.67 (SD = .01),
Conventional and Enterprising types index Data and the loadings averaged across Investigative
and Artistic types index Ideas, the correlation between the end poles would be ([0.66 + 0.71]/2) ×
([0.63 + 0.67]/2) = 0.45. In sum, the People-Things dimension meets the first criterion of
When the general interest factor, People-Things, and Data-Ideas was estimated, the total
variance accounted for was 61.6 percent and the Fit Index of 1.00 indicated perfect recovery of
Bipolar Dimensions 22
the variance accounted for. For the People-Things dimension, the average loadings for People
and Things were 0.31 (SD = .06) and -0.09 (SD = .02), respectively. The correlation between the
end poles is estimated to be -0.03. For the Data-Ideas dimension, the average loadings for
Conventional and Enterprising were 0.60 (SD = .05) and 0.02 (SD = .11); the average loadings
for Artistic and Investigative were -0.68 (SD = .09) and 0.17 (SD = .02). The estimated
correlation between the end poles for Data-Ideas is -0.08. In both dimensions, after accounting
for the general interest factor, neither dimension meets the second criteria of a negative
O*NET Interest Profiler. Initial exploratory factor analyses and selection of indicators
that are representative of each RIASEC type (factor loadings > 0.8) yielded a total of 64
Conventional. The reliabilities of this set of RIASEC scales ranged from 0.81 to 0.95 and these
indicators were used for the present analyses. As shown in Table 5, the multiple-factor model fit
was reasonable (χ2(313)= 1449.70, CFI = .92, TLI = .98, RMSEA = .059); although the CFI was
slightly lower than .95, the other indicators of fit were good. The RIASEC factor loadings ranged
from 0.72 to 0.94 (M = 0.85, SD = 0.06). This shows that the indicators index the underlying
RIASEC factors well. The factor correlations between the opposite RIASEC types were all
Examining gender groups also revealed that all the values were positive. These results showed
that neither the IF nor the IP meets the first criterion for bipolarity.
Fitting a bifactor model yielded a reasonable model fit (χ2(197)= 1113.03, CFI = .94, TLI =
.97, RMSEA = .066), which was comparable to the multiple-factor model. The RIASEC factor
loadings ranged from 0.31 to 0.90 (M = 0.70, SD = 0.12) whereas the general interest factor
Bipolar Dimensions 23
loadings range from 0.10 to 0.78 (M = 0.45, SD = 0.13). After controlling for this general interest
factor, the factor correlation was 0.00 for People and Things; males had a moderate positive
correlation of 0.24 and females had a correlation close to zero. For People-Things, no correlation
value exceeded a value of -0.40. For Data-Ideas, the Investigative-Enterprising correlation was -
0.67 (Males = -0.58; Females = -0.90), which was substantially higher than -0.40; whereas for
Artistic-Conventional, the correlation was -0.16 (Males = -0.20; Females = -0.16). There appears
to be indirect evidence that the Data-Ideas dimension possibly shows bipolarity because the
second bipolarity criterion is met. However, there was no evidence that the Investigative-
Targeted factor extraction results are shown in Table 5. The total variance accounted for
by this model was 53.2 percent. This value was similar to that found in the IF and within the
range found by Prediger (1982). The Fit Index was 1.00, indicating excellent fit. The average
loadings for the Social and Realistic interests on the People-Things dimension were 0.23 (SD =
.04) and 0.76 (SD = .06), respectively; thus, the correlation between the end poles are positive
0.23× 0.76 = 0.17 (Males = 0.02; Females = 0.32). The average loadings for Conventional and
Enterprising were 0.79 (SD = .06) and 0.38 (SD = .07), respectively; the average loadings for
Investigative and Artistic were -0.23 (SD = .07) and -0.19 (SD = .04), respectively. Using the
same formula for the IF, the estimated correlation between the end poles for Data-Ideas was -
0.12 (Males = 0.09; Females = -0.12). This analysis shows that People-Things did not meet the
Does the Data-Ideas dimension pass the second criterion for bipolarity? To examine this,
we used targeted factor extraction with a general factor. The total variance accounted for by this
Bipolar Dimensions 24
model was 62.0 percent, and the Fit Index was 1.00. We found that the average loadings for
Conventional and Enterprising were 0.46 (SD = .07) and 0.12 (SD = .07), respectively; the
average loadings for Investigative and Artistic were -0.44 (SD = .07) and -0.54 (SD = .04),
respectively. Therefore, the estimated correlation between the Data-Ideas end poles was -0.14
(Males = -0.13; Females = -0.13), which did not exceed the second criterion of -0.40. For the
People-Things dimension, the average loadings for People and Things were 0.46 (SD = .09) and
0.50 (SD = .04), respectively; therefore, the correlation between the opposite poles was 0.23
(Males = 0.20; Females = 0.02). Neither dimension reached the -0.40 criterion for bipolarity.
bipolarity of People-Things and Data-Ideas dimensions based on the targeted factor extraction.
We evaluated whether these bipolar dimensions sufficiently describe responses to the RIASEC
interests using SEM indexes of model fit. As displayed in Table 5, for the initial models where
only the People-Things and Data-Ideas dimensions were extracted, the SEM fit statistics was far
from good for both the Interest Finder (χ2(118)= 61695.64, CFI = .69, TLI = .59, RMSEA = .193,
SRMR = .087) and the Interest Profiler (χ2(215)= 7146.80, CFI = .54, TLI = .78, RMSEA = .174).
When a general factor was extracted in addition to the two factors, the fit statistics improved, but
still did not yield good model fit: Interest Finder (χ2(102)= 40000.16, CFI = .80, TLI = .70,
RMSEA = .168, SRMR = .062) and the Interest Profiler (χ2(254)= 5617.69, CFI = .64, TLI = .86,
RMSEA = .141). This replicates previous research by Rounds and Tracey (1993) where the two-
factor solution (NFI = .73, TLI =.73) was poorer than the three-factor solution (NFI = .79, TLI =
.80); nevertheless, both models do not fit the data well according to current standards for
assessing model fit. Overall, compared to the multiple factor model and the bifactor model, the
Aside from the SEM indexes, the variance accounted for was larger in the multiple factor
and bifactor models (M = 73.9 percent) as compared to that in the bipolar dimensional models
(M = 57.7). In addition, even though the Fit Index utilized by previous researchers (e.g.,
Prediger, 1982; Rounds & Tracey, 1993) evaluating the fit of the bipolar models indicated a
perfect fit; the SEM model fit indexes are clearly more sensitive to data misfit.
Summary. We did not find evidence to support the bipolarity assumption in either the IF
or the IP. The correlations in the People-Things and Data-Ideas dimensions were positive. After
controlling for systematic responding, the corrected correlations did not exceed -0.40 with the
were inadequate representations of the RIASEC data. Models that use a separate factor for each
RIASEC type yielded better fit on every model-fit index used; these models are clearly superior
Study 3
Cluster analysis
The zero-order correlation between People and Things quantifies the degree of bipolarity
but does not provide a detailed account of how individuals respond to vocational interest items.
Are there groups of individuals who have interests in both People and Things? If so, what
proportions do these groups account for? We apply a two-way clustering method -- known as
cluster heat maps – to visualize item responses to both the People and Things interests in the IP
and the IF. Cluster heat maps have had a long history in data visualization (Wilkinson &
Friendly, 2009) and are currently popular in biochemistry for organizing large volumes of
genotypic and phenotypic data in order to detect important trends (Weinstein, 2008). Because the
Bipolar Dimensions 26
Data-Ideas dimensions does not correspond to specific RIASEC types, this analysis was not
Cluster heat maps rely on the logic of hierarchical cluster analysis -- objects (e.g.,
cluster analysis has been utilized and advocated in counseling and vocational psychology for
grouping individuals and vocational interest items (Borgen & Barnett, 1987; Wolfe, 1978).
Cluster heat maps go beyond hierarchical cluster analysis by sorting both individuals and
vocational interest items. This method succinctly clusters both individuals and items, resulting in
groups of individuals who share a similar response pattern (e.g., bipolar responding or co-
endorsements) and groups of items which are most similar. In this manner, we can infer whether
a majority of individuals exhibit mutual exclusion in their responses to People and Things, or if
Method
Analysis. We apply clustering heat map using the package “gplots” with the statistical
software R version 2.92 (R Development Team, 2008). The heat map is produced from a
seriation algorithm; the rows and columns of the data matrix are permutated until a loss function
is minimized. This effectively produces clusters of individuals and clusters of items which are
most similar.
Interest Finder. Because the cluster heat map could not be easily computed and presented
for a large dataset of about 13,000 individuals, a random sample of 3000 individuals from the IF
was submitted for clustering (a replication based on another random sample of 3000 individuals
O*NET Interest Profiler. All 1061 individuals were used to generate the heat map.
Bipolar Dimensions 27
Results
Interest Finder. As shown in Figure 3, the vocational interest items were clustered into
Social and Realistic interests, on the left and right of the diagram, respectively. This is also
depicted in the vertical dendograms. The shaded regions in the heat map represent item
the horizontal dendograms and response patterns revealed four broad response patterns.
Individuals in Clusters 1 and 3 (approximately 50% of the sample) were interested in both
People and Things, although Cluster 1 generally had higher interest in Things and Cluster 3 had
higher interest in People. By contrast, Clusters 2 and 4 had interests that appeared more mutually
exclusive.
O*NET Interest Profiler. As shown in Figure 4, items asking about Social and Realistic
interests were clearly differentiated on the right and left, respectively. For this measure of
vocational interest, there were 5 broad groups of respondents. Approximately 20% of the sample,
as revealed in Cluster 3, had interests in both People and Things. Clusters 1, 2 and 5
(approximately 45%) had more bipolar interests, wherein the endorsement of Social items was
mutually exclusive to that of Realistic items. However, Cluster 1 had a substantial number of
individuals who had some interests in Things, though they were primarily interested in People.
Finally, there was a large group of individuals who had scattered interests in both People and
Summary. Overall, the clustering results revealed that interests in People and Things were
not necessarily exclusive, as bipolarity would suggest. There were individuals who endorsed
primarily Social interests or Realistic interests, but there were also a large number of individuals
Discussion
The current results from vocational interest inventories present clear evidence that
interests in People and Things, and interests in Data and Ideas, are not bipolar. Meta-analytic and
the two primary studies showed low to moderately positive correlations between the opposite
interest types. Even after controlling for general interest responding, the inverse correlations
were not large and almost all values did not reach -0.40. Across all the studies, the People-
Things and Data-Ideas dimensions did not meet both minimal criteria to claim bipolarity.
Furthermore, we found that models positing bipolar dimensions do not meet any of the current
SEM criteria for reasonable model-fit. Finally, the examination of vocational interest responding
with two-way clustering revealed a substantial proportion of individuals who were interested in
both People and Things, or neither People or Things. This leads us to conclude that there is
Theoretical contribution
bivariate perspective that allows, for example, interests in both People and Things, as well as
interests in only People or Things, or neither (cf. Cacioppo & Bernston, 1994). A bivariate
conception encompasses bipolarity but also allows for a range of other patterns of responding
found in our cluster analyses. By contrast, a bipolar view confounds individuals who are
indifferent (lack of interest) or ambivalent (strong dual interests) so that the middle of the bipolar
continuum is highly ambiguous (e.g., Tay et al., 2009). In effect, the bivariate view is more
More broadly, the structural results revealed that a bipolar conception of the RIASEC
space may not be suitable. Overall, the model fit indices revealed that a bipolar conception of the
Bipolar Dimensions 29
RIASEC interest space yielded worse model fit as compared to separate-factor models
representing each RIASEC type (i.e., multiple factor and the bifactor models). The present study
presents evidence that a multiple factor structure for vocational interest types may be more
appropriate (see Ashton & Lee, 2008); that is, each vocational interest type is represented by a
single factor, as with the Big-Five personality dimensions (McCrae & Costa, 1987).
Methodological contribution
the past (e.g., Prediger, 1982; Prediger & Swaney, 2004) – and MDS (Day, Rounds, & Swaney,
1998) for establishing bipolarity. With targeted factor extraction, the extracted dimensions do not
necessarily exhibit a strong negative correlation (or even a negative correlation) between the
opposite poles. Further, the variance accounted for and the Fit Index were less useful for
discriminating whether the model fits the data. Although the variance accounted for was similar
to what was obtained in the past – around 60 percent – and the Fit Index was perfect, the SEM
indices for assessing model-fit showed that targeted factor extraction does not provide adequate
fit across all the fit indices. Nevertheless, recent vocational interest studies continue to use
variance accounted for as a criterion for fit (e.g., Darcy, 2005). We propose that researchers
directly estimate the correlation between end poles. In addition, it is important to use SEM
model-fit indices to compare bipolar models with other competitive models (e.g., a multiple
factor model).
Traditionally, multidimensional scaling (MDS) has been the standard protocol for
investigating the structure of vocational interests (Rounds, Tracey, & Hubert, 1992; Rounds &
Zevon, 1983); a two-dimensional hexagonal (circular) structure of vocational interests has often
been revealed and researchers have interpreted the vocational interest types that lie in opposition
Bipolar Dimensions 30
in the MDS space as bipolar (Prediger, 1982; Prediger & Swaney, 2004). However, this
conclusion is premature as MDS only considers the relative size of the correlations and not the
absolute size. Therefore, descriptive dimensions (e.g., People-Things, Data-Ideas) applied to the
We have proposed two minimal criteria for bipolarity. First, it is important that there is
evidence for a negative correlation between two variables. Second, based on past bipolarity
research and effect sizes, a value of -0.40 is the minimal criterion that should be exhibited after
one corrects for general interest responding. We hope that future research examining this issue
can consider the use of these two simple rules-of-thumbs. Importantly, positive claims for
bipolarity would require more stringent criteria. A -0.80 correlation may be a reasonable value to
claim bipolarity.
We also present the use of cluster heat maps to show how raw data can be sorted and
organized to describe how subgroups of individuals respond to different sets of items. Although
important to consider the use of a person-centered approach (Meehl, 1992) where variability in
where variables are assumed to be the key interest, a person-centered approach focuses on the
of individuals (e.g., Hagenaars & McCutcheon, 2002; Tay, Diener, Drasgow & Vermunt, 2011;
Tay, Newman & Vermunt, 2011). We propose the use of person-centered approaches like cluster
heat maps and latent class modeling. In our analysis, we found distinct patterns in responding. It
is important to identify these subgroups and understand their differences to develop tailored
Practical contribution
The results from our study suggest the necessity of establishing a new approach for
opposite RIASEC types, it would be beneficial to have a more encompassing view of vocational
interests. It is important when discussing the RIASEC hexagonal model with clients that career
counselors present these interests as relatively independent types that are interrelated in a circular
fashion. The further apart two interest types are on the hexagon, the less strongly related they
are; interest types opposing each other on the hexagon are essentially independent rather than
“opposite.” It is not uncommon for clients to have both Realistic and Social, or Investigative and
Enterprising, or Artistic and Conventional interests. For example, many people in managerial
interests. Not assuming bipolarity will also broaden career opportunities for individuals.
There are implications for scoring vocational interest scales. Because individuals can be
highly interested in both People and Things or both Data and Ideas, counter-weighting opposite
interests in measuring vocational interests may be problematic. One issue would be that
individuals who are strongly interested in both People and Things and individuals who are
uninterested in neither will both be located in the middle of the RIASEC interest space.
Similarly, given the concerns about bipolar representation of the RIASEC space, mapping
individuals onto certain regions of the RIASEC space based on their high point codes may also
be questionable. For example, consider two individuals whose high point codes are RSI and SRI,
respectively. They would be placed on opposite regions on the RIASEC space even though their
researchers to appropriately distinguish individuals with high and low profiles and more
faithfully represent interest profiles. We suggest that alternative techniques be developed for
more accurate assessment of individual interests in career counseling. One direction to pursue
would be a factor approach to vocational interests, such as the use of basic interest scales (BISs;
Liao, Armstrong, & Rounds, 2008). BISs were developed as groups of unidimensional work
activities that share properties such as context, setting, object, or process that transcend specific
occupations (e.g., Day & Rounds, 1997). Compared to the RIASEC interest types, BISs reflect
more specific vocational preferences and provide finer connection to people‟s career choices (Su
& Rounds, 2010). Donnay and Borgen (1996) found the BISs to be the most valid predictors of
occupational group membership, and concluded that “basic interest scales more effectively deal
with the reality of a complex multivariate space.” The present results support a wider use of
occupations, another practical implication concerns the ratings and classification of occupations.
When RIASEC ratings of occupations are conducted, raters are typically trained to give ratings
according to Holland‟s hexagon, with the assumption that opposite RIASEC interests do not co-
Realistic interests, the presence of a Social or Enterprising component is often necessary. For
example, car mechanics (a stereotypical Realistic job) who lack people interests tend to service
cars, but car mechanics who are interested in socializing with coworkers and interacting with
customers tend to move into a supervisory position which apparently has a people component. It
is similar for other pairs of opposite RIASEC interest types. The issue of rating and classifying
Bipolar Dimensions 33
occupation is important because where we map occupations in the RIASEC space is directly
linked to career choices suggested to individuals in career counseling and impacts the validity of
interest measures.
Finally, the dimension of People-Things has received recent attention because it is tied to
like sciences, technology, engineering, and mathematics (STEM). Our results suggest that the
debate surrounding the gender-related interests in People or Things cannot be easily subsumed
into a single bipolar interest continuum. Instead, more accurate assessments would be derived by
assessing interests in People and Things separately. How do men and women differ in their
absolute and relative interests? Based on a bivariate perspective, one analytic strategy that can be
implemented in future studies is to use the absolute strength of interests, and the relative strength
of interests calculated as a difference between interests in People and Things. The separate
examination of People and Things interests on the basis of a bivariate view will provide a more
complete picture of gender differences in vocational interests and a more accurate assessment of
the issue of gender discrepancy in STEM, because it can identify occupational fields that
integrate the technical and interpersonal aspects and that attract both men and women. In
practice, many individuals (particularly females) with desirable abilities and skills for STEM shy
away from these fields because of their stereotypical “asocial” nature. Career counselors may
provide a closer examination of the STEM occupations based on a bivariate view for individuals
with high Social interests and point out the possibilities where their Social interests can be met
The present study used major interest inventories and heterogeneous samples to provide
evidence against the bipolarity assumption. It is important to extend the findings to other large
representative samples from highly regarded interest inventories. Additional research using
person-centered methodology, such as latent class modeling and cluster heat maps, needs to be
conducted on the SII (Donnay et al., 2005), the UNIACT (ACT, 2005), and the SDS (Holland et
al., 1997). Future research needs to replicate the results using the intended audience for the
inventories, for example, college students for the SII (Donnay et al., 2005), high school students
(1982). Based on a reviewer suggestion, one direction would be to examine the co-occurrence of
opposite RIASEC codes in the high-point coding of occupations. One issue that needs to be
considered and controlled for in this research is that ratings of occupations (e.g., O*NET) are
based upon Holland‟s theoretical model and it has generally been assumed that opposing
RIASEC codes tend not to co-occur. Therefore, in the development of occupational interest
profiles, raters are trained to assign high-point RIASEC codes that generally do not conflict with
this assumption (e.g., Rounds et al., 1999). A possible approach may be to implement an
used vocational interest instruments and found little evidence based on the correlations and the
pattern of responses. More generally, a bipolar dimensional model of the RIASEC space had a
substantially poorer fit than a multiple factor or a bifactor model of RIASEC types, indicating
Bipolar Dimensions 35
that interest profiles of individuals do not necessarily conform to a bipolar pattern expected from
opposite RIASEC types. These findings lead us to propose a bivariate perspective to interests in
People and Things, and similarly, Data and Ideas. We hope that future research can use these
References
*American College Testing Program. (1981). Technical report for the unisex edition of the ACT
*American College Testing Program. (1995). Technical manual: Revised Unisex Edition of the
*American College Testing Program. (2009). ACT Interest Inventory technical manual. Iowa
http://www.act.org/research/researchers/techmanuals.html
Ashton, M. C., & Lee, K. (2008). Gender-related occupational interests do not define a
10.1027/1614-0001.29.1.25
Bentler, P. M. (1969). Semantic space is (approximately) bipolar. The Journal of Psychology, 71,
33-40.
Bentler, P. M. (1990). Comparative fit indexes in structural models. Psychological Bulletin, 107,
238-246.
Bentler, P. M., & Bonett, D. G. (1980). Significance tests and goodness of fit in the analysis of
Borgen, F. H., & Barnett, D. C. (1987). Applying cluster analysis in counseling psychology
Buros, O. K. (Ed.). (1938). The nineteen thirty eight mental measurements yearbook. New
Buros, O. K. (Ed.). (1941). The nineteen forty mental measurements yearbook. Highland Park,
Buros, O. K. (Ed.). (1949). The third mental measurements yearbook. New Brunswick, NJ:
Buros, O. K. (Ed.). (1953). The fourth mental measurements yearbook. Highland Park, NJ:
Gryphon Press.
Buros, O. K. (Ed.). (1959). The fifth mental measurements yearbook. Highland Park, NJ:
Gryphon Press.
Buros, O. K. (Ed.). (1965). The sixth mental measurements yearbook. Highland Park, NJ:
Gryphon Press.
Buros, O. K. (Ed.). (1972). The seventh mental measurements yearbook. Highland Park, NJ:
Gryphon Press.
Buros, O. K. (Ed.). (1978). The eighth mental measurements yearbook. Highland Park, NJ:
Gryphon Press.
Cacioppo, J. T., & Bernston, G. G. (1994). Relationship between attitudes and evaluative space:
A critical review, with emphasis on the separability of positive and negative substrates.
*Campbell, D. P. (1974). Manual for the Strong–Campbell Interest Inventory: T325 (Merged
Campbell, D. P., Borgen, F. H., Eastes, S. H., Johannson, C. B., & Peterson, R. A. (1968). A set
of Basic Interest Scales for Strong Vocational Interest Blank for men. Journal of Applied
*Campbell, D. P., & Hansen, J. C. (1981). Manual for the Strong–Campbell Interest Inventory.
Cohen, J. (1992). A power primer. Psychological Bulletin, 112, 155-159. doi: 10.1037/0033-
2909.112.1.155
Conoley, J. C., & Impara, J. C. (1995). The twelfth mental measurements yearbook. Lincoln, NE:
Conoley, J. C., & Kramer, J. J. (1989). The tenth mental measurements yearbook. Lincoln, NE:
Darcy, M. U. A. (2005). Examination of the structure of Irish students' vocational interests and
10.1016/j.jvb.2004.08.007
Day, S. X, & Rounds, J. (1997). "A little more than kin, and less than kind": Basic interests in
vocational research and career counseling. The Career Guidance Quarterly, 45, 207-220.
Day, S. X, Rounds, J., & Swaney, K. (1998). The structure of vocational interests for diverse
Deng, C.-P., Armstrong, P. I., & Rounds, J. (2007). The fit of Holland's RIASEC model to US
Diener, E., & Emmons, R. A. (1984). The independence of positive and negative affect. Journal
Donnay, D. A. C., & Borgen, F. H. (1996). Validity, structure, and content of the 1994 Strong
*Donnay, D. A. C., Morris, M. L., Schaubhut, N. A., & Thompson, R. C. (2005). Strong Interest
Feist, G. J. (2006). How development and personality influence scientific thought, interest, and
2680.10.2.163
Eccles, J. S. (1994). Understanding women‟s educational and occupational choices: Applying the
Fouad, N. A., Harmon, L. W., & Borgen, F. H. (1997). Structure of interests in employed male
and female members of U.S. racial-ethnic minority and nonminority groups. Journal of
Geisinger, K. F., Spies, R. A., Carlson, J. F., & Plake, B. S. (2007). The seventeenth mental
Green, D. P., Goldman, S. L., & Salovey, P. (1993). Measurement error masks bipolarity in
Hagenaars, J. A., & McCutcheon, A. L. (2002). Applied latent class analysis. Cambridge:
University Press.
Bipolar Dimensions 40
*Hansen, J. C., & Campbell, D. P. (1985). Manual for the Strong Interest Inventory: Form T325
of the Strong Vocational Interests Blanks (4th ed.). Stanford, CA: Stanford University
Press.
*Hanson, G. R. (1975). Assessing the career interests of college youth: Summary of research and
applications (ACT Research Report No. 67). Iowa City, IA: American College Testing
*Harmon, L. W., Hansen, J. C., Borgen, F. H., & Hammer, A. L. (1994). Strong Interest
Inventory applications and technical guide. Stanford, CA: Stanford University Press.
*Harrington, T. F., & O‟Shea, A. J. (1981). Manual for the Harrington–O’Shea Career
*Harrington, T. F., & O‟Shea, A. J. (2001). Manual for the Harrington–O’Shea Career
Decision-Making System (Rev. ed.). Circle Pines, MN: American Guidance Service.
Hedges, L. V., & Nowell, A. (1995). Sex differences in mental test scores, variability, and
Hirschi, A. (2010). Vocational interests and career goals: Development and relations to
doi: 10.1177/1069072710364789
symposium on motivation 1982. Personality: Current theory and research (pp. 55-89).
45).
Bipolar Dimensions 41
*Holland, J. L. (1972). Professional manual for the Self-Directed-Search. Palo Alto, CA:
*Holland, J. L. (1977). Manual for the Vocational Preference Inventory (7th rev.). Palo Alto,
*Holland, J. L. (1979). The Self-Directed Search: Profession manual. Palo Alto, CA: Consulting
Psychologists Press.
*Holland, J. L. (1985). The Self-Directed Search: Profession manual – 1985 edition. Odessa, FL:
*Holland, J. L., Fritzsche, B., & Powell, A. (1997). Self-Directed Search: Technical manual.
Holland, J. L. (1997). Making vocational choices: A theory of vocational personalities and work
Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis:
10.1080/10705519909540118
Impara, J. C., & Plake, B. S. (1998). The thirteenth mental measurements yearbook. Lincoln,
Jackson, D. N. (1977). Manual for the Jackson Vocational Interest Survey. Port Huron, MI:
*Johansson, C. B. (1984). Manual for Career Assessment Inventory (2nd ed.). Minneapolis, MN:
*Johansson, C. B. (2003). Career Assessment Inventory: The Enhanced Version manual (Rev.
Junk, K. E., & Armstrong, P. I. (2010). Stability of career aspirations: A longitudinal test of
10.1177/0894845309350921
Kramer, J. J., & Conoley, J. C. (1992). The eleventh mental measurements yearbook. Lincoln,
Lapan, R. T., McGrath, E., & Kaplan, D. (1990). Factor structure of the Basic Interests Scales by
Leuwerke, W. C., Robbins, S., Sawyer, R., & Hovland, M. (2004). Predicting engineering major
Liao, H. Y., Armstrong, P. I. A., & Rounds, J. (2008). Development and initial validation of
public domain basic interest markers [Monograph]. Journal of Vocational Behavior, 73,
6494.2005.00326.x
Lipsey, M. W., & Wilson, D. B. (2001). Practical meta-analysis. Thousand Oaks, CA: Sage.
McCrae, R. R., & Costa, J. P. T. (1987). Validation of the five-factor model of personality across
instruments and observers. Journal of Personality and Social Psychology, 52, 81-90.
Meehl, P. E. (1992). Factors and taxa, traits and types, differences of degree and differences in
Meece, J. L., Parsons, J. E., Kaczala, C. M., Goff, S. B., & Futterman, R. (1982). Sex differences
324–348.
Mitchell, J. V., Jr. (1985). The ninth mental measurements yearbook. Lincoln, NE: University of
Nebraska Press.
Mount, M. K., Barrick, M. R., Scullen, S. M., & Rounds, J. (2005). Higher-order dimensions of
the Big Five personality traits and the Big Six vocational interest types. Personnel
Muthén, B., & Muthén, L. K. (2007). Mplus version 5.2 [Computer Program]. Los Angeles:
Overall, J. E. (1962). Orthogonal factors and uncorrelated factor scores. Psychological Reports,
10, 651-662.
Plake, B. S., & Impara, J. C. (2001). The fourteenth mental measurements yearbook. Lincoln,
Plake, B. S., Impara, J. C., & Spies, R. A. (2003). The fifteenth mental measurements yearbook.
Prediger, D. J. (1982). Dimensions underlying Holland's hexagon: Missing link between interests
8791(82)90036-7
Prediger, D. J., & Swaney, K. B. (2004). Work task dimensions underlying the World of Work:
Research results for diverse occupational databases. Journal of Career Assessment, 12,
Primavera, M. T., Church, A. T., Katigbak, M. S., Bruna, L., White, J. R., & Peradilla, I. (2010).
R Development Team (2008). R: A language and environment for statistical computing. Vienna,
Dawis (Eds.), Assessing individual differences in human behavior (pp. 177-232). Palo
Rounds, J., & Day, S. X (1999). Describing, evaluating, and creating vocational interest
meaning, measurement and use in counseling (pp. 103-133). Palo Alto, CA: Davies-
Black.
Rounds, J., Smith, T., Hubert, L., Lewis, P., & Rivkin, D. (1999). Development of occupational
http://www.onetcenter.org/dl_files/OIP.pdf.
Rounds, J., & Tracey, T. J. (1993). Prediger's dimensional representation of Holland's RIASEC
Rounds, J., Tracey, T., & Hubert, L. (1992). Methods for evaluating vocational interest structural
8791(92)90073-9
Rounds, J., Walker, C. M., Day, S. X., Hubert, L., Lewis, P., & Rivkin, D. (1999). O*NET
http://www.onetcenter.org/dl_files/IP_RVS.pdf.
Bipolar Dimensions 45
Rounds, J., & Zevon, M. A. (1983). Multidimensional scaling research in vocational psychology.
Russell, J. A. (1979). Affective space is bipolar. Journal of Personality and Social Psychology,
37, 345-356.
Russell, J. A., & Carroll, J. M. (1999). On the bipolarity of positive and negative affect.
Savickas, M. L., & Spokane, A. R. (1999). Vocational interests: Meaning, measurement, and
Spies, R. A., & Plake, B. S. (2005). The sixteenth mental measurements yearbook. Lincoln, NE:
Su, R., & Rounds, J. (2010, April). Gender differences in Basic Interests: A meta-analysis. Paper
presented at the annual conference of the Society for Industrial & Organizational
Su, R., Rounds, J., & Armstrong, P. I. (2009). Men and things, women and people: A meta-
10.1037/a0017364
Strong, E. K., Jr. (1927). Vocational interest test. Stanford, Calif.: Stanford University Press,
1927.
Tay, L., Diener, E., Drasgow, F., & Vermunt, J. K. (2011). Multilevel mixed-measurement IRT
Tay, L., Drasgow, F., Rounds, J., & Williams, B. (2009). Fitting measurement models to
vocational interest data: Are dominance models ideal? Journal of Applied Psychology,
Tay, L., Newman, D. A., & Vermunt, J. K. (2011). Using mixed-measurement item response
10.1177/1094428110366037
Tellegen, A., Watson, D., & Clark, L. A. (1999). On the dimensional and hierarchical structure
Tracey, T. J. G., & Hopkins, N. (2001). Correspondence of interests and abilities with
10.1037//O022-0167.48.2.178
*U.S. Department of Defense. (2005). ASVAB Career Exploration Program: Counselor manual.
*U.S. Department of Labor. (2000). O*NET Interest Profiler user’s guide. Retrieved from
http://online.onetcenter.org
*Wall, J. L., Wise, L. L., & Baker, H. E. (1996). Development of the Interest-Finder: A new
Wall, J. L., & Baker, H. E. (1997). The interest-finder: Evidence for validity. Journal of Career
Walsh, W. B., & Betz, N. E. (2001). Tests and assessment (4th ed.). Upper Saddle River, NJ:
Prentice Hall.
Bipolar Dimensions 47
Warr, P. B., Barter, J., & Brownbridge, G. (1983). On the independence of positive and negative
10.1126/science.1151888
Wilkinson, L., & Friendly, M. (2009). The history of cluster heap map. The American
Footnotes
1
Correcting for systematic error may produce a spurious negative relationship when the
relationship, an analogy may be helpful. A positive correlation between two interest types in the
data can become negative when one removes individuals in the data who have dual interest types
or no interests, so that only individuals who respond in a bipolar fashion remain. Inducing a
negative correlation in this fashion leads to a moot claim for bipolarity. Of course, there would
be stronger evidence against bipolarity if correcting for positive correlations yields only a small
negative relationship.
Bipolar Dimensions 49
Table 1
Table 2
RIASEC correlation matrices from the meta-analytic review and highly regarded inventories
Note. R, I, A, S, E, C denote Realistic, Investigative, Artistic, Social, Enterprising and Conventional interests, respectively. Correlations between opposite RIASEC interest
types are bold and shaded. CAI-V = Career Assessment Inventory – Vocational Version; UNIACT-R = Unisex Edition of ACT Interest Inventory – Revised Edition.
Bipolar Dimensions 51
Table 3
Meta-analytic correlations and highly regarded interest inventories: Opposite RIASEC types
controlling for the general interest factor
Table 4
Interest Finder and Interest Profiler: Means, standard deviations, reliabilities and correlations
Interest Finder
M SD R I A S E C
R 0.37 0.24 (0.93)
I 0.45 0.27 0.51 (0.94)
A 0.42 0.25 0.36 0.48 (0.94)
S 0.47 0.26 0.31 0.51 0.61 (0.94)
E 0.49 0.28 0.35 0.48 0.54 0.63 (0.95)
C 0.38 0.31 0.36 0.41 0.39 0.56 0.61 (0.96)
Interest Profiler
M SD R I A S E C
R 0.30 0.26 (0.93)
I 0.44 0.30 0.31 (0.94)
A 0.44 0.30 0.17 0.45 (0.94)
S 0.52 0.31 0.17 0.38 0.38 (0.95)
E 0.37 0.27 0.36 0.30 0.49 0.45 (0.93)
C 0.44 0.33 0.10 0.14 0.17 0.33 0.50 (0.96)
Table 5
Comparison of fit indices and estimated correlations from structural equation models
Correlations
Fit People- Data-
I-E A-C
Model Description χ2 df CFI TLI RMSEA SRMR VAF Index Things Ideas
Multiple factor Model 8892.24 120 0.96 0.94 0.072 0.039 74.9 - 0.35 - 0.53 0.43
-0.02 -0.26
Interest Bifactor Model 8013.38 114 0.96 0.95 0.071 0.045 75.5 - -0.33 -
Finder People-Things & Data-Ideas 61695.64 118 0.69 0.59 0.193 0.087 53.8 1.00 -0.15 0.45 - -
General Factor & People-
40000.16 102 0.80 0.70 0.168 0.062 61.6 1.00 -0.03 -0.08 - -
Things & Data-Ideas
Multiple factor Model 1449.70 313 0.92 0.98 0.059 - 72.4 - 0.17 - 0.20 0.14
Males 564.23 209 0.93 0.97 0.062 - 71.0 - 0.34 - 0.22 0.27
Females 684.03 248 0.95 0.98 0.053 - 73.0 - 0.28 - 0.19 0.06
Bifactor Model 1113.03 197 0.94 0.97 0.066 - 72.9 - 0.00 - -0.67 -0.16
Males 511.96 160 0.93 0.96 0.071 - 71.7 - 0.24 - -0.58 -0.20
Females 606.23 160 0.95 0.97 0.067 - 73.5 - -0.01 - -0.90 -0.16
Interest
Profiler People-Things & Data-Ideas 7146.80 215 0.54 0.78 0.174 - 53.2 1.00 0.17 -0.12 - -
Males 1906.26 158 0.66 0.81 0.159 - 54.0 1.00 0.02 -0.09 - -
Females 3205.46 182 0.66 0.81 0.163 - 55.4 1.00 0.32 -0.12 - -
General Factor & People-
5617.69 254 0.64 0.86 0.141 - 62.0 1.00 0.23 -0.14 - -
Things & Data-Ideas
Males 1407.64 186 0.76 0.89 0.123 - 62.3 1.00 0.20 -0.13 - -
Females 2390.51 214 0.74 0.88 0.128 - 62.9 1.00 0.02 -0.13 - -
Note. VAF represents the variance accounted for computed using the averaged communalities; Fit Index denotes the fit of the bipolar models
relative to an unrestricted model with the same number of factors; I-E and A-C denotes Investigative-Enterprising and Artistic-Conventional,
respectively. For the same structural model, MPlus 5.21 does not necessarily compute the same degrees-of-freedom (df) when WLSMV
estimation is used.
Bipolar Dimensions 54
Figure 1. Holland‟s hexagonal model of vocational interests and Prediger‟s (1982) bipolar
dimensions
Bipolar Dimensions 55
R1
R1
R R2 R R2
R3 R3
I1 I1
I I2 I I2
I3 I3
A1 A1
A A2 A A2
A3 A3
General
S1 Responding
S1
S S2 S S2
S3 S3
E1 E1
E E2 E E2
E3 E3
C1 C1
C C2 C C2
C3 C3
Figure 3. Cluster heat map of 3000 randomly selected individuals from the Interest Finder
Note. Items are labeled so that the first letter indicates whether items are Realistic (R) or Social (S). The second letter indicates whether items
assess activities (A), training (T) or occupations (O); the numbers index the items. The horizontal dendograms reflect clustering of individuals
based on their responses to Realistic and Social items; the vertical dendograms reflect the clustering of items. Shaded regions in the heat map
represent item endorsements and unshaded regions represent non-endorsements.
Bipolar Dimensions 57
Appendix
Note. a Based on a listwise deleted sample. ACT-IV = ACT Interest Inventory; ASVAB = Armed Services Vocational Aptitude Battery; CAI-E = Career Assessment
Inventory – Enhanced Version; CAI-V = Career Assessment Inventory – Vocational Version; CDM = Harrington-O‟Shea Career Decision-Making System; IF = Interest
Finder; O*NET IP = O*NET Interest Profiler; SDS-E = Self Directed Search (Form Easy); SDS-R = Self Directed Search (Form R); SDS-CE = Self-Directed Search Career
Explorer; SII = Strong Interest Inventory; UNIACT = Unisex Edition of ACT Interest Inventory; UNIACT-R = Unisex Edition of ACT Interest Inventory – Revised Edition;
UNIACT-S = Unisex Edition of ACT Interest Inventory – Short Edition; VPI = Vocational Preference Inventory.