Journal of Consulting and Clinical Psychology In the public domain.

1983, Vol. 51, No. 5, 730-742

The Structure of Psychological Distress and

Well-Being in General Populations
Clairice T. Veit and John E. Ware, Jr.
The Rand Corporation, Santa Monica, California

We describe the development of the Mental Health Inventory (MHI), a new 38-
item measure of psychological distress and well-being, developed for use in general
populations. The MHI was fielded in four large samples having quite different
characteristics (N = 5,089). One data set was used to explore the MHI's factor
structure, and confirmatory factor analyses were used for cross-validation. Results
support a hierarchical factor model composed of a general underlying psychological
distress versus well-being factor; a higher order structure defined by two correlated
factors—Psychological Distress and Well-Being; and five correlated lower order
factors—Anxiety, Depression, Emotional Ties, General Positive Affect, and Loss
of Behavioral Emotional Control. Summated rating scales produced high internal-
consistency estimates and substantial stability over a 1-year interval. Our results
provide strong psychometric support for a hierarchical model and scoring options
ranging from five distinct constructs to reliance on one summary index. This trade-
off, which is between the unique information contained in the subscales versus the
simplicity of a single score, should be evaluated further.

A review of general population mental- on symptoms such as anxiety and depression

health-survey instruments pointed out im- that reflect the more prevalent kinds of psy-
portant trends in questionnaire content and chological distress in general populations
conceptual issues that deserve empirical at- (Ware et al., 1979). Including these measures
tention (Ware, Johnston, Davies-Avery, & should increase the sensitivity of a general
Brook, 1979). Early instruments were very population survey in detecting changes in
heterogeneous in content (Gurin, 1960; Lang- mental health. However, a substantial pro-
ner, 1962; Macmillan, 1957). They included portion of people in a general population rarely
measures of physical and psychosomatic or never report occurrences of even the most
symptoms, functional status, other health prevalent psychological distress symptoms. To
problems or worries, and health habits, in ad- increase measurement precision, it may be
dition to measures of more straightforward necessary to extend the definition of mental
psychological constructs (e.g., symptoms of health beyond the mere frequency or intensity
anxiety and depression). More recent instru- of psychological distress symptoms to include
ments seem to focus almost exclusively on the characteristics of psychological well-being (e.g.,
more straightforward psychological constructs feeling cheerful, interest in and enjoyment of
(Bradburn, 1969; Cleary, Goldberg, & Kessler, life). Psychological well-being items have the
1982; Dupuy, 1972). Whereas the more het- potential to improve the precision of mental
erogeneous measures may be satisfactory for health measurement by distinguishing among
testing hypotheses about health status in gen- persons who receive perfect scores on measures
eral, they do not seem to do well in distin- of psychological distress. Whereas inclusion of
guishing changes in mental health from such items is a feature of more recent instru-
changes in physical health (Ware et al., 1979; ments (Bradburn, 1969; Dupuy, 1972; Gold-
Ware, Brook, Davies-Avery, et al., 1980b). berg, 1978; Ware et al., 1979), whether they
Another characteristic of more recently de- actually improve the power of hypothesis test-
veloped mental health surveys is their focus ing remains to be determined.
In 1975 we began work on a new general
Requests for reprints should be sent to Clairice T. Veit, population measure, the Mental Health In-
The Rand Corporation, 1700 Main Street, Santa Monica, ventory (MHI). This standardized survey,
California 90406. which was designed to measure general psy-
730
 PSYCHOLOGICAL DISTRESS AND WELL-BEING 731

chological distress and well-being, is the pri- of other instruments containing both positive
mary mental health outcome measure fielded and negative items (Goldberg, Steele, Johnson,
in the Rand Health Insurance Experiment & Smith, 1982; RadlQff, 1977; Ware et al.,
(HIE) and is also being used to predict use of 1979).
mental and general health care services (Man- Both the construction of the MHI and the
ning, Newhouse, & Ware, 1982; Wells, Man- techniques selected to analyze the data were
ning, Duan, Ware, & Newhouse, 1982).' Al- geared to better understand the dimensionality
though the MHI was designed for a particular and independence issues discussed above. Pos-
study, the rationale behind its development itive and negative items were included to ad-
and our findings regarding its underlying men- dress the issue of whether psychological distress
tal health dimensions should be useful in con- and well-being are distinct dimensions as de-
ceptualizing and measuring mental health for fined by the MHI and whether these dimen-
other general population studies. sions in turn are multidimensional (i.e.,
When work on the MHI began, a number whether they should be considered higher order
of conceptual issues with theoretical impli- mental health factors). Confirmatory factor
cations had not been resolved for general pop- analytic procedures provided us with an ob-
ulation measures. The first issue was whether jective basis for testing between alternative
mental health, as defined by instruments like measurement models and structural hy-
the MHI, has one or more dimensions. If more potheses. For practical applications of the
than one, do items seem to group into two MHI, we were also interested in whether sat-
distinct constructs—psychological distress isfactory reliability could be achieved with
(negative mental health states) and psycho- simple summated rating procedures for esti-
logical well-being (positive mental health mating a summary MHI score and scores on
states)—or do they tend toward bipolarity, with the separate factors.
factors defined by both negative and positive Whereas issues concerning models of mental
mental health states? Third, if more than one, health must ultimately be resolved in terms
are the dimensions correlated or independent? of which model best predicts mental health
Recently, Dohrenwend and his colleagues behavior and other important concepts, a psy-
have argued for a single-factor interpretation chometric prerequisite for using a multidi-
of screening instruments like those the MHI mensional specification for purposes of pre-
is based on, calling it "demoralization" (Doh- diction is the extraction and cross-validation
renwend, Oskenberg, Shrout, Dohrenwend, & of distinct mental health factors. Our concern
Cook, 1981; Link & Dohrenwend, 1980). here is with the psychometric basis of such a
Contrary to this position, a number of studies model.
have presented evidence for dimensionality
within the psychological distress dimension Method
(Costello # Comrey, 1967; Derogatis, Lipman,
Covi, & Rickels, 1971; Derogatis, Lipman, Construction of the MHI
Rickels, Chlenhuth, & Covi, 1974; Ware et
al., 1979; Edwards, Yarvis, Mueller, Zingale, The MHI was based substantially on the General Well-
& Wagman, Note 1). Being Schedule (GWB) developed by Dupuy (1972, Notes
Arguments for independent positive and 2-5). The GWB represents several psychological distress
negative dimensions have also been offered. constructs and includes items measuring well-being. Its
Bradburn (1969) has interpreted high interi-
tem correlations within groups of positive and 1
negative items and low correlations between The HIE is a social experiment designed to estimate
the effects of different health-care-financing arrangements
positive and negative items to mean that his on the demand for health care services as well as on health
Affect Balance Scale defines two distinct and status and patient satisfaction outcomes. Families are as-.
independent (uncorrelated) dimensions— signed randomly to plans differing in coinsurance and
positive and negative affect. Factor analytic deductible arrangements in the fee-for-service system or
evidence for a psychological well-being di- in a prepaid health plan. Their use of services and changes
in health and other variables are monitored over time
mension distinct from dimensions of psycho- (Newhouse, 1974; Newhouse, Manning, Morris et al.,
logical distress is also abundant from studies 1981).
732 CLAIRICE T. VEIT AND JOHN E. WARE, JR.

content emphasizes symptoms of psychological distress of increasing complexity do significantly better than simpler
that are most prevalent in general populations (e.g., anxiety models in accounting for the data? (e) Can the conclusions
and depression). Hence, the GWB provided a good starting based on one site be generalized across the other three
point for constructing the MHI. To address adequately sites? (f) Is sufficient reliability achieved using a simple
the dimensionality issues left unresolved by previous stud- summated ratings method to estimate factor scores? and
ies, it was necessary to add numerous other items. These (g) Are scores stable across a 1-year interval?
items were added and tested in two phases. These analyses were conducted in four stages. First, we
In the first phase, the factor structure of a 22-item version selected one site to address questions a through d above.
of the GWB was evaluated (Ware et al., 1979; Dupuy, Second, we performed independent cross-validation tests
Note 2). The hypothesis of six underlying factors—Anxiety, of the derived factor structure in each of the three re-
Depression, General Health, General Positive Affect, Loss maining HIE sites. Because results were very consistent
of Behavioral/Emotional Control, and Vitality—was clearly across all four sites, we also tested the alternative factor
supported. Correlations between four of these factors structures and estimated model parameters on data com-
(Anxiety, Depression, General Positive Affect, and Loss bined over all four sites. Third, we evaluated the reliability
of Behavioral/Emotional Control) and numerous other of factor scores obtained by the simple summated ratings
criterion variables further supported their validity. As an procedure. Finally, we estimated the stability of the MHI
inspection of their content anticipated, the GWB General score and subscale scores over a I-year interval.
Health and Vitality factors clearly failed discriminant tests Model development. We selected Seattle for the initial
of validity. Specifically, they correlated substantially with analyses because the MHI was first administered there and
both mental and physical health factors (Ware et al., 1979, because it had the largest number of respondents. A series
1980b). Hence, these two factors were eliminated from of analyses was performed on data obtained in this site.
the MHI. First, we performed factor analyses on two separate cor-
In the second phase, IS GWB items showing good dis- relation matrices in Seattle to assess whether the items
criminant validity were supplemented with 20 additional shown in Table 1 define an underlying mental health di-
items based on instruments developed by others (Beck, mension. Factor analyses were performed using SPSS soft-
1967; Costello & Comrey, 1967; Comrey, 1970; Dohren- ware (Nie, Hull, Jenkins, Steinbrenner, & Bent, 1975). In
wend, Shrout, Egri, & Mendelsohn, 1980) to enhance one analysis, we used a principal components factor ex-
measurement of these four GWB mental health factors. traction and a Varimax factor rotation to analyze a matrix
Three items identified by Ware et al. (1979) to represent of correlations among 38 MHI and 19 physical functioning
a fifth hypothesized factor, Emotional Ties, were also added items. The physical functioning items are described else-
to the battery. The first column in Table 1 provides a short where (Stewart, Ware, & Brook, 1981). The idea behind
description of each item (items are grouped according to this analysis was that, if the hypothesis of a distinct un-
final conclusions of this article); Column 2 identifies each derlying mental health dimension is appropriate, items in
item's originally hypothesized construct. the MHI should emerge on a separate factor than those
selected to define physical functioning. In the other analysis,
a principal components solution for just the 38 MHI items
Sample Characteristics and provided an assessment of the proportion of variance in
Data-Gathering Methods the interitem correlation matrix accounted for by the first
The MHI was fielded at all six HIE sites: Dayton, Ohio; extracted factor. The proportion should be substantial and
Seattle, Washington; Fitchburg and Franklin County, the magnitude of loadings roughly consistent across items
Massachusetts; Charleston and Georgetown County, South to support the hypothesis of a general underlying mental
Carolina.2 Characteristics of respondents (N = 5,089) are health component.
summarized in Table 2, With one exception, data for our Second, we used factor analyses (principal components
analyses came from a self-administered questionnaire that factor extraction, Varimax rotation, and a Promax oblique
was fielded when the HIE began. The exception was Day- rotation) as an aid in determining item clusters for the
ton, where MHI data came from surveys fielded after the two- and five-factor models. Conclusions about item clus-
experiment began. Longitudinal data for the 1-year stability ters were determined as follows. Generally, an item was
analyses came from identical versions of the MHI fielded hypothesized to define the factor on which it had the highest
factor loading. Items that had good convergent validity
approximately 1 year apart to a subsample at each site
(N = 3,525). Questionnaires were mailed out and mailed but poor discriminant validity3 were still hypothesized to
back and respondents were compensated for completing define their originally hypothesized construct (see Table
them (see Ware et al., 1980a for additional details). In- 1) on the strength of item content (face validity) and pre-
vious empirical data on the item.4 i
terviewer assistance was provided in person or by telephone
when needed. These procedures produced high return rates
(above 90% at all sites) and very few missing responses. 2
The two sites within Massachusetts and the two within
South Carolina were combined to achieve a high subjects-
Data Analysis to-variables ratio in the site-specific analysis.
3
Convergent validity was judged satisfactory when an
Our analyses addressed the following questions: (a) Does item's correlation with its hypothesized factor equaled or
a single mental health factor underlie the data? (b) Is there exceeded .40. The discriminant validity of an item was
empirical support for a two-factor (Psychological Distress considered poor when the difference in magnitude between
and Psychological Well-Being) higher order model? If so, its highest and second highest factor loading was less than
are the two dimensions correlated or independent? (c) Is .15.
4
there empirical support for the five lower-order factors that At this stage in the analysis we also studied the effects
were built into the instrument (see Table 1)7 (d) Do models of two kinds of response set (acquiescence and socially
 PSYCHOLOGICAL DISTRESS AND WELL-BEING 733

Table 1
Summary Information About Mental-Health Inventory Items

A priori scale
Factor groupings/item content8 hypothesis Item meanb Item SD

Anxiety (A)
Very nervous person A 4.96 1.12
Bothered by nervousness A 5.14 .99
Pelt tense or high-strung A 4.79 1.05
Anxious, worried A 4.78 1.04
Difficulty trying to calm down A 5.04 .95
Nervous or jumpy A 4.46 1.12
Restless, fidgety, impatient A 4.68 1.04
Rattled, upset, flustered A 4.52 .91
Hands shake when doing things A 5.46 .90
Relax without difficulty0
Depression (D)
Moody, brooded about things D 4.90 .95
Low or very low spirits D 5.01 .88
Felt downhearted and blue D 4.93 .97
Felt depressed D 4.10 .71
Strain, stress, pressure" A 4.48 1.25
Loss of Behavioral/Emotional Control (B)
Control behavior, thoughts, feelings B 5.00 .96
Concern about losing control of mind B 5.54 .91
Felt emotionally stable B 4.93 1.23
Nothing turns out as wanted D 4.48 1.10
Felt like crying D 5.00 1.05
Better off if dead D 5.76 .66
Down in the dumps D 5.26 .90
Think about taking own life D 4.91 .40
Nothing to look forward to D 5.20 1.07
General Positive Affect (G)
Happy person G 4.41 1.03
Happy, satisfied, or pleased G 4.27 1.03
Daily life interesting G 4.00 .17
Felt calm and peaceful G 4.07 .19
Felt cheerful, lighthearted G 4.10 .14
Generally enjoyed things G 4.43 .03
Relaxed and free of tension A 3.83 .20
Living a wonderful adventure G 4.29 .39
Expect an interesting day G 4.09 .23
Wake up fresh, rested G 3.95 .14
Future hopeful, promising G 4.18 .34
Emotional Ties (E)
.Felt loved and wanted E 4.69 1.28
Love relations full, complete E 4.42 1.49
Time felt lonely0 E 5.09 1.04

Note. All items in this table are positively scored.

• For the two-factor solution, all of the items on the A, B, and D factors and the "time felt lonely" item were hypothesized
a priori to load on Psychological Distress; remaining items were hypothesized to load on Psychological Well-Being.
This was also the definition of these two factors in the final two-factor solution.
b
Means and standard deviations are based on a combined sites analysis, N = 5,089.
c
These items had secondary loadings in the final five-factor solution and were not used to score subscales.

Third, the LISREL software for .testing structural equation

desirable) by including direct measures of these response models (Joreskog & Sorbom, 1981) was used to assess how
tendencies. They did not seem to influence our conclusions well the one-, two-, and five-factor models accounted for
about the structure'of the MHI and were dropped from the MHI data; for the two- and five-factor models, we
further consideration (see Ware, Veit, & Donald, in press, examined both correlated and uncorrelated factor struc-
for further details). tures. LISREL was also used to test for differences in fit
734 CLAIRICE T. VEIT AND JOHN E. WARE, JR.

Table 2
Summary of Respondent Characteristics

No. of
school years Family income'
Age completed
% % Range
Site N male nonwhite Range M Range M ($) M($)
Seattle, Wash. 1,755 48 4 13-61 32.0 4-25 12.8 0-22,240 9,536
Dayton, Ohio 1,046 46 10 14-69 33.8 3-24 12.7 0-23,850 12,465
Fitchburg/Franklin
County, Mass. 1,041 46 2 13-66 32.6 2-22 12.4 0-18,158 7,876
Charleston/Georgetown
County, S.C. 1,247 44 43 13-59 31.0 0-27 11.3 0-45,454 8,862
All sites combined 5,089 46 15 13-69 32.2 0-27 12.4 0-45,454 9,636
1
Expressed in 1973 dollars for Dayton, and in 1974 dollars for all other sites.

among the most viable models resulting from these as- we tested the competing factor solutions using
sessments.9 All of the analyses described above were per-
formed using data from Seattle respondents only (N =
confirmatory factor analysis.
1,755).
Cross-validation analyses.- The generalizability of the Testing for an Underlying Mental-Health
Seattle factor structure to the other three sites was evaluated Component
by testing for significant differences within those sites
among the viable models. The idea was that a generalizable A two-factor solution for physical function-
conclusion was one where the relative "goodness" of fits ing and mental health items with orthogonal
of the competing models remained invariant across sites.
Such a result would justify using a standard set of scoring rotation produced a simple structure with the
rules in future studies. These evaluations were performed mental health and physical functioning items
using confirmatory factor analysis. clustering on separate factors. The smallest
Reliability of summated rating scales. Internal-con- difference in magnitude between an item's
sistency reliability was evaluated for scale scores by esti- factor loadings on the physical and mental fac-
mating coefficient alpha (Cronbach, 1951). Total scale
scores were computed using the summated ratings pro- tors was about .30, indicating good discrim-
cedure (Likert, 1932). We also computed Pearson product- inant validity for all of the items. These results,
' moment correlations among the summated scale scores which are documented elsewhere (Ware, Veil,
and compared their magnitudes with the reliability esti- & Donald, in press), support our hypothesized
mates, Intercorrelations among the scales (factors) should
be substantially lower than the reliability estimates to con-
distinction between mental and physical func-
clude that the scales actually measure different constructs. tioning concepts and items.
Data from all four sites were combined for these analyses In the principal components analysis of just
(N = 5,089). the 38 mental health items, about 43% of the
Stability of factor structure. Pearson product-moment variance in the correlation matrix was ac-
correlations were used to evaluate the stability of the MHI
total score and the scale scores of the final higher and counted for by the first extracted factor in
lower order factor solutions over a 1-year interval for a Seattle. Loadings on this factor ranged from
subset of respondents (N = 3,525). .42 to .80, indicating that the first factor ac-
counted for a substantial proportion of the
Results variance in each item.6 This result supports
Model Development in Seattle 1
Decisions on fixed parameters and estimated parameter
Using data from the Seattle site, we initially input values for each confirmatory factor analysis were
examined the hypothesis that the MHI denned based on the corresponding exploratory factor analytic
a general mental health factor. Then, we as- solution and the assumption of a simple structure model.
6
sessed the plausibility of one-, two-, and five- Items with the lower loadings tended to be the more
extremely worded items that had highly skewed distri-
factor orthogonal and oblique structures using butions and restricted variances. They were retained be-
factor analysis. After item clusters had been cause they were found to have good discriminant validity
initially determined from the factor analyses, in other research.
 PSYCHOLOGICAL DISTRESS AND WELL-BEING 735

the notion of a general underlying mental combined the depression and loss of behav-
health factor; at the same time, it calls attentionioral/emotional control factors (Table 1) to get
to the fact that a substantial proportion of the a better idea of which solution—a four- or
variance would be left unexplained by a uni- five-factor oblique model—better defined the
dimensional model. MHI lower order factor structure.
The interfactor correlations for oblique so-
Two- and Five-Factor Orthogonal and lutions were substantial (ranging from .38 to
Oblique Solutions .59 in absolute magnitude), suggesting that
models with correlated factors should be tested
The total matrix variance explained in- in the confirmatory analyses. Complete tables
creased from 43% to 50% wheri a second factor of results from these solutions are reported
was extracted, and to 60% for a five-factor elsewhere (Ware et al., in press).
solution.7 The item groupings indicated by the
factor loadings observed in the two-factor so- Confirmatory Tests Among Seven Models
lution corresponded exactly to those predicted
by our two-dimensional (psychological distress Chi-square goodness-of-fit estimates were
and well-being) hypothesis. These hypothe- obtained for seven models—single-factor, two-
sized item groupings are defined in a footnote factor orthogonal and oblique, four-factor or-
to Table 1. All items hypothesized to measure thogonal and oblique, and five-factor orthog-
Anxiety, Depression, and Loss of Behavioral/ onal and oblique. Goodness of fit indexes for
Emotional Control correlated highest with the these models in the Seattle site are shown in
psychological distress factor. All items hy- Table 3.
pothesized to measure General Positive Affect The chi-squares in Table 3 indicate that
and Emotional Ties correlated highest (with none of the models accounts for the data with-
the exception of "time felt lonely") with the out significant deviations. However, the chi-
psychological well-being factor. Further, all square statistic is a direct function of sample
items exceeded our .15 discriminant validity size and thus, the probability of rejecting a
criterion in the two-factor solution, providing model increases as sample size increases. Since
support for a two-dimensional specification of we are dealing with a very large sample, a
mental health based on unipolar psychological "poor" fit to almost any model would be ex-
distress and well-being factors. pected (Bentler & Bonett, 1980; Joreskog &
The five-factor solution provided a basis for Sorbom, 1979). Thus, better assessments of
a five-dimensional specification of mental the models would result from inspection of
health. In that factor solution, 13 of the 38 the number of noteworthy residual correla-
items did not meet our .15 discriminant cri- tions and from differences 8
in the incremental
terion. These items were grouped with those fits of the models.
items defining their originally hypothesized The number of residual correlations greater
factor for confirmatory analyses. The resultant than or equal to . 10 and the magnitude of the
item groupings are displayed in Column 1 of largest residual correlation are noted in Col-
Table 1. As can be seen by comparing the item umns 5 and 6, respectively, of Table 3. The
groupings in Column 1 with our hypothesis
in Column 2 of Table 1, there was a high 7
The only difference between the orthogonal and oblique
degree of correspondence between our five- rotations for the two- and five-factor solutions was that
factor hypothesis and the pattern of item cor- the difference in magnitude among an item's factor loadings
relations for three of the five factors—Anxiety, was greater for the oblique solutions. The sixth, seventh,
General Positive Affect, and Emotional Ties. and eighth factors were relatively small, accounting for
only an additional 2% of the variance each. The inter-
However, it is apparent that the depression pretability of the five-factor solution was clearly more
and toss of behavioral/emotional control items straightforward than solutions with more factors. The ef-
did not correlate with their respective factors fects of these additional factors on the interpretability of
as hypothesized. In previous research, similar the structure is discussed in detail elsewhere (Ware et al.,
items have been grouped into a single depres- 8
A residual correlation is obtained by subtracting the
sion scale. This led us to retain (for our con- correlation between two items predicted by the model
firmatory analyses) a four-factor model that from the actual obtained correlation.
736 CLAIRICE T. VEIT AND JOHN E. WARE, JR.

Table 3 differences was highly significant (p < .001),

Goodness-of-Fit Indexes and Residuals for Seven indicating significant increases in information
Models: Seattle (N = 1,755) with the addition of each correlated factor.
No. of Maximum The index in the last column of Table 4 pro-
X2 residuals residual vides information about the incremental fit
Model df X2 (df) a . 10 correlation among the models, independent of sample size
(Bentler and Bonett, 1980).10 As can be seen
HI 665 10495.00 15.78 88 .42 from the last row in Panel 1 of this column,
H2O 665 9353.49 14.07 348 .65
H2C 664 7829.30 11.79 58 .35 the five-factor oblique model picked up about
H4O 665 9809.81 14.74 513 .65 48% more information in the correlation ma-
H4C 659 5685.58 8.63 71 .28 trix than the single-factor model (see Row 4).
H5O 665 11305.88 17.00 457 .65 However, the five-factor oblique model con-
H5C 655 5412.6 8.26 69 .28
tributed only a 3% increase in amount of in-
Note. All differences between models are statistically sig- formation over the four-factor oblique model,
nificant (p < .001). which combined the Loss of Behavioral Emo-
HI is a single-factor model; H2O is a two-factor orthogonal tional Control and Depression factors (see Row
model; H2C is a two-factor correlated model; H4O is a 3). Despite this small increase, both the four-
four-factor orthogonal model that combines items shown
under Behavioral/Emotional Control and Depression in and five-factor oblique models were retained
Table 1; H4C is a four-factor correlated model; H5O is a for cross-validation tests to better evaluate the
five-factor orthogonal model; H5C is a five-factor correlated best lower order MHI factor structure.
model.
Cross-Validation Tests
number of large residual correlations was sub- Comparisons among the four models de-
stantially greater for the orthogonal than the rived and tested in Seattle (Panel 1 of Table
oblique solutions. For example, in the two- 4) were performed independently in the Day-
factor orthogonal solution, 348 of the 703 re- ton, South Carolina, and Massachusetts sites,
siduals were greater than or equal to . 10, with where respondent characteristics were quite
a high of .65 (see Row 2). This number was different. A summary of these tests is presented
reduced to 58 (with a maximum residual of in Panels 2 through 4 of Table 4. In all sites,
.35) when the factors were allowed to correlate the order of goodness of fit of the four models
(see Row 3). Even larger reductions in number was the same. A two-factor oblique model
and magnitude of noteworthy residuals were added from 26% (Dayton) to 29% (South Car-
observed when factors in the four- and five- olina) additional information in the correlation
factor models were allowed to correlate. (Note matrix over a single-factor model; from 8%
that the five-factor uncorrelated model ac- (South Carolina) to 16% (Dayton) was gained
counted for the data noticeably less well than with a four-factor oblique model over the two-
the single-factor model.) A comparison of re- factor model; and a five-factor oblique model
siduals between the orthogonal models and gained from 2% (South Carolina) to 4% (Day-
their respective correlated solutions points to ton) more information than a four-factor
the importance of estimating correlations oblique model. All of the chi-square differences
among mental health factors. The chi-square were highly significant (p < .001). The five-
difference in goodness of fit between each or- factor oblique model showed gains of from
thogonal model and its oblique counterpart
was highly significant (p < .001; see Table 3). 9
These results argue for rejecting all orthogonal The oblique models are all special cases of the more
restricted single-factor model. A procedure for distin-
specifications. guishing among a set of nested models is to evaluate the
In Panel 1 of Table 4, we present differences difference in goodness of fit (Bentler & Bonett, 1980; Jo-
in goodness of fit among the remaining three reskog & Sorbom, 1979).
10
correlated models. The single-factor model is The statistic shown in Column 6 of Table 4 uses the
used as the basis for comparison.9 Column 4 chi-square associated with the single-factor model (used
as a baseline for comparison here) in the denominator;
shows the chi-square difference statistics for the difference in chi-squares between the models being
these comparisons. Each one of the chi-square compared serves as the numerator.
 PSYCHOLOGICAL DISTRESS AND WELL-BEING 737

39% (South Carolina) to 45% (Dayton) in in- A Five-Factor Overlapping Model

formation over the single-factor model. Be-
cause the five-factor oblique model consistently Up to this point, our models have assumed
did significantly better than the four-factor that each item measured a single factor. The
substantial interfactor correlations found
model in accounting for the data, we rejected
the four-factor specification as the lower order across sites and the number and pattern of
residual correlations exceeding .10 even with
factor solution. the best model led us to explore a five-factor
Combined Sites Analyses model that allowed some items to load on more
than one factor. Residuals in the correlation
The consistency of the results across sites matrix resulting from our best model (i.e., the
justified a summary combined-sites analysis. five-factor oblique solution) were helpful in
Results of chi-square difference tests among identifying items that might be appropriately
the four competing models in the combined- overlapped. In this matrix, 45 residual cor-
sites analyses are shown in the last panel of relations were greater than or equal to .10,
Table 4. As would be expected, these tests pro- and the largest residual was .26. Although these
duced similar results to those obtained in the noteworthy residuals were scattered among
four sites separately (compare with Panels 1- many different items, a few items were in-
4 of Table 4). The five-factor oblique model volved in a disproportionate number. Many
did best in accounting for the MHI data, cap- residuals involved depression and behavioral/
turing 49% more information in the correla- emotional control items. Close examination
tion matrix than the single-factor model. of the residuals led us to reason that a better

Table 4
Comparisons of Fit Between Models
Model Chi-square
Site comparison difference" df A
Seattle (N = 1,755) HI-H2C 2665.7 1 .25
H2C-H4C 2143.7 5 .20
H4C-H5C 273.0 4 .03
H1-H5C 5082.4 10 .48
Dayton (N = 1,046) H1-H2C 1433.9 1 .26
H2C-H4C 910.6 5 .16
H4C-H5C 201.6 4 .04
H1-H5C 2546.1 10 .45
South Carolina (N = 1,247) H1-H2C 1891.7 1 .29
H2C-H4C 512.9 5 .08
H4C-H5C 103.7 4 .02
H1-H5C 2508.2 10 .39
* Massachusetts (N = 1,041) H1-H2C 1676.2 1 .27
H2C-H4C 896.8 5 .15
H4C-H5C 161.5 4 .03
H1-H5C 2734.5 10 .45
Combined sites (N = 5,089) H1-H2C 7540.5 1 .30
H2C-H4C 4105.3 5 .16
H4C-H5C 705.7 4 .03
H5C-H5C' 1826.5 4 .07
H1-H5C' 14178.0 14 .56 .

Note. HI is a single-factor model; H2C is a correlated two-factor model; H4C is a correlated four-factor model;
H5C is a correlated five-factor model; H5C is a correlated five-factor model with secondary loadings for three items:
the "strain" Depression item, the "able to relax" Anxiety item, and the "lonely" Emotional Ties item. Recall that the
"lonely" item had secondary loadings on the Depression and Behavioral/Emotional Control factors.
* All chi-square comparisons were statistically significant (p < .001).
738 CLAIRICE T. VEIT AND JOHN E. WARE, JR.

solution might be one that allowed three items summated rating scales based on the five-factor
(starred in Table 1) to have secondary factor overlap model (top panel), the two-factor
loadings. An emotional ties item ("time felt model (middle panel), and the single-factor
lonely") was permitted a secondary loading model, that is, the MHI (lower panel). The
on both Depression and Loss of Behavioral items that we allowed to overlap in the final
Emotional Control; an anxiety item ("relax five-factor model (items with a superscript "c"
without difficulty") was permitted a secondary in Table 1) were not used when estimating
loading on General Positive Affect; and a scores for the five factors derived from this
depression item ("strain, stress, pressure'*) was model. Scores based on the one- and two-factor
permitted a secondary loading on Anxiety. models were estimated using all 38 items. (See
The five-factor oblique model that allowed footnote in Table 1 for items comprising the
estimates of these four additional parameters final two-factor solution.) Hereafter, we refer
was compared with the five-factor oblique to the factors as "scales."
model in the combined-sites data. This over-
lapping model reduced the number of residual
correlations greater than or equal to . 10 from Reliability, Stability, and Scale
45 to 21; the size of the maximum residual Intercorrelations
decreased from .26 to .17. Table 5 also presents reliability and stability
The third row of the last panel of Table 4 coefficients and intercorrelations among scales
shows the chi-square difference statistic for the based on the final one-, two-, and five-factor
simple five-factor oblique model versus the models. The reliability (internal consistency)
five-factor overlap oblique model. As can be estimates for all scales are more than satis-
seen, the five-factor overlap model picks up factory for group comparisons, ranging from
7% more information in the MHI correlation .83 to .91 for scales based on the five lower
matrix than the nonoverlapping five-factor order factors and from .92 to .96 for scales
oblique model; it represents an improvement based on the two higher order factors and the
over the baseline single-factor model of 56% MHI. The stability coefficients are in the .56
(Row 5 of the last panel)." Estimates of factor to .64 range, indicating that a substantial pro-
loadings for our final five-factor overlap model portion of the reliable variance in these scales
in the combined sites analysis are shown in is stable over a 1-year interval.
Table 5. The top panel of Table 5 presents correla-
tions among the five MHI subscales. As can
Summary of Results be seen, these range in absolute magnitude
from .34 between Emotional Ties and Anxiety
The cross-validational and combined-sites to .75 between Anxiety and Depression. All
analyses led us to conclude that the best in- of these intercorrelations are lower than the
terpretation of the MHI is a hierarchical factor reliability coefficients for the five subscales,
model composed of a general underlying men- which indicates that the subscales contain a
tal health factor, a higher order factor structure, noteworthy amount of unique reliable vari-
and a lower order factor structure as illustrated ance.
in Figure 1. The higher order factor structure
is defined by two correlated factors—Psycho-
logical Distress, which consists of all the items Discussion and Conclusions
that describe negative mental health states, and Our analyses produced a clear and consis-
Psychological Well-Being, which consists of all tent picture of the factor structure of the Men-
of the items describing positive mental health
states. Thus, both factors are unipolar in na-
ture. The lower order factor structure is defined " Another "null" model that may be of interest is one
by five correlated factors with three of the 38 that specifies 38 uncorrelated factors each measured by
items defining more than one factor. Item one item, that is, all zeros in the off-diagonals of the input
groupings for this lower order factor structure correlation matrix. In the combined-sites analysis, our sin-
gle-factor model picked up 77% more information in the
are shown in the first column of Table 5. correlation matrix than this "null" model that assumes
Table 6 presents summary statistics for correlations among the items are zero.
 PSYCHOLOGICAL DISTRESS AND WELL-BEING 739

Table 5
Factor Loading Estimates for the Final Model, Combined Sites (N = 5,089)
Factor groupings/item-content D B
Anxiety (A)
Very nervous person .80
Bothered by nervousness .80
Felt tense or high-strung .77
Anxious, worried .74
Difficulty trying to calm down .74
Nervous to jumpy .69
Restless, fidgety, impatient .69
Rattled, upset, or flustered .68
Hands shake when doing things .50
•Relax without difficulty -.29" .46
Depression (D)
Moody, brooded about things .74
Low or very low spirits .83
Downhearted and blue .82
.74
•Strain, stress, pressure .51 .14
Behavioral/Emotional Control (B)
Control of behavior, thoughts, feelings .60
Concern about losing control of mind .58
Felt emotionally stable .55
Nothing turns out as wanted .57
Felt like crying .66
Better off if dead .59
Down in the dumps .75
Think about taking own life .44
Nothing to look forward to .69
General Positive Affect (G)
Happy person .83
Happy, satisifed, or pleased .72
Daily life interesting .67
Felt calm and peaceful .79
Felt cheerful, lighthearted .77
Generally enjoyed things. .75
Relaxed and free of tension .74
Living a wonderful adventure .71
Expect an interesting day .67
Wake up fresh, rested .61
Future hopeful, promising .58
Emotional Ties (E)
Felt loved and wanted .85
Love relations full, complete ,82
•Time felt lonely .18 .29 -.26

"These items were overlapped in the final five-factor solution.

b
c
These loadings were fixed at zero.
It should be noted that although the overlap model substantially reduced the number of noteworthy residuals, the
secondary loadings are smaller than our covergent validity criterion of .4.

tal Health Inventory (MHI), This structure is The MHI Structural Model
a blueprint for scoring the MHI in future
studies. The analyses also clarified some of the Both the higher and lower order factor di-
theoretical issues debated in the literature on mensions identified in the exploratory and
general population mental health surveys. confirmatory factor analyses corresponded
740 CLAIRICE T. VEIT AND JOHN E. WARE, JR.

Anxiety

Psychological
Distress Depression

Loss of Behavioral/
Emotional Control

Mental Health '

General Positive
Psychological Affect
Well-Being

Emotional Ties

Figure 1. Mental Health Inventory structure.

well with those originally hypothesized. How- substantial gains with a two-dimensional spec-
ever, some discrepancies between the hypoth- ification of mental health. As hypothesized,
esized and obtained five-factor structure were items describing positive states clustered to-
observed. gether to define psychological well-being and
Figure 1 illustrates several of our major items describing negative states clustered to-
conclusions. First, a large mental health factor gether to define psychological distress. Of im-
underlies the MHI. Hence, there is a sound portance here is that these factors are clearly
psychometric basis for using the single MHI distinct. This result supports the practice of
summary score to define a bipolar psycholog- scoring them as separate factors (Bradburn,
ical distress versus well-being concept. Second, 1969; Goldberg, 1978; Radloff, 1977; Ware et
reliance on a single score is associated with aL, 1979). Third, we consistently identified
significant loss of information. We observed structures within both psychological well-being

Table 6
Summary Information About Mental Health Index Scales'
Scale intercorrelations
Re- Sta-
Scale M SD liability0 bility" A D B PD PWB

Anxiety (A) 9 19.15 6.85 .90 .63

Depression (D) 4 8.05 2.97 .86 .56 .76
Loss of Behavioral
Emotional Control
(B) 9 15.90 5.57 .83 .58 .71 .77
General Positive Affect
(0) 11 45.64 9. 56 .92 .62 -.62 -.70 -.69
Emotional Ties (E) 2 9.08 2,,56 .81 .59 -.39 -.50 -.53 .62
Psychological Distress
(PD) 24 47.54 15,,39 .94 .62 .93e .89* .90° -.73 -.52
Psychological Well-Being
(PWB) 14 59.16 12..16 .92 .63 -.63 -.71 -.71 .98' .74° -.75
Mental Health Index 38 177.56 25,.46 .96 .64 -185' -.86* -.87" .90' .66' -.95' .92'

' Scale scores are computed by summing scores over all items denning the scale (unlike Table 1, some item scores
were reversed, in scoring scales, so that a high score would be consistent with each scale name).
b
Number of items in scale.
°d Internal-consistency reliability estimated using Cronbach's alpha (N = 5,089).
Pearson product-moment correlation between scores obtained approximately 1 year apart (N = 3,525).
' Correlations inflated due to overlapping definitions.
 PSYCHOLOGICAL DISTRESS AND WELL-BEING 741

and psychological distress dimensions. This led reason for the discrepancy. A reanalysis of cor-
us to the hierarchical MHI model shown in relations among ABS items using confirmatory
Figure 1. These higher order psychological factor analysis led us to reject an orthogonal
distress and well-being factors were defined by model of the ABS in favor of one specifying
three and two lower order factors, respectively. correlated factors (Ware et al., in press). We
Fourth, the correlation between distress and found a correlation of -.285 between ABS
well-being was substantial, as were the inter- positive and negative factors. A two-dimen-
correlations among the lower order factors. sional correlated model of the ABS accounted
This hierarchical model was confirmed in for 15% more information in the correlation
cross-validation tests using samples of popu- matrix than a two-dimensional orthogonal
lations with quite different characteristics; this model. This difference was highly significant
supports the model's generalizability. (p < .001). These results argue for a correlated
rather than an orthogonal two-factor model
for the ABS.
Issues Debated in the Literature
The hierarchical MHI model provides in-
Dohrenwend and his colleagues (Dohren- teresting scoring options ranging from five dis-
wendetal., 1981; Link & Dohrenwend, 1980) tinct mental health constructs to reliance on
have argued for a single-factor interpretation a single summary index. An intermediate
for the GWB (the basis of the MHI) and other scoring option is two higher order psychological
instruments like the MHI. Their position is distress and well-being constructs. The trade-
in disagreement with the results of numerous off is between the unique information con-
studies of similar mental health instruments tained in the subscales versus the simplicity
(Bradburn, 1969; Cooke, 1980; Costello & of a single score. Ultimately, the basis for
Comrey, 1967; Derogatis et al., 1971; Gold- choosing among these options would be which
berg, 1978; Radloff, 1977; Wan & Livieratos, model best predicts mental-health-related be-
1978; Ware et al., 1979). Our results from the haviors. In a test between our one- and two-
MHI data are quite clear on this issue. Our factor specification of the MHI, Manning,
five-factor overlap model provided a 56 per- Newhouse, and Ware (1982) rejected a single-
centage point increase in information gained in favor of a two-factor specification (psycho-
in the intercorrelation matrix over a single- logical distress and psychological well-being)
factor model. This substantial gain in infor- for predicting ambulatory medical expendi-
mation with a multidimensional specification tures. Further research is needed to determine
of mental health was replicated in every site. whether a five-factor model predicts mental-
Further, the observed high reliability and sta- health-related behaviors significantly better
bility coefficients (relative to interscale corr than a single- or two-factor model and whether
relations) leaves little doubt that each of the the gains are worth the additional complexity.
MHI subscales contains reliable information
not contained in the other scales. These find- Reference Notes
ings constitute a strong psychometric basis for
a multidimensional specification of mental 1. Edwards, D. W., Yarvis, R., Mueller, D. P., Zingale,
health as defined by the MHI. H., & Wagman, W. J. Test-taking and the stability of
Another important issue is whether mental adjustment scales: Can we assess patient deterioration?
University of California at Davis, Department of Psy-
health factors are correlated or orthogonal. chiatry, 1977.
Bradburn (1969) concluded that his Affect 2. Dupuy, H. J. Developmental rationale, substantive de-
Balance Scale (ABS) could best be described rivative, and conceptual relevance of the General Weil-
by a two-factor model with independent pos- Being Schedule. National Center for Health Statistics,
Fairfax, Virginia, June, 1973.
itive and negative factors. We find no support 3. Dupuy, H. J. Utility of the National Center for Health
for such a model. Specifically, in our analyses Statistics General Well-Being Schedule in the assessment
of higher and lower order factor structures, of self-representations of subjective well-being and dis-
orthogonal models did substantially worse than tress. National Center for Health Statistics, Fairfax,
their oblique counterparts in explaining the Virginia, June, 1974.
variance in interitem correlations in all four 4. ical
Dupuy, H. J. Self-representations of general psycho-
well-being of American adults. Paper presented at
sites. Differences in data-analytic strategies and the annual meeting of the American Public Health As-
other methodological factors appear to be one sociation, Los Angeles, October, 1978.
742 CLAIRICE T. VEIT AND JOHN E. WARE, JR.

5. Dupuy, H. J. A brief description of the research edition Langner, T. S. A, twenty-two item screening score of psy-
of the General Psychological Well-Being Schedule chiatric symptoms indicating impairment. Journal of
(GWB). National Center for Health Statistics, Fairfax, Health and Human Behavior, 1962, 3, 269-276.
Virginia, May, 1979. Liken, R. A technique for the measurement of attitudes.
Archives of Psychology, 1932, 140, 1-55.
References Link, B., & Dohrenwend, B. P. Formulation of hypotheses
about the true prevalence of demoralization in the
Beck, A. T. Depression—Causes and treatment. University United States. In B. P. Dohrenwend, B. S. Dohrenwend,
of Pennsylvania Press, 1967. M. S. Gould, B. Link, R. Neugebauer, & R. Wunsch-
Bentler, P. M., & Bonett, D. G. Significance tests and Hitzig (Eds.), Mental illness in the United States: Ep-
goodness of fit in the analysis of covariance structures. idemiologic estimates. New York: Praeger, 1980.
Psychological Bulletin, 1980, 88, 588-606. Macmillan, A. M. The health opinion survey: Technique
Bradburn, N. M. The structure of psychological well-being. for estimating prevalence of psychoneurotic and related
Chicago: Aldine, 1969. types of disorder in communities, Psychological Reports,
Cronbach, L. J. Coefficient alpha and the internal structure 1957, 3, 325-339.
of tests. Psychometrika, 1951,16, 297-494. Manning, W. G., Newhouse, J. P., & Ware, J. E. The status
Cleary, P. D., Goldberg, I. D., & Kessler, L. G. Screening of health in demand estimation; or, beyond excellent,
for mental disorder among primary care patients. Ar- good, fair, and poor. In V. R. Fuchs (Ed.), Economic
chives of General Psychiatry, 1982, 39, 837-840. aspects of health. Chicago: University of Chicago Press,
Cooke, D. J. The structure of depression found in the 1982.
general population. Psychological Medicine, 1980, 10, Newhouse, J. P. A design for a health insurance experiment.
455-463. Inquiry, 1974, 11, 5-27.
Costello C. G., & Comrey A, L. Scales for measuring Newhouse, J. P., Manning, W. G., Morris, C. N., et al.
depression and anxiety. The Journal of Psychology, 1967, Some interim results from a controlled trial of cost
66, 303-313. sharing in health insurance. New England Journal of
Comrey, A. L. Comrey personality scales. San Diego, Calif.: Medicine, 1981,^05, 1501-1507.
Educational and Industrial Testing Service, 1970. Nie, N. H., Hull, C. H., & Jenkins, J. G., Steinbrenner,
Cronbach, L. J. Coefficient alpha and the internal structure K., & Bent, D. H., Statistical package for the social
of tests, Psychometrika, 1951, 16, 297-334. sciences, McGraw-Hill, 1975.
Derogatis, L. R., Lipman, R. S., Covi, L., & Rickels, K. Radloff, L. S. The CES-D scale: A self-report depression
Neurotic symptom dimensions. Archives of General scale for research in the general population. Applied
Psychiatry, 1971, 24, 454-464. Psychological Measurement, 1977, 3, 385-401.
Derogatis, L. R., Lipman, R. S., Rickels, K., Chlenhuth, Stewart, A., Ware, J. E., & Brook, R. H. Advances in the
E. H., & Covi, .L. The Hopkins symptom checklist measurement of functional status: Construction of ag-
(HSCL): A self-report inventory. Behavioral Science, gregate indexes. Medical Care, 1981, 19, 473-488.
1974, 19, 1-15. Wan, T. H., & Livieratos, B. Interpreting a general index
Dohrenwend, B. P., Oskenberg L., Shrout, P. E., Dohren- of subjective well being. Milbank Memorial Fund Quar-
wend, B. S., & Cook, D. What brief psychiatric screening terly, 1978, 56, 531-556.
scales measure. Health survey research methods (DHHS Ware, J. E., Johnston, S. A., Davies-Avery, A., & Brook,
Publication No. PHS 81-3268). Washington, D.C.: U.S. R, H. Conceptualization and measurement of health for
Government Printing Office, 1981. adults in the Health Insurance Study (Vol. 3, Mental
Dohrenwend; B. P., Shrout, P. E., Egri, G., & Mendelsohn, health; R-1987/3-HEW). Washington, D.C.: U.S. Gov-
F. S. Nonspecific psychological distress and other di- ernment Printing Office, 1979.
mensions of psychopathology. Archives of General Psy- Ware, J. E., Brook, R. H., Davies-Avery, A., et al. Con-
chiatry, 1980, 37, 1229-1236. ceptualization and measurement of health for adults in
Dupuy, H. J. The psychological section of the current the Health Insurance Study (Vol. 1, Model of Health
health and nutrition examination survey. Proceedings and Methodology; R-1987/1-HEW). Santa Monica,
of the Public Health Conference on Records and Statistics Calif.: The Rand Corporation, 1980. (a)
Meeting Jointly with the National Conference on Health Ware, J. E., Davies-Avery, A., & Brook, R. H. Concep-
Statistics. Washington, D.C.: National Conference on tualization and measurement of health for adults in the
Health Statistics, 1972. Health Insurance Study (Vol. 6, Analysis of Relationships
Goldberg, D. Manual of the General Health Questionnaire. Among Health Status Measures; R-1987/6-HEW).
Windsor, England: NFER Publishing Co., 1978. Washington, D.C.: U.S. Government Printing Office,
Goldberg, D., Steele, J. J., Johnson, A., & Smith, C. Ability 1980. (b)
of primary care physicians to make accurate ratings of Ware, J. E., Veit, C. T., & Donald, C. A., Refinements in
psychiatric symptoms. Archives of General Psychiatry, the measurement of mental health for adults in the Health
1982, 39, 829-833. Insurance Study (R-2131-HHS). The Rand Corporation,
Gurin, G., Veroff, J., & Feld, S. Americans view their in press.
mental health—A nationwide interview survey. New Wells, K. B., Manning, W. G., Duan, N., Ware, J. E., &
York: Basic Books, 1960. Newhouse, J. P. Cost Sharing and the Demand for Am-
Joreskog, K. G., & Sorbom, D. Advances in factor analysis bulatory Mental Health Services (R.2860-HHS). Santa
and structural equation models. Cambridge, Mass.: Abt Monica, Calif.: The Rand Corporation, 1982.
Books, 1979.
Joreskog, K. G., & Sorbom, D. LISRL: Analysis of linear
relationships by the method of maximum likelihood.
Uppsala, Sweden: University of Uppsala, 1981. Received April 25, 1983