Social Science & Medicine 66 (2008) 1334e1345

www.elsevier.com/locate/socscimed

Does a higher educational level protect against anxiety

and depression? The HUNT study*
Ingvar Bjelland a,*, Steinar Krokstad b, Arnstein Mykletun c, Alv A. Dahl d,
Grethe S. Tell e, K. Tambs f
a
Clinic for Child and Adolescent Mental Health Services, Haukeland University Hospital, Jonas Lies vei 65, N-5021 Bergen, Norway
b
Department of Community Medicine and General Practice, The Norwegian University of Science and Technology, Trondheim, Norway
c
Faculty of Psychology, University of Bergen, Bergen, Norway
d
Department of Clinical Cancer Research, Rikshospitalet, University of Oslo, Division The Norwegian Radiumhospital, Oslo, Norway
e
Department of Public Health and Primary Health Care, University of Bergen, Bergen, Norway
f
Division of Mental Health, The Norwegian Institute of Public Health, Oslo, Norway
Available online 12 February 2008

Abstract

The relationship of education to the experience of anxiety and depression throughout adult life is unclear. Our knowledge of this
relationship is limited and inconclusive. The aim of this study was to examine (1) whether higher educational level protects against
anxiety and/or depression, (2) whether this protection accumulates or attenuates with age or time, and (3) whether such a relation-
ship appears to be mediated by other variables. In a sample from the Nord-Trøndelag Health Study 1995e1997 (HUNT 2)
(N ¼ 50,918) of adults, the cross-sectional associations between educational level and symptom levels of anxiety and depression
were examined, stratified by age. The long-term effects of educational level on anxiety/depression were studied in a cohort followed
up from HUNT 1 (1984e1986) to HUNT 2 (N ¼ 33,774). Low educational levels were significantly associated with both anxiety
and depression. The coefficients decreased with increasing age, except for the age group 65e74 years. In the longitudinal analysis,
however, the protective effect of education accumulated somewhat with time. The discrepancy between these two analyses may be
due to a cohort effect in the cross-sectional analysis. Among the mediators, somatic health exerted the strongest influence, followed
by health behaviors and socio-demographic factors. Higher educational level seems to have a protective effect against anxiety and
depression, which accumulates throughout life.
Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Norway; Cohort study; Educational level; Anxiety; Depression; Mediators; Adult; Mental health

Introduction
*
The Nord-Trøndelag Health Study (The HUNT Study) is a collab-
oration between The HUNT Research Centre, Faculty of Medicine, A recent meta-analysis (Lorant et al., 2003) found an
The Norwegian University of Science and Technology (NTNU), increased risk of depression in people belonging to the
Verdal, The Norwegian Institute of Public Health, The National lowest socioeconomic status (SES) compared to those
Health Screening Service of Norway and Nord-Trøndelag County
Council.
in the highest. However, SES is an ambiguous concept
* Corresponding author. Tel.: þ47 90955893; fax: þ47 55586130. and in this meta-analysis it was represented by various
E-mail address: [email protected] (I. Bjelland). characteristics such as education, income, occupation,

0277-9536/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.socscimed.2007.12.019
 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345 1335

assets, and social class. As indicators of SES, such vari- levels increases with age (Miech & Shanahan, 2000).
ables are not recommended as interchangeable (Geyer, Another study with a similar design has demonstrated
Hemström, Richard, & Vågerö, 2006). Unlike income, the cumulative advantage of high educational level on
education is quite stable after young adulthood and is physical health (Ross & Wu, 1996).
likely to result in good mental health rather than result In contrast to the meta-analysis of Lorant et al.
from it. Education is a resource that is part of a person, (2003), a concurrent systematic review (Fryers, Melzer,
rather than being external to the person (Ross & Mirow- & Jenkins, 2003) addressed the relationship between
sky, 2006). Education precedes and influences other social inequalities and both depression and anxiety dis-
SES indicators, such as occupation and income. In orders. The three relevant studies (Andrews, Hender-
addition, education, as a source of human capital, may son, & Hall, 2001; Bijl, Ravelli, & van Zessen, 1998;
enable people to succeed more generally and may prove Kessler et al., 1994), which were cross-sectional, all
effective in pursuing fundamental ends that include found a significant negative association between educa-
emotional well-being (Mirowsky & Ross, 1998). Educa- tional level and anxiety disorders. In addition, we are
tion was also more frequently used than the other SES aware of only one longitudinal study addressing the as-
variables in the studies included in the aforementioned sociation between anxiety and education (Miech et al.,
meta-analysis (35 out of 56) (Lorant, et al., 2003). 1999). That study showed that the anxiety level was
Furthermore, the majority (48 out of 56) of the stud- negatively associated with both the educational level
ies (Lorant et al., 2003) had a cross-sectional design, of the respondent’s parents and their own educational
and thereby could not demonstrate whether differences attainment.
in SES were associated with increasing or decreasing Various studies have suggested that some mediators,
differences in depression with age or time. Such such as health behaviors (smoking, alcohol consump-
changes would indicate whether the protective effect tion, and physical activity) and psycho-social status
of a high SES is stable, increases, or decreases during (Kaplan et al., 1987), somatic health (Kaplan et al.,
life. Hence, the possible cumulative effect of educa- 1987; Miech & Shanahan, 2000), socio-demographic
tional level on mental health in principle can be demon- (Bracke, 2000; Kaplan et al., 1987) and work character-
strated, by longitudinal studies but not cross-sectional istics (Bracke, 2000; Kaplan et al., 1987; Ross &
ones. Four longitudinal studies that measured SES by Mirowsky, 2006; Zimmerman, Christakis, & Vander
educational level were reported in the meta-analysis Stoep, 2004) explain the association between educational
(Bracke, 2000; Eaton, Muntaner, Bovasso, & Smith, level and depression. However, the data are ambiguous.
2001; Kaplan, Roberts, Camacho, & Coyne, 1987;
Sargeant, Bruce, Florio, & Weissman, 1990). The re- Aims of the study
sults of these studies were inconsistent, with only one
study (Kaplan et al., 1987) reporting a clear inverse First, we used a cross-sectional design to examine
association between educational level and depression. whether higher educational level protects against anxi-
The only longitudinal study that addressed education ety and/or depression. Second, we examined whether
as the main predictor of depression reported that the as- such a protection accumulates or attenuates with age
sociation was weakened and no longer statistically sig- or time, partly by using an age stratification approach
nificant when adjusted for earlier depressive symptoms in a cross-sectional design, and partly by examining
(Sargeant et al., 1990). These results imply no cumula- the course of anxiety and depression symptoms over
tive effect of education on depression with age. A study an 11-year period. Third, we aimed to identify possible
not included in the meta-analysis that followed adoles- mediators in the relationship between education and
cents into early adulthood did not show an association anxiety/depression.
between educational attainment and depression at all
(Miech, Caspi, Moffitt, Wright, & Silva, 1999). Material and methods
Despite the limitation of the cross-sectional design in
this respect, stratifying the sample by age makes it pos- Sample
sible to study the effects of education on mental health
through life. Increasing associations between low edu- All inhabitants of Nord-Trøndelag County of Norway
cation and depression with age would imply that the aged 20 years and above (N ¼ 85,100) were invited to
protective benefits of high education accumulate participate in the Nord-Trøndelag Health Study 1984e
through life. One study has shown that the difference 1986 (HUNT 1) (Holmen, Midthjell, & Bjartveit,
in depressive level between high and low educational 1990; HUNT, 2007), of which 77,310 individuals
1336 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345

(90.9%) participated 11 years later in the HUNT 2 study. Was eligible for
The HUNT 2 study consisted of all inhabitants aged 20 HUNT 1
years and above in the same county (N ¼ 94,197) (Hol- N=85,100

men et al., 2003), and this time 65,648 individuals

Did not attend N=7,790
(69.7%) participated. Earlier studies have shown impor-
tant general health inequalities by educational level Attended HUNT 1
(Krokstad, Kunst, & Westin, 2002). In both HUNT 1 N=77,310

and HUNT 2, data were obtained from three sources:

Was not eligible for HUNT 2 N=17,177
(1) a first questionnaire that was a part of the invitation
Did not attend HUNT 2 N=12,809
was mailed to all eligible participants and returned
when they attended the study site; (2) physical examina- Did attend both
tion; and (3) a second questionnaire, handed out during HUNT 1 and 2
N=47,324
the examination and returned by mail (by approximately
80% of the participants in both studies).
Invalid information on
The Norwegian Data Inspectorate and the Regional education or Anxiety-Depression Index N=13,550a
Committee for Medical Research Ethics in Health re-
Longitudinal sample
gion IV of Norway approved HUNT 2. These agencies N=33,774
were not established in Norway when HUNT 1 was
conducted. Each participant in the HUNT 2 study was Fig. 1. The selection procedure for the longitudinal sample. aHigh attri-
asked to sign an informed consent, stating that his or tion rate because some of the anxiety and depression items were in a sec-
her data could be used for medical research (Holmen ond questionnaire handed out during attendance at the physical exam.
et al., 2003).
A cross-sectional sample aged 25e84 years from and missing frequency of light physical exercise by
HUNT 2 (N ¼ 60,210) with valid information on educa- heavy physical exercise (and vice versa) using linear re-
tional level, anxiety and depression (N ¼ 50,918) was gression. The remaining missing data were substituted
used to examine age specific effects. Among these, par- by sample means. In the mediator analyses, the covari-
ticipants below 70 years with information on a variety of ates were supplemented with a corresponding set of
somatic problems, health behaviors, socio-demograph- dummy variables that identified the subjects with a miss-
ics, and work characteristics from the second question- ing value (‘‘Mean imputation with a dummy’’) (Paul,
naire (N ¼ 41,819) were included in an examination of Mason, McCaffrey, & Fox, 2003). If the coefficient of
mediator effects. such a dummy is significant, then the data are not miss-
Longitudinal analyses were conducted on a cohort ing completely at random, thus, this procedure implies
from HUNT 1 (N ¼ 33,774) that was followed up at testing for selection bias (Anderson, Basilevsky, &
HUNT 2 (Fig. 1), and was identified by valid informa- Hum, 1983).
tion on educational level and an indicator of symp-
toms of anxiety and depression, termed the Anxiety/ Anxiety and depression
Depression Index (ADI). The participation rate was In HUNT 2, information on symptoms of anxiety and
decreased because some of the items were in a second depression was partly obtained from the Hospital Anxi-
questionnaire in HUNT 1. There was no established ety and Depression Scale (HADS), a self-administered
instrument tapping anxiety and depression in HUNT questionnaire consisting of 14 items, seven for anxiety
1, so the ADI was created in order to have a common (HADS-A subscale) and seven for depression (HADS-
anxiety/depression measure in HUNT 1 and HUNT 2 D subscale) (Zigmond & Snaith, 1983). Previous studies
(described below). have shown a two-dimensional covariance structure of
the HADS (Bjelland, Dahl, Haug, & Neckelmann,
Measures 2002; Mykletun, Stordal, & Dahl, 2001).
However, nearly all the HADS depression items tap
Treating of missing data the absence of positive emotion. Therefore, we wanted
When possible, information was replaced or pre- to supplement the HADS with some additional negative
dicted from a similar variable: Missing educational emotion items to obtain a more comprehensive measure
level at HUNT 1 was replaced by educational level at and to achieve a better correlation between the HADS
HUNT 2, while missing frequency of alcohol drinking and the ADI (described below), the latter which was
was predicted by alcohol consumption (and vice versa) used in both HUNT 1 and HUNT 2. Six new items
 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345 1337

were added to the HADS, three for each subscale, and it Education
was re-termed the Expanded Hospital Anxiety and De- Education was classified into five levels: primary
pression Scale (EHADS) (Cronbach’s alpha 0.85). school (<10 years), high school for 1 or 2 years (10e
These new items were not from a specific instrument, 11 years), complete high school (12 years), college or
but single questions assumed to tap elements of anxiety university less than 4 years (13e15 years), and college
or depression. The complete EHADS (EHADS-total) or university 4 years or more (16 years). Not all
was divided into anxiety (EHADS-A) (Cronbach’s participants in HUNT 1 had finished their education.
alpha 0.86) and depression (EHADS-D) (Cronbach’s Therefore, we composed a common educational level
alpha 0.82) subscales. The reported number and per- for HUNT 1 and HUNT 2 using the highest reported
centages of the missing values are according to all par- level. When information on education was missing at
ticipants in HUNT 2 (N ¼ 60,210). On average, there HUNT 1 (N ¼ 8100, the majority of which was due to
were 8.2% (N ¼ 4942) individuals with missing values not returning the second questionnaire, in which educa-
on each item of EHADS-A and 7.1% (N ¼ 4290) of tion was reported), information from HUNT 2
EHADS-D. However, substitution was performed only (N ¼ 7444) was used and vice versa (N ¼ 3830 and
when valid data already existed for at least half of the N ¼ 2708). When data were missing in both HUNT 1
items, resulting in replacement on average for only and HUNT 2 (N ¼ 656 in HUNT 1 and N ¼ 1122 in
1.1% (N ¼ 659) and 1.6% (N ¼ 941) (EHADS-A and HUNT 2), missing data were not substituted. In the
EHADS-D, respectively) of the subjects on each item. cross-sectional sample (HUNT 2), age was restricted
To achieve close to normal distributions, all three scores to 25 years and above to exclude most participants
were ln-transformed after adding the constants 32 with ongoing education. Mean educational level (1e5,
(EHADS-total), 13 (EHADS-A), and 12 (EHADS-D). 1 was the highest level) was 3.76 (standard deviation ¼
The logarithmic scores were then z-transformed in or- 1.29). In the sample used for longitudinal analyses, the
der to facilitate the interpretation of the regression youngest participants were 31 years old and were as-
coefficients. sumed to have completed their education at HUNT 2.
Because no established measurement instruments for Here, the mean educational level was 4.05 (standard
anxiety or depression were used at HUNT 1, an Anxiety/ deviation ¼ 1.21).
Depression Index (ADI) (Cronbach’s alpha 0.83) was
composed by selecting items that were identical (seven Other covariates
items) or almost identical (two items) in HUNT 1 and These variables from HUNT 2 were included as pos-
HUNT 2 (Data available on request from the authors). sible mediators of the effect of education on anxiety and
These were assumed to tap global features of anxiety, depression, i.e. in the models they were assumed to be
depression, and well-being. The reported number and a consequence of educational level and a cause of
percentages of the missing values are according to the anxiety and depression (intermediate variables). When
individuals who participated with both questionnaires not otherwise specified the data were obtained by self
in both HUNT 1 and HUNT 2 (N ¼ 35,766). On average, report. The covariates were grouped into four blocks.
there were 2.2% (N ¼ 780) individuals with missing The reported numbers and percentages of missing
values on each item at HUNT 1 and 2.5% (N ¼ 910) at data refer to the cross-sectional sample of HUNT 2
HUNT 2. Again, substitution was performed only selected for the mediator analyses (N ¼ 41,819).
when there already were valid data for at least half of The covariate block ‘‘Somatic health’’ included mus-
the items, resulting in replacement on average for only culo-skeletal complaints (from neck, shoulders, elbows,
0.8% (N ¼ 279) and 1.4% (N ¼ 507), of the subjects wrists/hands, chest/stomach, upper back, lower back,
on each item in HUNT 1 and HUNT 2. In order to hips, knees, ankles/feet, average missing data N ¼ 89
maximize the correlation between the two measures of or 0.2%), cardiovascular disease (myocardial infarction,
anxiety/depression, EHADS was regressed on each of angina pectoris, stroke, average missing data N ¼ 81 or
the ADI items and the regression coefficients were 0.2%), diabetes (missing data N ¼ 69 or 0.2%), gastro-
used to weight each ADI item before summation in intestinal symptoms (nausea, dyspepsia, diarrhea,
HUNT 1 and HUNT 2 (Data available on request from constipation, average missing data N ¼ 3481 or 8.3%),
the authors). The correlation between the ADI and the and chronic impairment of daily functioning concerning
EHADS-total in HUNT 2 was 0.83, and the correlation movements, vision, hearing, and somatic illness
between ADI in HUNT 1 and HUNT 2 was 0.53. The in- (average missing data N ¼ 1334 or 3.2%).
dex was z-transformed, with high values indicating high The covariate block ‘‘Health behaviors’’ included
anxiety/depression symptom level. smoking (number of cigarettes per day), alcohol
1338 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345

consumption, and physical activity. Cigarette smoking Statistical analyses

had N ¼ 1781 missing values (4.3%). Alcohol con-
sumption was calculated as a sum of drinking episodes We used linear regression analysis, adjusting for age
during the last month (SD ¼ 3.1) and consumed units of and sex, to examine main effects of educational level
beer, wine, and liquor during the last 2 weeks (each level as dummy variables and highest level as
(SD ¼ 4.8). The variables were coded zero for individ- reference) on anxiety and depression (EHADS-A,
uals who reported themselves as abstainers. In cases EHADS-D, and EHADS-total) in the cross-sectional
where there was no information on the frequency vari- sample. By the same procedure, we evaluated whether
able, this value was imputed using linear regression such effects were linear. Effect modifications were
analysis with the consumption variables as predictors. tested by specifying interaction terms between educa-
Likewise, missing consumption values were predicted tional level and both age and sex in analyses of variance
using the existing alcohol information. Missing values (UNIANOVA, SPSS). To examine the direction of the
on both the frequency and consumption variables possible effect modifications of age (10-year cate-
(N ¼ 1033 or 2.5%) were substituted with sample gories) and sex, linear regression analyses were strati-
means. Physical activity was calculated as a sum of fied by these variables.
graded responses (1e4) to questions about light and For the longitudinal sample, linear regression
hard physical activity outside job during the last year, models were used to predict the ADI level in HUNT 2
where the value of hard physical activity was weighted by educational level, adjusting for age, sex, and the
by a factor of two. In cases where there was no informa- ADI scores in HUNT 1. When entering the baseline
tion on hard physical activity, this value was imputed level of ADI (ADI at time 1) as a covariate, the effects
using linear regression analysis with light physical ac- of the other covariates should be interpreted as effects
tivity as a predictor, and vice versa. When there were on change in ADI from time 1 to time 2.
missing data on both variables, the values were In order to test the effects of the possible mediators,
substituted by sample means (N ¼ 2505 or 6.0%). pairs of regression models were specified with and with-
Height and weight were measured, and body mass index out certain covariates. The differences in the standard-
(BMI) was calculated as kg/m2. There were no missing ized regression coefficients of educational level
data on BMI. (within the pair) are assumed to be due to the influence
The covariate block ‘‘Work-related factors’’ in- of the covariate(s). To achieve such a difference, the co-
cluded variables such as whether the respondent had variate had to be associated with both educational level
lost his/her job, had considered changing jobs, re- and anxiety/depression. The linear regression models
ported that his/her job was so physically strenuous included adjustments for age and sex and were run for
that he/she was often physically tired at the end of the EHADS-total, the EHADS-A, and the EHADS-D.
the working day, reported that the work demanded so First, each covariate was included in a model separately
much concentration and attention that he/she often and consecutively. Then, each covariate block was in-
felt exhausted at the end of the working day, and cluded separately, and finally all covariates were added
whether he/she was satisfied with his/her job. The re- in the full model. All models were applied on the whole
sponse categories of these questions were graded and sample as well as on each sex and age stratum.
missing information (average N ¼ 8912 or 21.3%) Because the large number of statistical tests implies
was substituted by sample means. Having lost his/her a large risk of obtaining significant results at random,
job had no missing values. a strict two-tailed p-value, p < 0.005, was chosen. The
The covariate block ‘‘Socio-demographic’’ factors statistical analyses were conducted using the software
included whether the respondent was living alone and/ package of SPSS 11.5.
or was separated or divorced, his/her type of residence
according to an urbanerural gradient assigned to each Results
municipality (Statistics Norway, 1994), whether there
had been any economic hardship in the household the Characteristics of the study samples
last year, and whether the respondent felt that he/she
had enough friends. Information on the two former vari- Primary school was the most common level of educa-
ables was obtained from public population registries, in tion attained in both samples (Table 1). However, educa-
which there are practically no missing data. Missing tional level was characterized by a marked age gradient
data on the two latter variables (average N ¼ 820 or (Fig. 2). Hence, in the cross-sectional sample the
2.0%) was substituted by sample means. proportion of individuals with only primary school
 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345 1339

Table 1
Characteristics of the study samples and non-attendeesa
Cross-sectional sample Longitudinal sample Non-attendeesa
(N ¼ 50,918) (HUNT 2) (N ¼ 33,774) (HUNT 1/2) (N ¼ 8565) (HUNT 2)
Mean ageb (SD) 49.2 (14.0) 45.1 (20.4) 49.5 (18.7)
Men [N (percent)] 24,753 (48.6) 15,620 (46.2) 4479 (52.3)
Educational level [N (percent)]
Primary school 18,069 (35.5) 15,991 (47.3) 5356 (62.5)
High school <3 years 17,882 (35.1) 10,833 (32.1) 2165 (25.3)
High school 4202 (8.3) 1739 (5.1) 445 (5.2)
College/university <4 years 6407 (12.6) 3089 (9.1) 370 (4.3)
College/university 4 years 4358 (8.6) 2122 (6.3) 229 (2.7)
Mean (SD) Anxiety/Depression Index 0.00 (1.00) 0.10 (1.12)
(z-standardized)
The HUNT study.
a
Participated in HUNT 1 (with valid data on educational level and the Anxiety/Depression Index [ADI] in HUNT 1) and were invited to HUNT 2,
without participating.
b
At HUNT 2 in the cross-sectional sample; at HUNT 1 in the longitudinal sample.

increased monotonously from 9.3% in the youngest age with an increased level of all three anxiety/depression
group to 75.7% in the oldest. The anxiety level de- measures (Table 2). Compared to the highest
creased monotonously with increasing age, while the educational level, the lowest level was associated with
depression level showed an exact opposite pattern (Re- mixed anxiety and depression of an effect size (ES) of
sults not shown). Non-attendees (participants of HUNT 0.26 (95% confidence interval 0.23e0.29), which corre-
1 who were invited, but did not attend in HUNT 2) were sponds to a quarter of the standard deviation of the total
characterized by older age, lower educational level, pre- EHADS. The interaction terms between educational
dominance of males, and a higher level of anxiety/de- level and age and between educational level and sex
pression (Table 1). were all significant (p < 0.005). All the effects of
education on anxiety and depression decreased with in-
Cross-sectional analyses: associations between creasing age, with the exception of the age group 65e74
educational level and anxiety/depression years (Fig. 3). The figure shows standardized regression
coefficients with educational level entered as a continu-
Compared to the highest educational level, the three ous variable. In men, educational level was somewhat
lowest levels were significantly (p < 0.001) associated more strongly associated with depression, with and
without anxiety, and somewhat more strongly associ-
Primary school ated with anxiety in women (results not shown). The as-
High school < 3 years
sociation was most pronounced in the youngest and
High school 3 years
College/university < 4 years oldest age groups (results not shown).
College/university > 4 years
80
Percent within each age group

Longitudinal analyses: associations between

educational level and the ADI
60

The non-standardized regression coefficient of edu-

40 cational level predicting z-transformed ADI in HUNT
2, adjusting for age and sex, was 0.054 (p < 0.001)
(Table 3). After adjusting for ADI in HUNT 1, the coef-
20
ficient was reduced to 0.031, but was still highly signif-
icant (p < 0.001). Interaction terms between educational
0 level and (1) sex, (2) age and (3) age and sex were all sig-
25-34 35-44 45-54 55-64 65-74 75-84 nificant (p < 0.001). Stratified regression analyses re-
Age groups (years) vealed clear group differences regarding age, but not
Fig. 2. Distribution of educational level in the various age groups of sex. The oldest age group (three groups: <45 years;
the cross-sectional sample (N ¼ 50,918). The HUNT 2 study. 45e56 years; >65 years) had coefficients twice as large
1340 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345

Table 2
Mean differences in symptoms of anxiety and depression between groups with high and low education, adjusted by age and sex in HUNT 2
(N ¼ 50,918)
Mixed anxiety and depression Anxiety Depression
Effect size 95 % CI Effect size 95 % CI Effect size 95 % CI
College/university 4 years reference level reference level reference level
College/university <4 years 0.04 0.00e0.08 0.02 0.02e0.05 0.05* 0.01e0.09
High school 0.12** 0.08e0.17 0.08** 0.04e0.12 0.15** 0.11e0.19
High school <3 years 0.16** 0.13e0.19 0.10** 0.07e0.13 0.18** 0.15e0.22
Primary school 0.26** 0.23e0.29 0.18** 0.14e0.21 0.28** 0.24e0.31
*p < 0.010.
**p < 0.001.
Note: Effect sizes were calculated using non-standardized regression coefficients calculated for the normalized dependent variables, which gives
adjusted mean differences divided by standard deviations.

as the middle age group, and the youngest group approx- of covariates. In the first analyses, single covariates
imately 50% larger than the middle age group. were entered one by one. Then blocks of covariates,
To obtain the estimate of the increased difference be- including variables pertaining to somatic health, health
tween the highest and the lowest educational levels behavior, work-related factors, and socio-demographic
(coded 1e5) in anxiety/depression through 11 years, factors were entered consecutively. Finally, all covari-
the coefficient should be multiplied by a factor of four ates were entered simultaneously. Among the covari-
(0.31 SD  4 ¼ 0.124 SD). Thus, people with a low ate blocks, somatic health exerted the strongest
education on average moved towards higher anxiety influence in terms of reducing the observed effect of
and depression scores during the 11-year follow-up education; considerably so not only for anxiety (58%
period, whereas people with a high education changed reduction of the standardized beta coefficient), but
in a favorable direction. In other words, a moderate also for mixed anxiety and depression (43% reduction
protective effect of higher educational levels increased of beta), and for depression (34% reduction of beta).
somewhat during this period. Adjusting for health behaviors and socio-demographic
factors resulted in a moderate reduction of the betas,
Cross-sectional analyses: influence of possible for mixed anxiety and depression 25% and 23%, for
mediators on the educationeanxiety/depression anxiety 19% and 32%, and for depression 28% and
associations 17%. Work-related factors reduced the associations
more modestly with 13% for mixed anxiety and depres-
Table 4 shows the standardized regression co- sion, 2% for anxiety, and 17% for depression. Among
efficients for educational level predicting anxiety/ individual covariates, however, one of the work-related
depression scores in models including various types factors, ‘‘Being physically tired at the end of the day’’,

Table 3
Educational levela predicting the level of anxiety/depression (ADI Anxiety/Depression Index) during an 11-year period by linear regression (adjust-
ing for sex and age)
N Without adjustment for ADI at HUNT 1 After adjustment for ADI at HUNT 1
Bb 95% CI Bb 95% CI
All 33,774 0.054* 0.045e0.063 0.031* 0.024e0.039
Sex (educational level  sex*)
Men 15,621 0.051* 0.038e0.063 0.035* 0.024e0.046
Women 18,153 0.053* 0.040e0.066 0.025* 0.014e0.037
Age (educational level  age*)
<44 years 8214 0.061* 0.042e0.079 0.036* 0.019e0.053
45e64 years 14,930 0.043* 0.030e0.055 0.025* 0.014e0.036
>64 years 10,630 0.085* 0.065e0.106 0.058* 0.041e0.076
*p < 0.001. The HUNT study.
a
Ranked 1e5 (1 is highest level), mean ¼ 4.05, standard deviation ¼ 1.21.
b
Non-standardized regression coefficient.
 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345 1341

0.14 Discussion
Anxiety and depression
0.12 Anxiety
Depression
Main findings
0.10
We found highly significant, though moderate, as-
Beta

0.08 sociations between low educational level and anxiety

and depression in the cross-sectional as well as the
0.06 longitudinal samples. The results from the cross-sec-
tional data showed decreasing effects of education on
0.04 anxiety/depression with increasing age/older cohorts.
If we assume no cohort effects, this could suggest
0.02
25-34 35-44 45-54 55-64 65-74 75-84 a declining protective effect over time of a higher ed-
Age groups (years) ucational level. However, there was an opposite trend
in the longitudinal analysis, that is, the differences
Fig. 3. Associations (linear regression analysis, adjusted for sex)
between educational level and anxiety (EHADS anxiety subscale,
between low and high education for anxiety and de-
ln-transformed), depression (EHADS depression subscale, ln- pression increased with time. The non-standardized
transformed), and mixed anxiety and depression (EHADS total scale, regression coefficient was 0.031 (for one unit in-
ln-transformed) in the various age groups (N ¼ 50,918). The HUNT 2 crease in educational level), which should be multi-
study. plied by a factor of four (¼ 0.124) to obtain the
estimate of the increased difference between the
highest and the lowest educational level (coded 1e
caused the largest reduction in the effect of education 5) in standardized anxiety/depression scores through
(for mixed anxiety and depression 58%, for anxiety 11 years. The accumulated protection through the
75%, and for depression 49%). What may appear to 11-year period is the average effect across persons
be contradictory, that one work-related covariate re- highly different in age. Assuming an adult life span
duces the effect of education more than the whole block from age 25 to 80 and a linear increase in protective
of work-related covariates, is due to the fact that adjust- effect of education throughout life, the lifelong accu-
ing for other work-related variables, like ‘‘exhausted mulated protective effect could be around five times
due to concentration’’, actually increases the effect. the 11 year increase, or in the order of 0.124
Among other single variables that substantially reduced SD  5 ¼ 0.62 SD. Though somewhat theoretical,
the observed effect of education on mixed anxiety and this calculation gives at least a clue about the true
depression, anxiety, and depression were economic lifelong accumulated protection. Combined with the
hardship (38%, 53%, and 31%), musculo-skeletal com- results from the cross-sectional analysis, showing
plaints (26%, 35%, and 21%), and gastro-intestinal a difference between the highest and lowest education
symptoms (24%, 32% and 17%). group of 0.26 (Table 2, averaged across all ages),
The significance level of the dummy variables for these results indicate that there is a relatively weak
mean substituted missing values revealed that not all protective effect from high education in early adult
missing data were completely at random. All the life. However, the effect late in life may be far
dummies of work-related variables (except ‘‘lost from trivial.
job’’) were highly significant (p < 0.001). Compared In the models that included covariates that were
to the corresponding models that did not include assumed to mediate the effect of education on mental
dummies for missing values, the betas for work-related health, the most influential variable block was somatic
predictors of educational level in the models with health, reducing the effect with approximately 45%.
dummies were on average reduced by 13%. Dummies However, both health behavior and socio-demographic
for some other variables also reached significance level factors reduced the effect with approximately 25%.
(diabetes, angina, two gastro-intestinal symptoms, re- ‘‘Being physically tired at the end of the day’’ (app-
duced vision, and cigarette smoking) in some of the roximately 60% reduction), ‘‘economic hardship’’
models. However, in general, here the p-values were (approximately 40% reduction), ‘‘musculo-skeletal
higher and there were no or only minor differences complaints’’ (approximately 25% reduction), and
(on average 1.7%) in betas between the models with ‘‘gastro-intestinal complaints’’ (approximately 25%
dummies and the corresponding models without reduction) were the single covariates that weakened
dummies. the associations most.
1342 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345

Table 4
The influencea of the covariates (linear regression analyses) on the association between educational level and anxiety/depression
Covariates in the models Mixed anxiety-depression Anxiety Depression
Betab 95 % CI Betab 95 % CI Betab 95 % CI
Unadjusted 0.092 0.082e0.103 0.033 0.023e0.043 0.125 0.115e0.135
Age and sex 0.084 0.073e0.094 0.057 0.047e0.068 0.088 0.078e0.099
Somatic health
Musculo-skeletal complaints 0.062 0.051e0.072 0.037 0.026e0.047 0.070 0.060e0.080
Cardio-vascular disease 0.083 0.072e0.093 0.057 0.046e0.067 0.088 0.077e0.098
Diabetes 0.083 0.073e0.094 0.057 0.047e0.068 0.088 0.078e0.099
Gastro-intestinal complaints 0.064 0.054e0.075 0.039 0.029e0.049 0.073 0.063e0.083
Impairment of daily functioning 0.076 0.066e0.087 0.051 0.041e0.062 0.082 0.072e0.092
All somatic 0.048 0.038e0.058 0.024 0.014e0.034 0.058 0.048e0.068
Health behavior
Smoking 0.072 0.061e0.082 0.046 0.035e0.056 0.079 0.069e0.089
Alcohol consumption 0.086 0.075e0.096 0.062 0.052e0.073 0.087 0.077e0.098
Body mass index (kg/m2) 0.082 0.072e0.093 0.059 0.048e0.070 0.085 0.074e0.095
Physical activity 0.074 0.063e0.085 0.053 0.042e0.063 0.075 0.065e0.086
All health behaviors 0.063 0.053e0.074 0.046 0.035e0.057 0.063 0.053e0.074
Work-related factors
Lost job 0.067 0.056e0.077 0.043 0.032e0.053 0.101 0.090e0.111
Considered to change job 0.091 0.080e0.101 0.065 0.054e0.075 0.053 0.042e0.064
Physically tired 0.035 0.024e0.045 0.014* 0.003e0.025 0.045 0.034e0.056
Exhausted due to concentration 0.106 0.096e0.116 0.081 0.070e0.091 0.105 0.095e0.116
Satisfied with job 0.067 0.057e0.078 0.043 0.033e0.054 0.073 0.063e0.083
All work related factors 0.073 0.062e0.084 0.056 0.045e0.067 0.073 0.062e0.083
Socio-demographic factors
Economic hardship 0.052 0.041e0.062 0.027 0.016e0.037 0.061 0.051e0.071
Enough friends 0.095 0.085e0.105 0.067 0.057e0.077 0.099 0.089e0.109
Marital status 0.083 0.072e0.094 0.057 0.046e0.067 0.088 0.078e0.099
Urban-rural gradient 0.081 0.071e0.092 0.057 0.046e0.067 0.085 0.075e0.095
All socio-demographic factors 0.065 0.054e0.075 0.039 0.028e0.049 0.073 0.064e0.083

All covariates 0.045 0.034e0.055 0.034 0.023e0.045 0.044 0.034e0.055

The HUNT 2 study (N ¼ 49,048).
* p ¼ 0.015. p-values for all other betas were <0.001.
a
The influence of the covariate(s) should be read as the difference in the standardized regression coefficients of educational level between the
model adjusted for age and sex only, and the model including the covariate(s).
b
All the standardized regression coefficients (betas) are for the association between educational level and the anxiety/depression measure
(ln-transformed), adjusted for the covariate in question.

Comparisons with other studies follow-up time 1 year), Miech et al. (1999) (N ¼ 939
and follow-up time 6 years), and Nærde, Tambs, &
Our findings concerning depression and anxiety are Mathiesen (2002) (N ¼ 682 and follow-up time 3 years)
in accordance with the meta-analysis of Lorant et al. found an accumulating effect of educational level on
(2003), and the systematic review of Fryers, Melzer, & depression. The discrepancies between most of these
Jenkins (2003), which both included mainly cross- studies and our longitudinal results might be due to their
sectional studies. Contrary to the findings from our much smaller sample sizes and/or shorter follow-up
cross-sectional analyses, (but in line with our longi- time. If the true accumulated protective effect of
tudinal results) one of the included studies (Miech & a high educational level on anxiety and depression is
Shanahan, 2000) reported increasing associations be- quite moderate, as is implied by our results, it could
tween depression and educational level with increasing be difficult to observe in such studies.
age. Among the longitudinal studies, only Kaplan et al. Miech & Shanahan (2000) found that somatic
(1987) (N ¼ 4864 and follow-up time 9 years), in con- health problems accounted for most of the age-dependent
trast to Eaton et al. (2001) (N ¼ 693 and follow-up differences in depression between educational levels. In
time 15 years), Bracke (2000) (N ¼ 2223 and follow- our models, somatic health problems seemed to be the
up time 3 years), Sargeant et al. (1990) (N ¼ 423 and most important mediator as well, though to a lesser
 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345 1343

degree. Yet, physical disease is seldom considered in level most often restricts an individual’s occupational
such studies (Lorant et al., 2003). options to various forms of manual work. After some
years of one-sided physical strain, there is a risk of
Possible explanations for the observed relationships developing musculo-skeletal problems and general
between education and anxiety/depression fatigue.

The protective effect of educational level on mental Strengths and limitations of the study
health that attenuated, rather than accumulated with age
in the cross-sectional analyses, might reflect relation- The very large samples were population-based and
ships that differ from those reflected by the longitudinal covered a wide age range. Information regarding both
results. Fig. 2 reveals the dramatic development in edu- somatic and mental health as well as health behaviors
cational level from the oldest through the youngest age and work-related and socio-demographic factors was
groups, illustrating the profound changes that have available, and the follow-up period was relatively long.
taken place during the latter decades in Norwegian so- Educational level, as the main ‘‘exposure’’ variable
ciety. For the oldest part of the sample, family SES was established for most individuals as approximately
was probably the deciding factor in whether they would constant before the measure of possible mediators and
attain a lower or a higher education. With the develop- outcome variables.
ment of the Norwegian welfare state, a higher educa- Unlike previous studies, the use of the EHADS en-
tional level became more available regardless of abled us to study the effect of education not only on
economic ability, and, therefore, characteristics such symptoms of depression, but on anxiety as well. Though
as intelligence or specific abilities now play a greater there was no established measure of anxiety and depres-
role in educational choices. Heath et al. (1985) demon- sion in HUNT 1, the ADI should qualify as a satisfactory
strated in a large Norwegian twin sample that family en- substitute due to the high correlation with EHADS.
vironment was much more important, and genetic effect A possible selection bias in the longitudinal sample
less important for educational level among people born indicated by lower education and poorer mental health
in 1915 than among people born in 1960. It is likely that among non-attendees could have somewhat attenuated
the correlation between mental resources and educa- the true association between education and anxiety/
tional level is higher among the younger than among depression.
the older age groups. Perhaps such mental resources The most extensive substitutions of missing data
also give higher resilience in regards to strain or were performed for the work-related variables. On aver-
stresses, hence protecting against anxiety and depres- age 21.3% of the values were missing on each of the
sion. This cohort effect may have accounted for the four items. This proportion corresponds well with the
weakening of the association between educational level proportion of people in Norway in the age 25e69 years
and anxiety/depression with increasing age in the cross- that do not take part of the labor market due to old-age
sectional study. pension, disability pension, or unemployment. The
The stronger associations between educational level highly significant regression coefficients of the dummy
and the anxiety/depression measures in the age group of variables identifying the missing values of these vari-
65e75 years than in their adjacent age groups could be ables indicate that the role of these variables as media-
related to World War II, when Norway was occupied by tors should be interpreted with some caution. Most
Germany. Belonging to a higher social stratum, likely, the reduced variance for the predictor (mediator)
associated with taking a higher education, might have variables after mean substitution of missing data would
been protective to the general strain and hardship shrink the effect of these predictors, and hence, artifi-
most people experienced during those years (Grzywacz, cially reduce the observed mediator effect. However, in-
Almeida, Neupert & Ettner, 2004). In the years 1940e troducing the dummy variables for missing data besides
1945 individuals in this age group were between 8 and of the original mediator variables will probably produce
24 years old. mediator effects (in terms of reduced effect of education
The stepwise analyses including different sets of after adjusting for these variables) closer to what would
covariates indicated that somatic health problems, in be expected in a complete data set. The other dummy
particular musculo-skeletal and gastro-intestinal com- variables for missing data that were significant in the
plaints, and having a job that includes physically hard models did not notably change the coefficients of edu-
labor, might mediate some of the observed effect of cational level in the models. In addition, the substitu-
education on anxiety and depression. A low educational tions for these variables were not very extensive.
1344 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345

The possibility that the educational level was af- Fryers, T., Melzer, D., & Jenkins. (2003). Social inequalities and the
fected by preceding symptoms of anxiety or depression common mental disorders: a systematic review of the evidence.
Social Psychiatry and Psychiatric Epidemiology, 38(5), 229e237.
(the selection theory, social mobility) cannot be ruled Geyer, S., Hemström, Ö., Richard, P., & Vågerö, D. (2006). Educa-
out in the cross-sectional analyses. However, the down- tion, income and occupational class cannot be used interchange-
ward age restriction makes it probable that the partici- ably in social epidemiology. Empirical evidence against an
pants had completed their education when the data unquestioned practice. Journal of Epidemiology and Community
were recorded. Further, the results from the analyses Health, 60, 804e810.
Grzywacz, J. G., Almeida, D. M., Neupert, S. D., & Ettner, S. L.
of the longitudinal sample indicate a separate effect of (2004). Socioeconomic status and health: a micro-level analysis
educational level on anxiety/depression symptoms that of exposure and vulnerability to daily stressors. Journal of Health
increases with age. The cross-sectional design also de- and Social Behavior, 45(1), 1e16.
limits the interpretability of the examination of media- Heath, A. C., Berg, K., Eaves, L. J., Solaas, M. H., Corey, L. A., &
tor effects because the causal directions are assumed Sundet, J., et al. (1985). Education policy and the heritability of
educational attainment. Nature, 314(6013), 734e736.
rather than empirically established. Still, the single me- Holmen, J., Midthjell, K., & Bjartveit, K. (1990). The Nord-
diators that showed the strongest influence, i.e. eco- Trøndelag Health Survey 1984e1986. Purpose, background
nomic hardship, musculo-skeletal and gastro-intestinal and methods. Participation, non-participation and frequency
complaints, and work-related physical tiredness, were distributions. Verdal, Norway: Statens Institutt for Folkehelse,
plausible indirect consequences, or successors, of edu- Senter for samfunnsmedisinsk forskning.
Holmen, J., Midthjell, K., Krüger, Ø., Langhammer, A.,
cational level. Somatic complaints are often regarded Holmen, T. L., & Bratberg, G. H., et al. (2003). The Nord-Trøn-
as consequences of both anxiety and depression, but delag Health Study 1995e1997 (HUNT 2): objectives, contents,
in our models such a pathway would not influence the methods and participation. Norsk Epidemiologi, 13, 19e32.
association between educational level and anxiety HUNT. (2007). The Nord-Trøndelag Health Study: Verdal, Norway.
and/or depression. Available from. http://www.hunt.ntnu.no/index_nyforside.php?
side=english.
Our study supports the notion that higher educational Kaplan, G. A., Roberts, R. E., Camacho, T. C., & Coyne, J. C.
level, or the factors that are reflected by higher educa- (1987). Psycho-social predictors of depression. Prospective evi-
tion, may protect against anxiety and depression, and dence from the human population laboratory studies. American
the protective effect seems to accumulate throughout Journal of Epidemiology, 125, 206e220.
life. The mechanisms of the protective effect may in- Kessler, R. C., McGonagle, K. A., Zhao, S., Nelson, C. B.,
Hughes, M., & Eshleman, S., et al. (1994). Lifetime and 12-
volve factors such as personal characteristics related month prevalence of DSM-III-R psychiatric disorders in the
to various levels of resilience to stress and to differences United States. Results from the National Comorbidity Survey.
in occupational restrictions, which may be associated Archives of General Psychiatry, 51(1), 8e19.
with level of stress exposure and somatic health. Krokstad, S., Kunst, A. E., & Westin, S. (2002). Trends in health inequal-
ities by educational level in a Norwegian total population study.
Journal of Epidemiology and Community Health, 56, 375e380.
References Lorant, V., Deliège, D., Eaton, W., Robert, A., Philippot, P., &
Ansseau, M. (2003). Socioeconomic inequalities in depression:
Anderson, A. B., Basilevsky, A., & Hum, D. P. J. (1983). Missing a meta-analysis. American Journal of Epidemiology, 157, 98e112.
data: A review of the literature. In Rossi., Wright., & Anderson. Miech, R. A., Caspi, A., Moffitt, T. E., Wright, B. R. E., &
(Eds.), Handbook of survey research. New York: Academic Press. Silva, P. A. (1999). Low socioeconomic status and mental disor-
Andrews, G., Henderson, S., & Hall, W. (2001). Prevalence, comor- ders: a longitudinal study of selection and causation during young
bidity, disability and service utilization. Overview of the Austra- adulthood. The American Journal of Sociology, 104, 1096e1131.
lian National Mental Health Survey. British Journal of Miech, R. A., & Shanahan, M. J. (2000). Socioeconomic status and
Psychiatry, 178, 145e153. depression over life course. Journal of Health and Social Behav-
Bijl, R. V., Ravelli, A., & van Zessen, G. (1998). Prevalence of psychi- ior, 41, 162e176.
atric disorder in the general population: results of The Netherlands Mirowsky, J., & Ross, C. E. (1998). Education, personal control, life-
Mental Health Survey and Incidence Study (NEMESIS). Social style and health. A human capital hypothesis. Research on Aging,
Psychiatry and Psychiatric Epidemiology, 33(12), 587e595. 20, 415e449.
Bjelland, I., Dahl, A. A., Haug, T., & Neckelmann, D. (2002). The Mykletun, A., Stordal, E., & Dahl, A. A. (2001). The Hospital
validity of the Hospital Anxiety and Depression Scale. An up- Anxiety and Depression Scale (HADS): factor structure, item
dated review. Journal of Psychosomatic Research, 52, 69e77. analyses, and internal consistency in a large population. British
Bracke, P. (2000). The three-year persistence of depressive symptoms Journal of Psychiatry, 179, 540e544.
in men and women. Social Science & Medicine, 51, 51e64. Nærde, A., Tambs, K., & Mathiesen, K. S. (2002). Child related
Eaton, W. W., Muntaner, C., Bovasso, G., & Smith, C. (2001). strain and maternal mental health: a longitudinal study. Acta
Socioeconomic status and depressive syndrome: the role of Psychiatrica Scandinavia, 105, 301e309.
inter- and intra-generational mobility, government assistance, Paul, C., Mason, W. M., McCaffrey, D., & Fox, S. A. (2003). What
and work environment. Journal of Health and Social Behavior, should we do about missing data? (A case study using logistic re-
42, 277e294. gression with missing data on a single covariate). [California
 I. Bjelland et al. / Social Science & Medicine 66 (2008) 1334e1345 1345

Center for Population Research (CCPR): On-Line Working Paper depression in the community. Archives of General Psychiatry,
Series]. Los Angeles: University of California. 47, 519e526.
Ross, C. E., & Mirowsky, J. (2006). Sex differences in the effect of Statistics Norway. (1994). Standard classification of municipalities.
education on depression: resource multiplication or resource Oslo, Norway: Statistics Norway.
substitution? Social Science & Medicine, 63(5), 1400e1413. Zigmond, A. S., & Snaith, R. P. (1983). The hospital anxiety and
Ross, C. E., & Wu, C.-L. (1996). Education, age, and the cumulative depression scale. Acta Psychiatrica Scandinavia, 67, 361e370.
advantage in health. Journal of Health and Social Behavior, Zimmerman, F. J., Christakis, D. A., & Vander Stoep, A. (2004). Tin-
37(1), 104e120. ker, tailor, soldier, patient: work attributes and depression dispar-
Sargeant, J. K., Bruce, M. L., Florio, L. P., & Weissman, M. M. ities among young adult. Social Science & Medicine, 58(19),
(1990). Factors associated with 1-year outcome of major 1889e1901.