American Journal of Epidemiology Vol. 191, No.

8
© The Author(s) 2022. Published by Oxford University Press on behalf of the Johns Hopkins Bloomberg School of https://doi.org/10.1093/aje/kwac077
Public Health. This is an Open Access article distributed under the terms of the Creative Commons Attribution
License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted reuse, distribution, and Advance Access publication:
reproduction in any medium, provided the original work is properly cited. April 20, 2022

Original Contribution

Current Unemployment, Unemployment History, and Mental Health: A

Fixed-Effects Model Approach

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

Liina Junna∗, Heta Moustgaard, and Pekka Martikainen
∗ Correspondence to Liina Junna, Population Research Unit, Faculty of Social Sciences, University of Helsinki, P.O. Box 18,
FIN-00014 University of Helsinki, Helsinki, Finland (e-mail: liina.junna@helsinki.fi).

Initially submitted August 31, 2021; accepted for publication April 14, 2022.

Poor mental health among the unemployed—the long-term unemployed in particular—is established, but
these associations may be driven by confounding from unobserved, time-invariant characteristics such as
past experiences and personality. Using longitudinal register data on 2,720,431 residents aged 30–60 years,
we assessed how current unemployment and unemployment history predict visits to specialized care due to
psychiatric conditions and self-harm in Finland in 2008–2018. We used linear ordinary-least-squares and fixed-
effects models. Prior to adjusting for time-invariant characteristics, current unemployment was associated with
poor mental health, and the risk increased with longer unemployment histories. Accounting for all time-invariant
characteristics with the fixed-effects models, these associations attenuated by approximately 70%, yet current
unemployment was still associated with a 0.51 (95% confidence interval: 0.48, 0.53) percentage-point increase
in the probability of poor mental health among men and women. Longer unemployment histories increased
the probability among men in their 30s but not among older men or among women. The results indicate that
selection by stable characteristics may explain a major part of the worse mental health among the unemployed and
especially the long-term unemployed. However, even when controlling for this selection, current unemployment
remains associated with mental health.

hospitalization; longitudinal study; mental health; unemployment

Abbreviations: CI, confidence interval; FE, fixed effects; LPM, linear probability model; OLS, ordinary least squares.

Unemployment is associated with poor mental health. The a complex interplay between health and unemployment
unemployed—the long-term unemployed in particular (1– (18).
3)—exhibit higher levels of distress, psychiatric symptoms, A large share of the observational studies that suggest
and self-harm than the employed (1–11). However, it unemployment is detrimental to mental health have dealt
remains unclear to what extent the relationship between with selection and confounding by adjusting for preexisting
unemployment and mental health is causal. Unemployment mental health problems and various other personal and
could harm mental health through mechanisms such as contextual characteristics associated with unemployment
increased stress following the loss of income, social con- and mental health (2, 5, 6). Some of these confounders—
tacts, and meaningful activities associated with employment such as age, marital dissolution, and the labor market
(1, 6). These hardships also accumulate over time, which context (best measured as time-varying) and maternal
may explain why longer periods of unemployment have language and baseline education (time-invariant)—are
been found to be more detrimental to health than shorter relatively easy to address (2, 3, 5, 6). However, recent studies
spells (1–3). On the other hand, poor mental health may have suggested that a notable share of the association could
increase both the risk of becoming unemployed and longer be due to confounding by stable individual characteristics
unemployment spells through health-related selection (12– such as health histories (1), genetic background and
17). These explanations are not mutually exclusive. The childhood conditions (19–23), early career experiences (24,
poorer mental health among the unemployed likely reflects 25), and personality (26–30). These complex characteristics

1459 Am J Epidemiol. 2022;191(8):1459–1469

1460 Junna et al.

are demanding to measure accurately and therefore to fully sitive measures. We additionally explored heterogeneity by
control for. age and sex.
Some studies have used the fixed-effects (FE) approach
to tackle time-invariant confounders (31–38). In FE models,
METHODS
the individual is compared with him- or herself over multiple
measurement points with differing exposures, which Data
removes all confounding by time-invariant characteristics.
The results from these studies are inconsistent: Unemploy- The quarterly individual-level data on all Finnish resi-
ment has been found to be detrimental among the general dents aged 30–60 years and in the labor market (employed

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

population (36, 37), only among subpopulations unem- or unemployed) at some point in the years from 2008–
ployed due to reasons other than workplace closure (32), 2018 (n = 2,720,431) came from the population register of
or those below the age of 30 years (31, 33, 38). Variability Statics Finland. We excluded all individuals aged below 30
in these results is likely to relate to measurement or method- years as the share of students in these ages is large (39),
ological choices, adjusted time-varying characteristics, or and we could not reliably identify students. We identified
differences in study context and policy environments. all unemployment spells using onset dates from the Labour
Moreover, FE studies have primarily addressed current Market Data File of Statistics Finland. The included health
unemployment while ignoring past experiences (31, 32). conditions were derived from the Care Register for Health
This is a notable gap in the literature. Observational evidence Care of the National Institute for Health and Welfare and
suggests that longer periods of unemployment are more the Death Records of Statistics Finland. Using the personal
strongly related to mental health (1–3). However, complex identification code that all Finnish residents receive, we then
time-invariant characteristics such as personality are partic- linked these data, along with sociodemographic information
ularly important for selection into long-term unemployment from Statistics Finland.
and may therefore explain some of the prior findings (30, 31, The focus of this study was comparing the unemployed
36). Furthermore, little is known about the contribution of with the employed, and we therefore excluded quarter-year
time-invariant confounding across important demographic observations outside the labor market (n = 6,560,568). Of
dimensions such as sex and age. In general, the effects all the excluded observations, 50% were in retirement, 15%
of unemployment may be stronger among men, for whom were in education, and 35% were in other activities, such
employment is a more central social role (2, 36). It has also as taking care of the home. Consequently, individuals who
been suggested that unemployment could be more harmful were never in the labor market during the study period were
to the middle-aged, with more financial responsibilities and excluded (n = 202,477). We also excluded observations with
stronger career commitments (5). Only a small number of FE missing data on any of the measures, which led to excluding
studies have addressed unemployment histories or whether 12,164 (0.01%) of all observations. Individuals exited the
time-invariant confounders contribute similarly among men sample by aging out, by migrating, by leaving the labor force
and women and different age groups (31, 36). The prior permanently, or at death.
evidence is based on different outcomes and populations,
and the results are conflicting. Mental health scores deteri- Outcome
orated with cumulative exposure to unemployment among
Germans aged 30 or younger but not those older than 30 The outcome (Yt) was a binary measure of whether the
(men and women pooled together) (31). In turn, among Nor- individual experienced a mental health outcome (yes/no)
wegians aged 18–67 years, impeding unemployment was over the quarter-year observation period (t). We considered
associated with increased odds of initiating psychotropic visits to specialized care (booked and emergency inpatient
medications, and the risk then remained elevated during and outpatient care visits) due to psychiatric conditions
the first months of unemployment, increasing slightly at (International Classification of Diseases, Tenth Revision
4 months (36). The estimates were slightly higher among (ICD-10), main diagnosis, codes F01–F69) and self-harm
men and did not differ between age groups. Whether time- (ICD-10, X60–X84) as proxies for poor mental health. As
invariant confounding plays a larger role among those above it is possible for a person to die due to mental health–
the age of 30 is therefore still unclear, as are the reasons related causes before reaching a hospital (e.g., suicides) we
behind this possible age heterogeneity. also included deaths related to the same causes (based on
We contribute to the literature by considering both current underlying cause of death).
unemployment status and unemployment history as predic-
tors of mental health. We quantified the total contribution of Exposures
time-invariant confounding by using the within-individual
FE design (34). We used data on the total Finnish working- Current unemployment was measured at t – 1 and was
age population in 2008–2018. High-quality general popula- based on being registered as an unemployed jobseeker for
tion data with long follow-up periods and objective mental a minimum of 2 days at any point during the quarter-year
health outcomes have previously been underutilized for this observation period (yes/no). Spells shorter than 2 days were
purpose (36). Register data are a notable strength when not considered as unemployment because these are likely
studying unemployment and mental health, as they do not to represent workplace transitions. Unemployment history
suffer from problems related to loss to follow-up among was measured as the cumulative number of observations
individuals with poor mental health, or to misreporting sen- with unemployment in the 3 years preceding t − 1. This

Am J Epidemiol. 2022;191(8):1459–1469
 Unemployment and Mental Health 1461

was included in the models as linear and squared continuous group, we further estimated an interaction between current
measures, to allow for nonlinear associations (2, 31). unemployment and unemployment history for 10-year age
groups, for whom we then calculated predicted probabilities.
We fitted 2 interaction models: an ordinary-least-squares
Covariates
(OLS) model adjusting for year, quarter, and birth year
To adjust for time and age trends in visits to specialized (model 2 plus the interaction between current unemploy-
care and unemployment, we used both birth year and cal- ment and unemployment history) and a FE model adjusting
endar year as continuous measures and quarter as a binary for year, quarter, partnership status, and coresident children
(model 4 plus the interaction between current unemploy-

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

measure; year and quarter were included as time-varying.
A known protective factor is being in a relationship, as a ment and unemployment history). The reference category
partner may provide social and financial support as well was employed with no history of unemployment.
as social control (1, 5). Partnership status was registered We estimated heteroscedasticity-consistent standard
annually on December 31. We included it as a time-varying errors for all models (40). None of the models produced
categorical measure (married or cohabiting, single, divorced, predicted probabilities outside the unit interval (0,1). Stata,
or widowed) for the preceding year. We further controlled for version 16.1 (StataCorp LLC, College Station, Texas), was
having coresident children (yes/no), also measured annually used for all analyses.
and included as time-varying.
Sensitivity analyses
Statistical analysis To ascertain that our results were robust, we tested other
We estimated linear probability models (LPMs, ordinary- cutoffs for current unemployment and unemployment his-
least-squares models for binary outcomes) for the association tory (Web Table 1, available at https://doi.org/10.1093/aje/
between quarterly current unemployment, unemployment kwac077) and fitted the models using logistic FE (Web
history, and visits to specialized care. We modeled the asso- Table 2).
ciation both with and without controlling for the unobserved,
time-invariant characteristics using the FE estimator (34). RESULTS
As FE models rely solely on within-individual variance,
this removes all between-individual contribution from any Table 1 describes the characteristics of the analytical sam-
unchanging characteristics. The individual is used as his or ple. The maximum follow-up was 10 years (mean = 7.5).
her own control to compare deviations from the mean of Men were unemployed 13.0% of the quarter-year obser-
unemployment status with deviations from the mean of the vations and women 11.9%, while less than 1% of men
health outcome. The LPM-FE results can then be interpreted and 1.3% of women had a mental health-related visit to
as the difference in the probability of the outcome among specialized care (Table 1).
study subjects during periods of unemployment compared Considered separately, current unemployment and a
with periods in employment. longer unemployment history were detrimental to mental
We fitted all models for men and women separately. The health among both men and women (Table 2, Table 3;
LPM model 1 was modeled separately for current unem- model 1): The probability of a visit to specialized care
ployment and unemployment history, and adjusted for the due to psychiatric conditions and self-harm was 2.21 (95%
calendar year, quarter, and year of birth to account for the confidence interval (CI): 2.18, 2.25) percentage points
initial age. In model 2, we simultaneously included current higher among men during current unemployment and 1.98
unemployment and unemployment history, adjusting for the (95% CI: 1.94, 2.02) higher among women. Each additional
same covariates as in model 1. The LPM-FE model 3 con- quarter of unemployment history was also associated
trolled for all time-invariant characteristics by design, and with an approximately 0.3 percentage-point increase in
we thus excluded observed birth year from the model and the probability of a visit among both men and women.
again included current unemployment and unemployment Considering current unemployment and unemployment
history together. In the final model 4, we further adjusted for history together attenuated the individual impact of both
time-varying partnership status and having children. Model unemployment measures (Table 2, Table 3; model 2).
4 can then be presented as Controlling for all time-invariant factors (model 3), the
association between current unemployment and visits to spe-
cialized care attenuated by approximately 70%. However,
Yit = β1 CURRENT_UNEMPLOYMENTit current unemployment still increased the probability of a
+β2 UNEMPLOYMENT_HISTORYit visit by 0.51 (95% CI: 0.48, 0.53) percentage points among
+ β3 UNEMPLOYMENT_HISTORY 2it men and 0.51 (95% CI: 0.48, 0.53) among women. Including
+β4 YEARit + β5 QUARTERit time-varying partnership status and having children (model
+ β6 PARTNERSHIPit + β7 CHILDRENit + ai + εit 4) had little effect on the estimates.
We observed an interaction effect between current unem-
where α is the fixed but unknown individual-level intercept, ployment and unemployment history (Figure 1; Web Table
and ε denotes the error terms (35). 3). In the OLS models (Figure 1A and 1C) the combination
To estimate the probability of poor mental health for of current unemployment preceded by a longer unemploy-
different levels of unemployment history by sex and age ment history was particularly harmful among men and,

Am J Epidemiol. 2022;191(8):1459–1469
1462 Junna et al.

Table 1. Sample Characteristics, Men and Women in the Labor Force, Aged 30–60 Years, Finland, 2008–2018

Men Women
Characteristic
% of % of
No. No.
Observations Observations

No. of individuals 1,384,596 1,335,835

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

Quarterly observations 41,908,188 40,428,174
Age, yearsa 44.9 (8.9) 45.4 (8.9)
Married or cohabiting 54.0 56.6
Has coresident children 51.4 56.1
Unemployed 13.0 11.9
Unemployment history, quartersa 1.3 (3.0) 1.2 (2.8)
Visits to specialized care due to 380,229 0.9 542,445 1.3
psychiatric conditions and self-harm

a Values are expressed as mean (standard deviation).

up to approximately 24 months, also women. However, In sensitivity analyses, our results were robust to different
unmeasured confounding due to time-invariant characteris- cutoff points of current unemployment and unemployment
tics appears to explain a substantial share of the impact of history (Web Table 1), as well as for model specification
both current unemployment and unemployment history: In (Web Table 2).
the full FE models, the association between current unem-
ployment preceded by a long unemployment history and DISCUSSION
visits to specialized care strongly attenuated among men and
disappeared among women (Figure 1B and 1D). An extensive literature has demonstrated the associa-
The overall picture varied by age. In the OLS models, tion between unemployment and poor mental health (1–
current unemployment preceded by a long unemployment 11). Whether this association is driven by confounding from
history was particularly detrimental among men of all time-invariant characteristics is important for understanding
ages—but more so among the younger age groups—and the determinants of mental health as well as the association
women in their 30s and 40s (Figures 2–3; Web Tables 4–5). between unemployment and health. We analyzed unem-
In the full FE models, this was observed only among men in ployment and visits to specialized care due to psychiatric
their 30s. conditions and self-harm using longitudinal register data

Table 2. Change in the Probability of Visits to Specialized Care Due to Psychiatric Conditions and Self-Harm Among Men Aged 30–60 Yearsa ,
Finland, 2008–2018

Model 1b Model 2c Model 3d Model 4e

Exposure
Coefficient 95% CI Coefficient 95% CI Coefficient 95% CI Coefficient 95% CI

Employed 0.00 0.00 0.00 0.00

Unemployed 2.21 2.18, 2.25 1.51 1.48, 1.54 0.51 0.48, 0.53 0.50 0.48, 0.53
Unemployment history, quarters 0.29 0.28, 0.30 0.13 0.11, 0.14 −0.01 −0.01, 0.00 −0.01 −0.02, 0.00
Unemployment history2 −0.01 −0.01, −0.00 −0.00 −0.00, −0.00 0.00 0.00, 0.00 0.00 0.00, 0.00
ρ 46.54 46.49

Abbreviation: CI, confidence interval.

a Number of observations = 43,546,725; results presented in percentage points. Model estimates multiplied by 100.
b Ordinary least squares, current unemployment, and unemployment history modeled separately. Adjusted for year, quarter, and birth year.
c Ordinary least squares, current unemployment, and unemployment history modeled together. Adjusted for year, quarter, and birth year.
d Fixed effects, year, and quarter.
e Fixed effects, model 3 with the addition of partnership status and coresident children.

Am J Epidemiol. 2022;191(8):1459–1469
 Unemployment and Mental Health 1463

Table 3. Change in the Probability of Visits to Specialized Care Due to Psychiatric Conditions and Self-Harm Among Women Aged 30–60
Yearsa , Finland, 2008–2018

Model 1b Model 2c Model 3d Model 4e

Exposure
Coefficient 95% CI Coefficient 95% CI Coefficient 95% CI Coefficient 95% CI

Employed 0.00 0.00 0.00 0.00

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

Unemployed 1.98 1.94, 2.02 1.47 1.44, 1.51 0.51 0.48, 0.53 0.51 0.48, 0.53
Unemployment history, 0.31 0.30, 0.32 0.16 0.14, 0.17 −0.04 −0.05, −0.03 −0.04 −0.05, −0.03
quarters
Unemployment history2 −0.01 −0.01, −0.01 −0.01 −0.01, −0.01 0.00 0.00, 0.00 0.00 0.00, 0.00
ρ 47.60 47.58

Abbreviation: CI, confidence interval.

a Number of observations = 42,393,875; results presented in percentage points. Model estimates multiplied by 100.
b Ordinary least squares, current unemployment, and unemployment history modeled separately. Adjusted for year, quarter, and birth year.
c Ordinary least squares, current unemployment, and unemployment history modeled together. Adjusted for year, quarter, and birth year.
d Fixed effects, year, and quarter.
e Fixed effects, model 3 with the addition of partnership status and coresident children.

of the total population of economically active Finns. In age- and sex-related heterogeneity. We used a fixed-effects
addition to addressing current unemployment, we assessed design to control for all time-invariant characteristics such
3-year unemployment history to better understand the role as personality, past experiences, and initial health status,
of cumulative unemployment, and we further considered that may confound the association between unemployment

A) B)
0.035 0.035
Predicted Probability

Predicted Probability

0.025 0.025

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

C) D)
0.035 0.035 Employed
Unemployed
Predicted Probability

Predicted Probability

0.025 0.025

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

Figure 1. Predicted probability of visits to specialized care due to psychiatric conditions and self-harm among men and women aged 30–60
years, Finland, 2008–2018. Results are derived from an ordinary-least-squares (OLS) model adjusted for year, quarter, and birth year and a
fixed-effects (FE) model additionally adjusted for year, quarter, partnership status, and coresident children: A) men, OLS model; B) men, FE
model; C) women, OLS model; D) women, FE model.

Am J Epidemiol. 2022;191(8):1459–1469
1464 Junna et al.

A) B)
0.055 0.055

Predicted Probability

Predicted Probability
0.045 0.045

0.035 0.035

0.025 0.025

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

C) D)
0.055 0.055
Predicted Probability

Predicted Probability
0.045 0.045

0.035 0.035

0.025 0.025

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

E) F)
0.055 0.055 Employed
Unemployed
Predicted Probability

Predicted Probability

0.045 0.045

0.035 0.035

0.025 0.025

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

Figure 2. Predicted probability of visits to specialized care due to psychiatric conditions and self-harm among men according to 10-year age
group, Finland, 2008–2018. Results are derived from an ordinary-least-squares (OLS) model adjusted for year, quarter, and birth year, and a
fixed-effects (FE) model additionally adjusted for year, quarter, partnership status, and coresident children: A) OLS model, ages 30–39 years;
B) FE model, ages 30–39 years; C) OLS model, ages 40–49 years; D) FE model, ages 40–49 years; E) OLS model, ages 50–60 years; F) FE
model, ages 50–60 years.

and health. To our knowledge, this is the first FE study on associations were modest: In the full models, current unem-
unemployment histories and mental health–related visits to ployment was associated with a 0.50 (95% CI: 0.48, 0.53)
specialized care. percentage-point increase in poor mental health among men
We show that a large part—approximately 70%—of the and 0.51 (95% CI: 0.48, 0.53) percentage-point increase
association between current unemployment and mental among women. However, the outcome is rare, and while
health was due to confounding from time-invariant charac- the absolute difference was small, the relative difference
teristics. Yet current unemployment among men and women was not. Even in the full models, the increase in the
and unemployment history among men were associated probability was 1.6-fold among unemployed men with no
with an increase in probability of poor mental health even unemployment history and 1.4-fold among corresponding
after controlling for this unmeasured confounding. The women when compared with the employed. In other words,

Am J Epidemiol. 2022;191(8):1459–1469
 Unemployment and Mental Health 1465

A) B)
0.055 0.055

Predicted Probability

Predicted Probability
0.045 0.045

0.035 0.035

0.025 0.025

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

C) D)
0.055 0.055
Predicted Probability

Predicted Probability
0.045 0.045

0.035 0.035

0.025 0.025

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

E) F)
0.055 0.055 Employed
Unemployed
Predicted Probability

Predicted Probability

0.045 0.045

0.035 0.035

0.025 0.025

0.015 0.015

0.005 0.005
0 6 12 18 24 30 36 0 6 12 18 24 30 36
Unemployment History, months Unemployment History, months

Figure 3. Predicted probability of visits to specialized care due to psychiatric conditions and self-harm among women according to 10-year
age group, Finland, 2008–2018. Results are derived from an ordinary-least-squares (OLS) model adjusted for year, quarter, and birth year, and
a fixed-effects (FE) model additionally adjusted for year, quarter, partnership status, and coresident children: A) OLS model, ages 30–39 years;
B) FE model, ages 30–39 years; C) OLS model, ages 40–49 years; D) FE model, ages 40–49 years; E) OLS model, ages 50–60 years; F) FE
model, ages 50–60 years.

insofar as our estimates can be considered causal, if Current unemployment, unemployment history, and the
unemployment increased by 100,000 men, we would expect role of causation and selection mechanisms
500 more mental health–related visits to specialized care,
and if long-term unemployment increased by 100,000 men, Based on prior observational evidence, unemployment
we would expect 790 more visits (Web Table 4). These lasting for 6 months or more could be more detrimental to
visits represent both adverse individual-level consequences mental health than are shorter spells (2, 5). This is plausible,
and care costs to society. Moreover, visits due to psychiatric as there are various material and psychosocial mechanisms
conditions and self-harm are likely to represent the tip of the through which longer exposure could be more damaging.
iceberg. Less severe mental health measures, such as well- Employment is an important source of income, and higher
being scales, may react more strongly to unemployment than income in turn allows for personal choices that promote
our fairly severe outcome (1). mental health, such as the use of health and leisure services.

Am J Epidemiol. 2022;191(8):1459–1469
1466 Junna et al.

Financial insecurity is also stressful (41), and financial the unemployed is larger in countries with a high proportion
resources—both objective and perceived—are important of nuclear family households and lower female workforce
predictors of mental health (42). The Finnish universal participation (46, 49). However, other mechanisms are also
unemployment insurance system may somewhat soften likely to be at play.
the financial consequences of unemployment (43), but job Employment is also a source of social contact, status,
loss is still typically followed by an immediate income and purposeful activities that promote mental health (52–
drop, and subsequently another drop among the long-term 54). The loss of these psychosocial benefits may be more
unemployed once they exhaust their earnings-related bene- detrimental among men in the long run, as it has previously
fits. During our study period, the more-generous earnings- been suggested that women’s social status, identity, and

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

related benefits provided by unemployment funds were activities are tied to multiple domains of life, while men’s
available to those eligible (e.g., employed for a minimum are more strongly linked to employment (55, 56). Another
of 26 weeks over the preceding 28) for a duration of plausible explanation is that unemployment history may also
approximately 300 to 400 days (43, 44). The less-generous measure fragmented careers, which may be more the norm
basic allowance was available for 500 days for those who among women and thus less damaging to women’s health.
were not members of an unemployment fund. Thus, due to a Moreover, we showed that current unemployment was asso-
gradual loss of benefits, financial hardship may accumulate ciated with poor mental health among both sexes, albeit with
among the long-term unemployed. Moreover, accumulating relative differences that were stronger among men, which
unemployment has been linked to lifestyle factors, such raises the question of why these suggested mechanisms
as poor diet quality, increased alcohol consumption, and would function similarly for current unemployment but not
reduced health service use (45, 46), which may adversely unemployment histories. More studies are needed to confirm
affect mental health. Long-term unemployment may also these findings and to address the processes behind them.
influence health through various other mechanisms such as Our results also suggest that the association between
increased family stress and instability (47). unemployment and mental health varies by age. In accor-
With all of this in mind, unmeasured confounding may dance with the literature (5, 31), we did not find unem-
explain some of the association between longer unemploy- ployment to be less damaging to younger individuals. To
ment spells and mental health (2, 5). Previous studies have the contrary, among men in their 30s, the combination of
suggested that stable, individual characteristics such as per- current unemployment and a longer unemployment history
sonality (low emotional self-control, passivity) are particu- appeared particularly harmful. Meanwhile, among all unem-
larly important for selection into long-term unemployment ployed women and unemployed men in their 50s, current
(30). We show that controlling for time-invariant character- unemployment but not unemployment history was harmful.
istics strongly attenuated but did not fully explain the asso- The mechanisms underlying this interplay between age
ciation between current unemployment and mental health. and sex are unclear. The young unemployed may have less
Our results partially support the health-related selection into savings or be less able to depend on a partner’s income than
long-term unemployment hypothesis: Controlling for selec- the older working-age unemployed (57–59). They may less
tion based on time-invariant characteristics fully explained often be eligible for earnings-related unemployment bene-
the worse mental health of women with long unemployment fits, leading to more rapid financial distress as unemploy-
histories and notably reduced the impact among men. Few ment cumulates (43, 44). Unemployment may also lead to
studies have assessed unemployment duration controlling future wage scarring (60) and hinder accumulating financial
for unobserved confounding, and therefore comparisons are assets (e.g., home ownership, savings) (61). Both men and
difficult to make. Norwegians who were unemployed for a women in their 30s are likely to be starting their careers
minimum of 90 days had a slight, gradual decline in their risk and families, but men are again more likely to be the main
of initiating psychotropic medication over the first months breadwinners (50, 51), and therefore all of the mechanisms
of unemployment, followed by a peak at 4 months or more, together may make younger men more vulnerable to accu-
which indicates a more harmful effect for longer spells of mulating financial distress while unemployed.
unemployment (n = 2,348,552) (36). In partial accordance
with our results, these differences were more notable among Methodological considerations
men than women.
This study is based on large-scale, quarterly, and high-
Variation by sex and age quality register data and a long follow-up. We did not rely
on self-reports, which may suffer from higher loss to follow-
The reasons behind the gendered impact of unemploy- up, particularly among people with poor mental health (62).
ment history are unclear. One possible explanation for this Studies based on surveys may also suffer from retrospective
relates to income, as men’s salaries often contribute more to biases of past mental health reporting (63, 64), while unem-
family income (2, 48). Even in the context of Finland and ployment histories may be both over and underreported (65).
Norway, 2 Nordic countries with a relatively high level of However, there are limitations related both to the mea-
female workforce participation (49), men are more likely to sures and the method in this study. While we analyzed clini-
be the primary breadwinners (50, 51). Accumulating unem- cally relevant health outcomes, less severe changes in mental
ployment is a central threat to material well-being, which health (e.g., less severe depression) as well as untreated
may be more damaging to the primary breadwinners. Corre- mental health problems were not captured by our data.
spondingly, the mental health gap between the employed and We assume these less severe measures would also reflect

Am J Epidemiol. 2022;191(8):1459–1469
 Unemployment and Mental Health 1467

changes in employment status, but further studies are needed Martikainen); Max Planck Institute for Demographic
to evaluate unbiased measures of less severe changes in men- Research, Rostock, Germany (Liina Junna, Pekka
tal health. For the same reason, even though our employment Martikainen); Helsinki Institute for Social Sciences and
measures temporally precede the health outcomes, it is pos- Humanities (HSSH), University of Helsinki, Finland (Heta
sible that we may have not detected prior health deterioration Moustgaard); and Centre for Health Equity Studies,
that may have contributed to becoming unemployed. Using University of Stockholm and Karolinska Institute,
a relatively severe measure of mental health, however, may Stockholm, Sweden (Pekka Martikainen).
also be an advantage as it is less likely to reflect differences This work was funded by the Academy of Finland
in seeking and access to care. (grants 308247 and 345219). P.M. has also received

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

As we were interested in comparing the unemployed funding from the European Research Council under the
with the employed, we focused on individuals in the labor European Union’s Horizon 2020 research and innovation
market. Our exposure measurement was based on registered programme (grant 101019329) and the NordForsk grant for
unemployment, and therefore we were unable to consider the project WELLIFE (grant 83540).
discouraged jobless individuals who were no longer regis- Following the data-protection regulations of the national
tered as unemployed. Moreover, it is unclear whether these register holders, the authors cannot make the data available
individuals should be considered unemployed or outside the to third parties. Interested researchers may contact the
labor market, and therefore outside the scope of this study. following register-holding public institutions: Statistics
Based on these or any other observational analyses of Finland (https://www.stat.fi/tup/mikroaineistot/index_en.
unemployment and mental health, we cannot claim that html) and the Social Insurance Institution of Finland
the association between unemployment and mental health (https://www.kela.fi/web/en/data-permits-and-data-
is causal. Although we effectively controlled for all time- requests).
invariant confounding (30), time trends, and time-varying The study does not necessarily reflect the European
partnership status and having children, unmeasured time- Commission’s views and in no way anticipates the
varying factors (e.g., changes in social support, behaviors) Commission’s future policy in this area. The funders had
may still confound the association between unemployment no role in the study design, data collection and analysis,
and mental health. In addition, the FE model assumes— decision to publish, or preparation of the manuscript.
somewhat unrealistically—that past health conditions do Conflict of interest: none declared.
not affect current unemployment or the time-varying con-
founders (35). With the FE approach, we also cannot model
dynamic relationships between unemployment and health.
Lagged health outcomes are endogenous, and we therefore
cannot include them in our model. In other words, there REFERENCES
is a trade-off between modeling a dynamic association and
adjusting for the unobserved, time-invariant characteristics 1. Jin RL, Shah CP, Svoboda TJ. The impact of unemployment
on health: a review of the evidence. CMAJ. 1995;153(5):
(35); in this paper, we have chosen to focus on the latter. 529–540.
2. McKee-Ryan F, Song Z, Wanberg CR, et al. Psychological
Conclusions and physical well-being during unemployment: a
meta-analytic study. J Appl Psychol. 2005;90(1):53–76.
We showed that a large part of the difference in mental 3. Milner A, Page A, LaMontagne AD. Long-term
health between the employed and the unemployed appears unemployment and suicide: a systematic review and
to be due to time-invariant confounding. However, the asso- meta-analysis. PloS One. 2013;8(1):e51333.
ciation is not entirely driven by confounding. Even after 4. Milner A, Page A, LaMontagne AD. Duration of
adjusting for all time-invariant characteristics, current unem- unemployment and suicide in Australia over the period
1985–2006: an ecological investigation by sex and age during
ployment was associated with a 0.5 percentage-point higher rising versus declining national unemployment rates.
probability of psychiatric conditions and self-harm among J Epidemiol Community Health. 2013;67(3):237–244.
both men and women. Among younger men, the deteri- 5. Paul KI, Moser K. Unemployment impairs mental health:
oration of mental health gained momentum with longer meta-analyses. J Vocat Behav. 2009;74(3):264–282.
unemployment histories. Selection mechanisms related to 6. Milner A, Page A, Lamontagne AD. Cause and effect in
stable characteristics are thus likely to explain a large part studies on unemployment, mental health and suicide: a
of the impact of current unemployment on mental health meta-analytic and conceptual review. Psychol Med. 2014;
among both men and women, as well as strongly drive the 44(5):909–917.
association between long-term unemployment and mental 7. Blakely TA, Collings SC, Atkinson J. Unemployment and
health, particularly among women. suicide. Evidence for a causal association? J Epidemiol
Community Health. 2003;57(8):594–600.
8. Browning M, Heinesen E. Effect of job loss due to plant
closure on mortality and hospitalization. J Health Econ.
2012;31(4):599–616.
ACKNOWLEDGMENTS 9. Eliason M, Storrie D. Job loss is bad for your
health—Swedish evidence on cause-specific hospitalization
Author affiliations: Population Research Unit, University following involuntary job loss. Soc Sci Med. 2009;68(8):
of Helsinki, Helsinki, Finland (Liina Junna, Pekka 1396–1406.

Am J Epidemiol. 2022;191(8):1459–1469
1468 Junna et al.

10. Eliason M, Storrie D. Inpatient psychiatric hospitalization 28. Brown SA. Personality and non-suicidal deliberate self-harm:
following involuntary job loss. Int J Forensic Ment Health. trait differences among a non-clinical population. Psychiatry
2010;39(2):32–55. Res. 2009;169(1):28–32.
11. Cunningham R, Milner A, Gibb S, et al. Gendered 29. Klein DN, Kotov R, Bufferd SJ. Personality and depression:
experiences of unemployment, suicide and self-harm: explanatory models and review of the evidence. Annu Rev
a population-level record linkage study. Psychol Med. 2021; Clin Psychol. 2011;7(1):269–295.
1–9. 30. Kokko K, Pulkkinen L, Puustinen M. Selection into
12. van Rijn RM, Robroek SJW, Brouwer S, et al. Influence of long-term unemployment and its psychological
poor health on exit from paid employment: a systematic consequences. Int J Behav Dev. 2000;24(3):310–320.

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

review. Occup Environ Med. 2014;71(4):295–301. 31. Stauder J. Unemployment, unemployment duration, and
13. Shah D. Healthy worker effect phenomenon. Indian J Occup health: selection or causation? Eur J Health Econ. 2019;
Environ Med. 2009;13(2):77–79. 20(1):59–73.
14. Butterworth P, Leach LS, Pirkis J, et al. Poor mental health 32. Schmitz H. Why are the unemployed in worse health? The
influences risk and duration of unemployment: a prospective causal effect of unemployment on health. Labour Econ. 2011;
study. Soc Psychiatry Psychiatr Epidemiol. 2012;47(6): 18(1):71–78.
1013–1021. 33. Fergusson DM, Boden JM, Horwood LJ. Unemployment and
15. Claussen B, Bjorndal A, Hjort PF. Health and re-employment suicidal behavior in a New Zealand birth cohort. Crisis. 2007;
in a two year follow up of long term unemployed. 28(2):95–101.
J Epidemiol Community Health. 1993;47(1):14–18. 34. Allison PD. Fixed Effects Regression Models. Thousand
16. Heponiemi T, Elovainio M, Manderbacka K, et al. Oaks, CA: SAGE Publications; 2009.
Relationship between unemployment and health among 35. Imai K, Kim IS. When should we use unit fixed effects
health care professionals: health selection or health effect? regression models for causal inference with longitudinal
J Psychosom Res. 2007;63(4):425–431. data? Am J Polit Sci. 2019;63(2):467–490.
17. Vågerö D, Garcy AM. Does unemployment cause long-term 36. Kaspersen SL, Pape K, Ose SO, et al. Unemployment and
mortality? Selection and causation after the 1992–96 deep initiation of psychotropic medication: a case-crossover study
Swedish recession. Eur J Public Health. 2016;26(5): of 2 348 552 Norwegian employees. Occup Environ Med.
778–783. 2016;73(11):719–726.
18. Muntaner C, Eaton WW, Miech R, et al. Socioeconomic 37. Drydakis N. The effect of unemployment on self-reported
position and major mental disorders. Epidemiol Rev. 2004; health and mental health in Greece from 2008 to 2013:
26(1):53–62. a longitudinal study before and during the financial crisis. Soc
19. Dube SR, Anda RF, Felitti VJ, et al. Childhood abuse, Sci Med. 2015;128:43–51.
household dysfunction, and the risk of attempted suicide 38. Fergusson DM, McLeod GF, Horwood LJ. Unemployment
throughout the life span—findings from the Adverse and psychosocial outcomes to age 30: a fixed-effects
Childhood Experiences Study. JAMA. 2001;286(24): regression analysis. Aust N Z J Psychiatry. 2014;48(8):
3089–3096. 735–742.
20. Björkenstam C, Kosidou K, Björkenstam E. Childhood 39. Organisation for Economic Co-operation and Development.
adversity and risk of suicide: cohort study of 548 721 Education at a Glance 2014: OECD Indicators. Paris,
adolescents and young adults in Sweden. BMJ. 2017; France: OECD Publishing; 2014.
357:j1334. 40. Stock JH, Watson MW. Heteroskedasticity-robust standard
21. Björkenstam E, Kosidou K, Björkenstam C. Childhood errors for fixed effects panel data regression. Econometrica.
household dysfunction and risk of self-harm: a cohort study 2008;76(1):155–174.
of 107 518 young adults in Stockholm County. Int J 41. Lund C, Brooke-Sumner C, Baingana F, et al. Social
Epidemiol. 2016;45(2):501–511. determinants of mental disorders and the Sustainable
22. Woodward LJ, Fergusson DM. Childhood peer relationship Development Goals: a systematic review of reviews. Lancet
problems and later risks of educational under-achievement Psychiatry. 2018;5(4):357–369.
and unemployment. J Child Psychol Psychiatry. 2000;41(2): 42. Kaplan GA, Shema SJ, Leite CMA. Socioeconomic
191–201. determinants of psychological well-being: the role of income,
23. Power C, Stansfeld SA, Matthews S, et al. Childhood and income change, and income sources during the course of 29
adulthood risk factors for socio-economic differentials in years. Ann Epidemiol. 2008;18(7):531–537.
psychological distress: evidence from the 1958 British birth 43. Van Gerven M. Converging trends of social policy in
cohort. Soc Sci Med. 2002;55(11):1989–2004. Europe: social security benefit reform in the UK, the
24. Schmillen A, Umkehrer M. The scars of youth: effects of Netherlands and Finland. Eur J Soc Secur. 2008;10(3):
early-career unemployment on future unemployment 207–225.
experience. Int Labour Rev. 2017;156(3–4):465–494. 44. The Social Insurance Institution of Finland, KELA.
25. Thern E, de Munter J, Hemmingsson T, et al. Long-term Unemployment benefit from Kela? https://www.kela.fi/web/
effects of youth unemployment on mental health: does an en/unemployment-benefit-from-kela-?inheritRedirect=true.
economic crisis make a difference? J Epidemiol Community Published January 1, 2021. Updated November 11, 2021.
Health. 2017;71(4):344–349. Accessed December 9, 2021.
26. Nock MK, Hwang I, Sampson NA, et al. Mental disorders, 45. Smed S, Tetens I, Lund TB, et al. The consequences of
comorbidity and suicidal behavior: results from the National unemployment on diet composition and purchase behaviour:
Comorbidity Survey Replication. Mol Psychiatry. 2010; a longitudinal study from Denmark. Public Health Nutr.
15(8):868–876. 2018;21(3):580–592.
27. Klonsky ED, Oltmanns TF, Turkheimer E. Deliberate 46. Hammarström A, Gustafsson PE, Strandh M, et al. It’s no
self-harm in a nonclinical population: prevalence and surprise! Men are not hit more than women by the health
psychological correlates. Am J Psychiatry. 2003;160(8): consequences of unemployment in the Northern Swedish
1501–1508. Cohort. Scand J Public Health. 2011;39(2):187–193.

Am J Epidemiol. 2022;191(8):1459–1469
 Unemployment and Mental Health 1469

47. Nichols A, Mitchell J, Lindner S. Consequences of Long- 56. Solaz A, Jalovaara M, Kreyenfeld M, et al. Unemployment
Term Unemployment. Washington, DC: The Urban Institute; and separation: evidence from five European countries. J Fam
2013. https://www.urban.org/sites/default/files/publication/ Res. 2020;32(1):145–176.
23921/412887-Consequences-of-Long-Term-Unemployment. 57. Gruber J. The wealth of the unemployed. ILR Rev. 2001;
pdf. Published July 2013. Accessed March 16, 2022. 55(1):79–94.
48. Artazcoz L, Benach J, Borrell C, et al. Unemployment and 58. Plagnol AC. Financial satisfaction over the life course: the
mental health: understanding the interactions among gender, influence of assets and liabilities. J Econ Psychol. 2011;
family roles, and social class. Am J Public Health. 2004; 32(1):45–64.
94(1):82–88. 59. Mirowsky J, Ross CE. Economic hardship across the life

Downloaded from https://academic.oup.com/aje/article/191/8/1459/6571049 by Robert Gordon University user on 19 February 2024

49. Strandh M, Hammarström A, Nilsson K, et al. course. Am Sociol Rev. 1999;64(4):548–569.
Unemployment, gender and mental health: the role of the 60. Arulampalam W, Gregg P, Gregory M. Unemployment
gender regime. Sociol Health Illn. 2013;35(5):649–665. scarring. Econ J. 2001;111(475):577–584.
50. Kowalewska H, Vitali A. Breadwinning or on the breadline? 61. Gathergood J. Unemployment risk, house price risk and the
Female breadwinners’ economic characteristics across 20 transition into home ownership in the United Kingdom.
welfare states. J Eur Soc Policy. 2021;31(2):125–142. J Hous Econ. 2011;20(3):200–209.
51. Jalovaara M, Fasang AE. Family life courses, gender, and 62. Fischer EH, Dornelas EA, Goethe JW. Characteristics of
mid-life earnings. Eur Sociol Rev. 2020;36(2):159–178. people lost to attrition in psychiatric follow-up studies. J Nerv
52. Warr P. Work, Unemployment, and Mental Health. New York, Ment Dis. 2001;189(1):49–55.
NY: Oxford University Press; 1987. 63. Hardt J, Rutter M. Validity of adult retrospective reports of
53. Jahoda M. Reflections on Marienthal and after. J Occup adverse childhood experiences: review of the evidence.
Organ Psychol. 1992;65(4):355–358. J Child Psychol Psychiatry. 2004;45(2):260–273.
54. Goodman WK, Geiger AM, Wolf JM. Leisure activities are 64. Aneshensel CS, Estrada AL, Hansell MJ, et al. Social
linked to mental health benefits by providing time structure: psychological aspects of reporting behavior: lifetime
comparing employed, unemployed and homemakers. depressive episode reports. J Health Soc Behav. 1987;28(3):
J Epidemiol Community Health. 2017;71(1):4–11. 232–246.
55. van der Meer PH. Gender, unemployment and subjective 65. Morgenstern RD, Barrett NS. The retrospective bias in
well-being: why being unemployed is worse for men than for unemployment reporting by sex, race and age. J Am Stat
women. Soc Indic Res. 2014;115(1):23–44. Assoc. 1974;69(346):355–357.

Am J Epidemiol. 2022;191(8):1459–1469