Health-related selection into employment among the unemployed

We analyzed the total population of unemployed Finnish residents aged 30–60 during the years 2009 to 2018 identified from the population registers of Statistic Finland (Fig. 1; permission #TK-53-339-13). This age range was selected because a substantial share of Finnish residents under the age of 30 are still studying, while individuals above the age of 60 are near retirement age [36]. We linked information from various administrative registers using the personal identification codes that all residents receive at birth or upon immigration.

Unemployment was defined as having registered as currently jobless and seeking employment (n = 886,050, Fig. 1). This measure is highly sensitive, as registering immediately at the start of unemployment is required to receive unemployment benefits in Finland. We utilized data on exact unemployment onset and ending dates. The date of registering was set as the onset of unemployment. The end of unemployment was set at the date of registering an exit from unemployment, the end of the two-year follow-up, death or aging out, whichever came first (Table 1). In this analysis, we refer to an unemployment spell as the time in days that someone is registered as unemployed jobseeker. As it is possible to experience multiple spells of unemployment, we considered all spells from each individual (see Statistical Analyses).

We excluded spells lasting two days or less and ending in employment, as these are likely to represent workplace transitions (Fig. 1). We further excluded recurrent unemployment spells resulting in employment (five or more spells over a calendar year), as these are likely to be freelancers in between jobs. We also entirely excluded individuals who were considered to be weakly attached to the labor market (those with over 25 separate unemployment spells during the study period). Including these individuals in a sensitivity analysis had a minimal impact on our results (Supplementary Table 1, 2; Additional file 1). We excluded unemployment spells that had missing data on end dates, sociodemographic measures, or health status. Data were missing for less than 0.8% of all unemployment episodes and therefore this exclusion is unlikely to greatly affect our results. Furthermore, individuals with missing data for an unemployment spell could nevertheless enter the analytical sample if they had another unemployment spell for which these measures were available. In addition, individuals who only had unemployment spells with missing background data were excluded entirely; this constituted a total of 66,847 individuals (Fig. 1). Most of these cases consist of very recent immigrants or return migrants who only had one unemployment spell immediately after entering the Finnish registers, which meant that measures on health or other individual characteristics prior to the onset of unemployment could not be determined. The final analytical sample consisted of 814,085 individuals, with a total of 2,858,523 unemployment spells.

The outcome was the exact date of exit from unemployment due to finding employment. We also separately analyzed stable and unstable employment (i.e., if the unemployed with poor health opt for, or can only obtain, insecure or short-term jobs). Stable employment was defined as being employed at any workplace at the end of the calendar year following the initial date of employment. Unstable employment was defined as not being employed at the end of the calendar year following employment.

The exposure was the pre-baseline health status measured over the two years preceding the onset of unemployment; the information on health status and other covariates was therefore recorded in 2007–2019. All included health measures were based on information on visits to specialized health care from the Care Register for Health Care of the National Institute for Health and Welfare (hospitalizations, outpatient care, emergency care visits). For the specific physical conditions only, we additionally used information on special reimbursement rights for long-term medication for certain chronic diseases (the right to full or nearly full financial compensation upon purchase of medication), obtained from the registers of the Social Insurance Institution of Finland (Additional file 2). To obtain the right to special reimbursement, patients must provide a clinician’s confirmation of their need for the prescription due to a specific diagnosis [37]. The special reimbursement system covers medications that are primarily used to treat certain severe physical conditions and cannot be used to identify common mental health conditions, or physical conditions outside this reimbursement system. The care and medication registers are collected for administrative purposes of the publicly funded universal health insurance and cover all Finnish residents. They include the dates (visits to specialized care/beginning of special reimbursement right) and the main diagnoses (International Classification of Diseases, revision 10 (ICD-10)/Finnish special refund category).

The selection of the health conditions was based on their high prevalence and public health importance among working-age Finns [38]. First, we explored different general aspects of health. We considered visits to specialized care that were related to 1) any alcohol-related conditions and poisonings; 2) any psychiatric conditions and self-harm (excluding alcohol and other substance use disorders); 3) any injuries; and 4) any physical conditions (excluding visits related to pregnancy and childbirth, and alcohol-related physical conditions, see Additional file 2). We then separately considered common specific health conditions. Using data on both specialized health care visits and special reimbursement rights, we separated the physical conditions as follows: 1) cancer; 2) diabetes; 3) heart disease; and 4) neurological conditions. Based on specialized health-care visit data, we divided the psychiatric conditions into 1) depression; 2) anxiety disorders; and 3) substance use disorders. All health indicators – both the general aspects of health and the specific health conditions – were included as separate dummy measures of whether the individual had the given condition over the two calendar years preceding the onset year of unemployment (yes/no).

Based on previous literature [1, 11, 13], we controlled for the following additional characteristics: continuous age, native language (Finnish or Swedish, other), marital status (married or cohabiting, single, divorced or widowed), education (basic, secondary, or tertiary), occupational status (upper or lower-level white collar with managerial occupations, lower-level white collar, blue collar, farmer or self-employed, student, or other) [39], number of individuals living in the same household (continuous), home ownership (yes/no), unemployment history (number of months unemployed over the two calendar years preceding the onset of unemployment, continuous), running number of the unemployment spells (from the start of the study period, categorical), and the onset year of unemployment (categorical). As labor market prospects may be particularly limited in declining regions, we also included the region of residence (Nomenclature of Territorial Units for Statistics 3, categorical, 21 regions). All these covariates were measured annually on December 31 of the year preceding the onset of unemployment and included as time-invariant. For the analyses concerning the specific health conditions, we also generated a binary measure for any remaining comorbidity other than the specific health conditions considered in this study (any visits to health care over the preceding two years for any diagnosis excluding cancer, diabetes, heart disease, neurological conditions, depression, anxiety disorders, substance use disorders; yes/no). For women, we additionally controlled for births that occurred over the two preceding years (yes/no), which were identified from the hospital discharge register, as having a child may affect the risk and length of unemployment.

We conducted multivariate Cox proportional hazards models in Stata, version 16.1 (College Station, Texas, 2019). The analysis time was defined as the number of days from the onset of unemployment. As an individual could experience unemployment and find employment multiple times, we considered all spells for each individual, with standard errors clustered at the individual level. There was no loss to follow-up. Individuals were right-censored at the end of their unemployment spell (which could end due to employment, retirement, migration, exiting the labor market for other reasons, aging out, or death), or the end of the two-year follow-up, whichever came first. The proportional hazards assumption was visually estimated to hold well for all outcomes (results upon request from the authors).

We stratified all analyses by sex because health may have a gendered impact for finding employment, and because less is known about the employment among unemployed women [40]. We separately estimated the following models for the four general aspects of health, and the seven specific health conditions (exposure, see above). In Model 1, we included each health measure separately and controlled for age and the unemployment onset year. In Model 2, we added the other individual characteristics (see other covariates, above). In Model 3a, we considered all the general aspects of health simultaneously to control for the effects of comorbidity while controlling for the other covariates as in Model 2. In Model 3b, all the specific health condition were considered together and, in addition to the covariates introduced in Model 2, we further adjusted for a binary measure for any remaining comorbidity (see the other covariates above). We also separately considered the role of the different general aspects of health and the specific conditions for finding stable and unstable employment, while adjusting for the same variables as in Models 3a and 3b.

Some of the brief unemployment spells could reflect transitions from one workplace to another rather than involuntary unemployment. To address this, we excluded unemployment episodes from the analyses that lasted less than a month (Supplementary Table 1, 2; Additional file 1). As our exclusion criteria were arbitrary for weak labor market attachment and freelancing (over 25 unemployment spells, or more than 5 spells ending in employment over a calendar year), we also tested the inclusion of these spells (Supplementary Table 1, 2; Additional file 1) [41].

To ensure that our results were not due to competing risks such as retirement and death, we additionally conducted the Fine and Gray competing risk regression for finding any, stable, and unstable employment (Supplementary Table 3; Additional file 1) [36].

Our results from the multivariate Cox proportional hazard models were robust to competing risks. The exclusion of short unemployment spells and the inclusion of those with loose attachments to the labor market produced similar results to those of our main analyses (Additional file 1).

Few previous studies have investigated alcohol-related health in relation to employment among the unemployed, but we showedit to be an important risk factorfor employment, and in particular, for stable employment. A Norwegian study reported alcohol use disorders to be unassociated with 5-year employment among the long-term unemployed, but this is likely due to a lack of statistical power (n = 228), as none of the 18 subjects with alcohol use disorders became employed [9].

The prior evidence is more consistent for the association between psychiatric conditions and finding employment. In Norway, the chances of finding employment within two years was 70% lower (p-value < 0.001) for those unemployed with any psychiatric diagnosis [7]. Among the Danish unemployed, any psychiatric medication purchases were also negatively associated with employment (OR 0.46; 95% CI: 0.34, 0.63) [32]. These results are in line with our finding that the unemployed with any psychiatric condition, depression or anxiety disorders were approximately 30% less likely to be employed over the 2-year follow-up.

Previous research assessing employment in terms of physical health is rare and predominantly based on small samples that produce imprecise point estimates [7, 8]. Contrary to our results, two Norwegian studies reported no association between any physical conditions and finding employment (N = 277, 228). Conversely, purchasing any prescription medication related to a physical condition was demonstrated to be detrimental to employment among the Danish unemployed in 2006–2008 (n = 1806) [32]. Previous studies addressing specific physical health conditions are also rare. Our results for heart disease are congruent with a study conducted in Sweden, in which self-reported circulatory symptoms were negatively associated with employment among respondents aged 25–54 in 1992–1993 (n = 452) [10]. Altogether, our results based on a large sample and objective and consistent measurement of both unemployment and physical conditions demonstrates that physical health is predictive of transitions from unemployment to employment.

Two explanations are plausible for the health-related selection from unemployment into employment. First, employers may discriminate against applicants with health-related gaps in their employment history [42], no recent work experience [43], or those that display signs of poor health, especially if these cues point to alcohol-related or psychiatric problems [25, 26]. In support of this hypothesis, we demonstrated that alcohol-related and psychiatric conditions were particularly harmful to finding employment. It is unclear to what extent employment protection legislation in the Nordic welfare states hinders discrimination against applicants with health conditions in hiring [5, 24]. Second, poor health could make the unemployed individual less capable or motivated to seek employment, or it could limit the options available to them. Chronic conditions may hinder an individual’s education as well as building skills and experience, and thus limit employment prospects [14, 15]. This hypothesis was not supported by our results, as adjusting for education, occupational status and unemployment history did not explain the lower chances of becoming employed among those with poor health. It is therefore more likely that poor health may either narrow the available options down to positions with lower health-related demands (both physical and social), or that symptoms and health-related stress may hinder the job search [16‐19].

Regarding the general discussion on health-related selection into and out of employment, our results, together with the literature, suggest differing selection processes for different aspects of health. For example, while hospitalizations related to cancer and ischemic heart disease [11] and circulatory symptoms [10] did not seem to increase the risk of job loss among employed Swedes, our results suggest that physical conditions do hinder moving from unemployment to employment. Physical health may thus be less associated with the risk of job loss (e.g., due to employee protection legislation), and more with finding new work [5]. Psychiatric and alcohol-related conditions on the other hand appear to be associated with both job loss [11] and finding employment. Taken together, individuals with long-standing health conditions – physical or psychiatric – are likely to require additional support following their job loss.

Previous studies have hypothesized that poor health, especially alcohol-related and psychiatric conditions, may be a determinant of precarious employment [34]. We found some support for this hypothesis, as alcohol-related and psychiatric conditions were particularly harmful for stable employment. This indicates that these conditions may hinder both procuring employment as well as strong and stable attachments to the labor market. Prior studies have established that those with psychiatric conditions face discrimination when looking for work as well as during employment [44] and they also have a higher risk of job loss [11]. Some psychiatric conditions may also be episodic in nature, which translates to a time-varying capacity to work. Psychiatric health may therefore be particularly important for job loss and the retention of stable employment.

Our results also imply that health is likely to contribute to accumulating disadvantage [10]. Those with poor health who lose their employment may be in danger of lengthy unemployment spells and unstable career trajectories which, in addition to being costly to society, have been linked to individual level outcomes such as worsening labor market position [45] and subsequent poor health [1]. To support employment into stable jobs among all unemployed, we need to identify vulnerable population groups with special problems related to their workability [46, 47], to modify job demands to better accommodate people in poorer health, and to reduce health-related stigmatization of those unemployed with poor health. Additional studies are needed to identify the best practices for supporting swift and stable employment among those unemployed with health conditions.

The strengths of this study are the large-scale, high-quality longitudinal data that enabled us to separate specific health conditions and employment outcomes. The data also allowed us to form a measure of any alcohol-related conditions that includes both physical illness and alcohol use disorders, which is a notable strength because there is likely to be considerable overlap between the two. Another strength is that our unemployment and health measures include exact dates. In addition, the analysis does not suffer from problems related to population sub-group differences in reporting styles, or higher rates of loss to follow-up among those in poor health.

Our health measures capture a wide range of clinically relevant, chronic conditions. We utilize data from specialized health care, which includes hospitalizations, outpatient care and emergency care visits, as well as data on special reimbursement rights for certain physical conditions. The special reimbursement rights are registered irrespective of whether the patient purchases the medications and therefore do not reflect an adherence to treatment. However, the special reimbursement data only includes certain conditions, and this prevents us from identifying some mental health-related chronic conditions such as insomnia, or conditions that are primarily treated with pain medication. This study is also blind to untreated or less severe conditions that do not appear in the registers.

Similarly, we cannot identify those who do not register as unemployed (e.g., the discouraged unemployed, or those not financially dependent on unemployment benefits due to severance pay packages). Thus, those who do not register are likely either to be in an advantageous position (e.g., with regards to employment history or a better financial situation due to severance packages) or conversely, in an extremely disadvantaged position (discouraged workers) when compared to the registered unemployed. Inclusion of the discouraged workers would likely strengthen the association between health and employment. In contrast, inclusion of those on severance packages may weaken the association. This group is likely to be relatively healthy and affluent, with less financial pressure to find employment. Therefore, they are likely to experience longer unemployment spells [48]. However, little is known about the number of people who receive severance packages or the level of pay and benefits they receive, as these are considered private agreements with employers and employees. Overall, the magnitude and joint effects of these two tendencies are unclear, but they could potentially cancel each other out.

Finally, using register data, we cannot control for job attitudes such as motivation, willingness to relocate for work or to take up employment in a new field. The contribution of job attitudes to health-related selection therefore remains unknown.