The impact of an unemployment insurance reform on incidence rates of hospitalisation due to alcohol-related disorders: a quasi-experimental study of heterogeneous effects across ethnic background, educational level, employment status, and sex in Sweden

The Swedish welfare state has traditionally been characterised as generous [9]. Over time, however, policy changes and their accompanying reforms have resulted in a slow but persistent retrenchment [10]. A clear example of this can be found for unemployment insurance. For a long time, the Swedish system was based on a passive labour market policy with relatively generous and extended benefits without any strict qualifying requirements [11]. The voluntary unemployment insurance system was established in the 1930s and, while governed and subsidised by the state, it is managed by a relatively large set of unemployment insurance funds [12]. Important to note is that union membership automatically covers the fund membership fee. As of the late 1990s, the insurance system encompasses two schemes. First is the basic benefit scheme which is based on a daily allowance (flat-rate benefits). It is available to all job-seekers, including those who are not members of any fund, assuming that they meet certain requirements related to e.g. employment history. Second is the income-based scheme, which pays a replacement allowance (up to 80%) based on previous earnings but is conditioned upon membership in a fund (earnings-related benefits) [13].

Despite relatively low unemployment rates at the time, the public debate preceding the 2006 parliamentary election to a large extent circulated around the widespread perception of difficulties finding employment and discontentment with the idea of a high share of people receiving benefits [10, 14]. Making the reduction of unemployment one of their priorities, the alliance of center-right parties that won the election immediately introduced a substantial reform of the unemployment insurance which came into effect January 1, 2007. While the basic setup of the system remained largely the same, the state considerably decreased its involvement by reducing contributions to the unemployment insurance funds. As a consequence, union membership fees increased and the amount of benefits was reduced, which ultimately led to a significant drop in members of unemployment insurance funds. Although Sweden still has one of the highest unionisation rates in the world, union density fell by four percentage points between 2006 and 2007 (from an annual average of 77% to 73% among individuals aged 16-64) [15]. This was not only caused by individuals deciding to leave the unemployment insurance funds but also by a decreasing inflow of new members. Past research has shown that the drop was particularly evident among funds targeting the working class, for which the membership fees also increased more drastically [16]. Such funds generally have a higher representation of individuals with lower educational attainment and occupations generating less income, as well as foreign-born. Indeed, studies show that those who had resided in Sweden for a short period of time and originated from countries that are geographically far from Sweden were considerably more likely to drop out of the unemployment insurance funds [17]. Additionally, incentives for accessing the income-based scheme are most likely weaker among individuals with low income [18]. Concerning differences by sex, the decrease in members varied greatly across specific funds without any systematic differences between male-dominated and female-dominated funds [16].

The aforementioned changes of the unemployment insurance also involved requirements for benefits being made stricter [15]. This led to a decrease in the share of unemployed receiving unemployment insurance benefits, whether accessed through the basic benefit scheme or income-based scheme. For example, comparing the periods 2001-2006 and 2007-2012, the share of unemployed receiving such benefits diminished by approximately one half [19]. These individuals were instead referred to other sources of financial support, such as social assistance benefits [20]. Again, groups that were particularly affected included individuals with weak labour market attachment, such as foreign-born [19].

Previous research has identified several plausible pathways linking unemployment insurance to health-related outcomes [3]. By definition, unemployment benefits compensate (at least partially) for income loss among the unemployed. In addition to securing the economic resources required to satisfy the individual’s basic needs, it allows for investments in goods and activities that can further promote health [21, 22]. However, an unemployment insurance system that is generous in terms of benefit levels and coverage may be regarded as a collective resource [23] that is beneficial also for the employed. While unemployment insurance in itself cannot directly decrease the risk of becoming unemployed, the perception of economic security (in case of job loss) it entails could reduce feelings of stress and anxiety and thereby positively affect health [22]. In particular, such spill-over effects are expected to apply to groups who have employment but still experience limited economic resources and weaker labour market attachment [23].

More specifically applied to alcohol-related disorders, previous studies have found unemployment to be a risk factor [24]. Unemployment – or the threat of unemployment – may also worsen already existing substance use problems and increase the risk of relapse [25]. A reduction in the generosity and coverage of unemployment insurance could potentially exaggerate stress related to the anticipation of job loss or experiences of unemployment, and thereby increase the use of alcohol as a way to cope with stress, which would then magnify the risk of hospitalisation due to alcohol-related disorders [cf. 3, 17]. In this context, having a low educational level – and therefore potentially less socioeconomic resources – might reduce the individuals’ overall ability to manage their experiences of unemployment and exacerbate the negative impact on (alcohol-related) behaviours. Moreover, drawing upon previous studies looking into sex differences in the effects of unemployment on health-related outcomes [26], any such processes are likely to be more evident among men.

When it comes to individuals with foreign background living in Sweden, previous studies have generally demonstrated lower rates of alcohol consumption and alcohol-related disorders as compared to natives [27]. Individuals with foreign background is nonetheless a highly diverse group, and there are some marked differences between first-generation and second-generation immigrants as well as according to country of origin [28]. For example, rates tend to converge by time spent in Sweden and be relatively higher among groups originating from countries with an alcohol culture more similar to Sweden [27]. Regardless of the lower rates, however, there is reason to believe that stressful situations – such as facing unemployment – can be more detrimental for a population that experience more disadvantaged living conditions on average, and weaker labour market attachment in particular [29, 30]. Combining this with the observation that unemployment insurance coverage and replacement rates among foreign-born individuals were greatly affected by the 2006 unemployment insurance reform, it is plausible that influences of the reform on alcohol-related disorders might also be more pronounced in this group.

The overall aim of this study is to study (heterogeneity in) the effects of the 2006 unemployment insurance reform on hospitalisation due to alcohol-use disorders, using a quasi-experimental approach. Our study period is 2001–2012, with January 1, 2007 as the threshold of interest. The specific research questions are as follow:

In order to expand our understanding one step further, we will specifically examine heterogeneous effects among the unemployed individuals and the role played by unemployment benefit recipiency. Thus, here we ask:

The initial sample included individuals in ages 30 to 60 years at any point during the study period (n = 5,759,292). We right-censored individuals who died or migrated during or before the study period (n = 637,860) and excluded those with missing information on ethnic background or education (n = 103,159). Additionally, we excluded those for whom employment status could not be ascertained (n = 8,441). These exclusions led to 5,009,832 individuals being included in the study population (see Additional file 1, Supplementary Figure S1, for an overview). The unit of analysis in the current study is however not individuals but monthly aggregates of the individual-level data.

The stratification variables comprised ethnic background, educational level, employment status, and sex. In addition, among the unemployed, we included information about unemployment benefit recipiency. Hospitalisation due to alcohol-related disorders served as study outcome.

The analysis uses a quasi-experimental study approach based on a regression discontinuity design (RDD) [31]. RDD is well-established in economics and has become increasingly common in other research disciplines, including epidemiology. The basic idea of RDD is to assess the effects of an intervention (usually referred to as ‘treatment’) by assigning a threshold between the treatment and control group. Observations on either side of the threshold are compared and the average treatment effect is then measured with the emerging discrepancy (usually referred to as ‘discontinuity’) between these treatment and control units. RDD requires relatively mild assumptions compared to other nonexperimental research designs [32]: the basic model involves a continuous scoring variable representing the outcome, a forcing variable representing the exposure, and an (exogeneous) treatment to determine the cut-off (or ‘threshold’) when the discontinuity occurs. Furthermore, a key feature of RDD is the regression discontinuity (RD) plot. It is obtained by plotting the local means of incidence rates over units of time and estimating distinct regression lines for the treatment and control units. The gap between the intercepts of both regression lines around the cut-off-point designates the magnitude of regression discontinuity and the quasi-causal effect of the intervention of interest. The current study employs ‘Regression discontinuity in time’ (RDiT). RDiT is considered a subtype of RDD and involves units of time as the forcing variable [33]. The study period was 2001–2012 and the cut-off (threshold) was set at the date of the intervention (January 1^st 2007). All analyses were performed with Stata v17.

Based on the Stata package RDROBUST, a default RD model with uniform kernel function and evenly spaced mimicking-variance (ESMR) was used. The estimated regressions were based on local quartic (4^th order) polynomials that were derived from non-parametric regressions [34]. In the first set of models, we included both employed and unemployed individuals. The analysis was performed for the total study population and then separately by ethnic background, educational level, employment status, and sex. In the second set of models, the analysis was first performed for the unemployed population and then separately by ethnic background, educational level, and sex.

A corresponding RD plot was generated for each model to illustrate the results, with help of the user-written Stata packages RDPLOT [35] and HEATPLOT [36]. The RD plots are based on the local means of group-specific incidence rates of hospitalisation due to alcohol-related disorders at each month during the study period (represented by bins in the plots). The vertical dashed line denotes the cut-off point (January 1, 2007) that divides the time axis into a pre-intervention period and a post-intervention period. In the RD plots for the unemployed population, coloured bins were used to depict the altering proportions of between receiving benefits versus not receiving benefits at each cross-section; a higher value (and hence a warmer colour) indicates a larger proportion of non-recipients. Model-based regression lines were included in all RD plots to demonstrate the change around the threshold. Moreover, confidence intervals are included as shaded areas in the RD plots.

The descriptive characteristics of the study population are shown in Table 1. The estimates demonstrate higher incidence rates of hospitalisation due to alcohol-related disorders among men in relation to women, comparably high numbers in the European foreign-born population, but low rates among the Non-European foreign-born. Large differences in incidence rates by educational level and employment status are shown. Among individuals who were unemployed but did not receive unemployment benefits, incidence rates are more than twice as high compared to those unemployed who received such benefits. Unemployed and individuals with low education also show a notable increase across the two specified periods (2001–2006 and 2007–2012).

Additional files 2, 3 and 4 show the percentage of unemployed individuals with and without unemployment benefit recipiency by ethnic background, educational level, and sex. It can be noted that the percentage of unemployed individuals was generally stable or slightly declining in most groups before the reform was implemented. A more evident decrease is then seen in 2007–2008, followed by an increase starting in 2009. Concerning the results stratified by ethnic background (Additional file 2; Supplementary Figure S2), the orange lines disclose that the share of unemployed Swedish-born with Swedish-born parents who received benefits declined between the years 2005 and 2009 (two years before and after the reform was implemented). This pattern is consistent throughout all subgroups. The share of unemployed individuals without unemployment benefit recipiency is largely stable for both groups of Swedish-born, with some tendency towards an increase over time. For European foreign-born and non-European foreign-born, there is a marked increase starting around 2008, especially in the latter group. The blue bars depict how recipiency versus non-recipiency deviate from the overall mean among the unemployed. Particularly obvious is the mirror-inverted timely pattern with a breaking point in the year 2007 in the group of non-European foreign-born. The plots that stratify by education (Additional file 3; Supplementary Figure S3) and sex (Additional file 4; Supplementary Figure S4) show largely similar patterns. Among the low-educated, the increase of non-recipients after 2007 is somewhat more pronounced compared to the high-educated.

Figures 1, 2, 3, 4, and 5 show RD plots with incidence rates of hospitalisation due to alcohol-related disorders during the period 2001–2012 in the total study population, as well as by ethnic background, educational level, employment status, and sex. Again, it should be noted that the bins represent the local means of group-specific incidence rates of hospitalisation due to alcohol-related disorders at each month during the study period. The point estimates, p-values, and confidence intervals are available in Additional file 5 (Supplementary Table S1).

Regarding the results for the total study population (Fig. 1), the results show a slight reduction in incidence rates around the time that the reform was implemented (-0.49, p < 0.001), which is expected based on the overall decrease of hospitalisation due to alcohol-related disorders over time.

Next, the results are stratified by ethnic background (Fig. 2). Swedish-born with Swedish-born parents show a small decrease in incidence rates (-0.42, p < 0.001). This is also the case among non-European foreign-born individuals (-0.62, p < 0.001) although their absolute levels of hospitalisation are consistently much lower in comparison to the other groups. A more pronounced decrease can be seen for European-foreign individuals (-2.39, p < 0.001). In contrast, increased incidence rates posterior to the intervention are found among Swedish-born with foreign-born parents (1.41, p < 0.001).

Stratification by educational level (Fig. 3) discloses substantial differences in absolute levels between individuals with low and high education. However, both groups demonstrate a small decrease around the time of the reform (low education: -0.47, p < 0.001; high education: -0.43, p < 0.001).

For the results stratified by employment status (Fig. 4), the slight decrease in incidence rates for employed individuals (-0.56, p < 0.001) is contrasted by a relatively sharp increase among unemployed around the year 2007 (1.04, p < 0.001).

Large absolute differences in incidence rates can be observed between men and women (Fig. 5). The comparably high incidence rates among men follow the pattern seen for the total population. For both groups, however, decreases in the incidence rate of hospitalisation due to alcohol-related disorders can be observed (men: -0.77, p < 0.001; women: -0.21, p < 0.001).

The next set of figures (Figs. 6, 7 and 8) show the results specifically for the unemployed population, stratified by ethnic background, educational level, and sex. As previously mentioned, coloured bins were used to illustrate changes in the absolute difference between unemployed who received unemployment benefits and unemployed who did not receive such benefits during the study period.

Focusing on unemployed individuals’ ethnic background (Fig. 6), Swedish-born with Swedish-born parents show upward trends after January 2007 when the unemployment insurance reform was implemented (2.25, p < 0.001). An increase can also be noted for European foreign-born individuals (3.83, p < 0.001). Among the Swedish-born with foreign-born parents, the effect of the reform is particularly evident (7.41, p < 0.001). On the contrary, Non-European foreign-born individuals show very low incidence rates with a marked decrease around the time of the reform (-2.05, p < 0.001). Interesting to note is that the increase in the proportion of non-recipients during the post-intervention period seems to be largest for the non-European foreign-born individuals and smallest for the Swedish-born with foreign-born parents.

In terms of educational background (Fig. 7), incidence rates are comparatively higher among low-educated individuals. For this group, there is also a notable increase around the time of the reform (2.50, p < 0.001). The opposite is demonstrated for high-educated individuals (-1.52, p < 0.001). Men disclose incidence rates that are twice as high compared to women (Fig. 8). Both sexes demonstrate an increase in incidence rates due to alcohol-related disorders, although this is more pronounced among men (men: 2.77, p < 0.001; women: 0.51, p < 0.001). In general, a gradual increase in the proportion of non-recipients can be noted in all groups.

A major strength of the current study was the use of a total population sample which rendered sufficient statistical power to investigate a relatively rare outcome such as hospitalisation due to alcohol-related disorders while still being able to stratify the analysis into several subgroups. Another strength was that the study was based on register data with nearly complete coverage and limited loss to follow-up. The register-based exposure and study outcome was insusceptible to self-reporting bias or other sources of manipulation, which strengthens the validity of the RD design [32]. Moreover, even though the study was based on observational data, we could approach the issue of causality with more certainty by applying a quasi-experimental design.

Some important limitations should nevertheless be addressed. With regards to the outcome, hospitalisation data was the only feasible alternative since other national health registers in Sweden (e.g. the Prescribed Drug Register, and outpatient care data from the National Patient Register) did not commence until later years. As previously mentioned, this means that our outcome only captures the most adverse consequences related to alcohol use. Having information also on alcohol problems treated in outpatient care or alcohol consumption data might have provided a more complete picture of the effects of the 2006 unemployment insurance reform on alcohol-related outcomes.

The fact that the studied reform was implemented at the national level on a specific date (January 1^st 2007) is strength. However, it is unlikely that the reform caused any effects on incidence rates of hospitalisation due to alcohol-related disorders on the exact day of implementation. As with many reforms, the threshold is in reality ‘fuzzy’ – involving both anticipatory effects and lagged effects. We tried to account for this by truncating the regression lines three months before and after threshold in a series of sensitivity analyses (data not presented), but this did not cause the main conclusions of the study did not change. There are also other limitations with the regression discontinuity design that need to be discussed. To facilitate the interpretation, we included rather broad categories of ethnic background, educational level, employment status, and sex. For example, with regards to ethnic background, exploring differences across countries of origin is something for future research to consider, as it may provide a deeper understanding of the role of drinking cultures and level of alcohol use. Specifically, in this context, it may also be relevant to broaden the outcome to include other types of substance use.

There is always a possibility that effects found with regression discontinuity design are statistical artefacts, created by natural variation in the studied outcome over time, or that they are actually caused by macro-level changes other than the one targeted by the study [33]. To account for the former, we performed additional analyses with arbitrarily chosen cut-off points [44]. None of those showed any substantially different results in relation to the most prominent effects found in the main analyses, which strengthened our conclusions. In terms of the latter, it is indeed difficult to isolate the effect of a specific reform, since societies are in constant flux. We have not identified any similarly comprehensive reforms of relevance occurring at the same time as our specified threshold, neither in relation to unemployment insurance, alcohol policy, or health care. It should nonetheless be mentioned that several drastic macro-level changes occurred during the post-intervention period, including the financial crisis in 2008. Even if there were early global signs of a great recession already in 2007, the consequences did not hit the Swedish economy and labour market until late 2008. Compared to countries like, for example, Spain and Iceland, impacts on the Swedish economy were relatively mild [45, 46]. However, in order to decrease the possible influence of the financial crisis on our results, we re-examined our models using a narrower bandwidth (2005–2008). This did not cause our conclusions to change in any substantial way.