Study design and participants
A cross-sectional survey consisting of a questionnaire combined with register data from Statistics Sweden was conducted during 2016. From the 5,671,149 individuals who were between 20 and 64 years old and lived in Sweden at the time of the survey, 2500 individuals were selected with simple random sampling and were invited to participate in the study. The questionnaire, which was administered by Statistics Sweden, was sent to the home addresses of the invited individuals on May 2 with two reminders (May 18 and June 1). Participants consented to participate in the study when they returned their questionnaire. There were 967 individuals (39%) who responded to the questionnaire.
The questionnaire consisted of questions related to the participant’s labor market position, health status, health care consumption, and socio-economic and demographic conditions (including education level and marital status). We used self-reported questions related to labor market status to define exposure to unemployment, the EQ-5D as the outcome variable, and gender, age, education level, marital status, and previous health as potentially confounding variables.
Questionnaire responses were scanned and merged with register data that Statistics Sweden administers, and thereafter de-identified, by Statistics Sweden. Register data included, among other things, information about historical unemployment and demographics of the study participants. In the current study, we only used register data to validate the age and gender of the participants. Our choice of variables were consistent with variables in studies similar to ours, with previous health as the only exception [2]. The Regional Ethical Board in Umeå, Sweden, approved the survey.
Definition of labor market status
The study participants’ labor market status was determined based on the questions: “Which is your main employment” – with ten response alternatives – and “How long have you been unemployed in the last three years?” – with five response alternatives. Those who responded that they had at least six months of unemployment during the last three years were categorized as unemployed (n = 113), and those among the other respondents who had responded that their main employment was “gainfully employed” (n = 720) or “labor market activity” (n = 4) were defined as employed (n = 724). Labor market activity refers to a job that is subsided by the state to give the unemployed work experience in order to establish themselves in the labor market. We defined those engaged in labor market activities as employed because we argue that they, in line with what Jahoda has proposed, have a time structure for their waking day, regular contacts with people outside the nuclear family, a purpose of the day transcending their own, and an enforced activity [14]. Thus, their situation is similar to the gainfully employed. The 837 employed and unemployed participants were defined as active in the labor market. The remaining 130 individuals were excluded from our analyses because they had either a different main employment than “gainfully employed” or “labor market activity” or had not responded to this question. There were 52 participants who had experienced unemployment of less than six months, and of these 36 were defined as employed and 16 were excluded.
Health-related quality of life variables
The EQ-5D was used to measure health-related quality of life [11]. We used the descriptive system with five questions that measure different dimensions of health (mobility, self-care, usual activities (such as work, studies, housework, family, and leisure activities), pain/discomfort, and anxiety/depression) and three response alternatives (corresponding to no, some, or extreme problems). Responses to these questions were translated to QALY scores based on the United Kingdom value set for the EQ-5D, which was derived using linear regression and based on the responses to the EQ-5D descriptive system [12]. The main emphasis in our study was on the QALY scores, but we also present results for the other parts of the EQ-5D instrument, i.e. the dimensions themselves, and EQ-VAS. With the EQ-VAS, respondents valued their current health on a visual analogue scale ranging from 0 to 100. Responses to the EQ-5D dimensions were dichotomized into two groups for analyses of the dimensions themselves, with some and extreme problems combined into one of the two groups, while all three levels were used for QALY calculations.
Other variables
For gender, man was used as the reference group. We used age as a continuous variable. We also tested age categorized into three age groups (20–34, 35–49, and 50–64), but this did not provide results that seemed to improve the statistical model. Education level was divided into three groups based on the question “What is your highest education?” – 9 years at public school or less was categorized as “primary education”, “secondary education”, and university or college studies was categorized as “university” – with primary education being the reference group. For marital status, the response “living with wife/husband/cohabitant/partner” to the question “How do you live?” was coded as “married” and was used as the reference group, while other responses to the question were defined as “single” and used as the exposure group. Previous health was defined from the question “How was, in general, your health five years ago?”, where the responses “very good” and “fairly good” were defined as “good” and used as the reference group, while the other responses (“fair”, “fairly bad”, and “very bad”) were defined as “poor”.
Statistics
In our analyses, propensity score weighting was used [15]. Propensity scores were introduced in 1983 by Rosenbaum and Rubin [16], and they correspond to the conditional probability of being assigned to the exposure group based on baseline covariates. A more thorough description and explanation of the use of propensity scores in the current study is available in Norström et al. [6].
We used logistic regression with our potential confounders (gender, age, education level, marital status, and previous health) as covariates in order to estimate the propensity scores. In our study, the propensity score corresponds to the probability of being unemployed given his or her characteristics. Comparisons between an exposed and unexposed individual with the same propensity measure is therefore similar to analyzing exposure in a randomized controlled trial. Using the propensity score approach for observational studies is a quasi-experimental approach.
We used propensity scores weighting, as we favored this approach above matching, and stratification, which are other popular propensity score approaches [17]. We expected the other approaches to perform poorer as they would include fewer unemployed participants because of matching problems with employed individuals. However, there is still a lack of consensus about recommendations on when to use the different approaches. For further discussion about the propensity score approaches, see, for example, Schroeder et al. [17].
In our results, we used the risk difference, which corresponds to the marginal effect of becoming unemployed, with counterfactual arguments. We used an inverse probability weight estimator, as suggested by Lunceford and Davidian [15], to estimate the risk difference
$$ {RD}_{IPW}={\left(\sum \limits_{i=1}^n\frac{X_i}{PS_i}\right)}^{-1}\sum \limits_{i=1}^n\frac{Y_i{X}_i}{PS_i}-{\left(\sum \limits_{i=1}^n\frac{1-{X}_i}{1-{PS}_i}\right)}^{-1}{\sum \limits_{i=1}^n}_i\frac{Y\left(1-{X}_i\right)}{1-{PS}_i}, $$
where Y refers to the outcome (health-related quality of life). The marginal effect from this estimator corresponds to the average treatment effect [18]. The standardized difference was calculated, both with and without a weight, to assess the balance of covariates between the employed and unemployed groups for each potential confounder [6, 19].
To be part of our analyses of the descriptive system and the QALY scores, it was required that participants had responded to all variables, including the five health dimensions, with a valid response. Of the 837 individuals who were active in the labor market, 788 were part of our analyses. Of the 49 excluded participants, 14 had not answered at least one of the EQ-5D questions, 14 had no response to education level, 3 had no response to marital status, and 20 had no response to previous health. Two of them had no response to at least two of the variables. Valid responses were also required for all variables for the EQ-VAS analyses. There were 771 valid responses for the EQ-VAS analyses.
Descriptive statistics were used to present the characteristics of the sample, and stratified results were derived for each covariate for the outcome variables. Analyses were carried out for QALY scores, EQ-VAS, and three of the dimensions of the EQ-5D separately as outcome variables. For the first two dimensions of the EQ-5D (mobility and self-care), too few participants reported any problem, and these results were therefore only presented descriptively. Pearson’s χ2-test was used to test if the exposure variable (labor market status) was associated with potential confounders. Student’s t-test was used to test differences in age with respect to QALYs between the employed and unemployed.
There are some potential problems with low QALY scores, such as a large gap between QALY scores if responding with some or extreme problems, low QALY scores potentially being related to poor employability, and a low QALY score potentially implying poor health already ahead of unemployment. We therefore performed two different sensitivity analyses. In scenario 1, we chose to exclude those who had answered extreme problems to any of the first three EQ-5D questions (mobility, self-care, and usual activities), and in scenario 2 we excluded participants who had answered extreme problems to any of the EQ-5D questions. Sensitivity analyses were not performed for EQ-VAS.
For logistic regression, it is recommended that the number of individuals of the least occurring event, in our case unemployed, divided by the number of explanatory variables should be at least 10 [20]. This condition was not fulfilled in many of our stratified analyses, which we have indicated in our tables. Also, for other grouped analyses the interpretations should be handled with care due to the small number of unemployed. Interactions between variables were not considered in any of our analyses. We did not experience problems due to collinearity between variables, and hence all candidate variables were kept in the analyses.
R Studio was used for statistical analyses (R Studio, Boston, MA), with its GLM procedure used for logistic regression, where confidence intervals were derived with the profile likelihood [21]. The Bootstrap technique with replacement was used to derive the mean square error from 10,000 replicates. Confidence intervals corresponded to the 2.5 and 97.5% percentiles of the Bootstrap simulations [22]. Based on the Bootstrap simulations, p-values were derived. Statistical significance was defined at the 5% level.