2.1. Data Source
The analysis is carried out on data from the third release of EU-SILC survey of Eurostat, including information on citizens from 26 European countries followed from 2009 to 2012 (Hungary and Croatia are excluded from the study due to the missing information about the employment status for year 2009). The original dataset is made of 127,199 individuals entered in the study in 2009; the 63.02% of these subjects (corresponding to 80,159 individuals) remains in the panel along the full time period 2009–2012. More precisely, from this balanced subsample we only consider individuals who were in working age (i.e., 17–64 years old) in the time period 2009–2012 and we drop all those subjects suffering from limitations in activities because of health problems or from any chronic (long-standing) illness as well as individuals who are unable to work in order to reduce the sample selection effect. Individuals who did not provide information about the type of job contract (open-ended vs. temporary) and individuals with missing values on one or more covariates of our interest are also dropped. The final total sample size amounts to 26,898 individuals and 107,592 observations (i.e., four measurements for each individual). Details on modalities of data collection, comparability of data between countries and over time, response rates, and any other question concerning the quality of data are provided by the official EU-SILC documentation freely available at
http://ec.europa.eu/eurostat/web/income-and-living-conditions/overview.
SHS is described by an ordered polytomously-scored variable taking values 0 to 4 from
very poor to
excellent perceived health. As shown in
Table 1 (first row), the great majority of respondents (92.34%) declares that his/her health status is good or excellent.
The employment status is a categorical variable defining different profiles in the labor market. We account for the duration of the contract by distinguishing between
permanent employees, that is all dependent workers with an open-ended contract, and
temporary workers, who are all employees with non-standard labor contracts including persons with a seasonal job, persons engaged by an employment agency and hired out to a third party to carry out a work mission (unless there is a work contract of unlimited duration with the employment agency or business), and persons with specific training contracts (see item PL140 of the EU-SILC questionnaire at
http://ec.europa.eu/eurostat/web/income-and-living-conditions/overview).
We also consider a complementary definition for the labor market status according to the level of time effort and thereby discerning between full-time and part-time employees, alongside permanent and temporary employees. This definition allows us to compare the results of the empirical analysis with an alternative labor market profile that may capture a form of underemployment in terms of hours worked. Notice that, in this case, we are not able to prevent the bias from reverse causality to arise in our empirical analysis, as we do not have information on the reason for working fewer hours. The resulting classification allows us to distinguish four categories of employees: full-time permanent workers, part-time permanent workers, full-time temporary workers, and part-time temporary workers.
Individuals are defined as unemployed if they did not work during the year preceding the interview and are actively looking for a job. This definition of unemployment avoids that bias arises in the empirical analysis from the bidirectional nature of the relationship between SHS and unemployment: by considering only individuals looking for a job as unemployed, we are excluding the possibility of those persons that do not work because of bad health. The other modalities of the employment status distinguish between self-employed, that is, all self-employed workers with and without employees, and other individuals, which are inactive, such as students, home makers, retired workers, individuals in further training or unpaid work experiences, individuals in compulsory military or community service. In absence of further information, the last two categories can only be considered as residual in the empirical analysis since a reverse causality problem may emerge from workers that have chosen to be self-employed (with, for instance, more flexible hours compared to employees) or to stop working because of health limitations. Nevertheless, we keep these categories in our empirical analysis as their exclusion may result into a further selection bias problem.
The main interesting element arising from the distribution of SHS by employment categories concerns unemployed persons (
Table 1): more than 11% of them declares an at most fair level of SHS. Instead, the distribution of SHS for
temporary and
part-time is very similar to that of
permanent and
full-time employees, respectively.
We observe some differences in the distribution of SHS by country and by the other covariates included in the analysis, as illustrated in
Table 1. In particular, the 7.49% of total sample declares a fair level of SHS, however this percentage rises to 10.75% for citizens from Czech Republic, 13.93% for Poland, 18.87% for Portugal, 21.33% for Latvia, and 30.22% for Lithuania. Latvia and Lithuania are also those countries with the smallest percentage of persons with an excellent level of SHS: 7.74% and 7.10%, respectively, against the 36.54% of the total sample. Other countries with relatively few persons declaring an excellent level of SHS are Portugal (14.66%), Estonia (17.23%), Italy (21.99%), and Bulgaria (24.85%). On the other hand, countries with the best level of general self-perceived health are Cyprus (CY, 99.11% of interviewees evaluates as good or excellent the own health status), followed by The Netherlands (98.12%), Iceland (97.77%), Greece (97.74%), Spain (96.93%), Belgium (96.51%), Sweden (96.25%), United Kingdom (96.15%), and Austria (96.06%).
In the following inferential analysis, we control for the effect of the following set of covariates: household income per person, gender, age at baseline, year of interview (2009, 2010, 2011, 2012), education level (primary, secondary, tertiary), marital status (cohabitant on a legal basis, cohabitant without a legal basis, single). We also control for the unemployment rate (proportion of the labor force reporting unemployment) specific for each country and year (available at
http://www.oecd.org/employment/labour-stats/), which provides a measure of relative deprivation, since the well-being of temporary workers and unemployed individuals may decrease less if they live in environments with a high unemployment rate [
23]. From
Table 1 we observe that young persons and males, with a higher income and a higher level of education, living by yourself or cohabitant without legal basis and living in countries with smaller levels of unemployment rates tend to have a better perception of their own general health status.
In order to account for economic deprivation, beyond the monetary household income, we include in the model specification a set of items available in the survey that refer to two main aspects related to economic deprivation: the economic strain and the material deprivation. The economic strain attains to a general financial distress and it is measured by the ability to make ends meet (with difficulty, with some difficulty, fairly easily, easily). The material deprivation encloses the following items, according to the definition provided by Eurostat (for details see at
http://ec.europa.eu/eurostat/statistics-explained/index.php/Material_deprivation_statistics_-_early_results): presence of arrears on mortgages or rent payments, on utility bills, on hire purchase installments or other loans (no, yes), capacity to afford paying for one week holiday away from home (no, yes), capacity to afford a meal with meat, chicken, fish (or vegetarian equivalent) every second day (no, yes), capacity to face unexpected financial expenses (no, yes), possession of telephone, color TV, washing machine, car (yes, no-cannot afford, no-other reasons), and possession of an heating system to keep home adequately warm (no, yes). In addition to the mentioned items, we consider other three covariates that are available in the survey and refer to the presence of a leaking roof (no, yes), the level of financial burden of the total housing cost (heavy, slight, not burden at all), and the possession of a computer (yes, no-cannot afford, no-other reasons).
As shown in
Table 2, the more impaired is the economic situation of a person, the worse his/her self-perceived health condition. For instance, the percentage of persons with an at most fair level of SHS decreases from 10.58% for those making ends meet with difficulty to 3.67% for those making easily ends meet. Similarly, for the items related to material deprivation the cumulative percentage of responses in categories
very poor,
poor, and
fair is 4–7 percentage points lower for persons without economic deprivation compared to persons with an impaired economic status. Notice that we do not construct indicators for economic deprivation but just keep the detailed items provided by the EU-SILC. We do so in order to account for as much as the information on perceived and structural deprivation as available, while we do not infer on its direct interpretation, not being the focus of the present work.
Further, in order to explore the association between employment status and economic deprivation, we provide with
Table 3 showing the conditional percentage distribution of employment status given the economic deprivation. Generally speaking, we observe a more impaired economic situation of unemployed individuals related to workers with permanent jobs. A significant association (Chi-square independence test with
p-value less than 0.001) emerges with respect to all the economic deprivation variables taken into account. A particularly high association is related to ability to make ends meet (Cramer’s V = 0.1427), presence of arrears on utility bills (Cramer’s V = 0.1356), capacity to afford paying for one week holiday away from home (Cramer’s V = 0.2267), capacity to face unexpected financial expenses (Cramer’s V = 0.2042), possession of an heating system to keep home adequately warm (Cramer’s V = 0.1131), and level of financial burden of the total housing cost (Cramer’s V = 0.1364).
Other than the association between employment status and economic deprivation, data show that the distribution of the employment condition is also related to the country (
Table 4). For instance, the percentage of full-time permanent workers ranges between 28.68% (Greece) and more than 74% (Norway and Denmark) and the percentage of part-time permanent workers ranges from 0.17% (Romania) to 30.25% (The Netherlands). Moreover, the percentage of full-time temporary workers is minimum for Estonia and UK (less than 1.00%) and maximum for Poland (12.10%), whereas Norway and The Netherlands show the smallest percentages of unemployed individuals (less than 1.00%) and Spain and Latvia the highest percentages (12.18% and 13.20%, respectively). The association between employment status and country reflects the different welfare systems characterizing each country. Indeed, aggregating countries by welfare regime, according to the classification provided in [
24], a more homogeneous distribution of the employment status is observed within each group of countries (
Table 4). We explicitly take into account the association at issue in our inferential analysis, according to the approach illustrated in
Section 3.2.
Finally, it is worth to outline that, with the exception of country, gender and age at baseline, all the other variables are time-varying. However, for these variable we observe a high persistence of individuals in the modalities declared at the first year of interview. In particular, almost the 85% of interviewees remains in the same employment status over the period 2009–2012. We return on this point in the following section.
2.2. Statistical Analysis
The longitudinal structure of our dataset, with repeated measurements observed on a set of individuals along years 2009–2012, drives the choice of the statistical method for the data analysis to the class of random-effects (or multilevel or hierarchical) generalized linear models for panel data. In this section, we illustrate the class of models at issue; for more details see, among others, [
25,
26,
27,
28,
29]. More in detail, due to the ordinal nature of the dependent variable, we estimate a random-effects ordered logit model [
27,
28,
30,
31].
Let be the observed ordered response variable SHS, assuming values for individual i at time t, with and . Moreover, let be a vector collecting the time-varying observed variables and a vector for the observed time-constant individual characteristics (e.g., country and gender).
The random-effects ordered logit model is formulated according to a link function based on global logits [
32], as follows:
where
is a subject-specific random intercept specified in terms of fixed and random parameters:
where
is a random parameter that summarizes the unobserved individual characteristics, which are time-constant and affect the probability of answering
y across repeated measurements of SHS. As usual, the random effects
are assumed to be normally distributed with mean equal to 0 and constant variance
. We also consider the potential correlation between the employment status in
and the individual unobserved heterogeneity, as the latter may non-randomly group workers into different labor market categories and bias their effect on SHS (for instance, a risky health-related life-style may also affect the worker’s employability). In order to tackle this issue, we employ a correlated random-effects approach [
33,
34,
35], in which a parametric formulation of the time-constant correlation between the random effects and covariates of interest in
is specified. Vector
denotes a transformation of selected covariates in
, aimed at capturing the dependence between the corresponding elements in
and
. In particular,
includes four time-invariant covariates, one for each year in the sample, that represent the employment status of individual
i in each year (for instance, if individual
i is in employment condition
in 2009, where
is one of values taken by the employment condition covariate, then the first covariate we create will take value
in all the four time occasions; if
i is in condition
in 2010, the second covariate will take value
for the four time occasions, and so on). This is a standard strategy for correlated random effects when the variable of interest is qualitative [
34]. Furthermore, we also consider the individual income by including in
its average value, which represents a standard transformation for continuous covariates [
35].
Substituting Equation (2) in Equation (1), the reduced-form of the random-effects ordered logit model is obtained:
From Equation (3), it is clear that the probability of observing a given value of SHS depends both on the observed values of covariates in , , and and on the value assumed by the random component for the i-th individual: values of much higher (smaller) than zero imply a higher (smaller) level of SHS compared to the “average” individual, being constant all the observed covariates.
The model at issue is estimated through the marginal maximum log-likelihood approach, consisting in marginalizing out the distribution of the random effects and maximizing the resulting log-likelihood relative to the unknown model parameters. The integral involved in the log-likelihood function cannot be solved in closed form and it is approximated in a weighted sum, using the adaptive Gauss-Hermite quadrature method (for details see [
27]).
In order to investigate whether economic deprivation is able to partly capture the effect of the employment condition on SHS, we adopt a control function approach [
36] that accounts for the correlation between economic deprivation and employment status, under the assumption that the dependence between log-odds ratio in Equation (3) and the set of economic deprivation covariates has a linear form. In practice, we add to the set of explanatory variables the indicators of material deprivation and economic strain provided by the survey.
We stress that the approach we adopt in this paper is based on the assumption that the correlation between the employment status and the unobserved random effects has a linear parametric form, as opposed to the fixed-effects approach, in which the
are assumed to be fixed parameters, and therefore robust to violations of the implicit parametric assumption in Equation (2). Unfortunately, as mentioned at the very end of the previous section, we verified that the time-varying variables,
in primis the one denoting the employment status (but also those denoting the economic deprivation), are quite persistent. Since the conditional maximum likelihood estimation method, on which the fixed-effects approach is based, only relies on the information provided by the time variation in the covariates, weak identification problems may arise in presence of highly persistent time-varying variables, since the Hessian of the log-likelihood may be close to being not negative definite [
37]. Moreover, the fixed-effects approach does not identify the effects associated with time-constant covariates. A common solution to this problem consists in estimating a different model for each modality of each time-constant variable. In our study, we are specially interested in the effect of the country, which consists of 26 modalities. We also remind that formulating a specific model for each modality of a given time-constant variable corresponds to assuming interaction effects between the variable at issue and each other variable in the model, which is different from Model (3).