Theory and hypotheses

Decommodification is an outcome of generous welfare states. Strong social programs provide individuals with resources and options allowing them not to accept a job at any condition [12]. The welfare state also provides public transfers and services enhancing citizens’ autonomy and security, their capacity to stay in good health, and even their subjective well-being [19]. Social programs cushion the impact of individual life events on health [20]. Job displacement, for instance, is associated with poorer health in the long-term [21], but more generous unemployment insurance and more protective labour market regulations, clear forms of decommodification, reduce the association between unemployment on health [6, 22]. Similarly, Levecque et al. [23] find that the welfare state attenuates the consequences of economic hardship for health. However, Voßemer et al. [8] observe that generous unemployment insurance improves well-being but not self-reported health.

We know from the literature on population health that welfare state regimes [24] or social expenditures are associated with better health outcomes, whether one is comparing OECD countries, US states, or Canadian provinces [25–27]. Bambra [28] establishes a cross-sectional correlation between the welfare state generosity index and infant mortality, whereas Beckfield and Bambra [2] reach similar conclusions by extending the analysis with time-series cross-sectional modelling. Studies on specific social transfers such as social assistance or unemployment insurance reinforce these findings associating benefit generosity with positive health outcomes [4–7, 29].

The original work of Esping-Andersen on decommodification was criticized by feminist scholars for neglecting gender [30]. Esping-Andersen admitted the problem and introduced the concept of familialism to take gender into account [31]. Bambra et al. [32] established that indeed gender differences in self-reported health are related to welfare state regimes. More generally, we can expect gender differences in the relationship between decommodification and health. The association of welfare state generosity and health should be stronger for women than for men, since women rely more than men on health care, social services, programs to support families, and basic pensions [13]. Dahl and van der Wel [33], for instance, find that social expenditures are associated with lower health inequalities among women, more than among men. Women also receive less from employer-based contributory pension plans, are more likely to be poor in old age, and benefit more from decommodified basic pensions [34]. Björn Högberg [14] observes that while good standard pensions benefit men’s health, generous minimum pensions improve women’s health in old age, as does higher spending on eldercare (see also [35]). We thus formulate the following hypotheses about the direct and gendered relationship between decommodification and death rates:

H_1a: Welfare state generosity is associated with lower death rates

H_1b: And this association is stronger for women than for men

An intuitive interpretation of this hypothesis would relate the association between the welfare state and health to income redistribution [9, 20, 21]. Indeed, a vast and largely consensual literature establishes a link between the welfare state and redistribution [36]. Several studies, however, contend that there is no association between overall redistribution and population health [10, 11, 26]. We argue that the association between welfare state generosity and health stems not so much from the total redistributive effect of the welfare state but rather from the provision of better social protection. We thus expect to see a positive association between decommodification and population health, not mediated by income redistribution.

H_1c: Welfare state generosity is associated with lower death rates even when controlling for income redistribution

In addition to the direct association between decommodification and health, we expect two indirect connections. First, decommodification transforms the labour market: the institutional configuration of a welfare state regime structures the distribution and conditions of employment [12, 31]. We argue that decommodification reduces labour market polarization and lowers labour market risks such as large income losses due to unemployment, illness, or the loss of a spouse [18]. In turn, both factors are associated with better health.

Post-industrial, globalized, knowledge economies may be re-commodifying because they “demand flexibility and create insecurity” [31]. As market incomes and employment become more unstable and polarized, low- and middle-class earners increasingly stand “under pressure” [37]. A polarized labour market, defined by a growing gap between the highly skilled and highly paid who benefit from economic change and those at the bottom who do not, may be related to stress and poor health [11]. While the welfare state brings social protection associated with public health, a polarized labour market engenders insecurity and is associated, over time, with negative health consequences.

This relationship between labour market polarization and health has not been documented previously and it is therefore not easy to identify the exact mechanisms at play. One can think of a few plausible pathways. On one hand, polarization can be seen as an indicator of a less regulated labour market that does not prevent the creation of precarious, poorly paid, and stressful jobs, and probably makes it difficult for people with unsatisfactory working conditions to leave their jobs, at the risk of their own health [38]. On the other hand, the rise of top market incomes also reflects the political weight of the rich, who are less likely to favour universal, generous, and healthy social programs [39].

A welfare state that decommodifies social relations should mitigate the impact of globalization and technical change and prevent or moderate labour market polarization [31, 40]. Decommodification should also reduce the incidence of labour market risks. The welfare state works largely as an insurance mechanism against social risks such as employment termination, the loss of a spouse, sickness, disability, retirement, childbirth, or poverty [18]. We know from previous research that generous social policies can counter the consequences of adverse life events [41] and that large income losses associated with unemployment or economic hardships reduce population health [6, 7, 20, 22, 23, 29]. Welfare state decommodification is associated with lower labour market polarization and with a reduction of risk incidence, which in turn are related to population health. We thus propose the following hypotheses:

H₂: Welfare state generosity is associated with lower labour market polarization and a lower risk incidence

H_3a: Labour market polarization is associated with higher death rates

H_3b: The reduction of risk incidence is associated with lower death rates

In sum, we argue that one well-known outcome of generous welfare states, decommodification, is associated with public health directly, as it provides better social protection, and indirectly, as it favours working and living conditions that make for better, more secure, and healthier lives. The common trait shared by these two outcomes of decommodification is less the overall redistribution of incomes than the level of protection and security citizens obtain against social risks.

Data

Our dependent variable is the age-standardized death rate (from all causes) per 100 000 inhabitants, for both men and women, extracted from the OECD Health Statistics. Mortality is selected as an indicator because it is one of the most important measures of population health in comparative studies [5, 26]. In the Supporting information section, we present similar results with life expectancy at birth and at 65 years old.

The measurement of decommodification is subject to several debates and critiques. Many studies on the social determinants of health assess decommodification with welfare regime typologies contrasting three or four ideal typical regimes. This approach has the advantage of including difficult to measure regime-specific institutions [24]. Studies comparing health outcomes by regime types, however, are less likely to find significant relationships than those relying on time-variant measures [42], because regimes are time-invariant black boxes including several distinct policies [4, 33].

Others have focused on levels of public social spending, assuming that “the bigger-the better” [25, 27, 33]. Expenditures, however, are not always a good measure of decommodification as they fail to distinguish between the redistributive objectives and coverage of different programs [12, 31]. Expressed as a proportion of GDP, spending levels are also subject to fluctuations in the denominator unrelated to policy changes.

To improve upon Esping-Andersen’s early attempt at designing a measure of decommodification going beyond expenditures, Scruggs, Jahn, and Kuitto [16] constructed a welfare state generosity index that has become the standard measure of decommodification. This index is based on the generosity of three social programs addressing traditional social risks: loss of income due to sickness, unemployment, and old age. For sickness and unemployment insurance, the indices combine the income replacement rate for a typical person earning 100% of a country’s average income, the qualification period required to obtain the social benefit, the percentage of the population covered by this public insurance, and the duration and waiting period to receive the social benefit. The components for pensions include standard benefit replacement rates, standard qualification years, ratio of employer to employee pension contributions, and the coverage rates. Distinct measures thus cover sickness insurance, unemployment benefits, and pensions: we analyze their relationship to health when combined in an overall welfare state generosity index or when taken separately. A more detailed explanation of the construction of the generosity index is presented in the Supporting information section.

While we are convinced that the generosity index is the best available measure of decommodification, this index is not without shortcomings. It covers only parts of the welfare state, focusing on income protection rather than services and excluding, for example, family policies or social assistance. Moreover, the index is based on a specific group of the population: workers earning the country’s average wage. This focus led to a feminist critique of the decommodification index. Several post-industrial social programs aim instead to “re-commodify” women’s lives by favoring their participation to the labour market [30]. Aware of these critiques, we compare the effect of welfare state generosity on both men’s and women’s health.

Hypothesis 1b proposes that welfare state generosity has a direct association to population health that is not conditional on income redistribution and inequality. The Gini index for disposable income (after taxes and transfers), the most common indicator of income inequality, is used to measure income inequality. Following the standard approach, we calculate the size of redistribution as the Gini index before taxes and transfers (market Gini) minus the disposable income Gini (after taxes and services), divided by market Gini ([market Gini—disposable Gini] / market Gini). Note that we are interested in the total redistributive effect of the welfare state and that we recognize that specific programs, such as unemployment insurance, may have a redistributive effect on their own and be associated with population health.

It is not possible, within the confines of a single article on population health, to account satisfactorily for labour market polarization, a complex phenomenon with numerous determinants. The Gini index is not a good indicator for such polarization since it is more sensitive to changes in the middle of the income distribution and less affected by trends at the extremes (poor and rich) [43]. Labour market polarization and income inequality in the middle of the distribution are two distinct phenomenon driven by different factors [44]. To test hypotheses 2 and 3 about labour market polarization, we follow common practice and use the p90p10 income ratio, which is the ratio of market income between the top and bottom 10% of the distribution, before taxes and transfers [45]. To ensure that the relationship for labour market polarization does not simply reflect the association between the overall distribution of market income and health, we control for the market income Gini, which captures inequality in incomes from labour, capital, and private transfers (the results are presented in the Supporting information section).

To assess risk reduction, we use Jacob Hacker and Philipp Rehm’s new measure of the welfare state insurance function which has never been applied in studies of population health [18]. Their measure of risk reduction is based on panel surveys tracking household finances over time to calculate the importance of risk incidence mitigation by the welfare state. This measure is defined as the degree to which taxes and transfers reduce the frequency and severity of major losses in household income. Hacker and Rehm calculate the number of working age people who experience a 25% loss in market income but were insufficiently buffered by taxes and transfers to prevent a loss of 25% of disposable income, divided by the total number of working age people who experience a 25% loss in market income. In the Supporting information section, we replicate the models with a threshold of risk incidence set at losses of 50% of income.

We control for the percentage of the population aged 65 and over, the unemployment rate, and GDP per capita, three factors that are potential confounders between welfare state generosity, income distribution, or population health. An aging population and high unemployment rates are related to poorer health, increased social spending and to fiscal stress that may be associated to welfare state generosity, whereas GDP per capita decreases fiscal stress and could be related to population health. We also control for average per capita alcohol consumption, a good indicator of lifestyle habits that influence health and varies across countries and over time. Our dataset is primarily constrained by the availability of the generosity index, and it includes 21 OECD countries from 1971 to 2010. The dataset gets even smaller when considering risk reduction, a measure available for a shorter period. The descriptive statistics, the list of countries, temporal coverage, and the sources for the different variables are presented in the Supporting Information section.

Models and methods

Our models are based on Eq 1, which is an autoregressive distributed lag equation: (1)

Y is the death rate at time t in country i. It is predicted by a constant a, a lagged dependent variable Y_t-1, the independent variable X_t-5 (either the generosity index or measures of redistribution, labour market polarization, or risk reduction), a vector of controls Z_t-5, a country fixed effect φ_i, a country-specific time trend τ_it an error term ε_it. We use panel corrected standard errors to correct for heteroskedasticity in the error term, as well as an AR1 correction for serial correlation in the error term [46].

Several statistical considerations justify this choice of model. The strong serial correlation in the dependent variable militates for the inclusion of a lagged dependent variable (LDV). However, the inclusion of the LDV represents a trade-off between omitted variable bias and overfitting, since the lagged dependent variable explains a large proportion of the variance. While similar macro-level studies in population health are not using LDVs [2, 25, 47], in political science and economics, it is common to prevent omitted variable biases by using LDVs [48]. We thus present models with LDVs in the main manuscript and without a LDV in the Supporting information section.

We tested different lag structures and found that the relationship between our variables of interest and population health tends to take five years to unfold. Therefore, in the main analysis, we present models with five years lags (see [47] for a similar modelling strategy). These models assume there is no immediate relationship between X and Y. In the Supporting information section, we supplement the analysis with models including a one year and a five-year lag.

Our models include country fixed effects. These are technically justified by a Hausman test but also theoretically warranted since we cannot control for several unobserved factors influencing the health of a population. By including country fixed effects, our equations analyze the relationship between an independent variable and the dependent variable within a country. To model the negative, but country-specific, trend in age-standardized death rates, we include a country-specific time trend. Using fixed effects, country-specific time trends and LDVs to identify the relations with a slowly moving dependent variable is a very conservative modelling strategy, that purposely makes it difficult to identify significant relationships.

We run stationarity tests on each variable. They show that the generosity index, the p90p10 ratio, and the risk reduction variable are stationary, while the dependent variables are stationary around a trend. The disposable income Gini and the control variables (except the unemployment rate) are non-stationary, whereas the tests are uncertain for the redistribution measure. Regressing a non-stationary variable on a stationary series biases the equation [49]. We thus first difference the disposable income Gini and the control variables, except for the unemployment rate, to make them stationary [50]. We choose to present the redistribution measure in levels, to allow for a more meaningful comparison with the effect of risk reduction and the p90p10 ratio, but present it in first difference in the Supporting information section in case it would be non-stationary.

The next section presents the results. It starts with a test of H_2, associating the generosity index to risk reduction or labour market polarization. Then, we assess H_1a and H_1b by regressing welfare state generosity on the age standardized death rate following equation 1. Also using equation 1, we then test H_3a and H_3b by regressing either risk reduction, labour market polarization, disposable income inequality or income redistribution to predict the death rate. From these models, long-term relationships can be calculated, to represent the cumulative association of a change in the independent variable on the dependent variable as T reaches ∞. Their coefficient is determined by the following formula b₁+b₂/1-a₁, while their standard error is calculated with the Bewley transformation [50]. Finally, to consider H_1c (welfare state generosity is associated with lower death rates even when controlling for income redistribution), we present additional models including measures of income redistribution and generosity together in the same equation. We control the income redistribution at T-5 and T-1 allowing for the association of the generosity index and the death rate to be mediated by generosity’s association with the income redistribution over time. We also present models confirming that the generosity index maintains an association with mortality, even when controlling for the p90p10 ratio and for risk reduction.