Life expectancy inequalities in the elderly by socioeconomic status: evidence from Italy

A vast and growing literature shows that dramatic differences in health conditions and in longevity between individuals characterized by different socioeconomic conditions have emerged across the globe (e.g., [46, 69]).

In an effort to explain such a divide, many epidemiologic investigations on differential mortality take a “lifestyle” approach, focusing on individual choices. People who smoke, drink alcohol, are overweight, or engage in dangerous sexual behaviours have a higher risk of numerous diseases than others [22, 56], and all these dangerous behaviors are not randomly distributed within the population but tend to concentrate in the lowest social classes [31, 42, 67]. Moreover, the effects on health of these dangerous behaviors also differ according to socioeconomic status, being stronger for blue-collar workers and low-paid clerks than for managers, employers, and professionals [14]. In addition, those individuals who have suffered an economic, social, or physical disadvantage in the past are more at risk of further damage of greater intensity in the future, a phenomenon known as the “chain of disadvantages” because each negative event amplifies the negative effect on quality of life, health, and survivorship [7, 21].

As pointed out by Marmot [45], the social divide in health is a remarkably widespread phenomenon and is not confined to those in poverty. It involves all of society, from the top to the bottom, with poorer health conditions emerging at every step in the social hierarchy. Therefore, the social divide takes on the characteristics of a social gradient.

Many studies have confirmed the existence of this social gradient, using different proxies of individual “socioeconomic position,” such as education, occupation, or income,⁵ and some studies have also shown an increasing trend in the social divide (e.g., [24, 39, 44]). Specifically, in developed countries the type of job plays a fundamental role in the social gradient of health: employment conditions determine adult socioeconomic status (i.e., their relational network and social environment) and has a direct effect on psycho-physic health [49].⁶ The literature has identified two main mechanisms through which the work environment directly influences individual health: the “demand-control” model [4, 36] and the “effort-reward” model [61, 66]. Both models suggest negative effects on health for employees and positive effects for managers, employers, and the self-employed, mainly due to the different typologies of stress they face and to the alienation of results [14].

Initially, research into the relationships between socioeconomic position and health was primarily focused on the working-age population. Analyses of the socioeconomic determinants of health and survival at older ages are more recent, and the results are sometimes contradictory: some longitudinal and cross-sectional studies [13, 53, 65] have shown a strong association between the previous occupation and health and survival at older ages, while other studies did not find a statistically significant relationship [1, 19] or found that the association declines with age [20, 63]. Trying to disentangle such an apparent contradiction, Marmot and Shipley [48] argued that the occupational class may strongly influence pre-retirement mortality, whereas other socioeconomic determinants of health arise after retirement. However, most of the studies measured the occupational group as self-reported by the individuals in old age in a single moment. In contrast, both the life-course hypothesis and the chain of disadvantage theory emphasize the accumulation and interaction of advantages and disadvantages across the entire lifespan, as a sort of time “exposed to risk” [6].

As regards Italy, despite several efforts to investigate mortality differences by socioeconomic status at older ages, the available empirical evidence is still not exhaustive, due to limits in the available dataset.

A dataset based on the linkage of death events and the Italian census – where information on occupation and education of the deceased was recorded – was created by the Italian National Statistical Institute (ISTAT) in 1991, but it was not updated in the ensuing years. Therefore, other data sources have been used in order to estimate mortality by social class of the Italian elderly. Some studies [5, 18, 40] used the administrative archives collected by the Italian Social Security Institute (INPS), which, however, do not record important socioeconomic variables such as education or marital status and, in the versions delivered until a few years ago, did not record complete working histories. In order to overcome these shortcomings, a dataset merging INPS archives with death registries and census information has been developed, but this dataset only refers to the city of Turin or to the region of Piedmont, thus preventing scholars from extending the analyses to the entire Italian population.

Instead of relying on microdata, other scholars have developed analyses following a macro perspective, correlating average values in different geographical areas of death rates and socioeconomic and demographic variables of the local population [43, 51]. However, due to what is called the “ecological fallacy,” the aggregate-level relationship between socioeconomic status and the average level of mortality in specific areas may be quite different from the individual-level association between such variables.

We use an innovative dataset, called “AD-SILC,” built by merging the IT-SILC 2005 cross-sectional sample (i.e., the Italian module of the 2005 wave of the European Union Statistics on Income and Living Conditions – EU-SILC) and the administrative longitudinal records provided by the Italian National Social Security Institute (INPS). In detail, the cross-sectional variables collected in IT-SILC 2005 have been enriched by the longitudinal social security records since the entry in the labor market up to 2009.

Social security records offer a comprehensive picture of the working career of employees in the private sector (employees in the public sector were not enrolled with INPS until 2011) and all categories of the self-employed (professionals, though, are excluded because they are not registered with INPS) as they report, on a yearly basis and for each working relationship, region of residence, gross earnings, working weeks, the type of working relationship (thus allowing us to precisely distinguish the various categories of employees and the self-employed) and for pensioners, the type of benefit received (i.e., old-age pension, social pension, or disability benefit).

AD-SILC is innovative in the Italian context for two reasons: first, because of its retrospective (until 2005, the year of the EU-SILC interview) and prospective (from 2005 up to 2009) longitudinal design; and second, for its variety of available socioeconomic variables, unmatched before in any Italian study on social determinants of survival in old age.

From one side, AD-SILC is a retrospective unbalanced panel where longitudinal time-variant information on individuals’ working/pension histories is collected and linked to socioeconomic characteristics recorded in 2005. From another side, AD-SILC is a prospective dataset, with a base year in 2005 and a four-year follow-up. Indeed, we follow the histories of those interviewed in 2005 up until December 31, 2009. Therefore, the structure of the dataset allows us to observe the possible date of death until the end of 2009 of individuals who responded to the IT-SILC survey in 2005. Therefore, AD-SILC allows us to analyze mortality risks in the period 2005–2009 according to individual socioeconomic characteristics and according to their previous working history.

The AD-SILC panel size amounts to 1,162,045 observations, referring to 43,388 individuals interviewed by IT-SILC 2005 and recording at least once in INPS administrative archives. Since the aim of the present study is to estimate social determinants of mortality in old age, only individuals aged at least 60 on December 31, 2005, are included in the analysis (note that, according to OECD data,⁷ the effective retirement age was around 60 for males from the mid-1990s to the mid-2000s, while for females the effective retirement age was approximately one year lower). The sample used in the present study then amounts to 11,281 individuals: 5529 males and 5752 females. The age distribution in 2005 for males and females, respectively, is shown in Fig. 1a and b.

In order to carry out our analyses and to make the interpretation of the estimated coefficients easier, time-variant variables regarding working conditions over the whole life course are collapsed into a time-invariant variable, summarizing the prevalent condition along the career, where the prevalent condition has been defined as that with the longest time span over the working career. For instance, if an individual worked 10 years as an employee and 30 years self-employed, he/she is identified as self-employed. Note that our definition of the prevalent condition is independent of the final status before retirement; in the previous example, the individual would have been identified as self-employed even if he/she had spent the final part of his/her career as an employee.

We choose to rely on time-invariant covariates synthesizing the career trajectory instead of using more comprehensive information about the whole career for two reasons: i) since individuals might have had very complex employment trajectories (e.g., some years as a farmer, then an unemployment spell, then some years as an employee, then self-employed) it is extremely complicated, and in some sense arbitrary, to exogenously define – or to endogenously identify, for instance by using the sequence analysis – a wider set of categories summarizing these trajectories (and also the interpretation of the main findings of the empirical analysis would become less clear than if we consider a simple set of prevalent working statuses); ii) mostly we consider working categories that are rather invariant over the individual career; indeed, our data show that the prevalent condition is a very good proxy of the whole working trajectory: on average, in our sample, 87% of the individuals spent at least four-fifths of their working career in the status that we consider as prevalent and only 1% spent less than half of their working life in the status we consider as prevalent.

Working condition is identified in AD-SILC through information from the pension fund that the individual contributed to before they retired. Italian workers pay pension contributions to different funds depending on their working category: specific funds exist for employees in the private sector and for various categories of the self-employed, i.e., craftsmen, shopkeepers, and farmers (as mentioned, we do not have information on employees in the public sector or on professionals). Furthermore, for private employees, INPS archives also record occupation, which is coded through a dummy variable – white-collar versus blue-collar. Henceforth, exploiting information on the prevalent working category and the prevalent occupation during the working life, we distinguish five occupational groups: blue-collar employees in the private sector; white-collar employees in the private sector; self-employed craftsmen; self-employed shopkeepers; and farmers. Table 1 recaps the content of variables used in the present study.

Exploiting information recorded in INPS archives, we also distinguish individuals according to the type of pension benefits received, which is strictly linked to their previous working history. We distinguish four categories of pension benefits: old-age pensions (pensions based on previous work contributions); disability pensions (paid to those injured at work and unable to continue working); social disability pensions (paid to the disabled, independent of the source of their disability); and social pensions (where we group two means-tested benefits directed to poor pensioners, namely minimum pensions and social assistance benefits).

Therefore, the subgroups we consider are not representative of the whole Italian elderly population since we include neither professionals and employees in the public sector – two categories that are likely to be characterized by a relatively higher life expectancy – nor those individuals who never worked (often female housewives) or have always worked in the informal sector (and then do not receive an old-age or a minimum pension when retired) and are not entitled to a disability benefit.⁸

Finally, IT-SILC records information, observed in 2005 on educational attainments (coded through the ISCED classification and grouped into three categories: at most primary educated, lower secondary educated, and at least upper secondary educated) and marital status. To better capture permanent household economic well-being, we also include in our analysis the information provided by the subjective variable self-reported by individuals in 2005 on the “ability to make ends meet,” whose six responses are grouped into three categories – low, medium, and high ability – for simplicity.⁹

As argued above, the main aim of the present study is to estimate whether life expectancy of older Italians differs by their socioeconomic status. In particular, we investigate the association between having belonged to a certain occupational group during the working career and longevity at older ages, controlling for other individual demographics (i.e., gender, age, and marital status) and socioeconomic characteristics (i.e., education and household subjective economic condition).

We carry out our estimates by using the Cox semi-parametric proportional hazards model, which is the best-suited estimation model for samples containing truncated and censored data [2, 17].¹⁰

Alternative models to be applied to longitudinal microdata might be a parametric model (for example with a defined survival curve such as a Gompertz) or an accelerated failure time model (AFT). We prefer to use the Cox model because we do not want to invoke parametric assumptions about the shape of the baseline hazard function. Even if there is strong evidence in literature about the shape of the hazard function with respect to old-age human hazard, relaxing any assumption we will obtain survival data free to adapt to the real Italian survival curve, as officially estimated by ISTAT on the entire Italian population, which is our last step in this study, as explained further. However, we also estimated an alternative model using a Gompertz parametrization of the survival curve, obtaining substantially the same results with respect to the variable of interest in this study (see the results of the alternative Model 4b in Table 9, Appendix 2). On the other hand, AFT does not consent to perform a proper analysis when several censored and truncated data are present in the sample [54].¹¹

We assumed “age” as time of entry into and exit from the dataset, and we estimated the parameters of the model using a MLE, considering both left truncation and right censoring. Moreover, we applied the Efron [26] method to overcome the problem of presence of ties in the dataset.

Afterward, we estimate the baseline hazard and the baseline survival curve, which only depends on time, not considering cohort stratification. Finally, using the Kaplan-Meier estimator it is possible to simulate the survival curve of a hypothetical social group with a certain covariate profile [64].

Assuming the individual age as the time of entry into and exit from the dataset, the computation of simulated survival curves stops at age 86, due to the limited sample size (less than 100 observations) in our dataset of individuals alive and aged over 86 by December 31, 2009. Therefore, it is possible to compute life expectancy at 60, but we have to truncate life expectancy at age 86. This outcome is then useful for a comparison between social groups but is not comparable to the Italian life expectancy at 60 because of its truncation.

However, using the Brass relational model, which is often used in the presence of incomplete or untrustworthy survival curves [12, 70], and comparing the Italian survival curve as officially computed by the ISTAT with the Kaplan-Meier simulations, we can compute a complete survival curve of the hypothetical groups from age 60 to age 105.

Brass suggests the following logit and anti-logit transformations:

where \( {\boldsymbol{l}}_{\left(\boldsymbol{x}\right)}^{\boldsymbol{s}} \) are the values of a “standard” survival curve (complete and trustworthy, in this case the official Italian curve by gender, certified by ISTAT) and l_(x) are the values of the uncomplete survival curve (in this case, the Kaplan-Meier simulations for every group defined by a “k” covariates profile).

The model then supposes a linear relationship between the logit of every human survival curve, as expressed below:where the values \( {\boldsymbol{Y}}_{\boldsymbol{x}}^{\boldsymbol{S}} \) are the logit of the standard survival curve and the values Y _x are the logit of the uncomplete Kaplan-Meier simulations and α and β are the parameters estimated by a linear regression.

Following this simulation methodology, we can compare life expectancies of individuals belonging to different socioeconomic groups. Therefore, our study represents the first link for Italy between results based on a micro survival analysis with results related to the entire Italian population.

As a first outcome of our empirical analysis, we present a validation of both the sample and the methodology previously described. Figure 2a and b show, respectively for males and females, a comparison between the real survival curves of the Italian population and the curves simulated according to our data using the Kaplan-Meier estimator on Cox parameters. Results of the validation test confirm the goodness of both the AD-SILC sample and the methodology developed in the present study: the values of Italian survival curve by sex, as officially computed by the Italian National Statistical office, are always within 95% confidence interval (CI) of the simulated Kaplan-Meier curve.

Moreover, we performed a test to check the assumption of proportionality, analyzing the Schoenfeld scaled residuals [64]. From our analysis, the test results were not significant for any covariate except gender and disability pensions (p-value< 0.05). In particular, considering our preferred model where all sets of covariates are included and previous occupations are grouped in three categories (Model 4 in next section), the test results were not significant for both former job characteristics and self-perceived financial situation, which are the crucial variables of interest in this article.¹² To control for potential bias due to the violation of the assumption of proportionality, we also ran alternative models [9, 34, 62], stratifying by gender and typology of pension (results of alternative Model 4c are reported in the Appendix 2, Table 10), but we obtained substantially the same results, both considering the magnitude of the effects and the statistical significance of parameters.

Therefore, taking into account the validation we did using official data on the Italian population, the results of the Shoenfeld residuals test and the estimates from alternative models, we consider that our model is an appropriate representation of data.

We then show the results concerning the estimated association between individual characteristics – our main focus is on the previous occupation category – and mortality risk obtained by using the Cox semi-parametric model (Tables 2, 3, 4 and 5).

We start by estimating a first model (Model 1, Table 2) where we show the mortality hazard of different occupational groups, controlling for individual basic demographic variables such as gender, geographical area of residence, and marital status (age and cohort of birth are controlled by the structure of the model). Then, we estimate a second model (Model 2, Table 3), adding the type of pension benefit to the control variables, and a third model, where further socioeconomic characteristics are added to the regressors (namely, education and household socioeconomic conditions, as proxied by the subjective answer to the question about “ability to make ends meet”; Model 3, Table 4).

All models clearly show that those who were previously self-employed – as a shopkeeper or a farmer – are characterized by statistically significant lower mortality risks when elderly than those previously employed as blue-collar workers. Conversely, mortality risks when elderly for former white-collar employees are not significantly different with respect to former blue-collar employees (note that results are substantially the same when setting former white-collar employees as the reference group of our estimates). Likewise, the difference in mortality risks between those self-employed as craftsmen and employees is only statistically significant in the first model. Moreover, we find no significant difference in mortality risks between former self-employed shopkeepers and craftsmen.

According to these results, and for the sake of simplicity in order to run computations of the following sections, we have aggregated blue-collar and white-collar workers into a single “omitted” category for employees, and shopkeepers and craftsmen into a single category for the self-employed (Model 4, Table 5). As expected, the main findings on the link between mortality risks and previous working condition do not change.

However, the type of pension benefit might be related to health, since individuals in poor health could receive a disability benefit or could have worked only rarely and are then entitled to receive only a social pension. To deal with this issue and partially take into account a possible reverse causality between poor health and the type of working career, as a robustness check we also ran our estimates excluding from the sample those who receive a social pension or both types of disability benefits (Tables 11–12 in Appendix 2). Our main results about a significantly different mortality between self-employed and employees is confirmed also by these robustness checks.

Apart from the link between occupation and mortality, which is the main focus of this article, it is interesting to observe the estimated association between mortality risks and the control variables included in our regressions.

As expected, the typology of pension benefit has a clear association with mortality hazards of retired Italians. The disabled suffer a clearly higher mortality hazard with respect to those who receive a standard pension benefit (mortality hazard is 50% and 280% higher for those permanently injured at work and for the disabled, respectively, compared to those who receive an old-age pension), and these results also hold when controlling for further socioeconomic characteristics.

With respect to socioeconomic controls (Models 3 and 4), as expected, the better the household economic condition (captured by the “ability to make ends meet” variable), the lower the mortality hazard. The mortality hazard of those who report they make ends meet easily or very easily (the “high” group) is 45% lower than the hazard characterizing those who report making ends meet with great difficulty and with difficulty. Likewise, those in the “medium” group show a hazard that is around 20% lower than those belonging to the “low” group.

Conversely, less clear findings emerge for educational level, controlling for an individual’s occupation, type of pension, and household economic condition (variables that are, however, clearly correlated to education). Indeed, compared to those having attained at most a primary education, a significantly lower mortality characterizes those with a lower secondary education, while no further advantage – on top of those mediated by having achieved a better occupation, type of pension, and household economic status –characterize those with an upper secondary or a tertiary education.

Finally, as the main finding of our estimates, it must be pointed out that the prevalent type of working activity before retirement is strongly associated with the risk of death, even when we control for dozens of variables as proxies of individual demographic and socioeconomic characteristics. Aggregating working categories, the risk of death for self-employed individuals is 26% lower than the risk for employees and, likewise, farmers’ risk of death is 17% lower than that of employees.

As a final outcome, we use Kaplan-Meier simulations of the survival curves of different groups of individuals aged over 60 (as already explained, these simulations are truncated at age 86 due to problems in mortality estimation of groups with a limited sample size; in our sample less than 100 individuals are aged over 85, see Fig. 1a and b), distinguishing individuals in various socioeconomic groups according to the values of relevant covariates.

We simulate nine hypothetical social groups combining three different categories of former occupation groups (employees, self-employed, and farmers) and three levels of ability to make ends meet (high, medium, low) and compute their life expectancy (Table 6, where we consider married males, old-age retired, living in the center of Italy, with a primary education). Therefore, for example, we call “Low Employee” a social group that combines the former occupational class “employee” and a “low” ability to make ends meet.

We then show Kaplan-Meier simulated survival curves from age 60 to age 86 by occupation, for married males, old-age retired, living in the center of Italy, with a primary education and a low level of household wealth (Fig. 3a and b).

As already evident in the estimation results (Tables 2, 3, 4 and 5), the relation between former occupational class and mortality at older age does not disappear when controlling for financial situation (and other demographic variables).

According to these simulations, the self-employed have the highest life expectancy, and their life expectancy is 1.5 years higher than that of employees (Table 6). Moreover, the life expectancy of former employees with a medium level of ability to make ends meet, i.e., the “Medium Employee,” barely reaches the life expectancy of the “Low Self-Employed” (Table 6). Also, farmers have a higher life expectancy than employees, even though the divide is narrower (1 year, see Table 6).

As expected, economic conditions are clearly related to survival (Fig. 4a and b). All other variables being equal, the gap in life expectancy between the opposite positions is almost three years for men and almost two for women.

Combining the type of prevalent previous working activity and the self-reported economic condition of households, differences increase exponentially, as shown by the simulated Kaplan-Meier survival curves in Fig. 5. Indeed, the difference in life expectancy at 60 truncated at 86, between a “low employee” and “high self-employed,” is about 3.5 years (see Table 6).

However, as already explained, these life expectancies cannot be compared with the Italian national life expectancy at 60 because of the truncation of Kaplan-Meier survival curve at 86.

In order to overcome this problem, Figs. 6, 7 and 8 show Brass estimations computed putting in relation the Italian survival curve, computed by ISTAT, and the Kaplan-Meier simulations. Table 7 compares life expectancies at 60 computed on the Brass adjusted survival curves of different simulated groups, with the Italian national life expectancy.

Wide gaps between the average life expectancy for the Italian population and specific life expectancies for different population subgroups clearly emerge from our estimates and simulations. For instance, recalling our hypothetical groups, the other variables being constant, a “low employee” has a life expectancy two years shorter than the Italian value computed by ISTAT. In contrast, a “high self-employed” has a life expectancy three years longer than the average value for Italy (see Table 8).

Finally, the gap in life expectancy at 60 between opposite groups – i.e., those previously employees and currently in the poorest group (low-level households) and those previously self-employed belonging to the wealthiest group (high-level households) – even goes beyond five years.