3.1. SHARE wave 7-job history

The analysis of the career arduousness/longevity relationship at the core of this paper rests on a (quite important and time-consuming) preliminary work that quantifies the arduousness of the entire career of SHARE respondents. That task uses the 7th wave of SHARE. This wave was assembled in 2017(18) across 26 European countries plus Israel (Table 1). It contains several retrospective modules that provide detailed data about the respondent’s history. Extensive information is provided about job history at the ISCO4 level (Appendix A.1, Figure A1).

Table 1.

SHARE: Wave 7 (2017–18) respondents aged 50+ analysed in this paper. Count by country and gender

Source: SHARE 2004–2022 (Wave 7).

In the 7th wave of SHARE, respondents are asked to retrace their complete job history by providing the starting/ending year of each of their successive jobs/occupations and whether these were done on a full- or part-time basis. A participant’s history is reported retrospectively, and thus, a long time after the work occurred (i.e., a retiree in 2018 must recall their work history from 1970 if they started working at age 20). This can lead to memory biases. To address this issue, the SHARE surveyors employed a ‘Life History Calendar’ (LHC) approach to help respondents provide accurate reports. The LHC method utilises a calendar-like matrix to map out life events, giving visual cues to both the interviewer and the interviewee regarding the onset, duration, sequencing, and co-occurrence of these events. The calendar includes rows, which categorise life events, including schools attended, jobs, living arrangements, dating relationships, and other relevant details. Numerous innovations of the LHC offer benefits over traditional data collection methods, including questionnaires. The LHC’s columns encourage recall at the temporal level, while the rows encourage recall at the thematic level. The LHC has been tested extensively with respondents of varying ages and cultural backgrounds, including those with unstable lives and cognitive difficulties (DeHart, Reference DeHart2021). The LHC reports the occupation title for each successive job or occupation at ISCO-4 digits. We merge that information with arduousness indices estimated separately for each ISCO-4 occupation (see Section 3.2 below). The combination of SHARE job history data and arduousness data enables us to compute, inter alia, a career average arduousness indexFootnote ¹⁴ and examine how it correlates with a series of usual predictors, including gender, age, and GDP per capita. Additionally, the LHC allows for calculating the duration of their entire career, both in absolute years and in equivalent full-time years (which we use as a control hereafter).

3.2. O*NET : How to quantify the arduousness of jobs and occupations

SHARE wave 7 provides a wealth of information about people’s careers. But it falls short of providing information about the arduousness of jobs/occupations. To overcome that limitation, we turn to O*NET from the United States.Footnote ¹⁵ Appendix A.1 (Figures A1; A2; A3) contains a visual presentation of how we transit from the career ISCO4 code spells to the career arduousness index version of these spells.

The O*NET dataset used in this analysis differs from a typical individual-level survey, where each row corresponds to a male or female respondent reporting their personal experience of work arduousness. Instead, O*NET provides data at the occupation level, documenting the frequency of specific job characteristics—some of which reflect physical arduousness—based on responses from incumbent workers in particular occupations. These workers are surveyed at business establishments selected through a random sampling process. For example, characteristics such as exposure to vibration or high temperatures, as reported by workers, may be more prevalent in certain occupations than others. O*NET includes over 1,000 distinct items that describe the content of more than 900 occupations,Footnote ¹⁶ providing a rich source of information about the nature of work.

As discussed by Kroll Reference Kroll2011, utilising O*NET occupational-level frequencies assumes that these figures represent the fundamental characteristics of occupations rather than those of the individuals who hold these positions and were interviewed.Footnote ¹⁷ This assumption may become problematic if factors such as gender ratios, part-time employment rates, or average job tenure differ systematically across occupations and (non-randomly) affect the frequency of reported job characteristics. Unlike Kroll Reference Kroll2011, we do not have access to individual-level data underpinning the occupational frequencies in O*NET. Consequently, we cannot employ his hierarchical regression-based method for quantifying work arduousness. Kroll’s approach involves purging the arduousness responses for variation statistically driven by individual-level characteristics and then computing conditional occupation-level arduousness measures.Footnote ¹⁸ However, we believe that the compositional bias Kroll highlights is less relevant for O*NET. In contrast to the survey employed by Kroll to evaluate ISCO-specific arduousness, the items collected by O*NET are less explicitly or implicitly aimed at capturing work demands and stress (i.e., arduousness) and their implications for well-being or health. In short, due to their more descriptive and factual nature, rather than being oriented towards consequences, the O*NET responses are less likely to be affected by the respondents’ background characteristics.

The O*NET items are organised into different modules, including job tasks, work context, skills, abilities, and more. The data is continuously updated through a rolling survey process, providing detailed and comprehensive insights into the characteristics of different occupations. For this study, we focus on the Work Context module, as its variables appear to align well with the definition of physical arduousness found in the job demands and job quality literature (Bakker and Demerouti, Reference Bakker and Demerouti2007; Chen et al., Reference Chen, Li, Xia and He2017). They explicitly describe working conditions (e.g., exposition to contaminants, spending time bending or twisting the body, working in very hot or cold temperatures…), structural job characteristics (e.g., consequence of error, time pressure, freedom to decide), and interpersonal/managerial relationship at work (e.g., contact with others, responsibility for other’s health and safety, face-to-face discussions).

We use principal component (PC) analysis to obtain a summary indicator of occupational arduousness for each ISCO-4 occupation. More information (1^st and 2^nd principal components, eigenvalues and loading factors) is reported in the Appendix A.2. Only the 1st PC is used in the paper to quantify the physical arduousness of each occupation (Figure A4). We show in Table A1 in the Appendix that it correlates relatively well with working conditions items associated with physical arduousness (e.g., Exposed to Contaminants, Pace (of work) determined by the Speed of Equipment, Sounds noise levels are distracting or uncomfortable…). We also show that the 2nd principal component correlates more with managerial versus non-managerial work content, a dimension that is a priori less relevant here. In the Appendix A.2, Figure A5 presents our O*NET 1st principal component (PC) for a list of ISCO 2 level occupations. As expected, typical manual/outdoor occupations (e.g., building and related trades) translate into high arduous PC values. In contrast, more intellectual and indoor occupations (e.g., business and administration) display significantly lower values.

It is essential to emphasise how we utilise these occupation-specific O*NET arduousness indices. Once injected into SHARE, they are used to compute, for each respondent, a career arduousness score. For most of the results presented here, it is calculated as the weighted average of all O*NET-estimated PC for his/her consecutive ISCO-4-digit occupations self-reported in SHARE wave 7 (see Appendix A.1, Figure A3). The weights reflect the proportion of total full-time equivalent years forming the career during which each occupation was performed. To achieve full-time equivalence, the years have been multiplied by .5 if the occupation was always declared part-time, 1 if always full-time and .75 when variable.Footnote ¹⁹

Hereafter, we use a respondent’s decile in the international distribution of career average arduousness scores to predict his/her risk of ill health and mortality and to generate life tables specific to each decile of career arduousness.

Before turning to the fully-fledged econometric analysis of the relationship between these deciles and bad health or death, and the computation of life tables, Table 2 presents some interesting stylised facts regarding these arduousness deciles. First, we observe that they increase with the age of the SHARE wave 7 respondents (i.e., the age at the time of the interview). This is consistent with the notion that older cohorts (e.g., those aged 70 and above) may have experienced harsher working conditions than their relatively younger peers, reflecting a general decline in work-related arduousness over time.Footnote ²⁰

Table 2.

The determinants of career arduousness decile^a age,^b gender and GDP per head quartile (ref: 1st quartile)

Source: SHARE, O*NET. Our calculations

Standard errors are in parentheses;

* p < 0.1, **p < 0.05, ***p < 0.01.

^a Coefficients capture the marginal impact on the decile of the international arduousness distribution.

^b The respondent’s age at the interview.

^c GDP per head quartile 2017: output-side real GDP at chained PPPs (in mil. 2017US$) q1: Bulgaria, Croatia, Greece, Hungary, Latvia, Lithuania, Poland, Romania, Slovakia q2: Cyprus, Czech Republic, Estonia, Malta, Portugal, Slovenia q3: Belgium, Finland, France, Israel, Italy, Spain q4: Austria, Denmark, Germany, Luxembourg, Sweden, Switzerland

^d GDP per head quartile 1992: output-side real GDP at chained PPPs (in mil. 2017US$) q1: Croatia, Estonia, Latvia, Lithuania, Poland, Romania q2: Bulgaria, Czech Republic, Greece, Hungary, Malta, Portugal, Slovakia, Slovenia, Spain q3: Belgium, Cyprus, Finland, France, Israel, Italy q4: Austria, Denmark, Germany, Luxembourg, Sweden, Switzerland

We also find that career arduousness is systematically lower, on average, among female respondents. In Table 2, coefficients ranging from −1.83 to −1.84 suggest a gender gap in career arduousness of approximately two deciles within the pooled (international) distribution. This finding underscores the importance of distinguishing between men and women when estimating the impact of career arduousness econometrically and when computing life tables.

Finally, Table 2 hints at the strong link between career arduousness deciles and the GDP per capita of the respondent’s country. Countries are grouped by quartile of the international GDP per capita distribution, measured in 2017 PPP U.S. dollars. Using the first quartile as the reference category, we observe, for example, that living in a country in the fourth quartile is associated with a 1.18-point reduction in the career arduousness decile. The second model (M2) reports results using GDP per capita data from 1992, corresponding to a 25-year lag. This alternative measure is designed to more accurately capture the economic context during respondents’ active working lives. Importantly, the results obtained using lagged GDP per capita are very similar to those based on the more recent measure, reinforcing the robustness of the observed association.

3.3. Other SHARE data used

Another important source of data for this paper is the one that allows us to quantify the risk of death at different ages. By construction, all wave 7 respondents are alive when they answer questions about their careers that we use to quantify the degree of career arduousness they have been exposed to. Thus, it is key to record deaths as they transit from one year to another beyond wave 7. It is done here using SHARE data collected in Waves 8, 9, 10 and End-of-Life Interviews. Table 3 informs about the number of transitions we exploit in the paper. Table 4 contains a first indication of how these SHARE follow-up data can be used to deliver death indicators by gender and age.

Table 3.

Number of transitions beyond wave 7

Source: SHARE 2004-2020 (Waves 7-10 and End-of-Life survey).-

*p < 0.1, **p < 0.05, ***p < 0.01

Table 4.

SHARE wave 7 respondents follow-up: risk of death/bad health by age band (OLS estimates)

Source: SHARE 2004–2020 (Wave 7 follow-up).

* p < 0.1, **p < 0.05, ***p < 0.01.

a 1: Good, Fair and Poor, 0: Excellent and Very good.

Additionally, our estimates of healthy life expectancy are based on SHARE health data. The item used here is self-rated health. It is measured at the moment of a SHARE interview (wave 7 or in the closest previous wave if missing in wave 7), which means at an age ranging from 50 to 102 and more, depending on the respondent’s age. Our poor/bad health outcome variable is built using the answer to the question “How would you rate your health?” on a 5-item scale: Excellent, Very Good, Good, Fair and Poor. This variable is widely recognised as a reliable indicator of health (Bound, Reference Bound1991; Idler and Benyamini, Reference Idler and Benyamini1997; Han and Jylha, Reference Han and Jylha2006). It is frequently reassessed with the same conclusion about its validity (see e.g., Schnittker and Bacak, Reference Schnittker and Bacak2014).Footnote ²¹ Following Etilé and Milcent Reference Etilé and Milcent(2006), we dichotomised self-reported health into Good, Fair and Poor versus Excellent and Very Good. Table 4 details the frequency by age of death and bad health, underpinning our analysis.

Finally, this paper controls for the potential bias caused by the pre-labour determinantsFootnote ²² in driving the risk of poor health in late years (Trannoy et al., Reference Trannoy, Tubeuf, Jusot and Devaux2010) or death. SHARE contains data on educational attainment and health endowment. The latter comprises the health status during childhoodFootnote ²³ and information about parents’ death status.Footnote ²⁴ Our goal, when mobilising these items, is to assess the propensity of our results to over(under)estimate the contribution of career arduousness due to negative (positive) selection into arduousness.

4.1. Framework

Our econometric analysis (step 1) utilises the follow-up data to estimate a logit model where death and poor health indicators are regressed on age, age squared, and deciles of career arduousness, including an interaction term between age and arduousness. To control for other factors, we systematically include the (mean-centred) career duration and a COVID dummy variable.Footnote ²⁵ Including career duration helps account for the direct effects of unusually short or long careers on health and mortality risk. In a separate study (Vandenberghe, Reference Vandenberghe2023a), we show that individuals with shorter careers—often due to repeated interruptions such as periods of unemployment—tend, ceteris paribus, to exhibit poorer health after age 50. We interpret this finding as evidence that career instability, independent of career arduousness, can have lasting negative consequences for individual health and longevity outcomes.

In a model variant, we also control for the role of pre-labour endowment variables (parental death status, childhood health, and educational attainment). The econometric estimates are reported in Table 5. Note that they are derived from the estimation of logit models and have been exponentiated to correspond to odds ratios. They confirm that the age of the respondent is a key (generally non-linear) determinant of the risk of death and poor health. Point estimates suggest that increased work arduousness raises the risk of death and poor health. For example, an odds ratio of 1.15 for male mortality implies that each additional decile of arduousness increases the risk of death by 15%. However, this effect diminishes with age, as shown by statistically significant odds ratios below 1 for the interaction between age and arduousness.

Table 5.

Econometric analysis of the risk of death/bad health^a

Source: SHARE, O*NET work context items, our calculations. Standard errors are in parentheses;

* p < 0.1, **p < 0.05, ***p < 0.01. ^aExponentiated coefficients, ^bCentred on the age of 65, ^cCentred on the average (men: 37.43 fte years, women: 29.95 fte years)

Step 2 consists of calculating (healthy) life tables using the above point estimates. The notations hereafter are those used in the life table literature, with subscript x referring to the age band. We compute one life table for each career arduousness decile $k= 1,\dots,10$ (Madans and Molla, Reference Madans and Molla2008). Using predicted death probability by age $\widehat{q}_{x,k}$ and that of being in bad health $\widehat{bh}_{x,k}$, we compute the survivorship by age and arduousness decile: $l_{x,k}= l_{x-1,k} \;\times \; (1-\widehat{q}_{x-1,k})$. To be precise, the survivorship estimates are normalized using the Eurostat-EU 28, 2017 estimates of survival by age 50-102; indexed on Human Mortality Database (HMD).Footnote ²⁶ We thus impose survivorship for arduousness deciles 4 and 5 $l_{x,k}; k=4,5$ to correspond to the EU reference— $\tilde{l}_{x,k}=l_{x,EU}; k=4,5$. For the other deciles, we add to the EU reference the deviation from our survivorship estimates and the average of our estimates for arduousness deciles 4 and 5. In other words, $\tilde{l}_{x,k}=l_{x,EU} \;+\; [\;l_{x,k} - \overline{l}_{x}\;]; k\neq4,5$ where $\overline{l}_{x}$ is the average for deciles 4 and 5. The justification for this is that we are fully aware that longitudinal surveys, such as SHARE, tend to underestimate mortality rates (Bergmann et al., Reference Bergmann, birkenbach and Groh2020), which could introduce bias if used directly. To address this, we rely exclusively on the gradient of mortality across arduousness deciles within SHARE to differentiate mortality patterns, while anchoring the overall level to the HMD.

The healthy survivorship by age is computed as survival (EU/HMD normalized) times the likelihood to be in good health, i.e., $\tilde{hl}_{x,k}=(1-\widehat{bh}_{x,k}) \;\times \; \tilde{l}_{x,k} $ where $ \widehat{bh}_{x,k}$ is the econometrically estimated likelihood of bad health at age x for decile arduousness k. Finally, life expectancies $e_{x,k}$ and healthy life expectancies $he_{x,k}$ are computed as the integrals (over age ranging from 50 to 102) of the corresponding (EU-HMD normalized) survivorship functions $\tilde{l}_{x,k}; \tilde{hl}_{x,k}$.

4.2. Identification issues

The resulting life tables (Tables 6 and 7) are valid only if, using econometrics, we correctly identify the relationship between the risk of poor health/death and career arduousness. We see two key issues: one related to the measurement of arduousness and one rooted in selection.

Table 6.

Estimates of (healthy) life expectancy at the age of 50 (M), by decile of average career arduousness

Source: SHARE, O*NET work context items, our calculations.

* Pre-labour endowment variables (parental death status and childhood health, plus educational attainment)

^a Ref: 1st decile of career arduousness.

^b Life expectancy.

^c Healthy life expectancy.

^d Life expectancy handicap.

^e Healthy life expectancy handicap.

^f Bad health life expectancy handicap.

Table 7.

Estimates of (healthy) life expectancy at the age of 50 (F), by decile of average career arduousness

Source: SHARE, O*NET work context items, our calculations.

* Pre-labour endowment variables (parental death status and childhood health, plus educational attainment)

^a Ref: 1st decile of career arduousness.

^b Life expectancy.

^c Healthy life expectancy.

^d Life expectancy handicap.

^e Healthy life expectancy handicap.

^f Bad health life expectancy handicap.

Regarding measurement, the main issue is the time gap, i.e., the time that has elapsed between the moment SHARE respondents worked and the moment arduousness is assessed by O*NET. Our O*NET indices (i.e., estimates of the arduousness of occupations as they were in the early 2000s) may underestimate the actual arduousness of past occupations (i.e., when SHARE respondents worked), particularly among older respondents who appear in the highest arduousness deciles (Table 2). Assuming that the passage of time has led to the disappearance of the most challenging jobs, our 2020-based O*NET arduousness index might be lower than the index that they experienced. All this might lead to an overestimation bias known in the literature as an expansion bias due to censored regressors (Rigobon and Stoker, Reference Rigobon and Stoker2007). This is an X-related bias. However, it should not be confounded with the well-known attenuation bias driven by the presence of a (large) additive measurement error/random term to X. With the expansion bias, X is systematically upward bounded (the—in real-life—high values are de facto recorded—in the data—at a lower level), it is straightforward to show that, in such a case, the (absolute value of the) slope of the $Y,X$ relationship will be exaggerated (Rigobon and Stoker, Reference Rigobon and Stoker2007).

Regarding selection, we posit that non-random selection into arduous occupations may occur both before labour market entry and during individuals’ time in the workforce. Before entering the labour market, pre-existing health conditions can influence occupational choices, with healthier individuals more likely to select into more demanding jobs. During their careers, health deterioration, also known as health shocks, may cause individuals to leave jobs that are more challenging or stressful. This is just the flip side of what is known in the literature as the healthy worker survivor effect, whereby healthier individuals are more likely to enter or remain in physically demanding jobs.Footnote ²⁷ All these imply an attenuation of the actual causal relationship (Belloni et al., Reference Belloni, Carrino and Meschi2022).

To account for selectivity before labour market entry, SHARE offers interesting opportunities. It informs us on educational attainment, as well as health status up to the age of 15 (childhood health). These are endowment items whose level is determined before people pick their first occupation and enter the labour market. Moreover, SHARE informs us about the parents’ longevity and death status, which we use in all our baseline results to control for the more heritable part of people’s initial health endowment. In our key results tables, Tables 6 and 7 show that, contrary to what would be expected, the inclusion of these pre-labour market entry controls reduces the magnitudes of (health) life expectancy gaps between the 10th and the 1st decile of career arduousness. The point, however, is that the reduction is relatively small.

Turning to selectivity during people’s careers, SHARE also offers opportunities. It enables us to assess the impact of job arduousness at various stages of these careers. Using the O*NET PC score, it is indeed possible to retrieve each respondent’s first occupation and the corresponding index. In the robustness analysis Section 5.2, we report our key results when using these as a measurement of arduousness, and it seems reasonable to assume that the intensity of the selectivity bias is more substantial when using the career average arduousness. The point, however, is that we show, for example, in Tables 12 and 13 that using the arduousness of the first job does not matter econometrically.

In short, measuring career arduousness correctly is not trivial. Our measure is not perfect due to what we call the arduousness time gap. Non-random selection into arduousness matters also. While the latter problem might lead to an attenuation bias, the former could cause overestimation. Controlling for pre-labour selection and checking for during-the-career selection, we might have limited the magnitude of the selection bias, but not entirely. Simultaneously, the exaggeration bias due to the O*NET time gap remains. So, all in all, the odds are that the results presented hereafter might not be that different from the actual causal impact of the arduousness of (healthy) life expectancy.

6.1. Summary

This paper thoroughly investigates this issue through detailed microdata analysis, with a focus on Europe. It evaluates (entire) career arduousness by scrutinising retrospective ISCO4-digit career data from SHARE wave 7 respondents aged 50 plus, supplemented by US O*NET working conditions detailed data on each ISCO occupation. Subsequently, leveraging follow-up data from SHARE, which includes information on health status and mortality among wave 7 participants, the paper estimates (healthy) life expectancy across various levels of career arduousness, employing both econometric and life table methodologies.

The outcomes of this analysis reveal a notable discrepancy in life expectancy between individuals engaged in the least and most physically demanding careers, with the 10th vs 1st arduousness decile gap ranging from 4 (men) to 4.2 years (women), respectively. Disparities in healthy life expectancy are even more pronounced, ranging from 6.9 to 9.1 years for men and women, respectively.

6.2. Policy implications and feasibility

These findings suggest that policy interventions aimed at mitigating the burdens of demanding careers should permit retirement age differences of up to 10 years, with some variation by gender.

But retirement age differentiation also comes with important limitations. As highlighted by Baurin (Reference Baurin2021); Vandenberghe (Reference Vandenberghe, Bogrnard and Gosseries2023b), and Vandenberghe Reference Vandenberghe(2024), there are principled concerns tied to the unpredictability of (realised) health deterioration or mortality after age 50 (see Appendix A.4). While our analysis reveals statistically significant differences in outcomes across arduousness deciles, this approach does not fully resolve the issue of how to treat individuals in the remaining distribution. Indeed, the SHARE data (Appendix A.4, Figure A6)) clearly indicates that any broad implementation of retirement age differentiation would remain vulnerable to two types of errors. Type-F errors occur when individuals in poor healthFootnote ²⁹ are denied early retirement (i.e., a failure of treatment), and Type-E errors occur when individuals in good health are granted early retirement unnecessarily (i.e., excessive treatment), as originally conceptualised by Cornia and Stewart Reference Cornia and Stewart(1993). These risks point to a structural limitation of retirement age differentiation: it cannot, on its own, address the full heterogeneity in late-life health outcomes. Therefore, it would need to be complemented by a robust health and disability insurance scheme. However, this raises a critical policy question: might health and disability insurance alone be more effective in addressing longevity inequalities rooted in career arduousness, as suggested by Vandenberghe Reference Vandenberghe, Bogrnard and Gosseries2023b? Also, an alternative—and potentially more straightforward—approach to improving pension equity over the life course would be to frontload pension payments (Vandenberghe, Reference Vandenberghe2024).Footnote ³⁰

Another limitation stems from the statistical and practical feasibility challenges that policymakers would encounter in attempting to replicate and implement the methodology proposed in this paper. Analysing the relationship between career arduousness and (healthy) life expectancy is highly data- and time-intensive. The first step involves quantifying the arduousness of individuals’ entire career trajectories, as captured by the SHARE dataset. It requires collecting and continuously updating detailed data on workers’ occupational histories. The SHARE data show that individuals often transition between multiple occupations throughout their lifetimes, making it challenging to track and accurately assess each career spell. Each of these spells must be evaluated for its level of arduousness, which adds further layers of complexity and requires significant analytical and resource investment.

There is also the risk of an arduousness trap. Differentiated retirement ages could inadvertently discourage workers in demanding jobs from transitioning to less arduous ones—something that one would probably want as part of a health and safety prevention strategy to promote a longer career—for fear of losing their eligibility for early retirement.

Finally, legal and political feasibility must be considered. Retirement age differentiation is vulnerable to legal challenges because it invariably implies unequal treatment. For example, differentiating retirement ages differently by gender is likely unlawful in jurisdictions such as the United States or the EU. For instance, the European Court of Justice explicitly prohibits gender-based differences in the legal retirement age. Politically, such reforms could also face significant resistance. An equalisation policy that would entail up to a 10-year difference in retirement age across groups may be seen as inequitable or politically untenable. Moreover, acceptance depends heavily on establishing a robust—and ideally causal—link between career arduousness and adverse health outcomes or mortality. While the methodology presented in this paper offers valuable insights and a solid conceptual foundation, scaling it up for policy implementation would pose some credibility (and thus acceptability) challenges.