Too Healthy to Fall Sick? Longevity Expectations and Protective Health Behaviours during the First Wave of COVID-19

Abstract

Longevity expectations (LE) are subjective assessments of future health status that can influence a number of individual health protective decisions. This is especially true during a pandemic such as COVID-19, as the risk of ill health depends more than ever on such protective decisions. This paper examines the causal effect of LE on some protective health behaviors and a number of decisions regarding forgoing health care. We use data from the Survey of Health Ageing and Retirement in Europe, and we draw on an instrumental variable strategy exploiting individual level information on parental age at death. Consistent with the too healthy to be sick hypothesis, we find that individuals, exhibiting higher expected longevity, are more likely to engage in protective behaviours, and are less likely to forgo medical treatment. We estimate that a one standard deviation increase in LE increases the probability to comply always with social distancing by 0.6%, to meet people less often by 0.4% and decreases the probability to forgo any medical treatment by 0.6%. Our estimates vary depending on supply side restrictions influencing the availability of health care, as well as individual characteristics such as gender and the presence of pre-existing health conditions.

1. Introduction

Longevity expectations (LE) are subjective assessments based on specific information such as an individual’s genetic background and previous health investments (Perozek, 2008, Hakes, Viscusi, 1997). LE predict the time of one’s death (e.g. Smith et al., 2001) and, although some early literature revealed that they are consistent with life tables (Hurd and McGarry, 2002), Arni et al. (2021) suggest they exhibit some bias, especially when they are compared with individual level end-of-life data (Costa-Font and Vilaplana, 2022).

LE can influence behaviors through a variety of pathways. Most notably, LE can influence the subjective time horizon of a utility–maximizer consumer, which ultimately affects her individual behaviours including the decision of how much to save, to insure against old age risks, or when to retire (Hamermesh, 1985). So far, the economics literature has documented that LE influence a variety of investment decisions as well as health and consumption decisions (Khan, Rutledge, Wu, 2014, Salm, 2010, Costa-Font, Vilaplana, 2022). To date, it is unclear how LE influence similar protective health behaviours, especially during a pandemic such as the first wave of COVID-19, where the risk of contagion due to the virus was more salient, and dependent on such protective behaviours.

This paper studies the effect of individual LE on protective health and other behaviours, which are critical in the context of a pandemic, insofar as they influence not only the individuals' health, but also that of others they interact with.

More specifically, we examine protective behaviours such as frequent hand washing, physical distancing or staying-at-home recommendations, all of which have been at the forefront of public interventions aiming at limiting the spread of the virus across the population. Understanding individual explanations for compliance with recommendations and engagament in health protection in a pandemic is therefore critical to design current and future policy actions (Papageorge et al., 2021). So far, the evidence suggests that the uptake of physical distancing exhibits a non-linear effect and is influenced by social trust (Petherick et al., 2021). Other explanations for such behaviours are typically linked to pandemic fatigue, and the individual specific opportunity costs of each one of such behaviours.

In this paper, we offer an alternative explanation. We examine whether health related private information influences how individuals form expectations, and uniquely perceive the costs of limited protective behaviours to themselves and others. Furthermore, we study how LE affected decisions about avoiding necessary health care. During the COVID-19 pandemic, LE might have led to rescheduling medical visits or treatments, even though it might lead to a disinvestment in health. Consistently, Anderson et al. (2021) and Park and Stimpson (2021) show that the COVID-19 pandemic may have reduced or delayed access to medical care among Medicare beneficiaries for instance. However, to date, we do not know the underpinning behavioural mechanisms. To reduce the risk of infection, LE may influence the decision to cancel or postpone scheduled medical appointments. However, it is unclear whether individuals are fully capable to evaluate the negative long-term health effects of delaying medical care. To provide some light on this questions, this paper contributes to the literature by identifying a specific mechanism driving such effect, which plays an important role in the formation of some protective behaviours, namely LE.

LE are formed on an individual’s current knowledge, which typically comes from public and, especially, private information sources, and cannot be externally observed. Consistently, Smith et al. (2001) show that LE are able to predict actual deaths and are updated when a new health shock occurs accordingly. Among the different sources of private information, relatives’ longevity, especially parents’ longevity, plays a central role, as stressed in the literature (e.g. Hamermesh, 1985, Bonsang, Costa-Font, 2020, Costa-Font, Vilaplana, 2022). Given that LE might be affected by beliefs and cognitive biases, including optimism and overconfidence biases (Arni et al., 2021), assessing empirically the effect of LE on health behaviours is far from trivial. That is, are individuals expecting to live longer more likely to engage in protective behaviours because they perceive a higher opportunity cost of early death? Or alternatively, does higher subjective longevity breed a sense of overconfidence, and provide a feeling of optimism, that in turn discourages protective behaviours? This paper follows the economics literature (e.g. Bloom, Canning, Moore, Song, 2006, Fang, Keane, Khwaja, Salm, Silverman, 2007) and exploits differences in parental age at death to provide local average treatment effects (LATE) estimates of the effect of longevity expectations on both protective behaviours and the decision to forgo medical treatments. We use data from the Survey of Health Ageing and Retirement in Europe (SHARE) both from retrospective and regular waves as well as a special wave designed to understand how older Europeans coped with the pandemic.

The effect of LE is a priori ambiguous because it may induce individuals with better health status, who can avoid being sick, to be imprudent in the presence of overconfidence and optimism biases (Arni, Dragone, Goette, Ziebarth, 2021, Costa-Font, Mossialos, Rudisill, 2009). We might expect healthier people to value their health status more, and to perceive a higher-than-average opportunity cost of engaging in limited protective behaviors. We call the latter the too healthy to be sick hypothesis. This paper scrutinises the plausibility of each hypothesis in a pandemic. In examining the determinants of protective behaviours, we find evidence supporting the hypothesis that LE proxies an individual specific expectation of her future health status, hence reflecting the potential opportunity costs of failing to follow protective practices. However, such effects are likely to differ according to other characteristics, above and beyond LE such as age. The second part of this paper is devoted to examining the robustness of our findings; we further document heterogeneous effects related to supply driven health care availability, pre-existing conditions and gender specific effects.

The structure of the paper is as follows. Next, we report the related literature. Section three describes the data. Section four presents the empirical strategy, and section five contains the results and robustness analysis. A final section concludes.

2. Related Literature

2.1. Longevity Expectations and Household Behaviour

LE or subjective assessments of expected longevity play an important role in explaining a series of individual decisions including their health investments, labour supply, insurance purchase, education, occupation and mobility (Ben-Porath, 1972, Becker, 1994, Jayachandran, Lleras-Muney, 2009, among others). Indeed, health investments are more valuable in the long run, assuming a fixed annual return. Accordingly, individuals expecting to live longer should, other things equal, be expected to invest more in health (relative to those with a shorter LE). This is referred to as the too healthy to be sick hypothesis.

Using the Health and Retirement Study (HRS) data for the United States, Bloom et al. (2006) draw on an instrumental variable framework to document that an increase in subjective survival exerts a positive effect on household wealth accumulation, but no effect on the length of the working life. Similarly, Bíró (2013) shows that perceived longevity leads to lower consumption levels, slowing down wealth decumulation.

2.2. Longevity Expectations and Health Behaviours

LE play a central role in influencing health related behaviours. Individuals, holding higher longevity expectations, are expected to invest more in healthy behaviours to enjoy a better quality of life in those extra years an individual expects to live. This is because they face a higher opportunity cost of unhealthy behaviours. Picone et al. (2004) document that women, that overestimate their survival expectations, are more likely to perform breast self-exams and request Pap smears and mammograms. However, LE might also provide a disincentive to invest in health as long as individuals face a lower marginal value of additional years of life (Fang et al., 2007). Hence, it is an empirical question of whether one effect prevails over the other.

Consistently, Bertoni et al. (2019) document that an increase in LE decreases the probability of being overweight or obese, and smoking, and increases the likelihood of daily fruit and vegetable consumption and physical activity. Hence, suggesting that the “opportunity cost effect” dominates over the “lower marginal value of life effect”. However, one can argue that it is at times when individuals have to make critical health related decisions that such differences in expectations formation make a difference.

2.3. Our contribution

This paper adds to the prior literature by estimating the effect of longevity expectations on investments in health amidst a pandemic (the first wave of the COVID-19 pandemic). We concentrate on identifying the effect among the most vulnerable group of the population, e.g., older individuals, that are exposed to a greater risk of suffering complications from the disease.

We draw on rich European data collected through the first round of a special SHARE COVID-19 survey and the regular waves. The available information allows us to investigate not only the most important protective behaviours authorities have officially promoted to limit the virus proliferation, but also relevant decisions about forgoing medical treatment, to confront the fear of contracting COVID-19 and its potential long-term detrimental effects both to individuals and health care systems. Unlike studies relying on COVID-19 specific surveys, SHARE data contain a rich set of records that can help to identify the effect of longevity expectations and dealing with some challenges such as the effect of individual specific longevity optimism (Costa-Font and Vilaplana, 2022).

Preliminary evidence from SHARE COVID-19 data shows that, during the first wave, multimorbidity was associated with protective behaviours after controlling for age, gender, education and financial distress (Delerue Matos et al., 2021). Older Europeans responded strongly to official guidelines, especially, Sand and Bristle (2021) find a positive correlation between threat perceptions and optimistic attitude with protective behaviours. Consistently, Bertoni et al. (2021) examine retirement decisions and find that those who retired earlier responded to the pandemic by limiting their mobility and adopting stricter preventive behaviors in public. One potential explanation for such an effect is the one offered in this paper, namely that some individuals may perceive a high opportunity cost in not engaging in protective behaviors during a pandemic., namely being too healthy to fall sick. To test this, we examine some of the protective behaviours analysed by Bertoni et al. (2021) and we document the role of longevity expectations in shaping them.

In addition to changing certain protective behaviours, COVID-19 has exerted a non-negligible impact on access to health care in Europe. Studies using SHARE data show an association between fragile economic conditions and unmet healthcare needs - defined as voluntary forgoing care, having pre-scheduled treatments postponed and being unable to obtain medical appointments when needed (Arnault et al., 2021). This association varies depending on the health conditions of individuals before the outbreak and differs with respect to the cross-country differences in access to healthcare before the pandemic. Smolic et al. (2021) investigate the associations between unmet healthcare needs and micro-level characteristics together with macro-level factors. We add to this literature by examining how LE might mediate some of such effects.

3. Data

3.1. The COVID-19 SHARE sample

We use SHARE data, a rich longitudinal database, collecting information on different aspects of health, well being, retirement, socio-economic status and social networks of individuals aged 50 or over in Europe (Börsch-Supan, Börsch-Supan, Börsch-Supan).

The first SHARE regular wave was collected in 2004 and included samples of eleven European countries participating in the study in addition to Israel; the last available wave refers to 2020 and covers almost all European Union countries . In waves 3 (2008/2009) and 7 (2017), the survey additionally collected individuals’ retrospective information such as early-childhood conditions and labour market history to allow empirical analysis with a longer term perspective. The COVID-19 outbreak took place during the fieldwork of the eighth wave, and SHARE suspended the data collection process in March 2020 in all countries and in June 2020 collected the first wave of a special COVID-19 survey, which contains very specific questions about life during the first wave of the COVID-19 pandemic. More specifically, a sub-sample of SHARE longitudinal respondents was interviewed via a Computer Assisted Telephone Interview (CATI)1 covering different domains, including lifestyle changes as well as health, behaviours and healthcare use.

3.2. Measurement of Longevity Expectations

The central variable of interest in our analysis is longevity expectations. In regular SHARE waves, individuals are asked about what are the chances they will live to a specific target age . The age threshold, $T$, used in the question depends on the respondent’s age: if it is lower or equal then 65, $T$ is 75; if it is in between 66 and 70, $T$ is 80; if it is between 71 and 75, $T$ is 85; if it is in between 76 and 80, $T$ is 90. The question is asked also to individuals older than 80 years of age, but we do not consider individuals aged 81 or more to avoid sample selection concerns. Respondents can elicit their longevity expectations question using a scale that varies from zero to one hundred per cent. This is a standard question used in a series of previous studies (Hurd and McGarry, 2002).

Given that LE do not vary in short periods of time, in our analysis we use the most recent subjective longevity expectation assessment in SHARE, before the COVID-19 outbreak, generally either in wave 8 or 6. As we report in Table B.1 in Appendix, f the mean LE is 62 with a standard deviation of 29; we find no significant differences between men and women. However, consistent with the previous literature, we observe evidence of rounding and heaping (see Manski and Molinari, 2010, among many others), alongside focal points (Hurd and McGarry, 2002).

3.3. Sample Selection

The data release we are using is made of 52,310 respondents and it is described in terms of fieldwork monitoring and participation details in Sand (2021).

From the SHARE COVID-19 survey sample of 52,310 individuals, we consider individuals living in Germany, Sweden, Spain, Italy, France, Denmark, Greece, Switzerland, Belgium, Israel, Czech Republic, Poland, Luxemburg, Slovenia, Estonia, Croatia, Lithuania, Bulgaria, Cyprus, Finland, Latvia, Malta, Romania and Slovakia.2 Therefore, we do not include in our analysis the Netherlands, Hungary and Portugal for the following reasons. The Netherlands did not participate in the regular SHARE waves 6 and 7, but implemented a mixed mode survey that was different from that of all the other countries. Hungary is not available in the SHARE sample in wave 6, and for this reason, we are only able to define parental birth cohort for very few observations. Finally, we do not consider also Portugal due to the small sample size including records reporting parental education. Once we drop the three mentioned countries , the sample drops to 48,656 observations. We further select individuals aged between 50 and 80 years. The lower age bound is given and refers to the SHARE eligibility criteria3 , whereas the upper bound is set to avoid considering in our analysis a selected sample of very healthy individuals. Age constraints lead to a final sample of 48,403 respondents.

Our empirical strategy exploits the access to information about parental age at death and birth cohort of parents (the latter computed using the available information on the current age if the parent is alive in previous waves, or the year of death asked from wave 7 on-wards). We are not able to obtain this information for 6,693 individuals; therefore, we are left with 41,710 units.

To satisfy the exclusion restrictions of our empirical strategy (described in Section 4), we select individuals having dead parents, which leads us to a sample of 31,599 individuals. As shown in the Appendix (Table A.1), the excluded individuals are younger than those in our sample at this stage (on average, about 5 years younger). The longevity expectation question, by construction, depends on respondents’ age , which partly explains the significant difference in longevity expectations between excluded and included individuals: excluded - younger - individuals report higher longevity expectations. Looking at differences in protective behaviours between the two groups during the first wave of the pandemic, we can notice in Table A.1 panel (a) that, among the excluded, there is a lower fraction of individuals reporting Shopping less often, Walking less often, Meeting People less often, Keeping Always Distance. However, we can also observe that there is no significant difference between the two groups for Hands Washing & Sanitizer Usage and Forgone Treatment SP. Finally, among the excluded, there is also a lower fraction of individuals forgoing visits (to the general practitioner). We do not observe a clear behavioural pattern for the excluded pointing to a clear general direction in terms of potential bias. For this group, we need to consider also that some behavioural responses might be driven through having alive parents.

We further select respondents with valid answers to all the relevant variables included in our model.4 We provide additional evidence in the Appendix about differences in age, longevity expectations and protective behaviours between individuals included in our final sample and those excluded due to missing information (we label them Excluded due to item-non-response). Table A.1 panel (b) shows that the differences in terms of age and longevity expectations, while statistically significant, are smaller compared to panel (a); Excluded due to item-non-response individuals are slightly older, reporting lower longevity expectations. There are no significant differences between the two groups when looking at Shopping less often and all Forgone Treatment variables. The fraction of individuals reporting Walking less often is larger among the excluded but lower for Meeting People less often, Keeping Always Distance and Hands Washing & Sanitizer Usage. Also in this case, we do not observe a clear behavioural pattern for the Excluded due to item-non-response.

In Appendix A we provide also Table A.3 reporting weighted estimates based on inverse probability weighting which are in line with Table 2 baseline results.

Table 2.

The effect of expected longevity on health behaviours and forgone medical treatment.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Dep. Var.	Shopping	Walking	Meeting People	Always	Hands Washing &	Forgone	Forgone	Forgone
	Less Often	Less Often	Less Often	Distance	Sanitizer Usage	Treatment	Treatment GP	Treatment SP
OLS estimates
LongExp	0.001	0.001	0.004*	0.006**	0.003*	-0.006**	-0.002	-0.005**
	(0.003)	(0.003)	(0.002)	(0.003)	(0.002)	(0.003)	(0.002)	(0.002)
Obs	21,548	21,394	21,218	21,671	21,567	21,615	19,878	20,626
R${}^2$	0.131	0.278	0.076	0.056	0.027	0.044	0.032	0.040
TSLS estimates
LongExp	0.071*	-0.075*	-0.021	-0.008	0.062***	-0.001	-0.001	-0.031
	(0.041)	(0.043)	(0.028)	(0.037)	(0.022)	(0.033)	(0.023)	(0.028)
Endogeneity test (p-value)	0.088	0.078	0.379	0.709	0.007	0.850	0.949	0.351
F-stat (First-stage)	125.452	126.817	120.605	126.322	125.456	123.590	121.679	122.639
First-stage estimates
NpassedTage	0.080***	0.081***	0.079***	0.080***	0.080***	0.079***	0.082***	0.081***
	(0.007)	(0.007)	(0.007)	(0.007)	(0.007)	(0.007)	(0.007)	(0.007)
Reduced-form estimates
NpassedTage	0.006*	-0.006*	-0.002	-0.001	0.005***	-0.001	-0.001	-0.003
	(0.003)	(0.003)	(0.002)	(0.003)	(0.002)	(0.003)	(0.002)	(0.002)
R${}^2$	0.131	0.278	0.075	0.056	0.027	0.044	0.032	0.040

Notes: Standard errors in round brackets. Significance levels: p-value *** $\le$ 0.01, ** $\le$ 0.05, * $\le$ 0.1.

3.4. Protective Behaviours

In this study, we are especially interested in individuals self reported protective health behaviours and forgone medical treatment during the pandemic, which proxy investments in health. We focus on questions eliciting the extent to which respondents engaged in various types of protective behaviours, including whether they ever left home since the outbreak and how often they did specific activities such as going out for a walk as compared to before the outbreak. In addition, we include other protective behaviour that differ in how cognitively and economically costly they are, such as social distancing and washing hands or sanitizer/disinfection fluids usage too. The survey collects also information about whether respondents voluntarily stopped or delayed any planned medical treatment, and even identifies some detail about the specific type of treatment (general practitioner visit, specialist visit or other types of treatments) forgone.

Our final sample includes 22,602 individuals participating in the COVID-19 survey that report their LE in one of the previous recent waves, alongside valid information in all the relevant variables included in the model, especially parental age at death and birth cohort.5

Our outcome variables are defined as binary indicators. Individuals are first asked if they ever left home since the outbreak, and, among those having left home, the survey includes a follow-up question about the type of activity they went out for. We consider shopping, walking and meeting more than five people outside the household.6 Possible responses to the follow-up question include not anymore, less often, about the same or more often. Following the literature (Bertoni et al., 2021), we combine the first and the follow-up question, so that those who never left home are recorded as not anymore (and do not consider question 1). For instance, Going shopping less often takes the value of one if the individual reports never having left home since the outbreak, or having left home in the first question but not going out shopping any more or less often.

As depicted in Table 1, we find that since the outbreak, 74% of respondents have gone shopping less frequently, 53% went out walking less often and 91% met more than five people outside the household less often. As a result, we are interested in determining whether LE influenced such protective health-related behaviors during the pandemic by mediating the behavioral sequence.

Table 1.

Summary statistics - Health Behaviours and Forgone Medical Treatment (%).

	ALL	N	Men	Women	t-tests (p-value)
	(1)	(2)	(3)	(4)	(5)
Going Shopping Less Often	74.34	22,405	67.23	79.64	0.00
Walking Less Often	52.58	22,242	48.41	55.70	0.00
Meeting People Less Often	90.88	22,061	89.24	92.12	0.00
Always Distance	82.36	22,531	79.34	84.62	0.00
Hands Washing & Sanitizer Usage	95.05	22,424	94.19	95.70	0.00
Forgone Treatment	12.86	22,474	10.21	14.86	0.00
Forgone Treatment: GP	5.53	20,664	4.36	6.44	0.00
Forgone Treatment: SP	8.83	21,450	6.79	10.39	0.00
Forgone Treatment: PH	1.59	18,701	1.34	1.61	0.11
Forgone Treatment: OT	1.48	19,287	0.80	1.94	0.00

Those individuals who report having left home since the outbreak responded to two additional questions, one referring to social distancing and, another one referring to masks wearing. We exploit the responses to the first question , again defining a binary indicator taking value one if the individual never went out or kept always a social distance from other people when he or she went outside the home (Always distance). We do not consider mask wearing alone because when it was a voluntary decision it can be mediated by individual altruistic behaviour, as it captures how much respondents care about other individuals’ health. However, in almost all countries, masks wearing was compulsory and hence not a choice.7

We additionally define a dummy variable that takes the value of one if the respondent reports having washed their hands or used a special hand sanitizer or disinfection fluids more frequently than usual (Hands Washing & Sanitizer Usage). 82% of respondents did keep always social distance; and 95% reports having washed their hands or used a special hand sanitizer more frequently than usual.

Next, we measure forgone medical treatments, as a binary indicator that takes the value of one if individuals answer positively to the following question: since the outbreak of Corona, did you forgo medical treatment because you were afraid to become infected by the corona virus? This indicator identifies individuals who voluntarily choose to give up medical treatments. 13% of our sample decided to avoid medical treatments in the first wave of the COVID-19 pandemic. Table 1 reports a breakdown by type of forgone medical treatment. 6% of the sample avoid visits to the general practitioner, 9% avoid specialist visits, 1.5% avoid physiotherapy, psychotherapy, or rehabilitation, and 1% avoid other medical treatments. We consider the general forgone treatment question and forgone medical treatments regarding visits to the general practitioner or specialist in the empirical analysis. However, we do not examine other types of treatments due to low case numbers for those outcomes.

Table 1 reports descriptive evidence of differences in protective health behaviours by gender, which appear to be always statistically significant as shown in column 4. Among women, there is a higher percentage of individuals going shopping less often, walking less often, meeting people less often, keeping always social distance and washing hands or using a disinfection fluid more frequently. Regarding measures of forgone treatment variables, we find that men have a lower percentage of people forgoing all types of medical treatment. We will come back to such gender differences when we discuss the heterogeneity of our findings later in the paper. Country differences are reported graphically in the Appendix (Figure B.1) for all the outcomes of interest.

4. Empirical strategy

Our identification strategy lies in estimating the effect of LE on protective health behaviours during COVID-19 alongside forgone medical treatment, controlling for current health and other covariates that could exert a mediating influence. More specifically, we estimate the following linear probability model (LPM):

$$Y_i=\alpha +\beta LongEx{p}_i+{\mathbf{X}}_i^{\prime }\gamma +e_i.$$ (1)

where $Y_i$ refers to the relevant outcome - behaviours or forgone medical treatment - for individual $i$ interviewed within the SHARE COVID-19 survey. $LongEx{p}_i$ refers to the standardized longevity expectation elicited using a subjective survival probability question available in one of the previous SHARE regular waves and therefore pre-determined9 ; $\beta$ is the parameter of interest.

Although, ordinary-Least-Squares (OLS) estimates of equation  (1) might be biased in our framework due to measurement error in the longevity expectations and/or omitted variables (e.g. Bloom, Canning, Moore, Song, 2006, Fang, Keane, Khwaja, Salm, Silverman, 2007), the advantage of the LPM specification is that we can easily account for the potential endogeneity of longevity expectations but, for completeness, in the Appendix, we report equivalent probit marginal effect estimates when treating or not expected longevity as endogenous.

To account for endogeneity concerns, we use one instrument that is both relevant, i.e. correlated with $LongEx{p}_i$, and exogenous, i.e. influences health investments only through their effect on $LongEx{p}_i$. Consistently with the literature we exploit differences in parental age at death (Bloom, Canning, Moore, Song, 2006, Fang, Keane, Khwaja, Salm, Silverman, 2007). The validity of our strategy relies on the assumption that parental age at death captures individual specific private information of an individual’s genetic/hereditary health endowment affecting health investments only through LE. In other words, the key assumption in this framework, to identify the parameter of interest, is that health investments are conditionally mean independent of the genetic health endowment, given the controls included in the model (e.g. parental birth cohort).

Based on the parental age at death information, we exploit individual level data on the variation in the number of parents whose age at death is larger than $T$ - the threshold in the survival probability question. Yet, as a robustness check, available in the Appendix, we exploit different definitions of our instrument, taking into account for instance the potential differential maternal and paternal effect.10

We focus on individuals with deceased parents11 to exclude a direct effect of parental age on the outcomes: namely, individuals might adjust their health behaviours to protect their parents if they are still alive.12 We further drop individuals whose parents died recently - in 2018 or later -, given that for such individuals there could be a direct strong effect on the adoption of specific health behaviours due to parental bereavement.

In the next section we will show that our instrument is relevant based on the F-statistics for the excluded instruments that are well above the critical values for weak identification (Staiger, Stock, 1997, Lee, McCrary, Moreira, Porter, 2020). We further provide endogeneity test results based on Baum, Schaffer, Stillman, 2003, Baum, Schaffer, Stillman, 2007 to understand whether expected longevity can be treated as exogenous. Finally, we explore the plausible exogeneity of our instrument, and discuss the inference about our parameter of interest, $\beta$, when relaxing the exclusion assumptions along the lines proposed by Conley et al. (2012).

Our baseline two-stage least squares (TSLS) specification includes one endogenous variable and one instrument, the model is therefore just-identified. In the robustness checks, we propose over-identified models which allow us to run over-identification tests that we report and comment on in the results section. In addition to standard socio-demographic variables,13 we include among controls self-perceived health - current and referring to the wave in which individuals assess their survival probability -, a dummy covid taking value one if the respondent or anyone close to him/her experienced symptoms attributable to the COVID illness, a binary indicator for those who were diagnosed with a major illness or health condition since the last interview and a dummy ever hospitalized that identifies individuals being hospitalised before their survival probability assessment.

To account for potential cognitive biases, we include among covariates two indicators measuring individual COVID-19 related optimism attitudes as suggested by Sand and Bristle (2021). The two indicators refer to binary variables taking the value of one whether a respondent named any uplifting experience since the outbreak of COVID-19 (Optimistic attitude index 1) or something to look forward to, once COVID-19 abates (Optimistic attitude index 2), respectively. Regarding optimism, a thread to our analysis might be that we are observing a selected sample if the most pessimistic individuals with lower subjective longevity expectations (and health capital) happen to be those who have died, and did not complete the SHARE COVID interview. However, in such a case, we are likely to estimate a lower bound of the effect of longevity expectations on behaviours.

Another potential bias of our estimates might result from the effect of early life health. Hence, in our specification we control for poor childhood health, health related risk attitudes through a dummy for having ever smoked, parents’ birth cohort, parents’ education and an indicator for having received any vaccinations during childhood to capture parental prevention behaviour. Table B.1 reports all the related summary statistics.

5. Results

5.1. Baseline estimates

Table 2 reports our baseline results. In our framework, OLS estimates identify a partial association between subjective expected longevity and the outcome of interest, conditional on a number of controls described in the previous section, especially health indicators, parental education/prevention behaviour and birth cohort.

The estimated LE coefficients are statistically significant and revel a positive effect. That is, we find a positive association between longevity expectations and Meeting People Less Often, Always Distancing, Hand Washing & Sanitizer Usage, suggesting that the longer an individual life span, the more likely individuals are to invest in health protective behaviours during a pandemic such as COVID-19. Consistently, we find a negative association between longevity expectations and Forgone Medical Treatment in Column (6) in line with the opportunity costs explanation, namely individuals expecting to live longer are less likely to forgo medical treatment, and this holds specifically for visits to the specialist (Column 8). In contrast, we find no significant associations between LE and the Shopping Less Often, the Walking Less Often and the Forgone Treatment GP outcomes. The latter are basic daily decisions which can be done safely, either online or in safe environments.

Nonetheless, OLS estimates are potentially biased. This is either due to measurement error in the measurement of longevity expectations or, due to omitted variable bias. Therefore, the identification of the LE effect requires taking advantage of an exogenous source of variation. In this paper we exploit information on parental age at death and report TSLS estimates to retrieve the effect of LE on the protective behaviours of interest. Table 2 includes estimates of selected first-stage coefficients showing the effect of the instrument (NpassedTage) on the endogenous variable. NpassedTage is highly significant (at the 1% level) with the expected positive sign.14 The F-statistics on the excluded instruments reported in the table are well above the cut-off threshold, and the critical values for weak identification testing discussed in Lee et al. (2020). Hence, LE estimates are likely to be reasonably well identified.

Table 2 reports also endogeneity test results. The p-value is larger than the conventional 10% level for all outcomes with the exception of columns (1), (2) and (5), meaning that, in those cases, we reject the null hypothesis that longevity expectations may be treated as exogenous. We therefore rely on OLS estimates to assess the role of LE on the protective behaviours reported in columns (3), (4), and (6) to (8). More precisely, based on OLS estimates, we find that a one standard deviation increase in LE decreases the probability to forgo any medical treatment by 0.6% (column 6) and by 0.5% for specialist visits (column 8). Relative to the mean, the estimated effect for the forgo medical treatment outcome is 5% and 6% when focusing on specialist visits.

Consistently, a one standard deviation increase in LE increases the probability to comply always with social distancing by 0.6% and meeting people less often by 0.4%.

Next we examine the effect of LE on Shopping Less Often, Hands Washing & Sanitizer Usage and Walking Less Often, where the endogeneity tests suggest relying on TSLS estimates. Based on TSLS results, we estimate that a one standard deviation increase in LE increases the probability to shop less often by 7.1% and washing hands or use a sanitizer/disinfection fluids more often by 6.2%. According to column (2) a one standard deviation increase in LE decreases also the probability of walking less often by 7.5%.

The full set of estimates, including all covariates, is reported in Appendix C.

Most of the outcomes considered are undoubtedly important protective behaviours against the proliferation of COVID-19 infections (e.g. hand washing) and the related results are in line with the idea of investments in health, except for the estimated effect on Walking Less Often. Going out for a walk less often can, in turn, lower the risk of becoming infected (and have severe consequences on health) but, exercising less often can have negative effects on health too, especially in the long-term. Considering that 87% of individuals report to be walking less often declared to be always social distancing, the activity going outside for a walk (following social distancing guidelines) could be interpreted as a way of investing in (physical and mental) health during the pandemic.

5.2. Robustness Checks

In this subsection we report the robustness analysis of the results presented in the previous section. We first examine the effect of including recent parental deaths in our sample, results are reported in Table B.2 and confirm Table 2 estimates. The only exception is column (1) where, we do not find a significant effect of longevity expectations on the probability to go out shopping less often.

Tables B.3 and B.4 display how estimates change when we rely on different definitions of our instrument. In Table B.3 we use as instrumental variables two binary indicators: one dummy variable that takes the value one if both parents’ age at death is larger than $T$, and a second binary variable equal to one if only one of the two parents died after $T$ years of age. This provides us with one endogenous variable and two instruments and allows us to run over-identification tests which suggest that in all cases we do not reject the null of the J-test. In Table B.4 we use as instruments two dummies, one for the mother and one for the father having passed the age threshold $T$ respectively, to investigate potential heterogeneous effects related to the mother’s as compared to the father’s death based on J-test results as suggested in Angrist and Pischke (2009). Table B.4 does not reveal any systematic heterogeneous effects either (we fail to accept the null only in column 2).

For completeness, in Table B.5 we report probit estimates taking into account the binary nature of our outcomes. The estimated marginal effects (also when accounting for endogeneity issues) are in line with our LPM baseline results.15

We further check the robustness of our estimates by dropping observations from one country at a time, to verify that no single country is driving our results. Estimates available upon request are in line with what reported in Table 2.

5.3. Plausible Exogeneity

Next, we provide an analysis of the plausible exogeneity for the outcomes where test results point to endogeneity problems: Shopping Less Often, Walking less often and Hands Washing & Sanitizer Usage. In those cases, our TSLS estimates rely on the assumption that genetic factors do not affect directly health investments, once we control for all the covariates included in the model. To understand how relaxing exclusion restrictions can affect our estimates, we follow Conley et al. (2012). Conley et al. (2012) propose a procedure to show how the parameter of interest changes when relaxing strict exogeneity, allowing the instrument to have a direct - near zero - effect on the outcome.

Figures B.2, B.3 and B.4 report the 90% confidence intervals of $\beta$ in Eq. (1), according to the union of Symmetric CI and the local-to-zero methods. The two methods differ with respect to the prior information about the parameter capturing the direct effect of the instrument on the outcome - $\gamma$, according to Conley et al. (2012)’s notation. The union of Symmetric CI method, through $\delta$, allows to change the support of $\gamma$ which is [-2$\delta$, 2$\delta$]. According to the local-to-zero method, we assume $\gamma$ $\sim \mathcal{N}(0,{\delta }^2)$ instead.

Figures B.2, B.3 and B.4 show that the effect of longevity expectations on the probability of Shopping Less Often, Walking less often and Hands Washing & Sanitizer Usage more frequently than usual becomes insignificant when $\delta$ is about 0.002, i.e. the magnitude of the standard error estimated for the parameter of NpassedTage in the reduced-form specification.

5.4. Heterogeneity Analysis

Next, in this subsection we describe the heterogeneity analysis we have conducted. We first exploit the two questions about involuntary postponements of visits/treatments and unmet needs (i.e. impossibility to have an appointment for medical treatment since the outbreak) to stratify our sample according to living in a country with a health care system which might be under pressure due to the pandemic.

We first rank countries according to the percentage of individuals reporting involuntary postponements or unmet needs and consider, as one group, countries in the highest tertile (Sweden, Greece, Germany, Bulgaria, Cyprus, Latvia, Romania and Slovakia), against all the other ones.

Figure 1 shows how estimates differ between the two groups of countries when looking at the Forgone treatment outcomes. Significant results are estimated only for countries belonging to the second and third tertile. This result is consistent with idea that, in the former group, where health care systems have to postpone treatments or cannot accept treatment requests, individuals do not have an actual choice in terms of forgoing visits.

Fig. 1.

Heterogeneous effect of expected longevity on forgone medical treatment by group of countries. Notes: 95% confidence intervals displayed.

Figure 2 reports stratified estimates distinguishing between individuals that reported Fair/Poor health (labelled Poor Health in the Figure) versus Good/Very good/Excellent (labelled Good Health in the Figure), when the survival probability assessment is done; poor health is therefore predetermined. For each outcome, we report, OLS or TSLS estimates depending on endogeneity test results. Consistently with our main hypothesis, we find significant positive effects of LE on health behaviours especially for individuals in good health. It is worth noting that, the effect of LE on the probability to go out walking less often changes sign depending on individuals self-reported health: it is negative for individuals reporting being in poor health and positive for those in good health.

Fig. 2.

Heterogeneous effect of expected longevity on health behaviours and forgone medical treatment by self-reported health. Notes: 95% confidence intervals displayed.

The previous literature has documented heterogeneous behavioural responses to the pandemic by gender (e.g. Galasso et al., 2020). Here we investigate whether this is the case also when looking at the effect of longevity expectations on behaviours. Figure 3 reveals that our baseline results differ for men and women. They suggest gender heterogeneity, and more specifically show that LE exhibit a negative and significant effect on the probability to forgo medical treatments among women, but no significant effects among men. Furthermore, we do not observe any systematic differences in other behavioural responses. First-stage estimates and statistics suggest that women tend to react more in terms of expectations to parental deaths as compared to men. Point estimates of the $NpassedTage$ coefficient range from 0.097 to 0.102 (significant at the 1% level) for women, from 0.055 to 0.058 (significant at the 1% level) for men. The F-statistics on the excluded instruments range from 107 to 111 for women and from 25 to 27 for men.

Fig. 3.

Heterogeneous effect of expected longevity on health behaviours and forgone medical treatment by gender. Notes: 95% confidence intervals displayed.

6. Conclusions

We study how individuals’ subjective longevity influences protective behaviours during the first wave of COVID-19. More specifically, we empirically test the too healthy to be sick hypothesis. That is, we document that longevity expectations, namely subjective assessments of future longevity, are a proxy of individuals’ future health capital, which in turn influences decisions about both protective health behaviours and forgoing health care .

Exploiting evidence from SHARE, this paper estimates the causal effect of LE on protective behaviours. We exploit an instrumental variable strategy where individual differences in parental age at death provide exogenous variation in LE. More specifically, LE contain private information that explains heterogeneous behavioural reactions to pandemic restrictions across the population as well as the decision to forgo medical treatments during the COVID-19 pandemic. We have further tested whether they contain private information, or they are instead reflective of some cognitive biases such as health related optimism (Costa-Font and Vilaplana, 2022).

Our results are consistent with the idea that individuals holding higher LE continue to invest more in their health, in line with the too healthy to be sick hypothesis. We find robust evidence suggesting that a rise in LE increases the probability of individuals’ engagement in several protective health behaviours during the first wave of COVID-19 epidemic, and we document that it reduces the probability of forgoing medical treatment. We estimate that a one standard deviation increase in expected longevity increases the probability to comply always with social distancing by 0.6%, meeting people less often by 0.4% and decreases the probability to forgo any medical treatment by 0.6%. These estimates help explain the different behavioural reactions to a common health threat such as COVID-19 and suggest that incentives to increase compliance with restrictions should target specific individuals who do not expect such restrictions to influence their health, who do not perceive the health effects of COVID-19 as weakening their health status, and hence face a lower perceived opportunity cost. Interventions, aimed at increasing the salience of the health-related opportunity costs of not complying with pandemic restrictions, might include the use of reminders of an individual’s healthiness so they understand what is at stake.

Declaration of Competing Interest

We wish to confirm that there are no known conflicts of interest associated with this publication and there has been no significant financial support for this work that could have influenced its outcome.

We confirm that the manuscript has been read and approved by all named authors and that there are no other persons who satisfied the criteria for authorship but are not listed. We further confirm that the order of authors listed in the manuscript has been approved by all of us.

We confirm that we have given due consideration to the protection of intellectual property associated with this work and that there are no impediments to publication, including the timing of publication, with respect to intellectual property. In so doing we confirm that we have followed the regulations of our institutions concerning intellectual property.

We understand that the Corresponding Author is the sole contact for the Editorial process (including Editorial Manager and direct communications with the office). He/she is responsible for communicating with the other authors about progress, submissions of revisions and final approval of proofs. We confirm that we have provided a current, correct email address which is accessible by the Corresponding Author.