'More than One Red Herring'? Heterogeneous Effects of Ageing on Healthcare Utilisation

We study the effect of ageing, defined an extra year of life, on health care utilisation. We disentangle the direct effect of ageing, from other alternative explanations such as the presence of comorbidities and endogenous time to death (TTD) that are argued to absorb the effect of ageing (so-called 'red herring' hypothesis). We exploit individual level end of life data from several European countries that record the use of medicine, outpatient and inpatient care as well as long-term care.  Consistently with a 'red herring hypothesis', we find that corrected TTD estimates are significantly different from uncorrected ones, and its effect size exceeds that of an extra year of life, which in turn is moderated by individual comorbidities. Corrected estimates suggest an overall attenuated effect of ageing, which does not influence outpatient care utilisation. These results suggest the presence of 'more than one red herring' depending on the type of health care examined.


Introduction
Population ageing is commonly portrayed as a central determinant of health care spending (WHO, 2015;Marino et al, 2017) 1 . Given that the percentage of old age population in the countries of the Organisation of Economic Cooperation and Development (OECD) is projected to rise to 25% by 2050 (Lafortune et al, 2007), it is important to understand how ageing affects health care use. However, there are good reasons to argue that the effect of ageing on health expenditure is overestimated. One of the main explanations is that a significant share of expenditures takes place around the time of death. Some studies even go as far as to argue that the effect of ageing on health care reflects a 'red herring' given that when time to death (TTD) is accounted for, the effect of ageing disappears (Zweifel et al, 1999;Zweifel et al, 2004;Hall and Jones 2007;Shang and Goldman 2007) 2 .
In addition to the consideration of TTD, which is potentially endogenous, another source of overestimation (of ageing effects on health expenditure) results from the correlation between morbidity and individual's age 3 , as it is subject to omitted variable bias. The effect of such omitted variable bias can be analysed using individual longitudinal data, that captures the influence of early lifestyles. This paper addresses some of these econometric concerns by drawing on individual data that can explain both individual and country-level variation in morbidity and TTD.
Finally, another potential red herring results from the fact that ageing can change the composition of health care towards a more intense use of end of life care, hospital 4 care and long-term care 4 . Hence, the effect of ageing is likely to be heterogeneous across different types of health care, which especially differ in their intensity in the use of technology (Breyer et al, 2010). Finally, ageing can incentivise the utilisation of new technologies that specifically cater to the health care needs of an ageing population 5 .
Hence it is important to understand how ageing impacts on different types of health care that differ in their intensity of technology (e.g., medicines, hospital care, home care etc). This paper to examine the effect of ageing on different types of health care use, to disentangle the effect of additional confounding effects on health expenditure, namely (i) proximity to death, (ii) co-morbidities and lifestyles and (iii) differences in the composition of health care. Previous research so far has been country specific, and mainly relies on cross sectional insurance data records, often limited to hospital care. We exploit longitudinal end of life data that covers a long list of European countries for the period 2004-2017 Survey for Health, Ageing and Retirement in Europe (SHARE). The advantage of using a multi-country panel is that it allows for the inclusion of both individual and country fixed effects that net out specific institutional reasons for differences in the effect of ageing on health care (HC) expenditures. The survey contains an end of life module that identifies the cause of death of the individual and helps to distinguish between survivors and deceased, namely those that have died between two consecutive waves. We report both parametric and nonparametric specifications and address the problem of endogeneity of TTD by correcting the estimations with rich instruments for parental survival in the dataset.
Our findings suggest that corrected TTD estimates are significantly different from uncorrected ones and affect both the extensive and intensive margin of hospital 5 admissions and length of stay, as well as home and nursing home care use, consistently with a 'red herring hypothesis'. Second, the effect size of TTD exceeds that of ageing, which in turn is attenuated by the presence of comorbidities. Corrected estimates indicate that the effect size of ageing is far more attenuated when it is statically significant.
Finally, ageing does not explain (both the internal and external margins of) outpatient visits with doctors and nurses once TTD and comorbidities are controlled for.
The structure of the paper is as follows. Section two reviews the most relevant literature. Next, we describe the data and empirical exercise. Section five and six contains the results and a final section concludes.

Related Literature
Red herring hypothesis. The effect of ageing on health spending has been brought to question based on the fact that age is correlated with mortality. A seminal study used a sample of deceased patients from a Swiss sickness fund and found that the effect of age on healthcare expenditure disappears once it is controlled by the effect of time to death (TTD) (Zweifel et al., 1999). This opened a long list of contributions to the question of ageing and health spending, and this paper aims to add value to the same endeavour.
Econometric specifications. Almost all estimates of the effect of ageing on health expenditure have received a significant deal of criticism due to a series of econometric issues, mainly omitted variable bias, and the potential reverse causality of the TTD (Salas and Raftery, 2001;Seshamani and Gray, 2004). The logic is that if health care investments (e.g., such as new drugs) improve patient's health status, they could extend life. Therefore, estimates that fail to account the dynamic influence of current and previous health expenditures on life expectancy would overestimate the effect of ageing.
In a later study, Zweifel et al. (2004) confirmed his previous results after restricting the Electronic copy available at: https://ssrn.com/abstract=3603835 6 sample to a single year to ensure that HC expenditures only affect the probability of survival in cases in which death was very close, introduced the TTD as a single explanatory variable and considered both survival and deceased individuals in the sample. The results confirmed that age was not a significant variable in explaining the HC expenditures of the deceased and, in the case of survivors, the effect of age is much lower when the TTD variable is introduced. For their part, Seshami and Gray (2004) concluded that the omission of TTD from the analysis was found to overestimate the effect of ageing, and the number of trimesters before death is a significant explanatory variable, and its impact on cost is higher at the end of life.
Other more recent estimates suggest that TTD accounts for 16.7% and 24.5% of lifetime HC and LTC expenditures (French et al. 2017). Similarly, Breyer et al. (2017), estimates that HC expenditures in the last 4 years of life account for 30% of total expenditures over a lifetime, even though part of such effects result from the effects of life expectancy (Breyer et al, 2012). Hence, it seems TTD is not the only herd herring underpinning the effect of ageing on health care expenditures.
Endogeneity. TTD is likely to be affected by both reverse causality and omitted variable bias. Stearns and Norton (2004) use data from the Medicare Current Beneficiary Survey (1992)(1993)(1994)(1995)(1996)(1997)(1998) to document evidence of omitted variables, which is accounted for by adding individual specific fixed effects, which correct the effect of unobserved timeinvariant characteristics. However, such strategy does not deal with reverse causality. An alternative strategy lies in employing instrumental variables, namely a variable influencing health expenditure only via TTD, but not the age at which the individual is interviewed (Steinmann et al., 2007). OLS estimates would be biased if health care 7 expenditures (HCE) and medical innovations prolong life (Lichtenberg et al, 2012) 6 . Felder et al. (2010) address the problem of endogeneity using an instrumental variable strategy that employs lags as instruments. They document that TTD and its square retain their explanatory power in explaining HCE in its intensive and extensive margin.
However, as they recognize that are not able to fully purge TTD of its endogeneity. When errors are AR(1) distributed, the parameter is not estimated consistently from a lagged instrument 7 .
Heterogeneity. The effect of ageing might be heterogeneous to different types of health care use that differ in the intensity of use of technology. Werblow et al. (2007) eluded the problem of endogeneity and focused on relating the individual HCE in a given year with the remaining TTD. They find evidence of heterogeneous effects as the majority of the HC expenditure components (drugs, hospital outpatient and hospital inpatient) are found not be influenced by age, but by TTD. The most significant exception is the acute care provided to patients who also receive long-term care (LTC) regardless of their survival.
They explain these results by the fact that patients with limited survival prospects attract a large share of medical technology. Finally, Kelley et al. (2013) estimate that the increase in out-of-pocket expenditure in the last years of life shows a wide variability, which is explained by the increasing share of out-of-pocket expenditure that results from dementia or Alzheimer's diseases which is more than double that of gastrointestinal diseases or cancer. 9 equivalent changes in the structure of the population (+3.4%) 8 . Importantly Ishizaki et al. (2016) document a negative effect of age on the probability of hospitalization and that no significant effect of age on length of stay at hospital exists in the three months before death. Consistently, Howdon and Rice (2018) found that when morbidity is controlled for, it absolves two-thirds of the effect of TTD on HCE, which confirms the underestimation of the TTD effect when the potential endogeneity of this variable is not taken into account 9 .
Ageing and long-term care substitution. Finally, a set of studies examine the relationship between age and long-term care controlling for TTD. De Meijer et al. (2011) analysed the use of institutional LTC and home care from a Dutch dataset of individuals 55 years and older. They observed that once the effect of age was controlled by disability and morbidity, it remained significant, but TTD was no longer significant. Similarly, Larsson (2008) documents that whilst age is a significant variable in predicting the probability of receiving formal home care, TTD explained the probability of hospitalization, and both predict the use of nursing home care 10 . A final set of studies inclide Karlsson and Klohn (2011) addressing the problem of the endogeneity of TTD using instrumental variables, and Karlsson and Klohn (2014) which show that TTD is a diving in the use of institutional care whilst age was more important for home care use.

The Data and Descriptive Analysis
8 Other similar studies are Payne et al. (2013) who analysed hospital admissions among people aged 20 and over in Scotland and found that the presence of physical multimorbidity was strongly associated with a higher probability of hospitalization, especially related to diagnosed mental health conditions. Palladino et al. (2016) found a positive and significant relationship between the number of chronic diseases and the use of primary, specialized and hospital care, and Schneider et al. (2009) found a positive relationship between the use of Medicare fee-for-service without institutional claims and the number of chronic diseases. 9 Carreras et al. (2018) using Spanish data document that the inclusion of morbidity controls reduced the effect of TTD up to 92%. 10 More specific drawing on two instruments: (i) the absolute value of the difference between the mortality of men and women being 80 years and older divided by the total population of this age group, and (ii) the aggregate of this year's and next year's mortality rate of the middle-age population (25-55 years). The estimations show that age still has a strong impact on costs even after controlling for mortality rates, and that the impact of TTD is driven by the youngest cohort (70-74 years).
Electronic copy available at: https://ssrn.com/abstract=3603835 Longitudinal dataset. We use data from SHARE (Survey of Health, Ageing, and Retirement in Europe) corresponding to waves 1, 2, 4, 5, 6 and 7 11 . Our variation comes from representative samples of individuals aged 50 years or above followed through during 13 years (2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017). We exploit a cross-country variation of 17 countries, a sample of 288,555 observations. The following steps were taken to retrieve our sample (see Table 1). First, only individuals who we observe in at least two consecutive waves were selected given that only for those we can verify whether they were still alive in the next wave. This leaves a sample of 186,336 observations. To build the panel dataset we select individuals who are interviewed at least twice. This requirement allows us to determine accurately if the individual living status in the subsequent wave is survivor or deceased. Individuals who are only interviewed once are discarded because we cannot be sure of their living status in the subsequent wave. Nevertheless, in the robustness checks we study the effect of attrition on our estimates and we show no effect on the results. The final sample contains 156,979 observations corresponding to 54,549 individuals (51,789 survivors and 2,760 deceased).
Descriptive statistic and sample design. Table 2 reports the descriptive statistics for the dependent variables both in the extensive and intensive margins. In some cases, a high percentage of zeros is observed (hospitalisation, stays in other health care facilities, nursing homes and formal personal care). However, the duration or intensity of providing these services may be very high (overdispersion). Similarly, when we examine outpatient visits with a doctor or nurse and the consumption of prescription drugs, we document that the probability of an outpatient visit in the last year or the probability of consuming at least one medication is on average high (89% and 75%, respectively), but exhibits overdispersion.
11 Unfortunately, Wave 3 cannot be included as the questionnaire is not comparable to the other waves. The table A1 breaks down the descriptive statistics, differentiating between survivors and deceased. The percentage of deceased individuals in the 85+ age cohort is six times higher than that of survivors (25.17% vs. 3.65%). There is a higher percentage of men and individuals who have only completed primary education in the deceased sub-sample than in the survivor sub-sample. The deceased sample exhibits lower income and wealth (even adjusted for household size. However, to the extent that the differences between survivors and deceased are largely time invariant, they will be absorbed by our fixedeffects model. Our estimations controls for co-morbidity by using the Charlston Comorbidity Index (CCI) calculated as the sum of the scores that are obtained for seven items (Charslton et al. (1987) adapted for SHARE by Kusumastuti et al. (2017). The share of individuals without any comorbidity is 20 percentage points higher among the deceased 12 . Compared to survivors, the percentage of deceased respondents that report any of these comorbidities is significantly higher for all items except for arthritis and stomach/duodenal ulcers 13 .

Empirical Strategy
Empirical Specification. The analysis of the descriptive statistics suggests identify a significant group of people who never use these services, which is known as the zeromass problem. Second, the variance of health care use is higher than the mean variance (overdispersion), resulting in highly skewed (to the right) distributions of the variables because there are a few individuals with high consumption levels. Modelling a variable with excessive zeros and overdispersion and then introducing fixed effects is highly 12 One explanation lies in that deceased individuals with no initial comorbidities, the "End-of-Life" module reports that 33% had been sick for less than 1 month and 21% had been sick between 1 and 6 months. Hence, the majority deathly illnesses came about in a very short interval of time (less than 6 months). 13 Tables A2 to A6 report the descriptive statistics of the dependent variables. Comments are reported on Appendix A. 12 complex, and typically boils down to either running a negative Poisson or binomial model 14 (Hausman et al., 1984;Allison and Waterman, 2002 Gilles and Kim (2017) refined this approach within a framework where the true generation process is unknown and unobserved individual heterogeneity exists. Our empirical specification can initially be expressed as follows: where is the outcome variable, is a vector of explanatory variables, represents an individual fixed effect, is a time-fixed effect, and contains other unobservable shocks that are common to all individuals. We could take into account intra-region unobservable heterogeneity (at the NUTS (Nomenclature of Territorial Units for Statistics) level), and especially, an instrumental variable approach that considers the potential endogeneity of time to death (TTD). The main drawback is that a linear model does not fit well a count data-generating process, and negative and non-integer predicted values could be obtained (Wooldridge, 2002). Hence, an appropriate model for modelling count data is the Poisson model (2). However, if is modelled as a Poisson random variable with parameter , it is implicitly assumed that the conditional mean and variance of the outcome variable are equal to . The model is specified as follows: The Poisson model is preferred to the negative binomial because the latter does not eliminate the influence of unmeasured characteristics (Allison and Waterman, 2002). The consistency of the fixed effects estimator is conditional on the assumption that the potential sample selection operates only through the individual specific terms (Vella, 1998).

13
Individual fixed effects ( ) pose another problem as they cannot be net out as in linear models (i.e., first differences or mean deviations). Hence, if we proceed to estimate the Poisson model with fixed effects, the number of observations that are available to estimate each individual i remains fixed, which will produce inconsistent estimates of (Neyman and Scott, 1948). However, when panel data are available, it is possible to separate the and estimates from the fixed effects estimates, which allows retrieving consistent and estimates (Blundell et al., 2002). Yet, we face the additional challenge of the potential endogeneity of the TTD problem. To address this concern, we follow Imbens and Wooldridge (2007) and their proposed control function (CF) approach, which can be extended to panel data. To do this, a linear regression for the TTD is first estimated using all the exogenous regressors and the proposed instruments to obtain the residuals. Next, a Poisson model is estimated using all explanatory variables and the residuals 15 .
However, the Poisson model that is applied to panel data cannot account for the overdispersion that exists in many of the outcome variables. Therefore, the predictions that are made using these outcomes would only have a small percentage of zeros. For this reason, our panel data specification should allow us to separate two data-generating processes: an extensive margin process (probability of the outcome being positive) and an intensive margin process (change in the outcome frequency of use). Both are independent processes such that once the outcome is positive, it can be modelled using a truncated distribution (Cameron and Trivedi, 2013) 16 .
15 Guo and Small (2016) show that the control function (CF) estimator applied to non-linear models is more efficient that two-stage least squares (2SLS) provided that instrumental variables are valid. To test the convenience of the CF approach we have estimated both models (CF and 2SLS) and performed a Hausman test. For all variables, the null hypothesis cannot be rejected, which confirms the suitability of the CF estimator (results are available upon request. 16 We model the zero value (i.e., absence of consultations or hospitalizations, no consumption of any prescribed drug…) as a conscious decision rather than a missing observation as it is considered in the Heckman approach. In fact, the separation between patients and not patients overcome the requirement of an exclusion restriction which is needed in the Heckman approach in order to identify the correlation coefficient between the two margins. An additional advantage of the two-part

14
We estimate the extensive margin following a logit model with fixed effects as below: where is the outcome variable, is the explanatory variables, and is the unobservable heterogeneity of the individual i (i.e., the propensity of a person to use a health care service or long-term care service at least once in the period). The estimation of this model using conditional maximum likelihood is based on a restricted dataset that excludes all individuals whose outcomes (0 or 1) do not vary throughout the period (Chamberlain, 1980) 17 .
Next, the intensive margin is estimated using a truncated Poisson model with fixed effects in which only the positive portion of is considered as follows: We include the same explanatory variables ( ) in both steps of the model, but there is no reason to assume that the estimated coefficients ( and ) will be equal. Furthermore, the unobservable individual heterogeneity ( ) comes from those variables (resilience, desire for independence or level of concern about diseases) that influence the quantity of social and health care services that is consumed. This model is much more flexible than the Poisson model since it can model overdispersion and underdispersion: Endogeneity of TTD. The treatment of the endogeneity of the TTD in a truncated Poisson model remains to be addressed. Gillingham and Tsvetanov (2019) proposed an estimation procedure using the generalised method of moments (GMM), which provides consistent estimates of the parameters. This paper also uses this procedure and the STATA routine that they developed.
Instruments. We use parents' age at death as an instrument for the TTD. More specifically, a wealth of literature indicates that a long lifespan for a mother decreases the likelihood that her children will suffer from specific diseases, such as hypertension or lung disease (Goldberg et al. 1996;Gjonca and Zaninotto 2008). However, other studies, such as Ikeda et al. (2006), have found that the age at death of both the father and the mother are important, and a longer lifespan for the parents decreases the probability that their children will die between the ages of 40 and 79.
The SHARE data only reports the age at death of a mother or father for the deceased sample. Therefore, parental age at death is imputed for those respondents whose parents were alive when the survey was conducted. Since age is a continuous variable, we use a multiple imputation (MI) procedure proposed by Rubin (1987) to predict the age at death of living parents 18 .

Instrument validity.
To verify the validity of the instruments, we report in the appendix the results of a linear regression for the TTD using these instruments, the other explanatory variables and year fixed effects ( lower educational levels, but on the other hand, increases with wealth and in smaller municipalities 20 . One potential thread to the identification is the presence of intergeneration transmission of lifestyles, namely that behaviours that shorten parent's life expectancy were adopted by their children, who would also experience a reduction of TTD. To address this specific concern, we have regressed the effect of parents' age at death, as well as other explanatory variables, over the probability of having sedentary lifestyle, having overweight, being obese, having ever smoked daily, being smoker at present time and having consumed at least one alcoholic beverage during the last 7 days (Table B4).
Overall, our results suggest that the effect of parent's age of decease over TTD is not channelled through potential inherited habits from parents. Finally, although genetics are still is important, we do not expect a significant change overall in later life, and using a FE estimator we are implicitly addressing this potential drawback 21 .

Results
Baseline results. Table 3 reports the results of the logit model with fixed effects for the probability of using the service (extensive margin) and the truncated Poisson model for 20 Table B3 displays the direct effect of the instruments on the outcome variables and confirms that the instruments are not correlated with unobserved variables affecting the dependent variables at a 5% significance. The exception being the probability of hospitalisation or one instrument. We are concerned with respect to idiosyncratic heterogeneity, which arises when which arises when some of the explanatory variables are correlated with time-varying unobserved shocks. following Card (1999), the correlation between the instrument and the dependent variable through the unobservables can give rise to bias in IV estimates. To address this issue Lin and Wooldridge (2019) propose a test for idiosyncratic exogeneity based on the robustness properties of the Poisson fixed-effects estimator combined with the control function approach, that is robust to robust to distributional misspecification and serial dependence. First, we estimate a fixed effects model and retrieve the fixed effects residuals. Second, use a Poisson fixed effects model over the mean function and test the significance of the residuals through a Wald test. Applying this procedure to all the dependent variables, we conclude that the null of no idiosyncratic endogeneity cannot be rejected (results available upon request) 21 We have re-estimated the logit and truncated models for the subsample of respondents whose parents had already deceased by the time of the survey to account for the possibility that deceased parents transmit the worst characteristics to their children. However, estimated coefficients for age, TTD and CCI do not show significant differences with re (results are available upon request)". estimates for the probability of hospitalisation, we identify an increase in the effect of age (from 0.005 to 0.117) and TTD (from -0.016 to -0.376). Hence, we conclude that IV estimates correct for the underestimation of the two-reference variable. Yet, even more importantly, the magnitude of the TTD coefficient declines to one-seventh (-0.054) of its previous value when we control for co-morbidities in M5. We find that a closer time to death reduces the likelihood of hospitalisation, but an increase in CCI increases the 22 Descriptive statistics are shown on Table A7. Specifically, the number of hospital beds per 100,000 inhabitants is included in the probability of hospitalization and length of stay at the hospital. The number of beds in nursing and residential care facilities per 100,000 inhabitants in the regressions for the probability of staying in a nursing home and length of stay. Finally, the number of doctors and nurses per 100,00 inhabitants is included in the probability of outpatient visits and number of outpatient visits with doctor/nurse. For those individuals whose region of residence is unknown we have applied the country average. likelihood of hospitalisation, and the effect increases with TTD. It is important to note that, as expected, controlling for comorbidities using CCI (in M5) significantly reduces the effect of ageing. Without CCI, an additional year of life increased the probability of hospitalisation by 11.7 percentage points, whilst after controlling for CCI, an additional year of life only increases this probability by 1.4 percentage points.
When examining the extensive margin of doctor or nurse consultations, we find that IV estimates results in a series of changes in the relevant estimates. First, age is no longer a significant variable, indicating that ageing per se does increase outpatient visits to a doctor or a nurse. Second, the positive effect of the TTD increases (from 0.001 to 0.042).
Therefore, visiting a doctor or nurse is primarily driven by TTD and the presence of comorbidities, especially the latter.
When we turn to nursing home use, we observe that both age and the TTD reduce the probability of nursing home care use. That said, when our estimates are corrected using an IV strategy, the effect of the TTD is four times larger (increases from -0.002 to M2 to -0.008 in M4). However, when CCI is controlled for in M5, the effect of TTD decreases by 25% (until -0.006). The positive effect of CCI exceeds in absolute value the negative effect of age.
Next, home care is examined using the same strategy, the IV estimation produces a TTD effect that is almost 10 times larger (from -0.010 in M2 to -0.095 in M3), which reinforces the idea that TTD is underestimated when omitted variables bias and reverse causality are adjusted for. However, this effect decreases by half when CCI is included (-0.049), which supports the idea that the need for home care is spurred by TTD and the existence of comorbidities.

20
Lastly, when comparing the M2 and M4 estimates on medication consumption, it can be seen that the IV estimation amplifies the positive effect of age (from 0.049 to 0.236), and it also amplifies the negative effect of the TTD (from -0.005 to -0.607). Both effects decrease when CCI is introduced in M5, and the effect of age on the probability of consuming a medication decreases by five percentage points reduce.
The extensive margin five or more medication consumptions (polypharmacy) is then estimated using a sample that is limited to individuals who consume at least one medication. When comparing M2 and M4, we find that the effect size of TTD effect increases considerably. In M2, an additional year closer to death produces a barely perceptible decrease in the probability of consuming five or more medications. In contrast, in M4, each year closer to death decreases this probability by 14.7 percentage points. Finally, the inclusion of CCI in M5 suggests that an additional comorbidity increases the probability of polypharmacy by 12.6 percentage points, but the TTD variable is no longer significant and the effect of age reduces by two percentage points.
Intensive margin. According to the M5 model specification, it is observed that the TTD and CCI exhibit opposite effects on hospital length of stay, on the number of doctor or nurse outpatient visits, and on the number of prescription drugs that are consumed. An additional comorbidity increases the probability that a hospital stay will be extended by an additional day by 15.3%. Likewise, it increases the probability that an additional outpatient visit will occur by 31% and increases the probability of consuming an additional medication by 39.4%. Contrarily, a one-year increase in the TTD decreases the probability that a hospital stay will be extended for another day by 4.9%. Similarly, it also decreases the probability of an additional outpatient visit by 4.7% and decreases the probability of consuming an additional medication by 1%. When we turn to examine the effect of age, we find that including CCI in M5 reduces its effect since each additional year of life only increases the probability of an additional day of hospitalisation by 2.3% instead of the 5.3% in M4 (without CCI).
When home care is examined, we find that TTD produces the greatest differences. A one- year increase in the TTD decreases the probability of receiving an additional hour of personal care by 30.4%. Moreover, increasing CCI increases the probability of personal care by 8%. In some cases, the inclusion of the comorbidity variable (CCI) significantly decreases the effect of TTD. For example, TTD effect decreases from -16% to -4.9% for the hospital length of stay and from -34.2% to -4.7% for the number of doctor/nurse consultations. When we examine the effect of the length of stay in the use of nursing home and hours of home care, we find that this decrease is of a smaller magnitude.
Finally, when we examine the effect of the number of prescribed drugs, we find that TTD ceases to be significant when controlling for CCI (in estimates with and without IV). M5 estimates suggest that each year of additional life only increases the probability of higher medication consumption by 3.9% instead of 12.5% estimated retrieved in M4 (without including CCI).
Ageing effects. We find that each additional year of life has a positive effect on the hospital length of stay (+2.3%) and on the number of medications consumed (+3.9%).
This effect is six and 10 times lower, respectively, than the effect of an additional comorbidity. The significance of CCI emerges when examining the number of doctor/nurse consultations, since age ceases to be significant once CCI is introduced.
Moreover, we estimate that an addition year of life decreases the length of stay in nursing homes by 13%. The largest impact of age corresponds to the frequency of home-based assistance for personal care since each additional year increases the probability of receiving one more hour by 13.6%. Figure 1 depicted the predicted probability and duration as a function of the age cohort, TTD, and the value of the Charlson Comorbidity Index. We show that the probability of hospitalisation exhibits no differences with TTD horizon, but quadruples with a six-fold increase in CCI when it varies from 0 to 6/7. In contrast, the length of stay at hospital, is shorter for a TTD and increases as CCI rises, the probability of having a doctor/nurse consultation in the last year is higher for a TTD of more than three years. However, the number of consultations is lower for a TTD of more than three years compared to other TTD horizons. Figure 1 shows that the probability of a nursing home stay reaches 40-50% for the 75-84 and 85+ age cohorts with maximum levels of comorbidity and a TTD of 0-12 months. In contrast, it is almost zero for a TTD of more than three years and for all age cohorts and CCI values, and then length of stay has a U-shaped curve for all TTD horizons. The differences in the probability of home care use (for personal care) based on the TTD become more salient for the 75-84 and 85+ age cohorts. Increasing CCI increases the distance between the predicted probabilities, and it reaches the maximum with a TTD of 0-12 months. When we turn to examine the number of hours of received care, we find a significant increase in the number of formal caregiving hours for CCI=5, and then it decreases for CCI=6. Finally, the probability of medication consumption is greater than 50% for all age cohorts (80% after the age of 75). As CCI increases, the probability of consuming one medication or of consuming five or more medications (polypharmacy) is close to one for all age cohorts.

Robustness checks
Comparison between truncated Poisson and truncated negative binomial. Table C1 compares estimated odds ratios obtained for the truncated Poisson (the same shown on Table 3) and those obtained if the count data variables are modelled using a truncated negative binomial. Although the sign and significance of the estimated coefficients are the same in both estimations, the magnitude of the effect is always higher for the negative binomial. For example, a one unit increase in CCI, raises the probability that the number of outpatient visits increases between 40.9% (truncated negative binomial) and 31% (truncated Poisson) 23 . Nonetheless, the economic explanation underpinning our results is satisfied regardless of the estimator.
Attrition. Given that our estimates could be biased by potential non-random selection of the final sample, Table C2 compares the outcome variables between the initial sample and the final sample. Test statistics for equality of means between samples accept the null hypothesis of equal means for all variables. We have also employed the test for attrition suggested by Verbeek and Nijman (1992) which involves the estimation of all the crosssections introducing as explanatory variable, a binary indicator that takes the value 1 in case that the individual is present in the final sample, and 0 otherwise. Results are shown on tables C3 (binary outcomes) and C4 (count data variables). The variable "present in all samples" is only significant for the probability of staying at nursing home and length of stay at nursing home. The effect over the probability of staying in a nursing home is very small (1.3pp.), but the effect over the length of stay is more substantial (e.g., being in the final sample increases the probability that the length of stay raises 1 week by 14.4%). In any case, we consider that attrition does not blurs the validity of our estimations.

Instrument validity:
In order to dispel any cloud of doubt surrounding our instruments (parent's age of decease) and to show that the causal inferences about TTD on healthcare outcomes are credible, we rely on two bound methods proposed by Conley et al. (2012) that allow to obtain inferences even when the instrumental variables do not satisfy the exogeneity restriction (see Appendix D for explanation of both approaches). Figure D1 shows the results of testing both approaches for the instrument "male & father's age of decease" (similar results have been obtained for the other instruments; results available upon request). The solid line represents the 2SLS father's age of decease effect estimate for the respective outcome variable. The two dash lines represent upper and lower limits of the respective test scores. Overall the results confirm that even with substantial deviation from the exclusion restriction, the instrument has still a considerable effect over the outcome variable 24 .
Effect of CCI over estimations: To verify model fitting after introducing CCI, Figure E1 compares the residuals from the logit and truncated Poisson models (using IV for CCI) conditioned on including or not CCI, that is comparing M4 with M5. For all regressions, residuals are significantly lower in the models with CCI which confirms the overperformance of M5.

Heterogeneity
Finally, in this section we study whether results were driven by specific groups of people or countries all the models were re-estimated for men and women and for two groups of countries.
Differences between men and women: Table A8 shows descriptive statistics for outcome variables and Table E1 contains the model estimation results. A one-year increase in the TTD decreases both, the probability of hospitalization and hospital length of stay, more intensively for men than for women (-3.9 pp. vs. -2.5 pp. for the probability and -5.2% vs. -3.1% for length of stay). By the contrary, the effect is more intense for women in the following cases: (i) a one-year increase in the TTD decreases the probability of one 24 In the right column figures for union of confidence intervals are presented. The x axis measures how strong does the violation of the exclusion restriction needs to be in order for the instrument to turn insignificant. In all figures, the confidence intervals do not include the value 0 (red line), so we can infer that the IV estimations are robust to possible violations of the exclusion restriction additional outpatient visit with a doctor/nurse by 4.8% for men and by 10.1% for women, and (ii) a one-year increase in the TTD decreases the probability of receiving one additional hour of home care by 21.2% for men and 32.2% for women.
With respect to the effect of CCI, each additional CCI increases the probability of receiving and additional hour of personal care by 3% for men and 10.6% for women, but decreases the probability of extending length of stay at a nursing home by one week 12.1% for men and 5.5% for women. The largest effect size for age is observed for the number of prescribed medicines consumed: an additional year of life increases the probability of consuming one additional prescribed drug by 2.6% among men, and 5.8% among women. Figure E2 show the predicted probabilities and predicted values of count data variables distinguishing by gender, TTD and CCI. It is worth noting that hospital length of stay increases significantly from the age of 75 (for high CCI, but regardless of TTD). In contrast, the length of stay at nursing home describes a U shape, with a minimum length for the cohort age 75-84 years (regardless CCI and TTD). The number of home care hours exhibits a substantial jump for the oldest cohort. Finally, we appreciate that for men and women, as the individual gets older, the higher TTD is, the steeper is the probability of consuming any prescribed drug (for low CCI).

Northern and Southern European countries:
We have selected four northern countries (Denmark, Estonia, Poland and Sweden) and three southern countries (Greece, Italy and Spain). Table A9 shows descriptive statistics for outcome variables and Table E3 shows the model estimation results. The most striking result is the different impact of CCI on the probability of hospital use (and length of stay), which turns out to be two (three) percentage points lower in southern countries than in the northern countries. The effect of ageing on hospitalisation smaller in the southern countries. In northern countries, each additional year increases the likelihood that hospitalisation will be extended by one day by 13.5% compared to 11.2% in southern countries. Furthermore, in both groups of countries, the TTD variable is significant for the probability of hospitalisation, but not for length of stay. All estimates show that the absolute value of the coefficient of TTD decreases when including CCI.
The probability and the number of an outpatient visit with a doctor or nurse decrease with the TTD. The smaller effect is on the count variable. A one-year step towards death increases the number of outpatient visits by 12.8% in southern countries and 8.2% in northern countries. Another significant difference between both country groups is the effect of CCI, which is more intense in southern countries. An increase in comorbidity increases the probability of an outpatient visit by 6.9 percentage points in southern countries, and by 5.1 percentage points in northern countries. An increase in comorbidity increases the number of outpatient visits by 34.3% in southern countries and by 29.1% in northern countries. The results for home care are also interesting. The probability of receiving formal care at home increases slightly with the TTD for both country groups.
However, the TTD's effect on the number of formal, in-home care hours is different for each country group. A one-year step towards death increases the provision of personal care by one hour (+ 26.1% in southern countries and +16.9% in northern countries).
There are significant differences in the effects of age, TTD and CCI on medication. percentage points in northern countries compared to 11.4 percentage points in southern countries. Figure E4 in the appendix shows the predicted probabilities and predicted counts for the analysed outcomes based on the age cohort, the TTD (differentiating between the two extremes of 0-12 months and 3+ years), and CCI (considering only very low comorbidity profiles (CCI=0.1) and very high profiles (CCI=5, 6, or 7). The probability of hospitalisation and the hospital length of stay are higher for northern countries. In contrast, the probability of an outpatient visit with a doctor or nurse is higher for northern countries only when CCI is low. It is higher for a TTD of 0-12 months and decreases slightly for both groups of countries for the 85+ age cohort. In contrast, the number of outpatient visits, it is higher in southern countries, and a high CCI increases the distance between both groups. Furthermore, for both countries, a greater proximity to death is associated with fewer outpatient visits. When we turn to home care, we find differences among northern countries for a TTD of more than three years, a high CCI, and after the age of 75. Finally, when we examine medication consumption, the picture is very different depending on morbidity controls.

Conclusion
This paper studies the effect of ageing on health care utilization, to disentangle the effect of ageing, from other determinants of health care utilization. We exploit longitudinal individual end of life data that measures the effect of time to death (TTD) . We control for and measure a number of comorbidities, and consider the endogeneity of TTD.    The following categories of prescribed drugs are considered: (1) high blood cholesterol, (2) high blood pressure, (3) coronary or cerebrovascular diseases, (4) other heart diseases, (5) asthma, (6) diabetes, (7) joint pain or for joint inflammation, (8) other pain (e.g. headache, back pain, etc.), (9) drugs for sleep problems, (10) anxiety or depression, (11) osteoporosis (hormonal), (12) osteoporosis (other than hormonal), (13) stomach burns, (14) chronic bronchitis, (15) suppressing inflammation (only glycocorticoids or steroids), (16) other drugs, not yet mentioned. Source: SHARE waves (1, 2, 4, 5, 6, and 7). , except that IV is used for TTD (CF for logit with fixed effects and a GMM truncated Poisson). Marginal effects are offered for the logit models, and the incidence risk ratio are shown for the truncated Poisson models. Clustered robust standard errors (at the NUTS level) with 100 bootstrap replications are obtained in all models. * statistically significant at 10%; ** statistically significant at 5%; *** statistically significant at 1%. Charlston Comorbidity Index: level 6 also includes level 7. In the graphs for the probability of hospitalization: the probability for TTD (13-24 months) overlaps with probability for TTD (+3 years). In the graphs for length of stay at hospital: length of stay for TTD (0-12 months) overlaps with length of stay for TTD (13-24 months). In the graphs for the predicted probability of consultation: the probability for TTD (13-24 months) overlaps with the probability for TTD (25-36 months).

Online Appendix
Appendix A  Kusumastuti et al. (2017) The percentage of people who have been seen by a medical doctor or a qualified nurse in the last year increases from 86% for the cohort of 50-64 years to 94% for the cohort of 85 years and older and differences for survivors and deceased are not significant (Table A2). In contrast, the average number of visits in the last year is lower for the sample of survivors (8.02) relative to the deceased (13.11). For these, there is a slight increase as TTD decreases.

The percentage of people who have been hospitalized in the last year is not significantly different in the survivor and in
the deceased sample (31.02%) and increases progressively with age among survivors (14.56%) ( Table A3). The external margin of hospitalizations strikingly exceeds 40% among the youngest deceased cohorts (50-64 years) and those 75-84 years of age, in both cases when TTD = 0-12 months. The average length of stay at the hospital among the deceased is three times that of the survivors (6.30 days vs. 1.62 days). Importantly, the average stay increases as TTD decreases (4.98 days for TTD = +3 years vs. 7.86 days for TTD = 0-12 months). Significantly, stays longer than 10 days do not correspond to the older cohort.
The percentage of people who have taken any drug at least once a week is much higher among the deceased sample (89.13% vs. 73.69%) (Table A4). For both survivors and deceased, a progressive increase is observed with increasing age, and these increases are greater in the sample of survivors (e.g. 15.94 pp for survivors and 11.08 pp for deceased when going from the 50-64 to 65-74 years age bracket). In the deceased sample, the percentage of a drug consumed increases as TTD decreases, except for the youngest and oldest cohorts.
The average number of drugs consumed is higher among all age cohorts for the deceased sample (2.82 vs. 2.28 for survivors). We document an increasing pattern of consumption as TTD decreases, with the maximum consumption corresponding to the cohort of 85+ years and TTD = 0-12 months (3.29 drugs) and the cohort of 75-84 years and TTD = 13-24 months (3.04 drugs). Consistently, the percentage of polypharmacy (consumption of 5 or more drugs at least once a week) is 8 percentage points higher in the deceased sample, increasing from 15.13% to 22.21% as TTD decreases and reaches 25% for the two aforementioned groups.
Finally, we have examined the use of long-term care, the percentage of people who have been in a nursing home during the last year is 7.5 times higher among the deceased sample (2.73% vs. 0.36%) (Table A5). Considering TTD, this figure remains stable at approximately 2% when TTD is greater than one year and increases to 4.12% for TTD = 0-12 months. The cohort of 85+ years is the most common age of entry into nursing homes (5.37%). The analysis of the average length of stay reveals some interesting characteristics: (i) no significant differences are observed between survivors and deceased (27.57 weeks vs. 28.67 weeks) and (ii) the longest duration corresponds to the cohort of 65-74 years and TTD = 25-36 years (48.68 weeks). Table A6 reports the percentage that receives personal care at home which we find it increases with age for both survivors and deceased, with the increase becoming steeper among 75-84 years to 85+ years. We document an increase as the TTD declines (from 14.14% to 24.42% for TTD = 0-12 months). Two significant cases are the result of the largest number of hours of care received throughout the year: (i) when TTD = 0-12 months, the largest number of hours corresponds to the youngest cohort and (ii) for the cohort of 85+ years, the average number of hours is higher among survivors.   We consider the following drug categories: (1) high blood cholesterol, (2) high blood pressure, (3) coronary or cerebrovascular diseases, (4) other heart diseases, (5) asthma, (6) diabetes, (7) joint pain or for joint inflammation, (8) other pain (e.g. headache, back pain, etc.), (9) drugs for sleep problems, (10) anxiety or depression, (11) osteoporosis (hormonal), (12) osteoporosis (other than hormonal), (13) stomach burns, (14)     Source: own work using data from Eurostat. Regional healthcare indicators. Data for the number of beds in nursing and residential care facilities are not disaggregated by region in Greece. The following categories of prescribed drugs are considered: (1) high blood cholesterol, (2) high blood pressure, (3) coronary or cerebrovascular diseases, (4) other heart diseases, (5) asthma, (6) diabetes, (7) joint pain or for joint inflammation, (8) other pain (e.g. headache, back pain, etc.), (9) drugs for sleep problems, (10) anxiety or depression, (11) osteoporosis (hormonal), (12) osteoporosis (other than hormonal), (13) stomach burns, (14) chronic bronchitis, (15) suppressing inflammation (only glycocorticoids or steroids), (16) other drugs, not yet mentioned. Source: SHARE waves (1, 2, 4, 5,6, and 7). Standard errors between parenthesis. T-test assuming unequal variances. Satterthwaite's degrees of freedom is an alternative way to calculate the degrees of freedom that takes into account that the variances are assumed to be unequal. The following categories of prescribed drugs are considered: (1) high blood cholesterol, (2) high blood pressure, (3) coronary or cerebrovascular diseases, (4) other heart diseases, (5) asthma, (6) diabetes, (7) joint pain or for joint inflammation, (8) other pain (e.g. headache, back pain, etc.), (9) drugs for sleep problems, (10) anxiety or depression, (11) osteoporosis (hormonal), (12) osteoporosis (other than hormonal), (13) stomach burns, (14) chronic bronchitis, (15) suppressing inflammation (only glycocorticoids or steroids), (16) other drugs, not yet mentioned. Source: SHARE waves (1, 2, 4, 5,6, and 7). Standard errors between parenthesis. T-test assuming unequal variances. Satterthwaite's degrees of freedom is an alternative way to calculate the degrees of freedom that takes into account that the variances are assumed to be unequal.

Appendix B
We have selected characteristics of the deceased respondents (EoLQ) that are also seen in the parents of respondents (MQ) who are still alive. These variables are the following: (a) sex, (b) age at death of respondent at the time of death (EoLQ) and age of the father/mother (MQ), (c) number of children of a deceased respondent (EoLQ) and the number of children of a father/mother (MQ), (d) frequency of contact of a deceased respondent with their children (EoLQ) and frequency of contact of a father/mother with the respondent (MQ), (e) distance between a deceased respondent's home and his/her children (EoLQ) and distance of a father/mother from his/her children (MQ), and (f) country and year fixed effects. The reason for including whether or not a person has children is based on evidence that indicates greater longevity for people with children (Modig et al., 2017). The reason for including the spatial distance between parents and children and the frequency of contact is because loneliness has been positively correlated with morbidity and mortality (Stressman et al., 2014). Although it is perfectly plausible that a father or mother could have other relatives, the parent/child link is the only one for which information is available in both the EoLQ and the MQ. Five different random seed values have been selected that produce five different allocations and yield very similar results 27 . Figure B1 presents the density function corresponding to the reported age at death (for parents who died prior to the survey) and the imputed age at death (for parents who were still alive at the time of the survey).
Their age at death has been imputed using the procedure that was described above. The figure shows that for both fathers and mothers, the density function for the imputed age at death is to the right of the density function for the reported age at death. The table B1 separately presents (by interviewee gender) the descriptive statistics for the reported and imputed age at death of parents. For both men and women, the imputed age at death is two to three years older than the reported age at death of fathers and mothers. 27 The percentage of imputations amounts to 53.09% for mother's age of decease and 47.15% for father's age of decease (see Table B1).

Plausible exogeneity of the instruments:
The departure point is equation (1) in which we explicitly distinguish between the potential endogenous variable (TTD) and the other explanatory variables: The first method is the γ-Local-to-Zero (LTZ) approximation bounds method, which introduces some bias term (or exogeneity error) in the approximate distribution of ̂. In other words, it relaxes the exclusion restriction requirement by allowing for uncertainty in the priors about γ. According to Conley et al., (2012) this method provides robustness with respect to 2SLS approach under the assumption that the priors are correct.
Where Υ is the distribution of γ, Σ 2 is the variance-covariance matrix for the estimation 2SLS and is the vector of instrumental variables. The distribution of the exogeneity error (Πγ � ) depends on the sample moments of the matrix Π, which shows a negative relationship between the strength of the instrumental variable and the exogeneity error, and the distribution Υ. This exogeneity error is an indicator of the deviations of ̂ from the asymptotic standard distribution of the 2SLS estimator due to non-fulfilment of the exclusion restriction assumption.
It is assumed that γ follows a normal distribution with mean γ and variance-covariance matrix Ω γ . Then, the asymptotic distribution ̂ of can be expressed as: Following Conley et al. (2012), we implement the simplest form of priors for γ, that is, γ~(0, 2 ) and computed the 95% confidence intervals for for different values of . Under the assumption that priors are correct, this approach provides valid inference and robustness with respect to normal 2SLS approach.
The second method is the Union Confidence Interval (UCI), which allows us to analyse the robustness of the estimations in case of a direct relationship between the instrumental variables (parent's age of decease) and the outcome variables. Following Conley et al. (2012) equation (1) can be modified as follows: Where denotes parent's age of decease. In a normal 2SLS estimation the term ( γ 0 ) would not be present in equation (C.4). If the strict exogeneity assumption is satisfied, parents' age of decease does not have any effect over outcome variables and thus γ = 0. The innovation proposed by Conley et a. (2012) consist in relaxing the strict exogeneity assumption (γ ≠ 0) and checking its significance in the outcome equation. Then, allowing for non-zero γ, equation (C.4) can be expressed as (C.5) where we have assumed that γ = γ 0 : Considering that the outcome variable is now ( − γ 0 ), then can be consistently estimated using as an instrument for . Under the UCI approach, is estimated given any γ 0 belonging to the specific support interval for γ: γ ϵ [− , + ]. Conley et al. (2012) notes that given that γ belongs to that interval, the union will contain the true parameter value for at least 95% of the time (if using a 95% confidence interval).

28
These figures have been obtained using the command plausexog proposed by Clarke (2014) for STATA. The γ -Local-to-Zero (LTZ) approximation bounds are drawn for different values of δ under the assumption that γ~N(0, δ 2 ). All the reported bounds are for the 95% confidence intervals generated with robust standard errors. The Union for Confidence Intervals (UCI) bounds are drawn for different values of δ, which define the support of γ (i.e., the true direct effect of parents' age of decease on TTD). Dash lines around the 2SLS estimation represent the upper and confidence intervals or the respective tests. The solid red line represents the value λ = 0. The solid black line in the γ -Local-to-Zero (LTZ) approximation represents the 2SLS class size effect estimate. Estimations performed using the command plausexog from STATA. Each graph compares residuals obtained from logit (binary outcomes) or truncated Poisson (count data outcomes) instrumenting time-to-death(TTD) with parent's age of decease. Blue straight line corresponds to residuals from regressions after including as explanatory variables age, age squared, marital status, income and wealth, size of municipality, healthcare provision by NUTS and year fixed effects. Red straight line corresponds to residuals from regressions that includes the same explanatory variables as before and also Charlston Comorbidity Index (CCI). Blue and red dashed lines corresponds to confidence intervals at 95% significance level. M4 includes as explanatory variables age, age squared, TTD, marital status, income and wealth adjusted by the number of household members, municipality size, healthcare resources by NUTS and year fixed effects. IV is used for TTD (CF for logit with fixed effects and a GMM truncated Poisson). M5 uses the same explanatory variables and estimation procedure as M4, and also includes CCI index. Marginal effects are offered for the logit models, and the incidence risk ratio are shown for the truncated Poisson models. Clustered robust standard errors (at the NUTS level) with 100 bootstrap replications are obtained in all models. * statistically significant at 10%; ** statistically significant at 5%; *** statistically significant at 1%. Figure E2. Predicted outcomes conditioned on age, time to death and Charlston Comorbidity Index (CCI). Men   -0.000*** -0.000*** -0.000*** -0.000*** 0.999*** 1.000** 0.999*** 1.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) TTD -0.132*** -0.095*** -0.127*** -0.082 *** 0.706*** 0.972*** 0.633*** 0.878*** -0.000*** -0.000 -0.000*** -0.000*** (0.000) (0.000) (0.000) (0.000) TTD -0.139*** -0.071*** -0.159*** -0.108*** (0.008) (0 , TTD and morbidity effect on health care use on both the intensive and extensive margin. All models include the following explanatory variables: age, age squared, marital status, income and wealth adjusted by the number of household members, municipality size, healthcare resources by NUTS, TTD and year fixed effects. Additionally, CCI is included in the model shown in the right column (for each pair of columns). In all models IV is used for TTD (CF for logit with fixed effects and a GMM truncated Poisson). Marginal effects are offered for the logit models, and the incidence risk ratio are shown for the truncated Poisson models. Clustered robust standard errors (at the NUTS level) with 100 bootstrap replications are obtained in all models. * statistically significant at 10%; ** statistically significant at 5%; *** statistically significant at 1%. In the graphs for the probability of consultation with doctor/nurse and CCI=5, 6 or 7: the probability for Northern countries & TTD (+3 years) overlaps with the probability for Southern countries &TTD (0-12 months).