DOES LATER RETIREMENT CHANGE HEALTHCARE CONSUMPTION? EVIDENCE FROM FRANCE

This paper examines the causal impact of retiring later on the healthcare consumption of the French elderly. While most previous work has focused on the impact of the switch from employment to retirement, I here analyze that of delayed retirement on the healthcare consumption of retirees. I exploit the 1993 French pension reform in a two-stage least squares approach to address the endogeneity of retirement. This reform led to a progressive increase in the claiming age for the 1934 to 1943 birth cohorts. I ﬁnd that delaying the claiming age by three months signiﬁcantly reduces the probability of having at least one Doctor visit per year by 0.815 percentage points for retirees aged between 67 and 75, and the annual number of Doctor visits by 1.14%. I ﬁnd a similar drop in the consumption of prescription drugs but no signiﬁcant effect on dental visits and hospital stays. These results underline the potential spillover effects between pension funds and health insurance.


INTRODUCTION
The European Commission has noted the increasing pace of pension reforms in recent decades. One third of these reforms have changed the claiming age (Carone, Eckefeldt, Giamboni, Laine, and Pamies Sumner, 2016). Many of these reforms have been unpopular, with Unions often arguing that pension reforms will produce worse health. This has triggered a growing interest in applied research to estimate the impact of retirement on health. Most of this work has found that the transition to retirement leads to better selfreported health, fewer depressive symptoms, but worse cognition and an ambiguous effect on physical health. However research on the impact of delaying retirement is rarer, and reveals no effect on mortality, and a negative or insignificant effect on self-reported health. Nishimura, Oikawa, and Motegi (2018) and Garrouste and Perdrix (2021) provide detailed reviews of this literature. The detrimental effect of later retirement on health is thus not clear, due to the multiple mechanisms that have opposing effects on various aspects of health (for example, cognitive and physical-health outcomes).
Analyzing the impact of retirement on healthcare use is a useful complement to work on health itself, from both the accounting and public-health points of view, as it allows us to identify potential financial spillover effects between the retirement and National Health systems. From a public-health perspective, the combination of the results on healthcare consumption and health is potentially important. For example, if both worse health and less healthcare use is found, one interpretation is that more retirees are forgoing medical care.
The large majority of work on retirement and healthcare consumption has focused on the impact of the switch from employment to retirement, while that of delaying retirement is much less common. These two effects may not be the same, due to different underlying mechanisms . For example, the opportunity cost of time may lie behind changes in healthcare consumption pre-and post-retirement, but cannot explain any differences between retirees with different retirement ages.
From a public-policy point of view, the effect of switching to retirement on healthcare consumption raises the question of how the employment-retirement transition can be smoothed. Conversely, the effect of delaying retirement on healthcare consumption underlines the potential unintended side effects of reforms if designed without considering their consequences on health and the health-insurance system.
I here explore the causal impact of retiring later on healthcare consumption in France using 10 waves (2005 -2015) of HYGIE data, which links administrative data from the private pension scheme CNAV to the National Health System. Healthcare consumption is measured by Doctor visits, hospital stays, and the consumption of prescription drugs. The merged data provide information on both the past careers and the current healthcare consumption of retirees. I account for the endogeneity of retirement, using the 1993 French pension reform as an instrumental variable.
This 1993 reform was the first to produce later retirement in France, and affected individuals within and between cohorts differently. Those affected by the reform reacted strongly to the incentives by increasing their claiming age (Aubert, 2009;Benallah and Mette, 2009;Bozio, 2011b) and labor-force participation (Bozio, 2011b).
I focus on men who worked in the private sector who are not eligible for disability pensions. I find that an exogenous increase in the claiming age leads to a small but statistically-significant drop in retirees' probability of having at least one Doctor's visit per year, and a 1.14% fall in the number of visits for those who do see a Doctor. I find a similar drop in the probability of taking prescription drugs, but no effect on dental visits or hospital stays. This result complements those in Hagen (2018), where later retirement reduced prescription-drug consumption in Sweden among female civil servants. I make two main contributions. First, this is one of the first papers to document the impact of later retirement on healthcare consumption. Second, my granular data allows me to look more deeply into medical care as I observe a number of healthcare-consumption outcomes, and can differentiate between prices and volumes by measuring both the number of Doctor visits and their cost (whether reimbursed to the patient or not). I also document heterogeneous effect for individuals with and without co-payment exemption (the former are fully covered by National Health System) and chronic illness. Doing so helps to identify the mechanisms behind the observed decline in healthcare consumption. For example, by showing that this fall is similar for those with and without exemptions I conclude that this probably does not reflect retirees foregoing care for financial reasons.
The remainder of the paper is organized as follows. Section 2 describes the theoretical mechanisms and the literature. Sections 3 and 4 then cover the French pension reform used as an instrumental variable, the data and the empirical strategy. Last, Sections 5 and 6 present the main results and the underlying mechanisms, robustness checks and a discussion of the results.

THEORETICAL FRAMEWORK AND EXISTING LITERATURE
There is by now a large literature on retirement and health (see Nishimura, Oikawa, and Motegi (2018) and Garrouste and Perdrix (2021)). Two phenomena should be distinguished in this context : first, the change in healthcare consumption at the switch from employment to retirement (Section 2.1), and second the impact among retirees of the timing of retirement (Section 2.2). Different methods are required to identify these two phenomena. The results will not necessarily point in the same direction, as the underlying mechanisms and therefore public-policy implications may not be the same.
2.1. The Impact of the Switch from Employment to Retirement on Healthcare Consumption Figure 1 depicts the potential impact of switching from employment to retirement on healthcare consumption, with either a rise ( Figure 1b) or a fall (Figure 1a). Most often, the effect of the switch s in Figure 1 is estimated via a regression discontinuity design (RDD) or a fuzzy regression discontinuity design (FRDD). The mechanisms behind this switch are the abrupt changes in: the propensity to pay for care, the opportunity cost of Doctor visits, the perception of the usefulness of investment in health, and mandatory care (there no longer being a need for a Doctor's attestation for sick leave or a declaration of the ability to work).    Garrouste and Perdrix (2021).
Almost all empirical work on the impact of switching from employment to retirement finds lower healthcare consumption (see Table I: Coe and Zamarro (2015); Eibich (2015); Bíró and Elek (2018); Shai (2018); Nielsen (2019); Frimmel and Pruckner (2020); Kuusi, Martikainen, and Valkonen (2020); Gorry, Gorry, and Slavov (2018); Rose (2020)). Frimmel and Pruckner (2020) is of particular interest, as they are able to isolate the age effect by taking advantage of a reform that increased the early-retirement age. They can thus evaluate the impact of retirement on healthcare consumption regardless of the age at which it occurs. Frimmel and Pruckner (2020) find better health, and propose a number of explanations: administrative healthcare consumption (via sick-leave certificates), less preventive care (less benefit from healthcare investment), the income drop at retirement, and the sudden change in the opportunity-cost of time due to stopping work. Appendix A provides a more extensive discussion of these mechanisms.

The Impact of Delayed Retirement on Healthcare Consumption
I here do not focus on the impact of the employment-retirement switch on healthcare consumption, but rather on that of delayed retirement on the healthcare consumption of retirees ( Figure 2). The most-common empirical approaches are two-stage least squares (2SLS) regression and difference-in-differences.
The underlying mechanisms are different from those for a switch. For example, the opportunity cost of time is no longer relevant for the change in healthcare consumption: once retired, this opportunity cost only depends on the utility of leisure, and there is no reason why retirement-age reforms should affect this utility function.
There are four mechanisms behind the impact of delaying retirement on retirees' healthcare consumption. These relate to the changes in the life-time total duration in employment (mechanism 1, M1), the duration of retirement at a given age (M2), investment in health, due to incentives over the life-cycle and inter-temporal income (M3), and foregoing care, due to the change in inter-temporal income (M4).
The length of the working life could affect health and the related healthcare consumption (M1). There are two opposing hypotheses. The use-it-or-lose-it hypothesis assumes that physical and cognitive abilities decline when individuals do not use them, so that later retirement is health-preserving. Conversely, the physiological reserve hypothesis considers that individuals have a stock of health, and they run down their health by working. Later retirement would then be detrimental for health. As such, the expected sign of M1 is ambiguous. This effect may moreover differ from one individual to another, with heterogeneity by gender, occupation, wage, education and so on.
The second mechanism (M2) is related to the duration spent in retirement, which can affect the duration of exposure to behavior that is particular to retirees. Again, the impact may vary by individual characteristics. Previous work has found evidence of changes in alcohol and tobacco consumption at retirement (Zins, Guéguen, Kivimaki, Singh-Manoux, Leclerc, Vahtera, Westerlund, Ferrie, and Goldberg, 2011;Celidoni and Rebba, 2017), and physical activity (Barnett, van Sluijs, and Ogilvie, 2012;Celidoni and Rebba, 2017;Godard, 2016). A higher retirement age reduces the amount of time spent in retirement and so the exposure to these behaviors. The expected direction of M2 is again unclear.
Attitudes to investment in one's own health may also change with retirement (M3). The Grossman model (Grossman, 1972)  depreciation rate of health, and so healthcare consumption. As such, the expected impact of (M3) is ambiguous. Inter-temporal income will change with later retirement (M4). Income falls at retirement, as pensions are usually lower than pre-retirement wages, so that a later retirement age delays the timing of this drop and increases inter-temporal income. As healthcare consumption rises with income (Getzen, 2006), the M4 mechanism is expected to increase healthcare use. I will here not be able to disentangle mechanisms 1 to 3 but can exclude the fourth, related to foregoing care and inter-temporal income.

Notes:
The black line is the consumption of healthcare by age for those not affected by the reform, who retired at age SRA. The blue line is the consumption of healthcare by age for those affected by the reform, who retired at age SRA . Among retirees, ∆ T is the average difference in healthcare consumption between those affected by the reform, who retired at age SRA , and those not affected, who retired at age SRA.
Only two papers have considered the impact of the retirement age on healthcare consumption by retirees (Hagen, 2018;Hallberg, Johansson, and Josephson, 2015), as shown in Table I. Hagen (2018) analyzes the impact of later retirement on healthcare consumption, and Hallberg, Johansson, and Josephson (2015) that of earlier retirement. Hagen (2018) uses the Swedish pension reform for local-government workers as an instrumental variable, which increased the claiming age from 63 to 65 for women. He finds no significant change in inpatient care, but a significant drop in prescription-drug consumption between the ages of 65 and 69. Hagen (2018) concludes that work is healthpreserving for these female civil servants. Hallberg, Johansson, and Josephson (2015) exploits the Swedish early retirement possibility at age 55 introduced for military officers (which was at age 60 pre-reform) as an instrumental variable. They focus on the impact of this exogenously-lower claiming age on healthcare consumption. They find a significant fall in the consumption of inpatient care between ages 56 and 70, which is interpreted as the consequence of better health linked to earlier retirement.

Notes:
Main results are reported for Doctor visits, hospital stays, and prescription-drug use.

Abbreviations:
The country abbreviations are the 3-letter codes from the United Nations. "Sample" column: ERP = early retirement pension; DRP = disability retirement pension; "-" = No specific sample restriction.
"Method" column: (F)RDD = (fuzzy) regression discontinuity design; DiD = difference in difference; 2SLS = two-stage least squares; TPM = two-part model. "Instrument" column: ERA = early retirement age; SRA = statutory retirement age. "Results": NS = Not significant at the 5% level; n.a. = Not applicable (outcome not analyzed  In France, the public pension at retirement depends on the number of quarters contributed. A quarter is contributed if the individual earns 150 times the hourly minimum wage during the year (producing a figure of 1204.5N in 2005). To contribute four quarters in a year, the individual thus has to earn 600 times the hourly minimum wage during the year (4 818N in 2005). As such, almost all part-time and all full-time workers contribute four quarters per year. The maximum number of quarters contributed per year is four. In addition to the quarters contributed through employment, individuals are counted as having contributed during sick leave, maternity leave and unemployment. The full replacement rate is 0.5, which falls by 1.25 percentage points per missing quarter of contribution: the replacement rate with one missing quarter is 0.4875, with two missing quarters 0.475 etc. See Appendix B and Bozio, Garrouste, and Perdrix (2021) and Blake and Garrouste (2019) for more details on the pension system and how it has been reformed. There is no financial incentive in pension terms to continue contributing quarters after reaching the full replacement-rate threshold.
Prior to the 1993 pension reform, private-sector workers were eligible for full replacement-rate pensions if they were at least 65 years old, or at least 60 and had contributed at least 150 quarters.
The 1993 pension reform was the first in France to raise the claiming age; it only applied to private-sector workers. It was not anticipated as there was no announcement before the reform, and the reform was adopted very quickly (6 months after being voted). This reform increased the required number of quarters required to benefit from a full pension from 150 to 160, cohort by cohort, starting with the 1934 birth cohort (see Table II).
The effect of the reform on the retirement age differed by both birth cohort and number of quarters contributed at age 60. As such, individuals were affected differently within the same birth cohort. In Table III, those born in 1934 who had contributed 150 quarters at age 60 had to contribute one additional quarter due to the reform, while those born in the same year but with 151 quarters at age 60 were not affected. Those affected are individuals born in 1934 who had contributed between 131 and 150 quarters at age 60, those born in 1935 who had contributed 131-151 quarters at age 60, and so on up to those born in 1943 who had contributed 131-159 quarters at age 60.
This reform changed the incentive to claim a pension at a given age. This is different from a change in the mandatory claiming age. Individuals who are affected by the reform can retire at the same age as before the 1993 reform if they accept a lower pension.

DATA AND EMPIRICAL STRATEGY
I use the 1993 pension reform as an instrumental variable to estimate the causal impact of later retirement on healthcare consumption. This approach requires data on both retirement and healthcare.

Data
The data come from 10 waves (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) of the HYGIE survey from IRDES. 1 This links administrative data from the private pension scheme (CNAV) and National Health System. It includes 1/20th of all individuals born between 1935 and 1989 who have contributed at least once in the private-sector scheme. The data include information on individual careers (the number of quarters contributed, the claiming age and the pension).
The healthcare-consumption information includes: the number of Doctor visits and total expenditure on visits, the number of days in private hospitals (including both scheduled and emergency care) and the total associated cost, and the total cost of prescription drugs. Doctor visits are those provided outside of the hospital setting. Ambulatory surgery is excluded. More details on the French healthcare system are provided in Appendix C. All health information is aggregated at the annual level for each individual between 2005 and 2015. As such, the healthcare consumption of individuals born in 1935 is observed between the ages of 70 and 80, those born in 1936 between the ages of 69 and 79, and so on.

Sample Selection and the Variables of Interest
Sample Selection.
My analysis covers retirees only. I selected all men 2 who: i) were born between 1935 (the oldest observed) and 1943 (the last cohort affected by the 1993 reform but not by subsequent pension reforms); ii) contributed at least once during their career in the private sector, as the reform affected the private sector only; iii) had contributed between 131 and 160 quarters (i.e. for between 32.75 and 40 years) at age 60; and iv) were alive and retired before age 67.
I observe consumption between the ages of 66 and 76 for individuals born between 1935 and 1943 (a detailed summary appears in Appendix Table D1). Appendix D explains why I cannot look at younger and older ages.
The method takes into account the age differences between cohorts (see Section 4.3 below). Moreover, as a complementary analysis, I provide results by age sub-samples: the consumption between the ages 66 and 68 of those born between 1939 to 1943 (each individual being observed three times: at ages 66, 67 and 68), that of those born between 1938 and 1943 between the ages of 67 and 69, and so on up to the consumption of those born between 1935 and 1939 between the ages of 74 and 76, as summarized in Appendix  Table D2.

Variables of Interest.
I measure the probability of having at least one Doctor visit per year. For consumers, i.e. for those who made use of healthcare services, whether they paid for them or not, I measure the annual number of Doctor visits and the associated expenditure in 2016 Euros. This expenditure includes both the part that is reimbursed by insurance (both public and private) and out-of-pocket payments. Doctor visits and expenditure are expressed in logarithms. I distinguish between General Practitioners and specialists.
I also provide results on the number of dental visits and the associated expenditure on dental visits, expenditure on prescription drugs, and the number of days spent in private hospitals and the associated expenditure. As I only have information on private and not public hospitals, the impact of delayed retirement on hospital care can only be interpreted as that on total hospital care under the assumption that delaying retirement does not change preferences between public and private hospitals.
Last, I calculate the total expenditure on all of the types of medical care above to produce an overall amount. I consider retirees aged 66 and over, and compare the post-retirement healthcare consumption of individuals who had different financial incentives to retire at a given age due to the reform. I have panel data, but each individual is either treated in all of her observations, or not treated in all observations. As such, it is not possible to estimate panel regressions.

Descriptive Statistics.
Table IV provides the descriptive statistics on the health and careers of the sample individuals. 44% have or had a chronic condition after age 65 (column (1)). 71% of individuals benefited from co-payment exemption, so that they did not have to make any out-of-pocket payments for Doctor visits, and 94% of these individuals visited a Doctor (in column 5).
In the main sample, 24.3% of observations correspond to individuals who did not see a Doctor. This is higher than the 10% figure for non-consumers among those 75 and older in Calvet (2012) and Sourty-Le-Guellec (1999) of French registered in the National Health System. Moreover, 72% had at least one GP visit over the year, which is again lower than the National Statistics figure (Calvet and Montaut (2013) find that after age 60 only 7% do not see a GP). One explanation is that I here focus on men who worked in the private sector and who had contributed between 131 and 160 quarters at age 60: these men are usually in better health than the average French individual of the same age, as we exclude the self-employed, inactive, handicapped, and a large part of blue collar workers (who on average left school earlier, and who are more likely to have contributed over 160 quarters at age 60), who are on average in worse health than the average population.
There are no significant differences between samples in career characteristics (the whole sample, and the sub-samples of those with Doctor visits, with chronic conditions, and copayment exemption). Figure 3 shows the point estimates from the reduced-form. Each point represent the impact of the number of added quarters required on the number of Doctor visits by year of consumption ( Figure 3a) and age (Figure 3b), controlling for age, number of quarters contributed by age 60, the Department (there are 96 of these local administrations, called département), chronic conditions and the logarithm of the pension. This shows that the average level of consumption is higher among those not-affected (the red dots) for almost all years of consumption ( Figure 3a) and at all ages ( Figure 3b). It moreover seems that, on average, the more an individual was affected the lower their healthcare consumption.

Empirical Strategy
The analysis of the impact of later retirement on healthcare consumption raises two methodological issues: reverse selection and zero consumption. Reverse selection emerges as health could affect the retirement decision (Llena-Nozal Ana, Lindeboom Maarten, and Portrait France, 2004), and those who retire later may be in better health than those who retire earlier. Second, the relatively-large proportion of non-consumers prevents us from making the usual assumption about the distribution of variables. I address both issues by using a 2SLS in a two-part model, which allows the separate analysis of the marginal effect at the extensive and intensive margins.
In this Section, I first present the instrumental variable used and then the estimation 110   equations for the impact at the extensive and intensive margins. Both equations, as I control for birth cohort (Birth) and the number of quarters contributed by age 60 (RCL60), are equivalent to a generalized difference-in-differences.
The 1993 Reform as an Instrumental Variable.
To establish the causal impact of later retirement on healthcare consumption, I use the French 1993 pension reform as an instrumental variable. This reform affected individuals according to their birth year and their quarters contributed at age 60. The reform is thus exogenous to individual's healthcare consumption. Those affected can react to the reform by delaying their retirement.
I assume that the reform is independent of healthcare consumption (the exclusion restriction). This assumption is not testable, but is credible as the reform applied to every private-sector worker, without taking their health into consideration.
To be relevant, the reform must affect the claiming age (non-zero assumption). I show in the following section that this assumption holds (in the first-stage estimation).
I assume total independence with respect to the instrument: this means, in particular, there were no anticipation or bypass effects (there is no way for individuals to move from the treatment to the control group). This is very credible for the first cohorts affected, who could not have anticipated this reform as it was the first to increase the claiming age in France. On the other hand, the last cohort affected had been aware of the reform for about a decade before their retirement. However, they have little flexibility in changing their number of quarters contributed, as individuals contribute quarters even if in part-time work, sick-leave or unemployment. They cannot change the duration of their education so as to have entered the labor force earlier. Last, I assume that individuals react in the same direction to the incentive (i.e. there are no defiers, which is the monotonicity assumption). This assumption is also credible given the substantial financial penalty for retirement before reaching the full replacement rate.
Under these assumptions, the Wald estimator identifies the local average treatment effect. The effect of this reform on claiming age was illustrated in Bozio (2011b), and its validity as an instrumental variable was assessed in Bozio, Garrouste, and Perdrix (2021). These constributions used the 1993 reform as an instrumental variable to measure the impact of later retirement on mortality. Using similar data, they show that there were parallel trends before the reform, i.e. no change in claiming age by contribution quarters for the cohorts not affected by the reform.
The Causal Impact of Later Retirement on the Probability of Healthcare Consumption.
The impact of retiring later on the probability of consuming healthcare is estimated via two-stage least squares regression. The first step is the OLS estimation, among individuals across waves, of the impact of the number of additional quarters required (∆RCL) on the claiming age (A) (see Equation 1). The second step is a linear regression of healthcare consumption on the estimated claiming age due to the reform (the 2nd 2SLS step, as in Equation 2).
The vector X ∈ M n,3 includes a control for logarithm of the pension (as a proxy for income), to take into account potential heterogeneity by wage, the number of quarters contributed for sick-leave (to proxy for health heterogeneity), and the French Department of residence (at the NUTS3 level). 3 Finally, I also control for age. As I control for birth cohort, this is equivalent to controlling for wave. The main specification includes a weight inversely proportional to the number of observations, to avoid any bias associated with the number of times each individual is observed. (1) The Causal Impact of Later Retirement on the Level of Consumption.
The impact of retiring later on the level of healthcare consumption is also estimated using a 2SLS, among those who consume healthcare. The dependent variable is the logarithm of the number of Doctor visits made by those who consume healthcare. In complementary analysis, to distinguish between the price and volume effects, the price per visit and the expenditure in Euros are also considered as dependent variables. The control variables here are the same as those in the analysis of the extensive margin. (3) Note that in much of the existing literature it is not possible to disentangle the effect of retirement from an income effect, as both changes occur at the same time. In the current empirical analysis, due to the nature of the reform used as an IV, I can observe the effect of retiring later independently of the income effect of retirement. The reform affects the size of the pension of all retirees with the same intensity. As I observe the difference between cohort and within cohort for different numbers of contributed quarters, the pure effect of later retirement (independent of income) can be identified. Another income effect could arise since those who choose not to delay their retirement when affected will receive lower pensions. This is discussed in Section 6.2. Table V shows the impact of the reform on the claiming age (in quarters). The results are shown for a large set of samples (the main sample in Column (1), those with copayment exemption in Column (2), those with chronic conditions in Column (3)), and for sub-samples of individuals having at least one annual visit to a General Practitioner, specialist, or other type of healthcare. Requiring one additional quarter for a full pension increases the claiming age by 0.859 quarters. This effect is similar in all samples: between 0.688 quarters for those with chronic conditions with at least one hospital stay, and 0.891 for those with chronic conditions who have at least one annual instance of healthcare consumption. This result thus demonstrates that the 1993 reforms can serve as an instrumental variable, consistent with Bozio (2011a) and Bozio, Garrouste, and Perdrix (2021). Figure 4 shows that the average claiming age rises with the number of additional quarters required, following a linear trend with a coefficient close to one. The variation in intensity can then be assumed to be linear.

The Impact of Later Retirement on Healthcare Consumption
Column (1) of Table VI shows the causal impact of delaying retirement on the probability of: having at least one visit to a General Practitioner, a specialist, any Doctor visit, a Dentist, consuming prescription drugs, and staying in a private hospital. The naive analysis and the reduced-form appear in Appendix Table E2. An exogenous rise in the claiming age of one quarter statistically significantly reduces the probability of having at least one Doctor visit by 0.815 percentage points, i.e. one fewer individual out of 120. The drop in the probability of visiting a General Practitioner (−0.828 percentage points) is slightly larger than that for visiting a specialist (−0.724 percentage points). Probably as a consequence of fewer Doctor visits, the probability of consuming prescription drugs falls by 0.599 percentage points. However, there is no significant change in the probability of visiting a Dentist or a hospital stay.
Among those who consume healthcare, an exogenous one-quarter increase in the claiming age significantly reduces the annual number of Doctor visits by 1.14% (Table VI  Notes: Standard errors in parentheses. * p < 0.10, * * p < 0.05, * * * p < 0.01. This Table shows the OLS estimates of the impact of the number of added quarters required due to the reform on the claiming age. Sample: Men who contributed at least once in the private sector and had contributed between 131 and 160 quarters by age 60, and retired before age 67. Estimation for the main sample (Column (1)), the sub-sample of individuals with co-payment exemption (Column (2)), and those with chronic conditions (Column (3)). The first line is for all individuals, the second to fourth lines for those with at least one annual General Practitioner, Specialist Practitioner and any Practitioner visit, respectively. Lines five to seven refer to Dental visits, stays in private hospitals, and prescription drug consumption. The F-statistics of the instrumental variable and the adjusted R 2 's of the model appear in Appendix Table E1.  visits per individual, i.e. one fewer for one in 10 individuals. The overall consumption of Doctor visits drops significantly by 0.9%. Expenditure on Doctor visits is 1.64% lower, but the price per Doctor visit is unaffected. The change in expenditure is thus driven by fewer Doctor visits. The number of GP visits falls by 0.522%, and that to specialists by 0.797%. Neither of these figures are significant, due to statistical-power limitations (fewer observations and a smaller effect than for the analysis of the total number of Doctor visits).
There is no significant impact on the number of dental visits or expenditure on Dentists, nor on the number of days in private hospitals or expenditure on hospital stays. This is consistent with there being no substitution between ambulatory and hospital care.
The fall in healthcare consumption is similar in the whole sample and the sub-sample of individuals with chronic conditions (Table E3): among the latter, retiring later reduces the probability of a Doctor visit by 1.46 percentage points and the number of Doctor visits by 1.80%. These figures are only slightly higher than those in the main sample, so that the healthcare effect of later retirement is not driven by those with chronic conditions.
The size of the effect is fairly stable over a large set of sub-samples of different ages (see Figure E1. Each sub-sample correspond to specific cohorts and years of observation, as explained in Table D2).

The Size of the Effect and Comparison to the Literature
The results for hospital stays and prescription drugs complement those in Hagen (2018) for female civil servants, and are of similar size and sign. At the individual level, the size of the drop is only small.
Delaying retirement by one quarter leads to one fewer healthcare consumer out of 120 individuals; among consumers, the drop is of one fewer visit for one individual out of ten. The overall effect (calculated from the extensive and intensive margins in Ta Figure 5, this effect on healthcare consumption is smaller than that of switching to retirement found in previous work. While the magnitude of the effect is small at the individual level, it does still represent a large figure for the National Health System, which covers around 30,000 individuals per cohort at the ages affected by the reform. A rough calculation would be 1.8 Million Euros per year and per cohort saved due to the reform. However, this lower public expenditure has to be compared to any changes pre-retirement. Figure 6 represents schematically the effect of retirement on the consumption of healthcare by age for individuals who retire at age SRA (the black line) and at age SRA' due to the reform (the dashed blue line). From existing work, they are all exposed to the same drop in consumption S at retirement, but not at the same age (respectively SRA and SRA') (Coe and Zamarro, 2015;Eibich, 2015;Bíró and Elek, 2018;Nielsen, 2019;Frimmel and Pruckner, 2020). Moreover, those who retire later do not experience the same healthcare consumption trend during retirement, as shown in Hagen (2018) and in this paper. The overall effect over the life cycle of retiring later on healthcare consumption is greater expenditure between the ages of SRA and SRA' (the grey lozenge) and lower expenditure after age SRA' (the blue triangle).

Underlying Mechanisms
As mentioned in Section 2.2, there are four potential mechanisms behind this result: a longer working life (M1), a shorter retired life (M2), health investment (M3), and foregoing care (M4).
Mechanism M4 is unlikely to explain the lower healthcare consumption due to later retirement. Were the result to be explained by individuals foregoing care, we would expect no effect among those with co-payment exemption. However, the results turn out to be the same for this latter group: a fall of 0.740 percentage points in the probability of having a Doctor visit and of 1.30% in the number of Doctor visits (see Table F1). Moreover, there is no impact on Dentist visits, which is typically the first type of care that individuals forego.
We can also exclude an income effect, for at least two reasons. First, all individuals in a given cohort, regardless of whether they were affected, experienced the same change in their pension due to the reform. In addition, retiring later increases inter-temporal income: were this mechanism to apply, it would produce an opposite-signed effect to that we observe in this paper.
Mechanism M3 cannot be excluded. The fall may be attributable to less health investment, through a drop in preventive care by those who retire later, which could compensate for greater investment during the working life. This would be consistent with the lower preventive care found at the switch from employment to retirement (Frimmel and Pruckner, 2020;Eibich and Goldzahl, 2021). Unfortunately, I cannot test this hypothesis.
Mechanisms M1 and M2 cannot be excluded either, so that less healthcare consumption reflects better health. This could be due to the time spent in employment (as in the useit-or-lose-it hypothesis) or the time spent in retirement. However, I show that there is no significant change in the probability of having a chronic condition (Table F2) nor in the probability of having co-payment exemption.
The results are therefore not in favor of an income effect or foregoing care. The observed fall in healthcare consumption due to later retirement may be explained by both a health 119 Last, there may be an effect from this specific being both unexpected and quickly implemented, which may have forced employers to adapt the schedules of employees who were about to retire. This could produce a more gradual transition from employment to retirement that ends up being health-preserving, via arrangements with employers (e.g., lighter work schedules, working hours, part-time paid full time, and pre-retirement transition leave). This is not a general mechanism, like M1 to M4, but rather one that is specific to this reform, and due to the transitory arrangements made by employers and employees. I cannot here establish whether this mechanism explains the results.

Robustness Checks
Sample Selection on Quarters Contributed.
As the reform only affected individuals who had contributed between 131 and 160 quarters at age 60, the analysis sample is restricted to this population. Small changes in these threshold figures are not expected to drive the results, but large changes may do so, by including individuals who were not affected and have very different characteristics to those of the affected. For example, considering individuals with very long careers (more than 170 contributed quarters by age 60) leads to the inclusion of individuals who were not affected by the reform as they started working at around age 16 (the control group) while the treatment group started working between the ages of 23 and 27. Moreover, note that it is not possible to include individuals with either below 131 quarters or over 160 quarters as they are not affected by the reform but left work at very different claiming ages (the first group left at age 65 and the second at age 60). I show that small changes in the contribution-length threshold of individuals included in the study does not statistically significantly change the results, while large ones do (see Figure G1 in the Appendix).

Sensitivity to the Econometric Specification.
I show that the results are not sensitive to the econometric specification, as the use of an IV-probit or IV-GMM using linear or Poisson regression models, does not significantly affect the results at the extensive and intensive margins (Table G2).

Discussion
Using a Two-Part Model.
A relatively-large proportion of the observed population has zero consumption at a given point of time. There are two main ways of tackling this issue: sample-selection models (e.g., Heckman models) or two-part models. There is a large literature on the choice between these two (Leung and Yu, 1996;Jones, 2000), called the 'cake' debate (Mullahy, 1998). While both models have advantages and drawbacks, Dow and Norton (2003) and Madden (2008) provide choice criteria between the two-part and Heckman models. I choose the two-part model for a number of reasons. First, in the absence of a good exclusion variable, two-part models provide better estimates than sample-selection models (Manning, Duan, and Rogers, 1987;Hay, Leu, and Rohrer, 1987 find a convincing exclusion variable, and the use of the Inverse Mills Ratio leads to nonconvergence. Second, the sample-selection issue is not likely to be relevant when zero is not a missing value but a real zero (Dow and Norton, 2003). Third, this choice allows the use of the logarithmic transformation without any difficulties regarding individuals with zero consumption. Fourth, as emphasized by Belotti, Deb, Manning, and Norton (2015), this very flexible specification does not require any assumptions regarding the independence between the binary dependent variable and the conditional-on-positive continuous variable. Fifth, the instrumental-variable approach reduces the limitations associated with the use of two-part models. Last, delayed retirement might have different effects on the probability of consuming and the level of consumption. From a public-policy point of view, this distinction is of great interest, as the potential mechanisms behind zero consumption may not be the same as those behind changes in its level. For example, it can be argued that foregoing healthcare can only be measured through the probability of consumption, and not its level. Moreover, measuring only the overall effect could lead to incorrect interpretations in terms of public policy. 4 Assumption on the Exclusion Restriction.
The instrumental variable affects both compliers, by increasing their claiming age, and never-takers by reducing their pension, by 1.25% for each missing quarter. However, the exclusion restriction requires that the instrument does not affect the never-takers and my estimates could thus be biased. I show that this is unlikely, and even if there is bias, it is possible to estimate its magnitude to determine whether the results remain credible despite the bias.
Bias is unlikely as the effect of the reform on never-takers would pass via an income effect on healthcare consumption: they would reduce their healthcare consumption as their pensions are smaller than expected. However, the estimated effect is similar for those with co-payment exemption, for whom healthcare is free. As shown in Table F1, their effect for the probability of a Doctor visit is -0.00740 (-0.00815 in the main sample) and that among consumers is -0.0130 (-0.0114 in the main sample). If the result is biased by an income effect for never-takers, we suspect that this is only limited in size.
Moreover, if there is bias then the reduced-form would be a weighted average of the impact on compliers and never-takers. We do not know the share of compliers as our instrument is not dichotomous, but as the effect is close to a linear function with a coefficient of 1 (see Figure 4), the proportion is expected to be similar to the first-stage estimate (0.859 in the main sample and 0.839 for the sample of those with Doctor visits). We can thus calculate the theoretical decomposition between never-takers and compliers according to the proportion of compliers. Figure G2 shows that a fall of -0.007 in the probability of a Doctor visit (the main result in the reduced-form) can be decomposed into effects 4 Imagine that pension reform both increases the share of individuals foregoing healthcare for financial reasons but also worsens the health of those who can still afford care. The latter increase their consumption of healthcare as a consequence of their worse health. For example: consider a population in which 80% consumed healthcare before the reform, with each spending 10 Euros. Average consumption is then 8N. Post-reform, the share consuming falls to 60%, but expenditure of those who consume rises to 12N. Average expenditure is now 7.20N. We would wrongly conclude that the reform has reduced healthcare consumption. In reality, it increases expenditure among those who already consumed healthcare, but reduces the probability of healthcare consumption. on compliers and never takers, according to the share of compliers (with values presented being close to those in the first stage: 70%, 75%, 80%, 85%, 90% and 95%). For the 85% figure (as observed in the first-stage regression), the 1.25% pension fall for never-takers would have to produce at least a 5% lower probability of consumption and a 6% lower healthcare-consumption level for the estimated overall effect to conceal a zero or positive effect among compliers. The effect on never-takers would thus have to be very large in order to affect the main results.
External Validity.
The population of interest covered affects external validity. The reform did not affect those with long careers (over 160 quarters at age 60, i.e. who started working before age 20) or with large career gaps (under 130 quarters at age 60, i.e. who started working at age 27.5 if the career were continuous ). Moreover, the reform does not affect those who are eligible for disability pensions. This paper thus focuses on the impact of later retirement on the healthcare consumption of those who are in better health than the average French population, and we cannot extend its results to claiming-age reforms for the whole population.
Unobservable Effects before Age 66 and after Age 78.
The decrease we find holds between ages 66 and 76. In order to have an inter-temporal view of the impact of later retirement on healthcare consumption, it would have been useful to be able to measure the impact between ages 60 and 66, and after age 78.

CONCLUSION
In this paper, I used a two-stage least squares estimator in a two-part model to estimate the causal impact of retiring later on healthcare consumption at the extensive and intensive margins. I focused on a sample of men who worked in the private sector, aged from 66 to 76. I used the first French pension reform that increased the claiming age as an instrumental variable. This paper is one of the first to focus on the impact of retiring later on retirees' healthcare consumption, rather than on the impact of switching from employment to retirement. I show that delaying retirement due to the reform by one quarter significantly reduces the probability of having at least one Doctor visit, as well as the number of visits. This result refers only to a particular population: private-sector men who contributed between 131 and 160 quarters at age 60. These individuals have better than average health, so the results cannot be extended to the whole French population. Two main potential mechanisms can explain lower healthcare consumption. First, part of the population may benefit from better health, as in the use-it-or-lose-it hypothesis, or exhibit worse health-related behaviors during retirement. Second, retiring later could produce less health investment during retirement. On the contrary, income and price effects do not explain lower consumption, and nor does foregoing care.
We cannot make a conclusive statement regarding the impact on government spending of lower retiree healthcare consumption due to the reform, as any financial gain has to be compared to pre-retirement consumption for those who delayed retirement. Further research on the underlying mechanisms and potential heterogeneity would be useful in drawing public-policy recommendations. If lower healthcare consumption is associated with better health, via use-it-or-lose-it, the recommendation is for policies that increase 122 retirees' incentives to maintain their social, physical and cognitive activities. If the health improvement reflects investment in health during employment, the public-policy advice would be to increase incentives for health investment during retirement. If there is no change in health due to later retirement, but less use of preventive healthcare, public policy should increase the incentives to take advantage of preventive care during retirement. APPENDIX A. Literature on the Impact of the Switch from Employment to Retirement Research has evoked a number of mechanisms to explain the change in healthcare consumption at the switch from employment to retirement. The main mechanism is through improved health (Coe and Zamarro, 2015;Eibich, 2015;Shai, 2018;Frimmel and Pruckner, 2020).
However, Bíró and Elek (2018) believe that changes in health are slow, and that the sudden change in the healthcare consumption at retirement cannot be attributed to health. They suggest three other mechanisms. First, workers require Doctor visits for sick-leave certificates, while retirees do not (see also Nielsen (2019)). Second, individuals may be willing to invest more in their health while working. Along these lines, Frimmel and Pruckner (2020) find reduced participation in screening and preventive care at retirement, as do Eibich and Goldzahl (2021). Third, the income drop at retirement could diminish healthcare consumption. The sudden change in opportunity cost of time may also affect healthcare consumption, but the direction of the effect is ambiguous: individuals may consume more care during retirement, due to the lower opportunitycost of time. On the contrary, individuals may prefer using time off of work to visit a Doctor rather than their personal time for the visit during retirement (Nielsen, 2019).
Only three contributions find no fall in healthcare consumption at retirement (Bíró, 2016;Lucifora and Vigani, 2018;Zhang, Salm, and van Soest, 2018). Lucifora and Vigani (2018) use cross-country differences in the retirement age in Europe. Assuming a normal distribution for the number of Doctor visits, they find more visits upon retirement. They explain this result by the opportunity cost of time (which is implicitly assumed to fall, leading to more visits during retirement). Zhang, Salm, and van Soest (2018) also find an increase in healthcare consumption in urban China, explained by the particularly high opportunity-cost of time there. Last, Grøtting and Lillebø (2020) investigate heterogeneous effects and find a negative insignificant result for hospital stays in Norway, but a significant for those with the lowest levels of education.

B. The 1993 Pension Reform
The pension amount is calculated as following: P = τ × P C × W ref , with τ being the replacement rate, P C the proratisation coefficient and W ref the reference wage (which is the best N years of wages).
The replacement rate formula is τ = 0.5 − δ × max[0; min(4 × (65 − a); D − d)] with a being the claiming age, D the number of contributed quarters required to benefit from the full replacement rate, d the number of quarters contributed, and δ the minimization coefficient (equal to 1.25% per missing quarter).
The 1993 reform made the following changes: • D, the number of quarters required to benefit from a full pension: The change in the number of quarters required thus affects only a small percentage of individuals in each cohort as, independently of the change in D, all those with short careers will leave at age 65 and those with long careers will leave at age 60. The changes in N and indexation affect all cohorts in the same way. As a result, the difference between individuals affected and unaffected by the change in D within each cohort will capture only the effect of D in the 1993 reform.

C. The Health Insurance System in France
Almost all French have public mandatory health insurance. The reimbursement rate is defined at the national level for each type of care, except in the Alsace-Moselle region, where there is a higher coverage rate. Moreover, contributions vary by labor-force status (student, worker, unemployed, retired etc.).
On average, this mandatory public coverage reimburses 78% of healthcare expenditure. However, there is heterogeneity in the reimbursement rate. In particular, individuals suffering from chronic illness benefit from 100% reimbursement of their associated expenditure.
To cover expenditure which is not reimbursed by the National Health System, individuals can take out private supplementary health insurance. This can be through an individual contract or a collective employer contract. Before 2016, private-sector employees could benefit from supplementary health insurance via collective firm contracts paid for by both the employee and the employer. The firm had to pay at least 50% of the insurance premium. In 2015, 51% of private-sector firms offered this collective supplementary insurance (Lapinte and Perronnin, 2018), and 75% of private sector employees had collective supplementary insurance. At retirement, employees can continue to benefit from this coverage, but without any firm contributions. Following the Evin Law (1989), the insurance company has to apply the same fees for the first three years after retirement: these supplementary-insurance fees can change after this period. Most retirees are not affected by this price change as they change their health insurance at retirement. Franc, Perronnin, and Pierre (2007) show that 51% of the beneficiaries of mandatory collective firm contracts switch their contract at retirement, with analogous figures of 39% for beneficiaries of optional collective firm contracts, and 23% for new retirees with an individual supplementary health insurance.
The supply of healthcare is highly regulated in France, with different rules for hospital care and ambulatory care (those out of the hospital setting). Ambulatory care is provided by General Practitioners and specialists, and fees are regulated. GPs provide primary care and ensure the continuity of medical follow-up. They play a gatekeeper role. There are 26 main medical specialities in France, covering oncology, rheumatology, dermatology, ophthalmology, cardiology, etc. Dentistry is not one a French medical speciality. The standard regulated price for a GP visit in France in 2015 was 23N , while that for visits to specialists depends on the speciality (varying from 25 to 150N). The National Health System reimburses 70% of the regulated Doctor fees, with a few exceptions. Since the Douste-Blazy law (2005), the healthcare pathway encourages patients to visit first a GP in order to obtain a prescription to see a specialist. There is an incentive to do so, as the National Health System reimbursement rate is cut by 40% if the patient did not first see a GP.
Hospitals can be private for-profit, private non-for-profit, or public. The fees are composed of i) a fixed amount per day, called "forfait hospitalier", paid by the patient or their supplementary health insurance, and ii) fees that vary according to the care received. National Health System reimburses 80 to 100% of these fees; the patient or the supplementary health insurance covers the rest.
If an individual is administratively acknowledged to suffer from a long-term care illness, the reimbursement rate is 100% for all medical care related to that illness. There are a number of other circumstances in which individuals benefit from co-payment exemption: having an inpatient stay of over 30 days or having had a work accident, for example.

D. Data Details
The Selection of Sample Men. Figure D1 shows distribution by contribution length at age 60, first in the HYGIE data ( Figure D1a) and then in the exhaustive CNAV data from the private pension scheme ( Figure D1b). The male distributions are similar in the two datasets, but not those for women. The HYGIE data do not include information on the additional quarters received. As these additional quarters are mainly for childbirth, and for the years when parents stopped working in order to raise children, the number of contributed quarters for women who benefited from these added quarters is wrong. I underestimate women's contributed quarters in HYGIE data, and in particular those for women with many contributed quarters at age 60. This is unsurprising for two reasons. First, women with many contributed quarters at age 60 are those with lower education, and in France this group has on average more children (Davie and Mazuy, 2010) and so more quarters for childbirth. Second, they may also more often stop working to raise children, and thus have associated contributed quarters that I do not observe in HYGIE data. I observe individuals' healthcare consumption between the ages of 66 and 76. There are several reasons why I cannot look at younger ages. First, as individuals retire at up to age 65, the sample of those who retired between age 60 and 65 suffers from selection bias. Those most affected by the reform have an incentive to retire closer to age 65 and not 60, while those who are not affected have an incentive to retire at age 60. Second, post-retirement, many French change their supplementary health insurance. In the very short term, this can lead to many temporary manipulations of healthcare timing. However, once all individuals have been retired for at least one year, this manipulation no longer takes place. Third, at age 62, I observe only cohort 1943, and at age 63, I observe only cohorts 1942 and 1943, and so on. Thus, starting at age 66 allows the observation of at least five cohorts.

Selection by Year of Birth and Year of Observation.
I summarise in Table D1 the year of observation of each cohort at each age. The grey area covers the observations excluded in the main analysis. Table D2 shows the same Table,

E.1. Additional Details Concerning the Main Results
The tables below are: i) the main results with the adjusted R 2 , and the F-stats of the instrumental variable and the model in the first-stage, and ii) The reduced-form and naive analysis at the extensive and intensive margins. Notes: Standard errors in parentheses. * p < 0.10, * * p < 0.05, * * * p < 0.01. This table shows the impact of the number of added quarters required due to the reform on the claiming age. Sample: Men who contributed at least once in the private sector, had contributed between 131 and 160 quarters by age 60, and retired before age 67. Column (1) to (3): same as in Table 5. Column (4): Estimation in the sub-sample of individuals with a chronic condition. This estimation is without the controls for Department. It complements column (3) where the F-stat of the model cannot be calculated as the covariance-variance matrix is not full rank. The first line is with all individuals following the selection criterion of the column, the second and third lines for those with at least one annual GP and specialist Practitioner visit, and the last line for individuals who consume at least one type of care  To see if the effect is driven by individuals with particular healthcare consumption, I consider the impact of retiring later on the probability of Doctor visits and the level of consumption of individuals who had or have a chronic condition. I identify those with a chronic condition through the variable "Have administrative recognition of a chronic condition". Table E3 shows that for individuals with chronic conditions, a onequarter rise in the claiming age post-reform significantly reduces the probability of a Doctor visit by 1.46 percentage points and the number of Doctor visits by 1.80% among consumers. These figures are slightly higher than those in the main sample.  Results for Sub-samples of Three-year Age Groups. Figure E1 shows the main results for age-range sub-groups of individuals (66-68, 67-69, etc. up to 74-76). The difference between these age groups cannot be interpreted as an age effect as the estimates come from different samples: different cohorts, observed in different years, with different treatment intensities (see Table D2).  The Impact among Individuals with Co-payment Exemption.
To test for potential price-sensitivity, I focus on the sub-sample of those who benefit from co-payment exemption. The exempt in the sample are those who require care due to work accidents and occupational illness, and individuals with long-term illnesses. Table F1 shows the impact of retiring later on the probability and level of healthcare consumption. Delaying retirement by one quarter due to the reform significantly reduces the probability of a Doctor visit by 0.740 percentage points and the number of Doctor visits by 1.30% among consumers. These results are not statistically different from those in the main sample, which then do not seem to reflect price elasticity. Results on the Probability of Having a Chronic Condition after Age 65. Table F2 shows that delaying retirement by one quarter produces no significant change in the probability of having a chronic condition after age 65. This insignificant result holds for all age-group sub-samples tested ( Figure F1). There is no significant change in the probability of co-payment exemption.   Probability to have a chronic condition after age 65 6 6 -6 8 6 7 -6 9 6 8 -7 0 6 9 -7 1 7 0 -7 2 7 1 -7 3 7 2 -7 4 7 3 -7 5 7 4 -7 6 Selection on age at consumption CI at 95% Point estimate (b) Impact on the Probability to Co-payment Exemption Probability to have a copayment exemption 6 6 -6 8 6 7 -6 9 6 8 -7 0 6 9 -7 1 7 0 -7 2 7 1 -7 3 7 2 -7 4 7 3 -7 5 7 4 -7 6 Selection on age at consumption CI at 95% Point estimate Notes: This graph shows the average impact of retiring later on the probability of having a chronic condition after age 65 and co-payment exemption, by age range. 95% Confidence Intervals. Sample: Men who had contributed at least once in the private sector and had contributed between 130 and 180 quarters at age 60, and retired before age 67.  Notes: Individuals who contributed between 0 and 130 quarters at age 60 are not affected by the reform: they retire from age 65 at the full replacement rate, and started working at age 27.5 if they had a continuous career.

G.1. Sample Selection on the Number of Quarters Contributed
In the main analysis, I selected individuals who had contributed between 131 and 160 quarters by age 60. All those who had contributed between 131 and 151 quarters were affected by the reform; some of those who had contributed between 152 and 159 quarters were affected, depending on their year of birth (see Table G1). In a difference-in-difference design, I have to include at least one contributed quarter where nobody is affected by the reform. I can thus include all individuals who contributed fewer than 130 quarters or more than 160 quarters. However, I cannot include both as those with under 130 quarters leave with a full replacement rate from age 65 while those with over 160 leave at age 60. I include as the control group those with 160 quarters or more. This choice is more logical, as the affected are those who would have retired at age 60 but now retire later due to the reform. It arguably makes less sense to compare them to the group of individuals who leave at age 65 than to the group who leave at age 60.
Last, the identification strategy relies on those who are affected being similar to those who are not. This holds when comparing individuals with 159 and 160 contributed quarters. However, it may well not when comparing those with 159 and 180 quarters. Estimation is then less precise, or even incorrect, when including individuals many contributed quarters. Figure G1 shows the results when changing the number of contributed quarters in the sample. The dashed line separates the cases where the treated are compared to the non-treated who leave at age 65 (all points to the left of the line) and to the non-treated group who leave at age 60 (points to the right of the line). As expected, the results in the main sample (131 to 160 quarters) are not statistically different from those using fairly-similar contribution-length thresholds. Figures G1(a) and (b) show the impact of the reform on the claiming age. The points the most to the left of these figures reveal smaller average impacts. This is unsurprising as the further the dot is to the left, the more individuals who are not affected by the reform and leave at age 60 are included. Figure G1(c) shows the impact of delaying retirement by one quarter due to the reform on the probability of consumption. The effect here is not statistically different from that in the adjacent samples. However, adding a large number of individuals unaffected by the reform, who leave at age 65 (thus, later than those affected), leads to an insignificant impact or, in three cases out of eight, a positive significant impact that is close to zero. Adding individuals with very long careers, who leave at age 60 and are not affected by the reform yields a smaller but still significant effect, except in the three last samples tested (of individuals who started working at close to age 16). Figure G1(d) shows the impact of one quarter more due to the reform on the number of Doctor visits (in logs). The impact is never statistically different when adding individuals who leave at age 65. It is not statistically different for the two samples with individuals who leave at age 60 and that are the closest to the main estimation. However, once we add a large number of individuals not affected by the reform who leave at age 60, the effect becomes insignificant. Note that it is probable that the individuals with 170-180 quarters contributed at age 60 have different health conditions than those with 150-160 contributed quarters, and so are not entirely comparable.
I never consider individuals with more than 180 quarters, i.e. those who started work before age 16. These are found only in the cohorts born before 1943 (full-time employment under age 16 has been forbidden by Law since 1959). In my sample, these individuals are rare, and I consider them to be atypical as compared to the rest of the sample.   Notes: Standard errors in parentheses. * p < 0.10, * * p < 0.05, * * * p < 0.01. All coefficients are marginal effects. Retiring one quarter later due to the reform leads to an average rise in the probability of a GP visit by 0.7 percentage points estimated using a probit model. In the last two columns, only observations on consumers are included in the regressions.

G.3. Estimation of the Potential Exclusion-Restriction Bias
A potential exclusion restriction bias could arise as the reform affects compliers by delaying retirement but also never-takers via lower pensions. With this bias the reduced-form (estimated coefficient β) would be a weighted average of the effects on compliers (β C ) and never-takers (β N T ). Thus: β = Cβ C + N T β N T , with C and N T being the shares of compliers and never-takers.
When the instrumental variable is a dummy, the coefficient estimated in the first stage is the share of compliers. Here, the instrumental variable is not a dummy but the effect is close to a linear function with a coefficient of one. We thus know that the share of compliers should be close to the first-stage coefficient (in the main sample this is 0.859). We thus calculate β C and β N T for the reduced-form estimates at the extensive (-0.00700) and intensive (-0.00956) margins, as reported in Table E2 and for different shares of compliers with figures around those in the first stage: 70%, 75%, 80%, 85%, 90%. Figure G2a shows the betas at the extensive margin, and Figure G2b those at the intensive margin.
These show that with 80% of compliers, the impact on the never-takers has to be above -0.05 at the extensive margin ( Figure G2a) and -0.03 at the intensive margin ( Figure G2b). These represent large impacts on never-takers.