2.1 Early retirement and caregiving

The constraints to the supply of care are affected by individuals exit of the labour market. Consistently, some studies show a lower supply of round-the-clock care among employed individuals compared to retirees (Vlachantoni, Reference Vlachantoni2010). Caregiving increases the probability of retirement among men, and this is especially the case when the recipient of care is the spouse (Raab, Reference Raab2017).

Dentinger and Clarkberg (Reference Dentinger and Clarkberg2002) document that the probability of retirement is five times higher for women who care for their spouse or partner compared to those who don't. Similarly, other research finds that women providing at least 20 hours/week of informal care are 1–3 percentage points more likely to retire relative to other women (Jacobs et al., Reference Jacobs, Van Houtven, Laporte and Coyte2017). That said, such effects are more common among ‘intensive caregivers’ (Jacobs et al., Reference Jacobs, Laporte, Houtven and Coyte2014), and are driven by income effects (Schils, Reference Schils2008). However, to date there is limited evidence of an effect of caregiving on the probability for retirement (Debrand and Sirven, Reference Debrand and Sirven2009).

2.2 Gender differences in caregiving

A priori, it is not possible to discern whether gender matters for early retirement, as employment can, under some circumstances be a source of respite for caregivers (Mooney et al., Reference Mooney, Statham and Simon2002; Schneider et al., Reference Schneider, Trukeschitz and Ponocny2013). Dentinger and Clarkberg (Reference Dentinger and Clarkberg2002) found that men are less willing to exit the labour market, which in turn delays the decision to retire. This is because women are more likely to be the primary caregivers when their spouses need care (Allen, Reference Allen1994; Neal et al., Reference Neal, Ingersoll-Dayton and Starrels1997; Szinovac and Davey, Reference Szinovac and Davey2004). Consistently, Schneider et al. (Reference Schneider, Trukeschitz and Ponocny2013) document that providing care to an older person reduces women's exit from the labour force. However, in some countries, women exhibit an incentive to exit the labour market earlier, even when it has lower knock-on effects on their future pension entitlements, which is typically the case when statutory pension differs by gender (Carr et al., Reference Carr, Murray, Zaninotto, Cadar, Head, Stansfeld and Stafford2018).

2.3 Health shock and retirement

Caregiving decisions can be commonly influenced by health shocks in the household. Johnson and Favreault (Reference Johnson and Favreault2001) report that the probability of retirement increases (for both men and women) if the partner/spouse is already retired. However, Coile (Reference Coile2004) found evidence of the ‘added worker effect’ of disability subsidies among men only, that is, disability benefits ameliorate the labour supply response of a caregiver after a spousal health shock. In contrast, the opposite is found among women.Footnote ⁶

2.4 Care subsidies and supports

So far, there is little evidence that long-term care (LTC) reforms influence retirement decisions. Most of the existing evidence refers to employment effects resulting from reforms in Germany and Japan. Geyer and Korfhage (Reference Geyer and Korfhage2015) estimate that caregiving subsidies in Germany reduced the caregiver labour market participation, while supports exert a small positive effect. Similarly, Geyer and Korfhage (Reference Geyer and Korfhage2017) found that although the extension of LTC subsidies and supports did not shift the labour supply of women, it had a negative effect on male labout supply (−19.3 pp). Similarly, Korfhage (Reference Korfhage2019) reports that cash benefits encourage the provision of informal care and the exit from the labour market among lower income individuals.

Consistently, evidence from Japan points towards a negative effect of the provision of care on labour market participation, though it expands employment among caregivers (Shimizutani et al., Reference Shimizutani, Suzuki and Noguchi2008; Sugawara and Nakamura, Reference Sugawara and Nakamura2014). Probably, the closest study to ours is Fu et al. (Reference Fu, Noguchi, Kawamura, Takahashi and Tamiya2017), who examines the effect of the implementation of the introduction of LTC insurance in 2000 and the subsequent reform in 2006.Footnote ⁷ In contrast to the evidence form Germany, they find evidence of an expansion of male caregiver's labour supply and a reduction of labour market exit among female caregivers. Yet, the effect vanished completely after the reduction of benefits among recipients with mild care needs. So far, limited attention has been paid on the effect of caregiving subsidies on retirement intentions and behaviours. This is the main contribution of this paper.

4.1 Data

We use a sample for Spain obtained from five waves of the Survey of Health, Ageing and Retirement in Europe (SHARE) including Wave 1 (2004), Wave 2 (2006/2007), Wave 4 (2011), Wave 5 (2013), Wave 6 (2015) and Wave 7 (2017).Footnote ¹⁶ SHARE is the largest pan-European social science study on the effects of health, social, economic over the life-course of people aged 50 or older. The study began in 2004, and since then, 380,000 in-depth interviews from 28 European countries and Israel have been conducted.Footnote ¹⁷

Given that our sample is designed to contain enough information on both the receipt of LTC benefits and the caregiver's early retirement intentions, we have selected the sample as follows. First, from the initial sample containing the six cross-section waves (N = 28,814), we have identified all married or partnered individualsFootnote ¹⁸ (N = 20,699). Second, we have selected individuals who were employed (N = 3,337), discarding those who were self-employed, since early retirement is not available to them in Spain.Footnote ¹⁹ Third, we have considered only working individuals aged between 50 and 59 years, restricting the age range to those workers who still did not qualify for early retirement (N = 2,541).Footnote ²⁰

Given that the same individual may be interviewed in several waves, we allow for their economic activity to vary over time (e.g., going through periods of unemployment). For this reason, we have consistently checked which individuals are classified as ‘employed’ in several waves, because only for them we can measure their retirement intentions in the waves in which they are employed. To prevent these observations from distorting the sample, we have kept only those individuals who are consecutively employed in one or more waves (N = 2,499).

4.2 The empirical strategy

Our empirical strategy identifies the impact of SAAD related subsidies and supports on early retirement intentions of the caregiving partner or spouse. Our empirical strategy attempts to answer the following questions: (i) Do caregiving subsidies or home care supports influence early retirement intentions? (ii) How did the 2012 austerity cuts alter early retirement intentions? (iii) How do such estimates compare to the 2013 reform that incentivised early retirement among private sector employees?

Although the labour market effects of SAAD may be potentially heterogeneous, our focus is on the effect on retirement intentions, differentiating between subsidies and home care supports. The effect of each type of subsidy and support is far from trivial. Whilst the receipt of caregiving subsidies may increase the probability of intended retirement if subsidies are perceived as an extra household income, they might alternatively decrease retirement intentions, if the government pays additional social insurance contributions (which incentivises individuals to stay in the labour force). In the latter case, it is possible that spousal caregivers benefiting from home care supports may find it easier to reconcile work and care, and thus home care supports might discourage intended retirement as a result. Austerity cuts in subsidies and supports after 2012 might have also exerted a significant and counter impact on retirement intentions. The reduction in the subsidy and the suspension of the social insurance contributions arguably may revert the additional incentive to delay retirement intentions, unless earnings are perceived to be indispensable for the subsistence of the household. Similarly, the reduction home care supports may increase retirement intentions.

Given that access to early retirement is different among those employed in the public or private sector, we examine the effect of an early retirement reform for private sector employees after 1 January 2013, which significantly tightened the eligibility conditions and reduced the amount of pension that could be received. Caregivers working in the private sector were exposed to both SAAD budget cuts and the tightening of early retirement incentives. A unique contribution of this paper lies in examining the effect size of bothLTC and early retirement reforms on retirement intentions.

The literature examining the effect of informal caregiving on early retirement decisions has addressed the potential endogeneity of the supply of informal care resulting from reverse causality and omitted variable bias. Indeed, a search for a better work and life balance might incentivise caregivers to opt for early retirement, but early retirement might, in turn, make individuals more likely to become informal caregivers. Hence, it seems important to exploit the exogenous variation of a quasi-experimental reform such as SAAD in the probability of providing care. Furthermore, we examine early retirement intentions rather than ex-post retirement decisions, given that decisions might take some time to materialise, whilst intentions tend to adjust more immediately. We focus on the following question from the SHARE questionnaire: ‘Thinking about your present job, would you like to retire early from this job?’. This question is only addressed to respondents who are working at the time of the survey. In other words, we are not able to capture the early retirement intention of unemployed or inactive individuals.

Given that our sample refers to employed individuals between the age of 50 and 59 years old, it is reasonable to assume that individuals who respond affirmatively to a retirement intention question are clearly expressing their ‘true' early retirement intentions. We define the dependent variable ‘early retirement’ ER _it as a binary variable that takes the value 1 if it answers affirmatively to the previous question and 0 otherwise.

We exploit the introduction of SAAD, distinguishing two periods: (i) short-term effects: 2004–2011 (prior to the cuts in the SAAD and the reform of the retirement system) and (ii) long-term effects: 2004–2017 (which includes the two reforms mentioned above).

Short-term effects (2004–2011) are captured by the equation below:

(1)$$\eqalign{ER_{it} & = \beta _0 + ( {\beta_1 + \beta_2LTC_{it}} ) IC_{it} + [ {\beta_3 + \beta_4LTC_{it} + ( {\beta_5 + \beta_6LTC_{it}} ) IC_{it}} ] T_t^{7-11} \cr & \quad + \beta _7LTC_{it} + \beta _8UR_t + X^{\prime}_{it}\gamma + C_c + T_t + \varepsilon _{it}; \;\;LTC_{it} = \{ {CA_{it}, \;HC_{it}} \} .} $$

IC _it takes the value 1 if the interviewee is an informal caregiver of a partner/spouse in need of care LTC _it is a binary variable that refers to the reception of long-term care benefits. Since the care receiver cannot benefit from both the caregiving subsidy and home care supports simultaneously, we proceed to define two different binary variables, namely caregiving subsidies (CA _it) and public home care supports (HC _it).

$T_t^{7-11}$ is a binary variable that takes the value 1 in the post-reform period (2007–2011), and 0, otherwise $X^{\prime}_{it}$ denotes a vector of controls. Controls capture the effect of some alternative explanations for individual differences in early retirement intentions,Footnote ²⁶ such as (i) individual characteristics including age, sex, type of occupation,Footnote ²⁷ public/private sector employee income from employment, percentage of respondent's employment income as a share of total household income, number of consultations with doctor/nurse, having a retirement plan,Footnote ²⁸ (ii) characteristics of the partner/spouse (age, sex, state of dependence approximated by the Katz index, suffering from mental illness), and (iii) household characteristics (number of sons/daughters living at home and outside home, household wealth, household size, having mortgage) and the size of municipality. Finally, UR _t refers to the unemployment rate at the regional level (NUTS-2) to account for differences in regional labour markets. Finally, C _c and T _t denote regional fixed effects and year fixed effects, and ɛ_ict is an error term. We cluster our standard errors by region to capture heterogeneity in the implementation of the SAAD across regions.

We are interested in identifying (1), namely whether there exists an ‘added worker effect’, which is identified by β ₁ < 0, and (2) whether the introduction of SAAD changed early retirement intentions of caregivers. Specifically, we test whether caregiving subsidies (β ₆ < 0 for the model estimated with CA _it), or home care supports (β ₆ < 0 for the model estimated with HC _it) gave rise to a change of retirement intentions.

Long-term effects (2004–2017) are measured as follows:

Next, we examine not only the introduction of the SAAD but its subsequent austerity cuts, as well as a further early retirement reform among private sector workers. To identify private sector employees, we define a binary variable PR _it that takes the value 1 if the worker is employed in the private sector and 0 otherwise.

(2)$$\eqalign{ER_{it} & = \alpha _0 + \alpha _1LTC_{it} + ( {\alpha_2 + \alpha_3LTC_{it}} ) PR_{it} + [ {\alpha_4 + \alpha_5LTC_{it} + ( {\alpha_6 + \alpha_7LTC_{it}} ) IC_{it}} ] T_t^{7-11} \cr & \quad + [ {\alpha_8 + \alpha_9LTC_{it} + ( {\alpha_{10} + \alpha_{11}LTC_{it}} ) PR_{it} + ( {\alpha_{12} + \alpha_{13}LTC_{it} + \alpha_{14}PR_{it}} ) IC_{it}} ] T_t^{13-17} \cr & \quad + [ {\alpha_{15} + ( {\alpha_{16} + \alpha_{17}PR_{it}T_t^{13-17} } ) LTC_{it} + \alpha_{18}PR_{it}} ] IC_{it} + \alpha _{19}UR_t + {{X}^{\prime}}_{it}\delta + C_c + T_t + \epsilon _{it}; \;\cr LTC_{it} & = \{ {CA_{it}, \;HC_{it}} \} .} $$

We are interested in contrasting (i) the long-term ‘added work effect’ (α ₁₅ < 0), (ii) whether the cuts in SAAD had a significant effect on early retirement intentions, and (iii) whether such effects are different between public and private sector workers, given that the reform affected mainly early retirement of private sector workers (α ₁₇ ≠ 0 for CA _it and/or HC _it).

Assumptions of the model: To identify the effects of SAAD on early retirement intentions, we need to verify that two common assumptions: (i) stable unit treatment and (ii) common trends (Lechner, Reference Lechner2011). The stable unit treatment assumption states that the treatment should only affect the treatment group. That is, the control group should not be affected in any way by the interaction between members of the population, and the treatment group should not be affected by reforms taking place before the implementation period.Footnote ²⁹

Given that caregiving subsidies in SAAD can only be received under the assumption of co-residence between caregiver and care receiver, and that our sample is restricted to spouses (or cohabiting partners), it is unlikely that individuals strategically change their living arrangements to receive any caregiving benefit. In the robustness checks section, we show consistent evidence when we hold constant the composition of the household.

Similarly, we rule out that decisions regarding early retirement after the 2013 early retirement reform could have affected the equilibrium salary and, consequently, we have a stable control group. Notice that when we examine the effect of the early retirement reform, the control group consists of civil servants. Although these individuals have more advantageous conditions of early retirement, it does not seem plausible that private sector workers (over 50 years of age) switch to public employment to become eligible for early retirement.

The common trends assumption implies that both the treatment and control groups must be influenced by the same pre-trends. In other words, the potential non-treatment outcomes are expected to follow the same trend regardless of the group to which they belong. This assumption can be relaxed if the covariates give rise to different time trends that are identifiable. In such a case, the common trend assumption must be met conditional on these covariates.Footnote ³⁰ As the reform was presented to the parliament on 26 April 2006, and the legislative process of the reform was heavily amended in Parliament and it only was approved in the last month of the year, it was practically impossible to anticipate a specific outcome.Footnote ³¹ To further test the plausibility of the common time trends, we formally tested whether the treated individuals had deviating linear pre-trends by estimating the following specification:

(3)$$ER_{it} = \varphi _0LTC_{it}\cdot t + \varphi _1C_c\cdot t + \varphi _2UR_t + {X}^{\prime}_{it}\delta + C_c + T_t + \varepsilon _{it}$$

$$LTC_{it} = \{ {CA_{it}, \;HC_{it}} \} .$$

For both caregving subsidies and home care supports we find that the interaction φ₀LTC _it ⋅ t is not significant (results available upon request).

As for estimation method, in a non-linear model, the common trend assumption is only satisfied if there is no group-specific difference in the dependent variable. This implies that the treatment and control groups should start at the same average level of early retirement before treatment (Lechner, Reference Lechner2011). As this assumption is not met with our data, we have opted to estimate a linear model.

4.3 Robustness checks

To address potential endogeneity concerns, we proceed to estimate a first stage regression estimating the probability of being caregiver using as an instrument the number of informal caregivers outside the home. Instruments commonly used include health indicators of potential care receivers, which could be positively correlated with the demands for care; for example, parent's age (Bolin et al., Reference Bolin, Lindgren and Lundborg2008), parent's health (Bolin et al., Reference Bolin, Lindgren and Lundborg2008; Ciani, Reference Ciani2012; Van Houtven et al., Reference Van Houtven, Coe and Skira2013), parent's or parents-in-law living status (Bolin et al., Reference Bolin, Lindgren and Lundborg2008; Van Houtven et al., Reference Van Houtven, Coe and Skira2013), number of people who are sick or dependent in the home (Heitmueller, Reference Heitmueller2007; Ciani, Reference Ciani2012; Meng, Reference Meng2013). However, Hassink and Van der Berg (Reference Hassink and Van den Berg2011) cast doubts on this type of instruments as they ignore the fact that some care tasks are ‘time-bounded’ (e.g., eating, going to the toilet, getting in or out of bed), while others can be performed at different times of the day (e.g., bathing) or even moved from one day to another (e.g., shopping).

In this paper, we use the number of caregivers outside the household as an instrument, since help from non-co-residential caregivers could plausibly influence the spousal burden of care. An identification concern refers to whether receiving some benefit from the SAAD affects the composition of the household to reinforce the provision of care. For this reason, the models are re-estimated by restricting the sample to households whose composition has not changed throughout the period of analysis.

6.1 Short-term heterogeneity effects

Table B1 reports a re-estimation of the short-term model conditioned on different socio-demographic characteristics. Consistently, we find that the receipt of SAAD subsidies and supports in the post-reform period (2007–2011) reduces caregivers early retirement intentions, especially among women (−30.1pp for subsidies (CA); −24.9pp for supports (HC). We estimate the reduction in early retirement intention to be 121% higher for women than men when receiving CA and 74% higher when receiving HC.Footnote ³⁴

When we carry out a subsample analysis by employment sector, we find that our estimates are driven by blue-collar workers. That is, blue-collar workers exhibit a stronger early retirement intention once they become caregivers, but this intention decreases if they benefit from SAAD in the post-reform period (−14.7pp for subsidies (CA); −13.8pp for supports (HC)). In contrast, we find no significant effect for white-collar workers.

When we examine the heterogeneous effect by partner's mental health or disability, we document that being an informal caregiver increases early retirement intentins in 10–12pp among those with mental disorders and by 61pp when the Katz's index exceeds the value of 4. However, receiving a LTC benefit in the post-reform period reduces early retirement intention by 10pp in case of mental disorder or by 16–17pp in case of Katz's index exceeding the value of 4.

Next, we examine the heterogeneity of our estimates by household wealth. Estimates suggest that early retirement intention increases among households in the first (lowest) and second quartile, but such effect is smaller among households in the fourth quartile (highest).Footnote ³⁵ In contrast, being informal caregiver increases the probability of holding a retirement plan by around 25–26pp but receiving a LTC benefit in the post-reform period reduces such probability by approximately 60%. Having no mortgage increases the early retirement intention (22–23pp), although after SAAD, such intention is reduced by half.

Finally, when we examine the heterogeneity effects conditioned on the presence of children in the household we find that: (i) having no children increases the early retirement intention by 37.6pp among caregivers, but this intention declines in the post-reform period (by 40% among those receiving CA and by 36% among those with HC); (ii) having a co-resident daughter and co-resident son does not have any significant effect on the early retirement intention of the partner/spouse; (iii) the retirement intention decreases by 34.7pp if the caregiver receives CA and there is a co-resident daughter, or by 18.7pp if the caregiver receives HC and there is a co-resident son.

6.2 Long-term heterogeneity effects

Table B2 shows the results after re-estimating the long-term model considering some relevant sociodemographic characteristics. The austerity cuts in caregiving subsidies (CA) only reduced retirement intentions among women (−28.9pp), although the reduction in home supports (HC) increased retirement intentions among both genders (20.2pp for men and 10.6pp for women). For private sector workers, we find a reduction in early retirement intentions among those receiving CA (−12.1pp for men, −14.4 for women) or HC (−9pp for men and −13.6pp for women). These results imply that among private sector workers, the reduction in early retirement was higher among women (19% for CA and 51% for HC).Footnote ³⁶

When we compare the effect by income groups, we find that austerity cuts have discouraged early retirement among households in the lowest quartiles (−56.6pp for the 1st; −53.4pp for the 2nd) but increased early retirement among the most affluent households who received home care (+51.7pp). Similar effects are found among those individuals that have a retirement plan (−25.1pp for CA and +25.9pp for HC) or a mortgage (−20.9pp for CA and +25.1pp for HC). In contrast when we examine the effects by family composition, we find that the 2012 SAAD austerity cuts reduced retirement intentions among caregivers with no children receiving CA (−13.3pp), or HC with at least one co-resident daughter and son (−93.8pp). However, early retirement intentions increased for those receiving CA with no co-resident daughters (+81pp) or HC without children (+31.8pp).

Next, we compare the relative magnitude of the impact of the reform of the retirement for private sector workers compared to the SAAD cuts. This relative effect is 38% for men (−0.121 with respect to 0.313), and twice as large among women, (−0.144 with respect to −0.189). This result points to a long term effect among a specific household profile facing more difficulties to make ends meet, and where the caregiving subsidy was ‘welcomed’ as a supplementary household income (potentially shared with adult children) as discussed in Costa-Font and Vilaplana (Reference Costa-Font and Vilaplana-Prieto2017).

7.1 Treatment effect heterogeneity

De Chaisemartin and d'Haultfoeuille (Reference De Chaisemartin and d'Haultfoeuille2020) show that in a two-way fixed-effects design, the estimated effect is a weighted average of treatment effects in each group and time period, with weights that depend on group size and treatment variance. When the treatment effect is not constant over time and across groups, the estimated effect is biased. To overcome this issue, they propose a new estimator corresponding to the average treatment effect of all group-time cells whose treatment status changes between two consecutive time periods.

The estimator by De Chaisemartin and d'Haultfoeuille relies on two assumptions. The first one, which is a generalisation of the traditional common trend, requires that the mean outcome of groups having the same treatment status in wave t−1 would have the same trend between wave t−1 and wave t, in the absence of treatment. In our case, it implies that if no benefit had been awarded in a given year, early retirement intentions would have been similar in the treatment and control observations which were not treated in the preceding wave. The second assumption requires that, for each year, if one individual enters the treatment, then at least there is one observation that remains untreated. The first assumption is not testable, but it is weaker than the traditional common trend assumption. The second assumption is easy to check and holds as long as there is a sufficient number of benefits awarded each year. As a robustness check, we have used the De Chaisemartin and d'Haultfoeuille's estimator and the number of estimates with a negative weight. We find that only 0.61% of our estimates display a negative weight (the sum of the weights is −0.00019), which implies a very limited importance to the overall weighted average.

7.2 Stable household composition

As an additional robustness check, we have re-estimated the short and long-term models for the subsample of households whose composition has not changed. In the previous estimates, the set of explanatory variables included whether more than two people lived in the household, as well as the number of co-resident and non-co-resident sons and daughters. Now, we attempt to disentangle the effect of SAAD from other confounding effects related to living arrangements.Footnote ³⁷ Estimates are retrieved from a reduced sample size, given that households must have been interviewed several times and must not exhibit changes in its composition across different waves. The final sample for the short-term model consists of 230 observations (28.9% compared to the initial sample) and 837 observations for the long-term model (39.2% compared to the initial sample). Nevertheless, the results reported in the Table B3 are in line with those of Tables 1 and 4, both in magnitude and significance, consistent with prior estimates.

7.3 Endogeneity of informal caregiving

Testing whether the exclusion condition is satisfied is not possible, since the error term is unobservable, and must be conducted in a purely conjectural way (Roberts and Whited, Reference Roberts, Whited, Constantinides, Harris and Stulz2013). However, to assess the effect of violations of the exclusion restriction, we use the approach of van Kippersluis and Rietveld (Reference van Kippersluis and Rietveld2018), who extend the ‘plausibly exogenous’ method developed by Conley et al. (Reference Conley, Hansen and Rossi2012) and provides guidance on which prior to use in sensitivity checks. As van Kippersluis and Rietveld (Reference van Kippersluis and Rietveld2018) note, if there exists a sub-group of the sample for whom the first stage is zero, then the reduced form for this group can be used to back out a plausible estimate for the estimated coefficients of the instruments, instead of choosing arbitrary prior values. In our case, the sub-sample of individuals without children qualifies to estimate the direct effect of the instrumental variables on the dependent variable. Running the regression of the instrument, on the dependent variable, we find no significant effect, which increases our trust in our assumptions (results available upon request).

Table B4 shows the results of the first-stage regression for the probability of being informal caregiver using as instrument the number of non-co-resident informal caregivers. For the five different specifications, we find a positive association between supply of care and the probability of receiving help from other non-co-resident caregivers. We estimate that each additional non-co-resident informal caregiver increases the probability of spouse/partner providing informal by 4.4pp. In our specifications, the errors are also clustered by region. We use the traditional Hausman or Durbin–Wu–Hausman to test for endogeneity, and a variation of the Durbin–Wu–Hausman test that is robust to heteroskedastic and clustered errors (Cameron and Trivedi (Reference Cameron, Trivedi and Revised Edition2010) discuss the robust test (p. 190) as the ‘robustified’ Durbin–Wu–Hausman test).Footnote ³⁸

Similarly, we use the Kleibergen Paap's F-statistic to test for weak instruments, since we cannot formally test for weak instruments when errors are heteroskedastic, serially correlated or clustered (Pflueger and Wang, Reference Pflueger and Wang2015). Given that F-statistics may be high even under weak instruments, setting the confidence level to 5%, we compare the effective F test to the critical values under different values of τ (e.g., fraction of a ‘worst-case scenario’ situation in which the instruments are completely uninformative and first- and second-stage errors are perfectly correlated).Footnote ³⁹ The test rejects the null hypothesis of a weak instrument threshold of τ = 5%. These results show that the instrument is reasonably strong under all specifications.

7.4 Domestic care employment

Finally, in this section, we propose an additional mechanism which could drive our estimates: the effect of SAAD in the hire of domestic caregivers. If caregivers allocate part of the caregiving subsidies to hire a domestic carer (although this would be a misuse of the benefit), it might allow them to continue to participate in the labour force and delay their retirement. To test this effect, the short and long-term models are re-estimated using a binary variable that takes the value 1 if the household has hired a housekeeper, and 0 otherwise. The results are shown in Table B5.

Our results suggest that in the short-term, the probability of employment of a domestic carer increases by 11.7pp when the caregiver receives a caregiving subsidy, but we find no significant effect when home care supports are received. In the long-term, the 2012 budgetary cuts in the SAAD have led to a decrease in the probability of employing a domestic carer upon receiving a caregiving subsidy but suggest an increase after receiving home care supports (−6.0pp for CA; +6.1pp for home supports). Hence, these results provide suggestive evidence of some substitution between informal and formal caregiving, which might be offset by the role of domestic care employment.