Empowerment Effects and Intertemporal Commitment of Married Couples: Evidence from Japanese Pension Reform

Transcription

1 Empowerment Effects and Intertemporal Commitment of Married Couples: Evidence from Japanese Pension Reform Takahiro Toriyabe October 14, 2017 Abstract Following a 2007 pension reform, wives in Japan can now claim the pension records of their husbands upon divorce. This change has empowered wives relative to their husbands, but since the portion of the divisible records a wife can claim depends on employment status and length of marriage, the impact of the pension reform varies across households. Exploiting this exogenous variation in bargaining power that was unexpected for couples at the time of their marriage, we analyzed through a difference-in-differences approach the degree to which spouses could commit to their resource allocation plan. Estimation results suggest that spouses re-bargain their resource allocation when one member s options outside of marriage change significantly, and thus they cannot completely commit to the initial plan. This finding of incomplete commitment implies that household members face a hold-up problem, as they fail to achieve an ex ante efficiency. Keywords: family; bargaining power; risk sharing; pension; divorce JEL codes: D13, D91, H31, J12, K36 Doctoral student, The University of Tokyo, Hongo, Bunkyo-ku, Tokyo, Japan ttoriyabe@g.ecc.u-tokyo.ac.jp. 1

2 1 Introduction Economists have long been interested in the economic gains from marriage, from the joint use of public goods such as housing to the raising of children, specialization according to each spouse s comparative advantage, and self-insurance of household members through the process of family formation. This latter process of self-insurance has important implications, as Krebs (2007) points to the serious costs of business cycles when job displacement risks are not completely insured. While conventional unemployment insurance is one solution, an individual can also hedge these risks by marrying a partner whose income is negatively correlated with his/her own income, though this specific form of insurance through family formation does not seem salient in practice (Altonji et al and Hayashi et al. 1996). In addition to income risks, an individual could insure longevity risks through heritage from his/her partner (Kotlikoff and Spivak 1981). The validity of these arguments regarding gains from marriage, however, relies on the ability of marriage partners to commit to their initial resource allocation plan, given that their incentives potentially conflict. The traditional unitary model of a household in economics assumes there is no conflict between spouses by regarding a household as a single decision maker. Consequently, as any potential form of conflict is eliminated, the above family functions seem to work well in such cases. However, when the partners are allowed to have distinct preferences and incentives, a lack of commitment can prevent each member of the household from enjoying the fruits of marriage due to a hold-up problem, which occurs when they fail to achieve ex ante efficient outcomes. This is supported by studies such as Voena (2015), who finds that the dynamics of asset accumulation are affected by the divorce laws in the U.S., and Dufwenberg (2002), who shows that, in theory, a couple can fail to specialize and thereby under-accumulate human capital when the household members cannot commit in advance to their future behavior. These studies imply that a couple, without such commitment, could have difficulty appropriately sharing risks as well as childbearing and childrearing efficiently. In other words, in a dynamic model, a household s initial resource allocation can be inefficient, and the economic roles within a family depend on the degree of commitment of the family members. As noted by Chiappori (1988), the traditional unitary model of the household is inconsistent with individual welfare maximization and, consequently, many studies that analyze household behavior now use the collective model which captures the friction of individual interests and incentives by rec- 2

3 ognizing the bargaining positions of household members (Chiappori, 1992), and a model is further developed by Apps and Rees (1997), Chiappori (1997) and Blundell et al. (2005) to address household production and children. In the collective model, household members first determine the level of domestic production and household wealth allocation according to their bargaining position. Each individual then allocates his/her share of the wealth to maximize his/her own welfare. A main difference between the collective and unitary models is thus the role of bargaining among the household members. In the unitary model, bargaining does not occur, while in the collective model, household members allocate their wealth by bargaining. Mazzocco (2007), developing the collective model to study household intertemporal decision making, finds that household members may or may not commit to their resource allocation plan because they have incentives to deviate from the initial plan after uncertainty at the time of marriage is gradually resolved. In order to achieve complete commitment, the bargaining position of spouses is fixed at the time of marriage so as not to deviate from their initial state-contingent allocation plan. In this situation, within-household bargaining position is time-invariant and all the variation in bargaining position is between households. Under incomplete commitment, household members update their bargaining positions according to new information, which allows them to possibly deviate from their initial resource allocation plan. In this situation, we thus observe within-household variation in the bargaining position. The degree of commitment, therefore, can be identified by separating within-household and between-household variation in bargaining positions, and this test is available when there is variation in the relative bargaining power which cannot be anticipated during the initial period. In this paper, this is our main identification strategy. While it is difficult to directly observe the bargaining positions of household members, Browning and Chiappori (1998) argue that a distribution factor, a variable that does not affect the household budget set or preferences but affects the bargaining power, facilitates identification of the mode of household decision-making. Relative age, relative educational attainment, relative wage, relative income and local sex ratio are often used as distribution factor variables (Browning et al. 2014) and, although Aronsson et al. (2001) is an exception, studies using these distribution factor variables have typically rejected the static unitary model (Angrist 2002, Attanasio and Lechene 2014, Aura 2005, Cherchye et al. 2012, Duflo 2003, Francis 2011, Ponczek 2011). 3

4 However, while static studies of household decision-making have coalesced around the collective model, the results of the relatively few empirical studies of dynamic household decision-making are less conclusive. For example, Mazzocco (2007) rejects full commitment, but his test relies on the questionable assumption of separability between consumption and leisure in the utility function. Lise and Yamada (2014), who also reject full commitment, suggest that household members update their bargaining position only when the participation condition (or incentive compatibility) for marriage is binding, but this study still has several limitations. First, households are excluded from the sample if the wages of both spouses are not observed for several consecutive years, which could cause sample selection bias. Second, unexpected deviation from the relative wage path of spouses is chosen as a distribution factor, but this choice is questionable, as it is conceivable that a deviation unexpected by researchers could readily be anticipated by household members themselves. Third, relative wage, which is also chosen as a distribution factor in the study, may be endogenously determined through labor market participation. On the other hand, studies that suggest incomplete commitment by using quasi-experimental situations also have concerns. For example, Voena (2015) finds evidence of incomplete commitment in her study of the impact of divorce laws in the United States on the labor force participation and savings of marital households. However, since within-household changes in specific time allocations are not investigated (such as hours worked in the market and domestic sectors), the magnitude of the impact on the time allocation of spouses is speculative 1. Blau and Goodstein (2016), in their test of whether or not the source of an inheritance matters, also reject the full commitment model, as they find that any unexpected inheritance decreases the labor force participation of its recipient while not affecting that of his/her partner. Although their result is supported by a series of robustness checks, there is a potential concern about their key variable, the inheritance from parents-in-law, which is derived from the Health and Retirement Study (HRS). Laitner and Sonnega (2010) argue that HRS respondents tend to substantially under-report any transfers from parents in-law both in frequency and amount, which leads us to suspect that the estimation results of Blau and Goodstein (2016) suffer from serious non-classical measurement error problems. In such cases, the observed response to the partner s inheritance could be smaller than the response to one s own inheritance even in the absence of 1 Although Voena (2015) does analyze cross-sectional variation in time-allocation, it is difficult to derive intertemporal household decision-making decisions from cross-sectional analysis. 4

5 any re-bargaining issues, and so this leads to questions as to how to interpret their empirical evidence. Taking into consideration all of the above, further research of intertemporal household behavior is needed, and the current study aims to contribute to this literature by exploiting the natural experiment of a 2007 Japanese pension reform to examine intertemporal household decision-making. Our study examined the mode of household decision-making by focusing on the degree of commitment to an ex ante efficient allocation plan. The Japanese pension reform of 2007 provides exogenous variation in the bargaining position between spouses that would be unlikely to have been anticipated at the time of marriage. In particular, the pension reform now allows a dependent wife to claim a fraction of the pension records of her husband if she were to divorce, improving her economic situation should she divorce. Moreover, even if she does not divorce, her improved level of welfare upon divorce would still create a better bargaining position during the marriage. Due to its design, this pension reform has been most relevant to households in which the husband is a regular employee and the wife is a part-time worker or homemaker, but is less relevant to other types of households. As the reform did not change the total amount of pension benefits a household would receive, and because couples married before the reform was approved in 2004 were unlikely to have anticipated it, the pension reform enables us to test the degree of commitment by the spouses. Ours is not the first study to use the 2007 pension reform as a natural experiment. Sakamoto (2008), in his study of the reform, finds the policy impact to be small, but the study has several limitations. First, the study does not focus on the households presumably most affected by the reform, for while the reform is known to have had a greater impact on elderly households, the study sample consists of relatively young wives (48 years old, at most, in 2007). As a result, the policy effect is likely to have been underestimated. Second, the choice of dependent variables is questionable. The policy impact is estimated on the reduced forms of the ratio of savings, consumption and leisure between spouses, which makes interpretation of the estimates difficult, for when the relative bargaining power of household members changes, it is conceivable that the leisure of both the husband and wife, for example, could move in the same direction. Should this occur, the ratio of leisure could be unchanged even when the time allocated to leisure changes for both spouses. Finally, the Sakamoto (2008) study estimates a static household decision model and therefore does not address issues associated with dynamic household decision-making. 5

6 In order to exploit the exogenous variation in bargaining power associated with the Japanese pension reform, we employed difference-in-differences (DD) estimation. DD estimation requires that changes over time in an outcome variable within a control group correspond to counterfactual changes within a treatment group, in absence of treatment. If this assumption is satisfied, then the comparison of changes within a control group and within a treatment group reveals the causal impact of the policy. Put differently, we identified the average treatment effect on treated (ATT) as long as treatment and control groups faced the same time trends. Appropriate construction of the treatment and control groups was crucial for identification and, considering the characteristics of the pension reform, we assigned households in which the husband was a regular employee and the wife was not to the treatment group and all other households to the control group. We then implemented DD estimation for each age cohort of the wife so as to capture the heterogeneous impact of the policy. For our dataset, we used the Keio Household Panel Survey (KHPS), a representative Japanese survey that provides rich information about household behavior and background variables, which enabled us to analyze intertemporal household behavior. We found that in households in which the wife was aged 50 59, her empowerment due to the pension reform increased her allocation to leisure while decreasing her market and domestic labor supply, causing us to reject both the full commitment collective model and the unitary model, which is a special case, as it does not allow bargaining. Furthermore, those who were mainly affected were wives without positive housing values, which may represent the wife s value outside of her current marriage. Thus, this result implies that a wife with low outside options benefited most from the pension division system through re-bargaining, which is consistent with the model s prediction. By contrast, the impact of the reform on a wife under 49 was not significant at baseline, which may suggest that spouses do not re-bargain their allocation plan when a shock to their options outside of marriage is relatively small. We also found that the impact of the pension reform on divorce seemed to be marginal. These findings, which are underpinned by a series of robustness checks, have several implications. First of all, targeting a particular member can be more effective policy design than targeting a household, to take into account any behavioral responses associated with a change in the relative bargaining power of spouses. At the same time, policy can potentially have an unintentional impact on household 6

7 behavior due to the re-bargaining effect. Second, the incomplete commitment we found implies that ex ante efficient allocation is not necessarily ex post efficient and, as a result of this inconsistency, a hold-up problem may occur. This suggests that household behavior such as human capital accumulation and expenditure on public goods can possibly be inefficient within an intertemporal dynamic model, and the lack of full commitment also makes risk-sharing among household members difficult. Future research will address these specific issues associated with the hold-up problem. The remainder of the paper is structured as follows: Section 2 describes the model of household decision-making, while section 3 provides institutional background and section 4 analyzes the impact of the reform on divorce. Section 5 describes the dataset and our identification strategy, and sections 6 through 8 discuss our empirical result and its interpretation. Section 9 conducts subsample analysis using home ownership status, and sections 10 and 11 provide robustness checks to confirm our main findings. 2 Model Our study broadly follows the model of Mazzocco (2007). Consider a household with two adults, f and m, where f denotes a wife and m a husband. At each period s, member j spends his/her time on market labor supply, m j s, domestic labor supply, h j s, and leisure, l j s, with total disposable time normalized to one. We consider two commodities, a Hicksian composite market good, c, and a domestic production good, q. We assume that the market good is private and the domestic good is public, representing child welfare or perhaps the neatness of a couple s home. Domestic production technology is summarized by the production function F(g s,h m s,h f s ), where g s is the market good used for that production. The constraint for the domestic production is F(g s,h m s,h f s ) = q s, and the household budget constraint is { c j s + w j ( s l j s + hs)} j + gs + a s+1 = ws f + w m s + (1 + r s )a s (s = 0,...,T ) (1) j { f,m} where c j s is j s consumption at period s, l j s is leisure, h j s is domestic labor supply, g s is domestic production inputs, w j s is the wage rate, a s+1 is savings, r s is the risk-free interest rate, with the price of the market good normalized to one. 7

8 Denoting an individual utility function as u i (c i s,l i s,q s ), the utility maximization problem of this household at period t is thus max ( m m s,m f s,h m s,h f s,l m s,l f s,c m s,c f s,g s,q s ) T s=t E t [ T s=t β s t{ µ t su f (c f s,l f s,q s ) + (1 µ t s)u m (c m s,l m s,q s )} ], (2) subject to the constraints described above, where µ s t = µ s (y 0, y t, z 0, z t ), y t is the vector of information about the household budget available at period t, and z t is the vector of distribution factors, variables that do not affect the household budget or the preferences of the members but affect their bargaining power. Note that y t and z t may include information about past and future household budgets as expected values at period t. The Pareto weight, µ s( ) t [0,1], represents the relative bargaining position of the spouses. Hence, the objective function of the household problem gives more weight to a member with a better bargaining position, resulting in resource allocation in favor of that member. As suggested by Browning and Chiappori (1998), a distribution factor is useful in empirical studies as a proxy for the relative bargaining position of the spouses since it shifts the bargaining position without changing the household budget or preferences. The relative income of household members is often used as a distribution factor, since it does not affect the household budget set, holding the total household income constant, while it can possibly affect the bargaining position of each member. Relative income may, however, be correlated with individual preferences as well as market and household production technologies, so this variable is valid only if it is determined exogenously. As there is no guiding theory to determine whether a certain variable is a valid distribution factor or not, we need to carefully choose distribution factor variables in empirical analysis. The main implication from Mazzocco (2007) is that the spouses may or may not remain committed to their initial allocation plan. Under full commitment, the Pareto weight is fixed at the initial period, µ t s = µ s (y 0, z 0 ), (3) 8

9 and the solution to the maximization problem is m j s = m j s(y 0, y t, z 0 ), h j s = h j s(y 0, y t, z 0 ), l j s = l j s (y 0, y t, z 0 ), c j s = c j s(y 0, y t, z 0 ), (4) g s = g s (y 0, y t, z 0 ), q s = q s (y 0, y t, z 0 ), ( j { f,m};s {t,...,t }). (5) From the above, we can see that new information revealed later does not affect the bargaining position of spouses. In this case, ex ante efficiency implies ex post efficiency, and the state-contingent allocation plan of the household is constant regardless of any ongoing resolution of uncertainty over time. In this sense, we say that the spouses completely commit to their allocation plan. The unitary model is a special case of the full commitment case. Since it has no room for any conflict in incentives between spouses, the Pareto weight, µ s ( ), is unaffected by any distribution factors so that µ s t = µ s (y 0, z 0 ) = µ s. The typical interpretation of the unitary model is that the resource allocation is determined by a paternal member who is altruistic enough to consider the welfare of other members. In this case, the utility maximization problem of the household coincides with that of the paternal member. Under incomplete commitment, the bargaining position of the spouses is updated along with any new information, y t and z t. It follows then that the Pareto weight is µ t s = µ s (y 0, y t, z 0, z t ) (6) while the solution to the problem is m j s = m j s(y 0, y t, z 0, z t ), h j s = h j s(y 0, y t, z 0, z t ), l j s = l j s (y 0, y t, z 0, z t ), c j s = c j s(y 0, y t, z 0, z t ), (7) g s = g s (y 0, y t, z 0, z t ), q s = q s (y 0, y t, z 0, z t ), ( j { f,m};s {t,...,t }). (8) The spouses in this case update their bargaining positions according to any unanticipated changes in a distribution factor or option outside of marriage, and thereby re-bargain their resource allocation, thus deviating from the initial state-contingent plan. In short, in such a case in which there is less than full commitment within a household, ex ante efficiency does not necessarily imply ex post efficiency. Incomplete commitment does not imply no commitment, however. Consider two possible meth- 9

10 ods to update the spouses bargaining positions. One method is for spouses to always update their bargaining positions given new information related to the bargaining process. Namely, this corresponds to a household in which spouses re-bargain their resource allocation at each period after the realization of the state. As a result, the household seems likely to deviate from the ex ante efficient allocation, resulting in no degree of commitment. Another method is for spouses to update their bargaining positions only when the participation (or incentive compatibility) condition for marriage is binding. In this case, the spouses do not respond to a small shock, but when facing a shock large enough for one member to prefer divorce given the current resource allocation, they adjust the allocation so that both members are willing to stay married. (If the adjustment is not possible, they will divorce.) In other words, the spouses stick to the initial allocation plan as long as it makes both of them better off than the situation after divorce, and thus they commit to some incomplete extent. These observations enable us to identify the degree of commitment. Under the full commitment and unitary models, the vector of distribution factors at period t, z t, has no impact on household behavior. On the other hand, when the degree of commitment is limited, unexpected changes in distribution factors affect household behavior through an adjustment in the bargaining position. Thus, if there exists intra-household variation in a distribution factor which cannot be anticipated by the household members at the time of marriage, we can test the full commitment model by investigating how the spouses respond to that event. If they alter their behavior, the full-commitment and unitary models are rejected. Additionally, if there is variation in the magnitude of the unexpected shocks to the bargaining power, then we can empirically infer how the bargaining position has shifted. Specifically, if relatively small shocks to the bargaining power do not influence a household while relatively large shocks do, it implies that the household is committed to its allocation plan to some extent. In such cases, the spouses completely insure a small shock, but when they face a shock large enough for one member to consider divorce, they decide to re-bargain so as to enjoy the fruits of marriage. Finally, note that one may also consider that the husband and wife could make a contract on the distribution factors and, in such a case, the household consumption plan could be a function of the distribution factors even if the spouses are fully committed to their initial contract. However, this class of models is not testable unless we are able to directly observe the initial contract and all relevant distribution factors. Otherwise, any behavior could be attributed to unobserved variables. Alternatively, it 10

11 is possible that the spouses could make a contract to re-bargain in a certain situation and completely commit to that contract. As this situation is observationally equivalent to cases we have designated incomplete commitment, our empirical research could not exclude such possibilities. 3 Institutional Background 3.1 Japanese Pension System and the Reform of 2007 In order to better understand the context of the study, this section describes the Japanese public pension system, the reform of 2007, and a recent revision of the mandatory retirement system. The Japanese public pension system consists of three insurance policies, the Employee pension (kosei nenkin), the Mutual Aid pension (kyosai nenkin) and the National pension (kokumin nenkin). The first two policies cover regular employees 2, with the Employee pension insurance covering employees of private corporations and the Mutual Aid pension covering government employees. Since these two insurance policies are otherwise identical, we hereafter refer to them collectively as the Employee pension insurance. Within the Employee pension insurance, participants aged under 70 pay pension premiums as long as they earn labor income, with the amount of the premium proportional to their labor income. The age of eligibility for benefits is around 60, depending on sex and birth cohort, with the age of eligibility for men higher than women and for a recent cohort higher than an earlier cohort (Table 1). Employee pension benefits consist of two parts: a basic part and a proportional part, with the basic part depending only on the number of years for which the participant paid premiums, and the proportional part proportional to the earnings before retirement and premium payment duration. [Table 1 about here.] National pension insurance covers those who are not covered by Employee pension insurance, consisting mainly of the self-employed, part-time employees, and dependent wives, all of whom 2 In Japan, a typical regular employee is a full-time worker with an indefinite term (i.e. without a definite contract termination date). In addition to regular employees, the pension plans cover part-time employees who work more than three-quarters of the hours per day and days per month worked by a full-time employee. For example, if a full-time employee works eight hours per day and twenty days per month, then a part-time employee working more than six hours per day and fifteen days per month participates in either the Employee or Mutual Aid pension. 11

12 pay premiums until they turn 60 and become eligible for benefits at age 65. The National pension insurance is similar to the basic component of the Employee pension insurance, except for the age of eligibility and the participants in each respective plan. Before the pension reform of 2007, it was said that the pension system caused inequality within elderly households, as the wife was not allowed to claim any fraction of the pension benefits of her husband should they divorce. As a result, prior to the reform, the only benefits a housewife could receive after divorce were the National pension insurance benefits. According to the Ministry of Health, Labour and Welfare, the National pension plan paid out about 54 thousand yen (about 540 U.S. dollars) per month on average in 2007, while the Employee pension insurance paid out about 161 thousand yen (about 1,610 U.S. dollars) 3, making it difficult for a wife to live only on her National pension benefits. Due to this economic dependence, even if she were not satisfied in her marriage, she would have no choice but to remain married. Moreover, the collective model of household behavior predicts that her low welfare level after divorce would weaken her bargaining power, resulting potentially in poor treatment in the household even during marriage. In order to address this issue, pension reform on the division of a household s Employee pension records after divorce was approved in 2004 and initiated on April 1, This reform now allows spouses who divorce to divide their Employee pension records tracked during the marital period 4. Although a proportional division rate is determined by agreement between spouses, the spouse with the lesser pension record is not permitted to claim more than half of the total household pension record tracked during the marital period. Thus, if the Employee pension record amount for the wife is the same as that of her husband, there is no room for further pension division, and if the wife has no Employee pension records, the maximum proportional division rate is 50 percent. If the spouses fail to reach an agreement on the proportional division of records, a rate is provided by the courts. According to the Supreme Court of Japan, the courts decided on a 50 percent division rate in 99 percent of the cases in Since a divorced housewife is now assured of some income, her situation post-divorce is expected to have improved. Further, since the reform has provided her with a better bargaining position during marriage, it is reasonable to expect that the household resource 3 Data Source: 4 In particular, this pension division system applies to proportional part of the Employee pension records but not to the basic part. 5 Data Source: 12

13 allocation has also shifted in her favor 6. Several features of the 2007 pension reform help to facilitate our analysis. First, as the reform does not change the total amount of pension benefits received by a household, it does not affect the household budget unless the spouses divorce. Second, as discussed earlier, the reform presumably empowers a dependent wife by improving her economic situation after divorce, so the reform can be used to identify the mode of household decision-making. Finally, the reform of the Employee pension system leaves some households unaffected. This policy design creates treatment and control households, which allows us to make counterfactual inferences. Specifically, as the pension reform applies only to Employee pension records, the self-employed, who are covered by National pension insurance but not by Employee pension insurance, are unaffected by the reform. Additionally, when both spouses are regular employees, they each have their own Employee pension record, and so the pension splitting opportunities allowed by the pension reform does not substantially affect their relative bargaining power. Hereafter, we refer to this type of household in which both spouses are regular employees a dual-regular household. Dual-regular and self-employed households are unaffected by the reform and were included in the control group. Note that a dual-regular household excludes those households in which the husband is a regular employee while the wife works part-time. Since most part-time employees do not participate in the Employee pension insurance system, that type of household was included in the treatment group 7. To sum up, households with a self-employed spouse and dual-regular households were used to control for economic trends in absence of the reform. In addition to the employment status of the household members, the age of the wife can also be used to determine treatment and control households. First, because of the time value of money and the discounting to the present value of future pension benefits, a very young wife would be indifferent as to her pension benefits upon divorce. We show this throught the following example. Suppose that the interest rate, r, is 0.01, the wife s age of eligibility for pension benefits is 60, she marries at 28 6 A supplementary pension reform was implemented in 2008, which allows a dependent wife (or husband) to claim a half of the total household Employee pension records tracked after 1 st May 2008, if they were to divorce. This reform applies to households where a wife s annual income is less than 1.3 million yen (about 13,000 U.S. dollars) and that the proportional ratio is 50 percent, regardless of any agreement between spouses. However, considering that the first pension reform applies to the entire Employee pension records tracked during the marital period, the function of this second reform is, at most, supplementary, as it does not apply to pension records before 1 st May Consequently, as the second reform allows a couple to divide only a fraction of their household pension records, its impact on household behavior is likely to be negligible. For this reason, our analysis focused only on the impact of the first reform. 7 Even if a wife working part-time is covered by the Employee pension insurance, the amount of her records is likely to be much less than the amount of her husband because her earnings are presumably less than her husbands. 13

14 and dies at 86, reflecting the average lifespan of women in Japan 8, and the pension benefits divisible from her husband are proportional to the duration of the marriage. This last assumption reflects the fact that the pension division system does not apply to pension records tracked before marriage. With these assumptions, and normalizing the present value of divisible pension benefits for a wife aged 60 to one, we denote the present value for a wife aged T in 2007 as PV T = 26 t=0 ( ) (60 T )+t 1 T r (9) Comparing the present values of the future pension benefits for an elderly wife and a younger wife highlights how little a younger wife benefits from the pension reform. Figure 1, which shows the normalized present value of divisible pension benefits for a wife of a certain age, suggests that the present value grows almost linearly. The present value for a wife aged 40 is about 30 percent of the present value for a wife aged 60, and the present value for a wife aged 50 is about 60 percent. As a result, the impact of the pension reform on young households appears less striking than the impact on elderly households. In order to grasp the magnitude of the divisible pension benefits, we assume that the annual value of the divisible pension benefits for a wife aged 60 is 480 thousand yen (about 4,800 U.S. dollars), which was the average amount among those who were eligible for benefits in 2014 (i.e., wives who were older than 53 when the pension reform was enacted), according to the Ministry of Health, Labour and Welfare 9. The present value of benefits for a wife aged 60 is about 11.4 million yen (about 114,000 U.S. dollars), whereas the present value of benefits for a wife aged 50 is about 7.1 million yen (about 71,000 thousand U.S. dollars), and for a wife aged 40, only about 3.5 million yen (about 35,000 U.S. dollars). Note that these are not the present values of annual benefits but of the lifetime benefits. As the present value of future benefits seems too small for a young wife to count on to live independently, her welfare level upon divorce would seem to be largely unaffected by the pension division, indicating that the reform has had, at most, a marginal impact on young households. [Figure 1 about here.] It is worth noting that since the National Pension Act generally prohibits people from counting 8 Data Source: and mhlw.go.jp/toukei/saikin/hw/life/life07/index.html. 9 Data Source: 14

15 future pension benefits as collateral 10, a wife s economic situation after a divorce depends on whether or not she can receive pension benefits divided from her husband soon after the divorce. If she can, her relatively better economic situation upon divorce improves her bargaining power, but if she cannot, she will need other sources of income until she becomes eligible to receive pension benefits and, consequently, this continuing economic dependence causes her bargaining power to remain weak. To sum up the foregoing discussion, the pension reform has a greater impact on elderly households than on young households. 3.2 The Japan Elderly Employment Stabilization Law In this section, we describe the Elderly Employment Stabilization Law (EESL), which affects the mandatory retirement system in Japan. In 2006, the Japanese government, through the enactment of the EESL, required firms to attempt to raise the mandatory retirement age from 60 to 65 in order to fill the gap between the mandatory retirement age of 60 and the pension eligibility age of 65. Specifically, the EESL allows those born in 1946 to remain employed until age 63, those born in 1947 or 1948 until age 64, and those born in 1949 or later until age 65. Under the law, firms were not forced to raise the mandatory retirement age, but were required to follow one of three alternatives to continue to employ workers who would otherwise retire: (1) to raise the mandatory retirement age; (2) to extend or renew employment contracts; or (3) to abolish mandatory retirement. According to Ministry of Health, Labour and Welfare, most firms chose employment expansion or re-employment 11. Furthermore, most firms confirmed that workers were employed until they turned 65 through employment expansion or re-employment even if they were born between 1946 and The EESL is expected to influence the labor force participation rate of elderly people hired as full-time employees and, indeed, Kondo and Shigeoka (2017) show that the EESL has increased the ratio of salaried workers over 60. Those born in 1946 became 60 years old in Further, a main target of the EESL was full-time employees. Meanwhile, the pension reform of our interest was enacted in 2007, and presumably affected the same type of employees. Due to the similarity of the target populations, it seems difficult to estimate the effects of the pension reform on those who were 10 Exceptions to this include borrowing against future benefits through the Welfare and Medical Service Agency and the Japan Finance Corporation. For more detail, see and 11 Data Source: 15

16 targeted by the EESL. Therefore, our analysis focuses on households that were not directly affected by the EESL; that is, those who were between 30 and 59 years old. Another potential concern is that the EESL may indirectly affect younger employees. Since the EESL requires firms to continue to employ workers who would otherwise retire at age 60, it is conceivable that firms might accommodate this new requirement by decreasing the number of young employees, either by not hiring new graduates or by firing part-time employees. Conversely, if elderly workers and young workers are viewed as complementary, then firms could increase the number of young employees as a result of EESL. Ohta (2012) and Kondo (2016) report negative correlations between the proportions of employees aged over 60 and of female part-time workers, potentially suggesting that the EESL has created a crowding-out effect. However, since our control group includes part-time employees, such an indirect impact of EESL is controlled by our DD estimation to a certain degree. Furthermore, as long as the EESL affects young and old female part-time employees in a similar manner, DDD estimation (described later) partials out any potential indirect impact of the EESL so that any potential biases caused by the EESL law, if any, are likely to be small. 4 Divorce Also relevant to this study is the potential relationship between divorce laws and divorce rates, as a pension reform that improves the bargaining power of a dependent wife could potentially lead to an increase in the propensity to divorce. According to the Becker-Coase argument, which is a direct application of the Coase theorem, divorce laws do not affect household decisions to divorce because spouses choose the alternative that maximizes their total household resources and thus they always agree to either remain married or to divorce. However, Chiappori et al. (2007) show that this argument relies on unrealistic assumptions, including that household members are able to transfer their utilities before and after divorce, and also that private and public goods are consumed in the same way regardless of marital status. This latter condition requires, for example, that if spouses live together before divorce, then they continue to live together after divorce, a situation that rarely occurs in reality. From these considerations, Chiappori et al. (2007) conclude that, in theory, divorce laws are indeed likely to affect divorce rates. In consideration of this theoretical argument, before investigating the impact of the 2007 pension 16

17 reform on household resource allocation, we looked at its impact on divorce rates, finding that the impact was moderate at best. However, since we exploited the panel structure of our main sample to analyze intertemporal household behavior, any potential systematic sample attrition could possibly result in bias. Indeed, before its enactment, the pension reform was expected to increase the number of divorces, as it would improve the economic situation of dependent wives upon divorce. If the pension reform did substantially affect the divorce decision of couples, we would need to cope with any self-selection problems associated with divorce. We thus checked for this using Japanese Vital Statistics data, and found that the pension reform had only a small impact on divorce. Exploring this further, the pension reform could have had at least two possible effects on divorce, both of which would be likely to increase the number of divorces in First, the reform could induce a couple to divorce who would not have divorced in the absence of the reform. Second, the reform could induce a couple to adjust the timing of an impending divorce. While the husband would have an incentive to divorce before April 1, 2007 to avoid losing a portion of his pension to spousal division, the wife would have an incentive to divorce after that date in order to benefit from the division. Despite these conflicting incentives, it appears that any adjustment in the timing of a divorce would have been achieved in favor of the wife because the Japanese divorce law requires mutual consent. Unless the wife is shown to be responsible for the divorce (due to an affair, for instance), she can delay any divorce by refusing to sign the divorce registration. Consequently, one would expect that the number of divorces in 2007 would have increased, or at least not have decreased. Figure 2a shows the number of divorces in Japan, and we can see a declining long-term trend. If the pension reform increased the number of divorces, the slope would become flatter after the reform, while if couples adjusted the timing of divorce, the number of divorces in 2006 or before would decrease while the number in 2007 would increase, so the direction of the effect is the same in both cases. Although the slope seems to have become flatter around 2007, this does not necessarily imply an increase in the number of divorces, however, as this declining tendency towards divorce began before the pension reform. [Figure 2 about here.] To test this statistically, we estimated the following equation, using monthly Vital Statistics data 17

18 between 2005 and 2008: Divorce = α 0 + α 1 Year A fter + α 2 Year + 12 Month j + u, (10) j=2 where Divorce is the number of divorces each month, Month j is monthly fixed effects, and A fter takes one after the reform and is otherwise zero. In this equation, α 2 represents the yearly trend and α 1 allows the trend to be different before and after the reform. It follows that if the pension reform increased the number of divorces, we would observe that α 1 > 0. However, we found that ˆα 1 = 0.29 and was not statistically different from zero at any standard significance level. Thus, we did not find a substantial increase in the number of divorces, though we recognize that this crude analysis is not a precise estimate of the policy impact on divorce. Figure 2b shows the number of divorces each month in Japan between 2005 and A striking feature of this figure is that the number of divorces soars each March before declining sharply to its typical level of between 20,000 and 22,000 cases each April. This rise in divorces in March is likely an effect of the Japanese fiscal year, which runs from April 1 to March 31. Since the school year begins and many people begin their new jobs in April, March is typically a time when families prepare for changes in their living conditions in the new year. It is thus reasonable to assume that spouses may also see March as an appropriate time to divorce. Nevertheless, this tendency for divorces to surge in March was not observed in Instead, while the number of divorces did increase in March, it remained high in April and May after the pension reform was enacted. Therefore, the pension reform does seem to have had an immediate impact by increasing the number of divorces in April and May However, the overall number of divorces in 2007 did not appear to increase, though this effect could possibly be countered by the decreasing long term trend. We also found that the number of divorces between January and March was smallest in 2007 as compared to other years between 2000 and Consequently, from this analysis, we suspect that the pension reform mainly affected households that would have divorced even if the reform was not enacted, by inducing them to delay the timing of the divorce so that the wife could benefit from the pension division system. The overall impact of the reform on the total number of divorces in Japan thus appears to be limited. Our main concern is with any possible sample attrition bias that might be associated with divorce, for if many households decided to divorce due to the pension reform, the households within our 18

19 sample that have not been greatly affected by the reform would cause us to fail to consistently estimate the policy impact, with the estimated impact attenuated. However, it turns out that the reform impact was limited and possessed a predictable directionality. Furthermore, as we found (shown later) that our sample households did not show any significant increase in divorce in 2007, we believe that any potential sample attrition bias is negligible, though we recognize that more sophisticated research is needed to precisely estimate the policy impact. 5 Data and Identification Strategy 5.1 Data This study utilized data from the Keio Household Panel Survey (KHPS) provided by the Keio University Panel Data Research Center. KHPS is an annual panel survey of households beginning in 2004 and consisting of 4,000 households (3,000 married and 1,000 single households), with both spouses asked to respond. Each year, the survey is conducted at the end of January, and respondents are asked about their usual time allocation as well as background information such as age, sex and family composition. As Japan s fiscal year begins in April, each implementation of the survey inquires about the previous fiscal year. KHPS 2004, for example, asks about respondent behavior in the previous fiscal year, Accordingly, KHPS 2004 includes data on the socio-economic status of respondents before the 2004 pension reform approval, and KHPS 2008 and succeeding waves represent household behavior after the enactment of the reform. Although ten waves (KHPS 2004 through 2013) were available, we restricted our main sample to KHPS 2005 through As KHPS 2004 does not include information about time spent on childcare, the domestic labor supply in this wave is inconsistent with other waves, so we used this first wave to obtain background information. We also excluded KHPS 2009 through 2013 from our sample because of two external events: the global financial crisis of 2008, the impact of which seems difficult to distinguish from that of the 2007 pension reform, and the Great East Japan Earthquake of 2011, which very likely caused heterogeneous effects across households and so is not appropriate for DD estimation. We defined several variables for empirical analysis. Family size was defined as the number of 19

20 household members who live together, including the respondent, and the number of children was similarly defined. We also set a dummy variable which takes one if a household has any member aged less than or equal to six. Market labor supply and domestic labor supply 12 were measured as the average hours per week, and the residual hours were defined as leisure. The set of households chosen for the analysis sample was selected according to the following criteria: (1) spouses who got married before 2004, (2) spouses who live together, (3) spouses who were both aged in 2007, (4) a time allocation that meets the constraint of 168 hours, and (5) the lack of any missing key variables for our analysis. We did not include newly married couples in our sample because in 2004, these couples may have known about the pension reform at the time of their marriage and so within-household variation in bargaining position is not identifiable for those households. We also excluded those aged between 60 and 65 in order to avoid the direct impact of the Elderly Employment Stabilization Law, and while the pension reform could have potentially influenced those aged over 65, we excluded them as well due to a small sample size. Further, we set the lower bound on age to exclude full-time students, and while this cut-off rule may appear arbitrary, our analysis found that it was not critical for our results. DD estimation, which allows us to separate the impact of a specific policy from counterfactual time trends that would be faced by the treated in absence of treatment, requires that the analysis sample of households be divided into a treatment group and a control group. We included households in which a husband was regularly employed but a wife was not in the treatment group, as this type of household would have been most affected by the pension reform, and we included dual-regular households and households with a self-employed husband in the control group, as they would not have been materially influenced by the reform 13. Recall that dual-regular households consist of households in which both members are regular employees, so it does not include households in which the wife is a part-time worker. Note that the treatment status of each household was fixed throughout the sample period by using the employment status in 2003 as it would be unlikely to have been affected by the reform approved in To sum up, we assigned a household to the treatment group if in 2003 the husband was a regular employee but his wife was not, with the remaining households 12 Domestic labor supply consisted mainly of meal preparation, laundry, grocery shopping, cleaning and childcare. 13 Naturally, households in which the wife was a regular employee and the husband was not would also have been affected by the pension reform, but as the impact would be in the opposite direction of that of our treatment group, and since the size of this subset was very small, we included this type of households in the control group. 20