Public pension and senior worker employment rule in Japan

Japan’s public pension system consists of a basic pension (BP; first-tier, kiso-nenkin in Japanese) and employee pension insurance (EPI; second-tier, kosei-nenkin in Japanese). Self-employed persons and students over the age of 20 participate in the BP (also called the national pension or kokumin-nenkin in Japanese). The BP premium is fixed, and the benefits are adjusted proportionally based on the contribution periods. The BP premium is JPY 16,540 and its full benefit is JPY 65,141 per month in 2020, which requires 40 years of contribution. The average BP benefit is JPY 56,000 [37]. Participants with BP were excluded from the study. Private company employees and public servants participate in the EPI, which is our study target. After 2017, the premium is 18.3% of the remuneration, half of which is paid by employers. The EPI benefits are adjusted based on remuneration and length of contributions. The average EPI benefit is JPY 146,000, including the BP (fixed part of a pension) benefits [37]. The earliest age at which BP and EPI benefits can be claimed is 60 years, and the latest is 70 years (75 years after 2022).

The NPA is the standard starting age for public pension benefits defined by the government where full pension benefits can be received. The NPA of the BP and EPI has gradually increased. The NPA of BP for men increased from 60 to 61 in 2001 and continued increasing by one year every three years until it reached 65 in 2013. The NPA of EPI (wage-proportional part of the pension) increased from 60 to 61 in 2013 and will continue to rise by one year every three years until it reaches 65 in 2025. For women, the increase has been similar but with a five-year lag, which means that the NPA for the BP reached 65 in 2018 and will reach 65 by 2030 for the EPI. In addition to public pension, some private companies offer company pension plans (third-tier). A defined contribution pension plan was established in 2001, and a revised defined benefit pension plan was introduced in 2002. Individuals can also participate in the individual defined contribution (iDeCo) plan introduced in 2001.

The mandatory retirement age is the age at which a person must retire from the company, as defined by its employment rules; this age has been at least 60 years since 1998. The increase in NPA to 65 years may raise social concerns because household income may be unstable between the mandatory retirement age and NPA. The senior worker employment policy (Act on the Stabilization of Employment of Elderly Persons) was revised in 2006, and obligates employers to take at least one of the following three measures with certain exemptions: (1) raising the mandatory retirement age to the NPA, (2) setting up formal rules for employment extension and reemployment after the mandatory retirement age, or (3) abolishing the mandatory retirement age. The 2013 revision made this requirement mandatory. Generally, the mandatory retirement age and NPA are equal, or the NPA is higher. Our respondents were EPI beneficiaries aged 40–59 years. For most respondents, the NPA is 65, allowing them to continue working up to 65 if they choose, despite the mandatory retirement age being 60 years.

Earnings test rule in Japan

The ET in Japan (zaisyoku-rourei-nennkin) reduces EPI benefits if total income (including wage income and EPI benefits) exceeds a certain threshold. The current ET for individuals aged 65 years or older was introduced in 2002. The threshold in 2022 was JPY 470,000 for total monthly income. We describe the current ET rule in Japan along with our experimental setting (for details, see Japan Pension Service [38]). Actual calculations are applied every month; however, in our experiment, the calculations are made on an annual basis. The annual suspended EPI benefit (S) at 65 years of age (NPA) is as follows: (1) where W is the annual (adjusted standardized wage) income used to calculate the pension premium and I is the annual EPI benefit, where the subscripts indicate age. Eq (1) implies that half of the benefits above the threshold are reduced until all EPI benefits are suspended. BP benefits (B) are not subject to suspension. The annual total pension benefits (P), if received from the age of 65, is as follows: (2)

Even if pension benefits are voluntarily deferred, part of the actuarial adjustments is reduced for those earning more than the threshold. The pension benefit received from age 66 is as follows: (3) where g is the actuarial adjustment rate (8.4% in 2022 after the NPA) and P₆₅ is the pension benefit assumed to be received if claimed at age 65, as calculated by Eq (2). Eq (3) shows that no actuarial adjustment is applied to suspended benefits (S₆₅). As of 2018, approximately 410,000 individuals (17% of current working beneficiaries) have suspended their benefits. Approximately 200,000 individuals (8% of current working beneficiaries) have suspended their full benefits [39].

The details of our experiment are explained in the following section, where we consider different ET-elimination patterns. First, we propose eliminating the reduction of future benefits by actuarial adjustment by setting S₆₅ = 0 in Eqs (2) and (3); specifically, pension benefits at age 66 (delayed benefits) are calculated as follows: (4) (5) where is the unreceived (postponed) pension benefit at age 65. Second, we propose eliminating the suspension of immediate benefits by setting S₆₅ = 0 in Eq (2): (6)

Although Eqs (4) and (6) may appear identical, they differ in the timing of the benefit claims. In Eq (4), benefits are not claimed at age 65 but are postponed to 66. Conversely, in Eq (6), benefits are claimed and received starting at the age of 65.

The other current ET rule in Japan states that, for those aged 60–64 years, the threshold is the same as that for those aged 65 years or older after 2022. This rule will be discontinued in 2025 for men and in 2030 for women when the NPA reaches 65. The ET rule for those aged 65 years or older applies to these workers even when they claim benefits before the age of 65. Therefore, we focus on the ET rule for those 65 or older.

Base experimental setting and control group

In our survey experiment, respondents faced a hypothetical situation regarding retirement and pension-claiming decisions, similar to discrete choice experiments [25, 40, 41]. Respondents were asked to assume they were 59 years old and to consider retiring and claiming public pension benefits at age 65 or 66. We constructed five situations (or work plans): Plans A–E. These work plans comprised labor income and pension benefits: labor income at ages 65 and 66 and pension benefits at ages 65 and 66 or older. Labor income included two types: high and low income. Pension benefits were based on current standard amounts, applying the ET rule. Specifically,

1. Plan A denotes high income with pension benefits (HIWP),
2. Plan B denotes high income without pension benefits (HINP),
3. Plan C denotes low income with pension benefits (LIWP),
4. Plan D denotes low income without pension benefits (LINP), and
5. Plan E denotes no work with pension benefits (No work), respectively.

We asked the respondents which work plan they preferred most. Oshio et al. [2] use similar categories to represent the work and public pension benefit receiving statuses of senior workers in Japan. We suppose that high income corresponds to full-time work and low income corresponds to part-time work (reduced hours). Baker and Benjamin [8] and Disney and Smith [7] show a discrete shift from part-time to full-time employment. In a Japanese study, Higuchi and Yamamoto [32] analyze the labor supply as the difference between full-time and part-time workers. This study interprets part-time as reemployment after the mandatory retirement age, resulting in reduced working hours.

In reality, decisions to receive pension benefits and work (retirement) are highly associated, but they can also be considered separate decisions under certain conditions. For example, even after the mandatory retirement age, individuals can be rehired to work or move to another company; and they can decide whether to receive their benefits. However, they generally receive benefits once they retire. For our experiment, we assume that the participants can separately decide on working and receiving benefits (Plans A–D), excluding the full retirement decision at age 65 (Plan E).

We designed one control and three experimental treatments to assess the impact of eliminating the ET. Tables 1–4 outlines the experimental instructions provided to respondents. Table 1 introduces the experimental parameters for the control group, representing the status quo and reflecting the conditions under the existing ET rule, as previously discussed. This setup forms the foundation of our experiment. Below, we detail the characteristics of each plan, focusing on income levels, pension-claiming ages, and pension benefits:

• Income at age 65: Plans A (HIWP) and B (HINP) are characterized by high incomes (annual JPY 6.65 million (M)), while Plans C (LIWP) and D (LINP) feature low incomes (JPY 4.43 M).
• Pension claiming age (65 or 66): Plans A and C involve working until age 66 and receiving public pension benefits from age 65 (early pension claiming). In contrast, Plans B and D entail working until age 66 without receiving benefits at age 65 (delayed pension claiming). Plan E involves retiring at age 65 and receiving benefits from age 65.
• Pension benefits at age 65: Based on the Japan Pension Service [38], we set the EPI full annual benefits before the ET at age 65 (I₆₅) as JPY 1.2 M and BP full annual benefits (B₆₅) as JPY 0.78 M. The total full benefits before applying the ET rule sum to JPY 1.98 M. The benefits are suspended based on the ET rule. Under the ET rule, benefits in Plan A are reduced to JPY 0.88 M (= 0.78 M +1.2 M– 1.1 M) from Eq (2) due to high labor income. The suspended EPI benefits (S₆₅) amount to JPY 1.1 M from Eq (1), given W₆₅ = 6.65 M and I₆₅ = 1.2 M. Plans C and E at age 65 receive full benefits (JPY 1.98 M) due to low labor income (W₆₅ = 4.43 M).
• Pension benefits at age 66: Plan A has full benefits (JPY 1.98 M) because of no labor income or corresponding benefit suspension. Benefits for Plan C remain unchanged from age 65. For Plan B, benefits starting at age 66 (JPY 2.05 M) are lower than those for Plan D (JPY 2.15 M) due to decreased actuarial adjustments. From Eqs (2) and (3), the pension benefits at age 66 for Plan B are JPY 2.05 M = (1 + 8.4%) (0.78 M + 1.2 M– 1.1 M) + 1.1 M. Those of Plan D are JPY 2.15 M = (1 + 8.4%) (0.78 M +1.2 M).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Experimental instruction and parameter settings for control group.

Suppose you are now 59 years old. You are considering whether to retire at age 65 or work at age 65, and retire at age 66. Which work style and way of receiving the public pension (the sum of EPI and BP) would you prefer? Please choose the option that best applies to you.

https://doi.org/10.1371/journal.pone.0304458.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Enhancement of future pension (FP) treatment.

https://doi.org/10.1371/journal.pone.0304458.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Enhancement of current pension (CP) treatment.

https://doi.org/10.1371/journal.pone.0304458.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. FPCP treatment.

https://doi.org/10.1371/journal.pone.0304458.t004

In addition to outlining these parameters, a brief description of each work plan is provided.

Treatment groups

To examine the effect of ET elimination, we constructed three treatments in addition to the above control setting: enhancement of future pension (FP), enhancement of current pension (CP), and FPCP treatments. The FP treatment eliminated the reduction in actuarial adjustments, whereas the CP treatment eliminated the suspension of current (age 65) benefits. The FPCP treatment eliminated the reduction in both actuarial adjustments and current benefits, meaning that the ET was eliminated completely. Table 2 presents the FP treatment parameters. For the FP treatment, the reduction in future pension growth was eliminated. Actuarial adjustment (currently 8.4% annually) increases benefits if the benefit-claiming age is delayed after NPA. However, under the current ET rule, no actuarial adjustment was performed for a part of the benefits, as shown in Eq (3). The FP treatment eliminated this decreased actuarial adjustment, meaning that an individual could receive full delayed benefits at 66 years or older, regardless of labor income at 65 years, as shown in Eqs (4) and (5). The only difference between Tables 1 (control) and 2 (treatment FP) is the pension benefits for those aged 66 or older in Plan B. In Table 1, the benefits were JPY 2.05 M, whereas in Table 2, they were JPY 2.15 M because the reduction of the actuarial adjustment was eliminated.

Table 3 presents the CP treatment parameters. In the CP treatment, the suspension of current benefits was eliminated, meaning that full benefits at age 65 are obtained regardless of current labor income, as shown in Eq (6). The only difference between Tables 1 (control) and 3 (CP treatment) lies in the pension benefit for those over 65 years in Plan A. In Table 1, the benefits are JPY 0.88 M, whereas in Table 3, they are JPY 1.98 M due to the elimination of the benefit suspension at 65 years.

Table 4 shows the FPCP treatment parameters. In the FPCP, the suspension of both future and current pension benefits was simultaneously eliminated. That is, the ET was completely eliminated, and there was no reduction in current pension benefits or actuarial adjustments. Tables 1 (control) and 4 (treatment FPCP) differ in terms of pension benefits for individuals aged 65 years in Plan A and those aged 66 years or older in Plan B. In Table 4, the benefits at age 65 in Plan A are JPY 1.98 M, and the delayed benefits claimed at age 66 in Plan B are JPY 2.15 M.

Generally, controlling for the respondent’s situation in an experiment was preferable. For example, explaining the effect of early or late pension claims might have been necessary to make appropriate retirement decisions. However, additional explanations could have burdened the respondents. Thus, we provided only the minimum necessary explanations to our respondents; other situations were left to the respondents’ discretion.

Theoretical and experimental predictions on ET elimination

Following previous studies, we use a one-period labor supply model to predict individual behavior for ET elimination [1, 5, 7, 8, 42]. This simple model illustrates the theoretical impact of ET elimination on labor supply decisions. Fig 1 shows workers’ budget constraints relating to the allocation between work and leisure, where the decision to claim pension is not considered for the time being. Without a public pension system, the budget constraint is FI (for reference), where F signifies total leisure (0% work hours). This line’s slope reflects the wage rate–w, omitting non-labor income, as per Japan’s earnings test. In a public pension system with an earnings test, we observe kinked budget constraints EJGI (our base case). No benefit is reduced if the work hours are less than the threshold h₁ on segment JE. All benefits are taxed away beyond h₂ on GI. The slope is–w (1–τ) for JG, where, mirroring Japan’s earnings test, the withholding (tax) rate τ is 50%; the threshold h₁ and h₂ set at annual incomes of JPY 4.44 M and JPY 6.84 M in our experimental setting, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Retirement earnings test and budget constraint.

Note: Plans C and D are in the same location because the only difference is the timing of the actuarially fair benefit claims. The same applies to Plans A and B and Plans A^CP and B^FP. Here, h₁ corresponds to an annual labor income of JPY 4.44 M, while h₂ corresponds to that of JPY 6.84 M. The FP treatment eliminates benefit reductions in actuarial adjustments and moves Plan B to B^FP. The CP treatment eliminates the current benefit suspension and moves Plan A to A^CP. The FPCP treatment completely eliminates the ET and moves Plans A and B to A^CP and B^FP, respectively. Labor income for Plans A and B is JPY 6.65 M (h_H) and that of Plans C and D is JPY 4.44 M (h_L).

https://doi.org/10.1371/journal.pone.0304458.g001

Eliminating the ET shifts the budget constraints from EJGI to EK (our analytical target). The effects of this change differ depending on the initial location. Individuals initially on EJ face no change, while for JG, the impact is mixed: reduced labor supply by income effects and increased by substitution effects; the net effect remains an empirical question. Prior research has indicated that the substitution effect tends to prevail. For those on segment GI, eliminating the earnings test yields only negative income effects, suggesting less work. We recognize the limitation of this model—that is, labor adjustment may not be continuous and differ depending on family structure, spousal income, and financial asset holdings.

Next, the locations of each plan in our experiment are depicted in Fig 1. We initially assume that the pension-claiming age (65 or 66 years in our case) is indifferent because benefits are actuarially fairly adjusted. That is, the present values of benefits are equivalent on average regardless of claiming age because increased benefits in the future effectively neutralize forgone benefits through actuarial adjustments. We also assume no liquidity constraint, meaning life-contingent financial products can mimic delayed pension benefits, attaining the same cash flow even if the delayed pension claims are not available. Plan E is located where leisure is 100% (no work), and the individual receives pension benefits. Plans C and D are located on segment EJ (just right of J). Plans A and B are on JG (close to G). Labor income at age 65 in Plans C and D is JPY 4.43 M, corresponding to h_L in Fig 1. Similarly, labor income at age 65 in Plans A and B is JPY 6.65 M, corresponding to h_H. As mentioned, we assume that the respondent is indifferent toward the choice of either Plans C or D (A or B). The ET elimination in the FPCP treatment moves Plans A and B to A^CP and B^FP, respectively.

Finally, we consider individual variances in valuing actuarial adjustments. Respondents’ behavior varies because they may perceive actuarial adjustments as either favorable or unfavorable [7]. This variation is attributed to personal discount rates [43] and subjective life expectancy [44, 45]. Early or delayed claims decisions have long-term impacts on one’s life because benefits are fixed depending on the claiming age. We infer that respondents believing that actuarial adjustments favor them choose a plan with the delayed benefits (Plans B or D) and vice versa. Table 5 summarizes the factors affecting respondents’ plan choices based on Fig 1‘s theory and predictions from empirical studies according to our treatment setting.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Factors affecting respondents’ choice according to treatment.

https://doi.org/10.1371/journal.pone.0304458.t005

The CP treatment shifts Plan A to the A^CP located on JK in Fig 1, keeping Plan B at its original position. In the CP treatment (second row in Table 5), we expect respondents choosing Plan C or D in the first trial (control) to change to Plan A^CP by the substitution effect. In addition, respondents choosing Plan B in the first trial change to Plan A^CP because of its relative attractiveness. However, some respondents, considering longevity risk and believing that the actuarial adjustment is favorable to them, may remain in Plan B because it still offers increased benefits at age 66 or older (favorable actuarial adjustment).

Similarly, the FP treatment moves Plan B to B^FP on JK, keeping Plan A at its original position. We expect that the probability of choosing Plan B^FP will increase for those choosing Plans C or D in the first trial because of the substitution effect. In addition, some participants choosing Plan A in the first trial switch to Plan B^FP because of its relative attractiveness. However, some respondents, believing that the actuarial adjustment is unfavorable, remain in Plan A because it still offers the current benefits at age 65 (unfavorable actuarial adjustment). In the later section, we provide policy recommendations based on the respondents’ preferences, especially in the FP treatment, regarding increasing labor supply and delaying pension claims. Then, in the FPCP treatment, we expect that those choosing Plans C or D in the first trial change to Plans A^CP or B^FP by the substitution effect. Their choice depends on the subjective fairness of actuarial adjustment.

Note that, as our plans are discrete and respondents cannot adjust the labor supply continuously, we assume that respondents choose the closest plan for their labor supply preference. Strictly speaking, Plans C and D are on EJ and should not be affected by ET elimination. However, we assume Plans C and D are located around the kink. Empirical studies have shown that eliminating the ET increases the labor supply for those with incomes just below the kink point.

Although Plan E’s budget constraint remains unchanged in Fig 1‘s model, eliminating the ET could shift preferences within the FP, CP, and FPCP treatments owing to labor market dynamics. Individuals contending with labor market rigidity and the fixed costs of entering the labor market may initially opt for Plan E, as Baker and Benjamin [8] note. However, removing the ET might alleviate these barriers, potentially encouraging greater workforce participation in the FP, CP, and FPCP treatments, as predicted in the last column of Table 5 (Not choosing Plan E).

We recognize the difference between theoretical explanations and experimental settings. Our theory is conventionally explained sequentially, addressing labor supply and then pension claims. Respondents in our experiment may determine their plan in a different order. The theory outlined in this section is designed to simplify complex real-world phenomena, making them more comprehensible. However, to capture the nuances of real-world behavior more accurately, it will be important to consider more advanced theories in future research.

Survey implementation

An Internet-based survey was conducted in Japan in February 2022, and respondents were contacted using the registered individuals at MyVoice.com Ltd (https://www.myvoice.co.jp). We conducted the two surveys for this study: preliminary and main surveys. The recruitment period for the preliminary survey was from February 4, 2022, to February 9, 2022, followed by the main survey from February 10, 2022, to February 15, 2022. This study has been approved by the review committee for research involving human participants at Konan University (21–19). All participants provided (computer-based) written informed consent. Recruitment targeted registered members of MyVoice.com, reached through Internet advertising. Upon registration, members provided demographic and socioeconomic information, including gender, age, place of residence, marital status, occupation, annual income, and family structure (https://voice.myvoice.co.jp/front/sign_up/top/). We randomly selected approximately 5,000 potential respondents for the preliminary survey using this information—this survey aimed to update their demographic and socioeconomic information, acknowledging potential changes since their initial registration. From those who completed the preliminary survey, male participants aged 40–59 years, married, employed by a company, and with a household income ranging from JPY 4.0 M to less than JPY 9.0 M were randomly chosen to participate in the main survey. We used household income instead of individual income to exclude extrema households such that the household head and spouse income are both JPY 9.0 M or more, which is almost a top-income household in Japan. Respondents were randomly allocated to the three treatment groups (Groups 1–3) to mitigate the effects of the treatment experience, as explained below. They were further divided into subgroups that were almost equally distributed in five-year age intervals (40–44, 45–49, 50–54, and 55–59) and high- and low-income groups: the low-income group comprises those whose household income is more than or equal to JPY 4.0 M and less than JPY 6.5 M, while the high-income comprises more than or equal to JPY 6.5 M and less than JPY 9.0 M. We created age subgroups, considering that the population compositions of these age ranges are almost the same. We also created income subgroups so that respondents’ actual annual incomes do not differ significantly from these experimental settings. Given these sub-groups, we ensured that the respondents’ age and annual income distribution were not highly skewed. We tried to collect approximately 90 respondents in each subgroup and 360 respondents in each group. Each respondent underwent four trials to choose Plans A–E.

Table 6 displays the treatment order according to the respondent group. Each group faced treatments in a different order to mitigate the experience of preceding decisions in measuring the overall average treatment effects. The first trial was the control (status quo) for all respondent groups. The treatment order from the second to fourth trials differed according to the group. For example, Group 1 respondents faced the FP treatment in the second trial and the CP and FPCP treatments in the third and fourth trials, respectively. Respondents answered the questions as presented in Tables 1–4, followed by questions about their personal attributes. They were rewarded with a fixed cash equivalent. Given that public pension benefits vary based on employees’ income and length of service, we limited respondents to certain occupations and annual income levels to align with our experimental setting. In addition, those with large heterogeneity in public pension benefits, such as self-employed individuals, were excluded.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Order of treatment according to respondent group.

https://doi.org/10.1371/journal.pone.0304458.t006

Respondents were informed about the purpose of the survey: the authors conducted this survey to empirically analyze the employment of older workers and the pension system at their research institutions. The study sought to analyze the pension system design and explore the requirements for a prosperous retirement. Respondents were also informed that the response data would be statistically analyzed and used to prepare materials, reports, and academic papers and no other purpose. To complete the survey, they were instructed to read the questions online and select the answers that best represented their situation. Therefore, survey participants responded knowing the survey was about pensions and employment. However, respondents were not explicitly told that the study was designed to examine reforms to the ET in the pension system. As explained in Tables 1–4, we did not use the term “retirement earnings test” in our survey. Instead, we specifically indicated labor income and pension amounts to overcome the limitation of respondents, who were mostly young and did not have specific knowledge of the ET. At the same time, more knowledgeable respondents may gain an erroneous impression of the ET and fail to consider the labor income and pension amounts presented properly. Instead, they choose answers indicating ET elimination. Presenting labor income and pension amounts as work plan options without specifically mentioning the ET would be more likely to elicit an honest preference for employment and pension benefits.

As a result, respondents more interested in old-age pensions and employment issues were more likely to participate in the survey. By contrast, those less interested may be less likely to participate. However, given Japan’s declining birthrate and aging population, interest in employment and pensions among middle and senior workers is high. We anticipate that the number of respondents who opt out of the survey will not be significant due to their disinterest in these topics. Nonetheless, we acknowledge that some sample bias related to interest in the survey persists, warranting caution in interpreting our findings and highlighting the need for further research.

Empirical method

To estimate the treatment effects on senior workers’ employment and pension-claiming behaviors, we estimate the following regression: (7) where i represents a respondent. The outcome variable Y represents the respondent’s choice of work plans, which takes six forms: HIWP, HINP, LIWP, LINP, High income, or Work65. HIWP, HINP, LIWP, and LINP equal 1 if a particular work plan is selected, and 0 otherwise. Furthermore, we use the dummies High income and Work65. High income equals 1 if a respondent chooses the high-income plan (HIWP or HINP) and 0 otherwise. Work65 equals 1 if a respondent chooses plans for work at 65 (Plans A–D) and 0 otherwise (Plan E: retire at 65).

The main independent variables are FP, CP, and FPCP, which capture the treatment settings. FP (enhancement of future pension) equals 1 if a respondent faces treatment FP and 0 otherwise. In the FP treatment, the ET for reducing future pension growth based on actuarial adjustments is eliminated, given that all other parameters are unchanged, as shown in Table 2. CP (enhancement of current pension) equals 1 if a respondent faces the CP treatment and 0 otherwise. In the CP treatment, the ET for suspending current benefits is eliminated, given that all other parameters are unchanged, as shown in Table 3. FPCP equals 1 if a respondent faces the FPCP treatment and 0 otherwise. In the FPCP treatment, the ET is completely eliminated, as shown in Table 4. Variable Z represents individual characteristics explained below. β₀, β₁, β₂, β₃, and γ are regression coefficients, and ε_i is the residual.

Our empirical analysis is structured as binary choice models for each of the four distinct work plans and two combinations of plans (High income and Work) that respondents may select, where the dependent variable Y in Eq (7) for each model represents the binary choice (0 or 1) associated with selecting a specific plan. Accordingly, we use the ordinary least squares (OLS) with pooled data. We conduct six separate regressions, each examining the impact of the three treatment conditions (FP, CP, and FPCP) on the likelihood of selecting each (combination of) plan (linear probability model). Each respondent experienced the control (status quo), FP, CP, and FPCP treatments once in different implementation orders in each group to which they belonged (the status quo was experienced first by all respondents). Because each plan reflects the characteristics of wages and pension benefits, we can directly assess the effectiveness of each treatment on respondents’ preferences for different wage and pension claiming plans.

Each regression estimates the average treatment effects on the binary outcomes. The coefficients β₁, β₂, and β₃ capture these effects for the FP, CP, and FPCP treatments, respectively. Table 5 includes the empirical predictions according to the treatment. We expect that β₁ < 0, β₂ < 0, and β₃ < 0 in the low-income plans (LIWP and LINP) for the FP, CP, and FPCP treatments, respectively, by the substitution effect. For high-income plans (HIWP and HINP), the signs of β depend on the treatment. For treatment FP, we expect β₁ > 0 for HINP by the substitution effect and relative attractiveness, and β₁ < 0 for HIWP by relative (un)attractiveness. For the CP treatment, we expect β₂ > 0 for HIWP and β₂ < 0 for HINP for the same reasons. For the FPCP treatment, we expect β₃ > 0 for high-income plans and β₃ < 0 for low-income plans due to the substitution effect.

We recognized that the repeated choice aspect of the experiment might generate transition dynamics in the respondents’ choices across different trials. However, the treatment order changes can mitigate these issues to analyze overall treatment effects. Our robustness check analysis also includes the results from the mixed logit model, which measures the treatment’s effect by directly using variables representing wage and pension benefits as explanatory variables with the repeated choice aspect. These results drive qualitatively the same conclusion as our main analysis, as explained later.

The individual characteristics variables (Z) used in the regression analysis are as follows: trust in the public pension, subjective survival probability, financial literacy, subjective health condition, university graduation status, status of working in small companies, mandatory retirement age of 65 years or older, household income, financial assets, house ownership, and age. Trust in the public pension is measured as the respondent’s assessment of their trust in the public pension on a scale of one (almost unreliable) to six (quite reliable). Agnew et al. [46] find that defined contribution pension plan participants who do not trust financial institutions are more likely to quit automatic enrollment plans. We expect respondents with higher trust in the public pension to delay their pension benefit-claiming age. Subjective survival probability is measured as a respondent’s assessment of their probability of survival until the age of 85. Respondents chose options ranging from less than 5%, 5%, ⋯, 95%, and more than 95% in increments of 5%. We coded the choice of less than 5% as 2.5% and more than 95% as 97.5% for the survival probability variable. The literature shows mixed results regarding the survival probability of retirement-related decision-making. Payne et al. [47] find that subjective life expectancies are related to expressed preferences for hedging longevity risks. By contrast, Brown [48] shows that it has an insignificant effect on reducing longevity risk.

Financial literacy is measured as the average score of three standard financial literacy quizzes that test the (a) respondents’ understanding of compound interest rates, (b) the relationship between the value of foreign investments and exchange rates, and (c) the diversification effect. Regarding (b), instead of inflation, we use foreign investment, following Lusardi and Mitchell [49], who state that the relationship between the quiz and people’s daily financial decisions is essential. In Japan, the exchange rate is related to daily consumption, such as food and energy, and foreign investment is more aligned with this concept than the inflation rate. Financial literacy is the average of the correct answer rates for the three quizzes. Higher scores indicate higher financial literacy, which affects individuals’ retirement planning and implementation [50]. The literature shows mixed results for financial literacy regarding retirement and public pension-claiming decisions. Brown et al. [51] find that individuals with low financial literacy are less likely to value annuities, or public pensions in our study. They might not fully understand the benefits of delaying pension claims due to actuarial adjustments. By contrast, Agnew et al. [52] find that individuals with high financial literacy are less likely to depend on annuities and are more likely to choose self-management. Such individuals tend to be more confident about stock and mutual fund investments and prefer to receive public pensions at the NPA. Household income is the sum of the respondent’s and spouse’s incomes, obtained through multiple-choice questions with 14 choices. Financial assets are also obtained through a multiple-choice question with 22 choices, including bank savings, stocks, mutual funds, and money market funds, excluding the house. Income and financial asset choices are presented in ranges, and the variables use their medians.

Characteristics of respondents

Table 7 presents the descriptive statistics of the respondents’ characteristics. The trust for public pension is 3.0, which is neutral on average. The average survival probability at age 85 is 32.5%, far lower than the actual (objective) probability (42.5%) based on an average respondent age of 50 years. The average health condition is 4.3, which is relatively good. Table A1 in S1 Appendix presents the detailed responses for those categorical variables. The average correct answer rate for the financial literacy quizzes is 59.0%, which is within the range of results in the literature. Household income and financial assets are JPY 7.0 M and 8.9 M, respectively. Our respondents’ higher income and financial asset levels, compared to the national average, may be attributed to our targeted screening process during recruitment. The average income of respondents is JPY 5.8 M, which is on the segment GJ in Fig 1. We believe this income level seems appropriate for choosing between a high- or low-income plan at age 65. Table A2 in S1 Appendix presents the differences in individual characteristics between our sample and a large-scale governmental survey for the selected variables. The characteristics of our respondents align closely with those from the larger survey, except for education levels, a variation that could be attributed to the online nature of our survey. Table A3 in S1 Appendix presents the individual characteristics by respondent group. While minor differences are noted, we believe they do not significantly influence our analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Descriptive statistics of respondents.

https://doi.org/10.1371/journal.pone.0304458.t007

Transition probability by treatment effect

This subsection analyzes the respondents’ choices during the first and second trials to capture their initial behavior through straightforward analysis, as explained in Table 5. Each respondent’s first trial represents the control condition (status quo), and the second trial introduces their first experience with a treatment condition, varying according to the respondent’s group as outlined in Table 6. This setup ensures consistent comparison conditions across treatments. However, for the third and fourth trials, respondents encounter treatments based on their group allocation, introducing variability due to the experience effect, where prior exposure to treatments may influence subsequent responses. This variability presents challenges for direct comparison of later trials. We will explore the results of regression analyses using all data in the subsequent subsections, aiming to provide comprehensive insights into the treatment effects.

Table 8 displays the transition probabilities of the plan choices from the first to second trials according to the treatments. Panel A shows the transition probabilities of the FP treatment (Group 1). In the first trial (control), 92% of respondents chose to work at age 65 (8% did not work). The highest probability of being selected is HINP (29%), and the other working plans are almost similar (20–22%). In the second trial, the FP treatment increases the pension benefits at age 66 in HINP (Plan B). The probability of changing to HINP from low-income plans is high: 57% (32%) of respondents who select LINP (LIWP) in the first trial change to HINP in the second trial, which is consistent with the substitution effect. Among the high-income plans, approximately half of the respondents who select HIWP in the first trial change to HINP, which is consistent with the relative attractiveness. The remaining half maintains HIWP, consistent with the unfavorable actuarial adjustment. These results are consistent with the predictions presented in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Transition probabilities by treatment.

https://doi.org/10.1371/journal.pone.0304458.t008

Panel B shows the transition probabilities for the CP treatment (Group 2). The frequency of choosing each plan at the first trial is similar to that in Panel A. HINP has the highest probability of being selected (26%). The transition probabilities also have a similar tendency to those in Panel A. For the second trial, the CP treatment increases the pension benefits at age 65 in HIWP (Plan A). The probability of changing from low-income plans (LIWP and LINP) to HIWP is high (43% and 17%, respectively), consistent with the substitution effect. Approximately half of the respondents who select HINP change to HIWP, consistent with the relative attractiveness. The remaining half maintains HINP, consistent with the favorable actuarial adjustment.

Panel C shows the transition probabilities for the FPCP treatment (Group 3). For the first trial, the frequency of choosing each plan is similar to that in Panel A. For the second trial, the FPCP treatment eliminates the ET completely and increases pension benefits at age 66 or older in HINP and age 65 in HIWP. The probabilities of changing from low-income to high-income plans are high (5–49%), consistent with the substitution effect. The respondents exhibit a specific tendency in their choices: Those who choose LINP in the first trial tend to change to HINP (similarly, LIWP to HIWP). After the complete elimination of the ET, the only difference between HIWP and HINP is the timing of the claiming age for actuarially fair benefits. Therefore, the preferences for these plans should theoretically be indifferent. However, respondents’ pension benefits-claiming behavior is strongly related to the subjective evaluation of actuarial adjustments even under ET elimination, consistent with Disney and Smith [7]. For example, the sum of the transition probabilities from LIWP to plans that include pension claiming at age 65 (HIWP and LIWP) is 82%. Similarly, the sum of the transition probabilities from LINP to plans that include pension claiming at age 66 (HINP and LINP) is 89%.

Main results

Table 9 presents the estimation results based on Eq (7). The outcome is an indicator that represents the work plan and its combinations. The primary independent variables are FP, CP, and FPCP, which capture the extent of elimination of the ET. FP eliminates the reduction in actuarial adjustments, and CP removes the suspension of current pension benefits. FPCP eliminates both the reduction in actuarial adjustments and suspension of benefits. We employ a linear probability model to estimate Eq (7), with standard errors clustered at the respondent level.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Effect of eliminating earnings test.

https://doi.org/10.1371/journal.pone.0304458.t009

Regarding the intensive margin (change in labor income), the FP treatment significantly decreases the choice of low-income plans and increases the choice of HINP. The FP treatment reduces the probability of choosing LIWP by 4.9 percentage points in Column (3) and 13.2 percentage points for LINP in Column (4). It also increases the probability of choosing HINP by 23.0 percentage points in Column (2), which is consistent with the substitution effect. In Column (1), the FP treatment reduces the probability of choosing HIWP by 4.4 percentage points, consistent with its relative (un)attractiveness among high-income plans. Some respondents remain in the HIWP, as actuarial adjustments are unfavorable for them, and prefer a plan to receive pension benefits at age 65. The predicted probability of HIWP evaluated at FP = 0, CP = 0, and FPCP = 0 (control) is 19.4%, whereas that at FP = 1, CP = 0, and FPCP = 0 (treatment FP) is 15.0%. Indeed, ET reform accelerates earlier benefit claims [3]. By contrast, the FP treatment encourages labor supply and delayed benefit claims, which is a novel finding. In addition, those with higher trust in public pensions and a higher survival probability at age 85 tend to choose HINP. These respondents may be concerned about their longevity risk.

Second, the CP treatment significantly decreases the choice of low-income plans and increases the choice of HIWP. It reduces the probability of choosing LIWP by 7.1 percentage points in Column (3) and 5.9 percentage points for LINP in Column (4). It increases the probability of choosing HIWP by 23.7 percentage points in Column (1), which is consistent with the substitution effect. In Column (2), the CP treatment reduces the probability of choosing HINP by 9.4 percentage points, indicating relative (un)attractiveness. Some respondents remain in HINP, given favorable actuarial adjustments, and prefer a plan with delayed benefits. The predicted probability of HINP evaluated at FP = 0, CP = 0, and FPCP = 0 is 27.8%, whereas that of FP = 0, CP = 1, and FPCP = 0 (treatment CP) is 18.4%. Table A4 in S1 Appendix presents the results segmented by respondent group (Panel A). Notable differences are observed in the FP and CP treatments across groups, indicating the presence of experience effects (Panel B). Varying treatment sequences mitigate this concern.

Third, the FPCP treatment significantly decreases the choice of low-income plans and increases the choice of high-income plans. It reduces the probability of choosing LIWP by 7.6 percentage points in Column (3) and 13.2 percentage points for LINP in Column (4). It increases HIWP by 10.5 percentage points in Column (1) and HINP by 11.7 percentage points in Column (2), consistent with the substitution effect. In Column (2), the FPCP treatment weakly increases HINP (–0.113 = 0.117–0.230) compared with the FP treatment (p < 0.01) because the FP treatment includes relative attractiveness in addition to substitution effects.

Fourth, we observe distinct sensitivities among respondents to the different treatments. The FP treatment reduces the likelihood of selecting LINP by 13.2%, a more pronounced effect than the LIWP plan, which decreases by 4.9%, indicating a greater sensitivity among respondents opting for delayed benefits to the FP treatment. Conversely, the CP treatment shows a stronger reduction in preference for LIWP (–7.1%) over LINP (–5.9%) and a notable increase in preference for the HIWP, underscoring the sensitivity of respondents favoring immediate benefits towards CP. Similarly, the FPCP treatment exhibits a more significant decrease in the preference for LINP (–13.2%) compared to LIWP (–7.6%) and a larger increase in preference for HINP (+11.7%) over HIWP (+10.5%). These patterns suggest that decisions regarding pension benefit claims under the ET elimination are influenced by the perceived fairness of actuarial adjustments, aligning with the findings of Disney and Smith [7].

Finally, in Column (5), the FP, CP, and FPCP treatments significantly increase the choice of high-income plans. The effect of FP exceeds that of CP; the difference in coefficients is 0.043 (p < 0.01), indicating that the FP treatment is more likely to be a potential policy candidate for ET reform.

Next, regarding the extensive margin (work or not), Column (6) shows that the CP and FPCP treatments significantly increase Work65, which the labor supply model in Fig 1 does not predict. The result indicates that some respondents who choose not to work at age 65 (Plan E) may be subject to the fixed costs of work that prevent workers from freely choosing work hours. Eliminating the ET relaxes this constraint, as Baker and Benjamin [8] note. Contrary to these treatments, FP alone is insignificant for work, which aligns with the unfavorable actuarial adjustment. The respondents who choose not to work at age 65 and receive pension benefits do not prefer to delay the claiming age, even though the FP increases the delayed benefits.

Regarding the control variables, more financial assets decrease the choice of HINP, High income, and Work, indicating that those with sufficient financial assets are less willing to work. Financial literacy has an insignificant effect on the choice of work plans. Basic financial literacy may not impact the deferments of public pension benefits, and other skills, such as numeracy and knowledge of financial products, may be required to understand the benefits of delaying public pension claims [53]. Note that we estimate other specifications, such as excluding the FPCP treatment dummy variable, to check the robustness of our results. Nevertheless, our conclusions remain qualitatively the same.

Robustness checks

Analysis of data divided into subsamples.

We estimate Eq (7) for three subsample categories: household income, financial assets, and age. The effect of ET elimination may differ by the respondents’ income subgroups. Research shows that this effect is expected to be more significant for respondents just below the threshold [5, 9]. Panels A and B in Table 10 present the results regarding whether the household income subsample is lower and higher than the median. We report only the coefficients of FP, CP, and FPCP, and omit the coefficients of the control variables. The overall results in Panels A and B are similar to those in Table 9. FP (CP) increases HINP (HIWP) and decreases low-income plans. The FPCP increases and decreases high- and low-income plans, respectively. The magnitudes of these changes are similar in Panels A and B.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 10. Treatment effects by subsample.

https://doi.org/10.1371/journal.pone.0304458.t010

Next, holding sufficient financial assets can contribute to lowering longevity risk. We predict that respondents’ financial assets can affect their decisions. Panels C and D show the results for the subsamples wherein respondents’ financial assets are lower and higher than the median, respectively. Again, the overall results in Panels C and D are similar to those in Table 9.

Finally, we also predict that respondents’ age (the decision span until retirement) affects their decisions because of the high correlation between age and income. Panels E and F show the results for the subsamples with respondents in their 40s and 50s, respectively. Again, the overall results in Panels E and F are similar to those in Table 9.

In sum, these results indicate limited effects of household income, financial assets, and age differences on ET elimination. The results could be attributed to a relatively homogenous sample. While this may be desirable in our study, how respondents with other characteristics would respond to the ET reform remains to be explored.

Analysis using multinomial logit model.

To check the robustness of our results, we estimate a multinomial logit model, where the outcome is a dummy that represents a particular plan (Plans A–E), and the independent variables are the same as our main specification in Eq (7). Table 11 presents the marginal effect of the multinomial logit model. These results are similar to those presented in Table 9 and yield qualitatively the same conclusions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 11. Result based on multinominal logit model (marginal effects).

https://doi.org/10.1371/journal.pone.0304458.t011

Analysis using mixed logit model.

For our main analysis, we used treatment dummy variables to measure the treatment effects directly. We now examine the respondents’ wage and pension benefits preferences according to our treatment settings. Because a discrete choice experiment inspires our experimental settings, we employ a mixed logit model to evaluate respondent preferences, which can handle the transition dynamics of respondents’ choices. An analysis using a mixed logit model includes attribution (e.g., labor income and pension benefits) and levels (their values). As in our experiment, assuming labor income and benefits are received at different times (at ages 65 and 66 or older) complicates the direct application of these variables. In addition, a discrete choice experiment typically utilizes trade-offs, such as giving up a higher price to obtain better quality to infer the respondent’s preference. However, respondents tend to prefer higher labor income and pension benefits with no apparent trade-offs in our settings. Therefore, we use the reduction of labor income relative to high income (JPY 6.65 M) for our price attribute, which is defined as JPY 6.65 M–each plan’s labor income. We also introduce the concept of social security wealth (SSW), which is the present value of pension benefit streams, given survival for our non-price attribute, which can handle the difference in benefit receiving timings. The survival probability is obtained from the 23rd Life Table (Male). We assume the discount rate to be 4.88% for this section, such that the SSW of Plan C is the same as that of Plan D. We also use a dummy for delaying the age of pension claims to 66 years as well as that of work. Table 12 summarizes these attributes and levels for each life plan.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 12. Attributes and levels for mixed logit analysis.

https://doi.org/10.1371/journal.pone.0304458.t012

The conditional logit model is conventionally used to analyze choice experiment data [54]. Because its underlying assumption (i.e., independence of irrelevant alternatives [IIA]) is considered unrealistic and generally does not hold, Train [55] propose a less restrictive mixed logit model (MLM), also called a random parameter logit model, that allows for heterogeneity of individual preferences and a relaxation of the IIA assumption. In the MLM, some parameters are specified as random, implying that each respondent has a different parameter set. The outcome is the respondent’s choice of work plan (Choice), either 0 (not chosen) or 1 (chosen). The primary independent variables are the reduction of labor income, SSW, and the interaction term of SSW and treatment (SSW×FP and SSW×CP). We assume that the SSW coefficient follows a normal distribution. The variable FP for this section is an indicator that takes a value of 1 for FP and FPCP treatments, including the plan to eliminate the reduction of actuarial adjustment; it is 0 otherwise. The variable CP is also an indicator that takes a value of 1 for CP and FPCP treatments, including the plan to eliminate the suspension of current labor income; it is 0 otherwise. We use these variables because the SSWs have no variations among work plans in the FPCP treatment, and the MLM cannot be estimated with the FPCP indicator. We contend that these FP and CP indicators still capture the difference between the FP and CP treatments, although both include the FPCP treatment.

Table 13 displays the results using the MLM [56]. Column (1) displays the estimated coefficients and standard errors; Column (2) displays the willingness to pay (WTP), or subjective valuation, for each attribute: a marginal value of the attribute relative to reduced labor income (the price attribute). The coefficients of the reduction of labor income are negative, indicating that a plan with a higher reduction of labor income is less likely to be chosen. By contrast, those of SSWs are positive, indicating that a plan with a higher SSW is likely to be chosen. These results are consistent with predicted preferences of respondents. The coefficients of SSW×FP and SSW×CP are positive, meaning that a plan with higher SSW is more likely to be chosen compared with the status quo in the FP and CP treatments. In Column (2), the subjective valuations in the FP and CP treatments are 2.115 and 1.209, respectively, indicating that respondents value SSW, or pension benefits, relative to losing current income compared with the status quo. The difference in these subjective valuations confirms that respondents value pension benefits more in the FP treatment than in the CP treatment. These results are consistent with our main analysis. Column (3) adds indicators that delay claims and work at age 65, and the results are similar to that in Column (1). The coefficient of work at age 65 is negative, indicating that a plan involving work at age 65 is less likely to be chosen after considering the pension benefits under SSW. The lower part of the panel reveals that the estimated standard deviations for SSW are statistically different from zero, reflecting unobserved heterogeneities among our respondents.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 13. Result of mixed logit analysis.

https://doi.org/10.1371/journal.pone.0304458.t013