2.1. EPU and fiscal behavior

Several previous studies have examined how EPU influences both macroeconomic indicators (such as output, investment, employment, etc.) and fiscal behavior (such as government spending, taxation, budget deficits, etc.). Jalles and Karras (Reference Jalles and Karras2026), using local projections across 143 countries, found that increases in EPU reduce gross domestic product (GDP) per capita growth, with effects doubling from 0.05 percentage points at year 5 to 0.1 percentage points by year 10. Mumtaz and Ruch (Reference Mumtaz and Ruch2025) showed that fiscal and monetary policy uncertainty acts as a negative supply shock, lowering output, investment, and consumption while raising inflation across 54 economies. However, these effects are not uniform across countries, as Nusair (Reference Nusair2025) documented asymmetries in G7 countries, where rising EPU hurts industrial production while declining EPU boosts it, with particularly notable negative effects in France and Italy.

Studies further show that the effectiveness of fiscal policy depends on the prevailing EPU environment. Li and Wei (Reference Li and Wei2022) argued that government spending multipliers are larger during low EPU periods compared to high EPU periods. Goemans (Reference Goemans2022), using US data from 1890 onward, estimated a cumulative 1-year multiplier of 2 during uncertain periods, compared with 0.4–0.8 during normal times. Likewise, Rodrigues (2024) confirmed that fiscal effects become less intense when uncertainty reaches high levels, consistent with real options theory as agents become more cautious. Mao et al. (Reference Mao, Shen and Yang2024) showed that regime EPU dampens inflation and debt revaluation effects, reversing the debt decline that would otherwise occur under fixed policy regimes.

Several studies have extended the analysis to Japan, where persistent EPU and high public debt make the fiscal consequences of EPU especially salient. Montes and Salgado (Reference Montes and Salgado2024) found that EPU adversely affects business confidence in Japan, driven by both fiscal and monetary policy uncertainties. Arbatli Saxegaard et al. (Reference Arbatli Saxegaard, Davis, Ito and Miake2022) developed new EPU indices for Japan and showed that upward EPU foreshadows deteriorations in investment, employment, and output. Similarly, Fujitani et al. (Reference Fujitani, Hattori and Yasuda2023) noted that Japanese firms invest less and accumulate more cash when domestic EPU rises, with fiscal and exchange rate policy uncertainty driving the negative impacts. Lai and Yasuda (Reference Lai and Yasuda2025) also reported that Japanese banks respond to higher EPU by holding more government bonds, reducing stock holdings, and reallocating credit from small and medium-sized enterprises to large firms.

2.2. Pensions under EPU

The role of EPU in the pension context has been studied in two main strands. The first examines evidence outside Japan, focusing on reform timing, political factors, and pension fund performance, while the second centers on Japan, addressing pension reforms, aging, employment, retirement health, and several other related aspects.

Within the first strand, Andonoska and Xu (Reference Andonoska and Xu2023) confirmed that gubernatorial elections affect pension funding in US states, with plans better funded in election years. Kweon and Suzuki (Reference Kweon and Suzuki2022) showed that left incumbency relies more on public expenditure, while pension coverage is greater under right-leaning governments. Likewise, Dippel (Reference Dippel2022) demonstrated that benefit payments grow faster under Democratic mayors in US cities. Liu et al. (Reference Liu, Mikesell and Moldogaziev2021) noted that public corruption inversely relates to pension funding ratios, with reduced corruption saving states over 10% annually in pension benefits. Bagchi (Reference Bagchi2021) further reported that political pressure creates incentives for pension plans to use higher discount rates, artificially reducing reported liabilities by approximately 4.1%. Cumming et al. (Reference Cumming, Lu, Xu and Yu2024) similarly showed that US firms increase pension underfunding levels when facing higher EPU, effectively shifting risk to employees during uncertain times.

In contrast to the first strand, the existing literature on Japan’s pension does not link EPU to PPS. Instead, studies focus more on aging, pension reforms, and several other related aspects. For example, Kitao and Takeda (Reference Kitao and Takeda2025) established that raising the retirement age by 5 years would increase labor force participation among men in their sixties from 58% to 69%, while reducing the replacement rate by 20% would increase participation to 67%. Both reforms also encourage greater retirement savings. Yu (Reference Yu2024) showed that social security reforms account for over 77% of the observed rises in employment and hours worked by the 1950s cohort, with the retirement earnings test reforms being the most important. Kitamura and Nakashima (Reference Kitamura and Nakashima2025) similarly discovered that financial incentives effectively delay pension claiming. A 1.2% increase in pension benefits or a 19.4% additional yield on short-term insurance products delays claiming by 1 year on average, with insurance-based incentives proving more cost-effective for the government.

The literature also covers the automatic balancing mechanism, which has been shown to create trade-offs between benefit adequacy and financial sustainability in Japan. According to Kimura et al. (Reference Kimura, Kitamura and Nakashima2023), Japan’s automatic balancing mechanism makes moderate risk-taking optimal, as high risk raises benefit risk (benefits cut to restore solvency), while low risk locks benefits at a low level (insufficient returns). Beyond the design of specific mechanisms, uncertainty about the timing and structure of reform also carries costs. Kitao (Reference Kitao2018) modeled policy uncertainty and reform delay in a general equilibrium life-cycle model calibrated to Japan. The author found that reform delay leads individuals to save less, work less, and pay higher taxes to cover additional expenditures during the transition. The welfare losses from fiscal uncertainty fall most heavily on older individuals, who face greater income instability and weaker returns on retirement savings.

3.1. Political economy channel

Public choice theory provides the foundational framework in this study for understanding how EPU affects PPS. According to Buchanan and Tullock (Reference Buchanan and Tullock1962), political actors such as politicians, bureaucrats, and voters respond to incentives similarly to market participants, with self-interest dominating decision-making. The theory suggests that policymakers prioritize re-election and institutional power over long-term collective welfare. Given this, an increase in EPU can activate three political economy forces that may operate to affect the PPS. First, politicians facing uncertain policy environments may delay unpopular pension reforms, such as benefit cuts or contribution hikes, to avoid political backlash (Kitao, Reference Kitao2018; Weaver, Reference Weaver2019). In Japan, repeated delays in consumption tax hikes exemplify such political unwillingness. Second, bureaucrats may exploit periods of uncertainty to expand their institutional influence rather than optimize pension fund management, leading to inefficient resource allocation. Third, voters, particularly older cohorts who depend heavily on pension income, may oppose structural reforms, creating a ‘wait-and-see’ trap where short-term political costs outweigh long-term fiscal benefits (Teng et al., Reference Teng, Zhu, Nazar and Kanwal2024). However, countervailing forces may also operate, where the government may increase spending as a stabilizer to protect vulnerable populations during uncertain periods. Therefore, the net direction depends on the relative strength of the opposing forces, whether the political unwillingness to pursue unpopular cuts dominates, or whether the stabilizer motive overcomes. In the case of Japan, where fiscal policy dominates the EPU index, appearing in more than half of all EPU-related articles compared to only 24% for monetary issues (Arbatli Saxegaard et al., Reference Arbatli Saxegaard, Davis, Ito and Miake2022), the political economy channel carries particular weight. We test the net direction empirically through the coefficient on EPU in our main model, denoted as Hypothesis 1.

3.2. Macroeconomic channel

The second mechanism involves real economic outcomes as a transmission pathway. Baker et al. (Reference Baker, Bloom and Davis2016) established that EPU negatively affects employment, investment, and output at both firm and aggregate levels. At the firm level, firms with greater exposure to government purchases experience reduced investment rates and employment growth when EPU rises. At the macro level, EPU innovations foreshadow declines in investment (6%), industrial production (1.2%), and employment (0.35%). For Japan, Arbatli Saxegaard et al. (Reference Arbatli Saxegaard, Davis, Ito and Miake2022) confirmed similar patterns, showing that higher EPU lowers industrial production (by up to 2% in about a year) and also reduces employment.

In turn, macroeconomic deterioration directly transmits to PPS through the pension contribution base, inflation, and unemployment. Japan’s public pension system operates on a pay-as-you-go model, where current contributors fund current beneficiaries (JPS, 2024). However, the contributor base remains narrow and unevenly distributed across age-groups.Footnote ¹ The number of active contributors declines when EPU reduces employment, and fewer contributors mean lower contribution revenues, creating fiscal pressure that may lead policymakers to restrain PPS or reduce benefits.Footnote ² Nonetheless, one cannot rule out the possibility of an opposite effect. For instance, if EPU leads to higher inflation without simultaneously reducing employment, the purchasing power of pension benefits may decline. As noted in section 3.1, governments may also respond by increasing spending to protect household incomes during EPU-induced downturns (Ando et al., Reference Ando, Furukawa, Nakata and Sumiya2020). As a result, the net direction depends on whether the employment effect, the inflation effect, or the stabilizer effect dominates. We include employment and inflation rates as mediating variables to capture these competing forces.

3.3. Fiscal constraint channel

The third mechanism focuses on the possibility that the level of public debt conditions the relationship between EPU and PPS in Japan. Unlike the political economy channel, which emphasizes politician behavior, and the macroeconomic channel, which potentially operates through employment and inflation, this channel relates to the fiscal space within which policymakers operate. As highlighted by Baker et al. (Reference Baker, Bloom and Davis2016), governments face competing pressures when EPU rises. On one hand, high EPU may call for increased social spending, including pensions, to stabilize household incomes and maintain social cohesion. On the other hand, high public debt may constrain the government’s ability to expand discretionary spending, such as pensions. Japan’s general government debt-to-GDP ratio stood at 236% in 2024, one of the highest among advanced economies, which may encourage the government to prioritize fiscal consolidation over pension expansions when uncertainty is high (Beetsma et al., Reference Beetsma, Busse, Larch and Romp2026). The net effect of EPU on PPS may then depend on the level of public debt. At low-debt levels, the government may have room to increase spending as a stabilizer. When debt is high, fiscal constraints may dominate, leading to restrained or reduced PPS. According to Li and Wei (Reference Li and Wei2022) and Rodrigues ESFL (2024), fiscal policy effectiveness diminishes under high EPU, and Mao et al. (Reference Mao, Shen and Yang2024) also note that regime uncertainty dampens inflation and debt revaluation effects, making fiscal consolidation more difficult. We examine these directions through the moderating roles of the debt-to-GDP ratio and government effectiveness in our empirical analysis.

5.1. Data

We use annual data from 1997 to 2024, obtained from various sources, including the World Bank Development Indicators (WDI), the Organisation for Economic Co-operation and Development (OECD), the International Monetary Fund (IMF), and the Economic Policy Uncertainty database provided by Baker et al. (Reference Baker, Bloom and Davis2016). Although longer data series exist for some indicators, the analysis focuses on the 1997–2024 period to maintain balanced observations and consistent coverage across all variables.

5.2. Variables

5.2.3. Control variables

The study includes four control variables to isolate the effect of EPU while accounting for factors that influence PPS. Table 1 presents the variables selected for analysis. The control variables include life expectancy at age 65 (LE), defined as the average number of years a person of that age is expected to live, assuming constant age-specific mortality rates. GDP per capita (GDPPC) is included to capture the fiscal capacity to support PPS. Although the GDP growth rate was initially considered, negative values made it unsuitable for log transformation. In addition, the social security contribution rate (SSC) is included as a key policy variable influencing PPS, and net migration (MG) is incorporated to account for demographic changes that could affect PPS.

Table 1.

Data sources and variable details

Table 1 Long description

Columns list variable type, acronym, description, measurement, and source. The dependent variable is PPS, Public Pension Spending, measured as public pension spending as percent of gross domestic product. The independent variable is EPU, Economic Policy Uncertainty, measured by the Economic Policy Uncertainty Index. Control variables include LE, Life expectancy, measured as life expectancy at age 65; MG, Net Migration, measured as net migration; GDPPC, Gross Domestic Product Per Capita, measured as gross domestic product per capita in current US dollars; and SSC, Social Security Contribution, measured as social security contribution as percent of gross domestic product. Moderating variables include GOE, Government Effectiveness, measured by the Government Effectiveness Index, and DGR, Debt to GDP Ratio, measured as general government debt as percent of gross domestic product. Mediating variables include INF, Inflation Rate, measured as inflation consumer prices annual percent, and EMP, Employment of those aged 15 to 64 years old, measured as percent of the 15 to 64 age group.

5.3. Regression model

In line with Karimi and Fan (Reference Karimi and Fan2026) and Kweon and Suzuki (Reference Kweon and Suzuki2022), the regression model is specified as follows:

(1)\begin{equation} \begin{split} PP{S_t} & = \alpha + \mathop \sum \limits_{i = 1}^p {\gamma _i}PP{S_{t - i}} + \mathop \sum \limits_{i = 1}^q {\beta _i}EP{U_{t - i}} + \mathop \sum \limits_{i = 1}^q {\theta _i}L{E_{t - i}} + \mathop \sum \limits_{i = 1}^q {\delta _i}M{G_{t - i}} + \mathop \sum \limits_{i = 1}^q {\sigma _i}GDPP{C_{t - i}} \\ & \quad {}+ \mathop \sum \limits_{i = 1}^q {\varphi _i}SS{C_{t - i}} + {\varepsilon _t} \end{split} \end{equation}

where PPS denotes public pension spending, EPU refers to economic policy uncertainty, LE represents the life expectancy at age 65, MG signifies net migration, GDPPC is the GDP per capita, and SSC stands for the social security contribution rate. The subscript t indicates the period, spanning from 1997 to 2024, and t–i represents the lagged terms, where p is the lag length for the PPS and q is the lag length for all independent variables. The parameter α denotes the intercept, and ${\varepsilon _t}$ is the error term. To avoid multicollinearity, some variables are excluded from the analysis. The old-age dependency ratio and the proportion of the elderly population are excluded due to high multicollinearity with the SSC rate. Instead, LE at age 65 is retained as a more direct measure of longevity and retirement duration. Wage growth is not included because of its strong correlation with both inflation and GDPPC, and the labor force participation rate is left out due to its redundancy with the employment rate.

The study employs the Cobb–Douglas framework to model how EPU influences PPS, building on the methodological extensions adopted by Durani et al. (Reference Durani, Bhowmik, Sharif, Anwar and Syed2023) and Li (Reference Li2023). The standard form of the function is:

(2)\begin{equation}{\text{Y}} = {\text{ A}}.{L^{{\alpha }}}.{K^{{\beta }}}\end{equation}

where Y represents total output, and labor and capital (input factors) are denoted by L and K. The parameters α and β capture the output elasticities of labor and capital, respectively, and A reflects total factor productivity, commonly interpreted as technological progress. The Cobb–Douglas framework assumes a predictable and stable environment, particularly for firm-level production. However, regulatory shifts, political instability, ambiguity in fiscal policy, and other factors can increase uncertainty for both governments and firms, as noted by Baker et al. (Reference Baker, Bloom and Davis2016). To reflect this, we incorporate EPU into the production function, consistent with studies that include variables such as EPU (Durani et al., Reference Durani, Bhowmik, Sharif, Anwar and Syed2023), digital finance (Diao, Zhao, Xie, & Liu, Reference Diao, Zhao, Xie and Liu2025), etc. The augmented Cobb–Douglas function becomes:

(3)\begin{equation}{\text{Y}} = {\text{ A}}.{L^{{\alpha }}}.{K^{{\beta }}}.{E^{{\gamma }}}\end{equation}

Based on the extension in Equation (3), PPS is reformulated as a function of EPU and other explanatory variables introduced in Equation (1), yielding the following specification:

(4)\begin{equation}PP{S_t} = A{(EP{U_t})^{\beta 1}}.{(L{E_t})^{\beta 2}}.{(M{G_t})^{\beta 3}}.{(GDPP{C_t})^{\beta 4}}.{(SS{C_t})^{\beta 5}}.{\varepsilon _t}\end{equation}

To linearize the multiplicative form in Equation (4) and obtain the final specification, the equation is transformed using the natural logarithm, following the method applied by Li (Reference Li2023). As a result, the transformed model is expressed as follows:

(5)\begin{equation}lnPP{S_t} = {\beta _0} + {\beta _1}lnEP{U_t} + {\beta _2}lnL{E_t} + {\beta _3}lnM{G_t} + {\beta _4}lnGDPP{C_t} + {\beta _5}lnSS{C_t} + {\varepsilon _t}\end{equation}

where ${{{\beta }}_0}$ is the intercept and the coefficients ${{{\beta }}_1}to{\beta _5}$ are the slope coefficients corresponding to each explanatory variable.Footnote ³ The log-transformed specification, derived from the Cobb–Douglas function, is adopted to address large differences in scale across the variables and to leverage the advantages of logarithmic transformation. In our dataset, PPS as a share of GDP is relatively small, ranging from 6% to 11%, while EPU and other explanatory variables, such as GDPPC, span much larger ranges. Direct estimation in levels could produce coefficients that are difficult to interpret and potentially unstable, as large differences in variable scales may increase heteroscedasticity and reduce the reliability of the estimates. Applying the logarithmic transformation mitigates scale differences, reduces skewness and heteroscedasticity, and allows coefficients to be interpreted as percentage changes. In addition, the transformation improves estimation stability and reduces the risk of biased or inconsistent estimates (Wooldridge, Reference Wooldridge2019). Therefore, the log-transformed specification is adopted to provide more reliable estimates of the relationship between EPU and PPS.

5.2.6. Estimation strategy

The estimation proceeds in several stages, starting with preliminary econometric tests and moving to long-run estimation and subsequent analysis. A brief intuitive description of the preliminary econometric tests, such as stationarity and cointegration tests, is presented in the Appendix. For detailed technical procedures, refer to the sources cited in each corresponding description.

After conducting the preliminary econometric tests, we estimate the long-run relationship between EPU and PPS, which constitutes the core focus of this study, using a common autoregressive distributed lag (ARDL) model (Pesaran and Shin, Reference Pesaran, Shin and Strøm1999) and kernel-based regularized least squares (KRLS), a machine learning method (Hainmueller and Hazlett, Reference Hainmueller and Hazlett2014). The ARDL approach is well-suited for small samples and allows for a mix of I (0) and I (1) variables, unlike traditional cointegration techniques such as the Engle–Granger or Johansen methods, which require all variables to be integrated of the same order. Also, ARDL offers flexibility in lag selection and can simultaneously estimate short- and long-run dynamics, while its incorporation of lagged values captures dynamic adjustments over time (Pesaran and Shin, Reference Pesaran, Shin and Strøm1999).

Nonetheless, like most common time-series models, ARDL assumes a predefined linear relationship among variables. However, the relationships between variables may not always follow a predetermined pattern because real-world data often involve complex interactions. Factors such as diminishing returns, structural shifts, threshold effects, policy regimes, behavioral changes, etc., can generate varying patterns in economic data and allow variables to interact in ways that cannot be sufficiently captured by a straight-line model (Granger and Teräsvirta, Reference Granger and Teräsvirta1993). Similarly, some relationships may exhibit curvature, periodicity, or abrupt changes that require polynomial terms, alternative functional forms, or other advanced modeling techniques to accurately represent the underlying patterns. To address this, we employ KRLS, a machine learning method that does not impose a predefined functional form and learns patterns directly from the data by combining kernel methods with regularization. KRLS also allows marginal effects to vary across observations, which makes it useful for uncovering heterogeneous effects that the ARDL model may overlook (Hainmueller and Hazlett, Reference Hainmueller and Hazlett2014).

Moreover, after the long-run estimation, the study examines the quantile effects derived from the KRLS model to assess how the impact of EPU on PPS varies across the distribution of EPU. The analysis also considers the post-Abenomics period (2012–2024) by introducing an interaction term between EPU and a post-Abenomics dummy. Subsequently, the study examines the impact of major global events, such as the financial crisis and the COVID-19 pandemic, to assess whether the effect of EPU on PPS intensifies during periods of heightened global uncertainty. Besides, the two-stage least squares (2SLS) instrumental variable (IV) approach is employed to address potential endogeneity, followed by robustness and sensitivity analyses using an alternative proxy for the main variable and different model specifications.

6.1. Descriptive analysis

Table 2 presents the descriptive statistics of the main variables. PPS as a percentage of GDP has a mean of 9.13%, with values ranging from 6.31% to 11.09%, whereas EPU exhibits noteworthy variability, with a mean of 109.87 and a standard deviation of 25.99. LE averages 21.16 years, showing minimal fluctuation with a standard deviation of 0.89, and MG (in thousands) has a mean of 123.59, with a single negative value in 2021, which may be attributed to COVID-19-induced emigration. GDPPC, expressed in thousands of US dollars, averages 38.27, with values from 32.42 to 49.14, and SSC averages 11.03%, ranging from 8.82% to 13.32%.

Table 2.

Descriptive statistics

Table 2 Long description

Descriptive statistics are reported for six variables labeled PPS, EPU, LE, MG, GDPPC, and SSC, with rows for mean, standard deviation, minimum, maximum, and observations. Means are 9.13 for PPS, 109.87 for EPU, 21.16 for LE, 123.59 for MG, 38.27 for GDPPC, and 11.03 for SSC. Standard deviations are 1.12 for PPS, 25.99 for EPU, 0.89 for LE, 95.31 for MG, 4.29 for GDPPC, and 1.63 for SSC. Minimum values are 6.31 for PPS, 64.79 for EPU, 19.40 for LE, minus 34.99 for MG, 32.42 for GDPPC, and 8.82 for SSC. Maximum values are 11.09 for PPS, 183.46 for EPU, 22.45 for LE, 322.56 for MG, 49.14 for GDPPC, and 13.32 for SSC. Each variable lists 28 observations.

Note: The EPU index is from Arbatli Saxegaard et al. (Reference Arbatli Saxegaard, Davis, Ito and Miake2022), originally available on a monthly basis. Annual values are calculated by taking the average of the 12-monthly observations for each year. For clarity, MG is presented in thousands of people and GDPPC in thousands of US$. The time series spans the period from 1997 to 2024.

Figure 1 plots the trajectory of Japan’s PPS, alongside both Japan’s EPU and global EPU, across the sample period. Panel A plots annual PPS as a % of GDP in Japan and shows a gradual increase in PPS from 1997, reaching a peak around 2010. After 2012, the trend slightly declines and then stabilizes, with visible increases observed after the early 2000s and again in the period following 2007. Panel B visualizes the annual EPU index for Japan, and Panel C visualizes the global EPU index. Both panels show a general upward trend, with notable peaks around 2008–2009 and during COVID-19, corresponding to major global events.

Figure 1.

Annual trends in PPS and EPU (1997–2024).

Figure 1 Long description

A line graph titled, Public Pension Spending in Japan, 1997 to 2024. The horizontal axis label is, Year, with values from 1997 to 2024. The vertical axis label is, percent of gross domestic product, with values from 0 to 11. A single line with circular markers rises from about 5 in 1997 to about 9 by 2007, reaches about 10 around 2009 to 2011, then stays near 9 to 10 through 2024. The image B showing a line graph titled, Economic Policy Uncertainty in Japan, 1997 to 2024. The horizontal axis label is, Year, with values from 1997 to 2024. The vertical axis label is, Index Value, with values from 80 to 180. A single line with triangular markers fluctuates between about 90 and about 160 across the years, with higher points around the late 1990s, around 2008 to 2012 and several peaks after 2015, ending near about 120 to 140 by 2024. The image C showing a line graph titled, Global Economic Policy Uncertainty, 1997 to 2024. The horizontal axis label is, Year, with values from 1997 to 2024. The vertical axis label is, Index Value, with values from 50 to 300. A single line with diamond markers stays near about 80 to 120 from 1997 through the mid 2000s, rises to about 150 to 200 around 2008 to 2012, increases further after 2016, reaches a peak near about 300 around 2020, then declines and ends near about 200 to 250 by 2024.

Figure 2 presents ridge plots of monthly EPU distributions by year, where Panel A focuses on the Japan EPU and Panel B focuses on the global EPU. Panel A shows that the monthly EPU level varies over time in Japan. In the late 1990s and early 2000s, the EPU values mostly clustered at lower to moderate levels, whereas after 2008, the distributions become wider and peak values shift higher, indicating increased policy uncertainty. Similarly, the global EPU distribution (Panel B) mostly trends upward over time, with clearer increases from 2008 onwards.

Figure 2.

Monthly EPU distributions by year (1997–2024).

Figure 2 Long description

Distribution of Monthly EPU Japan. A ridge density plot of monthly EPU values by year. The y-axis label is, Year, with labeled ticks at 1997, 1999, 2001, 2003, 2005, 2007, 2009, 2011, 2013, 2015, 2017, 2019, 2021 and 2023. The x-axis label is, EPU, with labeled ticks at 0, 100, 200, 300 and 400. Units are not specified. A color scale labeled, EPU Level, shows tick labels 50, 100, 150, 200 and 250. Each year has one ridgeline distribution. Earlier years show ridgelines concentrated nearer the 0 to 100 range, while later years show ridgelines with more density beyond 100 and longer right tails extending beyond 200, with some tails approaching 400. Distribution of Monthly EPU Global. A ridge density plot of monthly GEPU values by year. The y-axis label is, Year, with labeled ticks at 1997, 1999, 2001, 2003, 2005, 2007, 2009, 2011, 2013, 2015, 2017, 2019, 2021 and 2023. The x-axis label is, GEPU, with labeled ticks at 0, 100, 200, 300 and 400. Units are not specified. A color scale labeled, GEPU Level, shows tick labels 100, 200, 300 and 400. Each year has one ridgeline distribution. Earlier years show ridgelines concentrated nearer the 0 to 100 range, while later years show ridgelines with more density beyond 100 and longer right tails extending beyond 200, with some tails approaching 400.

Note: Each vertical band represents a year, and the shape of the ridge shows the distribution of monthly EPU values within that year. The x-axis indicates the EPU level, while color intensity reflects the value range, from lower (blue) to higher (red) uncertainty levels.

6.2. Main results

The long-run estimates from both the ARDL and KRLS models in Table 3, which are our main results, reveal a negative association between EPU and PPS in Japan. Specifically, a 1% increase in the EPU index is associated with a 0.127% decrease in PPS in the ARDL model and a somewhat larger 0.203% decrease in the KRLS model. LE exhibits a strong positive association, with each 1% rise linked to a 0.313% increase in PPS under ARDL and a 0.336% increase under KRLS. Likewise, SSC shows a positive association, contributing around 0.231% (ARDL) and 0.286% (KRLS) to PPS growth per 1% increase. MG displays a negative association, where a 1% increase is associated with a 0.095% decrease in PPS in the ARDL model. GDPPC has a modest positive link, with a 1% rise associated with a 0.171% (ARDL) and 0.178% (KRLS) increase in PPS.

Table 3.

Main results: long-run estimation

Table 3 Long description

Long run coefficients are reported for two models, Autoregressive Distributed Lag and Kernel based Regularized Least Squares, with a coefficient column for each. In Autoregressive Distributed Lag, coefficients are EPU minus 0.127 with three stars and standard error 0.016, LE 0.313 with two stars and 0.129, MG minus 0.095 with one star and 0.044, GDPPC 0.171 with two stars and 0.072, and SSC 0.231 with three stars and 0.047. Autoregressive Distributed Lag also lists R2 0.84, Max. Lags 1, Constant or Trend Yes, and N 28. In Kernel based Regularized Least Squares, coefficients are EPU minus 0.203 with three stars and 0.032, LE 0.336 with two stars and 0.136, MG minus 0.081 with two stars and 0.036, GDPPC 0.178 with three stars and 0.045, and SSC 0.286 with three stars and 0.073. Kernel based Regularized Least Squares also lists Lambda 0.49, Tolerance 0.028, R2 0.73, and N 28. Standard errors appear in parentheses and stars appear next to several coefficients.

Note: Standard errors are in parentheses. Lambda (λ) controls regularization (higher λ = more smoothing). Tolerance sets convergence precision (smaller = stricter).

***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

6.3. Validation analysis

IV Approach: One of the fundamental issues in economic research is endogeneity, which may arise from omitted variable bias, measurement error, or reverse causality (Wooldridge, Reference Wooldridge2019). Endogenous relationships compromise the reliability of standard estimation techniques, producing biased and inconsistent parameter estimates. In response, scholars have developed several methodological remedies, including the IV approach, such as 2SLS, which resolves the problem by isolating the exogenous component of the endogenous variable. A key requirement of IV estimation is the availability of a valid instrument.Footnote ⁴ Following the IVs employed in the existing literature, we adopt EPU with a two-period lag and the EPU of the United States as IVs for our model, consistent with Dirks and Schmidt (Reference Dirks and Schmidt2024) and Cheng and Masron (Reference Cheng and Masron2023).

The U.S. EPU index is a potentially strong IV for this study, given the close trade and political linkages between the two countries. Strong comovement exists between U.S. and Japanese EPUs, as highlighted by Arbatli Saxegaard et al. (Reference Arbatli Saxegaard, Davis, Ito and Miake2022), who show that the Japanese EPU responds to developments in the U.S. Also, Fujitani et al. (Reference Fujitani, Hattori and Yasuda2023) show that the U.S. EPU negatively affects corporate investment in Japan. As a result, variations in U.S. EPU induce corresponding changes in Japan’s domestic policy uncertainty, supporting the relevance condition for a valid IV. Besides, the transmission of U.S. EPU occurs largely through economic expectations rather than direct fiscal channels, as shown by Montes and Salgado (Reference Montes and Salgado2024). Japanese firms and households adjust their investment, consumption, and hiring decisions in anticipation of changes in global trade, exchange rates, and financial conditions, with particular sensitivity to developments originating in the U.S. (Fujitani et al., Reference Fujitani, Hattori and Yasuda2023; Montes and Salgado, Reference Montes and Salgado2024). The resulting mechanism links U.S. EPU to Japan’s overall economic sentiment and shapes domestic EPU. However, the exclusion restriction requires that U.S. EPU affect Japan’s PPS only through Japanese EPU. This restriction is an identifying assumption that cannot be directly tested. The potential effects of U.S. EPU on Japanese macroeconomic conditions, as detailed in section 3.2, represent a limitation that should be considered when interpreting the 2SLS estimates.

Table 4 presents the 2SLS estimates, which align closely with the baseline KRLS estimates in Table 3 and provide a supplementary robustness check.

Table 4.

IV approach

Table 4 Long description

Results are reported with columns for Coefficient and Standard Error, followed by model tests with columns for Statistic and p-value. For EPU, the coefficient is minus 0.186 with three asterisks, and the standard error is 0.029. Wald chi2 has a statistic of 72.70 and a p-value of 0.000. Sargan Over-identification Test has a statistic of 0.78 and a p-value of 0.376. Joint significance of IVs indicated by F- statistic is 15.19 with a p-value of 0.000.

Note: Wald chi² indicates joint significance of the model (p-value < 0.01). The Sargan test does not reject the overidentifying restrictions (p-value = 0.376). The first-stage F-statistic indicates instrument strength (F = 15.19 > 10).

*** denotes significance at the 1% level.

Robustness Checks and Sensitivity Analysis: For the robustness check, the study first replaces the baseline EPU measure with an alternative index developed by Ahir et al. (Reference Ahir, Bloom and Furceri2022), which covers multiple countries, including Japan. The study then introduces an additional control variable, the long-term interest rate, to further account for macroeconomic influences on PPS and test the robustness of the results. At the last stage, the robustness of the ARDL results is checked by varying the lag structure to assess the sensitivity of the results to lag selection. We also perform diagnostic tests, including autocorrelation, normality, heteroscedasticity, and model stability tests using the cumulative sum (CUSUM) and CUSUM of squares tests, to confirm that the estimated parameters remain stable over time.

The results of the robustness tests, presented in Table 5, confirm the stability of our core findings in Table 3. The alternative EPU measure shows a consistent negative association, ranging from 0.212% (KRLS) to 0.135% (ARDL) reduction in PPS per 1% increase in the EPU index. Similarly, the diagnostic tests validate the results of the ARDL model, showing no evidence of autocorrelation, deviation from normality, and heteroskedasticity. The CUSUM tests further confirm the stability of the ARDL model results.

Table 5.

Robustness checks and diagnostic tests

Table 5 Long description

Panel A reports robustness check results with columns for coefficient and standard error, shown separately for krls and ardl. Under krls, alternative epu index has coefficient minus 0.212 with three stars and standard error 0.058, and adding more controls has coefficient minus 0.207 with three stars and standard error 0.033. Under ardl, alternative epu index has coefficient minus 0.135 with three stars and standard error 0.020, adding more controls has coefficient minus 0.131 with three stars and standard error 0.025, and changing lags has coefficient minus 0.129 with two stars and standard error 0.053. Panel B lists diagnostic tests for ardl with columns for statistic and p value. Breusch Godfrey lm test shows statistic 2.436 and p value 0.267, Jarque bera test shows statistic 0.271 and p value 0.814, and Breusch pagan godfrey test shows statistic 0.294 and p value 0.767. Cusum test is labeled stable with no p value shown, and cusum square test is labeled stable with no p value shown.

Note: The alternative EPU index is obtained from Ahir et al. (Reference Ahir, Bloom and Furceri2022). Since the index is normalized between 0 and 1, it is multiplied by 1,000 to make it comparable to the main index. The long-term interest rate is added as an additional control variable, and the number of lags is increased from 1 to 2 as robustness checks.

***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

6.4. Mediating and moderating analysis

We incorporate two mediating variables, the inflation rate (INF) and the employment rate (EMP), to examine the channels through which EPU influences PPS. EPU can affect INF by altering expectations, investment decisions, and monetary policy responses, which in turn impact the real value of pension benefits, especially when they are indexed to prices. Similarly, heightened EPU can reduce business confidence and slow job creation, thereby affecting EMP levels (Baker et al., Reference Baker, Bloom and Davis2016). A lower EMP rate may also reduce the number of contributors to the pension system, which would increase the fiscal burden.

The study further considers two moderating factors that may influence the strength or direction of the relationship between EPU and PPS. The government effectiveness index (GOE), which measures the capacity of public institutions to design and implement sound policies, conditions the extent to which EPU affects fiscal planning and PPS. Also, the high debt-to-GDP ratio (DGR), which represents the structural fiscal position, may amplify the sensitivity of PPS to EPU.

Table 6 reports the results of the moderating and mediating analyses. As shown in Panel A, the interaction term between EPU and DGR is negative (-0.074), indicating that the association between EPU and PPS becomes more negative as DGR increases. In contrast, the interaction term between EPU and GOE is positive (0.106), demonstrating that the EPU-PPS association becomes more positive as GOE increases.

Table 6.

Moderating and mediating analysis

Table 6 Long description

Coefficients and standard errors are reported for a moderating panel and a mediating panel, with outcomes labeled PPS and mediators labeled INF or EMP. Panel A lists DGR by EPU with coefficient minus 0.074 with three stars and standard error 0.019, and DGR with coefficient 0.139 with three stars and standard error 0.038. Panel A also lists GOE by EPU with coefficient 0.106 with two stars and standard error 0.043, and GOE with coefficient 0.053 with three stars and standard error 0.014. Panel B has three columns labeled PPS, INF or EMP, and PPS. In the INF rows, EPU is minus 0.203 with three stars and standard error 0.032 in the first PPS column, 0.216 with two stars and standard error 0.087 in the INF or EMP column, and minus 0.179 with three stars and standard error 0.049 in the second PPS column; INF in the second PPS column is minus 0.107 with one star and standard error 0.052. In the EMP rows, EPU is minus 0.203 with three stars and standard error 0.032 in the first PPS column, minus 0.177 with three stars and standard error 0.045 in the INF or EMP column, and minus 0.167 with three stars and standard error 0.045 in the second PPS column; EMP in the second PPS column is 0.195 with two stars and standard error 0.083.

Note: All results are based on the KRLS model. Interaction terms in Panel A examine whether the association between EPU and PPS varies with the respective moderating variables. A positive (negative) interaction coefficient indicates that the EPU–PPS association is more positive (negative) at higher values of the respective moderating variable. In Panel B, Column 1 reports the total effect of EPU on PPS without controlling for the mediator (INF or EMP). Column 2 estimates the effect of EPU on the mediator, and Column 3 includes the mediator in the model. Standard errors appear in parentheses.

***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

Moreover, the mediation pathways presented in Panel B show that a 1% increase in the EPU index is associated with a 0.216% increase in the INF rate, which in turn is associated with a 0.107% decrease in PPS. The direct effect of EPU on PPS then declines to 0.179% after including INF in the model. The same approach used for MG was applied to INF, with a constant added to the INF series to ensure that all observations were positive before applying the natural logarithmic transformation. Furthermore, EPU is negatively associated with EMP (-0.177%), while EMP shows a positive association with PPS (0.195%). As a result, including EMP in the model lowers the direct effect of EPU on PPS to 0.167%.

6.5. Heterogeneity analysis

Table 7 and Figure 3 present the heterogeneity analyses. In Table 7, Panel A focuses on short-run associations, Panel B reports the results of major events, and Panel C reports the post-Abenomics analysis. Figure 3 subsequently presents the quantile-based results from the KRLS model.

Figure 3.

Distributional effects of EPU on PPS.

Note: The results are derived from the KRLS model. Circle size represents the magnitude of the EPU coefficient across quantiles, with larger circles indicating higher coefficient values.

Table 7.

Heterogeneity analyses

Table 7 Long description

Coefficients and standard errors are reported for three panels: Short Run Effect with ARDL, Major Events Effects, and Post-Abenomics Effect. Panel A lists EPU with coefficient 0.043 and standard error 0.017, LE 0.117 and 0.031, MG minus 0.025 and 0.013, GDPPC 0.112 and 0.045, and SSC 0.153 and 0.059; significance markers appear next to each coefficient. Panel B lists EPU interaction with Financial Crisis 2008 to 2009 with coefficient minus 0.251 and standard error 0.071, and EPU interaction with COVID 19 Pandemic 2019 to 2020 with coefficient minus 0.237 and standard error 0.062; both coefficients have three asterisks. Panel C lists EPU interaction with Post-Abenomics 2012 to 2024 with coefficient minus 0.064 and standard error 0.033, with one asterisk. Column headers are, Coefficient and, Standard Error.

Note: Results presented in Panels B and C are derived from the KRLS model. The event indicators in Panel B equal 1 during 2008–2009 for the Financial Crisis and during 2019–2020 for the COVID-19 Pandemic, and 0 otherwise. In Panel C, the post-Abenomics indicator equals 1 for 2012–2024 and 0 otherwise. The interaction terms test whether the EPU–PPS association differs during the specified periods. A positive (negative) interaction coefficient indicates a more positive (negative) EPU–PPS association during the specified period than during the reference period.

***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

The short-run results reveal different patterns compared to the long-run results displayed in Table 3. A 1% increase in the EPU index is associated with a 0.043% increase in PPS in the short term, while LE maintains its strong positive association (0.117% per 1% increase). Panel B shows the impact of two major shocks – the global financial crisis and the COVID-19 pandemic. The results show that the association between EPU and PPS was more negative during both event periods, with interaction coefficients of -0.251 for the financial crisis and -0.237 for the COVID-19 pandemic. Moreover, the results reported in Panel C show that the negative association between EPU and PPS differs between the pre- and post-Abenomics periods. The interaction term, estimated at -0.064, is significant at the 10% level and indicates that the EPU–PPS association is 0.064 more negative in the post-Abenomics period than in the pre-Abenomics period. Figure 3 displays that the negative association between EPU and PPS varies across EPU quantiles. At the first quantile, a 1% increase in the EPU index is associated with a 0.194% decrease in PPS, growing to 0.208% at the median and reaching 0.253% at the third quantile.

6.6. Discussion of findings

The findings in Table 3 reveal that higher EPU is associated with a decrease in PPS in Japan. We provide further discussion of the results as follows.

1. 1. The negative long-run association between EPU and PPS in Table 3 supports the political economy and fiscal constraint channels outlined in sections 3.1 and 3.3. Consistent with the framework, the results suggest that Japan’s high public debt and aging population leave limited fiscal space to absorb EPU without restricting PPS. The employment and inflation mediation results (Table 6) further corroborate the macroeconomic channel (section 3.2), with EPU reducing the contribution base and compounding the direct fiscal pressures. Prior evidence shows that policy priorities shift toward immediate stabilization during uncertainty (Weaver, Reference Weaver2019; Wen et al., Reference Wen, Khalid, Mahmood and Yang2022), and Japan’s COVID-19 response, where public debt rose from 236% to 258% of GDP, exemplifies how immediate relief can crowd out longer-term commitments like PPS (Ando et al., Reference Ando, Furukawa, Nakata and Sumiya2020; Okamoto, Reference Okamoto2025).

2. 2. The KRLS model in Table 3 reveals a somewhat larger negative association than the ARDL model. Unlike ARDL, the KRLS method does not impose a predefined functional form and instead identifies patterns directly from the data. Moreover, the quantile analysis (Figure 3) shows variation in the association between EPU and PPS across the EPU distribution. The estimated negative association appears more pronounced at higher EPU levels, although this pattern cannot be formally tested within the KRLS framework. The observed pattern is consistent with the fiscal constraint channel (section 3.3), since PPS is high relative to GDP in Japan, leaving the government with tighter fiscal trade-offs and limited room to absorb additional EPU without adjustment. Japan’s rising old-age dependency ratio and shrinking contributor base have already placed upward pressure on PPS, which may contribute to the observed pattern at higher EPU levels. The important implication is that policymakers may benefit from building fiscal buffers during periods of relatively low EPU to strengthen fiscal capacity when uncertainty rises. Specifically, decision-making processes within the Japan Pension Service and the Ministry of Health, Labour and Welfare could be streamlined so that PPS adjustments are not delayed or amplified by bureaucratic uncertainty during crises. Improving fiscal communication through clear and binding schedules for PPS reviews could reduce uncertainty surrounding future pension policy. Policymakers could also prioritize debt reduction during low EPU periods to strengthen fiscal capacity and limit the need for abrupt fiscal adjustments during periods of heightened uncertainty.

1. 3. The post-Abenomics analysis reported in Table 7, Panel C, is consistent with the three factors identified in section 4, namely weakened fiscal credibility, the pension system’s exposure to policy reversals, and deepening demographic pressures. The results suggest that these factors, operating in combination, made PPS more sensitive to EPU in the post-Abenomics period than before. Moreover, both the global financial crisis and the COVID-19 pandemic produced larger negative effects on PPS compared to the average long-run effect in Table 3. Japan’s GDP contracted by -5.69% in 2009 and -4.15% in 2020, which reflects the severity of both shocks. Extreme EPU events leave governments with little choice but to redirect fiscal resources toward immediate stabilization, including economic relief and public health measures, at the expense of long-term commitments like pensions (Ando et al., Reference Ando, Furukawa, Nakata and Sumiya2020).

2. 4. Our findings in Table 3 align with Mumtaz and Ruch (Reference Mumtaz and Ruch2025), Li and Wei (Reference Li and Wei2022), Imrohoroglu et al. (Reference Imrohoroglu, Kitao and Yamada2019), and Kitao (Reference Kitao2018), who report that government spending tends to decline during periods of EPU unless supported by a higher consumption tax. Our short-run results align with Ando et al. (Reference Ando, Furukawa, Nakata and Sumiya2020), confirming that government spending rises in response to EPU in the short term. However, our long-run findings diverge from the conclusion that uncertainty and shocks, including the COVID-19 crisis, may lead to sustained increases in government spending, which may include PPS (Nguyen and Schinckus, Reference Nguyen and Schinckus2020). We argue that Japan’s ageing population, high public debt, and shrinking pension contributor base, despite policy efforts, suggest that EPU is likely to restrict PPS. As noted by the United Nations (2023b), ageing societies encounter fiscal pressure due to higher healthcare and old-age support costs, coupled with declining revenues from a smaller working-age population. The current environment, when combined with higher EPU, is unlikely to support long-term growth in PPS. In addition, although Japan has a mechanism to automatically adjust pension benefits in line with inflation and wage growth, the mechanism has rarely been activated. Since its introduction in 2004, the system has triggered an adjustment only twice, in 2015 and 2021, due to persistently low inflation to meet the activation threshold (Kitao, Reference Kitao2018). The absence of full price indexation means benefits have not kept pace with rising living costs. Also, the system reviews benefit levels only at 5-year intervals based on local cost-of-living surveys, which limits timely responses to rising EPU (OECD, 2023). Therefore, due to institutional rigidities and the persistent fiscal pressures of high debt and demographic changes, rising EPU is unlikely to translate into higher PPS over the long run.

7.1. Limitations and future research directions

While this study employed the aggregate EPU index, future research could utilize the available disaggregated indices, such as fiscal, monetary, and trade/exchange rate uncertainty, to assess their distinct impacts on PPS. The analysis also focuses solely on Japan, which limits the generalizability of the findings. Comparative studies involving other aging and high-debt economies, such as Italy or Greece, could clarify whether the effects of EPU are country-specific or generalizable.

Moreover, despite the inclusion of relevant control variables and the use of 2SLS estimation, we cannot completely rule out omitted variable bias. Therefore, we interpret our findings as correlational evidence rather than conclusive proof of causality. Future research can build on this study and employ suitable causal methods to further explore causal relationships. In addition, the analysis does not formally test whether differences across distributional points are statistically significant, as KRLS does not provide quantile-specific standard errors or p-values. Future research could address this limitation using alternative estimators or bootstrap-based methods that allow formal inference on distributional heterogeneity.

Furthermore, the relatively small sample size, due to data constraints, may limit statistical power and reduce the precision of estimates, particularly in heterogeneity analysis. Accordingly, the heterogeneity results should be interpreted as exploratory rather than confirmatory. Future work could benefit from longer time-series or higher-frequency data, where available, to enable more formal testing of heterogeneous effects. Finally, while we identify three primary channels linking EPU to PPS, other mechanisms such as individual retirement behavior, pension fund performance, etc., remain untested in this study due to data availability constraints. Future research with microlevel or individual-level data can examine these additional channels.