2.1 Aggregate saving rate

The contributions of workers to a PAYG system represent a pure tax, because they are used immediately to pay pension benefits to retirees. In contrast, the pension premiums in a funded system constitute savings that get invested in the capital market. A shift toward more funding therefore might increase the aggregate saving rate and, accordingly, may increase economic growth. For this effect to be operative in practice, three conditions must be fulfilled (Barr, Reference Barr2000): funding must lead to a higher saving rate than PAYG, the additional savings must induce more investment, and additional investments must lead to a higher economic growth rate.

Although the effect of funding on saving could be permanent, it might be highest during the transition from the PAYG system to the funded system. That is, during the transition, funds build-up and this leads to a net increase in pension fund assets. At some point in time although, the pension fund matures, and the net inflow of funds slows or even becomes negative, such that the outflow of funds to pension beneficiaries creates dissaving.

A few other issues may play a role as well. The transition to a funded system must be financed by either the government or the workers. In the former case, the government issues debt that might undo the possible positive effect on the aggregate saving rate. Furthermore, as Blanchard and Fischer (Reference Blanchard and Fischer1989) point out, a funded pension system can increase the aggregate saving rate only if the pension fund forces people to save more than they would have saved voluntarily. If saving rates are already high, people simply replace part of their voluntary saving with mandatory pension saving, and the aggregate saving rate might stay the same.

Davis and Hu (Reference Davis and Hu2006) argue that financial liberalization affects the relationship between funding and saving. That is, households living in countries with a relatively repressed financial sector might face liquidity constraints, so in these countries forced saving should have a greater effect on the national saving rate than it does in countries with a more liberalized financial system. However, there liquidity constraints could play a role among low-income households.

The question of whether funding leads to more saving has been examined extensively in the empirical literature. The comprehensive international study of Reisen and Bailliu (Reference Reisen and Bailliu1997) considers the link between pension fund assets and saving rates, using data from 11 OECD- and non-OECD countries. They conclude that the accumulation of pension fund assets has a positive and significant impact on private saving, although the effect is eight times greater for non-OECD countries than for OECD countries. Lopez-Murphy and Musalem (Reference Lopez-Murphy and Musalem2004) test whether the accumulation of pension fund assets influences national saving rates; they find that it increases national saving when the funds are part of a mandatory pension program but decreases it if the pension funds represent public programs to foster voluntary pension saving. Bosworth and Burtless (Reference Bosworth and Burtless2004) also provide evidence that pension saving substitutes for other forms of private saving in OECD countries.

According to Samwick (Reference Samwick2000), the effect of funding on saving depends primarily on the way the government chooses to finance the transition. There is a cohort that has paid pension premiums into a PAYG system, but will not be able to collect benefits from it. That is, the cohort immediately after them pays its pension premiums into the new funded system (i.e., their contributions are invested in the capital market, instead of being paid immediately as pension benefits). The government must finance this ‘gap’, which often entails higher taxes or government debt. Samwick (Reference Samwick2000) undertakes an empirical analysis: first, he analyzes the time-series behavior of several countries that have undergone a pension reform and finds that none of them, with the exception of Chile, experienced a significant increase in national saving after the reform. Second, he assesses saving rates in a cross-sectional data set and finds that countries with PAYG systems tend to have lower saving rates than those with funded pension systems. Thus, evidence on the link between funding and saving is mixed.

Moreover, a priori it is not clear whether higher saving automatically translates into more investment. On the one hand, a lot of pension funds invest members’ contributions in a portfolio of worldwide securities. For example, in 2003, Dutch pension funds invested 57% of their funds abroad (Kakes, Reference Kakes2006). Thus, a higher saving rate could have only limited influence on domestic investment. On the other hand, in some countries the government forces pension funds to invest a large part of their funds in their own country. Furthermore, the ‘Feldstein–Horioka puzzle’Footnote ¹ (Feldstein and Horioka, Reference Feldstein and Horioka1980) indicates that domestic investment is strongly related to domestic saving (Davis and Hu, Reference Davis and Hu2008). It also depends on the quality of investment. An example from the latter days of communism shows extremely high investment rates in the Soviet Union, with economic growth rates close to nil (Barr, Reference Barr2000). Finally, when pension funds use funds to finance government debt, it is questionable whether productive investments result.

It is important to stress that the saving channel reflects a possible long-term effect. It simply takes a few years before higher saving might translate into higher economic growth.

2.2 Other channels

The development of capital markets might also allow funding to influence economic growth. The literature has established a clear positive link between funding and financial development (Catalan et al., Reference Catalan, Impavido and Musalem2000; Impavido et al., Reference Impavido, Musalem and Tressel2003; Hu, Reference Hu2005), and financial development is positively associated with economic growth (Levine and Zervos, Reference Levine and Zervos1998; Beck and Levine, Reference Beck and Levine2004). Thus, it appears that funding leads to better developed capital markets, which enhance growth in turn. Yet, Barr and Diamond (Reference Barr and Diamond2006) argue that capital market development is a less relevant argument for developed economies, whose capital markets are already quite well developed. On the other hand, in developing countries with capital markets that are hardly developed yet, this could affect economic growth in the short-run already.

A shift from a PAYG system to a funded pension system decreases the amount of distorting taxes that the government has to collect (Disney et al., Reference Disney, Boeri and Jappelli2004). Also, the weak link between pension contributions and benefits in PAYG systems leads workers to retire earlier and engage in less job mobility (Disney, Reference Disney2002). Therefore, a shift from PAYG to funding might increase economic efficiency and lead to higher growth. However, simulation studies show that the effects are actually rather small (Raffelhuschen, Reference Raffelhuschen1993; Kotlikoff, Reference Kotlikoff1996).

Finally, another argument put forward in the literature indicates that funding of pensions could increase growth by improving corporate governance (Barr and Diamond, Reference Barr and Diamond2006; Davis and Hu, Reference Davis and Hu2008), perhaps due to pension funds’ demand for more transparency and accountability at the firm level, as well as pressures on pension funds to undertake socially responsible investments (Clark and Hebb, Reference Clark and Hebb2003). Despite clear evidence of such a positive impact at the firm level in the United States (Woidtke, Reference Woidtke2002; Coronado et al., Reference Coronado, Engen and Knight2003), only Davis (Reference Davis2002) argues that these effects may be economy-wide.

4.1 Short-run effects

To examine whether changes in the degree of pension funding affect economic growth in the short-run, we estimate the following regression model:

(1)$$\eqalign{ {\rm log}\lpar y_{it} \sol y_{i\comma t \minus \setnum{1}} \rpar \equals \tab \mu _{i} \plus \delta _{t} \plus \gamma _{\setnum{1}} \;{\rm log}\lpar y_{i\comma t \minus \setnum{1}} \sol y_{i\comma t \minus \setnum{2}} \rpar \plus \gamma _{\setnum{2}} \;{\rm log}\lpar {\rm PA}_{i\comma t \minus \setnum{1}} \sol {\rm PA}_{i\comma t \minus \setnum{2}} \rpar \cr \tab \plus \gamma _{\setnum{3}} \;{\rm ror}_{it} \plus \varepsilon _{it} \comma \cr} $$

where y is real GDP per capita of country i in year t, PA is the ratio of pension assets over nominal GDP, ror is the rate of return of the pension sector, μ_i is a country fixed effect, δ_t is a time fixed effect, and ε is the error term. With this approach we focus on a possible short-run effect of funding on growth during the transition from a PAYG to a (partly) funded system.

The country fixed effect μ_i is included to allow different countries to have different steady-state growth rates. We also estimate equation (1) without country fixed effects, using a simple pooled OLS model. Although the fixed effects model can only identify the possible effect of changes in the degree of funding within countries, without the country fixed effects it is possible to identify the effect of changes within and between countries. The drawback is that we then assume that all countries grow at the same rate, ceteris paribus.

The amount of pension assets is measured on the last day of the year. Note that we take the first lag of the growth in pension assets to prevent possible reverse causality. Furthermore, it is very hard to think of a situation in which changes in pension funding would immediately influence economic growth. Therefore, changes in the amount of pension assets during period t − 1 are assumed to influence economic growth in period t.

Widely used estimators for dynamic panel models include the Arellano–Bond and Blundell–Bond estimators (Arrelano and Bond, Reference Arrelano and Bond1991; Blundell and Bond, Reference Blundell and Bond1998), both of which were developed for small T, large n panels, where T refers to the time dimension and n to the cross-sectional dimension of the data. With 54 cross-sections our n is too small relative to our T (8), to employ these estimators.

As an alternative, we consider the within (or LSDV) estimator. Nickell (Reference Nickell1981) shows that this estimator is not consistent for finite T in dynamic panel data models, where the inconsistency for n → ∞ is O(T ⁻¹). In other words, the inconsistency is inversely proportional to T and disappears when T becomes very large. However, our T is 8, and the bias in our estimates will be considerable when we use the within estimator for our dynamic panel model.

Accordingly, we use a bias-corrected LSDV estimator for equation (1). Bun and Kiviet (Reference Bun and Kiviet2003) provide LSDV bias approximations, extended to unbalanced panels by Bruno (Reference Bruno2005a). The analytical approximation of the bias that we use has order O(T ⁻¹), the simplest bias approximation that Bun and Kiviet (Reference Bun and Kiviet2003) consider. It still accounts for most of the bias in the LSDV estimator when nT ⩾ 400 and n ⩾ 10. We initialize this bias correction procedure with a standard Blundell–Bond estimator with no intercept (see Bruno (Reference Bruno2005b) for details). In addition, we use a non-parametric bootstrap procedure to estimate the asymptotic variance–covariance matrix of the bias-corrected LSDV estimates. Kiviet and Bun (Reference Kiviet and Bun2001) show that this variance estimator is superior to the standard analytical variance estimator.

Table 2 presents the estimates of equation (1). We show results from four different specifications, with and without country fixed effects, and with and without the rate of return of the pension sector as control variable. All four specifications include a full set of time fixed effects. In columns 1 and 2, we show results with country fixed effects, estimated by the bias-corrected LSDV estimator. In columns 3 and 4, we show results without country fixed effects, estimated by pooled OLS. In order to account for possible serial correlation in the residuals, we cluster the standard errors across cross-sections (countries) in columns 3 and 4.

Table 2.

Dependent variable: real GDP per capita growth

Notes: Columns 1 and 2 are estimated by the bias-corrected LSDV estimator. The numbers in parentheses are bootstrap standard errors. Columns 3 and 4 are estimated by OLS (without country fixed effects). The numbers in parentheses are clustered standard errors.

** and *** denote significance at 5% and 1%. Growth variables and the rate of return are measured as fractions. The rate of return of the pension sector is calculated by using the MSCI World Gross Return Index, the Barclays Capital Global Aggregate Bond Index, and the 3-month T-bill rate plus data on the share of pension assets that is invested in stocks, bonds, and cash.

As discussed before, the relationship between pension assets and the rate of return might be two-sided. On the one hand, high rates of return increase the growth rate of pension assets beyond what is caused by changes in funding. On the other hand, if countries decide to increase the funding ratio of their pensions this might have a positive effect on the rate of return. Thus, if we include the rate of return as a control variable, we also correct for that part of the growth in pension assets that affects the rate of returnFootnote ². However, if we do not include the rate of return we ignore the possible effect of the rate of return on pension assets. Therefore, the true effect of changes in funding on economic growth will be somewhere in between the estimates with and without the rate of return included.

In all specifications, the coefficient estimate of the growth of pension assets is very close to zero and insignificant. The autoregressive parameter (γ₁) is between 0.4 and 0.7, while the rate of return has a positive effect, as expected, but is only significant in the model with country fixed effects. Furthermore, the inclusion of the rate of return hardly affects the relationship between changes in funding and economic growth, in contrast with our expectations.

All in all, we are led to conclude that changes in the degree of pension funding do not influence economic growth in the short-run. At least, not in the period between 2001 and 2010 in a large sample of OECD and non-OECD countries.

4.2 Long-run effects

Although we are not able to find an effect of changes in the degree of funding on economic growth in the short-run, there are several channels through which funding might impact upon growth that need somewhat more time. For example, it will probably take a few years before a larger pool of savings will translate into a higher rate of capital formation and thus higher economic growth. Therefore, we examine how changes in the degree of funding affect average economic growth in the years after the change. We start with a simple visual inspection of the data.

Figure 1 plots log(real GDP per capita) against real GDP per capita growth. As expected, and predicted by the convergence literature (see Barro (Reference Barro1997) ), the relationship is negative; poor countries grow faster than rich countries. However, from Figure 2 the relationship between log(real GDP per capita) and pension asset growth seems to be negative as well. Thus, mainly poor countries have high levels of pension asset growth. Then, simply looking at the bivariate relationship between pension asset growth and economic growth, as in Figure 3, reveals a positive relationship between the two; countries with high levels of pension asset growth tend to grow faster than countries with low (or negative) pension asset growth. The positive relationship becomes even stronger if we drop the obvious outlier.

Figure 1.

log(real GDP per capita) plotted against real GDP per capita growth (measured as a fraction)

Figure 2.

log(real GDP per capita) plotted against pension asset growth (measured as a fraction)

Figure 3.

Pension asset growth plotted against real GDP per capita growth (both variables measured as a fraction)

However, the question is whether this positive link between changes in funding and economic growth remains when we control for initial income. It might be that poorer countries do not grow faster because they reformed their pension system, but simply because they are catching-up with the developed world. In that case, it is a convergence effect that is at work and not an effect of pension funding on economic growth. Besides that, for countries with hardly any pension assets at all it is much easier to achieve enormous growth in pension assets as it is for countries with almost fully funded pension systems. This is simply due to the nature of the growth rate as a measure of changes in funding. Therefore, we will estimate regression models with and without initial income included. Thus, to find out whether there are long-run effects of changes in the degree of funding of the pension system on economic growth, we use a classical empirical growth model (see e.g., Barro (Reference Barro2001); Rodrik (Reference Rodrik2008)):

(2)$$\eqalign{ {1 \over T}{\rm log}\lpar y_{i\comma t \plus T} \sol y_{it} \rpar \equals \tab \gamma _{\setnum{0}} \plus \gamma _{\setnum{1}} \;{\rm log}\lpar y_{it} \rpar \plus \gamma _{\setnum{2}} \;{\rm log}\lpar PA_{i\comma t} \sol PA_{i\comma t \minus \setnum{1}} \rpar \cr \tab \plus \gamma _{\setnum{3}} \;{\rm ror}_{it} \plus \varepsilon _{i\comma t \plus T} \comma \cr} $$

where the year t is 2002, T equals 7, and all other variables are as defined before. We thus regress the average growth rate over the period 2002–9 on the (log) level of income in 2002, the growth rate of pension assets from 2001 to 2002, and the rate of return of the pension sector in 2002. As the amount of observations is very limited, we use the period 2002–9 instead of 2002–10 as economic growth rates for 2010 are not (yet) available for all countries.

Table 3 presents the results of estimating equation (2). We show the results for several specifications; with and without the rate of return as control variable, and with and without initial income included.

Table 3.

Dependent variable: average real GDP per capita growth 2002–9

Notes: Regressions are estimated by OLS. The numbers in parentheses are robust standard errors.

* , ** and *** denote significance at 10%, 5%, and 1%. Growth variables and the rate of return are measured as fractions. The rate of return of the pension sector is calculated by using the MSCI World Gross Return Index, the Barclays Capital Global Aggregate Bond Index, and the 3-month T-bill rate plus data on the share of pension assets that is invested in stocks, bonds, and cash.

Looking at the estimates for the growth rate of pension assets, it turns out to be significantly positive when we do not include initial income in the regression. However, when we do include initial income, which is significantly negative, the point estimate of the growth rate of pension assets decreases and becomes insignificant. Furthermore, the R ² dramatically increases. This is exactly what we discussed above; apparently relatively poor countries, with high growth rates of pension assets, grow faster than rich countries because they are poor and not because their pensions are being funded. Finally, the rate of return of the pension sector has a negative effect on growth, which seems to contradict theory, but does not qualitatively alter the effect of changes in funding on growth.

One of the drawbacks of the simple cross-sectional model is that it leaves us with only 33 observations. It might be that we do not find significant effects because we simply do not have enough data. We therefore estimate a model with overlapping observations as well. Average economic growth between 2002 and 2006 is regressed on income in 2002, the growth of pension assets from 2001 to 2002 and the rate of return of the pension sector in 2002, average economic growth between 2003 and 2007 is regressed on income in 2003, the growth of pension assets from 2002 to 2003 and the rate of return of the pension sector in 2003, and so on. This yields five observations for countries with no missing data, and a total of 199 observations for all countries together.

We estimate this model with pooled OLS, as the number of observations for each cross-section is too limited to estimate a fixed effects model. In order to account for serial correlation within clusters (countries), serial correlation caused by the overlapping observations, and correlation between groups we should calculate Hodrick-standard errors (Hodrick, Reference Hodrick1992). However, as these are not available for panel data, we calculate clustered bootstrap standard errors (see Cameron et al. (Reference Cameron, Gelbach and Miller2008) ) where we cluster by country. We thus assume correlation within countries over time. The results are shown in Table 4.

Table 4.

Dependent variable: average real GDP per capita growth (overlapping observations: full sample)

Notes: Regressions are estimated by OLS. The numbers in parentheses are clustered bootstrap standard errors.

** and *** denote significance at 5% and 1%. Growth variables and the rate of return are measured as fractions. The rate of return of the pension sector is calculated by using the MSCI World Gross Return Index, the Barclays Capital Global Aggregate Bond Index, and the 3-month T-bill rate plus data on the share of pension assets that is invested in stocks, bonds, and cash.

The growth rate of pension assets is significantly positive in all specifications although the point estimate becomes smaller when we include initial income, but it remains significant. The rate of return has a significantly positive effect on economic growth in all specifications. However, its inclusion seems to have hardly any effect on the estimate for change in pension funding.

Finally, as Figures 1–3 reveal, there are some obvious outliers present in the data. Therefore, we also estimated all models without these outliers. Most results are comparable to the results with the outliers included, with the exception of the model with overlapping observations. Leaving out five observations where the growth rate of pension assets exceeds 100% results in an insignificant estimate for the growth rate of pension assets when initial income is included (columns 2 and 4). We show these results in Table 5. Furthermore, we also estimated the models with a few variables included that should proxy for steady-state growth (i.e., trade/GDP, investment/GDP). We do not show the results here, but they are also comparable with the results in Tables 3 and 4.

Table 5.

Dependent variable: average real GDP per capita growth (overlapping observations: without outliers)

Notes: Regressions are estimated by OLS. The numbers in parentheses are clustered bootstrap standard errors.

** and *** denote significance at 5% and 1%. Growth variables and the rate of return are measured as fractions. The rate of return of the pension sector is calculated by using the MSCI World Gross Return Index, the Barclays Capital Global Aggregate Bond Index, and the 3-month T-bill rate plus data on the share of pension assets that is invested in stocks, bonds, and cash.

The evidence for a long-run effect of changes in funding on economic growth seems to be mixed. We only find a positive effect when we use a model with overlapping observations, but this positive effect disappears when we exclude a few outliers. The point estimate in the model with the outliers included is 0.018, which means that a 10 percentage points increase in the funding ratio would increase the average economic growth rate in the 4 years after the change with 0.18 percentage points, which seems to be rather modest.