Financial Development and Aggregate Saving Rates: A Hump-Shaped Relationship Pengfei Wang, Lifang Xu, and Zhiwei Xu Hong Kong University of Science and Technology First version: October 17, 2011 This version: June 06, 2012 Abstract This study has documented a hump-shaped empirical relationship between …nancial development and the national saving rate across 12 East Asian and 31 OECD economies. An incomplete-market model featuring both heterogeneous households and heterogeneous …rms is provided to explain this hump-shaped relationship. The key insight of the model is that …nancial development tends to reduce the precautionary-saving incentives of households but increase …rms’ability to borrow and invest. As a result, the aggregate saving rate may rise initially with …nancial development because of greater investment by …rms, but then declines with further …nancial development because of substantially reduced precautionary savings by households. The model also predicts that the market interest rate lies substantially below the rate of return to capital in emerging economies, but the gap diminishes with …nancial development, as observed in the data. Keywords: High Saving Rate Puzzle, Financial Development, Hetergenous Firms, Hetergenous Households, Borrowing Constraints JEL Classi…cation: E21, E22, E44 We thank Yi Wen for numerous conversations that have led to some of the ideas in this paper and for his detailed suggestions that have improved the quality of the paper. We also thank Charles Horioka for useful comments. Correspondence: Pengfei Wang, Economics Department, Hong Kong University of Science and Technology Phone: 852-2358-7612. Email: [email protected]. 1 1 Introduction How does …nancial development a¤ect saving rates? Understanding the link between …nancial development and saving rates is important for several reasons. First, the saving rate is a crucial determinant of a country’s long-run per capita income. The neoclassical exogenous growth model (Solow, 1956) suggests that the higher the rate of saving, the richer the country in per capita income. The endogenous growth theory pioneered by Romer (1986) and Lucas (1988) predicts that the saving rate determines long-run growth: a higher saving rate leads to a higher economic growth rate. Understanding the link between …nancial development and the saving rate is thus important for understanding how …nancial development a¤ects economic growth, which is a useful step towards understanding the cross-country dispersion in measured per capita income levels. Understanding the link between …nancial development and the saving rate is also important for policy. Aiyagari (1994) has shown that the optimal tax on capital income should be positive instead of zero in a …nancially underdeveloped economy, which contradicts the well-known recommendation of zero capital income taxation in the long run (Chamley, 1986). Krugger (2010) has further shown quantitatively that a positive capital income tax rate has sizeable welfare gains in a carefully calibrated general equilibrium model. However, such policy recommendations rely on the prerequisite that …nancial underdevelopment will lead to oversaving. If …nancial underdevelopment causes undersaving, such policies may lead to welfare losses. Finally, understanding the link between …nancial development and the saving rate is important for its own sake. The high saving rates in developing Asian countries have drawn signi…cant attention from both academics and policy makers. The high saving rates in emerging Asia are claimed to cause worldwide low real interest rates, to exacerbate asset price bubbles in the developed countries, and even to have triggered the recent …nancial crisis1 . One popular explanation provided by the global imbalance literature (e.g., Caballero, Farhi, and Gourinchas, 2008; Mendoza, Quadrini and Rios-Rull, 2009; Song, Kjetil, and Fabrizio, 2011; Wen, 2011) attributes the high saving rates to …nancial underdevelopment in those countries. However, Wei and Zhang (2011) recently criticize such explanations using the fact that there has been …nancial development in those countries but at the same time the saving rate continued to increase. This continuously rising saving rate along with …nancial development lead Wei and Zhang (2011) to conclude that other factors such as a growing sex-ratio imbalance are the more important causes 1 On June, 3, 2008, in a speach at an international monetary conference, US Federal Reseve chairman Ben Bernanke said, "In the …nancial sphere, the three longer-term developments I have identi…ed are linked by the fact that a substantial increase in the net supply of saving in emerging market economies contributed to both the U.S. housing boom and the broader credit boom. The sources of this increase in net saving included rapid growth in high-saving East Asian countries and, outside of China, reduced investment rates in that region; large buildups in foreign exchange reserves in a number of emerging markets; and the enormous increases in the revenues received by exporters of oil and other commodities. The pressure of these net savings ‡ows led to lower long-term real interest rates around the world, stimulated asset prices (including house prices), and pushed current accounts toward de…cit in the industrial countries–notably the United States–that received these ‡ows." 2 of the high saving rate in China (and some of other Asian emerging economies). Again, this criticism by Wei and Zhang is valid only if the relationship between …nancial development and the saving rate is monotonic. If the relationship is nonlinear, …nancial market development may induce high saving countries to save even more. The purpose of this study is hence twofold: First, to examine the relationship between …nancial development and saving rates empirically; and second, to explain the empirical relationship theoretically in a general equilibrium model. Our empirical analysis demonstrates a hump-shaped relationship between …nancial development and the saving rate. Using a panel of 31 OECD countries and 12 East Asian countries with annual time series observations from 1960 to 2008, …nancial development is shown to increase with the aggregate saving rate initially, but further …nancial development reduces saving, after controlling for demographic factors, income level, economic growth, in‡ation, and the interest rate, etc.2 A number of scholars have noted a statistical relationship between …nancial development and the saving rate. For example, King and Levine (1993) …nd that higher levels of …nancial development are associated with faster capital accumulation. Loayza, Schmidt-Hebbel and Serven (2000) and Horioka and Yin (2010) …nd a negative correlation. Park and Shin (2009) …nd the impact of …nancial development to be insigni…cant. Finally, Horioka and Hagiwara (2010) use data from 12 economies in developing Asian countries during 1996-2007 to demonstrate that the relationship between …nancial development and saving rate is nonlinear and hump-shaped. The empirical results of this study reconcile some of those con‡icting …ndings and con…rms the hump-shaped relationship in a larger cross-country sample. Con…rming the nonlinear relationship with a broader sample beyond East Asian countries is important, since one may argue that East Asian countries in the sample of Horioka and Hagiwara (2011) might be special, as they have historically had high saving rates. The existence of a uniform nonlinear relationship between …nancial development and the saving rate in a broader sample also calls for a uni…ed theory to explain the relationship. That is the main goal of this study. The hump-shaped relationship between …nancial development and the aggregate saving rate cannot easily be reconciled with the existing models of precautionary saving. The precautionary saving models developed by Leland (1968), Kimball (1990), Weil (1993), Aiyagari (1994), and many others typically predict a negative monotonic relationship between …nancial development and the saving rate. In these models households mainly rely on savings to self-insure against idiosyncratic risks due to incomplete …nancial markets. Greater …nancial development can hence reduce households’ reliance on saving for self-insurance and thus reduce the saving rate. Such a negative relationship has been used to explain the high saving rates in some …nancially underdeveloped countries. For example, Wen (2009, 2011) apply the precautionary savings story to explain the puzzle of high Chinese saving. He shows that with large uninsured risks and severe borrowing constraints, rapid income growth can increase (instead of decrease) 2 The ratio of domestic private credits to GDP is used as a measure of …nancial development following King and Levine (1993). Other …nancial indicators are also tested as measures of …nancial development, and the hump-shaped relationship are robust to these di¤erent indicators. 3 the household saving rate, thus explaining the puzzle that China’s household saving rate is one of the highest in the world despite a 10% per year average income growth rate over recent decades.3 The global imbalance literature also relies on a negative monotonic relationship between …nancial development and precautionary saving to explain capital out‡ows from emerging countries to developed ones. However, such a monotonic relationship is contradicted by the data. Take China as an example. The private credits to GDP ratio in China has increased from 50.9% in 1977 to 103.7% in 2008, but China’s national saving rate has climbed to 51.8% from 28.9% during the same period. A similar pattern can be found in South Korea. In the 1960s, the Koreans saved less than 10 percent (8.6%) of their gross national product, with the average private credits to GDP ratio equal to 20.6%. The average national saving rate in the 1990s had risen to 36.3% while at the same time the average private credits to GDP ratio has increased to 67.5%. The Korean saving rates do decline to 30% in 2008 as the average private credits to GDP ratio increased further to 108.8%. On the other hand, the existing models of …rm-level …nancial constraints typically imply a positive relationship between …nancial development and a …rm’s saving rate, in contrast to the previous literature on household savings. In his seminal work, Myers (1977) …rst points out that debt can prevent a …rm from making an investment project with positive net present value, a problem commonly referred to as the debt overhang or underinvestment problem. Most macroeconomic workhorse models of …nancial contracts feature such underinvestment problems. Examples include the credit rationing model of Stiglitz and Weiss (1981), the costly state veri…cation models of Townsend (1979), Gale (1985), and Bernanke and Gertelr (1989), the limited contract enforcement model of Kiyotaki and Moore (1997), and the moral hazard model of Holmstrom and Tirole (1998). Since investment equals saving at equilibrium, these models imply undersaving due to …nancial frictions. Financial development and saving can therefore exhibit a positive monotonic relationship. But, such a pattern is not fully consistent with the empirical evidence marshalled in this study. We combine the insights generated by the aforementioned two literatures to explain our empirical …ndings. In particular, we propose that the observed hump-shaped relationship between …nancial development and the saving rate can be explained by imposing …nancial frictions on both the household and the …rm side. Generally speaking, this assumption should complicate a dynamic general equilibrium model substantially, since to compute the general equilibrium in such a model one must keep tract of the distribution of household wealth and the distribution of …rms’ capital stocks simultaneously, and both distributions have typically in…nite dimensionality. In addition, the two distributions interact endoge3 We do not argue that no model of precautionary saving can explain the hump-shaped relationship between …nancial development and the aggregate saving rate. A hump-shaped relationship is possible if …nancial development can stimulate economic growth. Wen (2009, 2011) points out that economic growth can increase saving dramatically in a general equilibrium model. Financial development may increase saving initially due to a rising rate of income growth but decrease if eventually by reduction of the self-insurance motivation. The interaction of these two opposite forces may be able to generate a nonlinear relationship between …nancial development and the aggregate saving rate. But, con…rming these conjectures with an endogenous growth model is an interesting topic left for future research. 4 nously at general equilibrium. For example, the distribution of capital depends on the pricing kernel, which equals the marginal investor’s discounting factor, which depends on the wealth distribution of the households. Yet the households’wealth distribution is a¤ected by the …rms’distribution of capital as it a¤ects the returns on the households’portfolios. Solving such a model can be daunting. To overcome such technical di¢ culties, we borrow the household model of Wen (2009, 2011) assuming a quasilinear preference function. This allows us to characterize the household saving behavior in closed form even though households face uninsurable idiosyncratic liquidity shocks and borrowing constraints. For any given interest rate, households are found to save excessively compared to an economy without uninsurable risks, as Wen (2009, 2011) has shown. A higher level of …nancial development will thus lead to less incentive to save, everything else equal. Our major innovation lies on the …rm side. We add to heterogenous …rms to Wen’s heterogenoushousehold model. In the enriched model, the …rms discover investment opportunities randomly, which captures the idea that investment at …rm and plant level is lumpy. Hence, only a fraction of …rms invest in each period. This creates a need to transfer funds among the …rms. However, friction arises when moving funds between …rms, which is due limited enforcement problems. In other words, the …rms are …nancially constrained. Assuming constant returns to scales allows us to characterize the …rms’investment decision rules analytically and permit exact aggregation, so only the mean of the distribution of capital matters for aggregate equilibrium. 4 We show that borrowing constraints on the …rms create a gap between the return on capital and the e¤ective real interest rate for saving. A relaxation in the borrowing constraints on …rms will narrow this gap and increase the real interest rate. Hence a high level of …nancial development will also generate an incentive for …rms to increase their saving (or investment) rates. The overall e¤ect on aggregate saving rate will then depend on which side of the economy (the household side or the …rm side) dominates. Our model is able to generate a hump-shaped relationship between …nancial development and the aggregate saving rate as with reasonable parameter values. To understand the intuition, imagine an extreme case in which …rms have to borrow to invest but they cannot borrow at all. In such an extreme case, the total investment demand would be zero. Since at equilibrium total saving must equal to total investment, the aggregate saving rate will always be zero regardless of the saving incentives on the 4 Our modelling of …nancial frictions at the …rm level is motivated by a vast empirical literature that studies the interaction between …nancial constraints and corporate investment. Numerous studies have found that …nancial constraints on corporate investment are pervasive even in the developed economies. (See Hubbard (1998) for a recent survey on the earlier literature using the Q-theory of investment, and Inessa (2003) for recent international evidence from structural investment models). At the same time, an equally vast theoretical literature pioneered by Bernanke and Gertler (1989), Kiyotaki and Moore (1997), Bernanke, Gertler and Gilchrist (1999), Albuquerque and Hopenhayn (2004) and Cooley, Marimon, and Quadrini (2004) has developed full dynamic general equilibrium models of investment with …nancial constraints to examine the aggregate implications of …nancial frictions for business cycles ‡uctuations. Our heterogenous-…rm model with limited enforcement frictions and random investment opportunities is also closely related to the recent general-equilibrium-heterogeneous-…rm models of Wang and Wen (2010, 2011) in which the assumption of constant returns to scale allows them to characterize the …rms’investment decision rules in closed forms and permit exact aggregation. Wang and Wen (2010, 2011) do not consider heterologous households. Miao and Wang (2010) exploit the property of constant returns to scale to analytically model credit risk in a DSGE model. Miao and Wang (2011) study asset bubbles in a …nancially underdeveloped economy with heterogenous …rms. 5 household side. Although households have strong precautionary saving motives under large idiosyncratic risks, the e¤ective rate of return (the interest rate) on household saving is too low to induce them to save. As the borrowing constraints in such an economy gradually relax (for both …rms and households), the aggregate saving rate will initially go up as …rms start to invest heavily. Beyond a critical level, however, the downward trend in household saving starts to dominate. Thus, further relaxation in …nancial constraints will reduce the aggregate saving rate. The rest of the paper is organized as follows. Section 2 presents the empirical evidence of a humpshaped relationship between …nancial development and the aggregate saving rate. Section 3 presents our model in which both households and …rms may face …nancial constrains. Section 4 de…nes and characterizes the general equilibrium. Section 5 examines the impact of …nancial development on the aggregate saving rate for di¤erent scenarios in the model. Finally, section 6 concludes. 2 Empirical Evidence Anecdotal evidence usually suggests that the e¤ect of …nancial development on saving is negative. Indeed, many previous empirical studies, including those by Loayza, Schmidt-Hebbel and Serven (2000), Horioka and Yin (2010) and those surveyed by Schmidt-Hebbel and Serven (2002), con…rm such a negative relationship. However, those studies typically use the aggregate credits to the private sector as a measure of …nancial development, yet presume that the relationship between …nancial development and saving is linear, which may not be the case in reality. Beck et al. (2012) …nd that leading to …rms and lending to households have fairly di¤erent e¤ects on the outcomes of real sectors. For saving rate, on one hand, …nancially constrained household may save due to precautionary motives, hence expanding credits to the households can lead to a decreasing household saving rate5 ; on the other hand, …nancially constrained …rms cannot invest to their optimal level due to the limit of fund, thus, expanding credits to the …rms can lead to an increasing …rm investment and hence an increasing …rm saving rate. As a result, the e¤ects of increasing aggregate private credits on the aggregate saving rate are ambiguous, depending on which sector is favored and gets the expanding credits. Thus, at least in theory, a non-monotonic relationship is possible. Actually, when Horioka and Terada-Hagiwara (2010) relax the assumption of linearity, they …nd that the relationship between the saving rate and …nancial development (aggregate private credits) is concave for a sample of 12 East Asian countries. In this study, we collect time series data ranging from 1960 to 2008 on saving rates and its potential explanatory variables for the same 12 East Asian countries6 and extra 31 OECD economies, and use 5 In Appendix D, we use a sample of 26 OECD economies to show that household saving rate does decrease when the household credits expand or when the size of household …nance increases. 6 India and Pakistan are also included in the sample, in order to be comparable to the results of Horioka and TeradaHagiwara (2010). But it is not crucial for the results presented in this paper. 6 various subsamples and estimators to re-examine the possibility of nonlinearity in the relationship. 2.1 Data Our sample draws from the World Development Indicators and the Penn World Table. The data set consists of annual time series observations from 1960 to 2008 for 41 countries, which locate in East Asia or belong to OECD7 . A detailed description of the data is presented in Appendix B. Following Loayza et al. (2000), we set a threshold of 50% annual in‡ation rate to exclude the episodes of high in‡ation in the sample. To accommodate the joint endogeneity problem of explanatory variables, we will implement dynamic panel regression with GMM-IV estimators described below. Thus, we choose to work with the original data instead of phase-averaged data to retain relevant information. As a result, we have a maximum of 976 observations for 41 countries. The data show that both saving rates and levels of …nancial development display signi…cant variations across di¤erent countries and time. Saving rate ranges from less than 5% in Indonesia in the early 1960s to more than 50% in China today, private credits to GDP ratio also ranges from less than 10% in the early years of those developing countries to larger than 2 in currently developed economies. Figure 1 displays the scatterplots of gross saving rate against private credits to GDP ratio for the 12 East Asian countries, stacked with the quadratic …t and LOWESS …t. It is evident from the graph that the relationship between saving rate and …nancial development (private credits to GDP ratio) displays a humped shape, with …nancial development initially increasing and then decreasing the saving rate. This pattern is consistent with what Horioka and Terada-Hagiwara (2010) found in their sample of these same 12 East Asian economies. While East Asia might be a bit special due to its eye-catching high saving rates, our panel regression using a broader sample that also includes 31 OECD economies and using various speci…cations and estimators, further con…rms the hump-shaped relationship. 2.2 Empirical speci…cation Consider the following reduced-form regression equation: sit = 1 sit 1 0 2 Xit + + i + uit ; (1) where s is the gross domestic saving rate, X is a vector of explanatory variables re‡ecting …nancial development, age structure, rates of return, uncertainty, …scal policy, income level and growth, which have been shown to a¤ect the saving rate in the literature, denotes the unobserved country …xed e¤ects, and uit is the error term. 7 Japan and South Korea belong to both groups, thus the total number of countries studied is 41, instead of 43. The size of the sample is determined by availability of observations. 7 .5 .4 Gross saving rate .2 .3 .1 0 Scatterplots LOWESS f it 0 .5 1 1.5 Priv ate credits to GDP ratio Quadratic f it 2 2.5 Figure 1: Scatterplot of gross saving rate and private credits/GDP ratio, East Asia Various theories have emphasized on di¤erent factors in explaining the saving behavior. Life-cycle models imply that demographic factors play a non-trivial role in determining the saving rate, so each economy’s aged dependency ratio, youth dependency ratio and life expectancy are included in the regression. Precautionary saving theory emphasizes that people will save against future uncertainties, thus in‡ation rate is employed to capture macroeconomic uncertainty and public expenditure on health and education is used to re‡ect people’s concern about uncertainty on future health and education expenditures. The permanent income hypothesis suggests that income and its growth determine economic agents’ consumption and savings, so per-capita real GDP and its square, and the growth rate of per-capita real GDP are all included in the regressions. The square of income level in the model is included in an attempt to capture the potential nonlinear relationship between income and the saving rate. Both Park and Shin (2009) and Horioka and Terada-Hagiwara (2010) …nd that the relationship is convex in their samples of Asian countries. Several other common variables such as real interest rate and current account balance are also included. However, the main purpose of this regression is to investigate the presumable nonlinear e¤ects of …nancial development on the saving rate, after properly controlling those relevant factors mentioned above. There are various possible measures of …nancial development8 , some of which rely on the size of the whole …nancial sector, while the others focus on the role of …nancial intermediation. To be consistent 8 See the survey by Cook (2003). Thorsten Beck also writes about "two concepts of …nancial development" recently, one is the "…nancial intermediation view", the other is the "…nancial center view". For details, please refer to his article: http://www.voxeu.com/index.php?q=node/7185 8 with the argument in this paper and with the most popular measure used in the literature, this paper uses the private credits to GDP ratio as a measure of …nancial development9 . In order to capture the possible nonlinearity, we employ both of the level and its square as explanatory variables10 . 2.3 Econometric Methodology We employ this …xed-e¤ects dynamic panel speci…cation to account for unobserved country-speci…c e¤ects and to deal with the possible inertia in the annual time series. In our static panel regressions, Wooldridge test for …rst-order serial correlation of the errors always rejects the null hypothesis of zero autocorrelation, which implies that the time series display serious inertia. That’s why we include the lagged saving rate in the regression. However, …xed-e¤ects dynamic panel speci…cation brings about several problems in estimation. To eliminate the unobserved country-speci…c e¤ects, we need to work with the mean-di¤erence or …rstdi¤erenced model derived from equation (1). Unfortunately, neither the within estimator nor the IV estimation using lags is feasible in the mean-di¤erence model, because any lag sit is correlated with ui and hence (uit ui ), which is the error term in mean-di¤erence model11 . OLS estimator is also inconsistent in the …rst-di¤erenced model, because sit 1 is correlated with uit , which is the error term now. Thus, we need to resort to some other methods. Arellano and Bond (1991) describe a GMMIV estimator (Arellano-Bond di¤erence estimator henceforth) and the test for serial correlation of the errors that apply to the …rst-di¤erenced model in detail. The estimator assumes that the errors uit are serially uncorrelated and E(sis uit ) = 0 for s t 2. By this assumption, fsit j gj=2;3;::: can be used as instruments for sit in the …rst-di¤erenced model and GMM estimation using lags as instruments delivers consistent and e¢ cient results. The Arellano-Bond di¤erence estimator can also control for endogeneity of the other explanatory variables, which is particular relevant in our study here. We already know that introducing lagged dependent variable causes endogeneity problem. It is also highly possible that credits, income and its growth are jointly determined with the saving rate, so they might be correlated with the errors as well. Thus, IV estimator is necessary. However, since the model is potentially over-identi…ed, GMM-IV estimator is more e¢ cient than two-stage least squares (2SLS). 9 This measure refers to the "…nancial intermediation" view of …nancial development in Thorsten Beck’s article. To be sure, a larger …nancial sector does not necessarily imply a higher level of …nancial development. Thorsten Beck provides an example of Nigeria in the 1980s, where expanding of …nancial sectors was accompanied by "…nancial dis-intermediation". Another example is China, large-scale state-owned bank sector provide limited credits to private-owned …rms. On the contrary, aggregate credits to private sectors directly measure the activities of …nancial intermediation. Hence, we consider it as a more appropriate measure of …nancial development in the context of this paper. 10 While quadratic …t is enough to capture nonlinearity, it is also interesting to see what happens under semiparametric regression, in which …nancial development enters non-parametrically and other explanatory variables enter parametrically. In Appendix E, we apply the Baltagi&Li (2002) estimator for partially linear …xed-e¤ects panel data models to our full sample, and …nd that the relationship is indeed hump-shaped. 11 To be speci…c, within estimator is no longer consistent, because sit 1 is correlated with uit 1 and hence with ui . As a result, the regressor (sit 1 si ) in the mean-di¤erence model is correlated with the error term (uit ui ). Similarly, any lag sit is correlated with ui and hence with (uit ui ), thus IV estimation using lags is not feasible either. 9 Following Loayza et al. (2000), in the following regressions, we assume that those endogenous explanatory variables are correlated with past and current errors but not with future errors. That is, E(Xit uis ) 6= 0 for s t, but E(Xit uis ) = 0 for s > t. In other words, they are "weakly exogenous". As argued by Loayza et al. (2000), this is a fair restriction, because the past realizations of explanatory variables are less likely to be in‡uenced by future innovations to saving rates. By this assumption, weakly exogenous variables in Xit can be instrumented using their own lags fXit j gj=2;3;::: in the …rst-di¤erenced model. However, the Arellano-Bond di¤erence estimator do have some problems, because lags could be weak instruments in the …rst-di¤erenced model if the corresponding level variables display serious inertia over time. Thus, the system estimator that incorporates both of the level equation and its …rst di¤erence is proposed by Arellano and Bover (1995) and Blundell and Bond (1998). They add another moment condition beyond those in the di¤erence estimator, namely, E( sit instrumented using its own …rst di¤erence sit 1 1 uit ) = 0. As a result, sit can be in the level equation (1). Similar moment conditions can be imposed on those endogenous explanatory variables, so that they can be instrumented using their …rst lag di¤erences in the level equation. This system estimator is shown to be be more precise and to have better …nite sample properties. Thus, we …nally choose the system GMM-IV estimator. In the following regressions, we treat two dependency ratios, life expectancy and public expenditures as strictly exogenous variables, and assume that the others are all weakly exogenous. We present the main results in the next subsection. In deriving these results, we instrument the endogenous variables using their own …rst two lags in the …rst-di¤erenced equation and using …rst lag di¤erence in the level equation12 . We also employ the …nancial development index prepared by Abiad et al. (2008) as an external instrument in the level equation, because Roodman (2009a) …nds that lags of several popular …nancial development measures, including private credits used in this paper, all perform badly in the system GMM estimation, when he replicates the exercises done by Levine et al. (2000). It is worth mentioning that using lags as instruments typically generates numerous instrument count. Roodman (2009a) shows that "too many instruments" can over…t the endogenous variables and weaken the Hansen test of joint validity of instruments, making the results unreliable. Thus, as a robustness check, we also consider reducing the instrument count in our regression exercises by "collapsing" the instrument matrix, as suggested by Roodman (2009a). 12 The main results are robust if we use 3 lags as intruments in the …rst-di¤erenced equation, while the results are mixed if only one lag is used. To be speci…c, when only one lag is used, in all cases the coe¢ cients still imply a concave relationship, but they are only signi…cant in the East Asia subsample, no matter the IV matrix is collapsed or not. In the full sample, they are signi…cant only when the IV matrix is not collapsed. Nevertheless, Hansen tests for joint validity of instruments in Table 2-4 indicate that using 2 lags as instruments are valid, even when we take care of the problem of "too many instruments" using the method suggested by Roodman (2009a). Thus, we interpret those results as supportive for a humpshaped relationship between …nancial development and saving rate. Our semiparametric regression in Appendix E further con…rms that. 10 2.4 Empirical Results The coe¢ cients of most interest in model (1) are those on the …nancial development and its square. Table 1 summarizes the regression coe¢ cients on these two variables for our full sample of 41 countries, using three di¤erent estimators13 . Table 2 4 in Appendix C report the full results. We brie‡y discuss these results in this section. All of these regressions show that the relationship between aggregate private credits and the saving rate is hump shaped. When credit volume is low, the saving rate increases with credit expansion; when credit volume increases further, the relationship is reversed. They con…rm what has been observed in Figure 1 with our naked eyes, and are consistent with the results presented by Horioka and TeradaHagiwara (2010), who only considered the sample of East Asian countries. Indeed, the last two columns of Table 2 in Appendix C actually replicate Horioka and Terada-Hagiwara’s results. Table 1: Summary of the coe¢ cients on private credits and its quadratic term, full sample Estimator Collapsed IV Private credits Private credits sq. Country e¤ects Observations (Number of cn) (1) GMM-System No (2) GMM-System Yes (4) Within – (5) RE – 0.0210*** (0.00676) -0.00706** (0.00310) Yes 809 (37) 0.0321** (0.0144) -0.0127*** (0.00491) Yes 809 (37) 0.0925*** (0.0287) -0.0296*** (0.0100) Yes 976 (41) 0.105*** (0.0308) -0.0321*** (0.0101) No 976 (41) Note: *** p<0.01, ** p<0.05; robust standard errors in parentheses. The regression coe¢ cients on age dependencies, income growth, income level, real interest rate, in‡ation rate, public expenditure and current account surplus are in listed in tables 2 to table 4 in Appendix C. The regression results are broadly consistent with the these of Loayza et al. (2000) and Horioka and Terada-Hagiwara (2010) regarding other explanatory variables. For example, Two age dependencies show signi…cant negative e¤ects on the saving rate, as senior citizens tend to consume their previous saving and kids usually consume without their own income. The per-capita GDP growth has a highly signi…cant positive e¤ect in almost all of the cases. Real interest and in‡ation rates have ambiguous e¤ects; Higher public expenditure typically has a negative e¤ect, which may re‡ect households’precautionary saving to some extent. The current account surplus has a positive and signi…cant e¤ect. 13 In unreported tables, we also try the Arellano-Bond di¤erence estimator, and the main results are robust. 11 In short, our results are essentially the same with those of Loayza et al. (2000), expect for the e¤ect of …nancial development. Loayza et al. (2000) presume that the e¤ect is linear and …nd a negative e¤ect. However, when we explicitly take nonlinearity into account, we …nd that the relationship is hump-shaped. The hump-shaped relationship is consistent with what Horioka and Terada-Hagiwara (2010) have recently found in a sample of 12 East Asian countries. In fact such a hump-shaped relationship is not a surprise if both households and …rms face credit constraints, as our model will show. 3 The Model We consider an in…nite-horizon economy. There is no aggregate uncertainty. The economy has two types of agents, consumers and …rms with equal mass normalized to unity. We use i 2 [0; 1] to index households and j 2 [0; 1] to index …rms. Firms accumulate capital and combine labor and capital to produce consumption goods. Households supply labor to the …rms and own their stock. Both households and …rms are subject to uninsurable idiosyncratic risks and …nancial constraints to be speci…ed below. 3.1 Households The household side is similar to that modelled by Wen (2009, 2011). Household i derives utility from consumption cit and leisure nit . The instantaneous utility function is uit = it log cit an idiosyncratic preference shocks with the distribution function F ( ) = Pr [ i The mean of i is normalized to one, E i nit , where ] and support [0; it is max ]. = 1. Each time period is divided into two sub-periods. The idiosyncratic shocks are realized in the second sub-period. Each household i chooses labor supply nit in the …rst sub-period without observing it , and chooses consumption cit and savings in bonds (sit+1 ) and stocks (ait+1 ) in the second sub-period after observing it . With such an information and market structure, labor income cannot be used to fully diversify the idiosyncratic risk and savings become a bu¤er stock to smooth consumption. Taking as given the real interest rate Rf t and real wage Wt , the household i solves max fc;s0 a0 g where ( max Ei fng " 1 X t ( it log cit t=0 #) nit ) ; (2) 2 (0; 1) is the discount factor. The household i faces a period-by-period budget constraint, cit + sit+1 + ait+1 Qt Rf t sit + Wt nit + (Qt + Dt ) ait ; (3a) where ait+1 is the share of a portfolio of stocks purchased by household i. Assume each household is subject to a limited borrowing capacity: sit+1 + ait+1 Qt Rf t 12 Bt ; (4) where Bt is an exogenous borrowing limit. Denote the Lagrangian multipliers for (3a) and (4) as it ; it and respectively. The …rst-order condition for labor is Wt Ei [ it ] = ; (5) where Ei is an expectation operator over idiosyncratic risk serving i, the household will take the distribution of i. Since labor is determined before ob- into account. The …rst-order conditions for i fcit ; sit+1 ; ait+1 g are given, respectively, by it cit it it 3.2 it it = it ; (6) = Rf t Ei [ = Ei it+1 it+1 ] ; (7) Dt+1 + Qt+1 : Qt (8) Firms Each …rm j combines labor (Nj ) and capital (Kj ) to produce output employing Cobb-Douglas technology, 1 Yjt = Kjt Njt ; 2 (0; 1) : Following Kiyotaki and Moore (1997, 2008), we assume that …rms encounter investment opportunities randomly14 . In each period, …rm j encounters an investment opportunity with probability probability 1 : With ; the …rm has no investment opportunity. Assume the random investment opportunities are independent across …rms and over time. Denote as Ijt the investment level of …rm j in period t if it can invest, the capital accumulation rule then follows Kjt+1 = (1 ) Kjt + Ijt with probability (1 ) Kjt with probability 1 : (9) With heterogenous households, the …rm’s dynamic programing problem becomes slightly more complicated. The …rst step is to …nd the right discounting factor. We like Hansen and Richard (1987), Ingersoll (1988), and Cochrane (1991) assume that a sequence of prices Pt exist such that a …rm’s value is determined by Jjt = 1 X Pt+ =0 Pt Djt+ ; where fDjt+ g1=0 is the dividend ‡ow generated by …rm j. De…ne (10) t = Pt = t , where < 1, and we can rewrite the …rm’s value as 14 Miao and Wang (2011) use a similiar assumption to study asset bubbles in a …nancially underdeveloped economy with heterogenous …rms but a representative household. 13 Jjt = 1 X t t+ Djt+ ; (11) t =0 which can be rewritten recursively as Jjt = Djt + t+1 Jjt+1 : (12) t Here is a discount factor, but notice that because of heterogeneity on the household side, does not necessarily equal households’discount factor : With the the …rm value given by (12), each …rm’s problem is then to maximize its value Jjt by choosing its optimal labor use and investment. The optimal labor choice is static. It is straightforward to show that the …rm’s operating pro…ts are linear function of its capital stock, 1 Rt Kjt = max Kjt Njt Wt Njt ; Njt (13) 1 where Wt is the real wage and Rt = 1 Wt . The optimal labor demand is hence Njt = 1 1 Wt Kjt : (14) Denote the …rm value in period t as Jt (Kjt ), and it now can be de…ned recursively using the proper discounting factor t+1 t as Jt (Kjt ) = max Rt Kjt Ijt Ijt + t+1 [ Jt+1 ((1 ) Kjt + Ijt ) + (1 ) Jt+1 ((1 ) Kjt )] : (15) t Assume that …rm j can use both internal funds, Rt Kjt , and outside funds from borrowing, Ljt , to …nance investment, its maximum investment is thus subject to the constraint Ijt Ljt + Rt Kjt : (16) For simplicity, assume that the external funds are raised through intra-period loans: …rms borrow from …nancial intermediaries at the beginning of period t and pay back with zero interest rate at the end of period t through raising additional equity. (Dividends maybe negative at the …rm level.) One key assumption of this model is that loans are subject to collateral constraints, as in the model of Kiyotaki and Moore (1997). Firm j pledges a fraction t as collateral. The parameter 2 (0; 1] of its …xed assets Ktj at the beginning of period may represent the tightness of the collateral constraint or the extent 14 of …nancial market imperfections. At the end of period t, the market value of the collateral is equal to t+1 t Jt+1 ( Kjt ), which is the discounted expected market value of …rm j if it owns capital stock Kjt at the beginning of period t + 1 and faces the same investment and collateral constraints in the future. The amount of loans Ljt cannot exceed this collateral value, otherwise, the …rm would choose to default on its debt and lose the collateral value. This leads to the following collateral constraint: t+1 Ljt Jt+1 ( Kjt ) : (17) t To sum up, each …rm j solves problem (15) subject to constraints (16) and (17). 3.3 Financial intermediation The …nancial intermediation is quite simple. The …nancial intermediate holds a portfolio consisting of stock in all of the …rms and collects aggregate dividends from all of the …rms, Dt = Z (Rt Kjt Ijt ) dj: (18) [Qt+1 + Dt+1 ]: (19) The price of the portfolio Qt ; is thus Qt = t+1 t The presence of …nancial intermediation is just to simplify the notation of the households’maximization problem. One can also assume that households directly hold a market portfolio that consists of stock in all …rms, and the equilibrium results will be the same. 3.4 General Equilibrium R1 R1 R1 R1 R1 R1 Let Kt = 0 Ktj dj; It = 0 Ijt dj; Nt = 0 Njt dj; Yt = 0 Yjt dj; nt = 0 nit di; St = 0 sit di; and R1 Ct = 0 cit di be the aggregate capital stock, investment, labor demand, output, labor supply, household saving and consumption, respectively. General equilibrium is then de…ned as sequences of aggregate variables, (Kt ; It ; Nt ; Yt ; Dt ; nt ; St ; Ct ) ; individual …rms’choices, (Kjt ; Ijt ; Njt ; Ljt ; Yjt ) ; individual households’choices, (nit ; sit+1 ; cit ) and prices (Qt ; Wt ; Rt ; Rf t ) ; such that each …rm and each household solve their optimization problems, and all markets clear: N t = nt ; (20) St+1 = 0; (21) 15 Z ait di = 1; (22) Ct + It = Yt ; (23) Kt+1 = (1 4 ) Kt + It : (24) Characterization of the Equilibrium Let us …rst derive the optimal decision rules on a single …rm and a single household, then aggregate their individual choices and characterize the competitive equilibrium by a nonlinear equation system. 4.1 A Single Firm’s Decision Problem We conjecture that the value of a …rm has the following functional form: Vt (Ktj ) = vt Ktj ; (25) where vt is a to-be-determined variable that only depends on the aggregate states. De…ne qt = Et t+1 t vt+1 , which will proved to be the traditional Tobin’s q. With the conjectured value function, the …rm’s investment problem becomes vt Kjt = max Rt Kjt Ijt + qt [(1 Ijt ) Kjt + Ijt ] ; (26) with the liquidity constraint Ijt Ljt + Rt Kjt ; (27) qt Kjt : (28) and the collateral constraint Ljt The following Proposition characterizes the optimal decisions for …rm j. Proposition 1 If qt is greater than 1, the …rm always chooses to invest the maximum amount whenever it has an investment opportunity. In particular, the optimal investment decision is Ijt = qt Kjt + Rt Kjt : (29) In addition, vt = Rt + (1 )qt + (qt 1)(qt + Rt ); (30) where qt evolves according to qt = t+1 [Rt+1 + (1 )qt+1 + (qt+1 t 16 1)(qt+1 + Rt+1 )]: (31) Proof: See the Appendix A.1. Here, vt is the average market value of one unit of capital and qt is the ex-dividend value of one unit of installed capital, which is the marginal bene…t of new investment. Since the cost of investment is one, the additional gain from investing is positive if qt 1 > 0. In this case the …rm will want to borrow as much as possible to invest, so its borrowing constraint binds. The value vt consists of three parts on the right hand side of equation(30). First, one unit of capital can generate Rt units of operating pro…t in period t. Second, one unit of capital can carry 1 unit of capital to the next period with value (1 ) qt after depreciation and paying dividends. Finally, the capital can also be used as collateral. With probability the …rm will have an investment opportunity and one unit of capital is able to obtain qt units of loans, which can expand the expected net bene…t from investing by (qt 1)(qt + Rt ): Equation (31) comes from the de…nition of qt : In fact, after multiplying both sides by Kt+1 ; equation (31) is equivalent to the asset pricing formula (19). To see this, aggregating (29) gives total investment of It = (qt + Rt ) Kt : (32) The aggregate capital follows as Kt+1 = It + (1 )Kt : It is easy to see that the aggregate dividend is given by Dt = Rt Kt (33) It . Multiply both sides of (31) by Kt+1 ; and we have qt Kt+1 = t+1 [Rt+1 Kt+1 + (1 )qt+1 Kt+1 + (qt+1 1) (qt+1 + Rt+1 )Kt+1 ] t = t+1 [Rt+1 Kt+1 It+1 + (1 )qt+1 Kt+1 + qt+1 It+1 ]; [Rt+1 Kt+1 It+1 + qt+1 Kt+2 ]; t = t+1 (34) t where the second line comes from aggregate investment equation (32) and the third line comes from the law of motion of aggregate capital (33). Comparing it with (19) yields qt Kt+1 = Qt : 4.2 (35) A Single Household’s Decision Problem For household i, the …rst-order conditions are list by (5) to (8). Following Wen (2009, 2011) by denoting Hit = (Qt + Dt ) ait +Wt nt (i)+sit as the total income of household i in period t, the following proposition shows that the income distribution is degenerate Hit = Ht and there exists an unique cuto¤ the borrowing constraint (4) will bind if and only if 17 it t. t such that Proposition 2 Given Wt and Rf t , Ht and t = Rf t (Ht + Bt ) Et and Wt Given t are jointly determined by the following two equations: t Z ; Wt+1 (36) max( ; t ) f ( )d = : Ht + Bt (37) and Ht , the optimal consumption of household level can be summarized as cit = min it ; 1 (Ht + Bt ) ; (38) t and Lagrangian multipliers and it it are given by it it = max( it ; t ) ; Ht + Bt (39) max( it t ; 0) : Ht + Bt (40) = Finally the risk free rate and the asset price are determined, respectively, by 1 = Rf t Ht + B t Ht+1 + Bt+1 Z max( ; t+1 )f ( )d ; (41) Qt+1 + Dt+1 Ht+1 + Bt+1 Z max( ; t+1 )f ( )d : (42) t and t Ht + Bt Qt = Proof. See Appendix A.2. Similar to Wen (2009, 2011),15 with quasilinear preferences, the household can achieve a target wealth Ht in period t to bu¤er the idiosyncratic preference shocks through adjusting his labor supply. The quasilinear utility function implies that the marginal disutility of acquiring additional labor income is a constant (equal to Wt ) for all households. Since the preference shock is i.i.d, the expected marginal utility from consumption depends only on the total wealth in period t (not on the history of idiosyncratic shocks). The household supplies his labor to a point where the marginal disutility equals the expected gains in the marginal utility, which implies a common target wealth for all households. 15 Also see Wen (2010). 18 In the absence of aggregate uncertainty, equations (41) and (42) then imply Qt = Qt+1 + Dt+1 : Rf t (43) Namely, the risk free rate is the proper discounting factor for the …rms. With the decision rules of the …rms and the households in hand, we are ready to characterize the aggregate equilibrium by a set of nonlinear equations below. 4.3 Aggregation Equation (14) implies that the capital-labor ratio is the same for all …rms, so Kjt =Njt = Kt =Nt . It follows that the aggregate production is given by Yt = Kt Nt1 ; (44) and the two factor prices Rt , and wt are given by Rt = Yt Kt Wt = (1 ) (45) and Yt : Nt (46) Since the wealth distribution is degenerate (Hit = Ht ), aggregating Hit yields Z Hit di = Ht = (Qt + Dt ) Z ait di + Wt Nt + Z sit di = Qt + Dt + Wt Nt = qt Kt+1 + Yt The second line comes from the fact that R It : (47) ait di = 1 and fact that Qt = qt Kt+1 by equation (35), Dt = Rt Kt R sit di = 0; the third lines come from the It and Rt Kt + Wt Nt = Yt by equations (45) and (46). Aggregating equation (38) yields the aggregate consumption with heterogenous households and borrowing constraints as Ct = (Ht + Bt ) Z min( ; 1)f ( )d ; (48) t which resembles a Keynesian consumption function, with the propensity to consumption equal to R min( ; 1)f ( )d < t 1. The reason why the propensity to consume is less than one is that idiosyncratic preference shocks and 19 borrowing constraints give the households a precautionary motive to accumulate wealth. Equation (31) and aggregate (43) imply qt = 1 [Rt+1 + (1 Rf t )qt+1 + (qt+1 1)(qt+1 + Rt+1 )] (49) in the absence of aggregate uncertainty. The equilibrium can now be characterized in the model by the following proposition. Proposition 3 The equilibrium path of the model is characterized by dynamic movements of 8 aggregate variables {Ct ; It ; Yt ; Nt ; Kt+1 ; Rf t ; qt ; "t }, which can be solved by the system of 8 nonlinear equations: ( t+1 ) ( t) = Rf t ( Ct Ct+1 ( t ) ( t+1 ) (1 Ct+1 ) t+1 ) (50) Yt = Nt (51) Yt = Kt Nt1 Kt+1 = (1 (52) )Kt + It (53) Ct + It = Yt It = qt = Yt+1 1 [ + (1 Rf t Kt+1 Yt Kt Kt qt + )qt+1 + (qt+1 Ct = (qt Kt+1 + Yt where ( t) = R min( ; t )f ( )d , ( t+1 ) = R (54) (55) 1)(qt+1 + Yt+1 )] Kt+1 (56) It + Bt ) G( t ); max( ; 1)f ( )d > 1, and G( t ) = t (57) R min( ; 1)f ( )d . t Equation (50) follows from equation (48). Equation (51) is derived from (37) by replacing Ht + Bt by the aggregate consumption and replacing the real wage by equation (46). Equation (52) is identical to equation (44). Equation (53) describes how aggregate capital evolves, and equation (54) is the aggregate resource constraint. Equations (50) to (54) are similar to those of a standard RBC model. Equation (55) is the aggregate investment equation. The …nancial frictions introduce two new variables, qt and t, which are determined by equations (56) and (57). Equation (56) is obtained by substituting out Rt+1 using equation (45) on the right hand side of equation (49). Finally combining equation (47) and equation (48) yields equation (57). 20 5 The Saving Rate and Financial Development In this section, we discuss the relationship between the aggregate saving rate and …nancial development in the steady state. The saving rate is measured by the investment to output ratio, s = and the steady state ratio b I Y . The parameter B=qK measure the …nancial development on the …rm side and household side, respectively. To better understand the hump-shaped relationship between …nancial development and the saving rate, we consider four di¤erent scenarios: (i) the case without any …nancial constraints (the …rst-best allocation), (ii) the case with …nancial constraints only on …rms, (iii) the case with …nancial constraints only on households, and (iv) the case with both …rms and households subject to …nancial constraints. 5.1 The First-Best Allocation Consider …rst an economy without any …nancial frictions. The two borrowing constraints (4) and (17) then nerve bind. In such a case, equations (5), (6), and (7) imply that it cit = jt cjt ; (58) for any i and j. The aggregate consumption can be obtained by integrating cjt over j, which gives Ct = Z 1 cjt dj = it Ct . cit it 0 or cit = 1 Z 1 jt dj = 1 cit ; (59) it 0 Each individual’s consumption is proportional to the idiosyncratic shock he faces. Equation (7) then becomes 1 Ct 1 = Rf t Ct+1 and equation (5) can be rewritten as these two equations are equivalent to equations (50) and (51) if t 1 Ct Wt and t+1 other hand, if the borrowing constraint (28) does not bind, then qt = 1. imply that Rf t = Yt+1 Kt+1 + (1 = . It is easy to see that are equal to max . On the Equation (56) would then )qt+1 . Appendix A.3 shows that such an allocation is also the …rst-best allocation for a central planner. And at the steady state, Rf = 1 = Y +1 K : (60) The steady-state saving rate can then be determined as s = I K = = Y Y 1 which is identical to that of a standard RBC model. 21 (1 ) ; (61) The …nancial constraints on households and …rms create two wedges between the discount factor ; the interest rate Rf , and the marginal product of capital. First, notice that equation (50) gives a steady-state interest rate of 1 1 < ; ( ) Rf = since (62) ( ) > 1 as long as the borrowing of some households is constrained. As we will show in section 5.3, this tends to increase the saving rate above the …rst-best saving rate due to precautionary saving (as Aiyagari (1994) has shown). If q > 1 at the steady state, then …nancial constraints on the …rm side create a gap between the interest rate and the marginal product of capital. At the steady state, with q > 1 equation (56) implies Rf = Y 1 + (q K q 6= Y +1 K 1) +1 + (q 1) (63) : We will show that …nancial constraints on the …rm side alone will reduce the saving rate below the …rstbest level in section 5.2. The intuition is suggested by equation (63): …nancial constraints on …rms may reduce the interest rate for given marginal product of capital, so households have less incentive to save. 5.2 Financial Constraints only on Firms Assume that only …rms but not households are subject to …nancial constraints. As a result, Rf = 1 at the steady state as discussed in section 5.1. We will show that the aggregate saving rate in this case increases with the level of …nancial development measured by : Given that Rf = 1 at the steady state, equation (56) becomes q= [ Y + (1 K )q + (q 1)(q + Y )]: K (64) The following proposition shows that the saving rate is an increasing function of …nancial development level if it is below some critical value . Proposition 4 There exists a critical value q= ( = 1 1 1 + 1, such that if > 1 if 1 +1 > ; (65) and the steady state saving rate is given by I s= = Y ( 1= 1+ 1= 1+ 1= 1+ 22 if if > (66) Proof. See Appendix A.4. Recall that a …rm’s borrowing constraint (28) will be binding with q > 1. It is then clear that the saving rate is increasing in according to equation (66). Borrowing constraints on …rms will prevent them from undertaking investments that yield positive net present value, a problem commonly referred as the underinvestment problem as discussed in the introduction. To see that, …rst calculate the gross marginal production of capital MPK = Y +1 K 1 = (1 ) +1 + 1; (67) which is the social gross return on one unit of capital. One unit of investment can yield output of the next period, and after depreciation the consumption value of this unit is 1 MPK > if 1 = Rf ; Y K in . Notice that (68) . This di¤erence between capital’s social return and the interest rate prevents households from investing in capital to its socially optimal level despite that capital yields high social return. At steady state, the households need to pay q to purchase one unit of capital (indirectly through buying stock). While the social cost of building one unit of capital is one. In other words, the households pay a premium to acquire capital, which weakens their incentive to save. An increase in leads to a decrease in q, which increases the households’incentive to save. 5.3 Financial Constraints only on Households When only households are subject to the …nancial constraints, the Tobin’s q, qt , equals to 1, and equation (56) becomes Rf = M P K = Y +1 K at steady state. Solving for the saving rate as a function of the interest rate s= I = Y Rf 1+ : (69) The saving rate has a negative slope with respect to the user’s cost of capital (the interest rate). Since …rms are the demanders for savings, equation (69) can be considered the demand function for loanable funds. Equations (50) and (57) then lead to another relationship between saving and the real interest 23 rate from the suppliers of loanable funds. Equation (50) implies that Z max( ; 1)f ( )d = which de…nes an implicit function for the cuto¤ abuse of notation, de…ne such function as = 1 1 ; Rf (70) as a function of the interest rate Rf . With a slight (Rf ). It is easy to see that @ @Rf > 0: A higher interest rate induces households to save more and hence reduces the probability of being liquidity constrained. Equation (57) implies that at the steady state the saving rate is s= 1 1=G( where = ) 1 (1 + b) + ; (71) (Rf ) is increasing in Rf , and G( ) de…ned in proposition 3 is the marginal propensity to consume. The saving rate increases with the interest rate. Since the households are the suppliers of loanable funds, equation (71) describes the supply curve for funds with respect to the interest rate. The demand curve (69) and supply curve (71) are plotted in …gure 2. The steady state value of the interest rate Rf and the saving rate s can be found at the intersection of these two curves. An increase in b, which relaxes the households’borrowing constraint and reduces the incentive for households to save, shifts the supply curve inward and moves the equilibrium from point A to point B. So the steady state saving declines and the interest rate increases. The following proposition summarizes the result. Figure 2: The e¤ect of relaxing in the households’borrowing constraint on the saving rate Proposition 5 If only households are subject to …nancial constraint, the saving rate at steady state will decrease with increasing …nancial development: 24 5.4 Financial Constraints on both Households and Firms When both households and …rms are subject to …nancial constraint, the aggregate equilibrium is characterized by the system of equations from (50) to (57). In order to study the impact of …nancial constraints on the saving rate at steady state, we must …rst derive the relationships between saving/investment rate and the interest rate from the perspective of both households and …rms. As in the previous section, the demand and supply curves can again be used to study the impact of these two borrowing constraints on saving. Derive …rst the demand curve for saving using equation (53), (55), and (56). This gives the relationship between saving rate and the interest rate from the demand side: s= 1 I = Y 1 1 + = (Rf 1) ; (72) which decreases with interest rate Rf . An increase in Rf , representing a higher user’s cost of capital, depresses demand for investment (saving). The supply curve can be derived in similar steps as in section 5.3. First, equation (70) still holds, so = (Rf ) and @ @Rf > 0 for the reasons discussed in section 5.3. The next step is to use equation (57) to solve for the saving rate as a function of the interest rate from the supply side. After re-arranging terms, as an intermediate step, Y G( ) = q (1 + b) + ; K 1 G( ) where = (73) (Rf ), and G( ) is de…ned in proposition 3. q can then be derived as a function of the interest rate from equation (56) 1 q = q(Rf ) = + Rf 1 ; (74) which decreases with the interest rate. The intuition is simple: from the households’point of view, the value of one unit of capital, q; is the sum of the present value of all future dividend ‡ows. An increase in the interest rate reduces the present value and hence reduces q. Finally replace K by I= at steady state and q by (74), to obtain the supply of saving s as (s) Recall that 1 = G( )= q(Rf ) (1 + b) + 1: 1 G( ) is increasing with Rf and G( ) is decreasing with , so G( ) decreases with Rf . (75) G( )= 1 G( ) thus decrease with Rf . By equation (74), q is also decreasing with Rf . So the right hand side as a whole is decreasing with Rf . We can therefore conclude that the total saving rate, the inverse of the right hand side, is an increasing function of the interest rate. 25 Figure 3: The e¤ect of relax in household’s borrowing constraint on the saving rate Figure 3 shows the e¤ect of the households’borrowing constraints on saving rate. The relationship between interest rate and saving rate implied by equation Equation (72) is graphed by the downward sloping curve labeled D. The real interest rate, Rf , is shown on the vertical axis. Saving rate, s, is measured on the horizontal axis. Equation (75) is graphed by the upward sloping curve labeled S. The equilibrium interest rate and saving rate are at the intersection of the demand and supply curves. The initial steady state equilibrium is at point E. According to equation (75), an increase in b reduces the saving rate for any given interest rate Rf . As the borrowing constraint relaxes, there is less need to save to self-insure against the idiosyncratic preference shocks. In other words, an increase in b shifts the supply curve to the left from S to S 0 . The new steady-state equilibrium is at point E 0 . Interest rate Rf increases, but the saving rate, s, decreases. 26 Figure 4: The e¤ect of …rm’s borrowing constraints on saving rate The e¤ect of changing the …rms’ borrowing constraint is illustrated in Figure 4. First, by equation (72), an increase in increases …rm investment at any given interest rate. It is easy to understand since …rms’investments are constrained by their borrowing capacities. In other words, an increase in the demand curve to the right from D to equation (74), an increase in D0 . But an increase in shifts also changes the supply curve. By reduces Tobin’s q, the households’cost of obtaining one unit of capital, so it increases the households’incentive to save. As is more evident from equation (75), an increase in increases the saving rate from the supply side. In other words, an increase in also shifts the supply curve to the right from S to S 0 . The new steady state equilibrium is at point E 0 . The saving rate increases for sure, but the e¤ect on the interest rate is not clear. Proposition 6 summarizes these results. Proposition 6 The equilibrium saving rate can be expressed as a function of b and , namely, s = s( ; b) with @s @ > 0 and @s @b < 0 if borrowing constraints on both households and …rms bind. The above proposition also suggests that if …nancial development increases b and simultaneously, the overall e¤ect on the saving rate will be ambiguous. As increase in the households’borrowing reduces the saving rate but an increase in the …rms’borrowing increases it. A numerical example in section 5.5 will now illustrate how it is possible for this model to explain the observed hump-shaped relationship between …nancial development and the saving rate. 27 5.5 Explaining the Hump-Shaped Relationship: A Numerical Example For the purpose of a numerical example16 , set the parameters in a way such that the model economy is best interpreted as an emerging economy with a high average saving rate. Set capital share in production at 0:42, depreciation rate at 0:012 and the discount factor at 0:999. These generate a high saving rate and low interest rate in the steady state, broadly consistent with the data on emerging Asian. Set the probability that each …rm receives an investment opportunity at 0:1 and assume the idiosyncratic preference shock (i) follows Pareto distribution with shape parameter = 1:15. These two assumptions make …rms’ and households’ demand for borrowing large, so they can potentially face severe …nancial constraints if the …nancial market is imperfect. We assume b and are driven by a common force f that indicates the …nancial development level : b = f; = a1 f + a2 (76) (77) where parameters fa1 ; a2 g are non-negative. Set a1 = 0:15 and a2 = 0:005 to induce a hump-shaped relationship in data. Figure 5: Financial development and the saving rate Figure 5 plots the steady-state saving rate for di¤erent values of f 17 . The dashed line shows the …rstbest saving rate. It is evident from Figure 5 that …nancial underdevelopment can induce either oversaving 16 A quantitative exercise would be more desirable if it were possible. A careful calibration exercise, however, requires micro data on the dispersion of household consumption and incomes to measure the idiosyncratic risks and the consumer credit to GDP ratio to measure to what extent households are …nancially constrained. Due to data unavailability, we leave this as a future research topic. 17 Notice that f has one-to-one mapping with the credit to GDP ratio. 28 or undersaving. Figure 5 shows a clear hump-shaped relationship between …nancial development and the aggregate saving rate. The saving rate is about 25% when both households and …rms can barely borrow. The saving rate starts to increase as the borrowing constraints relax. The peak in the saving rate is close to 51%. Interestingly, with a higher level of …nancial development, the saving rate can double. This model thus puzzles of China’s saving rate discussed by Wen (2011). While the saving rate in China was less than 30% in late 70s, but has climbed to 51.8% in 2008 along with improvements in …nancial markets as many studies suggest.18 Figure 6: Financial development and saving rates for alternative cases Figure 6 shows that a hump-shaped relationship is not possible if only …rms or only households are subject to credit constraints. As expected, the left panel of …gure 6 shows that saving rates increases monotonically with the level of …nancial development if only …rms are constrained. While the left panel shows that saving rate monotonically decreases with …nancial development. Neither prediction is validated by the data, nor are they consistent with the recent pattern in emerging Asia. Although the left panel of the …gure is consistent with the fact that saving rates increases with …nancial development in emerging Asian, it suggests that saving rates in emerging Asia will be systemically lower than in economies with similar cultures but better …nancial development, which is counterfactual. It is well known that China has had a much higher saving rate than that in South Korea, Japan or Taiwan in recent years. The right panel shows that …nancial underdevelopment can indeed generate very high saving rates. But with only 18 Financial markets in China has developed in many ways. Ma and Wang (2010) document that bank loans to Chinese households have expanded substantially, reaching 15% of the total outstanding bank loans lately from less than 1% in the late 1990s. 29 households being …nancial constrained, such predictions are prone to the criticism by Wei and Zhang (2011) using time series evidence. 6 Conclusion This study has documented a hump-shaped relationship between …nancial development and the national saving rate. It has shown that …nancial development increases saving if only …rms are …nancially constrained, but reduces saving if only households are borrowing constrained. When both …rms and households are …nancially constrained, our theoretical model can predict the observed hump-shaped relationship between …nancial development and the aggregate saving rate. This model also shows that …nancial frictions can cause either undersaving or oversaving depending on the stage of …nancial development. A natural policy recommendation would be that the optimal taxation on capital income may either be negative (in the case of undersaving) or positive (in the case of oversaving). Given its enormous welfare consequences, a complete more characterization of the optimal capital income taxation should be a very important research topic for future. With both …rms and households …nancially constrained, the model also generates another interesting prediction: the coexistence of high returns on capital and low real interest rates, which is consistent with the evidence from emerging economies.19 An interesting implication of this is that if these economies were open to countries with more advanced …nancial markets, …nancial capital would ‡ow out from the developing countries to the developed countries, while physical capital will ‡ow (in terms of FDI) from the developed countries to the developing ones to enjoy high capital returns (see evidence by Ju and Wei (2011)). Explaining such two-way ‡ows in an open economy model will be another worthwhile research topic for future. 19 See Wen (2009) for an alternative model to explain this fact. Wen (2009) assumes an imperfectly competitive (stateowned) banking sector generating a spread between the deposit rate and the loan rate. Here we obtain the same e¤ect by allowing …nancial constraints on the …rm side. 30 References [1] Abiad, Abdul, Enrica Detragiache, and Thierry Tressel, 2008. "A New Database of Financial Reforms," IMF Working Paper. [2] Aiyagari, S Rao, 1994. "Uninsured Idiosyncratic Risk and Aggregate Saving," The Quarterly Journal of Economics, vol. 109(3), pages 659-84. [3] Albuquerque Rui and Hugo A. Hopenhayn, 2004. "Optimal Lending Contracts and Firm Dynamics," Review of Economic Studies, vol. 71(2), pages 285-315. [4] Arellano, M., and S. Bond, 1991. "Some Tests of Speci…cation for Panel Data: Monte Carlo Evidence and an Application to Employment Equations," Review of Economic Studies, vol. 58, pages 277-297. [5] Arellano, M., and O. Bover, 1995. "Another Look at the Instrumental Variable Estimation of ErrorComponent Models," Journal of Econometrics, vol. 68, pages 29-51. [6] Baltagi, B. H. and Li, D., 2002. "Series Estimation of Partially Linear Panel Data Models with Fixed E¤ects," Annals of Economics and Finance, vol. 3, pages 103-116. [7] Beck, Thorsten, Berrak Buyukkarabacak, Felix Rioja and Neven Valev, 2012. "Who Gets the Credit? And Does It Matter? Household vs. Firm Lending Across Countries," The B.E. Journal of Macroeconomics, Berkeley Electronic Press, vol. 12(1), pages 2. [8] Bernanke, Ben and Gertler, Mark, 1989. "Agency Costs, Net Worth, and Business Fluctuations," American Economic Review, vol. 79(1), pages 14-31. [9] Bernanke, Ben S. and Gertler, Mark & Gilchrist, Simon, 1999. "The …nancial accelerator in a quantitative business cycle framework," Handbook of Macroeconomics, in: J. B. Taylor and M. Woodford (ed.), Handbook of Macroeconomics, edition 1, volume 1, chapter 21, pages 1341-1393 Elsevier. [10] Blundell, Ricard and Stephen Bond, 1998. "Initial Conditions and Moment Restrictions in Dynamic Panel Models," Journal of Econometrics, vol. 87(1), pages 115-143. [11] Cameron, A. Colin and Pravin K. Trivedi, 2009. "Microeconometrics Using Stata," Stata Press. [12] Chamley, Christophe, 1986. "Optimal Taxation of Capital Income in General Equilibrium with In…nite Lives," Econometrica, vol. 54(3), pages 607-22. [13] Cochrane, John H, 1991. " Production-Based Asset Pricing and the Link between Stock Returns and Economic Fluctuations," Journal of Finance, vol. 46(1), pages 209-37. 31 [14] Cook, Christopher J., 2003. "Does Financial Depth Improve Aggregate Savings Performance? Further Cross-Country Evidence," Review of Development Economics, vol. 7(2), pages 248–265 [15] Cooley Thomas and Ramon Marimon & Vincenzo Quadrini, 2004. "Aggregate Consequences of Limited Contract Enforceability," Journal of Political Economy, vol. 112(4), pages 817-847. [16] Gale, Douglas and Hellwig, Martin, 1985. "Incentive-Compatible Debt Contracts: The One-Period Problem," Review of Economic Studies, vol. 52(4), pages 647-63. [17] Hansen, Lars Peter and Richard, Scott F, 1987. "The Role of Conditioning Information in Deducing Testable," Econometrica, vol. 55(3), pages 587-613. [18] Holmstrom, Bengt and Tirole, Jean, 1997. "Financial Intermediation, Loanable Funds, and the Real Sector," The Quarterly Journal of Economics, vol. 112(3), pages 663-91. [19] Horioka, C. Y., and A. Terada-Hagiwara. 2010. The Determinants and Long-term Projections of Saving Rates in Developing Asia. ADB Economics Working Paper Series No. 228, Asian Development Bank, Manila. [20] Horioka, C. Y., and T. Yin. 2010. “A Panel Analysis of the Determinants of Household Saving in the OECD Countries: The Substitutability of Social Safety Nets and Credit Availability.” Institute of Social and Economic Research, Osaka University. [21] Hubbard, R. Glenn, 1998. "Capital-Market Imperfections and Investment," Journal of Economic Literature, American Economic Association, vol. 36(1), pages 193-225, March. [22] Inessa Love, 2003. "Financial Development and Financing Constraints: International Evidence from the Structural Investment Model," Review of Financial Studies, vol. 16(3), pages 765-791. [23] Ju Jiangdong and Shang-Jin Wei, Domestic Institutions and the Bypass E¤ect of International Capital Flow, American Economic Journal: Economic Policy, November 2010, pages 173-204 . [24] Kimball, Miles S, 1990. "Precautionary Saving in the Small and in the Large," Econometrica, vol. 58(1), pp 53-73. [25] King, Robert G. and Levine, Ross, 1993. "Finance, entrepreneurship and growth: Theory and evidence," Journal of Monetary Economics, vol. 32(3), pages 513-542. [26] Kiyotaki, Nobuhiro and Moore, John, 1997. "Credit Cycles," Journal of Political Economy, vol. 105(2), pages 211-48. 32 [27] Lagos, Ricardo and Randall Wright, 2005. "A Uni…ed Framework for Monetary Theory and Policy Analysis," Journal of Political Economy, vol. 113(3), pages 463-484. [28] Leland, H. E. (1968): “Saving and Uncertainty: The Precautionary Demand for Saving.” Quarterly Journal of Economics, vol. 82, pages 465–473. [29] Levine, Ross, Norman Loayza and Thorsten Beck, 2000. "Financial intermediation and growth: Causality and causes," Journal of Monetary Economics vol. 46, pages 31-77. [30] Libois, Francois and Vincenzo Varardi, 2012. "Semiparametric Fixed-E¤ects Estimator," Department of Economics, University of Namur, Working Paper Series, WP 1201. [31] Loayza, Norman, Klaus Schmidt-Hebbel and Luis Serven, 2000. "What Drives Private Saving across the World?" The Review of Economics and Statistics, Vol. 82(2), pages 165-181. [32] Lucas, Robert Jr., 1988. "On the mechanics of economic development," Journal of Monetary Economics, Elsevier, vol. 22(1), pp 3-42. [33] Guonan Ma and Wang Yi, 2010, "China’s high saving rate: myth and reality", BIS Working Papers. [34] Miao, Jianjun and Wang, Pengfei, 2010, "Credit Risk and Business Cycles". [35] Miao, Jianjun and Wang, Pengfei, 2011, "Bubbles and Credit Constraints". [36] Mendoza, Enrique G., Vincenzo Quadrini and Jose-Victor Rois-Rull, Financial Integration, Financial Development, and Global Imbalances, Journal of Political Economy, June 2009, vol. 117(3), pp. 371416. [37] Modigliani Franco and Shi Larry Cao, 2004. "The Chinese Saving Puzzle and the Life-Cycle Hypothesis," Journal of Economic Literature, vol. 42(1), pages 145-170. [38] Myers, S. (1977), Determinants of Corporate Borrowing, Journal of Financial Economics, 5, 147-75. [39] Park, D., and K. Shin. 2009. Saving, Investment, and Current Account Surplus in Developing Asia. ADB Economics Working Paper Series No. 158, Economics and Research Department, Asian Development Bank, Manila. [40] Ricardo J. Caballero, Emmanuel Farhi, and Pierre-Olivier Gourinchas, An Equilibrium Model of “Global Imbalances” and Low Interest Rates, American Economic Review 2008, vol. 98:1, pages. 358–393. [41] Romer, Paul M., 1986, "Increasing Returns and Long-Run Growth", Journal of Political Economy, vol. 94(5), pages 1002-1037. 33 [42] Roodman, David, 2009a. "A Note on the Theme of Too Many Instruments," Oxford Bulletin of Economics and Sratsitcis, vol. 71(1), pages 135-158. [43] Roodman, David, 2009b. "How to Do xtabond2: an Introduction to Di¤erence and System GMM in Stata," The Stata Journal, vol. 9(1), pages 86-136. [44] Schmidt-Hebbel, Klaus and Luis Serven, 2002. "Financial Liberalization, Saving, and Growth," Unpublished manuscript. [45] Solow, Robert M., 1956. "A Contribution to the Theory of Economic Growth", The Quarterly Journal of Economics, vol. 70(1), pages 65-94. [46] Song Zheng, Kjetil Storesletten and Fabrizio Zilibotti, 2011. "Growing Like China," American Economic Review, vol. 101(1), pages 196-233. [47] Stiglitz, Joseph E. and Weiss, Andrew, 1981. "Credit Rationing in Markets with Imperfect Information," American Economic Review, vol. 71(3), pages 393-410. [48] Townsend, Robert M., 1979. "Optimal contracts and competitive markets with costly state veri…cation," Journal of Economic Theory, vol. 21(2), pages 265-293. [49] Wang, Pengfei and Yi Wen, 2009. "Financial development and diverging trends in economic volatility: A uni…ed explanation," Federal Reserve Bank of St. Louis Working Paper 2009-022C. [50] Wang, Pengfei and Yi Wen, 2012. "Hayashi meets Kiyotaki and Moore: A theory of capital adjustment costs," Review of Economic Dynamics, vol. 15, pages 207-225. [51] Wei Shang-Jin and Zhang Xiaobo, 2011. "The Competitive Saving Motive: Evidence from Rising Sex Ratios and Savings Rates in China," Journal of Political Economy, vol. 119(3), pages 511-564. [52] Wen, Yi, 2009. "Saving and growth under borrowing constraints: Explaining the high saving rate puzzle," Federal Reserve Bank of St. Louis Working Paper 2009-045C. [53] Wen, Yi, 2010. "Liquidity demand and welfare in a heterogeneous-agent economy," Federal Reserve Bank of St. Louis Working Paper 2010-009A. [54] Wen, Yi, 2011. "Making sense of China’s astronomical foreign reserves," Federal Reserve Bank of St. Louis Working Paper 2011-006A. 34 Appendix A A.1 Proofs Proof of Proposition 1 Suppose qt > 1, …rm j then always chooses to invest as much as possible when investment opportunities arise, which is Ijt = qt Kjt + Rt Kjt : (A.1) Substituting this decision rule to equation (26) and comparing terms, gives equation (30). Using the de…nition qt = t+1 t vt , easily yields qt = t+1 [Rt+1 + (1 )qt+1 + (qt+1 1)(qt+1 + Rt+1 )]; (A.2) t which is equation (31). Q.E.D. A.2 Proof of Proposition 2 Prove …rst that the total wealth Hit = (Qt + Dt ) ait + Wt nt (i) + sit is degenerate. In the second subperiod, the households’ consumption, saving, and stock holdings, can be written as a function of their wealth Hit , liquidity shock it , and the aggregate variables, it it = cit = it ct (Hit ; (A.3) it ) Equation (7) can then be written as it ct (Hit ; it ) = Rf t Ei [ = Rf t where the second line has used equation (5). De…ne it Hit + Bt Since cit Hit + Bt , then it ct (Hit ; it it ) Hit +Bt for it , + + it ; Wt+1 (A.4) it such that = Rf t it . it it+1 ] Wt+1 : By (A.4) and (A.5), (A.5) it > 0. Or the borrowing constraint (4) binds, and the household’s consumption cit = Hit + Bt . On the other hand, it ct (Hit ; it ) = 35 it Hit + Bt ; it = 0, so (A.6) or cit = (Hit + Bt ) it it : Since cit Hit + Bt , then it . it Finally, using the consumption rule derived above, we can rewrite equation (5) as Wt "Z it < it f ( )d + Hit + Bt Equations (A.5) and (A.7) jointly determine Z > it Hit + Bt # f ( )d = 1: and Hit . It is evident that Hit and it on aggregate variables in the economy. Hence Hit = Ht , and it = t. (A.7) it only depend Dropping the subscripts from equation (A.5) yields equation (36). Writing equation (A.7) more compactly and dropping the subscripts gives equation (37). Equations from (38) to (42) are straightforward to obtain. Q.E.D. A.3 The First Best Allocation The social planner’s problem is to maximize the total welfare: max cit ;nit ;It ( max nit Z Ei i " 1 X t # it log cit ) nit di t=0 (A.8) with the following resource constraint Z cit di + It Z Kt 1 nit di ; (A.9) and the aggregate capital evolves according to Kt+1 = (1 )Kt + It : (A.10) The social planner decides individual i0 s labor supply nit before observing individual i0 s idiosyncratic shock it as in the decentralized economy. Since the labor decision is made before the realization of idiosyncratic shock, for each household i, nit is the same. Denote the optimal aggregate labor as Nt . It satis…es = (1 ) t Yt : Nt (A.11) For the consumption cit ; it cit for all i. De…ne the aggregate consumption i; we have R = t; (A.12) cit di = Ct ; and integrate the last equation with respect to 1 = Ct 36 t: (A.13) Here we assume E ( it ) = 1: The …rst order condition with respect to capital is then Yt+1 +1 Kt+1 Ct Ct+1 1= : (A.14) At steady state, the saving rate can be calculated as I K = = Y Y 1 A.4 (1 : ) (A.15) Proof of Proposition 4 Suppose > : We aim to show that at the equilibrium, the steady state q equals 1; and the liquidity constraint on investment (27) is never binding. Equation (64) implies the saving rate is I K = = Y Y 1= In addition, it is easy to show that if In contrast, if < > ; I K < : 1+ ( + R), i.e. the liquidity constraint is not binding. ; q > 1 and aggregating the the …rms’optimal investment decisions we obtain It = [qt + Rt ] Kt Kt+1 = It + (1 (A.16) )Kt (A.17) These last three equations together with (64) imply that at the steady state R= Y = K 1 +1 1 + : (A.18) ; (A.19) We can immediately get the saving rate I = Y 1 1 + +1 1 +1 and q = = 1 ( R) (1 ) +1 This establishes the proof. Q.E.D. 37 : (A.20) B Sample description B.1 Variables used in the gross saving regression Variable Gross domestic saving rate Domestic credit to private sector Aged dependency ratio Youth dependency ratio Consumer price in‡ation GDP de‡ator in‡ation Real interest rate Per-capita GDP Per-capita GDP growth Public expenditure Current account balance Life expectancy B.2 Data source WDI WDI WDI WDI WDI WDI WDI WDI WDI PWT 7.0 WDI WDI Note % of GDP % of GDP Population aged 65 and above/total population Population aged 0-14/total population Annual rate Annual rate Annual rate In constant 2000 US dollar Annual rate % of GDP % of GDP, positive if surplus Life expectancy at birth, in years Variables used in the household saving regression Variable Household saving rate Household liabilities Aged dependency ratio Youth dependency ratio Consumer price in‡ation GDP de‡ator in‡ation Real interest rate Per-capita household disposable income growth Per-capita household disposable income Social expenditures Life expectancy Data source OECD OECD WDI WDI WDI WDI WDI OECD OECD OECD WDI Note % of household disposable income Loans Population aged 65 and above/total population Population aged 0-14/total population Annual rate Annual rate Annual rate Annual rate % of GDP Life expectancy at birth, in years Note: WDI stands for World Development Indicators from World Bank; PWT stands for PENN World table from University of Pennsylvania; OECD stands for OECD database from the OECD website. 38 B.3 Economies in the sample Country Australia Belgium Canada Czech Republic Denmark Estonia Finland France Germany Greece Hungary Iceland * Ireland * Israel * Italy Japan * Korea Luxembourg Mexico * Netherlands New Zealand Norway Poland Portugal Slovak Republic Slovenia Spain Sweden Switzerland United Kingdom United States China Hong Kong, China India Indonesia Malaysia Pakistan Philippines Singapore Thailand Vietnam World Bank Code AUS BEL CAN CZE DNK EST FIN FRA DEU GRC HUN ISL IRL ISR ITA JPN KOR LUX MEX NLD NZL NOR POL PRT SVK SVN ESP SWE CHE GBR USA CHN HKG IND IDN MYS PAK PHL SGP THA VNM Region or Group OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD, East Asia OECD, East Asia OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD OECD East Asia East Asia East Asia East Asia East Asia East Asia East Asia East Asia East Asia East Asia Note: OECD countries with asterisk are not included in the household saving regression, due to unavailability of data. 39 C Gross saving rate regression Table 2: Gross domestic saving regression: alternative estimators, East Asia subsample Estimator Regression Number of IV Collapsed IV Lagged saving rate Private credits Private credits sq. In‡ation rate Real interest rate Per-capita GDP growth Per-capita GDP (log) Per-capita GDP (log) sq. Current account surplus Public expenditures Aged dependency ratio Youth dependency ratio Life expectancy Wald test (p-val) Hansen-Sargan test (p-val) Wooldridge test (p-val) Arellano-Bond test (p-val): _1st-order _2nd-order _3rd-order Observations (No. of cn) (1) GMM-System Levels-Di¤s 312 No (2) GMM-System Levels-Di¤s 32 Yes (3) Within Deviation – – (4) RE Levels – – 0.851*** (0.0411) 0.0327** (0.0160) -0.00808* (0.00446) -0.0445 (0.0301) -0.0431 (0.0415) 0.241*** (0.0267) 0.00835 (0.0143) -0.000617 (0.000900) 0.121*** (0.0146) -0.0340 (0.0726) -0.106 (0.0759) 0.0172 (0.0390) -0.00293 (0.0338) 0.0000 0.936 – 0.805*** (0.0948) 0.109** (0.0531) -0.0271* (0.0160) -0.210** (0.102) -0.213* (0.127) 0.209 (0.129) -0.0205 (0.0645) 0.000999 (0.00402) 0.0275 (0.0676) -0.173 (0.183) -0.281 (0.233) 0.169 (0.111) 0.0193 (0.0491) 0.0000 1.000 – – – 0.140*** (0.0206) -0.0318*** (0.00817) 0.0696* (0.0420) 0.0257 (0.0544) 0.272*** (0.0546) 0.0986*** (0.0328) -0.000233 (0.00235) 0.227*** (0.0360) 0.214 (0.313) -1.490*** (0.158) -0.0317 (0.101) -0.296*** (0.0599) 0.0000 – 0.0002 0.261*** (0.0263) -0.0816*** (0.0115) 0.134** (0.0555) 0.0306 (0.0682) 0.673*** (0.0765) 0.183*** (0.0333) -0.0109*** (0.00207) 0.497*** (0.0464) -0.577*** (0.118) -0.470*** (0.157) -0.172** (0.0724) -0.225*** (0.0606) 0.0000 – 0.0002 0.025 0.831 0.248 265 (12) 0.009 0.615 0.320 265 (12) – – – 313 (12) – – – 313 (12) Notes: Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1; others see notes in Table 3-4. 40 Table 3: Gross domestic saving regression: alternative estimators, OECD subsample Estimator Regression Number of IV Collapsed IV Lagged saving rate Private credits Private credits sq. In‡ation rate Real interest rate Per-capita GDP growth Per-capita GDP (log) Per-capita GDP (log) sq. Current account surplus Public expenditures Aged dependency ratio Youth dependency ratio Life expectancy Wald test (p-val) Hansen-Sargan test (p-val) Wooldridge test (p-val) Arellano-Bond test (p-val): _1st-order _2nd-order _3rd-order Observations (No. of cn) (1) GMM-System Levels-Di¤s 701 No (2) GMM-System Levels-Di¤s 32 Yes (3) Within Deviation – – (4) RE Levels – – 0.794*** (0.0114) 0.0106** (0.00469) -0.00670*** (0.00208) 0.0231*** (0.00881) -0.0817*** (0.0106) 0.348*** (0.0166) -0.00769 (0.0195) 0.000521 (0.00102) 0.184*** (0.0140) 0.0143 (0.0158) -0.101*** (0.0159) -0.0395*** (0.00852) 0.0223 (0.0213) 0.0000 1.0000 – 0.830*** (0.0485) 0.0241 (0.0188) -0.0132** (0.00530) 0.00758 (0.0440) -0.0853 (0.0568) 0.279*** (0.0763) -0.0779 (0.0874) 0.00413 (0.00460) 0.142*** (0.0413) 0.0160 (0.0375) -0.0945*** (0.0318) -0.0155 (0.0259) 0.0966 (0.0964) 0.0000 0.492 – – – 0.0552** (0.0225) -0.0168** (0.00802) 0.0304 (0.0456) -0.136*** (0.0485) 0.200*** (0.0610) 0.329* (0.188) -0.0136 (0.0102) 0.285*** (0.0594) -1.101*** (0.263) -0.612*** (0.195) -0.278** (0.124) -0.767*** (0.274) 0.0000 – 0.0000 0.0513** (0.0230) -0.0165* (0.00848) 0.0239 (0.0465) -0.169*** (0.0516) 0.235*** (0.0578) 0.259 (0.176) -0.0107 (0.00996) 0.285*** (0.0570) -0.811*** (0.203) -0.583*** (0.190) -0.316*** (0.0996) -0.673*** (0.181) 0.0000 – 0.0000 0.0040 0.2748 0.1087 603 (27) 0.001 0.316 0.114 603 (27) – – – 731 (31) – – – 731 (31) Notes: i) Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1; ii) Two lags are used as intruments in the …rst-di¤erenced equation, …rst lag di¤erence is used as instruments in the level equation; iii) Financial development is also intrumented using the 41 …nancial development index prepared by Abiad et al. (2008). Table 4: Gross domestic saving regression: alternative estimators, full sample Estimator Regression Number of IV Collapsed IV Lagged saving rate Private credits Private credits sq. In‡ation rate Real interest rate Per-capita GDP growth Per-capita GDP (log) Per-capita GDP (log) sq. Current account surplus Public expenditures Aged dependency ratio Youth dependency ratio Life expectancy Wald test (p-val) Hansen-Sargan test (p-val) Wooldridge test (p-val) Arellano-Bond test (p-val): _1st-order _2nd-order _3rd-order Observations (Number of cn) (1) GMM-System Levels-Di¤s 806 No (2) GMM-System Levels-Di¤s 32 Yes (4) Within Deviation – – (5) RE Levels – – 0.783*** (0.0136) 0.0210*** (0.00676) -0.00706** (0.00310) -0.00451 (0.0107) -0.0814*** (0.0137) 0.313*** (0.0184) 0.00906 (0.00953) -0.000566 (0.000571) 0.169*** (0.0153) 0.0349* (0.0186) -0.0937*** (0.0191) -0.0256** (0.0114) 0.00596 (0.00902) 0.0000 0.381 – 0.846*** (0.0514) 0.0321** (0.0144) -0.0127*** (0.00491) 0.00472 (0.0613) -0.0662 (0.0844) 0.288*** (0.101) -0.0146 (0.0561) 0.000835 (0.00338) 0.0999* (0.0573) 0.0459 (0.0325) -0.0744 (0.0453) -0.00880 (0.0219) 0.0212 (0.0537) 0.0000 0.549 – – – 0.0925*** (0.0287) -0.0296*** (0.0100) 0.0332 (0.0424) -0.0644 (0.0510) 0.245*** (0.0495) 0.174** (0.0780) -0.00582 (0.00480) 0.226*** (0.0421) -0.824*** (0.271) -0.969*** (0.246) -0.110 (0.131) -0.265 (0.254) 0.0000 – 0.0000 0.105*** (0.0308) -0.0321*** (0.0101) 0.0403 (0.0431) -0.0866 (0.0553) 0.291*** (0.0484) 0.180*** (0.0658) -0.00779* (0.00404) 0.259*** (0.0403) -0.645*** (0.198) -0.861*** (0.203) -0.212** (0.0944) -0.251 (0.232) 0.0000 – 0.0000 0.0033 0.4178 0.3291 809 (37) 0.002 0.452 0.397 809 (37) – – – 976 (41) – – – 976 (41) Notes: i) Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1; ii) Two lags are used as intruments in the …rst-di¤erenced equation, …rst lag di¤erence is used 42 as instruments in the level equation; ii) Financial development is also intrumented using the …nancial development index prepared by Abiad et al. (2008). D Household saving rate regression In this section, we investigate how the household saving rate changes with household credits for the sample of OECD countries. We focus on OECD countries, simply because of data accessibility. For East Asian countries, there does not exist any single data source for us to retrieve all of the data needed in this regression. We collect the time series data of household saving rate and its potential explanatory variables for OECD economies from the OECD databases, whenever the data are available20 . As a result, we have a sample of 26 countries, with the time series ranging from 1995 to 2008, because the earliest observations on household …nance start from 1995. Consider the following reduced-form regression equation: shit = h 1 sit 1 + 0 h 2 Xit + h i + uhit ; (D.1) where sh denotes the household net saving rate, X h is a vector of explanatory variables re‡ecting volume of household credits, size of household …nance, household disposable income and its growth, age structure, social expenditures, rates of return and uncertainty, which we believe are important to household saving behavior, h denotes the country-speci…c e¤ects, and …nally uh is the error term. For the same reasons explained in the main text about gross saving rate regression, we choose to work with such a dynamic speci…cation. We use the the household liabilities to household disposable income ratio to measure the size of credits going to households. Table 5 summarizes the regression results, using three di¤erent estimators: the Arellano-Bond system estimator for the dynamic speci…cation in equation (D.1), the within estimator, and the RE estimator for static speci…cation without lagged dependent variable. It is evident from the table that household saving rate is decreasing in the volume of household credits. These results imply that households tend to save less when the level of household …nance development is high, which may justify our modeling of the household side in this paper. 20 We focus on OECD countries, simply because of data accessibility. For East Asian countries, there does not exist any single data source for us to retrieve all of the data needed in this regression. 43 Table 5: OECD household saving regression: with household credits, alternative estimators Estimator Regression Number of IV Collapsed IV Lagged household saving rate Household credits Per-capita household disp. inc. (log) Per-capita household disp. inc. (log) sq. Per-capita household disp. inc. growth In‡ation rate Real interest rate Social expenditures Aged dependency ratio Youth dependency ratio Life expectancy Wald test (p-val) Hansen-Sargan test (p-val) Wooldridge test (p-val) Arellano-Bond test (p-val): _1st-order _2nd-order _3rd-order Observations (Number of countries) (1) GMM-System Levels-Di¤s 19 Yes (2) GMM-System Levels-Di¤s 26 Yes (3) Within Deviation – – (4) RE Levels – – 0.724*** (0.109) -0.0262** (0.0107) -0.422** (0.215) 0.0224** (0.0101) -0.184 (0.284) 0.171 (0.128) 0.272*** (0.0844) -0.107 (0.158) 0.0234 (0.131) -0.119 (0.134) 0.469* (0.264) 0.0000 0.567 – 0.791*** (0.0598) -0.0255* (0.0138) -0.496** (0.231) 0.0265** (0.0109) 0.0711 (0.123) 0.233*** (0.0795) 0.328*** (0.0630) -0.0633 (0.141) -0.0349 (0.112) -0.0889 (0.131) 0.541* (0.280) 0.0000 0.560 – – – -0.0747*** (0.0183) -0.532 (0.694) 0.0353 (0.0374) 0.283*** (0.0653) 0.259** (0.119) 0.289** (0.124) 1.366 (0.875) -0.296 (0.320) 0.425* (0.214) -0.246 (0.491) 0.0000 – 0.0000 -0.0705*** (0.0160) -0.709 (0.534) 0.0422 (0.0300) 0.269*** (0.0597) 0.332*** (0.0979) 0.393*** (0.128) 0.618* (0.369) -0.296 (0.250) 0.271 (0.195) -0.0690 (0.449) 0.0000 – 0.0000 0.035 0.255 0.516 170 (24) 0.023 0.313 0.394 170 (24) – – – 229 (26) – – – 229 (26) Notes: i) Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1; ii) Results (1) in column 2 uses only one lag as instrument, Results (2) in column 3 uses only two lags as instrument in the di¤erenced equation; iii) The coe¢ cient on household credits in column 3 has p-value=0.064. 44 E Semiparametric estimation In this section, we apply the Baltagi-Li (2002) estimator for partially linear panel data models to our full sample. We choose this estimator among many others, simply because there exists a handy user-written Stata command xtsemipar to implement it. We consider the following semiparametric speci…cation: sit = 0 ~ it + g(zit ) + ~ i + u X ~it ; (E.1) ~ it contains those explanatory variables of saving rate explained in Section 2.2, except the measure where X of …nancial development (private credits), zit donotes the aggregate private credits to GDP ratio, ~ i is the unobserved country …xed e¤ects, and …nally u ~it is the error term. Thus, …nancial development enters non-parametrically while the other explanatory variables enter parametrically in this partially linear speci…cation. To eliminate the …xed e¤ects, we choose to work with the …rst-di¤erenced model: sit = where K 0 ~ it + [g(zit ) X g(zit 1 )] + u ~it ; (E.2) denotes the …rst di¤erence. Baltigi and Li (2002) propose using series pK (z) of dimension 1 to approximate g(z), where pK (z) denotes the …rst K elements of series fpj (z)gj=1;2;3;::: . The series of functions have the property that as K grows there exists a linear combination of pK (z) that can approximate any g(z) belonging to an additive class of functions arbitrarily well in mean square error. As a result, g(zit ) g(zit 1) can be approximated using pK (zit ) sit = Baltagi and Li (2002) show that 0 ~ it + X and pK (zit ) pK (zit pK (zit 1) 1) u ~it : + and equation (E.2) becomes: can be consistently estimated. Once (E.3) and are known, we can get the residuals and hence by equation (E.1), we have residuals = sit 0 ~ it X ~ i = g(zit ) + u ~it : (E.4) Thus, g(zit ) can be …tted by some standard univariate non-parametric regression of above residuals on zit . Table 6 reports the coe¢ cients on the other explanatory variables. It is evident that these results are consistent with what we get in Section 2, using parametric regression. Figure E.1 displays the …t of saving rate on private credits, using quartic B-spline smoothing. The graph looks similar if we set the power equal to 3, 5 or 6. It con…rms that the relationship between gross saving rate and …nancial development is hump-shaped. 45 Table 6: Gross saving rate regression: Baltagi&Li estimator, full sample In‡ation rate -0.0245 (0.0381) -0.0978*** (0.0351) 0.100*** (0.0297) 0.182 (0.109) -0.00201 (0.00606) 0.305*** (0.0349) -1.110*** (0.307) -0.761*** (0.231) 0.125 (0.121) -0.292* (0.149) 920 0.379 Real interest rate Per-capita GDP growth Per-capita GDP (log) Per-capita GDP (log) sq. Current account surplus Public expenditures Aged dependency ratio Youth dependency ratio Life expectancy Observations R-squared 0 -.05 -.1 Gross saving rate .05 .1 Note: Robust standard errors in parentheses***, p<0.01, ** p<0.05, * p<0.1 Scatterplots 0 .5 1 1.5 B-spline smooth 2 2.5 Private credits to GDP ratio Figure E.1: Nonparametric …t of gross saving rate on private credits, full sample 46
© Copyright 2024