World Saving1
Author:
Mr. Francesco Grigoli
Search for other papers by Mr. Francesco Grigoli in
Current site
Google Scholar
Close
,
Alexander Herman
Search for other papers by Alexander Herman in
Current site
Google Scholar
Close
, and
Klaus Schmidt-Hebbel https://isni.org/isni/0000000404811396 International Monetary Fund

Search for other papers by Klaus Schmidt-Hebbel in
Current site
Google Scholar
Close

Contributor Notes

This paper presents new evidence on the behavior of saving in the world, by extending previous empirical research in five dimensions. First, it is based on a very large and recent database, covering 165 countries from 1981 to 2012. Second, it conducts a robustness analysis across different estimation techniques. Third, the empirical search is expanded by including potential saving determinants identified by theory but not previously considered in the empirical literature. Fourth, the paper explores differences in saving behavior nesting the 2008-10 crisis period and four different country groups. Finally, it also searches for commonalities and differences in behavior across national, private, household, and corporate saving rates. The results confirm in part existing research, shed light on some ambiguous or contradictory findings, and highlight the role of neglected determinants. Compared to the literature, we find a larger number of significant determinants of saving rates, using different estimators, for different periods and country groups, and for different saving aggregates.

Abstract

This paper presents new evidence on the behavior of saving in the world, by extending previous empirical research in five dimensions. First, it is based on a very large and recent database, covering 165 countries from 1981 to 2012. Second, it conducts a robustness analysis across different estimation techniques. Third, the empirical search is expanded by including potential saving determinants identified by theory but not previously considered in the empirical literature. Fourth, the paper explores differences in saving behavior nesting the 2008-10 crisis period and four different country groups. Finally, it also searches for commonalities and differences in behavior across national, private, household, and corporate saving rates. The results confirm in part existing research, shed light on some ambiguous or contradictory findings, and highlight the role of neglected determinants. Compared to the literature, we find a larger number of significant determinants of saving rates, using different estimators, for different periods and country groups, and for different saving aggregates.

I. Introduction

What does consumption theory say about the main determinants of private saving decisions and which are the empirical measures that should be used to test for their relevance in explaining aggregate consumption/saving patterns? What determines the behavior of national, private, household, and corporate saving rates in the world? Did the exceptional depth of the Global Financial Crisis change the behavioral relationships of private saving and its determinants? And do saving determinants change across different country groups?

There is a small body of empirical saving studies using macroeconomic panel datasets that address some of these questions. A review of 15 empirical studies of mostly private saving rates reveals large differences in their sample size and coverage, data sources, saving rate definitions, model specifications, and estimation methodologies. Unsurprisingly, they also show large differences in empirical results that are difficult to reconcile.

This paper addresses limitations and contradictory findings of previous empirical research, extending it in five dimensions. First, it is based on a very large and more recent panel database for world saving, covering 165 countries from 1981 to 2012. This is almost four times the size of the most comprehensive panel study published to date by Loayza et al. (2000). Second, it conducts a robustness analysis across different estimation techniques. Third, the empirical search is expanded by including potential saving determinants identified by theory but not previously considered in the empirical literature. Fourth, the paper explores differences in saving behavior across time and space, nesting the 2008-10 crisis period and four different country groups. Finally, while this paper’s focus is on private saving, it also searches for commonalities and differences in behavior across national, private, household, and corporate saving rates.

Our results confirm some of the findings of the literature and unveil some novel features. Private saving rates are generally persistent and positively associated with income levels and income growth. Terms-of-trade improvements also contribute to a rise in saving through their effect on income. Permanent components of income and the terms of trade increase saving, and temporary parts of the terms of trade are saved to a larger extent than permanent parts. Saving is spurred by inflation, possibly due to precautionary motives. Increased credit availability, which is often associated with a process of financial liberalization, depresses private saving. A higher old-age dependency ratio reduces saving as the elderly finance their consumption needs with accumulated savings. Urbanization lowers private saving rates. Higher public saving reduces private saving, but exhibiting only partial Ricardian offsetting. Higher expected future growth has a positive effect on private saving, as does access to foreign borrowing. Finally, a higher share of young dependents reduces saving.

We also find that results differ across time periods and country groups. During the Global Financial Crisis, private saving inertia, income levels, and urbanization had diminished effects on saving. Compared to private saving in other country groups, saving in advanced economies is more responsive to income growth and almost non-sensitive to demographic variables, while low-income developing countries show a lower response of private saving to income growth and less persistent private saving rates. Private saving in oil exporters is positively associated with a larger old-age population share. High-growth Asian economies’ private saving rates are relatively more sensitive to real deposit rates.

We then replicate our empirical search for the national saving rate, with results that are largely in line with those reported for private saving. Finally, we check robustness of our results for private saving by estimating regressions separately on household and corporate saving. We also confirm many of the empirical findings reported for private saving at the household and corporate levels.

The paper structure is the following. In the next section we first review the determinants of private saving by briefly discussing the main consumption theories that drive the selection of empirical saving determinants. Then, we provide an overview of panel data studies on the behavior of private saving rates. Section III summarizes our data sources and construction and presents descriptive statistics, stylized facts on saving patters, and pairwise correlations with key potential determinants. Section IV outlines our empirical strategy, describing our choice of regression models. The empirical results are reported in Section V. Section VI concludes.

II. Private Saving Determinants

We review the determinants of private saving in two steps, extending previous surveys by Schmidt-Hebbel and Servén (1997) and Loayza et al. (2000). First, we briefly discuss the main theories of consumption and saving that drive the selection of potential regressors in empirical studies on aggregate consumption and saving. Then, we provide a compact overview of the empirical panel data literature on the behavior of aggregate private saving rates.

The starting point of modern theoretical research on consumption and saving is defined by two dominant models: the permanent-income hypothesis (PIH) and the life-cycle hypothesis (LCH). In contrast to the preceding Keynesian hypothesis (KH), in which consumption is determined by current income, the PIH focuses on a representative, infinitely-lived consumer who equates consumption to permanent income net of the present value of taxes (Friedman, 1957; Hall, 1978). As a variant of the PIH, the Ricardian-equivalence hypothesis (REH) derives permanent income as net of the present value of government spending, by making use of the representative consumer’s and the government’s budget constraints, which are linked by tax payments (Barro, 1974). If a large number of stringent (and empirically implausible) conditions are satisfied (Seater, 1993), the REH predicts that an increase in permanent government consumption is fully offset by lower private consumption.

The PIH assumption of homogeneous consumers contradicts observed consumer heterogeneity along several dimensions, including age, income, and access to borrowing. This leads to the main competitor of the PIH, the LCH, which introduces age-related consumer heterogeneity (Modigliani and Brumberg, 1954; Attanasio and Weber, 2010). Here, aggregate saving reflects the addition of saving by different age specific, finitely-lived cohorts who save for their old-age while working, dissave during retirement, and do not leave bequests. However, these LCH predictions are also at odds with the evidence. Planned bequests are empirically large and sensitive to income levels, implying elasticities of consumption to permanent income that are significantly lower than one.

Contradicting the PIH and LCH, consumption tends to exhibit excess sensitivity, i.e., its change is correlated with predictable changes in other variables.3 This is partly explained by the presence of durable goods (Caballero, 1991), consumption habits (external habits—Abel, 1990—or internal habits—Ferson and Constantinides, 1991), or consumer time inconsistency reflected in hyperbolic discounting (Laibson, 1997).4

Uncertainty can also explain in part the failures of the deterministic versions of the PIH-REH and LCH. Classical uncertainty or risk about future realizations of stochastic variables (but not about distributions of stochastic variables, which are assumed to be known and stationary) leads to precautionary saving by risk-averse consumers (Skinner, 1988; Zeldes, 1989). When risk-averse consumers face additional Knightian uncertainty (i.e., distributions of stochastic variables are unknown), precautionary saving is raised further (Miao, 2004; Hansen and Sargent, 2010).

Other theories substantially modify several key assumptions of the PIH-REH and LCH to derive behavioral predictions that are more consistent with the data. Borrowing constraints—the fact that interest rates on loans cannot be expected to rise to clear financial markets because they raise default risks (Stiglitz and Weiss, 1981) or because human capital cannot be used as collateral (Hayashi, 1982)—push consumers toward corner solutions and make borrowers’ consumption levels more sensitive to credit volumes and current income than to interest rates and wealth. When precautionary saving and borrowing constraints are combined, forward-looking, risk-averse consumers incur in buffer-stock saving, anticipating tighter future borrowing constraints (Schechtman, 1976).

According to the “capitalist spirit” model, which traces back to Smith and Marx, both consumption and wealth are valued by consumers (Cole et al., 1992; Fershtman and Weiss, 1993). If consumption and wealth are gross substitutes in utility, higher wealth does not raise consumption; instead, it is largely saved, contradicting the PIH-REH and the LCH.

Another dimension of consumer heterogeneity reflects differences in income and wealth across different population groups. The incidence of absolute poverty affects aggregate consumption because the poor save little. Then, utility is a positive function of the difference between current consumption and a subsistence consumption level (Christiano, 1989).5 Therefore, the saving rate declines with absolute poverty (given income distribution) and rises with the level of income—a refined version of autonomous consumption in a conventional KH model.

Post-Keynesian models stress the positive effect of functional income inequality on aggregate saving based on the observation that workers save less than capitalists (Lewis, 1954; Kaldor, 1957). More recent models focus on various channels from personal income inequality to saving, which, taken together, suggest that the effect of income distribution on saving is ambiguous.6

We end this brief survey of consumption theories by referring to the integration of household and corporate saving behavior. If a set of strict (and empirically implausible) assumptions are met, household owners of corporations are indifferent between saving as households or through their corporations. They are then able to “pierce the corporate veil,” offsetting one-to-one higher corporate saving by lower household saving. This hypothesis is the household-corporate saving analogue to the REH for government-private saving decisions.

Table 1 summarizes categories of saving determinants, specific variables in each category, expected signs of their saving effects according to consumption theories, and the empirical counterparts in country-panel studies based on aggregate saving data. It is important to note that each potential saving determinant is listed only once in Table 1, under the variable category to which it is most closely related by theory. However, both the expected sign and the signs reported in the empirical literature reflect the combined effects on saving predicted by different theoretical hypotheses. In fact, the latter relations could have opposite signs. For example, higher wealth leads to higher consumption according to the PIH but lowers consumption if wealth and consumption are substitutes in utility; hence, the expected effect of wealth on saving is ambiguous.

Table 1.

Determinants of Private Saving in Previous Studies

article image
Notes: The qualitative results listed in the last column of this table summarize signs of saving regressors reported in 16 panel studies of private saving. Positive and negative signs correspond to statistically significant coefficient estimates, while 0 denotes coefficient estimates that are not significantly different from zero. The sources are the corresponding tables and specific columns, rows, or regressions of the following studies: 1. Corbo and Schmidt-Hebbel (1991) (table 4); 2. Masson, Bayoumi, and Samiei (1995) (table 2, “restricted model” column); 3. Edwards (1996) (table 2, column 5); 4. Dayal-Ghulati and Thimann (1997) (table 4, column 2); 5. Bailliu and Reisen (1998) (table 1, columns 3 and 4); 6. Haque, Pesaran, and Sharma (1999) (table 6, columns 4 and 5); 7. Loayza, Schmidt-Hebbel, and Servén (2000) (table 4, column 6; table 7); 8. López, Schmidt-Hebbel, and Servén (2000) (tables 4 to 6); 9. Schmidt-Hebbel and Servén (2000) (table 6, columns 7 and 8); 10. De Serres and Pelgrin (2003) (table 2); 11. IMF (2005) (table 2.2, column 1); 12. Hondroyiannis (2006) (table 5, last row); 13. Gutiérrez (2007) (table 5, regression 9); 14. Horioka and Terada-Hagiwara (2012) (table 1, models 7 to 9); and 15. Bebczuk and Cavallo (table 3.1, columns 2 and 4). Significant coefficient signs are identified by a plus or a minus. Results identified by a zero mean either an insignificant coefficient in the corresponding column of the original study or, when the variable is omitted from the particular specification reported in the column, a significant or insignificant variable in a different column of the same table. When denoted by two signs separated by “or”, it denotes that the coresponding signs are reported in different columns. Real rates of return are measured either on deposits or loans. Each study is identified in the table by the corresponding number in parentheses.

We now present the variable categories and discuss expected signs of the specific variables on private saving. When discussing the expected signs of the saving determinants one by one, we hold constant the influence of other variables. For example, when analyzing the effects of inflation, we take the level of current income as given.

Income. PIH and LCH predict that current income should be largely saved. Consumption habits reinforce the saving effect predicted by PIH for higher current income. However, when fundamental assumptions of the PIH and the LCH are not satisfied or when consumption habits are weak, current income may raise consumption (in the extreme, one to one) when (i) it accrues to borrowing-constrained consumers, (ii) it signals higher future income, or (iii) it accrues to poor consumers that consume close to their subsistence income level. In these cases, marginal consumption of current income is high and marginal saving is low.7

Income reflects (unobserved) temporary and permanent income components. This distinction makes the prediction of the PIH and the LCH sharper (as long as the separation of current income into estimates for temporary and permanent components is statistically well-grounded), being that permanent income should be consumed while temporary income should be saved. Again, however, deviations from the latter prediction are observed when the assumptions of the PIH are not satisfied, as in the three cases described above.

Higher income growth could lead to an upward estimation of wealth, which reduces saving under the PIH. Under the LCH, the general effect of growth is ambiguous. Under the wealth-in-utility theory, higher wealth leads to less consumption and more saving. Therefore, growth has an ambiguous effect on saving.

Wealth. Consumer wealth comprises net financial assets, real assets (housing, consumer durables), and human wealth (the discounted present value of expected future labor income). If REH holds and if households pierce the corporate veil, consumer wealth is indistinguishable from national wealth after full consolidation of household, corporate, and government assets and liabilities. As discussed above, the saving effect of wealth and its components is ambiguous.

Rates of return on financial assets. According to the PIH and the LCH, a rise in the rate of return on financial assets held by consumers entails income, substitution, and human-wealth effects. If the consumer is a net creditor (a net holder of financial assets), the substitution and human-wealth effects of a higher rate of return on saving are positive, while the income effect is negative, hence the combined net effect is ambiguous. If the consumer is a net debtor, the income effect turns positive. Thus, higher bank deposit rates are likely to have an ambiguous effect on saving while higher lending rates are likely to reduce lending and raise saving. The overall effect of higher interest rates on aggregate private saving is ambiguous.

Relative prices. Relative prices of consumption and major consumption components affect saving because they entail intertemporal and intratemporal substitution effects, as well as income effects. Higher current consumer price inflation raises current prices of consumer goods relative to past prices, leading to higher saving. At the same time, current inflation signals macroeconomic instability, raising precautionary saving. Also, higher expected future inflation lowers the ex-ante real interest rate, inducing intertemporal substitution, income, and human-wealth effects that, on balance, imply an overall ambiguous (positive) effect on saving by savers with positive (negative) net financial asset positions.

Improved terms of trade entail a direct increase in net income from abroad, benefiting consumers by a proportional positive effect on income. Regarding the composition of consumption, imported goods typically represent a much larger share of consumption than exportable goods. Therefore, higher terms of trade are likely to reduce the average consumption deflator. In sum, higher current terms of trade are likely to affect saving positively. Like in the case of income, when distinguishing between estimated temporary and permanent components of the terms of trade, the former is expected to be largely saved and the latter to be consumed.

For consumption decisions, the real exchange rate is a relative price between different categories of consumption spending: domestic to imported goods or non-traded to traded goods. A change in the current relative price level of different consumption categories has an ambiguous effect on the consumption deflator (depending on the consumption basket) and therefore on saving. Similarly, an expected future appreciation of the real exchange rate has an ambiguous effect on saving.8

Risk and uncertainty. The precautionary saving theory predicts that higher levels of classical and Knightian uncertainty lead to higher precautionary saving. Therefore, higher financial risk measured by larger second moments of asset returns or larger market volatility indicators should lead to higher saving. However, when market volatility is extreme or financial, macroeconomic, and political forms of instability turn into crises, agents lose confidence in financial instruments and the institutions that issue or back them such that saving declines. Thus, the effect of risk and uncertainty on saving is ultimately ambiguous.9

Borrowing constraints. Tighter current borrowing constraints imply less access by consumers to credit, and therefore increase saving. This effect is magnified by anticipation of tighter future constraints by risk-averse consumers, giving rise to buffer-stock savings. Proxies of domestic borrowing constraints include money and credit flows in addition to current income.

A proxy of foreign borrowing constraints is foreign saving or the current account deficit, which is a valid saving determinant when the country faces a binding quantitative restriction in its access to foreign funding. In the absence of such a quantitative constraint or when a price variable such as the sovereign debt premium reflects the cost of external borrowing, the premium and the current account jointly respond to domestic saving and investment decisions. The sovereign debt premium is an important component of the cost of foreign funding and therefore affects saving like any lending interest rate that affects a debtor, i.e., positively.

Government restrictions imposed on international capital flows are likely to affect private saving. Restrictions on outflows limit capital outflows and restrictions on inflows limit issuance of foreign liabilities, hence both are likely to raise saving.

Financial depth. Development of deep and well-regulated financial and capital markets lead to a diversified supply of saving instruments that are similar to those offered in international markets. This could intensify home bias in domestic savers’ allocation of worldwide saving and, possibly, raise private saving flows. Proxies of financial depth, including bank credit stocks, financial assets, or broad money holdings could be positively associated to higher saving. However, the latter are also important components of consumer wealth. Therefore, the overall impact of the latter proxies of financial depth on saving is ambiguous.

Demographics. This variable category, as well as many subsequent categories discussed below, reflects potential aggregate saving effects that stem from differences in saving behavior across different population groups. Regarding demographic heterogeneity, the LCH predicts a hump-shaped pattern of saving along the life cycle. Standard proxies of a country’s demographic structure are the young- and old-age dependency ratios (the ratios of young and old people, respectively, to the active-age cohorts). The larger these groups, the smaller aggregate saving should be.

A country’s rate of urbanization could affect aggregate saving through several channels. First, a conventional hypothesis holds that the “city lights” reflected in larger consumption opportunities reduce saving of city dwellers compared to the rural population. Second, farmers are likely to face larger income uncertainty and less insurance and credit opportunities than urban dwellers, leading to higher saving in rural areas. Finally, farmers tend to be poorer than city dwellers, leading to lower rural saving. Hence, urbanization affects saving ambiguously.10

Poverty and distribution. As average consumption out of income declines with the distance between income and a subsistence level of consumption, saving declines with a larger share of people falling below absolute poverty (for given income inequality). The effect of the relative distribution of personal income or wealth on saving is ambiguous.

Fiscal policy. The REH predicts an offset of private saving to a change in public sector saving. Full offsetting is empirically unlikely, but it is expected that a higher government balance (or higher government saving, given government investment) lowers private saving. Government consumption has an ambiguous effect on saving, depending if public and private consumption are substitutes or complements in consumer utility (López et al., 2000).

Government spending components. Government spending on education, health, and other in-kind transfers reduces private consumption when the former spending categories are substitutes of similar private consumption categories, hence private saving rises. However, government transfers to consumers paid in cash raise disposable income, and have an ambiguous effect on private saving rates. Finally, government social spending and transfers lower uncertainty faced by consumers, reducing the need for precautionary saving. Thus, the overall effect of the latter on private saving is ambiguous.

Pension system. Pension benefits paid by a pay-as-you-go system raise pensioners’ consumption either fully or less than fully, having an ambiguous effect on private saving. Mandatory contributions to a fully-funded pension system reduce voluntary saving of contributors, but usually not one to one—hence, overall private saving is either maintained or increased.11 Fully-funded pension system assets held by individuals have an ambiguous effect on saving, like wealth and any of its components as discussed above.

Households and firms. When households pierce at least in part the corporate veil, higher corporate saving is partially offset by lower household saving.

The last Column of Table 1 lists the sign results of the estimated coefficients of saving determinants reported in 15 empirical studies on mostly private saving rates based on panel samples.12 Sample size and coverage, data sources, saving rate definitions, model specifications, and estimation methodologies vary significantly across studies.13 For example, sample sizes range between a low of 66 country-year observations (for 12 countries) in Horioka and Terada-Hagiwara (2012) and a high of 872 country-year observations (for 69 countries) in Loayza et al. (2000).

We conclude the following points from the heterogeneous empirical literature. First, most individual studies include few potential saving determinants in the specification—on average, six regressors out of 49 potential saving determinants listed in Table 1. This disappointingly small number reflects (i) the progress in consumption theory that has added over time new potential regressors for which empirical measures are only gradually made available; (ii) the lack of data, as many studies run out of degrees of freedom when adding more regressors; and (iii) many potential regressors listed in Table 1 are close substitutes of others. However, about half of the 49 potential regressors have not been used in any single previous private saving study.

Second, signs of several reported coefficients tend to be consistent with theory, either when they are expected to be unambiguously of one particular sign or when their expected sign is ambiguous. For example, 11 studies report statistically significant private saving offset coefficients for public saving or the public sector balance. A different example is the real interest rate, whose expected sign is ambiguous—an ambiguity reflected by a wide range of significant signs in 11 studies, from negative to zero and to positive.

Third, signs reported for other coefficients by several previous studies either contradict theory or results of other studies; e.g., the coefficients for inflation, credit flows, and old-age dependency. Fourth, the dispersion of parameter point estimates (and their confidence intervals) is very large—including those that are consistent with theory. Fifth, a core set of potential saving determinants is included in most studies. These are income level and income growth, real interest rate, inflation, terms of trade, demographic variables, and public saving. Few studies include non-standard variables like temporary/permanent components of income flows, income distribution, public consumption, and pension-system variables. Finally, variables for which empirical measures are not readily available or theory has been developed more recently are fully absent, including financial assets, violent conflicts, and government spending components.

III. Data and Stylized Facts

A. Sources and Construction

The world dataset constructed for this study is, to our knowledge, the most comprehensive on saving aggregates and their determinants. It contains a maximum of 4,137 observations, spanning from 1981 to 2012 and covering 165 countries. The panel dataset is unbalanced, with the number of time observations varying across countries.

The restrictions that shape our dataset come from several steps needed to improve its quality. We compile data from the IMF World Economic Outlook Database, the World Bank Worldwide Development Indicators, the UN National Accounts database, the OECD database, Haver Analytics, and several central bank web pages. The initial database then undergoes extensive cleansing to eliminate or replace faulty data, splice series, and fill gaps. For a complete list of the series compiled, the variables calculated, and the methods used for the construction of the database, see Appendix I.

B. Stylized Facts

We define the national saving rate as the ratio of gross national saving (GNS) to gross national disposable income (GNDI). Similarly, the private saving rate is defined as the gross private saving (GPS) scaled by gross private disposable income (GPDI).14 Tables 2 and 3 present descriptive statistics and pairwise correlations for the saving ratios and core saving determinants. In both tables, the sample is the one of the private saving baseline specification, which includes 3,254 observations for 153 countries over 31 years.

Table 2.

Descriptive Statistics

(Panel sample)

article image
Source: Authors’ calculations.
Table 3.

Correlation Matrix of Core Private Saving Determinants

(Panel sample in upper triangle, cross section in lower triangle)

article image
Source: Authors’ calculations.

Figure 1 presents the trends in world saving rates over the sample period 1981-2012.15 As shown in panel (a), the average national saving rate remained moderately stable around 19 percent of GNDI until the late 1990s. Since then, it climbed to 22.6 percent of GNDI in 2006, but then progressively fell to 19.5 percent in 2012. Panel (a) of Figure 1 also describes how the national saving composition changed over time. The average private saving component largely dominates the public counterpart, at about four-fifths of national saving. However, while the private saving rate remained virtually constant at 16.3 percent of GNDI, changes in public saving largely drove the changes observed in national saving after 2000. In particular, after fluctuating between 2 and 3 percent until the end of the 1990s, the public saving rate peaked at 6 percent in 2006 and remained high until the recent Great Recession, when it declined significantly.

Figure 1.
Figure 1.

World Saving Rates, 1981–2012

Citation: IMF Working Papers 2014, 204; 10.5089/9781475545517.001.A001

Panel (b) of Figure 1 shows the dispersion of private saving during the sample period. While the sample median private saving rate is remarkably stable at 20.5 percent of GPDI, the bands calculated for the percentile distribution of the private saving rate are wide, reflecting a large variability across countries. For example, about one half of the countries show, on average, private saving rates between 15.5 percent of GPDI and 28.5 percent of GPDI, and one-fifth shows private saving rates of more than 35 percent and less than 5 percent.

Figure 2 takes a closer look at private saving rates. Panel (a) depicts the average private saving rate across different country groups.16 Advanced economies had, on average, private saving rate of 27.2 percent of GPDI, about 7 percentage points higher than the sample average. Oil exporters experienced even higher private saving rates than advanced economies at times, but these countries are prone to a much higher volatility, associated to variations in oil prices. Over the sample period, average private saving rates for oil exporters fluctuated between 15.0 and 37.2 percent of GPDI. On the contrary, high-growth Asian economies show a steady upward trend since the 1980s. By the end of 2012, their average private saving rate stood at 34.7 percent of GPDI. Finally, the average private saving rate in low-income developing countries (LIDCs) is only 12.0 percent of GPDI over the sample period, at about 8 percentage points below the sample average.

Figure 2.
Figure 2.

World Private Saving Rates, 1981–2012

Citation: IMF Working Papers 2014, 204; 10.5089/9781475545517.001.A001

Panel (b) of Figure 2 presents the private saving composition. Available data is limited to 48 countries and 674 observations for household and corporate saving.17 The average private saving rate in this country subset of countries is 16.3 percent of GPDI, almost 4 percentage points below the average private saving rate for the whole sample, and declines slightly over time. The average household saving rate followed a downward trend since the mid-1990s, which was almost fully offset by an increasing corporate saving rate.

Finally, we show in Figure 3 scatter plots for pairwise panel correlations of cross-country averages between the private saving rate and its core determinants, and one scatter plot for correlations between household and corporate saving rates. The relationships reflect simple associations and certainly could be very different from partial correlations estimated in a multivariate regression. While some correlations are consistent with signs determined by theory (e.g., income level, income growth, and household and corporate saving), others do no show a clear pattern (e.g., terms of trade, public saving rate, real deposit rate), and others are inconsistent with theory (e.g., flow of private credit, old-age dependency ratio, and share of urban population).

Figure 3.
Figure 3.

Pairwise Panel Correlations

Citation: IMF Working Papers 2014, 204; 10.5089/9781475545517.001.A001

Notes: Depicted regression lines and 95 percent confidence intervals are based on linear regressions with a constant term using the unbalanced panel of 153 countries. Light-grey dots represent LIDCs, dark-grey dots represent advanced economies, and black dots represent the rest of the countries in the sample.Source: Authors’ calculations.

IV. Empirical Strategy

Let Yi,t denote the private saving rate. It can be modeled as:

Y i , t = b X i , t + d Z i , t + ɛ i t ( 1 )

where Xi,t includes the endogenous (and predetermined) covariates for country i at time t, Zi,t includes (strictly) exogenous variables and an intercept, b and d are the relative coefficients, and εit is a mean zero error term that captures unobserved heterogeneity.

The selection of variables Xi,t and Zi,t to be included in the baseline specification relies on consumption theory, previous empirical research (in particular, Loayza et al., 2000), as well as data availability. At a later stage, this set of regressors is complemented by other variables that are included in the specification to study their relationship with the dependent variable or to justify their exclusion from the baseline specification. In line with Loayza et al. (2000), we treat the log of real per capita GPDI in PPP terms, real growth rate of per capita GPDI in PPP terms, public saving in percent of GPDI, inflation, the real deposit rate, and the flow of private sector credit in percent of GPDI as endogenous variables, assuming that they are correlated with present, past or future error terms.18 On the other hand, we treat the log of the terms of trade, the old-age dependency ratio, the share of urban population, and the log of the real oil price as exogenous variables.

We estimate the static model (1) using ordinary least squares (OLS) applied to both a cross-section sample of country averages and a pooled panel sample of annual observations, correcting standard errors for heteroskedasticity and autocorrelation. Comparing OLS cross-section and pooled results is informative with respect to the between and within variation in the data, and one could interpret the results as long- and short-term coefficients, however OLS estimations suffer from potentially severe econometric problems: lack of dynamics, omitted variable bias due to absent country- and time-fixed effects, and endogeneity of the Xi,t variables.

Dynamics of the dependent variable are likely to be an important factor in the estimation because changes in private saving generally occur over a long period of time. More specifically, the observed private saving rate in a given year and for a given country yi,t may deviate from its target level Yi,t due to, for example, adjustment costs, consumption habits, consumption smoothing, or the lagged effects of the explanatory variables on private saving. Thus we specify a target adjustment model:

y i , t y i , t 1 = ( 1 γ ) ( Y i , t y i , t 1 ) ( 2 )

where γ is the adjustment speed. This means that if γ = 0, then yi,t = Yi,t, and the adjustment toward the target value takes place immediately. Combining equations (1) and (2), we specify the following dynamic model for the observables:

y i , t = γ y i , t 1 + β X i , t + δ Z i , t + u i t ( 3 )

where coefficients and the error term are defined as: (β, δ, ui,t) = (1 − γ) (b, d, εit).

Therefore incomplete adjustment (γ ≠ 0) leads to a form of state dependence where last period’s yi,t−1 determines this period’s yi,t. When this is not accounted for, correlation between yi,t and yi,t−1 could result from unobserved heterogeneity. Hence, a more general version of model (3) would include the time-invariant unobserved country-specific heterogeneity term ci, leading to the following specification:

y i , t = γ y i , t 1 + β X i , t + δ Z i , t + c i + u i t ( 4 )

This dynamic panel model also partially controls for possible reverse causality. For example, if past private saving performance yi,t−1 affects current income growth Xi,t, then this feedback is accounted for in model (4). Thus, we first estimate model (4) with the OLS estimator. To address endogeneity more comprehensively, we then estimate model (4) with the two-stage least squares (2SLS) estimator, which uses the lags of endogenous variables Xi,t as instruments.

Yet the dynamic structure of the model would make OLS estimates biased downward and inconsistent even with a fixed-effects estimator because the error term uit is correlated with the lagged dependent variable yi,t−1 (Nickel, 1981). An alternative approach is to first-difference the model in equation (4) to eliminate the fixed effects ci. However, the OLS estimator would still be inconsistent because this transformation does not affect the correlation between yi,t−1 and uit.

Under appropriate identification assumptions (or moment conditions), the difference Generalized Method of Moments (GMM) estimator would be consistent (Arellano and Bond, 1991). In particular, this estimator assumes that the idiosyncratic error uit is serially uncorrelated and that past values of the endogenous variables yi,ts are not correlated with the current error uit. These conditions allow using the second (and higher) lags of yi,t as instruments for yi,t−1, and second (and higher) lags of Xi,t as instruments for Xi,t.

Blundell et al. (2000) and Bond et al. (2001) show that the difference GMM estimator has poor finite sample properties and that the estimator performs weakly when the dependent variable is persistent. Arellano and Bond (1997) and Blundell and Bond (1998) propose the system GMM (S-GMM) estimator, which increases efficiency by estimating a system of two simultaneous equations, one in levels (with lagged first differences as instruments) and the other in first differences (with lagged levels as instruments). This estimator requires the additional identifying assumption that the instruments are exogenous to the fixed effects. Thus, we estimate model (4) with the asymptotically more efficient two-step S-GMM. The two-step variant presents estimates of the standard errors that tend to be severely downward biased (see Arellano and Bond, 1991, and Blundell and Bond, 1998). However, we implement the finite-sample correction of the two-step covariance matrix derived by Windmeijer (2005), which produces unbiased standard errors.

Finally, we estimate the following more comprehensive model that includes time-fixed effects τt (but excludes the real oil price from the set of strictly exogenous variables Xi,t) with the two-step S-GMM estimator:

y i , t = γ y i , t 1 + β X i , t + δ Z i , t + c i + τ t + u i t ( 5 )

While our preferred specification uses annual data with time-fixed effects, we also test the robustness of the results averaging all variables over five years. The latter specification has the advantage of abstracting from the business cycle and reducing the impact of measurement error, but at the cost of distorting and losing information about temporal dynamics of saving rates.

With the aim of testing for differences in saving behavior in specific time periods or selected country groups compared to the rest of the whole sample, we extend our preferred specification with interaction terms between our Xi,t and Zi,t variables, and a dummy variable Di,t, which takes a value of one for the specific time period or country group. More formally, we estimate the following nested model:

y i , t = γ y i , t 1 + β X i , t + δ Z i , t + ξ D i , t y i , t 1 + φ D i , t X i , t + ω D i , t Z i , t + c i + τ t + u i t ( 6 )

where ξ, φ, and ω are the coefficients of the interaction terms. The dummy variable Di,t is not included as a separate regressor because it would be perfectly collinear with time-fixed effects τt (in the case of time periods) or the country-fixed effects ci (in the case of country groups). The effect of the corresponding regressor Xi,t belonging to a specific time period or country group Di,t, on the dependent variable yi,t, is given by β + φ. Analogously, the effect of Zi,t (yi,t−1) belonging to the same country group or time period on the dependent variable yi,t is given by δ + ω (γ + ξ). We refrain, however, from analyzing other possible interactions or non-linearities as this is beyond the scope of this paper.

The S-GMM identification assumptions are tested applying a second-order serial correlation test for the residuals and the Hansen J-test for overidentifying restrictions. While the latter test is limited in that it hinges on the untestable assumption that at least one instrument is valid, it is still useful in spotting violations of validity.

V. Results

We present the estimation results in the following order. We start by reporting the regression results of a baseline specification for private saving, obtained by using different estimators (Table 4). Then, we extend the baseline specification by including additional regressors (Table 5). Subsequently, we analyze differential saving behavior in a particular time period and in selected country groups (Table 6). Then, we present empirical results for other saving aggregates, namely the national saving rate (Table 7), and household and corporate saving rates (Table 8). Finally, like our analysis of the private saving rate, we extend the baseline specification for the household saving rate by including additional regressors (Table 9).19

Table 4.

Determinants of Private Saving, Different Estimators

(Dependent variable: Private saving/GPDI)

article image
Notes: Standard errors in parentheses are corrected for heteroskedasticity and autocorrelation of the error term. System GMM estimations in columns (5) to (7) use a collapsed instrument matrix and perform the Windmeijer (2005) correction of the covariance matrix. Column (7) uses the sample 1983-2012 to have 6 periods of 5 years. The null hypothesis for the Hansen J−test is that the full set of instruments is valid. All estimations include a constant term. ***, **, * next to a number indicate statistical significance at 1, 5 and 10 Source: Authors’ calculations.
Table 5.

Determinants of Private Saving, Additional Explanatory Variables

(Two-step S-GMM; dependent variable: private saving/GPDI)

article image
Notes: Standard errors in parentheses are corrected for heteroskedasticity and autocorrelation of the error term. System GMM estimations use a collapsed instrument matrix and perform the Windmeijer (2005) correction of the covariance matrix. The null hypothesis for the Hansen J-test is that the full set of instruments is valid. All estimations include a constant term. ***, **, * next to a number indicate statistical significance at 1, 5 and 10 percent, respectively. Source: Authors’ calculations.
Table 6.

Determinants of Private Saving, Interactions

(Two-step S-GMM; dependent variable: private saving/GPDI)

article image
Notes: Standard errors in parentheses are corrected for heteroskedasticity and autocorrelation of the error term. System GMM estimations use a collapsed instrument matrix and perform the Windmeijer (2005) correction of the covariance matrix. The null hypothesis for the Hansen J-test is that the full set of instruments is valid. All estimations include a constant term. ***, **, * next to a number indicate statistical significance at 1, 5 and 10 percent, respectively. Source: Authors’ calculations.
Table 7.

Determinants of National Saving, Alternative Estimators

(Dependent variable: national saving/GNDI)

article image
Notes: Standard errors in parentheses are corrected for heteroskedasticity and autocorrelation of the error term. System GMM estimations in columns (5) to (7) use a collapsed instrument matrix and perform the Windmeijer (2005) correction of the covariance matrix. Column (7) uses the sample 1983-2012 to have 6 periods of 5 years. The null hypothesis for the Hansen J-test is that the full set of instruments is valid. All estimations include a constant term. ***, **, * next to a number indicate statistical significance at 1, 5 and 10 percent, respectively. Source: Authors’ calculations.
Table 8.

Determinants of Household and Corporate Saving, Baseline Specification

(Two-step S- GMM; dependent variable: household saving/GPDI, corporate saving/GPDI)

article image
Notes: Standard errors in parentheses are corrected for heteroskedasticity and autocorrelation of the error term. System GMM estimations use a collapsed instrument matrix and perform the Windmeijer (2005) correction of the covariance matrix. The null hypothesis for the Hansen J- test is that the full set of instruments is valid. All estimations include a constant term. ***, **, * next to a number indicate statistical significance at 1, 5 and 10 percent, respectively. Source: Authors’ calculations.
Table 9.

Determinants of Household and Corporate Saving, Additional Explanatory Variables

(Two-step S-GMM; dependent variable: household saving/GPDI, and corporate saving/GPDI)

article image
Notes: Standard errors in parentheses are corrected for heteroskedasticity and autocorrelation of the error term. System GMM estimations use a collapsed instrument matrix and perform the Windmeijer (2005) correction of the covariance matrix. The null hypothesis for the Hansen J-test is that the full set of instruments is valid. All estimations include a constant term. ***, **, * next to a number indicate statistical significance at 1, 5 and 10 percent, respectively. Source: Authors’ calculations.