The importance of social security revenue and expenditure as instruments of fiscal policy is on the rise in a number of developed and developing countries. Pressures are mounting to expand protection for old age and against current risks, especially through liberalized health-care and income-maintenance programs, often in the face of inadequate growth in revenue raised chiefly through payroll taxation. These pressures are brought about by the emergence of a more equity-oriented social philosophy, as well as by various demographic and economic trends, the main ones being the increase in the proportion of retired to employed persons and accelerated inflation. In many countries, the ensuing rise in the need for resources cannot be met with relatively inelastic social security revenue, given regressive payroll taxes and nearly full tax coverage of the working population.

To cope with these developments, governments must consider a number of policy options regarding changes in the structure and size of social security benefits, on the one hand, and in the sources of financing, on the other hand. In weighing the options, they have to take into account the potential repercussions on income distribution, stability, and the allocation of resources. With respect to the latter, a major question relates to the choice by households—in their dual capacity as beneficiaries and taxpayers—between savings and consumption in response to social security. But, despite the sizable literature on savings and consumption behavior, the answer cannot be provided through theoretical deduction alone. Rather, it is necessary to ascertain empirically whether social security schemes tend to encourage, discourage, or leave unaltered the savings propensity of households. In view of conflicting evidence reported in earlier research on the subject, this paper attempts to shed further light on the impact of social security on variations among countries in savings propensity of households, by introducing certain methodological refinements and using richer and more recent data.

To begin with, a model of long-run household savings consistent with the principal hypotheses of savings behavior is formulated. Then, appropriate social security variables are incorporated in the model. The theoretical sections are followed by a survey of past research on the cross-country relationship between social security and savings. In the remainder of the paper, the influence of social security benefits and taxes on household savings is estimated separately across industrial and developing countries for the period 1969-71. The concluding section presents a summary and discussion of some policy implications.

I. A Model of Household Savings

For any meaningful test of the influence of social security schemes on intercountry differences in savings propensities, it is necessary to adopt a widely accepted theoretical explanation of household savings activity. Clearly, the explanation must be consistent with the major hypotheses of long-run savings behavior, be able to capture international variations in such behavior, and be amenable to estimation on existing statistical information. Whereas adequacy under the last two criteria will become apparent in later sections, the affinity of this model with the accepted hypotheses is discussed next.

A model of household savings that appears to meet these characteristics can be stated in the structural form

where for any given country

SH	=	real household savings per capita
YH	=	real household disposable income per capita
G	=	growth rate of labor productivity
NRR	=	aftertax real rate of return on household savings
WH	=	real stock of nonhuman wealth per capita
RE	=	real flow of corporate retained earnings per capita
INQ	=	index of income inequality
LPA	=	labor-force participation of the aged
DA	=	dependency ratio for the aged
DM	=	dependency ratio for minors
LEA	=	life expectancy of the aged

View Table

SH	=	real household savings per capita
YH	=	real household disposable income per capita
G	=	growth rate of labor productivity
NRR	=	aftertax real rate of return on household savings
WH	=	real stock of nonhuman wealth per capita
RE	=	real flow of corporate retained earnings per capita
INQ	=	index of income inequality
LPA	=	labor-force participation of the aged
DA	=	dependency ratio for the aged
DM	=	dependency ratio for minors
LEA	=	life expectancy of the aged

Equation (1) depicts, in the absence of social security schemes, the average propensity of households to save (by country, over a number of years) as a function of economic, social, and demographic variables. It indicates that the savings rate may rise across countries above a constant value (1 > α₀ ≥ 0) in response to an increase in such economic determinants as per capita income (α₁ ≤ 0), the growth rate of productivity (α₂ ≥ 0), or the rate of interest (α₃ ≥ 0), and to a decline in wealth (α₄ ≤ 0) or in corporate retentions (α₅ ≤ 0). As for socioeconomic variables, an increase in the inequality of income or a contraction in the labor-force participation of the aged may lead to a higher savings rate (α₆ ≥ 0 and α₇ ≤ 0, respectively). Further, the savings propensity is likely to fall as a result of demographic forces in the form of an increase in the ratio of the aged or of minors to working-age population (α₈ ≤ 0 and α₉ ≤ 0) or of a decline in life expectancy at retirement age (α₁₀ ≥ 0). Accordingly, the model can accommodate any of three prototypes for which we express no preference a priori: the absolute-income hypothesis, the permanent-income hypothesis, and the life-cycle hypothesis.

The absolute-income hypothesis, derived from the Keynesian consumption function, states that savings is determined by the level of income. Moreover, household savings is negative at zero income (α₁ < 0), and the marginal propensity to save is always positive and larger than the average propensity to save (1 > α₀, > SH/YH) at any income level. The other determinants given in equation (1) are omitted from the original form (so that α₂, …, α₁₀ = 0), although the hypothesis does assign a minor role to the interest rate as well as to certain demographic variables.¹

According to the permanent-income hypothesis, developed by Friedman partly on the basis of the Fisherian theory of interest, long-run savings is a constant proportion of long-run income. This proportion is in turn determined by the interest rate, the ratio of nonhuman wealth to income, and various social and demographic factors.² Thus, in terms of equation (1), the savings ratio is affected negatively by the wealth ratio (α₄ < 0) and the dependency ratios (α₈ < 0, α₉ < 0), and positively by the interest rate (α₃ > 0)³ and the life expectancy of the aged (α₁₀ > 0). The influence of other characteristics of households may be reflected in the intercept, but the remaining variables shown in equation (1) are excluded from the hypothesis (α₁, α₂, α₅, α₆, α₇, = 0).

The life-cycle hypothesis, introduced and refined by a number of authors, also in the Fisherian tradition, provides perhaps the most versatile microeconomic explanation of savings behavior.⁴ In essence, it is based on the assumption that an individual accumulates wealth in his working years to compensate for the lack of earnings during retirement and thus to maintain a stable stream of consumption. Hence, under static conditions and barring intergenerational’ transfers, the net lifetime (long-run) sayings of the individual is zero; likewise, there is no net aggregate savings (α₀ = 0).⁵ However, in a growing economy, the long-run aggregate savings/income ratio becomes positive as each successive generation enjoys a higher income from which to save. Thus, in a dynamic context, the savings ratio is affected by the rate of growth of productivity, as well as by variations in the distribution of lifetime between work and retirement and in the distribution of the population between workers and dependents. In terms of equation (1), the average propensity to save increases owing to a rise in the growth rate of productivity (α₂ > 0) or the life expectancy of the aged (α₁₀ > 0), or owing to a fall in the stock of wealth (α₄ < 0), the labor supply of the aged (α₇ < 0), or dependency ratios (α₈ < 0, α₉ < 0).⁶

Other less frequently listed determinants of the household savings ratio, under the life-cycle hypothesis, are the interest rate, corporate savings, and income distribution.⁷ The positive effect of a rise in the interest rate (α₃ > 0) reflects the substitution of future for present consumption, dampened by any increase in corporate savings taken into account by shareholders in their personal savings decisions (α₅ < 0). A similar positive response to a rise in the unevenness of the distribution of income (α₆ > 0) is based on the trade-off between equity and efficiency, as relatively high savings propensities may tend to be associated with high levels of disposable income.⁸

In the life-cycle approach, there is no place for the absolute level of income as an explanatory variable (α₁ = 0). Yet in an extension of the hypothesis, it has been suggested that the savings ratio is influenced indirectly by income through the labor-force participation of the aged.⁹ According to this view, individuals will tend to lengthen their work life if their real income is relatively low, so as to generate sufficient purchasing power to complement leisure with an adequate level of consumption during the retirement years. Also, higher life expectancy at normal retirement age will prompt them to prolong participation in the work force, everything else being equal. Thus, labor-force participation of the aged is explained by the relationship

indicating the impact of income (β₁ > 0) and of life expectancy of the aged (β₂ > 0), above a constant rate (β₀ > 0), whereby the labor-force participation variable is to be treated endogenously in the savings equation (1). In sum, equations (1) and (5) comprise the analytical framework adopted to examine the link between social security and savings behavior.

II. The Role of Social Security

Individuals are associated with social security schemes as recipients of benefits, on the one hand, and as contributors, or rather as taxpayers, on the other hand. For purposes of this inquiry, namely, to assess the impact of these schemes on household savings, the benefits can be classified in three basic categories: old-age transfers, transfers related to current contingencies, and loans. Although normally the beneficiaries must be enrolled under the scheme, in some countries certain benefits are provided universally or upon meeting certain requirements.

Old-age transfers consist predominantly of pensions disbursed periodically to eligible retirees—in some countries qualifying automatically at a specified age, and in others subject to a retirement test, an income test, or a means test—in an amount more or less related to covered preretirement wages. Alternatively, in a few developing countries, a lump-sum payment is made to each retiree from a provident fund (normally operated on a funded rather than a pay-as-you-go basis), in an amount equivalent to past tax contributions into the fund plus accrued interest. Additional cash benefits are provided under old-age programs upon disability or death of eligible workers or retirees.

Other social security transfers to eligible households cover (fully or partially) medical contingencies, unemployment, work injury, and/or indigence. These benefits, in cash or in kind, are provided either to the entire population, or to enrolled workers and their dependents, or to individuals who qualify under a means test.

The third category of benefits, available in certain developing countries, involves the extension of loans to enrolled individuals—in some instances, limited on the basis of seniority or a means test. These loans are provided for housing or for other personal needs (such as expenditure on education, marriage, or certain durable goods).

Among the various forms of financing social security schemes, individuals are directly aware of payroll tax payments, which alone in many countries confer eligibility for benefits—fostering the popular belief that payroll taxes for social security are tantamount to insurance premiums or pension fund contributions. Other less common sources of financing, such as general revenue or revenue from earmarked taxes, are provided by taxpayers at large.

Social security schemes may influence the household savings ratio via three effects: the income effect, the wealth effect, and the retirement effect.¹⁰ In the framework of equations (1) and (5), any change in benefits or payroll taxes has an income effect on the savings ratio by altering the average level of disposable income and the degree of income inequality. This impact is contingent on the extent to which, over the long run, benefits raise disposable income while payroll taxes reduce it, from the level that would obtain otherwise, for households in different income brackets. In the event that parameters α₁, α₄, α₅, α₆, and β₁ turn out to be zero, the income effect vanishes. Also, insofar as benefits move closely with tax payments (as tax proceeds are used to finance benefits in unfunded or pay-as-you-go schemes) and as they have similar incidence on households across income categories, there is no income effect even in the presence of nonzero parameters.

The wealth effect indicates the direct savings response of individuals to expected future benefits. It is negative if individuals feel that they need to accumulate less wealth for retirement and current contingencies (such as sickness and unemployment) and thus reduce their desired ratio of savings to income, as a result of the protection afforded by social security. Alternatively, the effect is positive if the introduction of social security educates individuals as to the need to insure themselves for old age and against current risks, thereby raising their desired savings ratio.¹¹ The net wealth effect may be either negative or positive, depending on the relative strength of the individual’s perception of the substitutability between expected social security benefits and savings, and of his effort to complement such benefits with additional savings because of the educational impact.

In addition, an indirect repercussion of social security benefits on the savings ratio may take place through the retirement effect. The prospect of receiving an old-age pension or lump-sum transfer may encourage individuals to retire sooner than they would in its absence. Hence, the retirement effect on savings is expected to be positive, reflecting the increased savings ratio brought about by the benefit-induced reduction in the labor supply of the aged and by the consequent need to provide for the longer retirement period.

The wealth and retirement effects are incorporated into the savings model by modifying equations (1) and (5), respectively, as follows:

and

where the new variables are

SSP	=	social security pension per aged person
SSF	=	social security lump-sum transfer (from provident fund) per aged person
SSO	=	other social security transfers per capita
SSL	=	social security loans per capita
AGE	=	age of social security system

View Table

SSP	=	social security pension per aged person
SSF	=	social security lump-sum transfer (from provident fund) per aged person
SSO	=	other social security transfers per capita
SSL	=	social security loans per capita
AGE	=	age of social security system

Parameters α₁₁ through α₁₅ reflect the net wealth effect, and β₃ through β₅, multiplied by α₇, indicate the retirement effect. As discussed earlier, the net wealth effect of the rate of each major type of benefit can be positive or negative (α₁₁, α₁₂, α₁₃, α₁₄, ≷ 0), whereas the retirement effect of the old-age benefit rates cannot be negative (α₇β₃, α₇β₄ ≥ 0) on the savings ratio—where benefit rates are defined as the proportion of per capita benefit to per capita disposable income. But these effects are qualified by variations in the age of social security schemes across countries. The educational impact of social security on savings habits may register with a lag; analogously, individuals may delay the reduction of savings or the acceleration of retirement induced by increased benefits. In contrast, it can also be argued that the wealth and retirement effects may wear off as the scheme matures, if individuals no longer believe that they will receive an adequate old-age benefit stream upon retirement—especially in countries where the size of benefits is not corrected sufficiently for inflation or where the number of taxpayers supporting the scheme shrinks in relation to the number of beneficiaries. Thus, the possible direction of the relationship between the age of the system and either the savings ratio or the labor-force participation of the aged is ambiguous (α₁₅ ≷ 0, α₇β₅ ≷ 0).

Alternatively, savers may adjust their desired wealth/income ratio to the social security tax liability that they actually incur (that is, payroll tax payments of employers and employees) rather than to the social security benefits that they are entitled to receive in case of a contingency. To take into account this view of the wealth effect, equation (6) must be replaced by

where the effective payroll tax rate is calculated from the ratio of tax payments, SST, to gross wages and salaries, WS. Similarly, equation (7) is to be changed if individuals near retirement age are more apt to base their choice between work and leisure on the tax proxy for the expected stream of old-age benefits, so that

The parameters of the effective tax rate represent the wealth effect (α₁₁ ≷ 0)¹² and the retirement effect (α₇β₃ ≥ 0), qualified by the age of the system (α₁₂ ≷ 0, α₇β₄ ≷ 0).

A criticism that can be leveled at the foregoing analysis of wealth and retirement effects of social security is that it ignores the role of the extended family as a social security institution in certain countries. Gainfully employed individuals provide financial support to the retired, invalid, and sick members of the family in the expectation that they in turn will be similarly protected upon reaching old age or becoming invalid. This arrangement, which operates much like an unfunded social security scheme, may explain in part the relatively depressed savings ratio in a number of developing countries, implicitly casting doubt on the general applicability of the life-cycle hypothesis. However, notwithstanding the diminished importance of the negative wealth effect and the retirement effect in the presence of an extended family structure, the positive wealth effect reflecting the educational factor may still hold in these countries.

III. Review of Past Studies

There have been relatively few studies on the relationship between social security and savings behavior across countries—in contrast to the more extensive research based on U. S. time series.¹³ In the first attempt to analyze such a relationship, Aaron (1967) adopted a household savings function broadly in accord with the life-cycle hypothesis. Although he was concerned primarily with explaining social security expenditure (in terms of the age of the system and household savings, inter alia), he also explored the effect of the ratio of such expenditure to national income, the age of the system, the old-age benefit rate, and the growth rate of national income on the household savings ratio. Ordinary least-squares estimates on a cross section of 19 countries (most of them developed) for 1957 yielded a significant positive coefficient for the age of the system and a negative one for the ratio of social security outlays. But re-estimation by Pechman, Aaron, and Taussig (1968, Appendix D) on 1960 data for practically the same sample failed to confirm the significance of the inverse relationship between the social security expenditure ratio and the household savings ratio. In addition, they estimated, also on 1960 data, the labor-force participation of the aged as a function of the proportion of the population over retirement age, per capita national income, and the ratio of old-age benefits per aged person to per capita wages in manufacturing, all of which displayed the expected significant negative coefficients.

On the basis of the most comprehensive version of the life-cycle hypothesis, Feldstein (1977) tested the impact of old-age benefits and other characteristics of social security schemes with a two-equation model—which to a large extent has inspired our model. The two endogenous variables—the private savings ratio (including corporate savings) and the labor-force participation of the aged—are determined by the following exogenous variables: growth rate of real private national income, dependency ratios, real per capita national income, life expectancy of the aged, old-age social security benefit per aged person as a ratio of per capita income, age of the scheme, and the ratio of corporate retained earnings to national income. Alternatively, in an effort to identify more precisely the effects of old-age benefits, the benefit rate per aged person was replaced by the benefit rate per old-age pensioner, the coverage rate of aged persons, and the presence of a retirement test. Feldstein (1977) used averages from national income statistics for the years 1954-60 and social security data for 1958-60 for a sample of 12 developed and 3 developing countries to measure the variables. The model was estimated by applying single-equation as well as simultaneous-equation techniques: weighted ordinary least squares on the structural and reduced form equations, and weighted two-stage least squares—the weights consisting of the respective country’s population.

On the whole, Feldstein’s (1977) results gave support to the life-cycle hypothesis. The coefficients of the following exogenous variables of the savings function were significant (although indicating sensitivity to the inclusion or omission of particular variables) and carried the indicated signs: positive for the growth rate, the life expectancy of the aged, and the retirement test; and negative for the dependency ratios, various measures of the benefit rate, the per capita income level, and its inverse. As to the equation for the labor-force participation, varying degrees of significance were obtained with positive coefficients for per capita income and its inverse, and with negative coefficients for the components of the benefit ratio and, in some instances, the retirement test. The significance of the reciprocal of the income variable declines markedly (while that of other variables falls slightly) upon removal of the absolute level of income from either equation.¹⁴ The coefficients of the age of the social security system and the corporate retention ratio are insignificant in virtually all regressions.

In another recent study, Barro and MacDonald (1979) tried to explain the ratio of consumer expenditure to gross domestic product (GDP) with a pooled sample of annual observations for the period 1951-60 on 16 developed Western countries. They employed a one-equation model, also based essentially on the life-cycle hypothesis, in which the consumption ratio is determined by the growth rate in real per capita GDP, the reciprocal of real per capita GDP, the ratio of government expenditure to GDP, the proportion of old-age persons in total population, the old-age benefit rate, the unemployment rate, and the ratio of last year’s to the current year’s per capita GDP. Aside from the variables already discussed, the last two are supposed to take into account cyclical fluctuations, while the parameter of the government expenditure ratio is meant to capture the household’s perception of a proportion of government outlays as constituting an equivalent amount of imputed private consumption or saving.

Barro and MacDonald (1979) presented four sets of ordinary least-squares estimates. In one instance, the consumption function was weighted by the square root of population, in another it was not so weighted; further, each of these regressions was run alternatively with a constant term or with country dummy variables. In the equations with the intercept, the coefficients of all variables (except the ratio of past to current GDP) were at least marginally significant: carrying a negative sign for the government expenditure ratio, the growth rate, and the old-age benefit rate; and a positive sign for the proportion of elderly in the population, the unemployment rate, and the reciprocal of per capita GDP. Not surprisingly, the overall performance of the equation improved sharply upon introduction of country dummies (and exclusion of the growth rate), retaining the significance of the aforementioned variables although with reversal to a negative sign for the proportion of elderly and to a positive sign for the old-age benefit rate. The presence of dummies also contributed to a reduction in the serial correlation of residuals, which seems to be particularly severe otherwise (as revealed by the low value of the Durbin-Watson statistic). As for differences between weighted and unweighted regressions, the latter provided worse fits than the former, as expected.

Before interpreting the results, one should consider several problems associated with the measurement of variables and the estimation of the foregoing models. Measurement difficulties are likely to arise with respect to savings, per capita real income, and social security variables. The savings variable should be limited to the household sector,¹⁵ without including the corporate sector, so as to test rather than assume the transparence of the corporate veil in the personal choice between savings and consumption, and to acknowledge the possible distinction in the motives to save of households and corporations. Instead of translating per capita real income into U. S. dollars at the official exchange rate, as has been done in all the preceding models, the correct approach would be to adjust real income for differences in the purchasing power of currencies. Although one may criticize Aaron (1967) for using a crude social security variable, namely, the global social security expenditure ratio, it may also be questionable to break down, as Feldstein does, the benefit rate on the basis of rather soft data on the coverage ratio and the retirement test (obtained from statutory information that vaguely reflects actual practices), while ignoring the income and means tests used in some countries.

The estimation problems involve identification, specification, heteroscedasticity, distinction between long-run and short-run effects, and sample homogeneity. The first three are econometric issues, whereas the last two may affect only the interpretation of results. Estimation of most savings functions, including the ones discussed here, ignores the possible bias introduced by the difficulty of identifying such a function as distinct from the investment function, which also contains the growth rate and the interest rate as explanatory variables. This bias can be corrected through simultaneous equation estimation, but outside evidence suggests that the extent of the bias is not likely to be important.¹⁶

Another source of bias present in these models is the possible specification error brought about by the omission of certain key variables, notably, the interest rate, income distribution, the stock of wealth, and social security variables other than the old-age benefit rate. This error can result also from the linear form of savings function, given the generality of the underlying theory. Incidentally, the severe serial correlation reported by Barro and MacDonald (1979) may be symptomatic of the exclusion of certain exogenous variables or the choice of an incorrect functional form.

The weighting procedure employed by some authors to enhance the relative importance of larger countries in the regressions, by population size, may lead to heteroscedastic disturbances that (much like serial correlation) tend to inflate the significance of the estimated coefficients upon application of conventional tests of statistical significance.

Several studies fail to isolate long-run from short-run effects of social security. The Aaron (1967) and Pechman-Aaron-Taussig (1968) estimates, based on single-year cross sections, are undoubtedly influenced by short-run forces. Despite the explicit inclusion of certain cyclical variables, the Barro-MacDonald (1979) estimates do reflect short-run behavior, particularly in the presence of dummies that serve to eliminate intercountry variations (or shifts) among ten-year strings of country observations.

Most regressions were run on cross-sectional samples containing countries at different stages of development (although the majority are developed), under the assumption that they are drawn from the same population. A more realistic approach would have been to recognize that developed and developing countries may belong to different populations, by allowing differences in behavior to be reflected in differences in the values of corresponding parameter estimates.

Having discussed possible empirical shortcomings of previous research, one can summarize and interpret the findings on the impact of social security. In their estimates of the labor-force participation function, Pechman, Aaron, and Taussig (1968), as well as Feldstein (1977), found support for the postulated retirement effect: an increase in the old-age benefit rate may lead to a rise in household or private savings, via reduction in the labor supply of the aged. However, Aaron (1967) and Feldstein (1977) also produced evidence indicating that on balance the net negative wealth effect swamps the retirement effect, so that social security has a depressing impact on the propensity to save. Further, according to Aaron’s (1967) estimates, as the system matures, the negative influence of social security expenditure wears off. A plausible explanation is that individuals raise the savings ratio as they increasingly feel that the expected benefits will not be sufficient to provide protection against various contingencies, including old age.

Against these results, Barro and MacDonald (1979) found that increased old-age benefit rates are related to a rising savings propensity, suggesting that the retirement effect and/or the positive wealth effect more than offset the negative wealth effect. In the absence of a direct estimate of the labor-force participation equation, it is not clear whether indeed the retirement effect or a positive wealth effect, representing the educational role of social security, is the dominant influence. In sum, the conflicting findings of past studies, as well as the underlying empirical problems, warrant a fresh attempt to solve some of these problems and thus to provide more reliable evidence on the influence, if any, of social security on intercountry differences in savings propensities.

IV. Empirical Estimates of the Model

The theoretical model presented earlier is a two-equation system where the household savings ratio and the labor-force participation of the aged are endogenous. Having ruled out the application of either ordinary least squares to each structural equation¹⁷ or two-stage least squares to the savings equation,¹⁸ a third alternative was adopted to break the simultaneity. Specifically, the labor-force participation equation was substituted in the structural savings equation, and then ordinary least-squares estimation was applied to the resulting reduced form savings equation and the original labor-force participation equation. This approach is equivalent to indirect least-squares estimation, although without the advantage of being able to derive structural parameters from the reduced form coefficients.

The equations were estimated on cross sections of country averages of 1969-71 annual observations, grouped in samples of 14 industrial countries and 40 developing countries, the size of the samples being dictated chiefly by the availability of data.¹⁹ Other refinements over previous research include the measurement of per capita real income (taking into account intercountry differences in purchasing power) and the rate of growth of labor productivity (i.e., the difference between the growth rates of output and of the labor force). Further, several additional explanatory variables were introduced: various social security benefit and tax rates; the aftertax long-term interest rate on financial savings less inflation (for developing countries the inflation rate alone was used because of the lack of adequate data on interest and tax rates); and an index of income inequality consisting of estimates of the Gini coefficient. However, no attempt was made to deal with the savings/investment simultaneity and to quantify the wealth variable.

With regard to social security variables, special care was taken in selecting the appropriate economic bases to calculate effective benefit and tax rates. Four benefit rates were measured by the ratios to household disposable income of (1) old-age pensions, (2) old-age lump-sum payments (from provident funds), (3) other transfers, and (4) loans. Old-age transfers, in addition, were normalized for the proportion of the elderly in the total population. The tax rate was given by payroll taxes for social security as a ratio of compensation of employees plus entrepreneurial income. Because of insufficient data, for developing countries the tax rate could not be computed, while the benefit rates were taken as ratios to private disposable income.

Albeit unweighted and population-weighted regression estimates are shown in Tables 1 through 4, the following discussion focuses primarily on the unweighted estimates. Weighted estimates (obtained by multiplying each observation by population size) substantially raised the goodness-of-fit of the regressions, but at the cost of introducing a high degree of collinearity among several variables for all samples, and some heteroscedasticity for developing countries. Nonetheless, both sets of results have essentially the same implications regarding the effects of social security.

Table 1.

Industrial Countries: Unweighted Regression Results¹

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 1.

Industrial Countries: Unweighted Regression Results¹

	Dependent Variable
Independent Variables	SH/YH (1)	SH/YH (2)	SH/YH (3)	LPA (4)	LPA (5)
Constant	53.776 (46.748)	20.558 (7.481)	36.656 (10.045)	-11.786 (57.412)	17.828 (49.901)
1/YH	2.294 (5.364)			4.654 (5.614)	5.362 (5.370)
G	0.033 (1.900)	1.794 (0.507)	0.790 (0.516)
NRR	1.486 (1.142)	0.700 (0.583)	1.215 (0.540)
RE/YH	0.284 (0.438)
INQ	19.800 (24.578)
DA	-0.712 (0.494)		-0.601 (0.267)
DM	-0.242 (0.207)	-0.157 (0.093)	-0.270 (0.102)
LEA	-2.208 (2.940)			2.081 (3.360)	0.094 (2.888)
SSP/YH		0.092 (0.039)		-0.178 (0.090)
SSO/YH		-0.406 (0.202)
SST/CN			0.292 (0.107)		-0.522 (0.225)
AGE		-0.118 (0.053)	-0.086 (0.047)	-0.115 (0.127)	-0.220 (0.101)
Summary Statistics 2
${\overline{R}}^2$	0.210	0.725	0.759	0.317	0.386
$\overline{SEE}$	4.380	2.584	2.418	6.242	5.921
$\overline{X}$	11.375	11.375	11.375	12.464	12.464

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

View Table

Table 1.

Industrial Countries: Unweighted Regression Results¹

	Dependent Variable
Independent Variables	SH/YH (1)	SH/YH (2)	SH/YH (3)	LPA (4)	LPA (5)
Constant	53.776 (46.748)	20.558 (7.481)	36.656 (10.045)	-11.786 (57.412)	17.828 (49.901)
1/YH	2.294 (5.364)			4.654 (5.614)	5.362 (5.370)
G	0.033 (1.900)	1.794 (0.507)	0.790 (0.516)
NRR	1.486 (1.142)	0.700 (0.583)	1.215 (0.540)
RE/YH	0.284 (0.438)
INQ	19.800 (24.578)
DA	-0.712 (0.494)		-0.601 (0.267)
DM	-0.242 (0.207)	-0.157 (0.093)	-0.270 (0.102)
LEA	-2.208 (2.940)			2.081 (3.360)	0.094 (2.888)
SSP/YH		0.092 (0.039)		-0.178 (0.090)
SSO/YH		-0.406 (0.202)
SST/CN			0.292 (0.107)		-0.522 (0.225)
AGE		-0.118 (0.053)	-0.086 (0.047)	-0.115 (0.127)	-0.220 (0.101)
Summary Statistics 2
${\overline{R}}^2$	0.210	0.725	0.759	0.317	0.386
$\overline{SEE}$	4.380	2.584	2.418	6.242	5.921
$\overline{X}$	11.375	11.375	11.375	12.464	12.464

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 2.

Developing Countries: Unweighted Regression Results¹

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 2.

Developing Countries: Unweighted Regression Results¹

	Dependent Variable
Independent Variables	SP/YP (1)	SP/YP (2)	SP/YP (3)	LPA (4)
Constant	23.784 (5.770)	22.035 (8.821)	16.642 (2.925)	21.894 (3.730)
1/YP	-0.422 (0.166)	-0.423 (0.206)	-0.363 (0.171)	0.866 (0.267)
G	0.409 (0.360)	0.496 (0.428)	0.508 (0.316)
INF	-0.259 (0.145)	-0.232 (0.175)	-0.290 (0.139)
INQ	2.181 (9.223)	2.247 (10.631)
DA	-0.014 (0.033)	-0.022 (0.037)
DM	-0.047 (0.031)	-0.040 (0.042)
SSP/YP		-0.004 (0.074)		-0.257 (0.095)
SSF/YP		0.151 (0.168)		-0.199 (0.257)
SSO/YP		0.302 (0.394)	0.426 (0.272)
SSL/YP		-0.179 (1.725)
AGE		-0.041 (0.092)		0.250 (0.133)
Summary Statistics 2
${\overline{R}}^2$	0.303	0.226	0.341	0.326
$\overline{SEE}$	4.760	5.016	4.627	8.125
$\overline{X}$	15.584	15.584	15.584	29.594

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

View Table

Table 2.

Developing Countries: Unweighted Regression Results¹

	Dependent Variable
Independent Variables	SP/YP (1)	SP/YP (2)	SP/YP (3)	LPA (4)
Constant	23.784 (5.770)	22.035 (8.821)	16.642 (2.925)	21.894 (3.730)
1/YP	-0.422 (0.166)	-0.423 (0.206)	-0.363 (0.171)	0.866 (0.267)
G	0.409 (0.360)	0.496 (0.428)	0.508 (0.316)
INF	-0.259 (0.145)	-0.232 (0.175)	-0.290 (0.139)
INQ	2.181 (9.223)	2.247 (10.631)
DA	-0.014 (0.033)	-0.022 (0.037)
DM	-0.047 (0.031)	-0.040 (0.042)
SSP/YP		-0.004 (0.074)		-0.257 (0.095)
SSF/YP		0.151 (0.168)		-0.199 (0.257)
SSO/YP		0.302 (0.394)	0.426 (0.272)
SSL/YP		-0.179 (1.725)
AGE		-0.041 (0.092)		0.250 (0.133)
Summary Statistics 2
${\overline{R}}^2$	0.303	0.226	0.341	0.326
$\overline{SEE}$	4.760	5.016	4.627	8.125
$\overline{X}$	15.584	15.584	15.584	29.594

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 1 presents the unweighted results for industrial countries. In regression (1), fitted on the basic household savings function (without the social security variables) the coefficients of the growth rate of productivity, interest rate, income inequality, and dependency ratios carry the expected signs, whereas those of the reciprocal of per capita real income and retained corporate earnings are insignificant and have incorrect signs. Regressions (2) and (3), incorporating alternatively the social security benefit and tax rates-along with the age of the system-while excluding the weakest basic variables,²⁰ indicate a significant positive relationship between savings and old-age transfers or payroll taxation, and a negative relationship between savings and other social security transfers. Labor-force participation regressions (4) and (5) show the expected positive although insignificant coefficients for the reciprocal of income and life expectancy, and negative significant ones for the old-age transfers or payroll taxes and the age of the system.

Table 3.

Industrial Countries: Weighted Regression Results¹

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 3.

Industrial Countries: Weighted Regression Results¹

	Dependent Variable
Independent Variables	SH/YH (1)	SH/YH (2)	SH/YH (3)	LPA (4)	LPA (5)
Constant	64.085 (47.800)	12.154 (6.205)	31.335 (5.823)	198.837 (43.629)	238.263 (49.306)
1/YH	-3.353 (2.140)			-1.910 (2.261)	-2.287 (2.884)
G	1.552 (1.618)	2.486 (0.308)	1.037 (0.262)
NRR	1.681 (0.778)	0.288 (0.395)	0.868 (0.372)
REIYH	0.181 (0.233)
INQ	19.514 (23.625)
DA	-0.677 (0.582)		-0.470 (0.200)
DM	-0.386 (0.230)	-0.136 (0.062)	-0.239 (0.066)
LEA	-1.335 (2.739)			-10.568 (2.681)	-13.127 (2.955)
SSP/YH		0.106 (0.023)		-0.099 (0.069)
SSO/YH		-0.295 (0.146)
SST/CN			0.338 (0.084)		0.072 (0.309)
AGE		-0.079 (0.040)	-0.092 (0.049)	-0.427 (0.102)	-0.565 (0.095)
Summary Statistics 2
${\overline{R}}^2$	0.993	0.998	0.997	0.986	0.984
$\overline{SEE}$	53.900	31.800	35.420	139.500	151.600
$\overline{X}$	511.364	511.364	511.364	710.218	710.218

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

View Table

Table 3.

Industrial Countries: Weighted Regression Results¹

	Dependent Variable
Independent Variables	SH/YH (1)	SH/YH (2)	SH/YH (3)	LPA (4)	LPA (5)
Constant	64.085 (47.800)	12.154 (6.205)	31.335 (5.823)	198.837 (43.629)	238.263 (49.306)
1/YH	-3.353 (2.140)			-1.910 (2.261)	-2.287 (2.884)
G	1.552 (1.618)	2.486 (0.308)	1.037 (0.262)
NRR	1.681 (0.778)	0.288 (0.395)	0.868 (0.372)
REIYH	0.181 (0.233)
INQ	19.514 (23.625)
DA	-0.677 (0.582)		-0.470 (0.200)
DM	-0.386 (0.230)	-0.136 (0.062)	-0.239 (0.066)
LEA	-1.335 (2.739)			-10.568 (2.681)	-13.127 (2.955)
SSP/YH		0.106 (0.023)		-0.099 (0.069)
SSO/YH		-0.295 (0.146)
SST/CN			0.338 (0.084)		0.072 (0.309)
AGE		-0.079 (0.040)	-0.092 (0.049)	-0.427 (0.102)	-0.565 (0.095)
Summary Statistics 2
${\overline{R}}^2$	0.993	0.998	0.997	0.986	0.984
$\overline{SEE}$	53.900	31.800	35.420	139.500	151.600
$\overline{X}$	511.364	511.364	511.364	710.218	710.218

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 4.

Developing Countries: Weighted Regression Results¹

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

Table 4.

Developing Countries: Weighted Regression Results¹

	Dependent Variable
Independent Variables	SP/YP (1)	SP/YP (2)	SP/YP (3)	LPA (4)
Constant	19.026 (6.237)	26.170 (13.364)	23.246 (1.708)	17.018 (5.784)
1/YP	-0.313 (0.106)	-0.551 (0.286)	-0.438 (0.073)	1.117 (0.311)
G	0.479 (0.279)	-0.049 (0.521)
INF	-0.201 (0.124)	-0.255 (0.124)	-0.268 (0.091)
INQ	0.754 (7.844)	5.384 (9.952)
DA	-0.127 (0.169)	0.010 (0.163)
DM	-0.003 (0.038)	-0.027 (0.073)
SSP/YP		0.164 (0.070)	0.126 (0.039)	-0.328 (0.075)
SSF/YP		0.202 (0.138)		-0.789 (0.299)
SSO/YP		-1.447 (0.771)	-1.515 (0.406)
SSL/YP		2.773 (1.790)
AGE		-0.141 (0.140)		0.657 (0.196)
Summary Statistics 2
${\overline{R}}^2$	0.996	0.997	0.997	0.998
$\overline{SEE}$	67.240	59.740	57.760	170.000
$\overline{X}$	367.507	367.507	367.507	1,002.914

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

View Table

Table 4.

Developing Countries: Weighted Regression Results¹

	Dependent Variable
Independent Variables	SP/YP (1)	SP/YP (2)	SP/YP (3)	LPA (4)
Constant	19.026 (6.237)	26.170 (13.364)	23.246 (1.708)	17.018 (5.784)
1/YP	-0.313 (0.106)	-0.551 (0.286)	-0.438 (0.073)	1.117 (0.311)
G	0.479 (0.279)	-0.049 (0.521)
INF	-0.201 (0.124)	-0.255 (0.124)	-0.268 (0.091)
INQ	0.754 (7.844)	5.384 (9.952)
DA	-0.127 (0.169)	0.010 (0.163)
DM	-0.003 (0.038)	-0.027 (0.073)
SSP/YP		0.164 (0.070)	0.126 (0.039)	-0.328 (0.075)
SSF/YP		0.202 (0.138)		-0.789 (0.299)
SSO/YP		-1.447 (0.771)	-1.515 (0.406)
SSL/YP		2.773 (1.790)
AGE		-0.141 (0.140)		0.657 (0.196)
Summary Statistics 2
${\overline{R}}^2$	0.996	0.997	0.997	0.998
$\overline{SEE}$	67.240	59.740	57.760	170.000
$\overline{X}$	367.507	367.507	367.507	1,002.914

¹The figures in parentheses are standard errors of the regression coefficients.

²The statistics ${\overline{R}}^2$ and $\overline{SEE}$ represent the coefficient of determination and the standard error of estimate, respectively, corrected for loss in degrees of freedom; $\overline{X}$ is the arithmetic mean of the dependent variable.

The response of household savings to independent variables that exhibit significant coefficients is summarized in Table 5. The first column shows the elasticity of the household savings ratio with respect to these variables in industrial countries, based on unweighted regressions (2) and (3). Estimates of the elasticity with respect to the growth rate, interest rate, and dependency ratios are broadly consistent with those reported by other authors.²¹ As to the social security variables, it appears that a rise of 1 per cent in the rate of old-age pensions or of payroll taxes leads to an increase of nearly ½ of 1 per cent in the savings ratio, whereas a rise of 1 per cent in the rate of other benefits depresses the savings ratio by ½ of 1 per cent. The size of the negative influence of the age of the system depends on the estimated regression.

Table 5.

Elasticity of Savings with Respect to Selected Independent Variables¹

¹Elasticity estimates were computed using the arithmetic means of the relevant variables in each sample. The elasticity of savings with respect to variables whose coefficients are significant at the 10 per cent level or more is assumed to be zero.

²Estimates based on coefficients appearing in regressions (2) and (3) in Tables 1 and 3. The presence of two estimates for a given variable indicates a range of elasticities, derived from the coefficients of that variable in these regressions.

³Estimates based on coefficients from regression (3) in Tables 2 and 4.

Table 5.

Elasticity of Savings with Respect to Selected Independent Variables¹

	Unweighted Regressions		Weighted Regressions
Independent variables	Industrial countries 2(SH/YH)	Developing countries 3(SP/YP)	Industrial countries 2(SH/YH)	Developing countries 3(SP/YP)
1/YP		-0.19		-0.44
G	0.00/0.56		0.34/0.82
NRR	0.00/0.07		0.00/0.09
INF		-0.10		-0.13
DA	-1.07/0.00		-0.76/0.00
DM	-1.41/0.00		-1.27/-0.72
SSP/YH	0.47		0.43
SSP/YP		0.00		0.08
SSO/YH	-0.49		-0.28
SSO/YP		0.00		-0.24
SST/CN	0.34		0.34
AGE	-0.56/0.00		-0.32/0.00

¹Elasticity estimates were computed using the arithmetic means of the relevant variables in each sample. The elasticity of savings with respect to variables whose coefficients are significant at the 10 per cent level or more is assumed to be zero.

²Estimates based on coefficients appearing in regressions (2) and (3) in Tables 1 and 3. The presence of two estimates for a given variable indicates a range of elasticities, derived from the coefficients of that variable in these regressions.

³Estimates based on coefficients from regression (3) in Tables 2 and 4.

View Table

Table 5.

Elasticity of Savings with Respect to Selected Independent Variables¹

	Unweighted Regressions		Weighted Regressions
Independent variables	Industrial countries 2(SH/YH)	Developing countries 3(SP/YP)	Industrial countries 2(SH/YH)	Developing countries 3(SP/YP)
1/YP		-0.19		-0.44
G	0.00/0.56		0.34/0.82
NRR	0.00/0.07		0.00/0.09
INF		-0.10		-0.13
DA	-1.07/0.00		-0.76/0.00
DM	-1.41/0.00		-1.27/-0.72
SSP/YH	0.47		0.43
SSP/YP		0.00		0.08
SSO/YH	-0.49		-0.28
SSO/YP		0.00		-0.24
SST/CN	0.34		0.34
AGE	-0.56/0.00		-0.32/0.00

¹Elasticity estimates were computed using the arithmetic means of the relevant variables in each sample. The elasticity of savings with respect to variables whose coefficients are significant at the 10 per cent level or more is assumed to be zero.

²Estimates based on coefficients appearing in regressions (2) and (3) in Tables 1 and 3. The presence of two estimates for a given variable indicates a range of elasticities, derived from the coefficients of that variable in these regressions.

³Estimates based on coefficients from regression (3) in Tables 2 and 4.

These results confirm the extended life-cycle hypothesis, but they do not support the view that income inequality promotes savings propensities and that households count on corporate resources to meet their demand for wealth. The absence of a significant relationship between savings and inequality might be attributed chiefly to the crudeness of data on income distribution.²² However, the lack of a negative impact of corporate savings underscores the households’ perception that large, widely held corporations in industrial countries are independent of shareholder control, whereby households do not take into account corporate distribution policy in their decision to save.²³

More important, the coefficients of the social security variables, and the corresponding elasticity estimates, suggest that in industrial countries the retirement effect overwhelms the net wealth effect of old-age transfers, while other transfers tend to exercise a negative wealth effect on savings, reflecting asset substitution. If the influence of social security is tested with the payroll tax rate instead of the benefit rates, it seems that, on balance, the schemes have a positive impact on household savings—presumably for the most part owing to the retirement effect. Yet, as the social security system grows older, the inducement to save is apparently eroded as enrolled households step up asset substitution, perhaps because of increasing belief in the adequacy of future benefits to provide protection for old age and against current contingencies.

The statistical performance of the model in the unweighted form, shown in Table 2, is on the whole weaker for developing countries, which is not surprising in light of the less reliable underlying data.²⁴ Given the limited number of countries for which national income accounts are broken down between the household and corporate components of the private sector, it was necessary to use data on private instead of household savings and income. By the same token, it was not possible to test the relationship between the household savings ratio and the corporate retention ratio. This impediment, in itself, should not constitute a major drawback for two reasons: first, the corporate share in private savings or income is smaller in developing countries (i.e., for the dozen or so countries where the breakdown is available) than in industrial countries; second, much of the corporate activity is carried out by family-owned enterprises whose savings activity may well be regarded as an extension of household savings behavior.

The coefficients of all the basic exogenous variables display the expected signs, but those of the reciprocal of income, the growth rate, and the inflation rate alone have some significance. Among the social security variables, which—besides the old-age pension and other transfer rates—also include the rates of old-age lump-sum transfers and of loans, only the negative coefficient of the old-age pension rate and the positive coefficient of the age of the system are significant in the unweighted run (4) of the labor-force participation equation.

Reflecting to a greater extent the behavior of more populated developing countries, the coefficients of several social security variables in Table 4 are highly significant: the rates of old-age pensions and of other transfers with positive and negative signs, respectively, in weighted regression (3); the rates of all old-age transfers with a negative sign, and the age of the system with a positive sign, in regression (4).

Accordingly, the elasticity estimates (based on weighted regression results) for developing countries, shown in the fourth column of Table 5, are larger than the estimates in the second column (based on unweighted results). In either case, the elasticities of the private savings ratio with respect to social security benefit rates are substantially smaller than the corresponding estimates for industrial countries. The weighted estimates for developing countries indicate that an increase of 1 per cent in the old-age pension rate would lead to a rise of less than $\frac{1}{10}$ of 1 per cent in the savings ratio, and an equivalent increase in the rate of other transfers would result in a drop of ¼ of 1 per cent in the savings ratio. No savings response to changes in benefit rates is detected in the unweighted estimates.

In contrast to the results for industrial countries, where per capita income has virtually no impact on the propensity to save, in developing countries an increase in income depresses the labor supply of the aged and raises the savings ratio. Although the latter is in line with the extended life-cycle hypothesis, it can also be interpreted as a vindication of the Keynesian view. In any event, the difference in behavior between developed and developing countries may be ascribed to the extended family structure and to the greater variance in per capita income in developing regions. Moreover, in developing countries, the inflation rate performs fairly well as an inverse proxy for the real return on financial savings, the reason being that in many developing countries nominal interest rates—often subject to discretionary ceilings—fail to catch up with rapidly rising prices, so that real interest rates decline and may become negative.

With regard to the wealth and retirement effects of social security on private savings, the unweighted regressions lend support to the null hypothesis—notwithstanding the possible income effect suggested by the significance of the income variable. At the same time, estimates of the labor-force equation indicate a potentially strong retirement effect of old-age pensions, dampened by the age of the system. The insignificant coefficient of old-age benefits in the savings equation may be taken to mean that the retirement effect is fully offset by an equally powerful negative net wealth effect.²⁵ But perhaps a more plausible explanation is that neither effect operates, given the social security function of the extended family in many developing countries. In contrast, the weighted-regression results reveal some asset substitution in response to other than old-age transfers in countries with large populations.

All in all, the estimates of the savings function for developing countries are probably subject to considerable measurement error and to some misspecification.²⁶ The problem stems, at least in part, from the heterogeneous nature of the sample countries, which possess diverse cultural, social, and institutional characteristics not captured by the model.²⁷ More explicit account should also be taken of differences in economic environment among developing countries, particularly those involving variations in uncertainty and risk, which are reflected to a limited extent by the rate of inflation.

V. Summary and Conclusion

A model intended to explain differences in savings propensities of households among countries has been formulated in conformity with the absolute-income hypothesis, the permanent-income hypothesis, and an extended version of the life-cycle hypothesis. Comprised of two structural equations—the dependent variables being the average propensity to save of households and the labor-force participation of the aged—this eclectic model is especially useful in that it incorporates the effects of social security schemes. Such schemes provide protection for old age, through periodic pension payments or lump-sum disbursements from provident funds, and against current contingencies, through payments for medical expenses, unemployment, work injury, indigence, and, in some countries, the extension of loans. Although initially many social security schemes were funded, by now most of them are virtually unfunded, so that payroll taxes form the main source of revenue to finance benefits.

Social security institutions in their present form can have three distinct effects on the household savings ratio. First, the income effect consists of the possible repercussion of changes in benefit or tax rates transmitted through the average level of disposable income and the degree of income inequality. Second, the wealth effect represents the direct savings response of households to expected future benefits. Households may react either by cutting back savings if they perceive such benefits to be an adequate substitute for personal savings, or by raising their savings propensity if social security schemes enhance their awareness of the need to accumulate additional wealth as protection against various contingencies. Third, the retirement effect indicates the positive indirect relationship between old-age benefits and the savings ratio, via changes in the labor supply of the aged; an increase in old-age benefits is likely to induce earlier retirement, which in turn may lead to a higher propensity to save for the longer retirement period. Of these consequences of social security, particularly the wealth and retirement effects have attracted attention and generated controversy in the literature (as opposed to the income effect, which is likely to vanish over the long run, if benefits and taxes have similar incidence on households), and they are also the focus of this analysis.

Estimates of the model using cross sections of 1969-71 average observations, divided into samples of industrial and developing countries, suggest that social security schemes influence decisions on household savings, especially in industrial countries. In these countries, when allowance is made for a differential impact of various types of benefit, the (positive) retirement effect outweighs the (negative) wealth-substitution effect of old-age transfers, while the latter effect prevails for other transfers. An alternative test, performed under the assumption that households consider payroll taxes to be a proxy for future social security benefits, reveals that the positive retirement effect overrides the other effects. However, the results also indicate that the wealth-substitution effect strengthens as the social security system becomes older, apparently because of the gradual rise of household confidence in the system’s ability to deliver an adequate stream of benefits.

For developing countries, the statistical power of the estimates is generally lower and the nature of the findings is less conclusive. Despite the significant inverse impact of old-age transfers on the labor supply of the aged, the net effect of these transfers on savings is a draw, implying that the retirement effect and the wealth-substitution effect offset each other. But it is more probable that these effects, as well as the validity of the life-cycle hypothesis, are precluded by the extended family structure, which in many developing countries operates like an unfunded social security system. Nonetheless, transfers for other than old age seem to have a negative net wealth effect on the savings ratio in countries with a relatively large population.

From this evidence, it can be inferred that the expansion of the rate or coverage of old-age social security pensions and of payroll taxes does not have an adverse impact on household savings. If anything, such benefits and taxes are more likely to encourage savings in industrial countries—although this incentive may wear off as the social security institutions mature. There seems to be, however, substitution of other social security transfers for savings in industrial countries and more populated developing countries. Therefore, to some extent, social security can serve as an instrument of resource mobilization if the authorities invest the surplus of social security institutions (at least to the extent that they are funded) in projects with an adequate social marginal rate of return, without inhibiting private savings and investment activity.

Nevertheless, policy decisions with regard to social security benefits and taxation should be made in a broader setting. Investigation should also be undertaken of their consequences on vertical and horizontal equity (taking into account, for instance, the usual regressivity of the payroll tax across income brackets and its uneven coverage of various income sources), the country’s comparative advantage, and the choice of factor inputs, over the long run, as well as on internal and external stability over the short run.