Precautionary saving and earnings uncertainty in Japan: A household-level analysis

J. Japanese Int. Economies 17 (2003) 192–212 www.elsevier.com/locate/jjie

Precautionary saving and earnings uncertainty in Japan: A household-level analysis Yanfei Zhou ∗ National Institute of Population and Social Security Research, 2-2-3 Uchisaiwaicho, Tokyo, Japan Received 3 October 2000; revised 17 January 2003

Zhou, Yanfei—Precautionary saving and earnings uncertainty in Japan: A household-level analysis This paper improves upon the methodology of Dardanoni (1991. Appl. Econ. 23, 153–160) and applies it to household-level data from a Japanese Government survey in order to analyze the impact and importance of precautionary saving arising from earnings uncertainty. The major results can be summarized as follows: (1) earnings uncertainty has a significant impact on household consumption and saving; (2) precautionary saving arising from earnings uncertainty comprises 5.557% of the total saving of salaried worker households and 64.3% of the total saving of agricultural, forestry, fisheries, and self-employed households; (3) the prediction of Carroll and Summers’ buffer stock saving hypothesis that young households will be more likely to save for precautionary purposes than older households is confirmed; and (4) occupation, age, and educational attainment significantly affect the degree of household earnings uncertainty. J. Japanese Int. Economies 17 (2) (2003) 192– 212. National Institute of Population and Social Security Research, 2-2-3 Uchisaiwaicho, Tokyo, Japan.  2003 Elsevier Science (USA). All rights reserved. JEL classification: E270; E290 Keywords: Precautionary saving; Earnings uncertainty; Two-step estimation

1. Introduction Compared to the extremely strong consumption demand of American households, Japanese households behave as though they are reluctant to consume, and their saving * Present address: Institute of Social and Economic Research, Osaka University, 6-1 Mihogaoka, Ibaraki,

Osaka, Japan. E-mail address: [email protected]. 0889-1583/03/$ – see front matter  2003 Elsevier Science (USA). All rights reserved. doi:10.1016/S0889-1583(03)00012-1

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

193

rate is much higher than that of Americans. Excessive household saving is regarded as one of the main factors inhibiting the recovery of the Japanese economy. Although Japan’s high household saving rate is determined jointly by a large number of factors (Horioka, 1990), precautionary saving has been regarded as one of the most important ones, especially since the collapse of Japan’s bubble economy. Even so, few empirical analyses of the precautionary saving behavior of Japanese households have been conducted. The purpose of the present paper is to conduct just such an analysis using household-level data from a Japanese Government survey. The idea that people engage in precautionary saving dates back to Friedman (1957). Later studies by Leland (1968), Sandmo (1970), and Dreze and Modigliani (1972) show that precautionary saving in response to risk is associated with convexity of the marginal utility function, or a positive third derivative of utility function. Recent researches by Blanchard and Mankiw (1988), Zeldes (1989), and Caballero (1990) have bridged the gap between theoretical work and the empirical literature by successfully deriving a closedform solution of consumption function with earnings uncertainty. However, the empirical results concerning the importance of precautionary saving are inconclusive. Using data from the 1983 Survey of Consumer Finances (SCF), Carroll (1992) finds that 43% of respondents reported being prepared for emergencies as being their most important reason for saving. Skinner (1988) concludes that saving that arises as a precaution against future uncertainty is more than one-half of total life-cycle saving under certain assumptions. Carroll and Summers (1991) suggest that consumers do not save for retirement over most of their working lives, say until roughly age 45 or 50, and thus that the certainty equivalence LC/PIH can explain consumer behavior only between about age 50 and retirement. On the other hand, Guiso et al. (1992) use a self-reported measure1 of the subjective uncertainty of future earnings and finds that, on average, precautionary saving accounted for only 2% of Italian households’ net worth in 1990. Using the Family, Member, and Detailed Expenditure files of the CEX for the 1980–1993 period, Parker (1999) also finds no evidence that precautionary saving is responsible for the failure of consumption smoothing to hold in the USA. There has been very little empirical research done on precautionary saving in Japan. According to a household survey of Japan (Horioka and Watanabe, 1997), saving for illness, disaster and other unforeseen expenditures, and saving for peace of mind account for a total of 8.9% of average household disposable income. Ginama (1988) presents time series estimates of the ratio of precautionary saving to total personal saving in the USA and Japan. He finds that precautionary saving can explain Japan’s relatively high saving rate to some extent but that the precautionary saving motive of Japanese households was significant only during the period of the first oil crisis (1974–1976). For example, he estimates the share of precautionary saving to have been 5.81% in 1974, 4.13% in 1975, 2.57% in 1976, and less than 0.5% thereafter. Similarly, using time series data for the 1971–1987 period from the Survey of Consumption Trends in Japan, Ogawa (1991) investigates the importance of precautionary saving by using the variance of income 1 Every income recipient was asked to assign probabilities to various ranges of inflation and percentage

increases in nominal earnings one year from now. These two marginal distributions were then used to calculate the subjective uncertainty of real earnings.

194

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

growth rate expectations of Japanese households as a proxy for the degree of income uncertainty. Like Ginama, he finds that precautionary saving comprised a significant share of worker households’ saving only at the time of the first oil crisis. For example, he estimates the share of precautionary saving to have been 10.53% in 1974, 10.62% in 1975, 7.82% in 1976, and much lower thereafter. However, he finds precautionary saving to have been very important throughout the period of analysis in the case of farm households. Most previous analyses of precautionary saving use either simulation techniques (Skinner, 1988; Zeldes, 1989) or a measure of earnings uncertainty that is calculated from time series data (Ginama, 1988; Ogawa, 1991; Hahm, 1999), while some studies use self-reported information to construct a subjective measure of earnings uncertainty (Carroll, 1992; Guiso et al., 1992). While useful, all three approaches have their drawbacks: simulation analyses have been criticized because the results depend on their assumptions about risk aversion and the income-generating process; time series data provide only pooled statistics and thus cannot capture the precautionary saving behavior of individual households; and self-reported measures of earnings uncertainty are problematic because of the gap between perception and behavior and because of respondents’ lack of understanding about how to interpret earnings uncertainty. Possibly because of the difficulty in constructing a measure of earnings uncertainty based on cross section data, there have been very few studies of precautionary saving that employ cross section data with the sole exception of Dardanoni (1991) and there have been no such studies for Japan. However, measures of earnings uncertainty based on cross section data are valuable as a possible substitute for the three kinds of measures discussed above. For one thing, cross section data is much more readily available than panel data or the self-reported measures discussed earlier, and thus measures of earnings uncertainty that are calculated from cross section data would undoubtedly be widely used if available. The present paper improves upon Dardanoni’s (1991) approach of measuring earnings uncertainty by creating dozens of homogenous groups and regarding income variance within each group as the approximate income uncertainty of individuals in that group. The paper contributes to the literature on precautionary saving in the following ways: (1) it is the first attempt to analyze the impact of earnings uncertainty on the consumption and saving of Japanese households using household-level data, (2) it improves upon Dardanoni (1991) by employing a more complete model, by conducting a careful analysis of what attributes affect earnings uncertainty before deciding which grouping variables to use, by using not only grouped data but also household-level data in the estimations, and by estimating the consumption function not only for the full sample but also for young and old, salaried worker and agricultural, forestry, fisheries, and self-employed households separately, and (3) it uses the variance–covariance matrix adjusted estimates to make inferences about the consumption model in the second stage. The organization of this paper is as follows: the basic model and measurement of earnings uncertainty are set out in Section 2, the data used in the analysis are described in

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

195

Section 3, the main results are presented in Section 4, and our conclusions are summarized in Section 5.

2. Model and the measurement of earnings uncertainty 2.1. Theoretical conditions for precautionary saving It has been known since Leland (1968) and Sandmo (1970) that precautionary saving in response to risk is associated with convexity of the marginal utility function or a positive third derivative of a von Neumann–Morgenstern utility function. For example, Leland (1968) assumes that the consumers have no initial assets and that they attempt to maximize expected utility over two periods while facing earnings uncertainty in the second period. In order for precautionary saving to be positive, earnings uncertainty must reduce this gap by increasing saving above its initial level. More generally, Kimball (1990) notes that if an individual’s utility is a function of consumption, it should also be a function of a control variable S and an exogenous random variable Y . Thus, the individual solves the problem ∂U = 0. ∂S If ∂U /∂S is convex in Y , an increase in the variability of Y will result in an increase in the optimal choice of S. Kimball (1990) proves that this above theoretical condition of precautionary saving is isomorphic to that of the Arrow–Pratt measure of risk aversion, which means that a large body of knowledge about risk aversion can be applied to precautionary saving. For example, one measure of the strength of the precautionary saving motive is analogous to the Arrow–Pratt measure of risk aversion. Nevertheless, the above results do not provide a sensible way to test the sensitivity of consumption to wealth or transitory income; i.e., they do not deal with the slope of the consumption function. Caballero (1990) tried to bridge the gap between the theoretical work and the empirical literature by successfully deriving a closed-form solution of the consumption function with precautionary saving under constant absolute risk aversion (CARA) utility2 and a well-behaved income process. In particular, he emphasized that if labor income and its variance innovations are positively correlated, households’ average propensity to save implied by a CARA utility function is higher than that of the certainty-equivalent model. Besides, Blanchard and Mankiw (1988) and Dardanoni (1991)3 also contributed to the derivation of an optimal consumption function under earnings uncertainty. In brief, optimal consumption is jointly decided by the following max EUS (Y, S)

using the first-order condition E

2 Zeldes (1989) notes that a closed-form solution can no longer be obtained when the utility function is instead assumed to be of the standard constant relative risk aversion (CRRA) form. 3 Employing the utility function U (C) = − exp(−kC), where k is the risk-aversion parameter, Blanchard and Mankiw (1988) obtain a closed-form solution for optimal consumption: Ct = (1/(T − t + 1))Wt + Yt − (γ (T − t)/4)V where T is the time of death, γ is the coefficient of prudence, W is non-human assets, Y is permanent income, and V is earnings uncertainty. Dardanoni’s (1991) solution is essentially the same as that of Blanchard and Mankiw (1988).

196

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

set of variables:
 C = f Y p , ASSET, SSW, V (Y ) ,

(1)

where Y p is permanent labor income. ASSET is nonhuman assets and SSW is social security wealth. Real assets and financial assets are merged together as ASSET because the values of both are reported by households, whereas the value of SSW needs to be estimated.4 The last term V (Y ) is the variance of shocks to labor income and measures the necessity of precautionary saving. Equation (1) can be written in linear form as follows: C = a0 + a1 Y p + a2 ASSET + a3 SSW + a4 V (Y ) + u.

(2)

One of the main difficulties of estimating this consumption function is the unobservable nature of almost all of the variables it involves. As our data set is not a panel data, we do not have the information needed to estimate the future earnings variance of each household. However, we can eliminate this problem by using a proxy for individual labor income uncertainty, as done by Dardanoni (1991). 2.2. Estimating the consumption function There are at least two ways of estimating Eq. (2). The first is to use group averages as Dardanoni (1991) did by simply dividing the sample into many homogeneous groups and employing the income variance within each group as an index of earnings uncertainty for each household within that group. The consumption function can therefore be expressed as  + a2 ASSET + a3 SSW + a4 V (Y ) + u,  = a0 + a1 Y C

(2.1)

 is the average consumption of households in each group, ASSET is the average where C amount of real assets (RA) and financial assets (FA) in each group, and SSW is average ) social security wealth in each group. Average disposable income within each group (Y is used as a proxy for permanent labor income of households in that group. V (Y ) is the variance of labor income among households within each group and is used as a proxy for the time series variance of permanent income for individual households. The advantage of this method is that the calculations are simple. The weakness of this method is that by replacing individual data with group averages, the sample size will be sharply decreased and only the group fixed effect of earnings uncertainty will be estimated. Moreover, heteroscedasticity will arise because of the different numbers of households in each group. The second way of estimating the consumption function is to use individual data for all of the variables except for earnings uncertainty (V ): p + a2 ASSET + a3 SSW + a4 V (Y ) + v. C = a0 + a1 Y

(2.2)

The advantage of this method is that the sample size will be far larger so that the parameter estimates can be expected to be more stable and robust. The effect of income variance on the consumption of individual households can also be possibly obtained with the estimates. Equation (2.2) will actually be estimated by a two-step procedure. In the first stage, I first 4 See Appendix B for a detailed explanation of the calculation method for SSW.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

197

regress labor income on a vector of exogenous variables to obtain the predicted value of p , and then use it in the second stage regression of Eq. (2.2).5 permanent income, Y In the cases of both Eqs. (2.1) and (2.2), a strict fulfillment of precautionary saving theory requires that a1 > 0,

a2 > 0,

a3 > 0,

a4 < 0.

The coefficients of permanent income (a1 ), ASSET (a2 ), and SSW (a3 ) should all be positive because the higher the household’s permanent income, ASSET, or SSW, the more it should consume. In both equations, we should pay particular attention to whether or not a4 < 0 in order to see if there is a negative effect of earnings uncertainty on saving. 2.3. Measurement of earnings uncertainty Using cross section data for Britain, Dardanoni (1991) estimated income variances by grouping the sample into many groups whose household heads belong to the same industry, economic position, and skill level. The variance of income within each group is regarded as the approximate income uncertainty of each household in that group. Therefore, this measurement of earnings uncertainty will not be appropriate unless the wage variability of households within each group is homogenous enough and income variance varies significantly enough among groups. These are, of course, somewhat strong assumptions, and we need to perform further tests thereof. However, Dardanoni (1991) never provided any justification for assuming that the classifying households by industry, economic position, and skill level leads to groups of homogenous households, nor did he show whether income variances differ significantly among groups or not. This paper adopts Dardanoni’s (1991) method but also proposes a simple way of showing that our measure of earnings uncertainty is reliable. More specifically, we use regression analysis to identify the factors that affect the homogeneity or heterogeneity of households with respect to labor income (Y ), assets, social security wealth (SSW), and income variance (V ). 2.3.1. Grouping the data To keep households within each group as homogeneous as possible in the levels of Y , assets, SSW, and V , we must select criteria that are most likely to influence the above household characteristics. Attributes that affect household earnings uncertainty were selected in four stages. In the first stage, the labor income of the household head was regressed on various potentially important factors (Table 1). According to the results, the age and educational attainment of the household head as well as the household’s residence place seem to significantly affect the labor income of the household head. Moreover, as we stated before, the occupation of the household head (for example, whether he or she is a salaried worker or not) will also influence the household head’s labor income. In the second stage, we employed these 5 This method of two-step least squares will lead to biased standard errors. Therefore, I performed a variance covariance adjustment to the estimates later.

198

Table 1 Estimation of labor income by profession Explanatory variable

Full-time salaried workers (a)

AGE AGE ∗ AGE

39.84057

Agricultural, forestry, fisheries and self-employed workers

(b)

(c)

(5.327)***

−0.338402 (−3.817)***

(d)

35.072

(2.688)***

−0.269

(−1.494)

9.025

(e)

(f)

(3.364)***

2.636374

(0.524)

(3.659)*** 23.42052

(1.501)

(7.5)***

−4.985

(−0.446)

(3.645)***

96.1310

(3.88)***

85.530

(3.644)***

200.4903

(9.727)***

210.6804 (9.688)***

200.804

(9.732)***

Senior high school dummy—EDU2

117.5543

(3.91)***

101.5929

(3.2)***

117.038

(3.889)***

104.539

(1.582)

127.9193 (1.926)**

123.3998

(1.839)*

Junior college dummy—EDU3

175.6424

179.0085

(1.641)*

85.5387

Over 500 employees firm size dummy

30.3459

227.6372

(5.093)***

196.8997

(4.17)***

226.496

(5.057)***

175.600

(1.606)

College or graduate school dummy—EDU4 195.2733

(6.006)***

176.6167 (5.148)***

194.755

(5.983)***

344.819

(4.43)***

Big city residence dummy—Metro2

−45.2607

(−1.649)*

−39.0587 (−1.345)

−44.433

(−1.614)

−81.149

(−1.096) −87.0406 (−1.181) −84.97398 (−1.15)

Small city residence dummy—Metro3

−63.8671 (−2.112)** −41.7093 (−1.308)

−63.672 (−2.104)**

Health dummy (1 if unhealthy, 0 otherwise)

−72.5828

−72.429

Constant Sample size /Adj. R 2

−632.1318 (−4.141)*** 369.0438 (9.361)*** −554.619 (−2.398)** 69.938 962/0.2449 962/0.1564 962/0.2443 255/0.1037

(−1.553)

−50.7452 (−1.028)

(−1.549)

18.580

(0.235)

(1.616)

376.2480 (4.817)*** 369.3147 (4.655)***

31.1257

(0.395)

28.2438

(0.357)

−146.652 (−1.381) −139.5087 (−1.321) −143.1128 (−1.351) (0.454)

391.1105 (4.337)*** 290.2491 255/0.1109 255/0.1082

(1.366)

Notes. (1) The dependent variable is the labor income of the household head. In each column, the first number is the estimated OLS coefficient and the second is the t value calculated using robust standard errors. (2) Metro1 = 1 if reside in Tokyo and the other 11 largest cities; Metro2 = 1 if reside in other big cities with a population of more than 50,000; Metro3 = 1 if reside in cities and towns with a population of less than 50,000. (3) EDU1 = 1 if household head is primary or junior high school educated. *,**,*** Significant at the 10, 5, and 1% levels, respectively.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

26.5173

Cohort variable–c(A) 100–500 employees firm size dummy

Educational attainment

EDU = 1 (primary or junior high school)

EDU = 2 (senior high school)

EDU = 3 (junior college)

EDU = 4 (college or graduate school)

Age of the household head

20–30 31–40 41–50 51–60 20–30 31–40 41–50 51–60 20–30 31–40 41–50 51–60 20–30 31–40 41–50 51–60

PRO = 1 (full-time salaried workers in private companies with fewer than 100 employees)

5

10

31

74

54

77

114

88

7

16

15

9

16

51

42

26

PRO = 2 (full-time salaried workers in private companies with 100 or more employees)

1

3

21

39

39

112

127

122

12

17

16

5

55

129

118

55

PRO = 3 (full-time salaried workers in government agencies or other organizations)

2

1

3

7

11

32

63

40

3

9

7

6

10

49

61

25

PRO = 4 (agricultural, forestry, fisheries, and self-employed)

9

9

27

68

15

49

109

94

4

11

10

8

1

22

42

31

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

Table 2 Number of households in each homogeneous cell

199

200

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

four attributes (the age, educational attainment, and occupation of the household head and the household’s place of residence) as criteria when grouping the data. In order to insure that group averages are reliable, all groups containing less than five households were eliminated from the sample. In the third stage, we investigated whether or not the five attributes we selected have a significant impact on household income variance. We found that the income variance of individuals within each group differs significantly by the occupation, age, and educational attainment of the household head but that it does not differ significantly by the household’s place of residence. Therefore, in the fourth stage, we grouped households by the occupation,6 age, and educational attainment7 of the household head only (Table 2). Grouping by these three criteria results in 56 groups with five or more observations. The largest group, which contains households with a head who is in his 30s, who is college-educated, and who works for a large private company, contains 129 observations. Households in the same group are assumed to face the same level of earnings uncertainty, and households in different groups are assumed to face different levels of earnings uncertainty. 2.3.2. Income variance—earnings uncertainty Dardanoni (1991) suggests employing the income variance within each homogeneous group as a proxy for the earnings uncertainty of all households in that group. This index can be expressed mathematically as 2 1 
Yi − Y , n−1 n

V (Y ) =

i=1

where Y is the average of income in the group and n is the number of individuals in the group. However, we must take into account of the income of household head as well as that of the spouse. Hence, we utilize an identity, Y = Yh + Ys , where Yh and Ys are the head’s income and the spouse’s income, respectively, in order to calculate the earnings uncertainty measure: V (Y ) = V (Yh ) + V (Ys ) + 2 cov(Yh , Ys ) n n  
2
2 = Yih − Y h /(n − 1) + Yis − Y s /(n − 1) i=1

i=1

6 The occupation variable classifies each household according to whether the head is a full-time salaried

worker in a small private company with less than 100 employees, in a large private company with 100 or more employees, in a government agency or other organization, is self-employed, or is an agricultural, forestry, or fisheries worker. It is likely that agricultural, forestry, fisheries, and self-employed workers face more earnings uncertainty than salaried workers and that employees of small companies face more earnings uncertainty than employees of large companies because their jobs are generally less stable in Japan. 7 Four age groups of the household head were used: 20–30, 31–40, 41–50, and 51–60. The relatively high unemployment rate of young and old workers suggests that they may face more earnings uncertainty than middleaged workers. Four categories of educational attainment of the household head were used: junior high school, senior high school, junior college, or college graduate. The higher one’s educational attainment, the less earnings uncertainty one is likely to face in the future because of the effect of human capital accumulation.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

201

  n 

 Yih − Y h Yis − Y s (n − 1), +2 i=1

where n is the number of observations in a specific group.

3. Data This study uses data on 2441 Japanese households taken from the 1996 Survey on the Financial Asset Choice of Households (SFACH),8 a survey of 3942 households whose heads are aged 20 or older (including single households) from throughout Japan. The number of observations was reduced to 2441 in four stages. All households headed by females, students, unemployed workers, part-time workers and aged individuals are excluded from the sample. First, we excluded all households headed by a female (n = 357) regardless of whether she is single or married. A substantial fraction of such females are widows whose permanent income is determined primarily by the lifetime earnings of their deceased husbands on whom no information was available. Households headed by a married female are mainly poorer households and their average income is only about half that of all households. The permanent income and saving behavior of young single females may depend on her future husband, information on whom is not yet available. Next, households headed by students (n = 16) or by unemployed (n = 52) or by part-time workers (n = 24) were also excluded because most students have no income and the income profiles of parttime workers are difficult to predict. We also excluded households that reported negative disposable income (n = 2). Finally, we dropped all households with a head who is older than 60 because their present labor income will generally not reflect their pre-retirement income (n = 1050). In Japan, most people choose to retire around 60, while some of them stay in the labor market but are obliged to accept a much lower wage rate than before. Since our main objective is to investigate whether people save for life cycle or precautionary motives, we focus on the behavior of working households.

4. Empirical results 4.1. First step: income function estimation We take the predicted labor income of the household head from his income function as a proxy of his permanent income. This income function uses such explanatory variables as age, education, health condition, and place of residence (Table 1). The income profiles of full-time salaried workers are estimated separately from those of agricultural, forestry, fisheries, and self-employed group because generally only salaried workers will face an age-based reverse U earnings profile where earnings keep rising until one’s 50s and decline thereafter. As a result, we included the square of the individual’s age as an explanatory 8 A more detailed description of the data can be found in Appendix A.

202

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

variable in the earnings profile of full-time salaried workers. Moreover, inasmuch as the wage of salaried workers will be influenced by the size of their employers, firm size dummies are also employed when estimating the earnings of salaried workers. The estimation results in Table 1 (case a) indicate that all explanatory variables except the health dummy are statistically significant in explaining the labor income of salaried workers. Furthermore, age and firm size have a positive effect on the labor income of salaried workers, as we would have expected. The negative coefficient of AGE ∗ AGE affirms our reverse U earnings profile hypothesis. On the other hand, although age still shows a significantly positive effect on the labor income of agricultural, forestry, fisheries, and self-employed households, few other variables are statistically significant (case d). The adjusted R 2 is as low as 0.1037, implying that only about 10% of the variation in the labor income of agricultural, forestry, fisheries, and self-employed households can be correctly explained by our model. Regarding the cohort effect on labor income, we use the method of King and DicksMireaux (1982) and create a cohort variable c(A) that reflects the fact that, ceteris paribus, younger cohorts are better off than older cohorts because of technical progress and capital accumulation. The cohort effect was captured by a piecewise linear function for c(A) that has a zero value at the standard age (40). It was assumed that one-half of the growth rate of real earnings was accounted for by the cohort effect and that the other half was due to other factors such as improvements in educational attainment and changes in occupational structure. Hence, the average annual growth rate of real earnings attributable to the cohort effect is 3.995% in 1956–1965, 6.565% in 1966–1975, 5.535% in 1976–1985 and 1.495% in 1986–1996.9 The cohort effect variable c(A) is defined as follows: Age: Age
30: 30 < Age
40: 40 < Age
50: Age > 50:

c(A)  − 0.05535 + 0.01495(30 − Age) −0.05535(40 − Age) 0.06565(Age − 40) 0.06565 + 0.3995(Age − 40)

When the age variables are replaced by the cohort variable c(A) in the estimation of salaried workers’ income (case b), we obtained a significant coefficient for c(A) but a much lower adjusted R 2 . Employing both c(A) and age variables in the estimation (case c) leads to a totally insignificant estimated coefficient of c(A). Thus, the cohort effect on salaried workers’ labor income is apparently very minor or different to distinguish from age effects. In light of these estimation results, we employ predicted labor income without cohort effect10 (case a) as a proxy for the permanent income of salaried workers in the second step consumption function estimation. For agricultural, forestry, fisheries, and selfemployed households, we utilize their present labor income directly in the consumption 9 Data on the growth rates of earnings were obtained from “Long-term Economic Statistics 2001” (Cabinet Office, Government of Japan). 10 The cohort effect should be negligible in Japan for two reasons: first, as the cohort effect affects not only income but also consumption, the consequences of not adjusting for it in the first step estimation will not be so serious. Second, using historical data to estimate the cohort effect will inevitably lead to a substantial bias, especially if the high growth period is included.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

203

function estimation because the predictive power of the estimated income function is so low. 4.2. Second step: consumption function estimation It is still not known whether the empirical failure of the LC/PIH is due to earnings uncertainty or to other factors such as liquidity constraints11 or bequest motives, and thus testing for the existence and magnitude of the precautionary saving motive is of crucial importance. This paper tests for the validity of the precautionary saving model by estimating Eqs. (2.1) and (2.2). As we discussed before, the precautionary saving model requires the estimated coefficient of V (Y ) to be negative. We estimated Eq. (2.1) using weighted least squares (WLS) with the number of observations in each group as the weight because the use of grouped data will lead to heteroskedasticity if the variance of the error term is inversely proportional to group size. Table 3.1 shows the WLS estimates of Eq. (2.1). The consumption functions of all households, older households (households with a head aged 50 or older), and younger households (household with a head aged 49 or younger) are estimated separately. Although the magnitude of each coefficient varies from case to case, we find that the coefficients of all of the explanatory variables are statistically significant and the signs of the estimated coefficients are consistent with the predictions of the precautionary saving model with the exception of the coefficient of SSW. Dropping V (Y ) from Eq. (2.1) lowers its adjusted R 2 significantly from 0.7575 to 0.7506 (t value = 8). This suggests that the effect of the precautionary motive on household consumption/saving is quite significant but that its magnitude is not as large as we would have expected. In sum, the estimation results based on grouped data suggest that although earnings uncertainty has a significant negative Table 3.1 Parameter estimates of consumption function (Eq. (2.1)) based on grouped data (a) All Y ASSET SSW V (Y ) Constant Sample size/ Adj. R 2

(23.191)***

0.1748467 0.0164019 (28.848)*** −0.0117343 (−9.659)*** −0.0000330 (−8.351)*** 229.387 (81.024)*** 56 groups (2423 households)/ 0.7579

(b) Young (13.184)***

0.1623995 0.0208917 (21.35)*** −0.0066212 (−3.347)*** −0.0000337 (−6.847)*** 212.270 (56.863)*** 40 groups (1620 households)/ 0.6796

(c) Old 0.1610979 (16.965)*** 0.0074928 (8.692)*** −0.0042178 (−2.196)** 0.0001215 (9.462)*** 236.375 (56.839)*** 29 groups (803 households/ 0.8689

Notes. (1) The dependent variable is average yearly consumption within each group. (2) In each column, the first number is the estimated WLS coefficient (weight = number of households within each group) and the second is the t value calculated using robust standard errors. (3) All of the explanatory variables are the average values within each group. (4) “Young” refers to households whose heads are aged 49 or younger, “Old” refers to households whose heads are aged 50 to 60, and “All” refers to all households whose heads are aged 60 or younger. **,*** Significant at the 5 and 1% levels, respectively. 11 For Japanese households, liquidity constraint may not be a serious problem because most households reported a positive accumulated amount of financial assets.

204

Salaried worker households (a) All

(b) Young

PI 0.1291526 (3.39)*** 0.0933240 (2.191)** ASSET 0.0044480 (2.994)*** 0.0067062 (2.97)*** SSW 0.0125841 (5.168)*** 0.0131956 (4.539)*** V (Y ) −0.0000806 (−1.909)** −0.0000947 (−2.238)** (9.581)*** Constant 1.983 (9.911)*** 2.116 Sample size/ 497/ 368/ 0.1297 0.1279 Adj. R 2

Agricultural, forestry, fisheries, and self-employed households (c) Old 0.2522090 0.0021888 0.0101948 0.0003785 1.251 129/ 0.1225

(a) All

(b) Young

(c) Old

(2.641)*** 0.0858008 (1.862)* 0.215043 (3.107)*** −0.0012 (−0.036) (1.399) 0.0030034 (1.611) −0.00304 (−1.459) 0.0058 (5.8)*** ** *** *** (2.161) 0.0271093 (3.765) 0.026743 (3.659) 0.0269 (2.297)** (1.01) −0.000301 (−2.693)** −0.00021 (−0.69) −0.0003 (−1.408) (2.231)** 2.658 (8.875)*** 2.310 (5.354)*** 2.742 (4.696)*** 140/ 78/ 62/ 0.2562 0.343 0.3126

Notes. (1) The dependent variable is the yearly consumption. (2) In each column, the first number is the estimated FGLS coefficient (weight = least square residual) and the second is the t value calculated using robust standard errors. (3) Expected permanent income (PI) is employed in estimating the consumption function of salaried worker households, while for agricultural, forestry, fisheries, and self-employed households, current labor income (Y t) is employed instead of PI. (4) t values for the estimates of salaried worker households are calculated using the adjusted standard errors. *,**,*** Significant at the 10, 5, and 1% levels, respectively.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

Table 3.2 Parameter estimates of the consumption function (Eq. (2.2)) based on individual data

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

205

Table 3.3 Parameter estimates of consumption function based on individual data and group fixed effects Salaried worker households All PI ASSET SSW Sample size/Adj. R 2

0.2820926 0.0033452 0.0121928 497/0.1852

Young (3.66)*** (2.39)** (4.76)***

0.2931546 0.0062614 0.0136158 368/0.2033

Old (3.6)*** (3.12)*** (4.59)***

0.2974846 0.0021974 0.009101 129/0.0766

(2.74)** (1.08) (1.73)*

Notes. (1) The dependent variable is the yearly household consumption. The explanatory variables are PI, ASSET, SSW and group dummies. (2) In each column, the first number is the estimated FGLS coefficient (weight = least squares residual) and the second is the t value calculated using robust standard errors. (3) The estimated coefficients of the group dummies and intercepts are omitted. *,**,*** Significant at the 10, 5, and 1% levels, respectively.

impact on the consumption of Japanese households, the magnitude of its effect is not so large. Table 3.2 presents the estimation results of the precautionary saving model based on household-level data (Eq. (2.2)). In practice, we estimate Eq. (2.2) by profession and the age of the household heads. The consumption function of agricultural,12 forestry, fisheries, and self-employed households (n = 529) is estimated separately from that of salaried worker households (n = 1864) because their consumption behavior is likely to be different. We also estimate the consumption function separately for young and old households to test the buffer stock saving hypothesis of Carroll and Summers (1991), who suggest that the saving behavior of households younger than 45 or 50 is motivated largely by precautionary and other motives rather than by the retirement motive. Due to the tendency of heteroskedasticity to be present in consumption functions based on householdlevel data, we chose to estimate Eq. (2.2) using Feasible Generalized Least Squares (FGLS) with the least squares residual as the weight. However, there still exist two econometric problems with the estimation results for Eq. (2.2). First, as we include the within-group variance of labor income in the estimations, there will be unobserved group specific heterogeneity and Eq. (2.2) might suffer from omitted variable bias. Although this bias exists to some extent, we found that it is not very serious because when we replaced V (Y ) with group dummies (see Table 3.3 for details), we found that the estimated coefficients of the other explanatory variables do not change substantially.13 An alternative approach is to use Ramsey’s RESET(2) test, which is a very useful test because it combines relatively good power with simplicity of computation. This procedure simply involves augmenting the original regressors with the square of the predicted value of the dependent variable 12 Agricultural, forestry and fisheries households (n = 43) refer to those that are earning more than half of their overall incomes by working on agricultural, forestry or fisheries business. 13 For young households, however, the coefficient on PI has tripled with the fixed effects. One of the possible reasons could be that PI and V (Y ) are positively correlated for young households. Hence, replacing V (Y ) with group dummies in the consumption function could result in a somewhat larger effect of PI. Nevertheless, the correlation coefficient between PI and V (Y ) is quite small (r = 0.1077). So there could be some other underlying reasons.

206

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

in the original model and then testing the significance of the estimated coefficient of the additional variable (Godfrey et al., 1987). The test of our linear consumption model is based on the auxiliary regression:
2  +ε p + a2 ASSET + a3 SSW + a4 V (Y ) + b C C = a0 + a1 Y (3)  is the predicted value of C in Eq. (2). In the case of Eq. (3), the RESET(2) test in which C tests the null hypothesis that b (which approximates a t distribution if the disturbances are approximately normal) is zero. The RESET(2) test results show that the null hypothesis of b = 0 should be accepted (t = −1.51) and that the specification of our model should be regarded as appropriate.14 In other words, the omitted variable bias about which we were worried is not very serious. Another problem relates to the standard errors of the second step estimates for salaried worker households. This two-step (2S) procedure produces consistent but inefficient estimates, and the standard errors from this two-step procedure will be inconsistent and generally understated. The primary reason for this result is that the presence of a generated regressor causes the true disturbances to be serially correlated and heteroskedastic (McKenzie and McAleer, 1997). Two approaches15 have been developed for deriving the properties of the efficiency and consistency of the standard errors of 2S estimators of linear models. The first method (Pagan, 1984; Newey, 1984) is based on the derivatives of the likelihood function and the information matrix. The second method employs information on the sampling distribution of the first step estimates to adjust the estimated covariance of the second step equation (Murphy and Topel, 1985; McAleer and McKenzie, 1991). In this paper, we employ the second approach since it illustrates clearly and more intuitively. More specifically, we construct consistent estimates of the standard errors using the following equation as an expression for the covariance matrix: V2∗ = V2 + V2 [CV1 C  − RV1 C  − CV1 R  ]V2 , where V1 and V2 are the asymmetric variance covariance matrix of the first and second step estimations, respectively. The R and C matrices are obtained by summing the individual observations on the cross products of the derivatives about a vector of estimates of the log likelihood function. Table 3.216 shows that the consumption of agricultural, forestry, fisheries, and selfemployed households is significantly discouraged by earnings uncertainty. However, when the consumption function is estimated separately by age group, the estimated coefficient of V (Y ) is negative but totally insignificant. In addition, labor income, assets, and social security wealth (SSW) all have a positive impact on the consumption of agricultural, 14 The RESET(3) test yielded similar results to those of RESET(2). 15 Another alternative might be to use the Newey and West approach to construct a heteroscedasticity and

autocorrelation consistent covariance matrix, but in this case the standard errors are quite often downward biased (McKenzie and McAleer, 1997). 16 140 of 529 (26.5%) samples of agricultural, forestry, fisheries, and self-employed households have been employed in the estimations of Eq. (2.2). 389 samples lost for missing values about the necessary variables. In general, the 140 employed samples averagely have a slightly higher VYD and household income than the lost samples.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

207

forestry, fisheries, and self-employed households, as expected, but only labor income and SSW are statistically significant. Table 3.2 also shows the consumption function estimates for salaried worker households17. According to the variance–covariance matrix adjusted estimation results,18 V (Y ) has a significantly negative effect on the consumption of young households only. This finding combined with the finding about older households based on grouped data is in accordance with the hypothesis of Carroll and Summers (1991) that young households are more likely to save for precautionary purposes and that old households save primarily for retirement. Dropping V (Y ) from the explanatory variables lowers the adjusted R 2 from 0.1297 to 0.1266. This confirms what we found from our estimates based on grouped data that the impact of precautionary motives on household saving might be significant but its magnitude is somewhat low. The estimated coefficients of the other main variables such as expected permanent income, assets, and SSW are positive and statistically significant in almost all cases. Table 4 presents the estimation results of the consumption function (Eq. (2.2)) when a liquidity constraints variable19 and various household attributes are added. Profession and marital status of the household head, homeownership, and the number of dependent children are characteristics that could possibly affect the consumption of households. The Table 4 Parameter estimates of consumption function with household characteristics Individual data (Eq. (2.2))–FGLS (e) All PI −0.0660491 (−1.64)* ASSET 0.0018784 (1.12) SSW 0.0123301 (5.08)*** V (Y ) −0.0001548 (−2.75)*** Liquidity constraint 0.1356063 (5.27)*** PROd (1 if salaried workers) 0.0676989 (0.47) Homeownership dummy −0.0296761 (−0.28) Marriage dummy 0.3481621 (1.54) Number of dependent children 14.41229 (3.59)*** Constant 2.301105 (7.98)*** 2 Sample size/Adj. R 533/0.2263

(f) Young −0.1060052 −0.000231 0.0136307 −0.0001499 0.1951791 −0.0307163 0.0969049 0.348924

(−2.26)** (−0.11) (4.85)*** (−2.35)** (5.65)*** (−0.18) (0.72) (1.33)

13.10983 (2.7)*** 2.244711 (6.73)*** 372/0.265

(g) Old 0.0185892 (0.2) 0.0031833 (1.8)* 0.0096283 (2.19)** −0.0001427 (−1.13) 0.0758412 (2.29)** 0.1332352 (0.43) −0.2638334 (−1.45) 0.0719656 (0.21) 10.20279 (1.17) 2.591244 (4.46)*** 161/0.1979

Notes. (1) The dependent variable is the yearly household consumption. (2) Liquidity constraint is defined as ratio of household gross income to wealth. *,**,*** Significant at the 10, 5, and 1% levels, respectively. 17 497 or 26.7% of 1864 observations on salaried worker households were employed in the estimation of Eq. (2.2). 1367 observations had to be dropped because they had missing values for one or more of the variable needed for our analysis. In general, the distributions of the major household variables were very similar to those for the full sample. 18 Theoretically the variance-covariance adjusted standard errors will be larger than the unadjusted ones. Our results confirm this prediction but the gap is quite small. For example, the unadjusted standard error of V (Y ) is −0.0000421, while but the adjusted one is −0.00004221 (case a, Table 3.2). 19 The liquidity constraints variable is defined as the ratio of household gross income to wealth.

208

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

estimated coefficients of the main variables such as income variance and social security wealth are all broadly consistent with the earlier results shown in Tables 3.1 and 3.2. In particular, the estimation results (see Table 4 for details) indicated that the precautionary saving hypothesis is supported even after controlling for liquidity constraints. In addition, the consumption of younger households seems to be negatively related to liquidity constraint, with household with a higher income/wealth ratio having a higher consumption level. Nevertheless, consumption of older households is not significantly affected by the liquidity constraint. Besides, number of dependent children has a positive impact on household consumption. Finally, we estimate the magnitude of precautionary saving by calculating what our estimation results imply about the share of precautionary saving in total household saving. We can calculate the share of precautionary saving (λ) from a4 , the estimated coefficient of V (Y ), as follows:  n  n n n     λ= P Si Si = (Yi − Ci ), n = 1, 2, . . . , N. −a4 Vi (Y ) i=1

i=1

i=1

i=1

λ equals 0.05557 in case (a) of Eq. (2.2) (a4 = −0.0000806), which means that precautionary saving comprises about 5.557% of the total saving of salaried worker households. On the other hand, for agricultural, forestry, fisheries, and self-employed households, the estimated magnitude of precautionary saving is as high as 64.3% of their total savings.20 This finding confirms the finding of Ogawa (1991) that agricultural, forestry, fisheries, and self-employed households save much more for precautionary purposes than salaried workers.

5. Conclusions This paper uses household-level data from a Japanese Government survey to analyze whether precautionary saving is a potential source of cross-household differences in consumption and saving and to test the hypothesis that households with greater earnings uncertainty will have systematically higher saving rates. Our estimation results show that the precautionary saving model should be fully accepted and that income uncertainty has a statistically significant impact on Japanese household saving. Although there still remains some uncertainty about the magnitude of the impact of earnings uncertainty on household saving, we find that precautionary saving may account for 5.557% and 64.3% of the total saving of salaried worker households and 20 The 64.3% share of precautionary saving for agricultural, forestry, fisheries and self-employed households should be interpreted cautiously for at least two reasons: firstly, it is well known that these households are very likely to under-report their income for tax avoidance and other reasons (Deaton, 1997). Besides, these households are likely to have a significant amount of self-consumption, which could lead to some unintended under-reporting of income values. The intended or unintended under-reporting of household income could cause an upward bias in VYD and hence in the proportion of precautionary saving; secondly, a large proportion of these households have been dropped in the estimations, which could somehow lead to an upward bias in VYD and hence in the share of precautionary saving.

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

209

agricultural, forestry, fisheries, and self-employed households, respectively. This finding is broadly consistent with the findings of Ginama (1988) and Ogawa (1991). The 5.557% figure may seem low, but it must the borne in mind that, earnings uncertainty is only one of numerous uncertainties faced by households and that the total amount of precautionary saving attributable to all sources of uncertainty could well be much higher. Indeed, survey respondents in Japan and elsewhere consistently indicate that saving for illness and other emergencies is one of their main reasons for saving. For example, according to the 1994 Survey on the Financial Asset Choice of Households, conducted by the Japanese Ministry of Posts and Telecommunications, well over half of all Japanese households report that they are saving for precautionary motives and precautionary saving comprises a full 56% of their total saving (see Horioka and Watanabe, 1997). Similarly, according to the 1996 Comparative Survey of Savings in Japan and the United States, also conducted by the Japanese Ministry of Posts and Telecommunications, precautionary saving comprises a full 62% and 31% of total household saving in the United States and Japan, respectively (see Horioka et al., 2000). Finally, the prediction of Carroll and Summers (1991) that precautionary saving is more important for younger households than for older households is also supported by the results based on household-level data. This result has a very meaningful policy implication in Japan’s salaried worker household case: tax cuts or public spending that alleviate the risk borne by younger cohorts will be more effective in stimulating overall household consumption than the same policies that target older cohorts or the population as a whole. Turning to policy implications, our finding that precautionary saving arising from earnings uncertainty is of some importance implies that policies aimed at alleviating earnings uncertainty (such as reform of the unemployment insurance and employment systems) will reduce saving and increase consumption to some degree, thereby helping to spur economic recovery. Turning finally to directions for further research, we have focused on precautionary saving arising from earnings uncertainty in this paper, but the evidence suggests that precautionary saving arising from other uncertainties is of considerable importance and thus warrants careful analysis. Another important task is to verify the empirical results of this paper from other perspectives. For example, public policies such as the introduction of long-term insurance and the cut in unemployment compensation benefits could have a direct or indirect impact on the level of household income uncertainty. It would be interesting to investigate the effect of these so-called “natural experiments” on household saving.

Acknowledgments This paper was prepared under the supervision of Colin R. McKenzie (Professor, Osaka School of International Public Policy, Osaka University) and Charles Y. Horioka (Professor, Institute of Social and Economic Research, Osaka University, and former Special Guest Research Officer, Institute for Posts and Telecommunications Policy). I am greatly indebted to them for their kind advice, comments and encouragement throughout the process of writing this paper. I am also grateful to Yukinobu Kitamura, Wataru Suzuki,

210

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

members of Professors Horioka and McKenzie’s graduate seminars at Osaka University, and especially two anonymous referees and Editor-in-Chief Takeo Hoshi for their helpful comments. I appreciate the Institute for Posts and Telecommunications Policy of the Government of Japan for permitting me to use the micro data from the 1996 Survey on the Financial Asset Choice of Households. This paper is based on a chapter of my PhD dissertation (Zhou, 2001).

Appendix A. Data The Survey on the Financial Asset Choice of Households (SFACH) has been conducted every two years since 1988 by the Ministry of Public Management, Home Affairs, and Posts and Telecommunications (formerly the Ministry of Posts and Telecommunications) of the Government of Japan. The sample design is such that the sample is representative of the total Japanese population. The sample is selected on the basis of a multiple-stage stratified sampling procedure. In the first stage, the country is divided into 12 postal areas. In the second stage, cities and towns within each postal area are divided into five groups according to population. In the third stage, households are drawn by a random sampling procedure from the list of all resident households in a given city or town. A household is defined as a group that shares a common dwelling, is related by blood or marriage, and pools income. 6000 households (including single households) whose heads are older than 20 were selected from throughout Japan, and of these 6000 households, 3695 households responded to the questionnaire. The surveyors distributed and explained the questionnaires to subjects in person, and several days later, the surveyors visited the subjects again to collect the questionnaires. The 1996 SFACH collects very detailed data on household holdings of financial assets, real assets, and loans, home ownership, retirement plans, pension participation, bequests, etc. In addition, the 1996 SFACH provides some unique information that can seldom be found in other surveys. For example, it collects data on expected monthly living expenses after retirement and the percentage of living expenses after retirement that the respondent expects to be able to finance using public pension benefits, neither of which is collected by other surveys. Thus, the SFACH enables us to calculate the value of public pension wealth (SSW) much more easily and more accurately than in the case of other surveys (see Appendix B). Because of its sampling design, the distributions of the major household indexes in the 1996 SFACH are very similar to those of the Census of Population and other large-scale national surveys performed at about the same time in Japan. A detailed comparison of the major household indexes in the 1996 SFACH, the 1996 Family Income and Expenditure Survey (FIES), and the 1994 National Survey of Family Income and Expenditure (NSFIE) is presented in Table A. All of these data sets have a very similar household distribution with respect to the age of the head (44–45) and family size (3.5–3.7). Moreover, the average values of the annual labor income of the household head, annual consumption, the average propensity to save, and the distribution of households by place of residence in the SFACH are similar to those in the FIES, and net financial assets, the homeownership rate, etc., in the SFACH are very similar to those in the NSFIE. However, compared to the NSFIE,

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

211

Table A Descriptive statistics for worker households (1996) SFACH Yearly labor income of head (10,000 JPY) Yearly labor income of spouse (10,000 JPY) Yearly consumption (10,000 JPY) Age of head Real asset holdings (10,000 JPY) Net financial assets = Financial assets-liabilities (10,000 JPY) Social security wealth (10,000 JPY) Persons per household Average propensity to save (%) Proportion living in cities with a population of 50,000 or more (%) Homeownership rate (%) Number of observations (maximum)

589.8 105.1 358.4 44.1 1753 565.8 3208.1 3.7 32.9 69.9 62.2 2441

1996 FIES

1994 NSFIE

569.5 66.1 422.04 45.8 – – – 3.5 28 75.7 51.4 10,000

442.7 61.6 428.0 44.9 2935 613 – 3.7 19.6 86.5 69.2 7025

Sources: 1996 FIES (Family Income and Expenditure Survey), pp. 86–88, Statistics Bureau, Japan; 1994 NSFIE (National Survey of Family Income and Expenditure), pp. 48–50, 69, 184, Statistics Bureau, Japan.

the proportion of respondents residing in cities with a population of less than 50,000 is somewhat higher, their average propensity to save is higher, and their real asset holdings are lower in the SFACH. The relatively low value of real assets in the 1996 SFACH compared to that in the 1994 NSFIE may be due to the dramatic decline in land prices in the 1990s, the relatively low homeownership rate in the SFACH, and so on. The average propensity to save from the SFACH (32.90%) is much higher than that from the NSFIE (19.6%) because the SFACH figure refers to the entire year, whereas the NSFIE figure refers to the September–November period, which excludes periods when bonuses are paid and therefore exhibits a much lower saving rate than the year as a whole.

Appendix B. The estimation of social security wealth The 1996 SFACH contains data not only on expected living expenses per month during retirement (CRET ), but also on the proportion of living expenses during retirement the respondent plans to finance using social security (ρ). Therefore, as long as we can obtain a value for the retirement span (RETP), we can easily estimate the value of SSW as follows: SSW = (CRET 12)ρRETP

(B.1)

and RETP is estimated as: RETP = AGE + l exp(−RET ),

(B.2)

where RET is the planned retirement age and l exp is the life expectation of the household which was obtained from Abridged Life Table for Japan, 1999 (Ministry of Health, Labor and Welfare, Japan). The RETP of the spouse of the household head was calculated separately and the larger of the two values was employed as the actual value of RETP in Eq. (B.1), following Horioka and Okui (1999).

212

Y. Zhou / J. Japanese Int. Economies 17 (2003) 192–212

References Blanchard, O.J., Mankiw, N.G., 1988. Consumption: Beyond certainty equivalence. AEA Pap. Proc., May. Caballero, R.J., 1990. Consumption puzzles and precautionary savings. J. Monet. Econ. 25, 113–136. Carroll, C.D., 1992. The buffer stock theory of saving: Some macroeconomic evidence. Brookings Pap. Econ. Act. 2, 61–135. Carroll, C.D., Summers, L.H., 1991. Consumption growth parallels income growth: Some new evidence. In: Bernheim, B., Shoven, J.B. (Eds.), National Savings and Economic Performance. Chicago University Press. Dardanoni, V., 1991. Precautionary saving under income uncertainty: A cross-sectional analysis. Appl. Econ. 23, 153–160. Deaton, A., 1997. The Analysis of Household Surveys: A Microeconometric Approach to Development Policy. Johns Hopkins University Press. pp. 29–32. Dreze, J., Modigliani, F., 1972. Consumption under uncertainty. J. Econ. Theory 5, 308–335. Friedman, M., 1957. A Theory of the Consumption Function. Princeton, NJ. Godfrey, L.G., McAleer, M., McKenzie, C.R., 1987. Variable addition and Lagrange multiplier tests for linear and logarithmic regression models. Rev. Econ. Stat. 70 (3), 492–503. Guiso, L., Jappelli, T., Telizzese, D., 1992. Earning uncertainty and precautionary saving. J. Monet. Econ. 30, 307–337. Ginama, I., 1988. Estimation of precautionary saving in the U.S. and Japan. PhD Dissertation, Univ. of California, San Diego. Hahm, J., 1999. Consumption growth, income growth and earnings uncertainty: Simple cross-country evidence. Int. Econ. J. 13 (2). Horioka, C.Y., 1990. Why is Japan’s saving rate so high? A literature survey. J. Japan. Int. Econ. 4, 49–92. Horioka, C.Y., Watanabe, W., 1997. Why do people save? A micro-analysis of motives for household saving in Japan. Econ. J. 107 (442), 537–552. Horioka, C.Y., Fujisaki, H., Watanabe, W., Kouno, T., 2000. Are American more altruistic than the Japanese? A U.S.–Japan comparison of saving and bequest motives. Int. Econ. J. 14, 1–31. Horioka, C.Y., Okui, M., 1999. A U.S.–Japan comparison of the importance and determinants of retirement saving. Econ. Letters 65, 365–371. Kimball, M.S., 1990. Precautionary saving in the small and in the large. Econometrica 58, 53–73. King, M.A., Dicks-Mireaux, L., 1982. Asset holdings and the life-cycle. Econ. J. 92, 247–267. Leland, H.E., 1968. Saving and uncertainty: The precautionary demand for saving. Quart. J. Econ. 82, 465–473. McAleer, M., McKenzie, C.R., 1991. When are two step estimators efficient? Econometric Rev. 10, 235–252. McKenzie, C.R., McAleer, M., 1997. On efficient estimation and correct inference in models with generated regressors: A general approach. Japan. Econ. Rev. 48, 368–389. Murphy, K., Topel, R., 1985. Estimation and inference in two step econometric models. J. Bus. Econ. Statist. 3, 307–379. Newey, W.K., 1984. A method of moments interpretation of sequential estimators. Econ. Letters 14, 201–206. Ogawa, K., 1991. Income uncertainty and precautionary saving. Keizai Kenkyu 142 (2). In Japanese. Pagan, A.R., 1984. Econometric issues in the analysis of regressions with generated regressors. Int. Econ. Rev. 25, 221–247. Parker, J.A., 1999. The reaction of household consumption to predictable changes in social security taxes. The Amer. Econ. Rev. 89, 959–973. Sandmo, A., 1970. The effect of uncertainty on saving decision. Rev. Econ. Stud. 37, 353–360. Skinner, J., 1988. Risky income, life cycle consumption, and precautionary savings. J. Monet. Econ. 22, 237–256. Zeldes, S.P., 1989. Optical consumption with stochastic income: Deviations from certainty equivalence. Quart. J. Econ. 104, 275–298. Zhou, Y., 2001. Japanese household’s precautionary saving and public pension choice. PhD dissertation. Osaka School of International and Public Policy, Osaka Univ., Japan.