A. Channels of transmission for stock market wealth

Broadly speaking, there are five different transmission channels from changes in stock market wealth to changes in consumption⁵:

1. Realized wealth effect: If the value of consumers’ stock holdings increase and consumers realize their gains then consumption will increase. This result will be a direct effect as a consequence of higher current liquid assets.

2. Unrealized wealth effect: An increase in stock prices can also have an expectations effect where the value of stocks in pension accounts and other locked-in accounts increases. When these assets go up in value but the increase is not realized, it results in higher consumption today on the expectation that income and wealth will be higher in the future.

3. Liquidity constraints effect: An increase in stock market prices increases the value of a portfolio for an investor. Borrowing against the value of this portfolio in turn allows the consumer to increase consumption.

4. Stock option value effect: An increase in stock prices can lead to higher consumption for stock option owners as a result of an increase in the value of households’ stock options. Again, this increase in consumption may come independently of whether the gains are realized or unrealized.

These transmission channels are thus associated with a positive impact of changes in prices on consumption. It is worth noting the difference between the realized and unrealized wealth effect. As Poterba (2000) puts it: “it seems particularly likely that the marginal propensity to consume out of wealth gains in (locked) retirement accounts is lower than the propensity to consume out of directly held assets since the former are often thought of as ‘long term assets’” (see also Thaler (1990)).⁶ Furthermore, there appears to be a difference in how households (and firms) adjust their consumption to changes in different types of assets. For non-technology stocks Edison and Sløk (2001a and 2001b) find a stronger reaction in consumption to changes in stock prices in market-based financial systems (United States, Canada, and United Kingdom) than in bank-based financial systems such as continental Europe. For changes in technology stock prices, however, the consumption reaction is more similar, reflecting that the technology sector worldwide seems to function and carry out business in a more similar way. Finally, as a fifth transmission channel, consumption of households who do not participate in the stock market may be indirectly affected by changes in stock market prices. Such indirect effects between stock prices and consumption have for example been highlighted in Romer (1990).

B. Channels of transmission for housing wealth

There are also five different transmission channels from changes in housing prices to changes in consumption, but some of these channels are somewhat different from the channels for stock prices listed above:

1. Realized wealth effect: For consumers who are house owners, the increase in house prices leads to an increase in net wealth, which can raise consumption today. If house prices increase it is possible for consumers to take out equity in the form of refinancing or selling of the house. Such a realized gain must be expected to have a positive impact on private consumption.

2. Unrealized wealth effect: If house prices increase but households do not refinance or sell the house it may still have a positive impact on consumption due to the increase in the discounted value of wealth. Hence consumers can spend more today on the expectation that they are “richer” than they were before.

3. Budget constraint effect; For consumers who are house (or apartment) renters an increase in house prices has a negative impact on private consumption. As house prices go up the budget constraint becomes tighter for renters, which must be expected to result in lower private consumption. This channel works through a realized capital loss since the increase immediately leads to higher prices, which have to be paid by the renters.⁷

4. Liquidity constraints effect: A fourth factor that is also important for the consumption impact of house price changes is how well functioning the financial system is. If house prices change it may require access for consumers to credit markets in order to take loans against the increase in house prices. If credit is constrained or the financial system is not able to support such a wish for loans, households may experience that they cannot react accordingly to higher house prices.

5. Substitution effect: An increase in house prices may imply that households who are planning to buy a home may lower consumption when faced with increasing higher prices since they may increase down payments and future loans. This might force households to either buy a smaller house or to lower private consumption.

Both the realized and unrealized gains from increases in house prices must be expected to increase private consumption, but as with stocks, the marginal propensity to consume out of unrealized gains in housing wealth might be lower. But both the budget constraint and the substitution effect work in the opposite direction. The latter was neglected in the discussion on stock prices since the degree of divisibility is much higher for stocks and the decision to buy a house is driven by other motivations than the decision to invest in the stock market.

There are various other dimensions along which housing wealth differs from stock market wealth. For example, a similar argument to Poterba’s (2000) argument about the lower marginal propensity to consume out of unrealized wealth can be made for housing wealth since it is considered as long-term asset. Moreover, the two types of assets have different risk characteristics as for example mirrored in lower loan-to-value ratios for housing wealth and there might be higher costs imbedded to getting information on housing wealth than on stock market wealth.⁸

In sum, while the effects of changes in stock market prices seem to unambiguously point to an increase in consumption, the impact of changes in housing prices is ambiguous. Further, the discussion of the various channels of transmission from movements in both wealth components to consumption suggests that the financial system plays a central role in this process.

C. The role of the financial system

The discussion above suggests that the design of the financial system plays an important role in the transmission mechanisms from changes in the two wealth components to changes in consumption. Differences in the design of the financial systems are particularly pronounced between on the one side most of the continental. European countries which have bank-based financial systems and then on the other side the United States, Canada, and United Kingdom which have market-based financial systems. The design of the financial systems has important implications for the strength of the wealth effect. Of key importance for the strength are a) differences regarding sizes of financial markets and b) how widespread stock ownership is and c) the use of stock options as a means of payment by firms. These factors all influence the magnitude of the impact of an increase in wealth, and these different components will now be analyzed in turn.

First, the market-based system has resulted in different sizes of stock markets. In general, the size of stock markets is much higher in Anglo-Saxon countries than in the bank-based systems in Continental Europe (Figure 2). According to this criterion, the Netherlands, Japan, and Sweden might also be considered to belong to the group of market-based economies, Japan is a bit special, which has to do with significant amounts of crossholding and cross ownership of stocks. Moreover, Japan has experienced the well-known sharp increase in asset values during the eighties followed by a sharp decrease in values in the early nineties.

Second, the market-based system has also lead to a higher degree of stock market participation by households in countries with market-based financial systems than in countries with bank-based systems. Figure 4 describes some of these differences. According to this criterion, again the Netherlands and Sweden could be considered as market-based systems whereas Japan could not.

View Full Size

Percentage of adult population holding shares in 2000

Citation: IMF Working Papers 2002, 001; 10.5089/9781451841640.001.A001

Source: Proshare (2000)

• Download Figure
• Download figure as PowerPoint slide

Third, stock options as a means of payment is at least until now more widespread used in Anglo-Saxon countries than in countries with bank-based financial systems. Also, numbers for venture capital clearly suggest that there is a stronger market-based tradition for venture capital in the Anglo-Saxon countries.⁹

Borio (1996) provides an analysis of the credit characteristics of the non-government sector in fourteen industrialized countries. His criteria and the discussion above suggest assigning Australia, Canada, the Netherlands, Ireland, Sweden, the United Kingdom and the United States to an extended group of market-based economies whereas Belgium, Denmark, Finland, France, Germany, Italy, Japan, Norway, and Spain are treated as bank-based.¹⁰ The Japanese bubble economy, however, could add a lot of noise to our data and therefore the estimation was also carried out when Japan was excluded from the sample. These results do not differ from the estimates reported in section IV and are shown in tables A2 to A4 in the Appendix.

While the distinction of these two groups of countries according to stock market capitalization and the degree of participation in the stock market is more or less distinct, such an ordering matches countries with very different homeownership rates into one group. The comparison of homeownership rates shows that three of the countries in the group of market-based economies (Ireland, the United States and the United Kingdom) have fairly high homeownership rates, whereas two others (Sweden and the Netherlands) do not (Table 1). Such an alternative grouping of the data will be considered in section IV.

Table 1:

Homeownership in different countries

Source: Hypostat (2001) and Fannie Mac Foundation (2001). Data are from the most recent year (and all data are from the 1990s).

Table 1:

Homeownership in different countries

Ireland	80.0%
Spain	78.0%
Italy	78.0%
United States	67.5%
United Kingdom	67.0%
Belgium	65.0%
EU Average	61.0%
Finland	60.8%
Sweden	60.0%
Austria	55.0%
France	54.0%
Denmark	52.3%
Netherlands	52.0%
Germany	41.0%

Source: Hypostat (2001) and Fannie Mac Foundation (2001). Data are from the most recent year (and all data are from the 1990s).

View Table

Table 1:

Homeownership in different countries

Ireland	80.0%
Spain	78.0%
Italy	78.0%
United States	67.5%
United Kingdom	67.0%
Belgium	65.0%
EU Average	61.0%
Finland	60.8%
Sweden	60.0%
Austria	55.0%
France	54.0%
Denmark	52.3%
Netherlands	52.0%
Germany	41.0%

Source: Hypostat (2001) and Fannie Mac Foundation (2001). Data are from the most recent year (and all data are from the 1990s).

Overall, these three factors support the notion that changes in stock market wealth have a larger impact on consumption in countries with market-based financial systems. At the same time, the responsiveness of consumption to changes in stock market prices must be expected to have increased substantially over time. For housing wealth, however, the impact is more uncertain. In particular, as discussed above, there is no a priori reason to expect a positive effect of changes in housing prices on consumption. But due to the deregulation of financial markets across countries and a corresponding increased information level and transparency in financial markets, it must be expected that the positive influences of housing prices on consumption have become more important over time. This effect is likely to be more pronounced in countries where households have easier access to their housing wealth, e.g. refinancing of loans. In addition, the analysis of the differences between the two wealth components suggests that consumption reacts stronger to changes in stock market wealth than to changes in housing wealth.

Based on the considerations above, Table 2 summarizes the main hypotheses to be tested in the empirical section.

Table 2:

Summary of main hypotheses

Table 2:

Summary of main hypotheses

Hypothesis 1	Positive effect of stock market prices on consumption
Hypothesis 2	Stock market effects are higher in countries with market-based financial systems
Hypothesis 3	The impact of housing prices on consumption is ambiguous
Hypothesis 4	Stock market effects on consumption have increased over time
Hypothesis 5	The positive effect of changes in housing prices has become more important over time
Hypothesis 6	The sensitivity of consumption to changes in stock market wealth is higher than the sensitivity to changes in housing wealth

View Table

Table 2:

Summary of main hypotheses

Hypothesis 1	Positive effect of stock market prices on consumption
Hypothesis 2	Stock market effects are higher in countries with market-based financial systems
Hypothesis 3	The impact of housing prices on consumption is ambiguous
Hypothesis 4	Stock market effects on consumption have increased over time
Hypothesis 5	The positive effect of changes in housing prices has become more important over time
Hypothesis 6	The sensitivity of consumption to changes in stock market wealth is higher than the sensitivity to changes in housing wealth

A. Data

Below, equation (3) will be estimated for the entire sample as well as for groups of countries according to the discussion in Section II. Data covers 16 OECD countries. Specifically, Belgium, Denmark, Finland, France, Germany, Italy, Japan, Norway and Spain will be included into the group of economies with bank-based financial systems while Australia, Canada, Ireland, the Netherlands, Sweden, the United Kingdom and the United States are treated as market-based economies. Data availability of the unbalanced panel is shown in Table A1 of the appendix.

Given the broad coverage of the study there are certain data limitations. Therefore the approach adopted below deviates a bit from the simple theoretical model formulated above. First, stock market and housing prices are here used as proxy variables for the wealth components. Note, however, that a high correlation traditionally has been found between stock market prices and wealth measures as documented in Lettau and Ludvigson (2001) and Deutsche Bank (2001). The (in)direct impact of stock market prices on aggregate consumption has for example been investigated in the studies by Romer (1990) and more recently in Poterba and Samwick (1995). The role of housing prices on consumption is the focus in, among others, Miles (1992), Miles (1995, chapter 4) and more recently Brady et al. (2000) and Girouard and Blöndal (2001).

To underscore the validity of using price data as proxy variables, the analysis is extended in section IV. F by using market capitalization as a more direct measure of stock market wealth. But even with the use of market capitalization data as a proxy for stock market wealth of households, one problem remains: international capital mobility. However, still a high proportion of stocks within each country are held by domestic residents and therefore the use of such variables as proxies for domestic stock market wealth seems to be a valid approximation.

Second, focus is on total aggregate consumption and there is no distinction between non–durable and durable consumption. Conventional theories on consumption apply to the flow of consumption. Since durable consumption can be thought of as a replacement and addition to a capital stock, the conventional approach is to only use non-durable consumption in wealth effect studies.¹⁷ However, and as pointed out in Mehra (2001) total consumption is the parameter of interest when studying movements in stock market prices. Particularly stock market crashes are more likely to lead to a postponement of durable consumption while the reduction of nondurable consumption might be of minor importance (see e.g. Romer (1990)). Regarding the role of housing prices on aggregate consumption, durable consumption goods are among the major entities on which resources raised by mortgage refinancing are spend as Brady, Canner and Maki (2000) show. One shortcoming with using total consumption is that it also includes expenditures on housing services (see section II). Besides, it might be insightful to nevertheless distinguish between the two components of consumption but for reasons of data availability this is difficult for a panel of countries.

Third, total disposable income is used and not only labor income as would be suggested by the traditional permanent income hypothesis. Data availability constrains us to do so. But also economically it is more sensible to use total income rather than labor income. This is suggested by an extended view of the life-cycle theory as suggested by Attanasio (1999), and also sensible if households are on average more myopic than the life-cycle theory of consumption would suggests (see Campbell and Mankiw (1991) and more recently Mankiw (2000)).

Data for consumption and disposable household income was taken from the OECD Analytical Database (OECD, 2001b). Data on stock market price indices are taken from International Financial Statistics (IMF, 2001b), which provides us with a relatively broad coverage and enough time series observations. Data on housing price indices are taken from the Bank for International Settlement’s house price database (BIS, 2001).¹⁸ The data on housing prices are in annual frequency, and in order to interpolate the data linear interpolation was applied. Below this interpolation method is tested in order to check the robustness of the estimated coefficients, and it turns out that the interpolation method is in general not important for the main results found.

All variables are in local currencies and deflated by the consumer price index taken from the OECD Analytical Database. Consumption and income are expressed in per capita units using United Nations population data (UN, 2000), which are linearly interpolated between annual observations. Logs have been taken of all variables and hence the estimates reported below are the estimated elasticities of consumption in changes of the right hand side variables. We therefore control for the different size of movements in stock market and house prices across countries.

As Engle and Granger (1987) have pointed out, the long-run elasticities in equation (3) cannot be consistently estimated if all the single variables have unit roots unless the variables in the long-run relationship are cointegrated. Therefore, one has to examine the statistical properties of the data and test whether a cointegrating equilibrium relationship between consumption, income and the two price indices exists. Recently, tests for unit roots of individual series and cointegrating relationships between series have been developed for panel data.¹⁹

B. Unit root tests

Among the various tests proposed in the literature, the Im, Pesaran and Shin (1997) (IPS) panel unit root test is suitable here.²⁰ The IPS t-bar test is based on an average of individual country augmented Dickey Fuller (ADF) tests while allowing for heterogeneous coefficients under the alternative hypothesis and different serial correlation patterns across groups. Under the null hypothesis all groups exhibit a unit root while under the alternative this is not the case for some i,²¹ A more detailed discussion of the test can be found in Baltagi and Kao (2000).

More specifically, the following model is tested for all variables.

Under the null hypothesis the autocorrelation coefficient ρ_i equals one for all i while under the alternative it docs not for at least one i. The test statistic is then computed as the average of the individual ADF statistics as

Im, Pesaran and Shin (1997) show that this test statistic converges to a standard normal distribution. Under the alternative hypothesis, the IPS panel t-bar test diverges to minus infinity and therefore the left tail of the standard normal distribution is used to reject the null hypothesis. Table 3 summarizes the results for the unit root tests of the four variables where the lag lengthρ_i in equation (5) is chosen by the Schwartz Bayesian criterion (SBC) with a maximum number of four lags.²²

Table 3.

IPS (1997) panel unit root tests

Note: *(**) indicate significance at the 10 (5) percent level. Test results are for the sample period of 1960-2000.

Table 3.

IPS (1997) panel unit root tests

Variable	Test Statistic
Log of private consumption (per capita)	-1.59983
Log of household disposable income (per capita)	3.36438
Log of IFSS stock market price indices	0.43943
Log of housing price indices	-2.23799**

Note: *(**) indicate significance at the 10 (5) percent level. Test results are for the sample period of 1960-2000.

View Table

Table 3.

IPS (1997) panel unit root tests

Variable	Test Statistic
Log of private consumption (per capita)	-1.59983
Log of household disposable income (per capita)	3.36438
Log of IFSS stock market price indices	0.43943
Log of housing price indices	-2.23799**

Note: *(**) indicate significance at the 10 (5) percent level. Test results are for the sample period of 1960-2000.

These results confirm that the null of a panel unit root is not rejected for most of the series. At a five percent level of significance it is rejected for the housing price data. Note that this might result from the linear interpolation of annual housing price data. Not surprisingly, the test statistic is -0.75741 when a time trend is included in equation (5).

C. Cointegration tests

Pedroni’s (1999) tests for cointegration are used to test for the null of no cointegration in the panel of 16 OECD countries.²³ Pedroni’s tests allow for a considerable degree of heterogeneity between groups with regard to the intercept, the error structure, and the cointegrating relationship. Specifically, Pedroni’s tests are residual-based tests on a cointegrating regression of the following type:

The seven test statistics presented in Pedroni (1999) can be grouped into two types of statistics. The first type of statistics is based on pooling along the within-dimension of the panel; the second is based on pooling along the between-dimension (see Pedroni, 1999, p. 657). The former is constructed by first summing the numerator and the denominator of the statistics over theN groups dimension separately and then dividing while the later are average statistics in that they are constructed by first dividing the numerator and denominator and then summing over theN group dimension. This feature results in a more flexible correlation pattern of the residuals from equation (7) acrossi and as discussed in Baltagi and Kao (2000) allows for an easier interpretation of the statistics if the null is rejected for the second type of statistics.²⁴

Therefore, Table 4 only summarizes results for the second type of Pedroni panel cointegration tests for alternative choices of cointegratingx variables.²⁵ Under the null hypothesis all statistics asymptotically converge to a standard normal distribution. Under the alternative hypothesis, the statistics diverge to negative infinity. Therefore, the left tail of the normal distribution is used for a rejection of the null.

Table 4:

Pedroni (1999) panel cointegration tests

Note: *(**) indicate significance at the 10 (5) percent level. Test results are for the sample period of 1960-2000.

Table 4:

Pedroni (1999) panel cointegration tests

Regression Specification	Group ρ-Statistic	Group t-Statistic (non-parametric)	Group adf-Statistic
y: Income (Yd) x1: Stock Market Price Index (SMPl) x2: Housing Price Index (HPI)	3.58687	2.28910	-0.23974
y: Consumption (C) X1: Income (Yd) x2: Stock Market Price Index (SMPI) x3: Housing Price Index(HPI)	-4.21876**	-5.24211**	-4.63842**

Note: *(**) indicate significance at the 10 (5) percent level. Test results are for the sample period of 1960-2000.

View Table

Table 4:

Pedroni (1999) panel cointegration tests

Regression Specification	Group ρ-Statistic	Group t-Statistic (non-parametric)	Group adf-Statistic
y: Income (Yd) x1: Stock Market Price Index (SMPl) x2: Housing Price Index (HPI)	3.58687	2.28910	-0.23974
y: Consumption (C) X1: Income (Yd) x2: Stock Market Price Index (SMPI) x3: Housing Price Index(HPI)	-4.21876**	-5.24211**	-4.63842**

Note: *(**) indicate significance at the 10 (5) percent level. Test results are for the sample period of 1960-2000.

When estimating the specification in equation (7) it is first investigated whether a cointegrating relationship already exists among the x’s on average across countries and then the specification of interest is tested. The results reported in Table 4 confirm a cointegrating relationship between the variables in the specification of interest while the cointegrating relationship between the independent variables is rejected.

D. Estimating consumption equations

The investigation of the data properties above imply that estimation of equation (3) with variables expressed in log levels provides reliable inferences about the long and short term influences of the stock market and housing price indices and income on consumption. The considerations in Section II suggest that the impact of stock market and housing prices on consumption has changed over time. To test this hypothesis, equation (3) was estimated first for the sub-period 1985Q1 to 2000Q4.²⁶ Regression results for the combined sample and for the two groups, market-based and bank-based economies are summarized in Table 5. In the table C is consumption, Y^d is disposable income, SMPI is the stock market price index, and HPI is the house price index. The Schwartz criterion is applied for selection of the lag length for each individual country with a maximum number of four lags is allowed. Given the different lag lengths applied, the short run coefficient estimates are not representative for all countries.

Table 5:

Estimating consumption functions (1985-2000)

Note: See Table A1 for a description of the variables. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. *(**) indicate significance at the 10 (5) percent level.

Table 5:

Estimating consumption functions (1985-2000)

Variables	All Countries	Bank-based Economies	Market-based Economies
	Long run coefficients
Yd	0.7031** (0.0227)	0.6444** (0.0471)	0.7056** (0.0275)
SMPI	0.0802** (0.0059)	0.0305** (0.0077)	0.0815** (0.0072)
HPI	0.0362** (0.0122)	0.0154 (0.0186)	0.0403** (0.0215)
	Averages of heterogeneous short run coefficients
Adjustment coefficient	-0.096** (0.062)	-0.140* (0.123)	-0.128** (0.043)
d(C(-l))	0.006 (0.1)	-0.016 (0.173)	0.099 (0.109)
d(Yd)	0.276** (0.083)	0.346** (0.138)	0.205** (0.085)
d(Yd(-1))	-0.075 (0.080)	-0.109 (0.141)	-0.058 (0.090)
d(SMPl)	-0.003 (0.002)	-0.005 (0.004)	-0.003 (0.003)
d(SMPI(-l))	-0.002 (0.002)	0.000 (0.000)	-0.006 (0.006)
d(HPI)	0.162** (0.057)	0.142* (0.087)	0.180** (0.086)
d(HPI(-1))	-0.111** (0.057)	-0.084 (0.068)	-0.108 (0.108)
Intercept	0.063** (0.019)	0.209** (0.070)	0.029** (0.014)
Diagnostic Statistics
LR test (p-value)	245.47 (0.00)	149.93 (0.00)	83.39 (0.00)
Joint Hausman test (p-value)	5.62 (0.13)	4.09 (0.25)	7.67 (0.05)

Note: See Table A1 for a description of the variables. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. *(**) indicate significance at the 10 (5) percent level.

View Table

Table 5:

Estimating consumption functions (1985-2000)

Variables	All Countries	Bank-based Economies	Market-based Economies
	Long run coefficients
Yd	0.7031** (0.0227)	0.6444** (0.0471)	0.7056** (0.0275)
SMPI	0.0802** (0.0059)	0.0305** (0.0077)	0.0815** (0.0072)
HPI	0.0362** (0.0122)	0.0154 (0.0186)	0.0403** (0.0215)
	Averages of heterogeneous short run coefficients
Adjustment coefficient	-0.096** (0.062)	-0.140* (0.123)	-0.128** (0.043)
d(C(-l))	0.006 (0.1)	-0.016 (0.173)	0.099 (0.109)
d(Yd)	0.276** (0.083)	0.346** (0.138)	0.205** (0.085)
d(Yd(-1))	-0.075 (0.080)	-0.109 (0.141)	-0.058 (0.090)
d(SMPl)	-0.003 (0.002)	-0.005 (0.004)	-0.003 (0.003)
d(SMPI(-l))	-0.002 (0.002)	0.000 (0.000)	-0.006 (0.006)
d(HPI)	0.162** (0.057)	0.142* (0.087)	0.180** (0.086)
d(HPI(-1))	-0.111** (0.057)	-0.084 (0.068)	-0.108 (0.108)
Intercept	0.063** (0.019)	0.209** (0.070)	0.029** (0.014)
Diagnostic Statistics
LR test (p-value)	245.47 (0.00)	149.93 (0.00)	83.39 (0.00)
Joint Hausman test (p-value)	5.62 (0.13)	4.09 (0.25)	7.67 (0.05)

Note: See Table A1 for a description of the variables. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. *(**) indicate significance at the 10 (5) percent level.

Table 5 shows a relatively low estimate of the income elasticity. An income elasticity less than one is suggested by economic theory in a life-cycle model inter alia by Ando and Modigliani (1963) and Gali (1990, p. 439).²⁷ In a regression with income as the only dependent variable, a value close to (and insignificantly different from) one was found for the entire group sample. This suggests that the presence of the wealth measures in the specification takes out some of the co-variation of income and consumption.

For the combined sample, both the estimated house price elasticity and the stock market price elasticity are positive and significant while the size of the coefficient estimate on house prices is about half the size of the coefficient estimate on stock market prices. The difference is significant with a t-ratio of 3.28. Splitting the sample into the two groups reveals that the estimated coefficients are roughly similar for the market-based economies while they are substantially reduced for the bank-based economies. This similarity of the coefficient estimates for the market-based economics to the estimates of the combined sample of all countries indicates that the market-based economies have a higher relative weight than the bank-based economies in the combined sample.²⁸ For stock market prices, the difference between the estimated elasticities for the two groups of countries (and its t-ratio) is 0.051 (4.89). The estimated housing price elasticity for the bank-based economies is insignificant and also lower than the significant estimate for the market based economies but the difference in the estimates between the two groups of countries is insignificant.

Thus, the estimated coefficients on stock market wealth are about twice as large as the estimated coefficients on housing wealth. This finding is consistent with our earlier analysis on the differences between the two wealth components (Section II). Yet it contrasts with the study of Case et al. (2001) who find a remarkably strong sensitivity of consumption to changes in housing wealth across countries (11 to 17 percent) whereas their estimated elasticity of consumption to changes in stock market wealth is rather low and unstable across different econometric approaches. In five out of six specifications in Case et al. the hypothesis that housing wealth has a stronger impact on consumption than stock market wealth is accepted. With regard to the different approaches between the two studies discussed in the introduction this difference in evidence must be interpreted with care. Yet Case et al. do not provide any explanation for their finding and given that housing wealth is supposedly less liquid, viewed more as a long term asset and that information on its exact value is not as widespread as for the value of stock market wealth, a higher sensitivity of consumption to changes in housing wealth seems counterintuitive and might have resulted from the way housing wealth was constructed in their data set.

The size of the adjustment coefficient is higher for the two separate samples than for the combined sample of all countries. While the individual group estimates of the adjustment coefficient remain unaltered for the market-based economies they increase for four countries (Finland, France, Norway and Spain) among the group of bank-based economies. Moreover, the mean group estimate of the adjustment coefficient is higher for the bank-based economies. The difference is insignificant but nevertheless prior considerations would have suggested the opposite. However, the estimated adjustment coefficients are positive for Australia and Canada and taking the average of the remaining five countries results in an average adjustment coefficient of-0.18. Both observations illustrate the sensitivity of mean group estimation to outliers in the small sample. Table 5 also shows that the mean group estimator on the first short run income coefficient is higher for the market-based economies. This result was expected and is insensitive to outliers.

Table 5 does not show any measure for the goodness of fit of the regression specifications since achieving a best possible fit is not a primary issue of this investigation. However the un-weighted average R² of the individual restricted country regressions is around 50 percent for all the regressions. The individual estimates of the fit vary a lot ranging from around 21 percent in the case of Canada to 91 percent in the case of Belgium (and this pattern again is independent of the estimated specification).

The likelihood ratio statistic that tests for the homogeneity assumption of the long run coefficients is rejected for all groups. A comparison of the sum of the two independent likelihood statistics for the two groups with the statistic for the combined panel reveals a modest gain from grouping the data into the two groups. However, equality of the mean group and pooled mean group estimators is accepted by the Hausman test for all groupings of the data. This means that the pooled estimates are not biased by the imposition of homogeneity. In sum, the evidence supports the hypothesis of a positive responsiveness of consumption to changes in the two wealth components on average over the last 15 years. The impact of price changes is higher for the group of market-based economies than it is for the group of bank-based economies and consumption is more sensitive to changes in stock market wealth than housing wealth.

E. Sensitivity analysis

Next, it is investigated whether the coefficient estimates shown in Table 5 have changed across time by estimating equation (3) for the period 1960Q1 to 1984Q4. Again, the effective time period is shorter and the number of observations range from 28 in the case of Ireland to 93 for Australia. As before, the Schwarz criterion is applied for the selection of the lag length for each country regression. Estimation results are shown in Table 6.

Table 6:

Estimating consumption functions (1960-1984)

Note: See Table A1 for a description of the variables. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. * (**) indicate significance at the 10 (5) percent level.

Table 6:

Estimating consumption functions (1960-1984)

Variables	All Countries	Bank-based Economies	Market-based Economies
	Long run coefficients
Yd	0.9171** (0.0159)	0.9045** (0.0191)	0.9144** (0.0273)
SMPI	0.0152** (0.0045)	0.0059 (0.0057)	0.0263** (0.0079)
HPI	-0.0544** (0.0131)	-0.0377** (0.0185)	-0.0282 (0.0230)
	Averages of heterogeneous short run coefficients
Adjustment coefficient	-0.213** (0.051)	-0.218* (0.074)	-0.223** (0.078)
d(C(-l))	0.130 (0.073)	0.202 (0.124)	0.039 (0.046)
d(Yd)	0.355** (0.077)	0.386** (0.105)	0.306** (0.115)
d(Yd(-1))	-0.095 (0.067)	-0.176 (0.117)	0.006 (0.006)
d(SMPI)	0.012** (0.005)	0.007 (0.006)	0.018** (0.008)
d(SMPI(-l))	0.0001 (0.003)	0.00 (0.00)	0.001 (0.007)
d(HPI)	0.009 (0.014)	-0.006 (0.006)	0.022 (0.030)
d(HPI(-I))	0.00 (0.00)	0.00 (0.00)	0.00 (0.00)
Intercept	0.094** (0.026)	0.116** (0.041)	0.043* (0.022)
Diagnostic Statistics
LR test (p-value)	188.95 (0.00)	86.33 (0.00)	99.5588 (0.00)
Joint Hausman test (p-value)	n.a.	n.a.	n.a.

Note: See Table A1 for a description of the variables. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. * (**) indicate significance at the 10 (5) percent level.

View Table

Table 6:

Estimating consumption functions (1960-1984)

Variables	All Countries	Bank-based Economies	Market-based Economies
	Long run coefficients
Yd	0.9171** (0.0159)	0.9045** (0.0191)	0.9144** (0.0273)
SMPI	0.0152** (0.0045)	0.0059 (0.0057)	0.0263** (0.0079)
HPI	-0.0544** (0.0131)	-0.0377** (0.0185)	-0.0282 (0.0230)
	Averages of heterogeneous short run coefficients
Adjustment coefficient	-0.213** (0.051)	-0.218* (0.074)	-0.223** (0.078)
d(C(-l))	0.130 (0.073)	0.202 (0.124)	0.039 (0.046)
d(Yd)	0.355** (0.077)	0.386** (0.105)	0.306** (0.115)
d(Yd(-1))	-0.095 (0.067)	-0.176 (0.117)	0.006 (0.006)
d(SMPI)	0.012** (0.005)	0.007 (0.006)	0.018** (0.008)
d(SMPI(-l))	0.0001 (0.003)	0.00 (0.00)	0.001 (0.007)
d(HPI)	0.009 (0.014)	-0.006 (0.006)	0.022 (0.030)
d(HPI(-I))	0.00 (0.00)	0.00 (0.00)	0.00 (0.00)
Intercept	0.094** (0.026)	0.116** (0.041)	0.043* (0.022)
Diagnostic Statistics
LR test (p-value)	188.95 (0.00)	86.33 (0.00)	99.5588 (0.00)
Joint Hausman test (p-value)	n.a.	n.a.	n.a.

Note: See Table A1 for a description of the variables. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. * (**) indicate significance at the 10 (5) percent level.

The results confirm a number of the hypotheses discussed above. First, the estimated stock market price has increased both for the entire sample as well as for both groups. In particular, the stock market elasticity has been very low and insignificant for the bank-based economies during the first observation period. As for the second observation period, the estimated elasticity is significantly higher for the market-based economies. Second, the estimates of the housing price elasticities are all negative. This suggests that during the period from 1960 to 1984 the negative impact of housing prices on consumption dominated the positive impact. The change in sign of the housing price coefficient is a striking evidence for chances in financial markets, particularly in the mortgage markets during the late 1980’s, which has made it easier for households to access their housing wealth (see Brady, Canner and Maki (2000) and Girouard and Blöndal (2001)).

The specification of the consumption function used here leaves out some variables that also influence consumption. Most importantly, only rather narrow measures of wealth were used. Including other independent variables like the short interest rate, government consumption and measures of the population age distribution either as exogenous regressors or in the long run relationship had only minor effects on the results reported above. However, Lettau, Ludvigson, and Barczi (2001) critically discuss the inclusion of such variables as exogenous regressors in a cointegrating framework like above since this results in a biased adjustment coefficient in a two-step Engle-Granger procedure if these variables arc not weekly exogenous. In a maximum likelihood approach as adopted here, including such variables might not only bias the adjustment coefficient but also the cointegrating relationship itself.

Beyond such robustness checks the robustness of the housing price estimates with regard to alternative measures of housing prices was investigated. In particular, it was checked whether the earlier estimates are robust with regard to alternative interpolation strategies like cubic spline. Also for a small sub-sample of countries regressions were run with actual quarterly housing price data that were provided by the ECB.²⁹ For this small set of countries (Germany, Italy, Spain, Sweden and the United Kingdom) and generally shorter time series, the estimates confirmed the positive influence of changes in housing prices on consumption that was found for the second observation period. An alternative grouping of the data according to homeownership rates as discussed in section II found a negative estimate for the long run elasticity of consumption to housing prices for the first period in both groups. In the second period it was, however, positive for the group of countries with relatively high ownership rates and negative for the other group, supporting our earlier findings of important interactions between financial markets and ownership of assets. Further, the robustness of the stock market price elasticities with data on stock market capitalization as a more direct measure of stock market wealth was investigated. The next subsection will turn to an analysis of these results.

F. From elasticities to marginal propensities to consume

In this subsection the elasticities estimated above are translated into estimates of marginal propensities to consume out of stock market wealth. For this, the regressions shown in Table 5 were estimated using logs of real stock market capitalization per capita instead of the price data. The stock market capitalization data are in quarterly frequency and taken from Datastream. Since comparable data on housing wealth arc not available it is impossible to perform the same exercise for housing wealth.

It is compelling how close the estimates for the long run elasticities to consume out of stock market wealth are to the earlier estimates in Table 5. The additional quantity dimension of the market capitalization data does not affect these estimates but changes the coefficient estimates of the long run income elasticities and the housing price coefficient, which is now higher for the bank-based economies (the difference is significant with a t-ratio of 4.3). The size of the estimated coefficient of housing wealth for the bank based economies is at the lower end of the range of estimates found in Case et al. (2001). However, compared to the other estimates it is unsystematically high and probably due to bias induced by the measurement error resulting from the interpolation of the housing price data. The estimated adjustment coefficient is now higher for the market-based economics while the difference between the two estimates is insignificant (t-ratio of 0.43). This is due to the fact that Australia and Canada now show negative coefficients and thus more reasonable coefficient estimates.

In order to translate the estimated elasticities into long run marginal propensities to consume (MPC) out of wealth, the estimates are multiplied with the recent aggregate consumption to stock market capitalization ratio. The short run estimates are accordingly obtained by multiplying the long run estimates with the estimated adjustment coefficient.³⁰

Table 7:

Estimating consumption functions - Using stock market capitalization data (1985-2000)

Note: See Table A1 for a description of the variabfes. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. *(**) indicate significance at the 10 (5) percent level.

Table 7:

Estimating consumption functions - Using stock market capitalization data (1985-2000)

Variables	All Countries	Bank-based Economies	Market-based Economies
	Long run coefficients
Yd	0.4759** (0.0218)	0.4551** (0.0276)	0.4690** (0.0232)
SMPI	0.0950** (0.0029)	0.0382** (0.0032)	0.0963** (0.0031)
HPI	0.0353** (0.0133)	0.1069** (0.0096)	0.0309** (0.0145)
	Averages of heterogeneous short run coefficients
Adjustment coefficient	-0.131** (0.035)	-0.175* (0.065)	-0.221** (0.060)
d(C(-l))	0.067 (0.099)	0.061 (0.152)	0.114 (0.114)
d(Yd)	0.256** (0.085)	0.325** (0.126)	0.153* (0.087)
d(Yd(-1))	-0.108 (0.077)	-0.148 (0.110)	-0.048 (0.092)
d(SMCAP)	-0.005 (0.002)	-0.003 (0.003)	-0.009* (0.005)
d(SMCAP(-l))	-0.002 (0.002)	0.000 (0.000)	-0.005 (0.005)
d(HPI)	0.147** (0.040)	0.173** (0.074)	0.203** (0.064)
d(HPI(-l))	-0.081 (0.054)	-0.045 (0.052)	-0.111 (0.111)
Intercept	0.086** (0.026)	0.282** (0.105)	0.074** (0.023)
Diagnostic Statistics
LR test (p-value)	236.73 (0.00)	156.44 (0.00)	50.09 (0.00)
Joint Hausman test (p-value)	14.90 (0.00)	3.49 (0.32)	n.a.

Note: See Table A1 for a description of the variabfes. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. *(**) indicate significance at the 10 (5) percent level.

View Table

Table 7:

Estimating consumption functions - Using stock market capitalization data (1985-2000)

Variables	All Countries	Bank-based Economies	Market-based Economies
	Long run coefficients
Yd	0.4759** (0.0218)	0.4551** (0.0276)	0.4690** (0.0232)
SMPI	0.0950** (0.0029)	0.0382** (0.0032)	0.0963** (0.0031)
HPI	0.0353** (0.0133)	0.1069** (0.0096)	0.0309** (0.0145)
	Averages of heterogeneous short run coefficients
Adjustment coefficient	-0.131** (0.035)	-0.175* (0.065)	-0.221** (0.060)
d(C(-l))	0.067 (0.099)	0.061 (0.152)	0.114 (0.114)
d(Yd)	0.256** (0.085)	0.325** (0.126)	0.153* (0.087)
d(Yd(-1))	-0.108 (0.077)	-0.148 (0.110)	-0.048 (0.092)
d(SMCAP)	-0.005 (0.002)	-0.003 (0.003)	-0.009* (0.005)
d(SMCAP(-l))	-0.002 (0.002)	0.000 (0.000)	-0.005 (0.005)
d(HPI)	0.147** (0.040)	0.173** (0.074)	0.203** (0.064)
d(HPI(-l))	-0.081 (0.054)	-0.045 (0.052)	-0.111 (0.111)
Intercept	0.086** (0.026)	0.282** (0.105)	0.074** (0.023)
Diagnostic Statistics
LR test (p-value)	236.73 (0.00)	156.44 (0.00)	50.09 (0.00)
Joint Hausman test (p-value)	14.90 (0.00)	3.49 (0.32)	n.a.

Note: See Table A1 for a description of the variabfes. The LR test tests for the homogeneity restriction of the long run coefficients. The Hausman test tests for equality of the MG and the PMG estimators. The Hausman test statistic is indeterminate if the difference between the variance-covariance matrices of the MG and PMG estimators is not positive definite (see Pesaran et al. (1999) for more details). The unrestricted short run coefficient estimates are the MG estimates under the restriction of long run homogeneity. Standard errors of the estimated coefficients are in parenthesis. *(**) indicate significance at the 10 (5) percent level.

Table 9 reports such estimates for a group of countries. The estimated long run MPC for the United States is 0.04. A dollar increase in stock market capitalization therefore causes consumption to increase by 4 cents in the long run which translates into a short run adjustment of about 0.8 cents increase in consumption per quarter. These estimates are in the range of estimates conventionally suggested in the literature.

For comparability with the estimate of the United States’ MPC, the estimates of the MPCs of all other countries are adjusted by the variation in the stock market capitalization to GDP ratio of each particular country relative to the United States. Note that this is just another way to express the estimated elasticities from before. The average of the adjusted long run MPCs is 4.3 cents for the market-based economies and 2.6 cents for the bank-based economics.

Table 8:

Marginal propensities to consume out of stock market wealth

Note: The marginal propensities to consume out of stock market wealth are calculated by multiplying the estimated elasticities with the most recent consumption to wealth ratios. Stock market wealth in each country is measured by stock market capitalization data. The adjusted marginal propensities to consume are calculated by correcting the original marginal propensity of countryi by the variation in the first differenced series of the stock market capitalization to GDP ratio of countryi relative to the US.

Table 8:

Marginal propensities to consume out of stock market wealth

	AU	CA	UK	US	JAP	FR	GE	IT
Adj. L.R. MPC	0.043	0.040	0.049	0.040	0.040	0.014	0.020	0.030
Adj. S.R. MPC	0.0094	0.0089	0.0109	0.0089	0.007	0.0025	0.0035	0.0053

Note: The marginal propensities to consume out of stock market wealth are calculated by multiplying the estimated elasticities with the most recent consumption to wealth ratios. Stock market wealth in each country is measured by stock market capitalization data. The adjusted marginal propensities to consume are calculated by correcting the original marginal propensity of countryi by the variation in the first differenced series of the stock market capitalization to GDP ratio of countryi relative to the US.

View Table

Table 8:

Marginal propensities to consume out of stock market wealth

	AU	CA	UK	US	JAP	FR	GE	IT
Adj. L.R. MPC	0.043	0.040	0.049	0.040	0.040	0.014	0.020	0.030
Adj. S.R. MPC	0.0094	0.0089	0.0109	0.0089	0.007	0.0025	0.0035	0.0053

Note: The marginal propensities to consume out of stock market wealth are calculated by multiplying the estimated elasticities with the most recent consumption to wealth ratios. Stock market wealth in each country is measured by stock market capitalization data. The adjusted marginal propensities to consume are calculated by correcting the original marginal propensity of countryi by the variation in the first differenced series of the stock market capitalization to GDP ratio of countryi relative to the US.