The Impact of Stock Prices and House Prices on Consumption in OECD Countries

Transcription

1 WP/01/xx The Impact of Stock Prces and House Prces on Consumpton n OECD Countres Alexander Ludwg and Torsten Sløk Prelmnary Verson, October 26, 2001 Comments most welcome.

2 2001 Internatonal Monetary Fund WP/01/xx IMF Workng Paper Research Department The Impact of Stock Prces and House Prces on Consumpton n OECD Countres Prepared by Alexander Ludwg and Torsten Sløk 1 Authorzed for dstrbuton by Tamm Bayoum Abstract The vews expressed n ths Workng Paper are those of the author(s) and do not necessarly represent those of the IMF or IMF polcy. Workng Papers descrbe research n progress by the author(s) and are publshed to elct comments and to further debate. Abstract: Ths paper quantfes the dfferent mpact of stock and house prces on consumpton usng data for 16 OECD countres. The analyss fnds that the long-run mpact of an ncrease n stock prces s n general hgher n countres wth a market-based fnancal system whereas the long-run mpact of an ncrease n housng wealth s more or less dentcal n countres wth market-based and bank-based fnancal systems. Splttng the sample nto the 1980s and 1990s shows that all economes have moved towards a hgher degree of responsveness to changes n stock prces and house prces on average. For stock prces the dfference between the market and bank-based economes remans sgnfcant. JEL Classfcaton Numbers: E20, E44 Keywords: Consumpton, Wealth effect, Asset prces Authors E-Mal Address: tsloek@mf.org and ludwg@econ.un-mannhem.de 1 Ths paper was wrtten whle Alexander Ludwg was a summer ntern at the IMF. Helpful comments by Tamm Bayoum, Hal Edson, Zhu Feng, Robert Flood, Doug Laxton, and several semnar partcpants at the IMF Research Department are gratefully acknowledged. All remanng errors are of course ours.

3 - 3 - Contents Page 1. Introducton...4 II. Transmsson mechansms from wealth to consumpton...6 III. The Econometrc Model...12 IV. Emprcal Results...15 VI. Concluson...30 Tables 1. Table 1: Homeownershp n dfferent countres Table 2: Summary of man hypothess Table 3: IPS (1997) panel unt root tests Table 4: Pedron (1999) panel contegraton tests Table 5: Estmatng consumpton functons ( ) Table 6: Estmatng consumpton functons ( ) Table 7: Estmatng consumpton functons ( ) Table 8: Estmatng consumpton functons Usng stock market captalzaton data ( ) Table 9: Margnal propenstes to consume out of stock market wealth...29 Fgures 1. Fgure 1: Real stock prce ndex, selected countres Fgure 2: Stock market captalzaton as percentage of GDP Fgure 3: Real house prce ndex Fgure 4: Percentage of adult populaton holdng shares n Bblography...34 Appendx Table A1: Data avalablty...38

4 - 4 - I. INTRODUCTION Ths paper examnes the role of stock prces and house prces as determnants of prvate consumpton. The mpact of stock market wealth on consumpton has receved ncreased attenton among both academc researchers and polcymakers, partly because of the dramatc ncrease and subsequent fall n stock prces experenced n the 1990s (Fgure 1). 2 Measured n percent of GDP the Anglo-Saxon countres have experenced the strongest gans (Fgure 2) and have consequently also been the focus of most studes. For many contnental European countres the ncrease n stock market wealth has also been qute substantal. In Japan, however, asset prces have fallen through the 1990s, reflectng the protracted downturn for the Japanese economy. Whle movements n fnancal wealth have been domnated by movements n stock market wealth, housng wealth s the sngle most mportant component of non-fnancal wealth n households portfolos (Deutsche Bank, 2001). Due to pronounced ncreases n housng wealth (Fgure 3) and deregulaton of mortgage markets, the mpact of housng wealth on consumpton n OECD countres has therefore also receved ncreased attenton among researchers and polcy makers. 3 As Greenspan (2001) has recently suggested, the margnal propensty to consume out of housng wealth mght dollar for dollar be hgher than the margnal propensty to consume out of stock market wealth whle the overall mpact of the latter must expected to be hgher. The vews on the role between asset prces and real economc actvty dffer wdely n the lterature. At one end of the spectrum t has been argued that the observed correlaton between asset prces and consumpton expendtures are due to the role of asset prces as leadng ndcators (see e.g. Morck et al. (1990); Poterba and Samwck (1995)). Accordng to ths vew, asset prces reflect future output growth and are therefore correlated wth consumpton expendtures. At the other end of the spectrum s the vew that the observed correlaton s due to real wealth effects the vew that has recently been emphaszed and s also shared n ths contrbuton. The dstncton between these two perspectves s of hgh mportance regardng polcy conclusons from the observed correlatons. It wll therefore be further dscussed n the concludng remarks n secton V. Ths paper contrbutes to the exstng lterature by takng a broader perspectve n nvestgatng the relatve mportance of two wealth components housng and stock market wealth - n a sample of 16 OECD countres. These countres are grouped nto bank-based and market-based economes and dfferences between the relatve mportances of the two wealth 2 See IMF (2000) and (2001a), Greenspan (2001), Edson and Sløk (2001a, b), Mak and Palumbo (2001), Davs and Palumbo (2001), Ludvgson et al (2001), and Mehra (2001). 3 See for example Grouard and Blöndal (2001), Deutsche Bank (2001) and Brady, Canner and Mak (2000).

5 - 5 - Fgure 1: Real stock prce ndex, selected countres (1995Q1=100) FR US GE UK JA Q1 1975Q1 1980Q1 1985Q1 1990Q1 1995Q1 2000Q1 Source: Internatonal Fnancal Statstcs, IMF (2001b) Fgure 2: Stock market captalzaton as percentage of GDP UK US AU JA IT GE FR Q1 1979Q2 1985Q3 1991Q4 1998Q Q1 1979Q2 1985Q3 1991Q4 1998Q1 Source: Datastream Fgure 3: Real house prce ndex UK US FR JAP GE Source: Bank for Internatonal Settlements, (BIS, 2001)

6 - 6 - components n these sub groups are analyzed. A new panel data technque for contegrated panels developed by Pesaran, Shn and Smth (1999) s appled n order to address these and related questons. Prevewng the results, the analyss leads to the followng conclusons. Frst, there s, as expected, a sgnfcant dfference n the role of stock market wealth n the countres wth market-based fnancal systems and countres wth bank-based fnancal systems. In regons wth market-based fnancal systems, the role of house prces and stock prces s hgher n the long run. In terms of numbers, the estmated long run elastcty of consumpton n stock market wealth s about 2.5 tmes hgher n Anglo-Saxon countres compared wth the smlar coeffcent for the contnental European countres. Whle the dfference between the two groups of countres remaned stable, the estmated long run elastctes have sgnfcantly ncreased over tme mrrorng not only the ncreased mportance of stock market wealth but also the changes n fnancal systems. For house prces the evdence on the relatve mportance between the two groups s mxed and not robust aganst alternatve specfcatons. However, a consstent fndng s that contrary to the sample perod the mpact of changes n housng prces on consumpton was postve for the perod n both groups ndcatng that the mportance of housng wealth has ncreased over tme. The paper s organzed as follows. Secton II dscusses varous transmsson mechansms from stock market prces and housng prces to consumpton. The econometrc model s ntroduced n secton III and the emprcal results of varous specfcatons are dscussed n secton IV. Secton V concludes. II. TRANSMISSION MECHANISMS FROM WEALTH TO CONSUMPTION Ths Secton analyzes the varous transmsson mechansms from changes n the two varables on consumpton. Tracng out the ndvdual channels also helps to derve the man hypotheses that wll be tested n the emprcal nvestgaton. A. Transmsson channels of stock prces Broadly speakng, there are fve dfferent transmsson channels from changes n stock market prces to changes n consumpton 4 : 1. Realzed wealth effect: If the value of consumers stock holdngs ncrease and consumers realze ther gans then consumpton wll ncrease. Ths result wll be a drect effect as a consequence of hgher current lqud assets. 4 Some of these channels were frst dentfed n the lfe-cycle and permanent ncome theory (Fredman (1957), Ando and Modglan (1963), Modglan and Brumber (1979)).

7 Unrealzed wealth effect: An ncrease n stock prces can also have an expectatons effect where the value of stocks n penson accounts and other locked-n accounts ncreases. When these assets go up n value but the ncrease s not realzed t results n hgher consumpton today on the expectaton that ncome and wealth wll be hgher n the future Lqudty constrants effect: An ncrease n stock market prces ncreases the value of a portfolo for an nvestor. Borrowng aganst the value of ths portfolo n turn allows the consumer to ncrease consumpton. 4. Stock opton value effect: An ncrease n stock prces can lead to hgher consumpton for stock opton owners as a result of an ncrease n the value of households stock optons. Ths effect s most pronounced n sectors and frms such as the technology sector where stock optons are granted more often as a means of payment (partly because of the hgh rsk nvolved n undertakng new projects). Agan, ths ncrease n consumpton may come ndependently of whether the gans are realzed or unrealzed. 5. Substtuton effect: An ncrease n stock prces may lead to further nvestment and therefore to a postponement of consumpton. Whle transmsson channels 1 through 4 are thus assocated wth a postve mpact on consumpton, channel 5 works n the opposte drecton. It s very lkely that the substtuton effect of stock market prces s of mnor mportance snce an nvestment decson wll always be undertaken wth the purpose of ncreasng wealth (and thereby consumpton) n the long run. Thus, n the aggregate economy the negatve substtuton effect wll probably n far most cases be outweghed by the postve wealth effects. It s worth notng the dfference between the realzed and unrealzed wealth effect. As Poterba (2000) puts t: t seems partcularly lkely that the margnal propensty to consume out of wealth gans n (locked) retrement accounts s lower than the propensty to consume out of drectly held assets snce the former are often thought of as long term assets (see also Thaler (1990)). Accordng to Poterba and n lght of the growng relatve mportance of retrement accounts ths dfference may have reduced the margnal propensty to consume out of stock market wealth n the Unted States. Indeed, Ludvgson and Stendel (1999) as well as Mehra (2001) estmate a lower margnal propensty to consume out of total wealth for samples of later perods. Furthermore, there appears to be a dfference n how households (and frms) adjust ther consumpton to changes n dfferent types of assets. For non-technology stocks Edson and Sløk (2001a and 2001b) fnd a stronger reacton n consumpton to changes n stock prces n market-based fnancal systems (Unted States, Canada, and UK) than n bank-based fnancal systems such as contnental Europe. For changes n technology stock prces, however, the consumpton reacton s more smlar, reflectng that the technology sector worldwde seems to 5 Smlarly, f the value of assets goes down and s unrealzed, t may force consumers to lower ther consumpton as ther expected stream of future ncome (when they retre) goes down.

8 - 8 - functon and carry out busness n a more smlar way. Fnally, consumpton of households who do not partcpate n the stock market may be ndrectly affected by changes n stock market prces. Such ndrect effects between stock prces and consumpton have for example been hghlghted n Romer (1990) and Zand (1999) and are closely related to the dstncton between the wealth vew and the leadng ndcators vew of stock market prces that wll be further dscussed n the conclusons. B. Channels of transmsson for housng wealth There are also fve dfferent transmsson channels from changes n housng prces to changes n consumpton, but some of these channels are somewhat dfferent from the channels for stock prces lsted above: 1. Realzed wealth effect: For consumers who are house owners, the ncrease n house prces leads to an ncrease n net wealth, whch can rase consumpton today. If house prces ncrease t s possble for consumers to take out equty n the form of refnancng or sellng of the house. Such a realzed gan must be expected to have a postve mpact on prvate consumpton. 2. Unrealzed wealth effect: If house prces ncrease but households do not refnance or sell the house t may stll have a postve mpact on consumpton due to the ncrease n the dscounted value of wealth. Hence consumers can spend more today on the expectaton that they are rcher than they were before. 3. Budget constrant effect: For consumers who are house (or apartment) renters and ncrease n house prces has a negatve mpact on prvate consumpton. As house prces go up the budget constrant becomes tghter for renters, whch must be expected to result n lower prvate consumpton. Ths channel works through a realzed captal loss snce the ncrease mmedately leads to hgher prces, whch have to be pad by the renters Lqudty constrants effect: A fourth factor that s also mportant for the consumpton mpact of house prce changes s how well functonng the fnancal system s. If house prces change t may requre access for consumers to credt markets n order to take loans aganst the ncrease n house prces. If credt s constraned or the fnancal system s not able to support such a wsh for loans households may experence that they cannot react accordngly to hgher house prces. 5. Substtuton effect: An ncrease n house prces may also mply that households who are plannng to buy a home may lower consumpton when faced wth ncreasng hgher prces 6 Ths budget constrant effect s also but to less extend relevant for house owners snce an ncrease n housng prces mght not only ncrease the rent but also other expendtures on housng servces lke fuel and power.

9 - 9 - snce they may ncrease down payments and future loans. As a result, when house prces ncrease consumers who are not yet on the housng market wll face hgher loans and hence wll ether choose a smaller house or to lower prvate consumpton. Both the realzed and unrealzed gans from ncreases n house prces must be expected to ncrease prvate consumpton but just as wth stocks the margnal propensty to consume out of unrealzed ncreases n housng wealth mght be lower. Despte the fact that an ncrease n housng prces also ncreases the expendtures on housng servces and that housng wealth s less lqud, the wealth effect of an ncrease n housng prces s smlar for homeowners as an ncrease n stock prces for stockowners. But snce the decson to nvest n housng wealth manly s drven by other motvatons than the decson to nvest n the stock market, the negatve substtuton effect due to changes n housng prces s lkely to be much more mportant. In addton, the budget constrant effect negatvely affects consumpton. There are varous other dmensons along whch housng wealth dffers from stock market wealth. For example, a smlar argument to Poterba s (2000) argument about the lower margnal propensty to consume out of unrealzed wealth can be made for housng wealth snce t s consdered as long-term asset. Moreover, the two types of assets have dfferent rsk characterstcs as for example mrrored n lower loan-to-value ratos for housng wealth and there mght be hgher costs mbedded to gettng nformaton on housng wealth than on stock market wealth. Furthermore, the ndrect wealth channel dscussed above mght not work for housng wealth. In sum, whle the effects of changes n stock market prces seem to unambguously pont to an ncrease n consumpton, the mpact of changes n housng prces s ambguous. 7 Further, the dscusson of the varous channels of transmsson from movements n both wealth components to consumpton suggests that the fnancal system plays a central role n ths process. C. The role of the fnancal system The role of the fnancal system s to amelorate market frctons. Fnancal ntermedares do that by provdng fnancal servces whereby: a) they assess nvestment opportuntes and provde corporate control, b) they ease rsk management (ncludng lqudty rsk), and c) they lower the costs of moblzng resources. Three factors affect the strength of the wealth transmsson channels dscussed above. Frst, when wealth ncreases, consumers may want to borrow aganst the ncrease n the wealth component. Such a wsh on the consumer s sde requres the exstence of banks or markets that 7 The arguments on the varous mpacts of housng and stock market wealth on consumpton, abstract from possble lnks between the two wealth components. The evdence on such lnks however s mxed (see Poterba (2000) and the lterature cted there).

10 lend to ndvduals and are able to assess the rsk nvolved n the fnancal contract. For example, f house prces ncrease, consumers may want to borrow aganst the ncreased value of ther wealth and ths can be done ether through the bankng system or through the mortgage market. Second, fnancal ntermedares such as banks may ease rsk sharng and poolng by lowerng transacton costs. Fnancal ntermedares may lower the costs of holdng a standardzed portfolo of assets f there are fxed costs to each purchase. Moreover, fnancal ntermedares may facltate the ntertemporal smoothng of rsk. Rsks that cannot be dversfed at a partcular pont n tme, such as macroeconomc shocks, can be dversfed across generatons. Thrd, fnancal ntermedares facltate savngs moblzaton poolng by economzng on the transacton costs assocated wth moblzng savngs from many dsparate agents and by overcomng the nformatonal asymmetres assocated wth makng savers comfortable n relnqushng control of ther savngs. Dfferences n the desgn of the fnancal systems are partcularly pronounced between on the one sde the contnental European countres and Japan whch have bank-based fnancal systems and then on the other sde the US, Canada, and UK who have market-based fnancal systems. The market-based and bank-based fnancal systems have produced dfferent outcomes. Of key mportance for the strength of the wealth effect are dfferences regardng szes of fnancal markets and how wdespread stock ownershp s and the use of stock optons as a means of payment by frms. These factors all nfluence the magntude of the mpact of an ncrease n wealth, and these dfferent components wll now be analyzed n turn. Frst, the market-based system has resulted n dfferent szes of stock markets. In general, the sze of the stock markets s much hgher n Anglo-Saxon countres than n the bankbased systems n Contnental Europe (Fgure 2). Accordng to ths crteron, the Netherlands, Japan, and Sweden mght also be consdered to belong to the group of market-based economes. Japan s a bt specal, whch has to do wth sgnfcant amounts of crossholdng and cross ownershp of stocks. Second, the market-based system has also lead to a hgher degree of partcpaton n countres wth market-based fnancal systems than n countres wth bank-based systems. Fgure 4 descrbes some of the dfferences. Accordng to ths crteron, agan the Netherlands and Sweden could be consdered as market-based systems but for now they wll be ncluded n the group of countres wth bank-based fnancal systems. Thrd, stock optons as a means of payment s at least untl now more wdespreadly used n Anglo-Saxon countres than n countres wth bank-based fnancal systems. Also, numbers

11 for venture captal suggest that there s a stronger market-based tradton n the Anglo-Saxon countres. 8 Fgure 4. Percentage of adult populaton holdng shares n USA Australa Sweden Netherlands Canada UK Denmark Germany Fnland Austra France japan Norway Belgum Source: Proshare (2000) Table 1: Homeownershp n dfferent countres Ireland 80.0 % Span 78.0 % Italy 78.0 % Unted States 67.5 % Unted Kngdom 67.0 % Belgum 65.0 % EU Average 61.0 % Fnland 60.8 % Sweden 60.0 % Austra 55.0 % France 54.0 % Denmark 52.3 % Netherlands 52.0 % Germany 41.0 % Source: Hypostat, (2001) and Fanne Mae Foundaton (2001). Data are from the most recent year (and all data are from the 1990s). Boro (1996) provdes an analyss of the credt characterstcs to the non-government sector n fourteen ndustralzed countres. Hs crtera and the analyss so far suggest assgnng Australa, Canada, the Netherlands, Japan, Sweden and the Unted Kngdom to an extended group of market-based economes whereas Belgum, Denmark, Fnland, Germany, Ireland, Italy, Norway, and Span are treated as bank-based. A dstncton between countres accordng 8 See e.g. Edson and Sløk (2001) for a more general dscusson of ths ssue.

12 to more objectve measures and formal crtera as n Beck and Levne (2001) would be warranted. See the conclusons for a more general dscusson on ths ssue. Whle the dstncton of these two groups of countres accordng to stock market captalzaton and the degree of partcpaton n the stock market s more or less dstnct, such an orderng matches countres wth very dfferent homeownershp rates nto one group. The comparson of homeownershp rates shows that three of the countres n the group of marketbased economes (Ireland, the Unted States and the Unted Kngdom) have farly hgh homeownershp rates, two others (Sweden and the Netherlands) do not (Table 1). Such an alternatve groupng of the data wll be among the senstvty checks carred out n secton IV. Overall, these factors support the noton that changes n stock market wealth have a larger mpact on consumpton n countres wth market-based fnancal systems. At the same tme, the responsveness of consumpton to changes n stock market prces must be expected to have ncreased substantally over tme. For housng wealth, however, the mpact s more uncertan. In partcular, as dscussed above, there s no a pror reason to expect a postve effect of changes n housng prces on consumpton. But due to the deregulaton of fnancal markets across countres and a correspondng ncreased nformaton level and transparency n fnancal markets, t must be expected that the postve nfluences of housng prces on consumpton have become more mportant over tme. Ths effect s lkely to be more pronounced n countres where households have easer access to ther housng wealth, e.g. refnancng of loans. Based on the consderatons above Table 2 summarzes the hypothess to be tested n the emprcal secton. Hypothess 1 Hypothess 2 Hypothess 3 Hypothess 4 Hypothess 5 Table 2: Summary of man hypothess Postve effect of stock market prces on consumpton Stock market effects are hgher n countres wth market-based fnancal systems The mpact of housng prces on consumpton s ambguous Stock market effects on consumpton have ncreased over tme The postve effect of changes n housng prces has become more mportant over tme III. THE ECONOMETRIC MODEL Consumpton s tradtonally explaned by ncome and wealth, and a man focus of ths paper s the relatve mportance of dfferent wealth components n dfferent countres. A smple model of an aggregate consumpton functon wth household (labor) ncome and wealth as the only determnants s motvated by several theores, ncludng the permanent ncome theory by Fredman (1957) and the lfe cycle theory by Modglan and Brumberg (1954) and Ando and

13 Modglan (1963). 9 In most emprcal studes of the wealth effect of consumpton, a common trend among the three varables s assumed and tested for. 10 Gal (1990) provdes a theoretcal foundaton for a common trend approach between these three macroeconomc aggregates that s derved from an open-economy verson of Blanchard s (1985) overlappng generatons model. In ths paper t s assumed that (and tested f) such a contegratng relatonshp exsts between consumpton, ncome, and the two wealth measures. Next, an error correcton specfcaton of a consumpton functon as frst proposed n Davdson et al. (1978) s estmated for a sample of 16 OECD countres usng panel data technques. The lterature on nference n dynamc and contegrated panels has evolved rapdly over the past few years. 11 Among the varous estmators suggested n the lterature, the pooled mean group (PMG) estmator proposed by Pesaran et al (1999) s partcularly attractve snce t pools the long run relatonshp between the countres whle the short run responses are flexble and unrestrcted across countres. Ths lkelhood-based estmaton procedure s an ntermedate procedure to poolng the panel data and a mean group estmator (MG) n the fashon of Pesaran and Shn (1995). Economc theory suggests that the long run ncome elastcty of consumpton s equal to one. In a panel data analyss of a group of countres, t s therefore reasonable to assume that ths coeffcent s equal across countres. On the other hand, f ths assumpton s not vald, then poolng the cross secton nformaton mght stll have some merts snce t yelds to more effcent estmates than runnng ndependent regressons for each group and computng an average of the estmated coeffcents, the mean group estmator. Moreover, when N s small as s the case here, the PMG estmator s less senstve to outlers snce t weghs the ndvdual unrestrcted country coeffcents accordng to ther precson (see Pesaran, Shn and Smth (1999) for a more detaled dscusson). An dentcal form of the long run consumpton functon s assumed for all countres, where the long run relatonshp between consumpton, ncome and the two wealth measures s gven by d sw hw (1) c α + α y + α w + α w + ε, = 1,2,..., N, t 1,2,...T, t, = 0 1 t, 2 t, 3 t,, t = and the subscrpts and t denote the country and tme respectvely. c s the log of prvate per capta consumpton, y d the log of per capta dsposable household ncome and w hw and w sm refer to the logs of housng and stock market wealth respectvely. ε s the error term capturng the 9 For a further dscusson of the theores underlyng the consumpton functon see Deaton (1992), Brownng and Lusard (1996), Attanaso (1999), Mehra (2001) and Lettau and Ludvgson (2001). 10 See e.g. the recent studes by Mehra (2001) and Lettau and Ludvgson (2001). 11 An excellent overvew s gven n Baltag and Kao (2001).

14 effects of unexpected shocks to consumpton. Note that one may also thnk of the LHS of equaton (1) as planned consumpton (compare to Mehra, 2001). Devatons from the long run relatonshp gven by equaton (1) are possble n the short run. There are varous reasons for such devatons ncludng adjustment costs, habt persstence and lqudty constrants (See Mehra (2001), Poterba (2000, p. 112) and also Campbell and Mankw (1991)). Attanaso (1999) stresses the mportance of adjustment costs for durable consumpton, whch s ncluded n the measure of consumpton consdered here (see secton 4). It s assumed that (short run) consumpton functons dffer across countres. Ths assumpton wll n the followng be mplemented by determnng the lag length of each varable by conventonal statstcal crtera. For ease of presentaton, t wll be assumed that the frst lag of each varable s an mportant determnant of the short run relatonshp n each country. The ARDL(1,1,1,1) specfcaton of equaton (1) therefore becomes d d sw sw hw hw (2) ct δ + β yt + β yt + β wt + β wt + β wt + β wt + γ ct + t, = 10, 11 1, 20, 21 1, 30, 31 1, 1, η,. The error term s assumed to be ndependently dstrbuted across t and but the varances may be heterogeneous across countres. The cross-sectonal ndependence assumpton of the error term s rather strong and restrctve. For example, t s not hard to magne shocks that affect all countres at the same tme. Ths assumpton s standard n the dynamc panel lterature and ts mplcatons for ths analyss wll be dscussed n the concluson. 12 Moreover t s assumed that the error term s ndependent of all the other varables n equaton (2), an assumpton that s nvaldated f other mportant varables nfluence consumpton that s not contaned n equaton (2). Rewrtng equaton (2) gves the error correcton specfcaton as: d sw hw d sw hw (3) ct = φ ( ct α α yt α wt α wt ) + β yt + β wt + β wt + η t. where, 1, 0 1 1, 2 β β + β β + β β + β (4) φ = (1 γ ), α =, α =, α =, α = γ 1 γ 1 γ 1 γ 1, In ths framework, Pesaran et al. (1999) have suggested to restrct the coeffcents of the error correcton term n equaton (3) to be equal across countres whle all other short run coeffcents are allowed to vary. The equaton s then estmated by maxmum lkelhood. 13 The 12 Pesaran et al. (1999) dscuss cures to be undertaken f ths assumpton s volated. One s to nclude observable varables as exogenous regressors that are common for all groups. Another approach would be to remove such factors by runnng the regresson on the demeaned data. 13 The approach thus dffers from the conventonal two-step procedure appled to sngle equaton contegratng relatonshps as n the semnal contrbuton of Engle and Granger (1987). 3 1, 10, 20, 30,,

15 restrcton of equal long run coeffcents across countres can be tested by a conventonal lkelhood rato (LR) test. As Pesaran et al (1999) pont out, t s very lkely that ths homogenety restrcton s rejected n emprcal applcatons. One obvous explanaton for such a rejecton s smple: the restrcton s wrong. Another explanaton s that there mght be samplespecfc omtted varables n the ndvdual country regressons or measurement errors that are correlated wth the regressors. Whle t mght be possble to correct for such bases n ndvdual country regressons, t s mpossble to do so n a panel of countres. If such bases average to zero across groups, then poolng s sensble snce t removes such random varaton. Thus poolng mght provde a more reasonable estmate of the true coeffcent. If the restrcton s wrong and one s nterested n the average affect across a certan group of countres, then poolng and thereby ex ante mposng homogenety mght be more reasonable n small samples. Whle the MG estmator (the un-weghted average of ndvdual country specfc estmates) where homogenety s mposed ex post s very senstve to outlers n small samples, poolng whch weghs the ndvdual country specfc heterogeneous coeffcents accordng to precson, reduces bas such bas. Along ths lne the estmated coeffcents can be nterpreted as the weghted averages of ndvdual group estmators whle the weghts are determned by the nverses of ther varance covarance matrces. 14 A few further remarks on the econometrc procedure are n order: () The coeffcents on the lagged dependent varables n equaton (2) are subject to the famlar small sample (small T) downward bas. Snce ths downward bas s n the same drecton for each group, averagng or poolng does not remove the bas. Kvet and Phllps (1993) have proposed a procedure to remove ths bas, whch apples to the short run coeffcents. Snce the long run coeffcents are nonlnear transformatons of the short run coeffcents such bas correctons wll leave the long run coeffcents unbased (see equaton (4)). We are not aware of any procedure n the lterature that has resolved ths problem. 15 () As Pesaran and Smth (1995) have ponted out, falsely mposng homogenety n panels leads to an upward bas n the estmates of the coeffcents on the lagged dependent varables, a bas that s not reduced when both T and N grow large. It s possble to determne f such an upward bas s serous. Under slope homogenety, the PMG estmators are consstent and effcent whle the MG estmator s consstent but neffcent. Therefore a Hausman-type test for comparson of the MG and the PMG estmators can be appled. () The fact that the long run coeffcents are (nonlnear) functons of the short run coeffcents also mples that the long run coeffcents change whenever the lag order s ncreased or supposedly exogenous regressors are added and f these varables are correlated wth the long run relatonshp,.e. the added varables are not weakly exogenous. In a contegraton framework a la Engle and Granger (1987) t s only the adjustment coeffcent that changes n such a case snce the contegratng vector s estmated ndependently from the system. 14 See the dscusson n the 1997 workng paper verson of Pesaran et al. (1999) on page Pearan and Zhao (1997) have suggested a correcton of the bas for the MG estmator.

16 IV. EMPIRICAL RESULTS A. Data Below, equaton (3) wll be estmated for the entre sample as well as for groups of countres accordng to the dscusson n Secton II. Data covers 16 OECD countres. Specfcally, Belgum, Denmark, Fnland, France, Germany, Italy, Norway and Span wll be ncluded nto the group of economes wth bank-based fnancal systems whle Australa, Canada, Ireland, Japan, the Netherlands, the Unted Kngdom, the Unted States and Sweden are treated as market-based economes. Data avalablty of the unbalanced panel s shown n Table A1 of the appendx. Gven the broad coverage of the study there are certan data lmtatons. Therefore the approach adopted below devates a bt from the smple theoretcal model formulated above. Frst, stock market and housng prces are here used as proxy varables for the wealth components. Note the hgh correlaton between stock market prces and wealth measures as documented n Lettau and Ludvgson (2001) and Deutsche Bank (2001). The (n)drect mpact of stock market prces on aggregate consumpton has for example been nvestgated n the event studes by Pennar (1988) and Romer (1990) and more recently n Poterba and Samwck (1995) and Zand (1999). The role of housng prces on consumpton s the focus n, among others, Mles (1992), Mles (1995, chapter 4) and more recently Brady et al. (2000) and Grouard and Blöndal (2001). To underscore the valdty of usng prce data as proxy varables, the analyss s extended n secton IV.F by usng market captalzaton as a more drect measure of stock market wealth. But even wth the use of market captalzaton data as a proxy for stock market wealth of households, one problem remans: nternatonal captal moblty. However stll a hgh proporton of stocks wthn each country are held by domestc resdents and therefore the use of such varables as proxes for domestc stock market wealth seems to be a vald approxmaton. Second, focus s on total aggregate consumpton and there s no dstncton between nondurable and durable consumpton. Conventonal theores on consumpton apply to the flow of consumpton. Snce durable consumpton s a replacement and addton to a captal stock, the conventonal approach s to only use non-durable consumpton n wealth effect studes. 16 However and as ponted out n Mehra (2001) total consumpton s the parameter of nterest when studyng movements n stock market prces. Partcularly stock market crashes are more lkely to lead to a postponement of durable consumpton whle the reducton of non-durable consumpton mght be of mnor mportance (see e.g. Romer (1990)). Combned wth the theory of adjustment costs to durable consumpton ths mght also be an explanaton for the hgher volatlty of durable consumpton than of non-durable consumpton (see Attanaso, 1999). Regardng the role of housng prces on aggregate consumpton, durable consumpton goods are 16 See Lettau and Ludvgson (1999, 2001) for a dscusson.

17 among the major enttes on whch resources rased by mortgage refnancng are spend as Brady, Canner and Mak (2000) show. However, one shortcomng wth usng total consumpton s that t also ncludes expendtures on housng servces (see secton II). Besdes, t mght be nsghtful to nevertheless dstngush between the two components of consumpton but for reasons of data avalablty ths s nfeasble n a panel of countres. Thrd, total dsposable ncome s used and not only labor ncome as would be suggested by the tradtonal permanent ncome hypothess. Data avalablty constrans us to do so. But also economcally t s more sensble to use total ncome rather than labor ncome. Ths s suggested by an extended vew of the lfe-cycle theory as Attanao (1999) advocates for and also sensble f households are on average more myopc than the lfe-cycle theory of consumpton would suggests (see Campbell and Mankw (1991) and more recently Mankw (2000)). Data for consumpton and dsposable household ncome was taken from the OECD Analytcal Database (OECD, 2001b). Data on stock market prce ndces are taken from Internatonal Fnancal Statstcs (IMF, 2001b), whch provdes us wth a relatvely broad coverage and enough tme seres observatons. Data on housng prce ndces are taken from the Bank for Internatonal Settlement s house prce database (BIS, 2001). 17 The data on housng prces are n annual frequency, and n order to nterpolate the data lnear nterpolaton was appled. Below ths nterpolaton method s tested n order to check the robustness of the estmated coeffcents, and t turns out that the nterpolaton method s n general not mportant for the man results found. All varables are n local currences and deflated by the consumer prce ndex taken from the OECD Analytcal Database. Consumpton and ncome are expressed n per capta unts. Populaton data s also taken from the OECD Analytcal Database and are lnearly nterpolated between annual observatons. Logs have been taken of all varables and hence the estmates reported below are the estmated elastctes of consumpton n changes of the rght hand sde varables. We therefore control for the dfferental sze of movements n stock market and house prces across countres. As Engle and Granger (1987) have ponted out, the long-run elastctes n equaton (3) cannot be consstently estmated f all the sngle varables have unt roots unless the varables n the long-run relatonshp are contegrated. Therefore one has to examne the statstcal propertes of the data and test whether a contegratng equlbrum relatonshp between consumpton, ncome and the two prce ndces exsts. Recently, tests for unt roots of ndvdual seres and contegratng relatonshps between seres have been developed for panel data The comparablty of these ndces across countres s dscussed n Grouard and Blöndal (2001, p. 36). 18 Baltag and Kao (2001) provde a revew.

18 B. Unt Root tests Among the varous tests proposed n the lterature, the Im, Pesaran and Shn (1997) (IPS) panel unt root test s sutable here. 19 The IPS t-bar test s based on an average of ndvdual country augmented Dckey Fuller (ADF) tests whle allowng for heterogeneous coeffcents under the alternatve hypothess and dfferent seral correlaton patterns across groups. Under the null hypothess all groups exhbt a unt root whle under the alternatve ths s not the case for some. 20 A more detaled dscusson of the test can be found n Baltag and Kao (2000). More specfcally, the followng model s tested for all varables. (5) yt, = θ + ρ yt 1, + ϕ j x, t j + υ, t. p j= 1 Under the null hypothess the autocorrelaton coeffcent ρ equals one for all whle under the alternatve t does not for at least one. The test statstc s then computed as the average of the ndvdual ADF statstcs as N (6) t = t _ = 1. ρ Im, Pesaran and Shn (1997) show that ths test statstc converges to a standard normal dstrbuton. Under the alternatve hypothess, the IPS panel t-bar test dverges to mnus nfnty and therefore the left tal of the standard normal dstrbuton s used to reject the null hypothess. Table 3 summarzes the results for the unt root tests of the four varables where the lag length p n equaton (5) s chosen by the Schwartz Bayesan crteron (SBC) wth a maxmum number of four lags. 21 These results confrm that the null of a panel unt root s not rejected for most of the seres. At a fve percent level of sgnfcance t s rejected for the housng prce data. Note that 19 The Chang and Kao (2001) NPT 1.2 program s used to mplement these tests. 20 Cho (2001) has proposed varous tests that relax upon ths restrctve formulaton of the alternatve hypothess n allowng for tests aganst the alternatve where some groups exhbt a unt root and others do not. 21 We also expermented wth the Akake (1973) nformaton crteron (AIC) whch showed that the results are not senstve to the lag selecton procedure. Also, we do not report ndvdual country specfc ADF statstcs here. These results are avalable from the authors upon request.

19 ths mght result from the lnear nterpolaton of annual housng prce data. Not surprsngly, the test statstc s when a tme trend s ncluded n equaton (5). Tabel 3: IPS (1997) panel unt root tests Varble Test Statstc Log of prvate consumpton (per capta) Log of household dsposable ncome (per capta) Log of IFSS stock market prce ndces Log of housng prce ndces ** Note: * (**) ndcate sgnfcance at the 10 (5) percent level. Test results are for the sample perod of C. Contegraton tests Pedron s (1999) tests for contegraton are used to test for the null of no contegraton n the panel of 16 OECD countres. 22 Pedron s tests allow for a consderable degree of heterogenety between groups wth regard to the ntercept, the error structure, and the contegratng relatonshp. Specfcally, Pedron s tests are resdual-based tests on a contegratng regresson of the followng type: (7) y, t = φ + λ j, t x j, t + ζ, t. k j= 1 The seven test statstcs presented n Pedron (1999) can be grouped nto two types of statstcs. The frst type of statstcs s based on poolng along the wthn-dmenson of the panel, the second s based on poolng along the between-dmenson (see Pedron, 1999, p. 657). The former s constructed by frst summng the numerator and the denomnator of the statstcs over the N groups dmenson separately and then dvdng whle the later are average statstcs n that they are constructed by frst dvdng the numerator and denomnator and then summng over the N group dmenson. Ths feature results n a more flexble correlaton pattern of the resduals from equaton (7) across and as dscussed n Baltag and Kao (2001) allows for an easer nterpretaton of the statstcs f the null s rejected for the second type of statstcs. 23 Therefore, Table 4 only summarzes results for the second type of Pedron panel contegraton tests for alternatve choces of contegratng x varables. 24 Under the null 22 We are grateful to Peter Pedron for provdng us wth hs program wrtten n RATS. 23 For a more detaled dscusson on the statstcs the reader s referred to Pedron (1999) and Baltag and Kao (2000). 24 Results for the frst type of tests also support the fndngs n table 4 and are avalable from the authors upon request.

20 hypothess all statstcs asymptotcally converge to a standard normal dstrbuton. Under the alternatve hypothess, the statstcs dverge to negatve nfnty. Therefore, the left tal of the normal dstrbuton s used. When estmatng the specfcaton n equaton (7) t s frst nvestgated whether a contegratng relatonshp already exsts among the x s on average across countres and then s tested the specfcaton of nterest. The results reported n Table 4 confrm a contegratng relatonshp between the varables n the specfcaton of nterest whle the contegratng relatonshp between the ndependent varables s rejected. Regresson Specfcaton y: Income (Y d ) x 1 : Stock Market Prce Index (SMPI) x 2 : Housng Prce Index (HPI) y: Consumpton (C) x 1 : Income (Y d ) x 2 : Stock Market Prce Index (SMPI) x 3 : Housng Prce Index (HPI) Table 4: Pedron (1999) panel contegraton tests Group ρ-statstc Group t-statstc (non-parametrc) Group adf-statstc ** ** ** Note: * (**) ndcate sgnfcance at the 10 (5) percent level. Test results are for the sample perod of D. Estmatng consumpton equatons The nvestgaton of the data propertes above mply that an estmaton of equaton (3) wth varables expressed n log levels provde relable nferences about the long and short term nfluences of the stock market and housng prce ndces and ncome on consumpton. The consderatons n secton II suggest that the mpact of stock market and housng prces on consumpton has changed over tme. To test ths hypothess equaton (3) was estmated frst for the sub-perod 1985Q1 to 2000Q4. 25 Regresson results for the combned sample and for the two groups, market-based and bank-based economes are summarzed n Table 5. In the table C s consumpton, Y d s dsposable ncome, SMPI s the stock market prce ndex, and HPI s the house prce ndex. The Schwartz crteron s appled for selecton of the lag length for each ndvdual country wth a maxmum number of four lags. Gven the dfferent lag lengths appled the short run coeffcent estmates are not representatve for all countres. A frst observaton from Table 5 s the relatvely low estmate of the ncome elastcty. An ncome elastcty less than one s suggested by economc theory n a lfe-cycle model nter 25 Effectvely, ths tme perod s shorter and dffers from country to country. The shortest tme perod s 42 observatons for Belgum and the longest are 59 observatons for Australa, Canada, France, Germany and the Unted States.

21 ala by Ando and Modglan (1963) Gal (1990, p. 439). 26 In a regresson wth ncome as the only dependent varable, a value close to (and nsgnfcantly dfferent from) one was found for the entre group sample. Ths suggests that the presence of the wealth measures n the specfcaton takes out some of the co-varaton of ncome and consumpton. For the combned sample, both the estmated house prce elastcty and the stock market prce elastcty are postve and sgnfcant whle the sze of the coeffcent estmate on house prces s about half the sze of the coeffcent estmate on stock market prces. Splttng the sample nto the two groups reveals that the estmated coeffcents are roughly smlar for the market-based economes whle they are substantally reduced for the bank-based economes. For stock market prces, the dfference between the two estmated elastctes (and ts t-rato) s 0.04 (-2.8). The estmated housng prce elastcty for the bank-based economes s also lower but the dfference n the estmates between the two groups of countres s nsgnfcant. A bt surprsngly, the mean group estmate of the adjustment coeffcent s hgher for the bank-based economes. The dfference s nsgnfcant but nevertheless pror consderatons would have suggested the opposte. However, the estmated adjustment coeffcents are postve for Australa and Canada and takng the average of the remanng three countres results n an average adjustment coeffcent of Ths exemplfes the senstvty of mean group estmaton to outlers n the small sample. Table 5 also shows that the mean group estmator on the frst short run ncome coeffcent s hgher for the market-based economes. Ths result was expected and s nsenstve to outlers. The dagnostc statstcs show that the un-weghted average R 2 of the ndvdual restrcted country regressons s around 50 percent. The ndvdual estmates of the ft however vary a lot rangng from 12 percent n the case of Canada to 91 percent n the case of Belgum (and ths pattern s ndependent of the estmated specfcaton). The lkelhood rato statstc that tests for the homogenety assumpton of the long run coeffcents s rejected for all groups. But as a comparson of the sum of the two ndependent lkelhood statstcs for the two groups wth the statstc for the combned panel reveals, the gan of splttng the sample s substantal and roughly 100. Equalty of the mean group and pooled mean group estmators s accepted by the Hausman test for all groups. Ths means that the pooled estmates are not based by the mposton of homogenety. 26 Its exact value depends among other thngs on the age structure of populaton and the dstrbuton of ncome.

22 Table 5: Estmatng consumpton functons ( ) Varables All Countres Bank-based Economes Market-based Economes Long run coeffcents Y d ** (0.0215) ** (0.0384) ** (0.0273) SMPI ** (0.0057) ** (0.0072) ** (0.0072) HPI ** (0.0113) * (0.0143) ** (0.0212) Averages of heterogeneous short run coeffcents Adjustment coeffcent ** (0.062) * (0.123) ** (0.043) d(c(-1)) (0.1) (0.182) (0.128) d(y d ) ** (0.083) ** (0.149) ** (0.079) d(y d (-1)) (0.080) (0.159) (0.079) d(smpi) (0.006) (0.007) (0.003) d(smpi(-1)) (0.002) n.a (0.005) d(hpi) (0.078) (0.100) ** (0.072) d(hpi(-1)) ** (0.057) (0.073) (0.095) Intercept ** (0.051) * (0.199) ** (0.025) Dagnostc Statstcs LR test (p-value) (0.00) (0.00) (0.00) Jont Hausman test (p-value) 5.94 (0.11) 4.26 (0.23) 5.52 (0.14) Average R Note: See Table A1 for a descrpton of the varables. The LR test tests for the homogenety restrcton of the long run coeffcents. The Hausman test tests for equalty of the MG and the PMG estmators. The Hausman test statstc s ndetermnate f the dfference between the varance-covarance matrces of the MG and PMG estmators s not postve defnte (see Pesaran et al. (1999) for more detals). The unrestrcted short run coeffcent estmates are the MG estmates under the restrcton of long run homogenety. Standard errors of the estmated coeffcents are n parenthess. * (**) ndcate sgnfcance at the 10 (5) percent level.

23 In sum, the evdence supports the hypothess of a postve responsveness of consumpton to changes n the two prce ndces on average over the last 15 years. At the same tme, the mpact of prce changes s hgher for the group of market-based economes than for the bank-based economes. E. Senstvty Analyss A natural queston to ask s f the coeffcent estmates shown n Table 5 have changed across tme and ths s nvestgated by estmatng equaton (3) for the perod 1960Q1 to 1984Q4. Agan, the effectve tme perod s shorter and the number of observatons range from 28 n the case of Ireland to 93 for Australa. As before, the Schwarz crteron s appled for the selecton of the lag length for each country regresson. Estmaton results are shown n Table 6. The results confrm a number of the hypotheses dscussed above. Frst, the estmated stock market prce has ncreased both for the entre sample as well as for both groups. In partcular, the stock market elastcty has been very low for the bank-based economes durng the frst observaton perod. As for the second observaton perod, the estmated elastcty s sgnfcantly hgher for the market-based economes. Second, the estmates of the housng prce elastctes are all negatve and sgnfcant at the 10 percent level. Ths suggests that durng the perod from 1960 to 1984 the negatve mpact of housng prces on consumpton domnated the postve mpact. The change n sgn of the housng prce coeffcent s a strkng evdence for chances n fnancal markets, partcularly n the mortgage markets durng the late 1980 s, whch has made t easer for households to access ther housng wealth (see Brady, Canner and Mak (2000) and Grouard and Blöndal (2001)). Moreover, the absolute value of the estmates are n both perods hgher for the market-based economes (whle the dfferences are nsgnfcant). Ths may ndcate that the relaxaton of market restrctons was stronger n the group of marketbased economes. Studes on the wealth effect on consumpton have almost exclusvely focused on a very long perods of tme, mostly rangng from The results found here suggest that such a vew s lkely to underestmate current exposure of the demand sde of the economy to shocks n the stock market. To nvestgate ths pont, estmates for the entre observaton perod 1960Q1 to 2000Q4 are presented. 28 At the same tme ths exercse provdes a robustness check of earler estmates. The lag length of the ndvdual country regressons s agan chosen by the Schwartz crteron. 27 An exempton s Mehra (2001) who extends hs observaton perod by fve years n order to nvestgate changes n the wealth effects across tme. 28 The effectvely ncluded number of observatons range from 84 for Ireland to 156 for Australa.