FIACE RESEARCH VOL., O., JAUARY 202 5 Fnancal Development and Economc Growth: Evdence from Heterogeneous Panel Data of Low Income Countres A. Qayyum, R. Sddqu and M.. Hanf Abstract Ths paper examnes emprcal relatonshp between fnancal development and economc growth whle ncorporatng the nflaton rate effect on fnancal development for low ncome countres. The study focuses on both the ndrect fnance and the drect fnance, separately as well as collectvely. We apply an approprate econometrc methodology of Wenhold (999) and ar-rechert and Wenhold (200) for causalty analyss n heterogeneous panel data. Two sets of results are reported. Frst, the relatonshp between fnancal development and economc growth from contemporaneous non-dynamc fxed effects panel estmaton can at best be nterpreted as mxed. egatve and statstcally sgnfcant estmates of coeffcent of the nflaton and fnancal development nteracton varable ndcate that fnancal sector development s actually harmful for economc growth when nflaton s rsng. Second, n contrast wth the recent evdence of Beck and Levne (2003), use of more approprate econometrc methodology of dynamc heterogeneous panel for causalty analyss and a refned model reveal that there s no defnte ndcaton that fnance spurs economc growth or growth spurs fnance. Our fndngs are n lne wth the Lucas (988) vew on fnance that the mportance of fnancal matters s very badly over-stressed n popular and even much professonal dscusson. Index Terms Fnancal Development, Economc Growth, Panel Data, Low Income Countres T I. ITRODUCTIO HERE s a longstandng tradton n economcs wth the ssue of fnancal development and economc growth (Krkpatrck 2000). Bagehot (873) and Hcks (969) argued that fnancal system played a crtcal role n gntng ndustralzaton n England by facltatng the moblzaton of captal for mmense works. Schumpeter (934) emphaszed the mportance of the bankng system n economc growth and hghlghted crcumstances when banks can actvely spur nnovaton and future growth by dentfyng and fundng productve nvestments. Wth the contrbutons of McKnnon (973) and Shaw (973), the relatonshp between fnancal development and economc growth has been an mportant ssue of debate, and durng the last thrty years these studes have fostered a fresh research nterest n ths relatonshp. However, obel Laureate, Lucas (988), dsmsses fnance as a major determnant of economc growth callng ts role overstressed by economsts. The road from the early work on fnance growth nexus to where we are now, however, has not been a straght one Manuscrpt receved January 27, 202. Ths work s a part of PhD dssertaton of M.. Hanf (the correspondng author; phone 0092-32- 243674; emal: muhammadnadeemhanf@yahoo.com,) completed under the supervson of A. Qayyum and R. Sddqu at Pakstan Insttute of Development Economcs, Islamabad. (Krkpatrck 2000). Levne (997) acknowledges that some recent work has extended our knowledge about the causal relatonshps between fnancal development and economc growth but fnds that the emprcal studes have not unambguously resolved the ssue of causalty. Resolvng the debate and advancng our understandng about the role of fnancal factors n economc growth, f any, wll help dstngush among competng theores of the process of economc growth (Levne, 2003). Khan and Senhadj (2000) stresses that the relatonshp between fnancal development and economc growth needs to be refned and approprate estmaton methods employed. Ths paper s an attempt on both of these fronts. There s both theoretcal and emprcal lterature suggestng that ncreases n the rate of nflaton can adversely affect fnancal market condtons (Khan, Senhadj, and Smth [2003]). Followng Harrs and Glman (2004) we assume that fnancal development effect (the coeffcent of the proxy for fnancal development) s a functon of nflaton rate, we ntroduce an nteracton (of fnancal development and nflaton) varable n the model relatng fnancal development and growth and thus take the proxy for fnancal development and nflaton rate both ndvdually as well as n product n the emprcal model we estmate. Furthermore, n most of the emprcal lterature ether ndvdual country tme seres analyss or cross-sectonal methodology has been used. Where tme seres analyss s confned to ndvdual country studes, cross-sectonal methodology has been crtczed on ts falure to control effectvely for cross country heterogenety. Some studes have used a panel GMM estmator to assess the fnance and growth relatonshp. Ths approach mproves upon pure cross-country work n varous respects. However, Kvet (995) shows that panel data models that use nstrumental varables estmaton often lead to poor fnte sample effcency and bas. Consderng the heterogeneous nature of the relatonshp between fnancal development and economc growth across countres, we use an approprate methodology of panel causalty analyss for heterogeneous panel data. Our man objectve s to nvestgate the causal relatonshp between fnancal development and economc growth by usng panel data of 9 Low Income Countres (LIC) for the perod 973-2002. Ths paper contrbutes to the exstng lterature relatng to fnance growth nexus n at least two ways. Frst, ths study uses an advanced and approprate econometrc methodology for causalty analyss n heterogeneous panel data. Second, we use a refnement n the econometrc model, generally used for emprcal research related to fnance growth
FIACE RESEARCH VOL., O., JAUARY 202 6 nexus, by takng care of nflaton rate effect on fnancal development. Our emprcal fndngs suggest that the relatonshp between fnancal development and economc growth from contemporaneous non-dynamc fxed effects panel estmaton can at best be nterpreted as mxed. egatve and statstcally sgnfcant estmates of coeffcent of the nflaton and fnancal development nteracton varable ndcate that fnancal sector development s actually harmful for economc growth when nflaton s ncreasng n low ncome countres. Furthermore, n contrast wth the evdence of Beck and Levne (2003), use of more approprate econometrc methodology of dynamc heterogeneous panel for causalty analyss and a refned model reveal that there s no defnte ndcaton that fnance spurs economc growth or growth spurs fnance. Our fndngs are n lne wth the Lucas (988) vew on fnance that the mportance of fnancal matters s very badly over-stressed n popular and even much professonal dscusson. ext Secton revews some of the theoretcal and the emprcal work relatng to the relatonshp between fnancal development and economc growth. Secton 3 begns by embarkng on the model we use n our emprcal work n ths paper. Here we show how we attempt to follow the advce of Khan and Senhadj (2000) to refne the relatonshp between fnancal development and economc growth. Here we detal the data ssues related to the emprcal work n ths paper and then we dscuss the methodology of dynamc heterogeneous panel approach of ar-rechert and Wenhold (200) for causalty analyss. In Secton 4 we provde emprcal results. Last Secton, whle concludng, gves a summary of the overall pcture and throws lght on polcy mplcatons. II. REVIEW OF LITERATURE Economsts hold startlngly dfferent vews about the mpact of fnancal sector, ncludng banks and markets on long- run economc growth. The vews over fnance-growth nexus can be grouped nto four schools of thoughts. Frst, fnance promotes growth. Banks are the best engnes that ever were nvented for creatng economc growth [Bagehot (873), Schumpeter, (934), Hcks (969), McKnnon (973), Shaw (973)]. Second, fnance hurts growth. As s explaned n Levne (2003) ths followers of ths school has the opnon that banks have done more harm to the moralty, tranqulty, and even wealth of ths naton than they have done or ever wll do good. It s argued that although fnancal nsttutons facltate rsk ameloraton and effcent allocaton of resources, ths wll not necessarly boost growth because better fnance means greater returns to savng (whch may lower the savngs rates) and lower rsk (whch may also result n lower savngs) and both may yeld lower growth. Thrd, fnance follows growth - where enterprse leads fnance follows [Robnson (952)]. Economc growth creates demand for fnancal arrangements and fnancal sector responds automatcally to these demands. Fourth, fnance doesn t matter. Accordng to Lucas (988) economsts overstress the role of fnance n economc growth. Emprcal work on fnance and growth has been done n varous dmensons. A number of papers studed the ssue n a cross-country framework. A lot of studes made purely tmeseres nvestgatons. Some others used cross-sectonal approach. Where tme seres analyss s confned to ndvdual country studes, cross-sectonal methodology has been crtczed on ts falure to control effectvely for cross country heterogenety. Studes lke Levne, Loayza and Beck (2000) and Beck, Levne, and Loayza (2000) have used a panel GMM estmator to assess the fnance and growth relatonshp. The use of panel GMM estmator mproves upon pure crosscountry work n varous respects, However, Kvet (995) shows that panel data models that use nstrumental varables estmaton often lead to poor fnte sample effcency and bas. Some studes allow for heterogenety but restrcted to the ntercept and not permtted n the slope coeffcents. Pesaran and Smth (995) shows that f n a dynamc panel data model slope coeffcents are assumed to be constant but n fact they vary across countres, the tradtonal panel estmators (fxed effects or GMM estmators) yeld nconsstent estmates. Furthermore studes that use perod average, whereby tmeseres characterzng each varable s collapsed nto sngle observaton, are also crtczed because of possbly nonstatonary nature of these data. Van den Berg and Schmdt (994) and Van den Berg (997) argue that nonstatonarty of many tme seres makes the use of such perod averages napproprate. Varables are often nonstatonary contanng stochastc or determnstc trends. Such varables ether have a mean that s changng through tme or have expandng varance. Regresson estmates from cross secton data created from averages of such tme seres are not well suted for characterzng prospectve long run relatonshp among varables. One possble soluton to the problems dscussed above s the use of tme-seres, cross-secton panel data estmaton. Ths allows the researchers to control for country-specfc, tmenvarant fxed effects, and nclude dynamc, lagged dependent varables whch can also help to control for omtted varable bas. The ablty to lag explanatory varables may also help control for endogenety bas. But the tradtonal panel data fxed effects estmators (FEE) mposes homogenety assumptons on the coeffcents of lagged dependent varables when n fact the dynamcs are heterogeneous across the panel. Ths msspecfcaton can lead to serous bases that cannot be remeded wth nstrumental varable estmaton. Then we have Mean Group Estmators of Pesaran and Smth (995). The MG estmator gves us an un-weghted average of the country specfc coeffcents and s thus partcularly senstve to outlers. A smple Random Coeffcent (RC) estmator, on the other hand, calculates a varance weghted average, but unfortunately t s not possble to estmate dynamc RC models [ar-rechert and Wenhold (200)]. The Mxed Fxed Random (MFR) effects approach of Hsao et al (989) whch has been exploted by Wenhold (999) and Levne (997), Levne (2003), and FtzGerald (2006) provde a comprehensve survey n ths regards
FIACE RESEARCH VOL., O., JAUARY 202 7 ar-rechert and Wenhold (200) falls somewhere n between the two extremes of FEE and MGE n terms of allowng for heterogenety. Ths method mposes more structure on the coeffcent values of the exogenous varables than the MGE (after all, f the relatonshp s completely dosyncratc across countres then t s dffcult to meanngfully nterpret the results from an economc or polcy perspectve). As compared to FE estmator wth small T, MFR coeffcents approach produces consderably less based parameter estmate [ar-rechert and Wenhold (200)]. Wenhold (999) shows that the MFR coeffcents model performs well compared to nstrumental varables (GMM) approaches as well. In addton, the MFR coeffcents model has other features whch make t deally suted to the task of testng for causalty n heterogeneous panel data sets. In partcular, Wenhold (999) allows for a dstrbuton of causalty across the panel, rather than mposng an assumpton that causalty occurs everywhere, or nowhere, n the panel. We may use the dstrbutonal nformaton to gan a general dea of the degree of heterogenety. The combnaton of a less-based mean estmate and an dea of the degree of heterogenety gves a researcher more nformaton about the underlyng process than tradtonal panel causalty tests. ausser and Kugler (998) uses heterogeneous panel data approach but ths study s only for a lmted number of developed countres of OECD and that after dong panel contegraton analyss ths study uses ndvdual country Granger causalty methodology for causalty analyss. Chrstopoulos and Tsonas (2003) use panel unt root tests and panel contegraton analyss to examne the relatonshp between fnancal development and economc growth n ten developng countres. But for causalty analyss they use tmeseres tests to yeld causalty nferences wthn a panel context. After showng that the relatonshp between fnancal development and economc growth s heterogeneous across countres, we use an approprate methodology of panel causalty analyss for heterogeneous panel data. Other than the methodologcal ssues the lterature on fnance growth relatonshp has gnored (except Harrs and Glman (2004)) the nflaton rate effects on fnancal development. Char, Jones and Manuell (996) argue that fnancal regulatons and ther nteracton wth nflaton have substantal effects on growth. There are some other studes whch dscuss how nflaton s lnked wth the fnancal sector. Cho, Smth, and Boyd (996) argue that nflaton reduces real return to savngs and makes more severe the adverse selecton problems n captal markets nducng a hgh degree of credt ratonng and have negatve mpact on fnancal development. In a monetary growth model Huybens and Smth (999) show that, at the steady state, hgher rates of money creaton reduces the real return on all assets and, under certan condtons, lead to a reducton n the volume of tradng n equty markets. Boyd, Levne and Smth (200) consder alternatve theory regardng the relatonshp between nflaton and fnancal sector performance and that s a fscal story. Governments combne hgh nflaton wth varous restrctons on the fnancal sector to help fund expendtures. As a result, they have both poorly developed fnancal systems and hgh nflaton. Barro (997) fnds that permanent ncreases n the rate of nflaton have sgnfcant negatve effects on the long run real growth rates. Khan, Senhadj, and Smth (2003) asserts that the real effects of nflaton derves from the consequences of nflaton for fnancal markets condtons. Thus analyss of fnance-growth relatonshp s ncomplete wthout consderng the nflaton rate effects on fnancal development. We contrbute by ncorporatng a fnancal development and nflaton nteracton varable n the model relatng economc growth and fnancal development, though n lnear manner. 3. Model III. MODEL, DATA, AD ECOOMETRIC METHODOLOGY Followng Kng and Levne (993) the growng body of emprcal work models the relatonshp between fnancal development and economc growth usng the lnear regresson equaton gve below. G F X e (3.) where G s the proxy for economc growth, F s the proxy for fnancal development and X s the set of condtonng nformaton to control for other factors assocated wth economc growth. e s error term. In sprt of model (3.) above, ths paper starts wth a smlar model for our heterogeneous panel data G t where F X e (3.2) t,2,...,, and t,2,...,t t t. refers to the number of countres, and T refers to the number of observatons over tme for country n the panel. G denotes proxy for economc growth and F denotes proxy for fnancal development. The parameter s the country specfc ntercept, or fxed effect parameter, whch of course s also allowed to vary across ndvdual countres 2. Slope coeffcent s also allowed to vary across natons to take nto account the possble heterogenety 3 among the varous countres n a panel. 2 Country specfc fxed effects heterogenety s assumed on the bass of dfferences n technology. 3 Even though we have grouped countres accordng to ther level of ncome, there s stll heterogenety between the countres n the panel. There are dfferent sources of such heterogenety lke dfferences n populaton sze, dfferences n poltcal and economc nsttutons, dfferences n geography,
FIACE RESEARCH VOL., O., JAUARY 202 8 Followng the recent lterature on the analyss of fnancal development and economc growth, four varables are ncluded n the condtonng set to control for other factors assocated wth economc growth, n addton to the ntal real GDP per capta. These nclude measures of (n)stablty (nflaton), fscal polcy (government consumpton to GDP rato), trade polcy (overall trade to GDP rato), and educaton (secondary school enrollment rato). We use secondary school enrollment rato wth 5 year lag because people n secondary school at tme t wll generally be enterng the labour force n some latter tme and wll not be productve for 5 years or so. We proxy the ntal level of ncome by real GDP per capta and we use ths wth year lag as we take annual growth rates on LHS of the regresson equaton. Thus we relate the real per capta growth (GRGPC) to ntal level of educaton, the ntal level of GDP, rate of nflaton (IFL), the rato of government consumpton to GDP (GCGR), and the rato of exports plus mports to GDP (TRGR). Prevous studes have shown that these varables correlate sgnfcantly wth real per capta GDP growth (Barro, 997). GRGPC s negatvely related to Inflaton (IFL), government consumpton to GDP rato (GCGR), and ntal level real per capta GDP (RGPC); and s postvely related to overall trade to GDP rato (TRGR) and ntal level of secondary school enrollment rato (SSER). We can wrte (3.2) as: GRGPC 3 6 t t RGPC F IFL t 4 GCGR TRGR t t t 2 5 t SSER t5 t are assumed to be dosyncratc errors. (3.3) Before addng the proxy for fnancal development we wll estmate general model that s contemporaneous non dynamc fxed effects panel model of economc growth by regressng the GRGPC on all ts determnants: IFL, GCGR; TRGR; SSER; and ntal RGPC. Droppng the nsgnfcant varables (f any) from among these we wll be left wth a parsmonous basc model for economc growth. To ths basc model we wll add the proxy for fnancal development and have an ntermedate model to see what fnancal development contrbutes to the economc growth. In order to capture the (adverse) mpact of ncreases n the rate of nflaton on fnancal market condtons, followng Harrs and Glman (2004), we assume that fnancal development effect, s a (lnear) functon of nflaton rate effect. A smply way to allow for such an effect s to wrte as we get. 7 IFL t. By substtutng t back nto (3.3) GRGPC 3 6 t t RGPC F IFL t 4 GCGR TRGR 7 t t 2 5 ( F * IFL) t t SSER t5 t (3.4) In ths way, we arrve at our fnal model whch ncludes the proxy for fnancal development and nflaton both ndvdually as well as n product to our basc model. To provde a sense of whether there s a causal relatonshp between economc growth and the fnancal development we turn to the dynamc panel form of (3.4) n whch GRGPC s modeled as a functon only of lags of tself and of all other rght hand sde varables n (3.4). That s: GRGPC F 4 6 t t TRGR RGPC GRGPC IFL 2 t t2 5 7 t SSER 3 t6 t GCGR ( F * IFL) t t t (3.4a) To take care of the lnear nfluences of the remanng rghthand sde varables n (3.4a) on the canddate causal varable, we orthogonalze the canddate causal varable and thus our fnal model n dynamc form becomes: GRGPC F 4 6 o t t TRGR RGPC GRGPC IFL 2 t t2 5 7 t SSER 3 t6 t GCGR ( F * IFL) t t t (3.5) All the varables n the model are assumed to be statonary. 3.2 Data One of the mportant ssues pertanng to the analyss of the fnance growth nexus s that of selecton of proxes to measure fnancal development and economc growth. For economc growth, followng Kng and Levne (993), we use the real per capta GDP growth. We denote t by GRGPC 4. There does not exst a sngle accepted emprcal defnton of fnancal development [Beck, Demrguc-Kunt, and Levne (200)]. Prevous studes have used varous ndcators of fnancal ntermedary and stock market sze and actvty to measure the fnancal development. Followng Kng and Levne [993]; Levne and Zervos [996]; and Beck, Demrguc-Kunt, and Levne [200] we use varous ndcators of sze and actvty of the ndrect as well as drect fnance as a proxy for fnancal sector development. We also combne the sze and actvty measures of drect and ndrect fnance to proxy the overall fnancal sector development. As a whole, we have sx measures of fnancal sector development whch wll be used one by one n ths study. These measures are dscussed below. and dfferences n culture. Thus we take slope coeffcents to be heterogeneous n the causalty analyss we do. 4 For complete lst of data varables and sources of data see Appendx.
FIACE RESEARCH VOL., O., JAUARY 202 9 The sze of ndrect fnance To measure sze of the fnancal ntermedares we use currency plus demand and nterest bearng labltes of banks and other fnancal ntermedares dvded by GDP whch s generally known as lqud labltes to GDP rato. We denote t by LLGR. Ths s the broadest avalable ndcator of fnancal ntermedaton. The proxy for sze of the fnancal sector may not accurately measure the functonng of the fnancal system. Here we also consder a measure whch takes nto account the actvty of the fnancal sector. The actvty of ndrect fnance To measure the actvty of fnancal ntermedares we consder prvate sector credt by depost money banks and other fnancal nsttutons to GDP rato. We denote t by PCGR. There s a postve sgnfcant correlaton between real per capta GDP growth and the extent to whch loans are drected to the prvate sector (Levne, 997). The sze of drect fnance As an ndcator of the sze of drect fnance we use the stock market captalzaton to GDP rato, denoted by MCGR, whch equals the market value of lsted shares dvded by GDP. The actvty of drect fnance As an ndcator of the actvty of drect fnance we use total value of the shares traded n the stock market to GDP rato, dented by VTGR. The sze of overall fnancal sector To have an overall sze measure of the fnancal sector we combne the two sze measures and call t as fnancal depth to GDP rato, denoted by FDGR, whch s sum of the LLGR and MCGR. The actvty of overall fnancal sector To have an overall actvty measure of the fnancal sector we combne the two actvty measures and call t as fnancal actvty to GDP rato, denoted by FAGR, whch s sum of the PCGR and VTGR. We have two types of measures: frst, the rato of a stock varable to a flow varable that s LLGR; and second the ratos of two flow varables that s PCGR. Whereas stock varables are measured at the end of a perod, the flow varables are defned relatve to a perod. Ths presents a problem n the frst type of measures, both n terms of correctng tmng and n terms of deflatng correctly. To address these problems, we deflate the end-of-year fnancal aggregates by end-of-year consumer prce ndces (CPI e ) and deflate the GDP seres by annual consumer prce ndex (CPI a ). Then we compute average of the real fnancal aggregate n year t, and t and dvde ths average by real GDP measured n year t. The end-of-year CPI s ether the value for December, or, where December-CPI s not avalable, for the last quarter. The formula, for LLGR, s the followng: LLBt LLGR 0.5* CPI e, t LLBt CPI e, t GDPt CPI a, t (3.6) In case of the rato of two flow varables measured n the same tme deflatng s not necessary. We use a dataset of 9 LIC countres lsted n the Appendx. The countres have been selected from the overall lst of Low Income Countres for whch World Bank publshes ncome classfcaton n ts World Development Indcators 5. The countres ncluded are selected on two crterons: there s data both on ndrect as well drect fnance; and that data are avalable for at least 5 observatons for both type of fnance. The tme dmenson of the dataset s that we use annual data startng from 973 whch s the year n whch heroc peces of work by MacKnnon and Shaw were publshed. 3.3 Methodology In recent emprcal research there has been an upsurge of nterest n the development and use of methods for nonstatonary panels, ncludng panel unt root and panel contegraton tests. In partcular, there exst some nterestng contrbutons on heterogeneous panels. Before movng to regresson analyss we test for statonarty of the varables we use. For ths purpose we apply Im Pesaran and Shn (2002) panel unt root test for dynamc heterogeneous panels whch s based on the average (across countres) of the (augmented) Dckey-Fuller statstcs. 3.3. Panel Unt Root Tests Frst we consder the calculaton of ndvdual country unt root (augmented) Dckey-Fuller test-statstcs denoted by. The process starts by estmatng the followng (augmented) Dckey-Fuller regresson y t t y t p j y t j t t T (3.7) for each of the cross sectonal unt n the panel and estmatng the value of the t-statstcs and then averagng them. The decson of the number lags of the dependent varables to be ncluded depends on statonarty of the error term and here we wll be usng step down procedure by startng at maxmum lag of four. The null hypothess for the IPS panel unt root test s H : 0 for all (3.8) 0 aganst the alternatves H : 0, for,2,...,, and 0, 2,...,, for (3.9) 5 The World Development Indcators for year 2002 has been used. The country classfcaton s based on World Bank estmates of per capta GI durng 2000. Countres for whch estmates of per capta GI are US$ 755 or less are classfed as Low Income Countres.
FIACE RESEARCH VOL., O., JAUARY 202 20 Ths formulaton of alternatve hypothess allows for dfferng across groups It allows for some (but not all) of the ndvdual seres to have unt roots under the alternatve hypothess. Essentally, the IPS test averages the ADF ndvdual unt root test statstcs that are obtaned from estmatng (3.7) for each (allowng each seres to have dfferent lag length, p f necessary); that s: t bar T whch s referred to as t T t bar statstc. (3.0) IPS shows that under the assumpton that,,2,...,, t,2,..., n (3.7) are ndependently t T and dentcally dstrbuted for all and t wth mean zero and fnte heterogeneous varances, 2 are ndependently (but not dentcally) dstrbuted for T 9 and that the standardzed t bar statstc t bart E( t ) T, Z (3.) tbar VAR( t ) T t T converges to standard normal varate 6 ndefntely. as ncreases Whle testng for panel unt roots at level we take both unobserved effects and heterogeneous tme trend n our equaton as n equaton (3.7). If n no case we can reject the null hypothess that every country has a unt root for the seres n levels, we then test for a unt root n frst dfferences. If we fnd that the man varables of nterest that s the proxy for economc growth and that for fnancal development are of same order of ntegraton and that none of the control varables s of hgher order than that of the dependent varable then we move towards testng for possble contegraton between fnancal development and economc growth. Otherwse we say the order of ntegraton of seres of nterest does not support to move to contegraton analyss. Snce, on the bass of the evdence documented n Lee, Pesaran and Smth (997) and n Cannng and Pedron (999), we expect our dependent varable (growth n real GDP per capta) and the varables of nterest to be statonary and hence we do not expect to be n need of the applcaton of panel contegraton analyss and thus we do not dscuss t. 3.3.2 Contemporaneous Fxed Effects Model Estmaton After ensurng the statonarty of the varables of nterest we move to the estmaton process. Assumng the slope coeffcents to be homogeneous we estmate model n (3.4) 6 IPS standardzed ther test statstcs based on smulatons of the mean and varance (wth dfferent values obtaned dependng on the lag length used n the ADF tests and the value of ). These smulated values are gven n IPS (2002). usng fxed effects methodology n wth the country specfc fxed effects are wped out and each varable s replaced by ts devaton from cross-sectonal means. To ths transformed data OLS method s appled. However, for calculatng the estmated t-values robust varance estmator proposed n Arellano (987) s used to address the ssue of possble heteroscedastcty. 3.3.3 Panel Causalty Analyss for Dynamc Heterogeneous Panel Data Model We then examne the drecton of causalty, f there s any, between fnancal development and economc growth usng an advanced and approprate econometrc methodology of panel causalty analyss for dynamc heterogeneous panel data models gven by Wenhold (999) and ar-rechert and Wenhold (200). Ths methodology s based upon mxed fxed random (MFR) coeffcents approach of Hsao et al (989). We examne the drecton of causalty between fnancal development and economc growth, and vce versa, usng methodology ntroduced by Wenhold (999) and ar- Rechert and Wenhold (200) for causalty analyss n heterogeneous panel data whch s based upon mxed fxed random (MFR) coeffcents approach of Hsao et al (989). Followng ar-rechert and Wenhold (200), we consder the model y t where y x x o t t 2 2t t (3.2) j j. 2 (, j j j ) s a random dsturbance. Here. The varable o x t denotes the orthogonalzed canddate causal varable after the lnear nfluences of the remanng rght-hand sde varables have been taken nto account. Orthogonalzaton 7 provdes for approprate nterpretaton of the estmated varances by makng sure that the coeffcents are ndependent. Unobserved effects ( ) and the coeffcent of the lagged dependent varable are fxed and country specfc; and the coeffcents on the exogenous explanatory varables are drawn from a random dstrbuton wth mean and fnte varance 8. j Let Y be dependent varable; Z contans vector of s for ntercept, and the lagged dependent varables,.e. those for whch we have fxed coeffcents; X has orthogonalzed causal canddate varable, and other control varables,.e. all other rght hand sde varables for whch we have random coeffcents. We denote the vector of all the rght hand sde varables (ncludng unobserved effects) by W,.e. t 7 For the purpose of orthogonalzaton of the lagged causal canddate varable, we regress the lagged causal canddate varable upon constant, lagged dependent varable and all other explanatory varables. We use errors of ths regresson as orthogonalzed (lagged) causal canddate varable. 8 Wenhold (999) explans why to model ths partcular combnaton of fxed ndvdual specfc coeffcents on the lagged dependent varable and random coeffcents on the lagged ndependent varables
FIACE RESEARCH VOL., O., JAUARY 202 2 contans all the varables that are n Z and X. Let 2 be vector of fxed coeffcents (whch are f n number) and be vector of random coeffcents (whch are r n number). Let denotes the vector of all fxed as well as random coeffcents. We estmate by [ [ X X X Y X Z ( Z Z ) Z X ] * X Z ( Z Z ) Z Y ] (3.3) whch s the GLS estmate of under MFR coeffcents assumpton. Here 2 X X ˆ I ) (3.4) and ( r T 2 ˆ s OLS estmate of error varance of ndvdual Y upon W,.e. Y W error, and r regresson of s the covarance matrx whch s sub-matrx for random coeffcents from where ( ˆ ˆ )( ˆ ˆ) (3.5) ˆ s the OLS estmate from ndvdual regresson of Y upon W,.e. of such Y W error and ˆ s the average ˆ s for the ndvduals countres n the panel. We estmate ndvdual coeffcents under MFR effects approach by [ { X 2 X ˆ [ { X 2 X ˆ and X Z ( ZZ ) X Z ( ZZ ) ( ZZ ){ Z( Y ZX } } ˆ ZX ] X )} 2 We have r ] r * (3.6) (3.7) X ut Yt 2Zt and mean square error s 2 2 t ( u ) { T ( f * r)} and Var ( errors ( t 2 ) ( W W) from whch we can have standard ) of the MFR effects estmates. For causalty testng, we have to buld confdence nterval around zero 9 (here we wll use the frst element n the vector whch s [] ) for whch the lower and upper bounds are gven below: Lower Bound (Confdence Interval): {( 2) * [ ]} r [] Upper Bound (Confdence Interval): { 2 * [ ]} r [] The area that falls wthn ths nterval s nterpreted to correspond to observatons that are not sgnfcantly dfferent from zero 0. IV. EMPIRICAL AALYSIS 4. Statstcal propertes of the data Table 4.A shows summary statstcs of varous varables we have used n ths study. An mportant analyss from ths table relates to the comparson of wthn-country standard devaton and between-country standard devaton for all the varables we have. Ths analyss reveals that for all the varables most of the varablty n the data occurs between countres whch shows the heterogenety between the countres for all these varables. The par-wse correlatons matrx s presented n the Tables 4.B. The growth n real per capta GDP correlates postvely wth secondary school enrollment rato n addton to all the ndcators of fnancal development. In accordance wth the Barro (997) s fndng that bg government s bad for growth, government consumpton to GDP rato s negatvely correlated to real GDP per capta growth. Smlarly, n lne wth the Barro s results, the rate of nflaton has negatve correlaton wth real GDP growth rates. Only unexpected sgn s that of the correlaton between openness (proxy by TRGR) and real per capta GDP growth and that may be because ether the trade n the low ncome countres s not fully lberalzed or the ntal condtons for trade lberalzatons were not met when the lberalzaton process started n such countres. Fnally, nflaton rate s negatvely correlated wth all the measure of fnancal development except MCGR whch s very near to zero. An nterestng feature s that the (absolute) correlaton coeffcents between nflaton and fnancal development, n most of the proxes of fnancal development, are hgher f we compare them the correlaton coeffcents between fnancal development and economc growth. 9 Theoretcally speakng; for populaton parameter under the null hypothess that [] s zero. 0 For panel causalty analyss, we use SAS verson of the program (whch calculates estmate of the coeffcent of the causal varable, ts standard error, the confdence nterval and the estmate of the varance of the estmated random coeffcent) developed by Dana Wenhold and avalable on her ste lnked wth that of London School of Economcs, UK. Ths SAS program does not orthogonalze the canddate causal varable, however, we dd t.
FIACE RESEARCH VOL., O., JAUARY 202 22 4.2 Im-Pesaran-Shn Panel Unt Root Test In the Table 4.2 we present the results of Im-Pesaran-Shn (2002) panel unt root (IPS PUR) test on all varables used n ths study. It s evdent that all the varables are statonary at level except LLGR whch s nonstatonary and becomes statonary after frst dfferencng. Whle testng for panel unt roots at level we take both unobserved effects and heterogeneous tme trend n our equaton as n equaton (3.7) n Secton 3. One may argue, partcularly n the case of growth rate of real GDP per capta and nflaton, that there s no reason to nclude the heterogeneous tme trend whle testng for unt root but t s observed whle dong analyss that the orders of ntegraton of growth and nflaton are nsenstve to whether or not we nclude the heterogeneous tme trend. 4.3 Contemporaneous Fxed Effects Model Estmaton In order to explore the relatonshp between fnancal development ndcators and economc growth, we start wth the estmaton of contemporaneous non dynamc fxed effects panel estmaton of the general form whch relates growth rate of GDP per capta to nflaton, government consumpton to GDP rato, overall trade to GDP rato, (ntal) secondary school enrollment rato and the (ntal) level of per capta GDP 2. We drop the varables wth nsgnfcant coeffcents and arrve at the basc model. To the basc model we nclude the proxy for fnancal development and have ntermedate model. Our fnal model s one where we have nflaton and fnancal development both ndvdually and n product form ncluded n the basc model. 4.3. Indrect Fnance and Economc Growth Table 4.3. gves the results of smple contemporaneous non dynamc fxed effects panel estmaton. The results show that all the four explanatory varables n the basc model have approprate sgn. These results are consstent wth standard growth theory. Inflaton depresses growth due to ts adverse mplcatons for workng markets lke rsng prce varablty whch makes the log term plannng dffcult. Government consumpton s observed to affect growth negatvely. It may be because of well know neffcences assocated wth the larger sze of the government. egatve sgnfcant coeffcent of ntal level of per capta GDP s n accordance wth the condtonal convergence growth theores. Intal secondary school enrollment has postve effect on growth rate of GDP per capta. As regard to mpact of fnancal development on growth, the results show that coeffcents of the proxes of both the sze and the actvty of fnancal sector are negatve and statstcally nsgnfcant. However, when the nteracton of fnance wth nflaton s ntroduced, then the coeffcents of the proxes of both the sze and the actvty of the fnancal sector become postve but reman nsgnfcant. We wll be usng frst dfferences of LLGR n the panel causalty analyss n next secton 2 All the varables are n log form. From here we observe that for the LIC fnance does not matter for growth and the data we use support the Lucas vew and our results are n lne wth the fndngs of Barro and Sala-Martn (2004). It s nterestng to note that both the nteracton varables are hghly sgnfcant and have negatve sgn. It mples economc growth returns of fnancal sector development actually declnes wth the ncreased nflaton for LIC. In other words a negatve sgnfcant coeffcent on the nteracton term means that fnancal development accelerates the negatve effect of nflaton on growth rate of GDP per capta. Another mportant observaton s that the magntude of the partal effect of nflaton on growth rate of GDP per capta s much larger n the fnal model as compared to that n the basc model whch shows that nflaton may be a much serous ssue n fnancally developed stage of economy as ts mpact s larger than that can be at the lesser (fnancally) developed stage of the economy n case of Low Income Countres. 4.3.2 Drect Fnancal Development and Economc Growth ow we examne the lnks between economc growth and fnancal development consderng drect sources of fnance,.e. stock market. We wll follow all the same step as we dd for ndrect fnance. Table 4.3.2 gves the results of smple contemporaneous non dynamc fxed effects panel estmaton. The basc model s the same as we dscussed above. By ncludng proxes for drect fnance as regressors we re-estmate the smple contemporaneous non dynamc fxed effects panel regresson and results are shown n the column under ntermedate model. The coeffcents of the proxes of both the sze and the actvty of fnancal sector are statstcally nsgnfcant n the ntermedate model whch becomes sgnfcant n the fnal model when we nclude nteracton varables. Ths shows that sze and actvty of drect fnance has strong postve relatonshp wth economc growth for LIC. The nteracton of nflaton wth sze of drect fnance has a negatve sgnfcant coeffcent whch has the nterpretaton that growth return of ncrease n the sze of fnancal sector decreases wth nflaton. If we consder the postve sgnfcance of the sze measure of drect fnance we cannot gnore the fact that the magntude of the estmated coeffcent of the nteracton varable s larger than that of the sze of the drect fnancal development and hence even wth the low level of nflaton the total mpact of fnancal sector development has negatve mpact on growth rate of GDP per capta. The nteracton of nflaton wth actvty of drect fnance has a negatve sgnfcant coeffcent at 0% level. All the other explanatory varables have expected sgns n the fnal model as well as n basc and ntermedate models whch are consstent wth the theory. Here also, lke n case of ndrect fnance, we observe that the magntude of the partal effect of nflaton on growth rate of GDP per capta s much larger n the fnal model as compared to that n the basc model. It agan shows that nflaton may be a much serous ssue n fnancally developed stage of economy as ts mpact s larger than that can be at the lesser (fnancally) developed stage of the economy n case of Low Income Countres.
FIACE RESEARCH VOL., O., JAUARY 202 23 4.3.3 Overall Fnancal Development and Economc Growth Table 4.3.3 gves the results of smple contemporaneous non dynamc fxed effects panel estmaton for the panel. The basc model s the same as we have dscussed already. By ncludng proxes for overall fnance as regressors we re-estmate the smple contemporaneous non dynamc fxed effects panel regresson and results are shown n the column under ntermedate model. The coeffcents of the proxes of both the sze and the actvty of overall fnancal sector are statstcally nsgnfcant n the ntermedate model. When the nteracton of fnance wth nflaton n ntroduced, then the coeffcent of the proxy of the sze turns to be postvely sgnfcant and that of the actvty of overall fnancal sector reman nsgnfcant. It s nterestng to note that both the nteracton varables are hghly sgnfcant and have negatve sgn. It mples that economc growth returns of further fnancal sector development actually declnes wth the ncreased nflaton n the case of Low Income Countres. Lke n above cases of ndrect and drect fnance we observe that the magntude of the partal effect of nflaton on growth rate of GDP per capta s much larger n the fnal model as compared to the basc model. It agan shows that nflaton may be a much serous ssue n fnancally developed stage of economy as ts mpact s larger than that can be at the lesser (fnancally) developed stage of the economy n case of Low Income Countres. 4.4 Panel Causalty Analyss Entre analyss of contemporaneous non-dynamc fxed effects panel estmaton presented above s based on underlyng assumpton about the homogenety of the relatonshps n questons across countres n the respectve panels. Heterogenety s restrcted to the ntercept but s not permtted n the slope coeffcents. ow we wll be movng to the causalty analyss based on our dynamc model. We apply Rechert and Wenhold (200) panel causalty method to our fnal model n dynamc form n equaton (3.5). In ths model the coeffcent on the lagged dependent varable s country specfc and the coeffcents on the other RHS varables are allowed to have normal dstrbuton. We choose a lag length of one due to the large number of explanatory varables and relatvely short tme seres for each country. The results are presented n Table 4.4 where we report the mean of the estmated coeffcent, standard error of the mean of the estmated coeffcent, and the varance estmate of the estmated coeffcent on the causal varable. For causalty testng, we buld confdence nterval around zero (here we wll use the frst element n the estmated vector whch s [] whch s to be tested to be zero) to test for mean of the estmated coeffcent on the causal varable to be zero. The lower and upper bounds are gven below: LB (Confdence Interval): {( 2) * [ ]} r UB (Confdence Interval): [] } r { 2 * [ ] [] The area that falls wthn ths nterval s nterpreted to correspond to observatons that are not sgnfcantly dfferent from zero. We do not fnd evdence that the mean of the estmated coeffcent of the orthogonalzed causal canddate varable s sgnfcantly dfferent from zero. Thus the results of the tests of causalty from ndrect fnance to growth as well as that of causalty from growth to ndrect fnance show that both are ndependent of each other and hence we fnd support for Lucas vew that the economsts overstress the role of fnance. In cases of drect fnance and overall fnancal development also we do not fnd any evdence of causal effect of fnancal development on economc growth as the estmated coeffcent of the orthogonalzed causal canddate varables are sgnfcantly nsgnfcant. However, when we conduct the reverse causalty analyss we fnd that economc growth has negatve mpact upon the actvty n fnancal sector n the cases of drect fnance and overall fnancal development. V. COCLUSIO Ths study examnes emprcal relatonshp between fnancal development and economc growth whle ncorporatng the nflaton rate effect on fnancal development hghlghted n the lterature by Huybens and Smth (999); De Gregoro and Sturzenegger (994a, b); Boyd, Levne, and Smth (200); and Khan, Senhadj, and Smth (2003). We present evdence usng panel data of Low Income Countres. We apply panel causalty analyss for heterogeneous panel data gven by Wenhold (999) and ar-rechert and Wenhold (200). Our study focuses both ndrect fnance and drect fnance, separately as well as collectvely. Smple statstcal analyss made n Secton 4, wth the comparson of wthn-country standard devaton and between-country standard devaton for all the varables we have, revealed that for all the varables most of the varablty n the data occurs between countres whch shows the heterogenety between the countres for all these varables. one of the varables have larger wthn-country varaton. Ths justfes our use of heterogeneous panel methodology for causalty analyss. The evdence of the relatonshp between fnancal development and economc growth from contemporaneous non-dynamc fxed effects panel estmaton can at best be nterpreted as mxed. We do not fnd any postve sgnfcant relatonshp between ndrect fnance and economc growth. We do fnd that the drect fnance s sgnfcantly postvely related to economc growth. It s nterestng, however, to note that we fnd sgnfcant and postve relatonshp between sze of the overall fnancal development and economc growth aganst the evdence of no relatonshp between actvty of the overall fnancal development and economc. egatve sgnfcant estmates of coeffcent of the nflaton and fnancal development nteracton varable ndcate that fnancal sector development s actually more harmful for economc growth wth the ncreased nflaton n such countres
FIACE RESEARCH VOL., O., JAUARY 202 24 or puttng n a smple way: hgher nflaton s more harmful for economc growth for these countres at more developed stage of the fnancal system as compared to the less developed fnancal system. Monetary authortes of such countres have to take care of ths possble threat whle ther countres fnancal sector grows. In cases where we fnd the nteracton term to be sgnfcant, the magntude of the partal effect of nflaton on growth rate of GDP per capta s found to be larger n the fnal model as compared to that n the basc model whch shows that nflaton may be a much serous ssue n fnancally developed stage of economy as ts mpact s larger than that can be at the lesser (fnancally) developed stage of the economy. The contemporaneous analyss s based on underlyng assumpton about the homogenety of the relatonshps n questons across countres n the respectve panels. However, we fnd a bt of heterogenety n such relatonshps. Rechert and Wenhold (200) explots MFR coeffcents approach of Hsao et al (989) to develop a panel causalty method allowng for heterogeneous dynamcs across countres and for a dstrbuton over the coeffcents on the other explanatory varables. We apply Rechert and Wenhold (200) panel causalty method to our fnal model n dynamc form, to a panel dataset of low ncome countres. In contrast wth the evdence of Beck and Levne (2003), use of more approprate econometrc methodology of dynamc heterogeneous panel for causalty analyss and a refned model reveal that there s no ndcaton that fnancal development spurs economc growth or growth spurs fnancal development. Our fndngs are n lne wth the Lucas vew on fnance that the mportance of fnancal matters s very badly over-stressed n popular and even much professonal dscusson. The emprcal proxes of the fnancal development, whch most of the past emprcal studes have used, and followng these we have used n ths study, may not measure accurately the concepts emergng from theoretcal models. Theores focus on partcular functons provded by the fnancal sector, lke producng nformaton, exertng corporate governance, facltatng rsk management, poolng savngs, and easng exchange and how these functons nfluence recourse allocaton decsons and economc growth. Future research that concretely lnks the concepts from theory wth the data may substantally mprove further our understandng of the fnance growth lnk. In ths study we have not touched upon the ssues related to research on the relatonshp between fnancal development and economc growth employng the ndustry-level and frmlevel data. However, to further mprove our understandng of the fnance growth relaton future research work may focus to model ths relatonshp whle ncorporatng the nflaton effect on fnancal development usng the ndustry-level and frmlevel data whle applyng heterogeneous dynamc panel methodology for causalty analyss. We may hope some nterestng outcomes from such research. APPEDIX TABLE 3.: DATA DESCRIPTIO AD SOURCES Varable Data Descrpton and Source CPIa Annual Consumer Prce Index from IFS (Lne 64) CPIe End-of-year CPI from IFS (Lne 64M, or 64Q where 64M s not avalable) GDP Gross Domestc Product from IFS (Lne 99B) LLB Lqud Labltes from IFS (Lne 55L or 35L, f 55L s not avalable) MCP Market Captalzaton from Global Fnancal Data Base PCR Clams of Prvate Sector from IFS [Lnes 22D.MZF, 22D.TZF, 22D.ZF, 42D.FZF, 42D.GZF, 42D.LZF, 42D.ZF, and 42D.SZF are ncluded] POP Populaton (Lne 99Z) VTD GCE TRD GRGPC LLGR PCGR MCGR VTGR FDGR FAGR IFL GCGR TRGR SSER RGPC Value Traded from Global Fnancal Data Base Government Consumpton Expendtures from IFS (Lne 9F) Sum of Exports and Import (Lne 90C+98C from IFS) of Goods and Servces Annual percentage growth rate of GDP per capta based on constant local currency from WDI-2004. (Dependent Varable) Lqud Labltes to GDP rato Prvate sector credt to GDP rato Stock market captalzaton to GDP rato Stock market total value traded to GDP rato (Overall) fnancal depth to GDP rato (Overall) fnancal actvty to GDP rato Inflaton Rate Calculated from CPIa Government Consumpton Expendtures to GDP rato Internatonal Trade (sum of Exports and Import of Goods and Servces) to GDP rato Gross Secondary School Enrollment Rato from UESCO GDP per capta based on purchasng power party from WDI-2004 TABLE 3.2: COUTRIES ICLUDED Country Tme Span and Observatons From To Bangladesh 987 200 5 Cote d'ivore 98 200 2 Inda 977 200 25 Indonesa 977 2000 24 Kenya 976 200 26 Korea, South 974 2002 29 gera 978 200 24 Pakstan 984 200 8 Zmbabwe 98 999 9 Total 20