Author contact details Farid Khan, Curtin University, Ruhul Salim, Curtin University,

Transcription

1 The Public R&D and Produciviy Growh in Ausralian Broadacre Agriculure: A Coinegraion and Causaliy Approach Farid Khan and Ruhul Salim Auhor conac deails Farid Khan, Curin Universiy, faridecoru@yahoo.co.uk Ruhul Salim, Curin Universiy, Ruhul.Salim@cbs.curin.edu.au Conribued presenaion a he 59h AARES Annual Conference, Roorua, New Zealand, February 10-13, 2015 Copyrigh 2015 by Auhors. All righs reserved. Readers may make verbaim copies of his documen for non-commercial purposes by any means, provided ha his copyrigh noice appears on all such copies.

2 The Public R&D and Produciviy Growh in Ausralian Broadacre Agriculure: A Coinegraion and Causaliy Approach Absrac: This sudy invesigaes he nexus beween research and developmen expendiure and produciviy growh in Ausralian broadacre agriculure using counry-level ime-series daa for he period 1953 o Using sandard ime-series economerics daa are analysed o examine he dynamic relaionships beween research and developmen expendiure (R&D) and oal facor produciviy (TFP) growh. Findings here provide economeric evidence of a co-inegraing relaionship beween R&D and produciviy growh, and a unidirecional causaliy emergen from R&D o TFP growh. Moreover, employing variance decomposiion and impulse response funcion he dynamic properies of he model are explored beyond he sample periods. Findings sugges ha R&D can be readily linked o he variaion in produciviy growh beyond he sample periods. Furher, forecasing resul suggess a significan ou-of-sample relaionship exiss beween he public R&D and produciviy in broadacre agriculure. We used a novel mehod MIRR which is concepually superior han he convenional IRR o obain a credible esimae of reurns on public research invesmen. We found MIRR of 10.06% per year for he reinvesmen rae of 3% per year. Therefore, resuls esablishing long run relaionship beween produciviy and R&D in Ausralian agriculure shed ligh on he fuure policies in R&D invesmens in Ausralia. Keywords: Public Research & Developmen (R&D), Produciviy, Ausralian Broadacre Agriculure, Coinegraion, Inernal Raes of Reurn. JEL Classificaion: C32, Q16 1. Inroducion Research invesmens in agriculure are he cenral o he improvemens in agriculural produciviy growh, which is a crucial means for achieving economic prosperiy and developmen in an economy (Pardey e al., 2006; Mullen, 2010). A number of sudies have 1

3 been examined he effecs of research and developmen (hereafer, R&D) on oal facor produciviy (hereafer, TFP) in he agriculural secor. Many of hem provide empirical evidence ha R&D, boh domesic and foreign, is one of he main sources of produciviy growh (Hall and Scobie, 2006; Griliches, 1979, 1988; Coe and Helpman, 1995). In recen decades, he concern has been ha produciviy in agriculure is falling paricularly in developed economies. The declines in he agriculural produciviy have renewed ineres in he produciviy analysis, paricularly in he esimaion and explanaion of he effecs of R&D in agriculure. Few sudies examining he possible causes of he recen declines in he agriculural produciviy growh find he falling public R&D invesmen in agriculure over pas decades as one of he possible causes (Mullen, 2010; Alson and Pardey, 2001; Bervejillo e al., 2012). For example, Piesse and Thirle (2010) menion a slowdown and reargeing of public R&D as one of he key facors ha is primarily causing a slowdown in TFP growh in he Unied Kingdom. Similar evidence of slowing produciviy is also found in he US agriculure (Ball e al, 2013) in recen periods. Sudies also provide empirical evidence of long run relaionship beween research expendiure and agriculural produciviy growh in he developed counries such as UK agriculure (Thirle e al., 2008; Schimmelfenning and Thirle, 1994) and US agriculure (Wang e al., 2013; Alson e al. 2011). The falling produciviy growh is also eviden in Ausralian agriculure. Recen sudies found a slowdown in produciviy growh in Ausralian agriculure over he recen decade compared o earlier periods (Nossal and Sheng, 2010; Sheng, Gray and Mullen, 2011; Khan, Salim and Bloch, 2014). Keaing and Carberry (2010) saed ha one recen challenge for Ausralian agriculure is ha i has been facing slow agriculural produciviy growh in recen periods. They sugges ha his decline in produciviy growh can be aribued o he lagged impac of he public invesmen in agriculural research, which is sagnaed since 1970s. Some previous sudies esimaed he rae of reurn o R&D expendiure in Ausralian broadacre agriculure and indicaed ha public invesmen in agriculural R&D is conribuing o TFP growh. In he early 1990s, Mullen and various co-auhors conduced a series of economeric research wih agriculural R&D and produciviy in Ausralia. Using a unique daa se, hey found R&D is a major source of produciviy in Ausralian agriculure. Exending heir previous daa se, Mullen (2007) revisied heir previous sudy and found no evidence ha raes of reurn were declining over he years

4 Though previous sudies on Ausralian broadacre agriculure have esimaed he growh of TFP over recen decades, he empirical evidence wih regard o wha deermines he slowing TFP growh is apparenly very lile. Besides, mos of heir sudies have emphasised reurns o agriculural research and hus could no confirm he exisence of a sable long-erm co-inegraing relaionship beween research and produciviy growh. To dae, here have been very few sudies underaken in Ausralia ha examine he long-run relaionship beween R&D and produciviy growh in Ausralian broadacre agriculure. To he bes of our knowledge, we only find Salim and Islam (2010) explored long-run relaionship beween R&D and agriculural produciviy in broadacre agriculure in Ausralia. They applied sandard ime series echniques o invesigae he long-erm and causal relaionship beween R&D and TFP bu heir resuls are limied for Wesern Ausralian broadacre agriculure and do no based on a large ime-series daa. This sudy, herefore, aims o fill his empirical gap examining he relaionship beween public R&D spending and produciviy growh in Ausralian broadacre agriculure. To achieve his objecive his sudy applies coinegraion and Granger causaliy in order o invesigae he relaionship beween R&D and TFP and he direcion of causaliy running beween hem. Moreover, i applies variance decomposiion, impulse response funcion and a forecasing exercise o explore he dynamic properies of he relaionship beyond he sample periods. The res of he sudy proceeds as follows. The nex secion gives a shor overview of public R&D and agriculural produciviy in Ausralia. Secion 3 presens economeric mehodology of coinegraion and causaliy ess. A model is specified on wha facors affec oal facor produciviy in secion 4. A discussion on daa source is followed by in secion 5. Secion 6 presens empirical esimaes and analysis of resuls. The penulimae secion esimaes he benefis of research. Finally, Secion 8 concludes he sudy. 2. The Public R&D and Broadacre Agriculural Produciviy in Ausralia Ausralian agriculure is primarily based on exensive cropping and livesock farming aciviy, which is generally ermed as broadacre agriculure. Broadacre agriculure is a significan conribuor o he counry s agriculural and economic growh. I generaes more 3

5 han 85% of he counry s gross value of agriculural producion. The economic prosperiy of he rural communiy depends upon he growh of he counry s agriculure. Moreover, Ausralia expors around 60% of is agriculural producion, which represens 10.9% of oal expor earnings in The public secor plays a dominan role in R&D invesmen in Ausralian agriculure, which accouns generally more han 90 per cen of oal agriculural R&D. This saisic srongly conrass o oher OECD counries where he share of privae R&D is more han half of he oal invesmen in agriculural R&D (Sheng e al., 2011). Thus, he level of public invesmen in agriculural R&D and is impac on agriculural produciviy have been an imporan candidae in erms of public policy issue in Ausralia. However, he concern is ha i has been falling in recen periods apparenly since Before 1994, broadacre has experienced abou 2.2 per cen of growh in produciviy a year, bu i has faced a slowdown in produciviy growh hereafer. Since 1994, i has declined o 0.4 per cen a year. However, some recen sudies indicae ha he sluggishness in public R&D since he mid-1970s may have conribued o he slowdown in agriculural produciviy growh in recen periods (Sheng e al., 2011; Mullen, 2010). 3. Economeric Mehodology: Coinegraion and Causaliy 3.1. Tesing for he Order of Inegraion of he Variables To es he presence of uni roos, wo mos popular mehods applied in recen lieraure are he Augmened Dickey-Fuller (ADF) es and he Phillips-Perron es. The hree differen forms of simple relaionships allowing various possibiliies in economic ime series are he random walk, random walk wih a drif and rend saionary processes. The equaion ha nesed all he hree models is Y u (1) 1 2 Y 1 This equaion is used for he Dickey-Fuller uni roo es where he null hypohesis is ha = 0, i.e. here is a uni roo and hus he ime series Y is non-saionary. If is significanly differen from zero, here will be no uni roo and Y will be saionary in he levels, or inegraed of order zero, I(0). If Y is non-saionary in he levels, bu i becomes saionary 4

6 a firs differences, hen he series is o be inegraed of order one, I(1). However, if Y is no a firs-order auoregressive process, hen more lagged values of he dependen variable will need o be added o ensure ha he error erm is a whie noise. By adding m lagged values of dependen variable he equaion for he augmened Dickey-Fuller (ADF) es is Y m 1 2 Y 1 i i 1 Y u (2) i Phillips and Perron have developed a more comprehensive es of uni roo non-saionariy. Their ess are similar o ADF ess, bu hey address he issue of auocorrelaion by incorporaing an auomaic correcion o he Dickey-Fuller -es saisic, which allows for unspecified auocorrelaion in he disurbance process. Mos of he cases he ess give conclusions similar o he ADF ess Tesing for Coinegraion The Johansen echnique based on VAR Model This VAR-based coinegraion es proposed by Johansen (1995) uses he Maximum Likelihood esimaion mehodology o es for he coinegraion rank r, which represens he number of independen coinegraing vecors. I is more generally applicable han he radiional Engle Granger wo-sep mehodology o explore a single coinegraing relaionship. The VAR approach models every endogenous variable wihin he sysem. The following mahemaical form gives he VAR of order p in sandard form: y A y Ap y p (3) where y is a k vecor of endogenous variables ha are inegraed of order one, I(1), and A 1 Ap are (k x k) marices of coefficiens o be esimaed, and is a vecor of disurbances ha are serially uncorrelaed wih all he righ-hand side variables. The issue of simulaneiy does no arise in his specificaion as all endogenous variables of (3) are only predeermined lagged variables. Hence, each equaion in he sysem can be esimaed using OLS echnique, which gives consisen and asympoically efficien esimaes. 5

7 In order o use Johansen es, he VAR model is reparameerized ino a vecor error correcion model (VECM) of he following form: y y p y 1 1 i i (4) i 1 where p i 1 i and p i A j j i 1. The Johansen es examines he coefficien marix, as he key ineres o noe is he rank of he marix. According o Engle and Granger (1987), if all variables of he vecor y are inegraed of order one, I(1), he coefficien marix has rank 0 r <k, where r is he number of linearly independen coinegraing vecors. If rank ( ) = 0, here is no coinegraing vecor. Bu, if 1 r <k, here is a single or muliple coinegraing vecor in he sysem. If all variables of he vecor y are inegraed of order one, he coefficien marix has reduced rank r < k. The number of coinegraing vecors can be obained based on significance of he number of characerisic roos of he coefficien marix, as he rank of a marix is equal o he number of is characerisics roos. Johansen proposes wo ypes of likelihood raio es: he race es and maximum eigenvalue es for he number of characerisic roos using he following wo saisics: k T ln( 1 ) (5) race i r 1 ˆ i max T ln( 1 ˆ r 1) (6) where ˆ is he esimaed values of he characerisic roos (also called eigenvalues) obained from he marix and T is he number of usable observaions. The null hypohesis for he race es is r coinegraing vecors, and he alernaive is k coinegraing vecors. The maximum eigenvalue ess he null hypohesis for he race es is r coinegraing vecors agains r+1 coinegraing vecors. 6

8 3.3. Vecor Error Correcion Model The evidence of coinegraion only suggess an exisence of a long-erm, or equilibrium relaionship 1 beween ime series variables under consideraion. I does no consider he shor-erm dynamics of he model explicily. However, he presence of coinegraion among variables does no necessarily rule ou shor-erm disequilibrium among hem. The Granger represenaion heorem saes ha a coinegraed sysem of variables can be expressed as an error correcion model (ECM) (Engle and Granger, 1987). The ECM reconciles he shor-run behaviour of variables wih is long-run behaviour using he error erm of he coinegraing equaion, which is also ermed as equilibrium error. As a simple example, for he wo-variable case wih only one lagged difference he ECM can be wrien as: y1 1( y2 1 1y1 1) 11 y y2 1 1 (7) y2 2( y2 1 1y1 1) 21 y y2 1 2 (8) where denoes he difference operaor, y1 and y2 are he wo variables of inegraed of order one, and is a random error erm which is independenly and idenically disribued. The inclusion of lags of he dependen variable as he explanaory variable o he regression is necessary as he dependen variable iself may be correlaed wih is lags. Noe ha he error correcion erm ( y2 1 1y1 1) is one-period lagged value of error u 1 from he coinegraing equaion, which equals zero in a long-run equilibrium relaionship. However, if i is non-zero, variables adjus in he shor run o correc he equilibrium error o make he model equilibrium. In he shor-run, he error correcion erm is non-zero and each variable adjuss o resoring he equilibrium. The coefficiens 1 and 2 are he adjusmen parameers, which represen he speed of adjusmen in error correcion mechanism. The ECM has boh long-run propery, which is buil in error correcion erm, u 1 and shor-erm propery, which is capured by he error correcion coefficien Granger Causaliy 1 Long-erm relaionship measures a he level form of he variables while shor-run dynamics measure a he firs-differences of he variables. 7

9 Granger causaliy is used o shed ligh on he direcion of possible causaliy beween pairs of variables. According o he Granger represenaion heorem, here will be Granger causaliy in a leas from one direcion if wo variables inegraed of order one, I(1), are coinegraed. In a simple model wih wo variables, y 1 and y 2, Granger causaliy ess wheher pas values of y 1 help in predicing y 2 given he effecs of pas values of y 2 on y 2 are accouned for. If hey do, hen y 1 is presumed o Granger causes y 2. Granger causaliy can be examined using following VAR framework of order-p: y y p y1 p 11y p y2 p 1 (9) y y p y2 p 21y p y1 p 2 (10) The equaion (9) models y 1 as a linear funcion of is own lagged values, plus lagged values of y 2. If lagged values of y 2 have non-zero effecs on y 1, hen y 2 Granger causes y 1 condiional on he effecs of is own lagged accouned for. In his simple VAR, Granger causaliy esing ses he null hypohesis ha y 2 does no Granger causes y 1. H0 : p 0. This join hypohesis can be esed using a sandard Wald F or 2 es, since each individual se of parameers resriced is drawn from only one equaion. Similarly, in equaion (10) he null hypohesis ha y 1 does no Granger causes y 2 can be expressed as H0 : p 0. If y 1 cases y 2, lags of y 1 should be significan in he equaion for y 2. If i does so and no vice versa, hey indicae ha here exiss unidirecional causaliy from y 1 o y 2. On he oher hand, if y 1 cases y 2, lags of y 1 should be significan in he equaion for y 2. If i does so and no vice versa, hey indicae ha here exiss unidirecional causaliy from y 1 o y Model Specificaion On modelling he relaionship beween oal facor produciviy and research expendiures, his paper employs a producion funcion approach of he following form: TFP A R & D FR & D ENROL (11) i 8

10 where is a ime index; R&D is lagged domesic public R&D expendiures (or R&D socks) in broadacre agriculure; FR&D is foreign public R&D, which is proxied by US R&D in agriculure; ENROL is a measure of farmer educaion, which is proxied by school enrolmen; and TFP is oal facor produciviy. A is he par of TFP no caused by he included variables and s are he respecive weighs o he facors menioned. The funcional form is specified as log-linear - he four variables are all in logarihmic erm. Given limied guidance in he economic heory regarding he shor-run and he long-run dynamic relaionships beween TFP and R&D, we adop a modelling sraegy based upon he informaion provided by he ime-series daa. Hence, we use an unresriced VAR model ha allows daa o speak o he possible links and direcions among he variables of ineres. To conrol he spillover effecs of foreign research his sudy uses R&D expendiure in US agriculure as a proxy for he foreign R&D expendiure. US play a significan role in global agriculural R&D in relaion o is invesmen and in erms of research spillovers (Alson, 2002; Sheng e al., 2011). Besides, Ausralia mainains a considerable economic and rade relaion wih US. Moreover, assuming he effecs of foreign research and developmen usually depend on how he counry is exposed o foreign rade, we consruc and use an impor-share-weighed US R&D variable o he model following Coe and Helpman (1995) raher han simply using US R&D as a crude proxy for foreign R&D. Because, i is ofen assumed ha he ransfer of knowledge and echnology beween counries depend on rade channel, which faciliaes access o he oupus of foreign R&D, hereby enhance produciviy (Ang and Madsen, 2013). Anoher conrol variable is farmers educaion, which is proxied by school enrolmen (ENROL) i.e. he proporion of primary school-age sudens in he oal populaion enrolled in primary schools in rural areas. Inclusion of human capial is naural in he TFP regressions because educaion makes people beer o organize work, communicae, and help o be innovaive, all of which conribue o a higher produciviy level. 5. Daa This sudy uses he counry-level ime-series daa for he period 1953 o The broadacre TFP index is measured by he Ausralian Bureau of Agriculural and Resource Economics 9

11 Sociey (ABARES), which is esimaed as he raio of a Fisher quaniy index of oal oupu o a Fisher quaniy index of oal inpu. Empirically, TFP growh is measured as a par of farm oupu growh, which is no conribued by growh of he facor inpus o he conrol of farmers (Solow, 1957). TFP hus includes he effecs of advances of knowledge or echnological progress along wih oher facors affecing i (Jorgenson and Griliches, 1967). A complee descripion of how ABARES consrucs TFP index for he broadacre indusries can be found in Gray e al. (2011). The domesic public invesmen in R&D in broadacre agriculure series builds on daa calculaed by Mullen (2010) and from he Ausralian Bureau of Saisics (ABS) biannual Ausralian Research and Experimenal Developmen Survey. Mullen assembled he daa from various public sources, including Ausralian Bureau of Saisics (ABS) R&D daa, and from a previous daase developed by Mullen e al. (1996). The real public R&D expendiure is in 2009 dollars based on he GDP deflaor. This daa considers invesmen on plans and animals and excludes for he fisheries, foresry, environmen and processing. Finally, based on broadacre agriculure s share of he oal value of producion in agriculure, he R&D in broadacre alone is derived from he R&D invesmen in Agriculure. Toal R&D expendiure on agriculural producion in US o proxy for foreign R&D expendiure is colleced from US Deparmen of Agriculure (USDA). This daa is weighed by rade openness, he percenage of he agriculural impors o he agriculural gross value of farm producion (GVP) in Ausralia. Agriculural GVP is obained from ABARES and impors of agriculural crops and livesock producs are obained from FAO saisics. However, rade openness daa is exrapolaed backwards for he period 1953 o 1960 using acual daa from 1961 o Similarly, school enrolmen is also exrapolaed backwards for he period 1953 o 1970 using he acual daa. This sudy uses he World Developmen Indicaors daabase o obain daa on he proporion of primary school-age sudens in he oal populaion enrolled in primary school in Ausralia o proxy for he level of educaion of broadacre farmers. 6. Empirical Resuls and Discussion 6.1. Uni roo es 10

12 We invesigae he ime-series properies of he variables using wo widely used uni roo ess, he Augmened Dickey-Fuller (ADF) and he Philips-Peron ess. Table 3.1 repors he es saisics for he ime-series daa covering he period in heir naural form. The resuls show ha all variables TFP, public agriculural R&D expendiures, farmer educaion and foreign R&D expendiures are non-saionary in heir levels, bu hey are saionary in he firs differences, or inegraed of order one, I(1). We also find similar inegraion order for all variables by Phillips-Perron ess saisics. Table 3.1 Uni Roo Tess: ADF and Phillips-Perron Variables ADF Tes Phillips-Perron Tes Order of Inegraion P-value Inercep, Trend and Inercep P-value Inercep, Trend and Inercep TFP 0.75 Inercep 0.67 Inercep TFP 0.00 Boh 0.00 Boh I(1) R&D 0.36 Boh 0.99 Boh R&D 0.00 Boh 0.00 Boh I(1) FR&D 0.42 Inercep 0.08 Boh FR&D 0.00 Boh 0.00 Boh I(1) ENROL 0.20 Inercep 0.54 Boh ENROL 0.01 Boh 0.01 Boh I(1) Noe: In case of Boh es saisics are repored for Trend and Inercep. Table 3.2 Zivo Andrews Uni Roo Tess Series Level Break a Firs diff. Break a Lag lengh TFP *** *** R&D ** FR&D *** ENROL *** Criical values: 1%: and 5%: -5.08; *** significan a 1% level, ** significan a 5% level. Noe: Breaks are considered boh in inercep and in rend. All variables are in logarihm form. However, he sandard uni roo ess may no be appropriae if he concerned series conain any srucural breaks (Bloch e al., 2012; Shahiduzzaman and Alam, 2012). The resuls of ADF or PP ess migh lead o conclude a non-saionary series as saionary because of no allowing breakpoin in he series if any. Considering he possibiliy of a srucural break in he daa series his es can be reaed as a cross check of he oher usual 11

13 uni roo ess. Table 3.2 shows he resuls from he Zivo-Andrews ess (Zivo and Andrews, 1992) considering srucural breaks in he series if any. Similar o he Dickey-Fuller es, he Z-A es also mainains he null hypohesis of a uni roo in he process, i.e., non-saionary series. The Z-A es suggess o rejec he null of I(1) for all variables as he -saisics are larger han he criical values, which subsaniae he uni roo resuls of saionariy in firs difference found in wo oher ess ADF and PP. However, for TFP and Enrol variables, we canno rejec he null of I(0) suggesing hey are inegraed in he levels while we consider he srucural break in he series Coinegraion and VEC Model Coinegraion es: Johansen Approach based on VAR To es for coinegraion using Johansen approach, we need firs o specify how many lags o include in he VAR model wih I(1) variables. Table 3.3 presens he saisical resuls for deermining opimal lag lengh. As here is no explici heory o guide opimal lag lenghs, we rely on differen saisical echniques commonly applied o he lieraure in selecing he opimal lag for he VAR model. Resuls indicae ha he sequenial modified likelihood raio (LR) es, he Schwarz informaion crierion (SC) and he Hannan-Quinn informaion crierion (HQ) sugges for only one lag in he model, as indicaed by * in he able. Resuls repored in Table 3.3 show ha according o LR and AIC mehods he number of opimal lag is hree hough wo oher ess SC and HQ favour wo lags. Table 3.3 Selecion of he number of VAR lags Endogenous variables: LnTFP LnR&D LnFR&D LnEnrol Lag LR AIC SC HQ 0 NA * * * * * indicaes lag order seleced by he crierion a 5% level Deermining he common inegraion properies of all he variables in he model as well as selecing he number of opimal lag, we can proceed o es he presence of 12

14 coinegraing vecor. However, as all he variables are saionary in he firs difference i.e. I(1), here may presen a coinegraing relaionship in he model. We use mulivariae maximum likelihood approach of Johansen and Juselius (1990) which allows esimaion of muliple coinegraing relaionships. The resuls for race es and eigenvalue es are presened in Table 3.4. The resuls sugges rejecing he null hypohesis of no coinegraing vecors bu canno rejec he hypohesis of a mos one coinegraing equaion according o he ess saisics. Boh Trace es and Max-eigenvalue es indicae one coinegraing equaion a 5% significance level. Table 3.4 Coinegraion Tess: Johansen and Juselius Approach Series Tesed: LnTFP LnR&D LnFR&D LnEnrol Hypohesized 5% No. of CE(s) Eigenvalue Saisic Criical Value Prob.** Trace Tes None * A mos A mos A mos Max-Eigenvalue Tes None * A mos A mos A mos * denoes rejecion of he hypohesis a he 0.05 level **MacKinnon-Haug-Michelis (1999) p-values Vecor Error Correcion Model: Johansen and Juselius Mehod Having esablished coinegraion, we can proceed o es he shor-run dynamic relaionship beween variables. Table 3.5 presens he es resuls for error correcion by using Johansen- Juselius vecor error correcion mehod for differen lag specificaions of R&D. In he able, Panel A shows resul for 12 years of lag value of R&D, following a sudy by Thirle e al. (2008) in UK agriculure where hey used 12 years lag srucure. This ype of lag srucure has been fied o oher sudies as well, including Salim and Islam (2010), Piesse and Thirle (2010), and Schimmelpfennig and Thirle (1994). The resul of saisically significan and 13

15 non-zero equilibrium error erm provides he evidence of he adjusmen of he shor-run disequilibrium condiion owards he long-run equilibrium for he model. In addiion, Panel B and Panel C repor resuls based on R&D socks consrucing by wo alernaive specificaions of R&D lag srucure: perpeual invenory mehod (PIM) and gamma disribuion, respecively. Under he PIM mehod, R&D socks are calculaed assuming a depreciaion rae fixed a 5%. In panel C, R&D socks are calculaed assuming a gamma disribuion wih 30-year research lag lengh. Given he daa limiaion and considering he relaively applied naure of public agriculural R&D in Ausralia, we allow 30-year lagged specificaions of he research impacs on produciviy, which is consisen wih previous sudies in Ausralian broadacre agriculure e.g., Cox e al., (1997). Following Alson e al. (2011) he parameers of he gamma lag disribuion are assigned wih values of = 0.70 and = Resuls show ha in TFP equaion he equilibrium error erm is saisically significan and non-zero reflecing adjusmen of he shor-run disequilibrium condiion owards he long-run equilibrium. The negaive value o he adjusmen coefficien, which gives he required sign, suggess ha TFP will be negaive abou resore he equilibrium for he sysem. This implies ha agriculural TFP growh responds o shocks from he R&D spending. Table 3.5 Error-correcion model Panel A. 12-year Lag R&D alpha Coef. Sd. Err. z P>z [95% Conf. Inerval] TFP _ce1 L R&D -12 _ce1 L FR&D _ce1 L ENROL _ce1 L Panel B. Wih R&D Socks (PIM) alpha Coef. Sd. Err. z P>z [95% Conf. Inerval] TFP _ce1 L R&DS PIM _ce1 L FR&D _ce1 L ENROL 14

16 _ce1 L Panel C. Wih R&D Socks (Gamma disribuion) alpha Coef. Sd. Err. z P>z [95% Conf. Inerval] TFP _ce1 L R&DS gamma _ce1 L FR&D _ce1 L ENROL _ce1 L The deail resuls of he coinegraing equaions are repored in appendix Table A.3.1 wih Johansen s normalizaion resricion is imposed on TFP o be uniy. The esimaed parameers of he coinegraing vecor are exacly idenified, and he model fis well. Overall, he oupus indicae he exisence of an equilibrium relaionship beween he TFP and R&D. The resuls of normalized coinegraing coefficiens are presened in he following coinegraing relaionship for differen specificaions: LnTFP LnTFP LnTFP LnR & D *** LnFR & D 0.84LnENROL *** (12) LnR & DS PIM *** 0.019LnFR & D 1.91LnENROL ** (13) gamma *** LnR & DS 0.105LnFR & D 3.022LnENROL ** (14) The normalized coinegraing equaion (12) considers 12 years of R&D lag. Equaions (13) and (14) specified wih research socks based on PIM and gamma disribuion, respecively. In all specificaions, he bea coefficiens for R&D are posiive and saisically significan across differen R&D lag lengh srucure. This bea coefficien indicaing posiive relaionship beween lagged R&D and TFP can be considered as long-erm marginal effecs on TFP. As we used double logarihmic funcional form, he bea coefficiens can be inerpreed as long-erm elasiciy. In addiion, foreign R&D is posiively relaed o TFP, hough he coefficiens are no significan. However, hough i is likely ha he enrolmen coefficien is posiively relaed o TFP in he long run, bu in he model, he resul shows a negaive relaionship beween hem. We use he LR es for linear resricions o see wheher he bea coefficiens are significan in he coinegraing relaionship. Table 3.6 repors he chi-squared es saisics for zero resricions (coefficien resriced o zero) ess o see wheher each of he variables 15

17 can be excluded from he coinegraing space. Resuls sugges ha R&D (boh 12-year lagged R&D and research sock based on gamma disribuion) conribues significanly o he coinegraing relaionship. This oucome does valid our model ha R&D has a long-run impac on he TFP. The resul also shows ha TFP and Enrol variables ener he coinegraing relaionship significanly since each resricion is rejeced a he 5% level. In addiion, foreign R&D canno be excluded from he model a 10% significan level. Table 3.6 LR es for exclusion of variables from coinegraing space (zero resricion) 12 Years R&D Lag R&D Socks Gamma disribuion chi2 p-value Chi2 p-value LnTFP LnR&D LnFR&D LnENROL z saisics in he parenheses Specificaion esing We conduc a series of diagnosic ess o check specificaion of he model, which is crucial for he validiy of he esimaes and inferences of he model. Table 3.7.a repors resul for checking he sabiliy condiion of VECM esimaes. The resuls sugges ha we have correcly specified he number of coinegraing equaions as we find K r (K endogenous variables and r coinegraing equaions) uni moduli in he sabiliy ess and he remaining moduli are sricly less han one. In addiion, we also perform LM es for auocorrelaion in he residuals. Resul repored in Table 3.7.b suggess ha we canno rejec he null hypohesis ha here is no auocorrelaion in he residuals a eiher lag order one or wo. Thus, es indicaes no evidence of auocorrelaion in he model. Table 3.7.a Eigenvalue sabiliy condiion Eigenvalue Modulus

18 i i The VECM specificaion imposes 3 uni moduli. Table 3.7.b Lagrange-muliplier Tes lag chi2 df Prob > chi H0: no auocorrelaion a lag order 6.3. Granger Causaliy Tess To explore he direcion of he causaliy among he variables in he coinegraed vecor, we applied Granger causaliy es. The presence of one coinegraing vecor implies ha here should be Granger causaliy in one direcion. Table 3.8.a presens he Granger causaliy Wald es based on vecor auoregressions o esablish he direcion of causaliy of he coinegraed vecor. The chi2 saisics in he firs row ess if R&D, foreign R&D and enrolmen are Granger-prior o TFP, he dependen variable in his case. The probabiliies in he nex row show ha R&D is Granger-prior o TFP, and his is also rue for all explanaory variables ogeher, which is an expeced oucome. We run similar es for each of he remainder dependen variables such as R&D, foreign R&D, and enrolmen o find if hey are Grangercaused by any variables. The resuls sugges no evidence of any feedbacks in he opposie direcion, which esablish he presence of one granger causaliy running from R&D o TFP. Table 3.8.a Granger causaliy Wald ess Vecor auoregressions Dependen Excluded Variables Variable TFP R&D FR&D ENROL All chi2 TFP Prob > chi * * 0.000* chi2 R&D Prob > chi chi2 FR&D Prob > chi chi2 ENROL Prob > chi * denoes rejecion of he hypohesis a he 0.05 level 17

19 Table 3.8.b Toda-Yamamoo Granger non-causaliy ess Dependen Variable Excluded Variables TFP R&D FR&D ENROL All TFP *** 5.079** *** R&D FR&D *** 3.591* ** ENROL * ***, ** and * denoe rejecion of he hypohesis a he 0.01, 0.05 and 0.10 level, respecively. This sudy also follows he Toda-Yamamoo (TY) procedure o es for Granger causaliy for sensiiviy check, i.e. o make sure ha he causaliy esing is done properly. Toda and Yamamoo (1995) indicae ha economic series likely o be eiher inegraed of he differen orders or non-inegraed or boh. Hence, he usual Wald es saisic does no follow is usual asympoic disribuion, which could lead o a flawed inference. Toda and Yamamoo (1995), herefore, developed an alernaive augmened Granger causaliy es, which is useful when series are even no inegraed in he same order. Table 3.8.b repors he resuls of he TY augmened Granger Non Causaliy es. The es s resuls suppor he view ha R&D Granger-causes he TFP and evidence of no feedback in he opposie direcion. From he able, we find in he case of he dependen variable TFP, he resul suggess rejecing he null hypohesis of Granger non-causaliy implies he presence of Granger causaliy running from R&D o TFP. On he conrary, when R&D considers as a dependen variable resul does no sugges rejecing he null, he presence of no Granger causaliy of TFP o R&D. This implies Toda-Yamamoo procedure also suggess ha he R&D Granger cause TFP Variance Decomposiion and Impulse-Response Funcion The variance decomposiion and impulse response funcion provide more informaion of he dynamic properies o he model and allow predicing he relaive imporance of he variables beyond he sample period (Salim and Islam, 2010). Variance decomposiion measures he proporion of variaion in he dependen variable ha is induced by heir own shocks or shocks emanaing from oher variables. Table 3.9 presens he variable decomposiion esimaes for TFP for 30 years of he ime horizon. The resul shows in he case of TFP, abou 90% of he forecas error variances a he fifh-year horizon are accouned for by is own 18

20 shock, and he R&D, foreign R&D, and Enrolmen conribue he remaining 10% of shocks. The R&D explains abou 7.5% and 14.2% in 10 h and 20 h year, respecively, which remain almos persisen over he fuure period. The resuls indicae he fuure variabiliy o TFP largely originae from is own shocks, which is hus appeared o be exogenous. In 30 years, 71.1% of fuure variaion in TFP is due o is own innovaions and R&D explains abou 18.1%. On he oher hand, oher variables such as foreign R&D and enrolmen do no considerably explain in he long run. Table 3.9 Variance Decomposiion of LNTFP Period S.E. LnTFP LnR&D LnFR&D LnENROL Cholesky Ordering: LnTFP LnR&D LnFR&D LnENROL This sudy, furher, uses Cholesky one sandard deviaion impulse response funcion as par of he robusness checks of he coinegraion findings beyond he sample period. The impulse response funcions provide he response of he dependen variables o he shocks o each of he variables in he VEC model. Figure 3.2 shows he impulse response funcions based upon he VAR esimaes. As he main ineres of his sudy is o examine he responses of TFP, we only presen he effecs of shock in all variables o he variable TFP. The impulse response funcions for he res of he variables are presened in he appendix Figure A.3.1. Figure shows ha he response of produciviy growh o a one sandard deviaion innovaion in research and developmen is posiive and persisen. The graph suggess, in response o a shock in R&D, fuure TFP iniially increases, and hen i remains posiive and nearly permanen for he fuure periods a 3%. Figure also shows a negaive and ransiory response of produciviy o he shocks boh in foreign R&D and in enrolmen as he effecs die ou in he fuure. 19

21 Response o Cholesky One S.D. Innovaions Response of LNTFP o LNRDD Response of LNTFP o LNFRND Response of LNTFP o LNENROL Figure 3.2 Generalized Impulse Response Funcions in LNTFP Equaion 6.5. Forecasing Exercise This secion presens a forecasing exercise in order o evaluae wheher changes in R&D socks conain informaion abou fuure changes in he produciviy of Ausralian broadacre agriculure. We produce forecass from he esimaed VEC model where boh lagged values of TFP and R&D socks are used for forecasing. Model also includes foreign R&D and enrolmen as wo exogenous variables. Figure 3.3 shows esimaed forecass of TFP for he forecas period 2010 o 2020 along wih confidence error bands. Based on he esimaed VEC model he graph shows ha produciviy declines over he forecass period. We use dynamic forecasing approach for his ou of sample forecasing. This approach uses he forecased value of he lagged dependen variable. As a resul, he confidence error bands widen owards o he end of he forecas sample because he forecass errors end o compound over ime. 20

22 LNTFP ± 2 S.E Acual LnTFP Forecas lntfp (VEC) Figure 3.3 Ou of Sample Forecass of TFP for sample To see ou-of-sample performance of he VEC model we esimae forecas evaluaion and compares wih oher models. To obain ou-of-sample forecasing evaluaion we reserve par of our sample by no including i in he esimaion sample. We esimae VEC and oher models for he sample period 1953 o 2002 (reserving seven years of acual daa for he evaluaion purpose) and perform ou of sample forecasing for he period 2003 o Following, Apergis (2014), we compare he VEC-based TFP forecass wih hose of he random walk model (RW) and basic forecasing model (wih consan and rends) by using wo saisics: roo mean squared errors (RMSE) and he Theil coefficiens. Table 3.10 repors and compares forecas evaluaions across differen forecasing models. The resuls indicae ha he VEC model ha includes R&D knowledge socks performs beer han oher wo models giving smaller values of RMSE and Theil coefficien. These resuls necessarily imply ha inclusion of informaion on R&D knowledge socks gives beer predicive abiliy of fuure TFP. Table 3.10 Ou of sample forecasing of TFP for he period RMSE Theil Inequaliy Coefficien 21

23 VEC Model RW Model Basic Inernal Rae of Reurn In his secion, we invesigae economic performance of he public invesmens in R&D in Broadacre agriculure by applying he measures of benefi-cos raios, IRR, and MIRR. Three main ingrediens required o calculae hese economic performance measures are he elasiciy of produciviy wih respec o a change in he R&D sock, esimaes of he real value of agriculural oupu and esimaes of R&D socks ha include a simulaed increase in research invesmens. Following Andersen and Song (2013), we compue economic performance measures applying a sraighforward mehod ha uses aggregae naional-level daa and a single esimae of he elasiciy of produciviy wih respec o a change in he R&D sock. A simulaed percenage increase in he R&D sock for he period can be defined as: ( ) (15) where is he acual knowledge sock and is he simulaed knowledge sock afer including a hypoheical increase of $1,000 in R&D invesmen in 1954, he year ha represens he presen value in he analysis a which = 0. For consrucing knowledge sock, we assumed gamma lag disribuion wih he research lag lengh of 30 years including implici gesaion period. compued as: The presen value of benefis from he $1,000 invesmen in public R&D can be ( ) (16) where denoe he real value of agriculural oupu in period, r denoe a real ineres rae, N is he research lag lengh and is he elasiciy of produciviy wih respec o a change in he knowledge sock in Eq. (11). 22

24 Now, he benefi-cos raio for ha $1,000 invesmen is compued by dividing he presen value of benefis, PVB, by he presen value of cos, PVC which is simply he iniial increase in invesmen of $1,000 in 1954 is: ( ) (17) In addiion o benefi-coss raio, we compue inernal rae of reurn (IRR) which is he ineres rae received for an invesmen ha makes he ne presen value equal o zero. Nex, he fuure value of benefis afer N years is defined as: (18) Finally, he modified inernal rae of reurn is defined as: [ ] (19) According o Alson e al. (2011) and Andersen and Song (2013), in evaluaing he reurn o public invesmens in R&D a MIRR is superior o a convenional IRR for a concepual reason. Specifically, he convenional IRR implicily assumes ha he flows of benefis ha accrue over ime can be reinvesed in he same iniial invesmen. However, i may no be suied for he public agriculural R&D where he benefis ha accrue over ime go o producers and consumers of farm producs by reducing producion coss and food prices. The IRR measure is bes suied for an invesmen siuaion where he invesor reaps all of he reurns. We compue he modified inernal rae of reurn as an alernaive of convenional inernal rae of reurn, which has an advanage ha i allows for alernaive reinvesmen raes of he sream of benefis. Table 3.11 Benefi-cos raios, IRR and MIRR 30 years research lag lengh 50 years research lag lengh Reinvesmen Benefi-Cos IRR MIRR Benefi-Cos IRR MIRR rae Raio Raio (1) (2) (3) (4) (5) (6) (7) Percen per Year Percen per Year 5%

25 3% % Our esimaes of benefi-coss raio, convenional inernal rae of reurn and modified inernal rae of reurn are repored in Table Resuls show ha measure of benefi-cos raios range from o depending on he assumed maximum lag lenghs and discoun raes. In case of 30 years of research lag lengh and a an assumed real discoun rae of 3% per year he benefi-cos raio is The benefi-cos raios are consisen wih oher recen sudies. For example, in US agriculure, Alson e al. (2011) found benefi-cos raios 17.5 and 21.9 for 50-years and 35-years research lag lengh, respecively. Similarly, Andersen and Song (2013) also found he esimaed benefi-cos raio for he base model wih he preferred esimaion procedure is in he US agriculure. We also calculae he convenional IRR repored in column (3) and (6) in Table Alhough IRR is no a preferred measure, bu is common in he lieraure. Mos of preceden lieraure esimaed IRR as i is useful for purposes of comparison. The esimaed IRR for he maximum research lag lengh of 30 years and 50 years are respecively 26% per year and 23% per year. This resul is consisen wih some recen sudies in US agriculure, where Alson e al. (2011) and Andersen and Song (2013) found he esimaed IRR are approximaely 22.7% per year and 21% per year, respecively. Similarly, in case of Ausralia, Mullen (2007) found he real rae of reurn of he public research of 15% per year in Ausralian broadacre agriculure. Recenly, for all agriculure Sheng e al. (2011) also compued an average esimae of real rae of reurn of 28.4% in Ausralian agriculure. However, his rae is repored o be relaively smaller compare o he resuls from surveys of he numerous sudies over he years where he esimae of he raes of reurn is in he range of approximaely percen per year as repored in Alson e al. (2009). Also, in a meaanalysis of 292 sudies, Alson e al. (2000) repored an overall mean inernal rae of reurn of 64.6%using a sample of 1,128 esimaes. A grea number of he previous lieraure used inernal rae of reurn as a common summary measure of invesmen performance in he agriculural R&D evaluaion despie of is mehodological criicisms by economiss for more han half a cenury. We compue he modified inernal raes of reurn (MIRR) which addresses he mehodological concern wih using he IRR (Hurley e al., 2014). The esimaes of MIRR are repored in column (4) and 24

26 (7) in Table 3.11 under he assumpion of a real reinvesmen rae of 1%, 3% and 5% per year. Depending on he maximum research lag lengh and he assumed reinvesmen rae, resuls indicae ha he esimaes of MIRR is somewhere in he range of 8.87% per year o 16.31% per year. For 50-years maximum lag lengh, he esimaed MIRR is 10.06% per year when he reinvesmen rae is 3% per year. The esimaed MIRR in his sudy is consisen wih some recen sudies wih US agriculure. For example, Alson e al. (2011) compued an average of 9.9% per year across US saes. Similarly, Andersen and Song (2013) found ha he MIRR is 9.84% per year for he public invesmen in agriculural R&D in he Unied Saes. Our esimaes of MIRR of 10.06% per year is also consisen wih a recen sudy by Hurley e al. (2014) which re-examined he repored raes of reurn from 372 separae sudies from 1958 o They found ha he median MIRR varies from 9.7% o 10.4% per year for a range of reinvesmen raes of benefis from 0 o 50%. 8. Conclusions This sudy invesigaes he long-run relaionship beween he public R&D and he TFP in broadacre agriculure in Ausralia over he period of five decades. A producion funcion approach is used as an analyical model by making oal facor produciviy a funcion of research and developmen expendiure. This model also incorporaes he variables ha conrol foreign echnology ransfer (foreign public R&D) and human capial (farmers educaion level). To ensure ha coinegraion is possible, firs we use he Augmened Dickey Fuller and he Phillips Perron uni roo ess o deermine ime-series properies of he variables. Then, using he coinegraion and causaliy analysis, we find economeric evidence of coinegraing relaionship beween research and developmen expendiure and produciviy growh. Resuls also show he evidence of a causal relaionship beween R&D o TFP growh. Wih respec o he direcion of causaliy, he empirical evidence indicaes a unidirecional causaliy running from R&D o TFP growh. In oher words, research and developmen expendiure Granger causes oal facor produciviy as curren and pas values of R&D improve predics of TFP above he pas values of TFP alone. This resul is robus according o he Toda-Yamamoo Granger non-causaliy es. Having esablished coinegraion, an error correcion model consruced, which shows ha lagged R&D is significan in explaining changes in oal facor produciviy. This resul 25

27 implies an increased R&D expendiure leads o beer oucomes for produciviy in Ausralian broadacre agriculure. Furher, we explore he dynamic properies of he model using variance decomposiion and impulse response funcion. The resul suggess ha beyond he sample periods, he public R&D considerably explain he variaion in produciviy growh in Ausralian broadacre agriculure. In addiion, TFP responses posiively and persisenly for he fuure period as he effec of shock in he public R&D does no die ou over ime. This sudy, herefore, esablishes he exisence of a long-run unidirecional causal relaionship beween R&D and produciviy growh in a more dynamic fashion. Furher, his sudy, hrough an ou-of-sample forecasing exercise, also indicaes ha invesmen in public R&D in agriculure does maer in forecasing produciviy growh. Resuls show ha informaion on R&D invesmen improves produciviy forecass significanly. Moreover, his sudy also compued and compared differen measures of economic performances for he public invesmens in agriculural R&D. The resuls show ha he benefi-cos raio is 26.94, he inernal rae of reurn is 23% per year, and he modified inernal rae of reurns is 10.06% per year. The measures of convenional inernal raes of reurn are consisen wih some recen sudies, e.g., Alson e al. (2011) and Andersen and Song (2013), ye relaively lower han some previous sudies. The esimaed modified inernal rae of reurns is approximaely % per year, depending on he research lag lengh and reinvesmen rae of benefis. This esimaed modified reurn o public R&D is lower han he repored convenional IRR and is mehodologically more jusified and plausible. These resuls, indeed, sugges ha research affec agriculural produciviy in he long run as an imporan source of produciviy growh. The insigh behind he relaionship beween he public R&D and produciviy in broadacre agriculure in Ausralia is sraighforward. An increase in he public expendiure in R&D is likely o lead o higher produciviy growh in he long run. Finally, R&D should arac more public aenion in governmen agriculural policy as an increase in R&D expendiure has a posiive and sizable rae of reurn hrough conribuing produciviy growh. The resuls may, however, be limied by he naure of he research and developmen daa. The model focuses solely on public R&D in broadacre agriculure. Moreover, only he effec of US R&D is represened for he effecs of foreign R&D on TFP. Hence, he resuls may be limied by any effecs of he R&D expendiure in privae secors and in oher secors in Ausralia. Given hese pracical limiaions, our resuls are sill perinen as our main 26