Learning-by-exporting in Korean Manufacturing: A Plant-level Analysis

Transcription

1 Chapter 8 Learnng-by-exportng n Korean Manufacturng: A Plant-level Analyss Chn Hee Hahn Korea Development Insttute Chang-Gyun Park College of Busness Admnstraton, Chung-Ang Unversty March 2009 Ths chapter should be cted as Hahn, C. H. and C.-G. Park (2009), Learnng-by-exportng n Korean Manufacturng: A Plant-level Analyss, n Corbett, J. and S. Umezak (eds.), Deepenng East Asan Economc Integraton. ERIA Research Project Report , pp Jakarta: ERIA.

2 CHAPTER 8 Learnng-by-exportng n Korean Manufacturng: A Plant-level Analyss CHIN HEE HAHN 1 Korea Development Insttute CHANG-GYUN PARK 2 College of Busness Admnstraton, Chung-Ang Unversty The paper analyzes whether frms that start exportng become more productve utlzng recently developed sample matchng procedures to control the problems from self-selecton nto the export market. We use plant level panel data on Korean manufacturng sector from 1990 to We fnd clear and robust emprcal evdence n favor of the learnng-by-exportng effect; total factor productvty dfferentals between exporters and ther domestc counterparts arses and wdens durng several years after export market entry. We also fnd that the effect s more pronounced for frms that have hgher skll-ntensty, hgher share of exports n producton, and are small n sze. Overall, the evdence suggests that exportng s one mportant channel through whch domestc frms acqure accesses to advanced knowledge and better technology. Also, the stronger learnng-by-dong effect for frms wth hgher skll-ntensty seems to support the vew that absorptve capacty matters to receve knowledge spllovers from exportng actvty. 1 2 Chn Hee Hahn: Senor Research Fellow, Korea Development Insttute; chhahn@kd.re.kr. Chang-Gyun Park: Assstant Professor, College of Busness Admnstraton, Chung-Ang Unversty; cp19@cau.ac.kr. 279

3 1. Introducton One of the most frequently asked queston n trade and growth lterature s whether and how nternatonal trade or openness of tradng regme promotes productvty growth of countres. Although numerous studes, both theoretcal and emprcal, have been conducted on ths ssue, there seems to be no clear consensus yet. Recently, a growng number of studes have started to utlze frm or plant level data and reexamned ths ssue, partcularly focusng on exportng as a channel of nternatonal technology dffuson or knowledge spllover. One emprcal regularty emergng from these studes s that exporters are more productve than non-exporters. The postve correlaton between exportng and productvty n cross-sectonal context, however, provdes lttle useful nformaton on the drecton of causalty. On one hand, ths could reflect self-selecton nto export market: only productve frms can expect to recoup the sunk entry cost of enterng nto the export market and jon the export market. In ths case, the causalty runs from productvty to exportng. On the other hand, t s also plausble that the postve correlaton between exportng and productvty reflects learnng-by-exportng effect: frms that become exporters could gan new knowledge and expertse after enterng export market and mprove ther productvty relatve to average player n the same ndustry. The self-selecton hypothess s supported by most studes, but the evdence on learnng-by-exportng seems less clear-cut (Tybout 2000). Ths paper examnes the exportng-productvty nexus utlzng the plant level panel data on Korean manufacturng sector (Survey of Mnng and Manufacturng, SMM henceforth) from 1990 to The man queston to be addressed s whether 280

4 exportng actvty mproves productvty performance of plants. The emphass on learnng-by-exportng n the paper stems from the recognton that t s the area where exstng lterature presents mxed emprcal results and, nevertheless, whether or not the learnng-by-exportng effect exsts has an mportant mplcaton on the formulaton of approprate polcy stance toward openness. As dscussed by Bernard and Jensen (1999a), f the gans do accrue to frms once they become exporters, then the approprate polcy nterventons would be those that reduce barrers to enterng foregn markets ncludng macroeconomc trade polces to promote openness to trade and mcroeconomc polces to reduce entry costs, such as export assstance, nformaton programs, jont marketng efforts, and trade credts. On the other hand, f there are no post-entry rewards from exportng, these polces desgned to ncrease the numbers of exporters are more lkely to end up wastng resources. 3 Furthermore, ths paper attempts to clarfy the condtons, f at all, under whch the learnng-by-exportng may or may not take place, utlzng nformaton on some plant or ndustry characterstcs. As plant characterstcs, we consder skll-ntensty, export propensty, plant sze, and R&D ntensty. Most exstng studes utlzed nformaton only on whether a plant exports or not and focused on the exstence of learnng-byexportng effect. However, t s plausble that the degree of learnng-by-exportng could be related to, for example, how mportant exportng actvty s to the plant nvolved, n as much as learnng-by-exportng arses through nteractons wth foregn buyers whch requres costly resources. Thus, we examne whether plants wth hgher export propensty enjoys more benefts of learnng-by-exportng. Meanwhle, f knowledge spllovers from exportng actvtes requre domestc absorptve capacty, 3 See Bernard and Jensen (1999a) for detaled dscusson. 281

5 then we could expect that plants wth hgher absorptve capacty wll exhbt stronger learnng-by-exportng. We use the skll-ntensty of plants as a proxy for the domestc absorptve capacty. We also examne whether the destnaton of exports matter n learnng-byexportng a là Loecker (2007). He shows that the degree of learnng-by-exportng depends on destnaton of exports, usng plant level nformaton on the export destnaton n Slovenan manufacturng. The analyss s based on the presumpton that learnng-by-exportng effect wll be stronger for plants that start exportng to more advanced countres. In case of Korea, however, the plant level nformaton about the export destnaton s not avalable. So, we examne nstead whether plants n ndustres wth hgher share of exports to advanced countres tend to exhbt stronger learnng-by-exportng. Examnng these ssues n the Korean case s partcularly mportant n several respects. Above all, as well recognzed, Korea s one of the few success countres that has narrowed the ncome gap wth advanced countres by adoptng an outwardorented trade strategy. 4 So, examnng and clarfyng the openness-productvty nexus n the Korean case could provde valuable lessons on other developng countres that hope to catch-up wth advanced countres. Furthermore, Korea s a country wth large external exposure n trade that stll needs to make a transton toward a fully developed country. Thus, n so far as learnng-by-exportng, f t exsts, reflects traderelated un-drectonal knowledge spllovers from advanced to less-advanced countres, Korea s the approprate place to examne these ssues. There are some emprcal studes that scrutnze the causal relatonshp between 4 See Krueger (1997), for example. 282

6 exportng and productvty. Most studes report that exporters are more productve than non-exporters before they start to export, suggestng that cross-sectonal correlaton between exportng and productvty partly reflects a self-selecton effect. For example, Clerdes, Lach and Tybout (1998) fnd very lttle evdence that prevous exposure to exportng actvtes mproves performance, usng the plant-level panel data from Colomba, Mexco, and Morocco. Smlar results are reported by Aw, Chung, and Roberts (2000) and Aw, Chen, and Roberts (2001) for Tawan, Bernard and Jensen (1999b) for U.S. By contrast, the evdence on a learnng effect s mxed. Earler research such as Bernard and Jensen (1999b) fnd lttle evdence n favor of learnng. They report that new entrants nto the export market experence some productvty mprovement at around the tme of entry, they are skeptcal about the exstence of strong learnng-by-exportng effect. However, several recent studes utlzng more refned emprcal technque to deal wth self-selecton problem such as matched samplng technques provde some emprcal evdence n favor of learnng-byexportng. See Grma, Greenaway, and Kneller (2002) for UK, Loecker (2007) for Slovena, and Albornoz and Ercolan (2007) for Argentna. Related prevous studes on Korea nclude Aw, Chung, and Roberts (2000) and Hahn (2004). Aw, Chung, and Roberts (2000), usng plant-level panel data on Korean manufacturng for three years spaced at fve-year ntervals, does not fnd evdence n favor of ether self-selecton or learnng-by-exportng. It dffers from smlar studes on other countres n that even the self-selecton hypothess s not supported. Aw, Chung, and Roberts (2000) argue that Korean government s nvestment subsdes ted to exportng actvty rendered plant productvty a less useful gude on the decson to export. By contrast, followng the methodologes of Bernard and Jensen (1999a, 283

7 1999b), Hahn (2004) fnds some supportng evdence for both selecton and learnng n Korean manufacturng sector, usng annual plant-level panel data from 1990 to However, Hahn (2004) suffers from the same techncal dffcultes as Bernard and Jensen (1999a, 1999b) n that the uncontrolled self-selecton problem n export market partcpaton may have contamnated the result. In ths paper, we re-examne the learnng-by-exportng hypothess n Korean manufacturng sector controllng for the self-selecton n export market partcpaton wth a recently developed statstcal tool: propensty score matchng. The organzaton of ths paper s as follows. The followng secton explans the data set and the calculaton of plant total factor productvty. Secton 3 brefly dscusses the estmaton strategy to overcome the dffcultes arsng from self-selecton n decson makng for export market partcpaton and to obtan a better estmate for the effects of learnng-by-exportng. Secton 4 dscusses our man emprcal results and the fnal secton concludes. 2. Data and Plant Total Factor Productvty 2.1. Data Ths paper utlzes the unpublshed plant-level census data underlyng the Survey of Mnng and Manufacturng n Korea. The data set covers all plants wth fve or more employees n 580 manufacturng ndustres at KSIC (Korean Standard Industral Classfcaton) fve-dgt level. It s an unbalanced panel data wth about 69,000 to 97,000 plants for each year from 1990 to For each year, the amount of exports 284

8 as well as other varables related to producton structure of plants, such as producton, shpments, the number of producton and non-producton workers and the tangble fxed nvestments, are avalable. The exports n ths data set nclude drect exports and shpments to other exporters and wholesalers, but do not nclude shpments for further manufacture Plant Total Factor Productvty Plant total factor productvty (TFP) s estmated followng the chanedmultlateral ndex number approach as developed n Good (1985) and Good, Nadr, and Sckles (1997). Ths procedure uses a separate reference pont for each crosssecton of observatons and then chan-lnks the reference ponts together over tme. The reference pont for a gven tme perod s constructed as a hypothetcal frm wth nput shares that equal the arthmetc mean nput shares and nput levels that equal the geometrc mean of the nputs over all cross-secton observatons. Thus, output, nputs, and productvty level of each frm n each year s measured relatve to the hypothetcal frm at the base tme perod. Ths approach allows us to make transtve comparsons of productvty levels among observatons n panel data set. 5 Specfcally, the productvty ndex for frm at tme t n our study s measured n the followng way. 5 Good, Nadr, and Sckles (1996) summarze the usefulness of channg multlateral productvty ndces. Whle the channg approach of Tornqvst-Thel ndex, the dscrete Dvsa, s useful n tme seres applcatons where nput shares mght change over tme, t has severe lmtatons n cross-secton or panel data framework where there s no obvous way of sequencng the observatons. To the contrary, the hypothetcal frm approach allows us to make transtve comparsons among cross-secton data, whle t has an undesrable property of sample dependency. The desrable propertes of both channg approach and the hypothetcal frm approach can be ncorporated nto a sngle ndex by chaned-multlateral ndex number approach. 285

9 lntfp N t t lny lny lny lny 1 ( S 2 nt t S nt t )(ln X nt 2 ln X ) 1 n1 2 n1 nt t N 1 ( 2 S n S n 1 )(ln X n ln X n 1 ) (1) where Y, X, S, and TFP denote output, nput, nput share, TFP level, respectvely, and symbols wth an upper bar are correspondng measures for the hypothetcal frm. The subscrpts and n are ndces for tme and nputs, respectvely. The year 1990 s chosen as the base year. As a measure of output, we use the gross output (producton) of each plant n the Survey deflated by the producer prce ndex at dsaggregated level. The captal stock used n ths paper s the average of the begnnng and end of the year book value of captal stock n the Survey deflated by the captal goods deflator. As for labor nput, we use the number of workers, whch ncludes pad employees 6, workng propretors and unpad famly workers. We allowed for the qualty dfferental between producton workers and all other types of workers. The labor qualty ndex of the latter was calculated as the rato of non-producton workers and producton workers average wage at each plant, averaged agan over the entre plants n a gven year. The sum of major producton cost and other producton cost reported n the Survey was taken as the measure of ntermedate nput. Major producton cost covers costs arsng from materals, parts, fuel, electrcty, water, manufactured goods outsourced and mantenance. Other producton cost covers expendtures on outsourced servces such as advertsng, transportaton, communcaton and nsurance. The estmated ntermedate nput was deflated by the ntermedate nput prce ndex. 6 Pad employees s the sum of producton and non-producton workers. 286

10 We assumed constant returns to scale producton technology so that the sum of factor elastctes equals to one. Labor and ntermedate nput elastctes for each plant are measured as average factor cost shares wthn the same plant-sze class n the fve-dgt ndustry n a gven year. Here, plants are grouped nto three sze classes accordng to the number of employees; 5-50, , and over 300. Thus, the factor elastctes of plants are allowed to vary across ndustres and plant sze classes and over tme Defnton of Exporters Followng conventon n the lterature, we defne an exporter n a gven year as a plant reportng postve amount of exports. Accordngly, non-exporters n a gven year are those plants wth zero exports. Wth ths defnton of exporters, t s possble to classfy all plants nto fve sub-groups: Always, Never, Starters, Stoppers, and Other. 7 Always s a group of plants that were exporters n the year that they frst appear n the data set and never changed ther exportng status. Smlarly, Never s a group of plants that were non-exporters n the year that they frst appear n the data set and never swtched to exporters. Starters ncludes all plants that were nonexporters n the year that they frst appear, but swtched to exporters n some later year and remaned as exporters thereafter. Stoppers conssts of all plants that were exporters n the year that they frst appear, and then swtched to non-exporters, never swtchng back to exporters thereafter. All other plants that swtched ther exportng status more than twce durng the sample perod are grouped as Other. 7 We elmnated plants that swtch n and out of the dataset more than twce durng the sample perod. Thus, we keep only those plants that do not have a splt n tme seres observatons. Ths procedure elmnates about 10 percent of the sample n terms of number of plants. 287

11 2.4. A Prelmnary Analyss: Performance of Exporters and Non-exporters Table 1 shows the number of exportng plants and average exports as percentage of shpments, or export ntensty, for each year durng the sample perod. Exportng plants accounted for between 11.0 and 15.3 percent of all manufacturng plants. The share of exportng plants rose slghtly between 1990 and 1992, but snce then steadly declned untl However, wth the outbreak of the fnancal crss n 1997, the share of exportng plants rose somewhat notceably to reach 14.8 percent n The rse n the share of exportng plants can be attrbuted mostly to the closure of nonexportng plants, rather than ncrease n the number of exportng plants. Note that the ncreases n the number of exporters n 1997 and 1998 were modest, whch are broadly consstent wth the severe contracton of domestc demand and huge deprecaton of Korean Won assocated wth the crss. Table 1. Number of Exporters and Export Intensty Year Total number of plants (percent) Nonexporters (percent) Exporters (percent) Exports/shpments rato (percent) unweghted weghted ,690 58,392 10, (100) (85.0) (15.0) ,213 61,189 11, (100) (84.7) (15.3) ,679 63,241 11, (100) (84.7) (15.3) ,864 77,514 11, (100) (87.2) (12.8) ,372 80,319 11, (100) (87.9) (12.1) ,202 85,138 11, (100) (88.5) (11.5) ,141 86,502 10, (100) (89.0) (11.0) ,138 80,963 11, (100) (87.9) (12.1) ,544 67,767 11, (100) (85.2) (14.8) Source: Hahn (2004). 288

12 Consstent wth the hgh export propensty of the Korean economy, the share of exports n shpments at plant level s qute hgh. Durng the sample perod, the unweghted mean export ntensty s between 43.6 and 54.8 percent, declnng from 1990 to 1996 but rsng wth the onset of the crss n The average export ntensty weghted by shpment shows a smlar pattern, wth generally lower fgures than the unweghted average, suggestng that smaller exportng plants have a hgher export ntensty. It s a well-establshed fact that exporters are better than non-exporters by varous performance standards. Table 2 compares varous plant attrbutes between exporters and non-exporters for three selected years. Frst, exporters are on average much larger n the number of workers and shpments than non-exporters. The dfferental n shpments s more substantal than that n the number of workers. So, the average labor productvty of exporters measured by ether producton per worker or value added per worker s hgher than that of non-exporters. Compared wth the cases of value added, the dfferental n producton per worker between exporters and nonexporters s more pronounced. Ths mght reflect a more ntermedate-ntensve producton structure of exporters relatve to non-exporters. Although exporters show both hgher captal-labor rato and a hgher share of non-producton workers n employment than non-exporters, they do not fully account for the dfferences n labor productvty. As a consequence, total factor productvty levels of exportng plants are, on average, hgher than those plants that produce for the domestc market only. Some dfferences n the total factor productvty may be attrbuted to the dfferences n R&D ntensty. Note that, controllng for the sze of shpments, exporters spent about twce as much on R&D as non-exporters. From a worker s pont of vew, exporters 289

13 had more desrable attrbutes than non-exporters. That s, the average wage of exporters s hgher than that of non-exporters. Although both a producton worker s wage and a non-producton worker s wage are hgher n exporters than n nonexporters, the dfferental n the non-producton worker s wage s more pronounced. Table 2. Performance Characterstcs of Exporters vs. Non-exporters exporters nonexporters exporters nonexporters exporters nonexporters Employment (person) Shpments (mllon won) producton per worker (mllon won) value-added per worker (mllon won) , , , , , TFP captal per worker (mllon won) non-producton worker/ total employment (percent) average wage (mllon won) Average producton wage (mllon won) average non-producton wage (mllon won) R&D/shpments (percent) Source: Hahn (2004). 290

14 3. Emprcal Strategy: Propensty Score Matchng It s now well-recognzed n the lterature that the decson to become an exporter s not a random event but a result of delberate choce, requrng specal efforts to correctly dentfy the true effect of becomng an exporter on ts productvty (Loecker 2007, Albornoz and Ercolan 2007). The partcpaton decson n the export market s lkely to be correlated wth the stochastc dsturbance terms n the data generatng process for a frm s productvty, so that the tradtonal smple mean dfference test on productvty dfferences between exporters and non-exporters does not provde the correct answer. The matchng method has been ganng popularty among appled researchers snce t s vewed as a promsng analytcal tool wth whch we can cope wth statstcal problems stemmng from an endogenous partcpaton decson. The underlyng motvaton for the matchng method s to reproduce the treatment group (exporters) out of the non-treated (non-exporters), so that we can reproduce the experment condtons n a non-expermental settng. Matched samples enable us to construct a group of pseudo-observatons contanng the mssng nformaton on the treated outcomes had they not been treated by parng each partcpant wth members of the non-treated group. The crucal assumpton s that, condtonal on some observable characterstcs of the partcpants, the potental outcome n the absence of the treatment s ndependent of the partcpaton status. y 0 d X (2) where 0 y s the potental outcome n the absence of the treatment, d s the dummy to ndcate partcpaton, and X s the vector of condtonng varables. The basc 291

15 dea of matchng s to construct a sample analog of a counter factual control group by dentfyng the members of a non-partcpatng group that possess condtonng varables as close to those of treatment group as possble. In practce, t s very dffcult to construct a control group that satsfes the condton n (2), especally when the dmenson of the condtonng vector X s hgh. Rosenbaum and Rubn (1983) propose a clever way to overcome the curse of dmensonalty n the tradtonal matchng method. Suppose that the condtonal probablty of frm s becomng an exporter can be specfed as a functon of observable characterstcs of the frm before the partcpaton; p X d 1 X Ed X Pr (3) Rosenbaum and Rubn (1983) call the probablty functon n (3) propensty score and show that f the condtonal ndependence assumpton n (2) s satsfed t s also vald for p X that y 0 d p X (4) We have replaced the mult-dmensonal vector wth a one-dmensonal varable contanng the same nformaton contents so that the hghly complcated matchng problem n (2) s reduced to a smple sngle dmensonal one n (4). One can defne the average treatment effect on the treated () as; E E y y d 1 EEy y d 1, px 1 0 E y d 1, px Ey d 1, px (5) where 0 y s the potental outcome that would have been observable had partcpatng 292

16 frm decded not to partcpate n an export market and y 1 s the observable outcome for partcpatng frm. Note that s not the measure for the effect of exportng on all frms but on frms that start to export. Snce 0 y s not observable, the defnton (5) s not operatonal. Gven that the unconfoundedness condton under propensty score (4) s satsfed and the propensty score (3) s known, the followng defnton s equvalent to (5) y y d 1 EEy d 1, px Ey d 0 px E, (6) Snce both 0 y and y 1 are observable n (6), one can construct an estmator for by constructng ts sample analog. As the frst step, we estmate the probablty functon n (3) wth the followng probt specfcaton. p 2 1 z (7) 2 2 ' X X :, 1 exp dz Log of total factor productvty, log of the number of workers employed, log of captal per worker, 9 yearly dummes, and 10 ndustry dummes are ncluded n the condtonng vector X. As for condtonng varables, we use the values from one year before the frm starts to export n order to account for the tme dfference between decson to partcpate and actual partcpaton. Based on estmated verson of (7), one can calculate propensty score for all observatons, partcpants and non-partcpants. Let T be the set of treated (exportng) unts and C the set of control (non-exportng) unts, respectvely, and denote by C 293

17 the set of control unts matched to the treated unt wth an estmated value of propensty score of p. Then, we pck the set of nearest-neghbor matchng as; C mn p p (8) j j Denote the number of controls matched wth a treated unt T by C N and defne the weght w j 1 N f C C j and w 0 otherwse. Then, the propensty j score matchng estmator for the average treatment effect on the treated at tme t s gven by; * 1 t T N T y 1, t jc w j y 0 j, t (9) where 1 y,t s the observed value on frm n the treatment group at tme t and 0 y j,t the observed value on frm j n the matched control group for frm at tme t. Moreover, one can easly show that the varance of the estmator n (9) s gven by; Var 1 1 Vary * 1 t t T, T 2 N N T jc w j 2 Var y 0 j, t (10) One can estmate an asymptotcally consstent estmator for (10) by replacng two varance terms for the treatment and control groups wth correspondng sample analogs. We use two dfferent versons of the propensty score matchng procedure wrtten n STATA language; attn.ado explaned n Becker and Ichno (2002) (BI, hereafter) and psmatch2.ado provded by Leuven and Sanes (2008) (LS, hereafter). The two 294

18 procedures follow an dentcal approach n estmatng propensty score and constructng the control group, except for the fact that the former tres to verfy the unconfoundedness condton n the sample by dvdng the entre regon of estmated propensty scores nto several blocks and construct the matched control group wthn the block to whch the treated observaton belongs. In order to allow for the possblty that the effect of learnng by exportng works at dfferent ntenstes dependng on a frm s characterstcs and ndustry, we dvde the entre sample nto several categores accordng to plant or ndustry characterstcs, such as the export ntensty of plants, skll ntensty of plants, plant sze measured by the number of workers, R&D ntensty of plants, and export destnaton of ndustres. We measure the average treatment effect of the treated for each sub-sample. 4. Emprcal results: Learnng-by-exportng Effects 4.1. Starter vs Non-exporter Table 3 reports the estmated productvty gan from partcpatng n an export market when heterogenety n treatment effect s not taken nto account. The estmated coeffcents ndcate percentage productvty dfferentals between plants that start exportng and ther domestc counter-parts s years after enterng the export market. We report results from the two dfferent versons of propensty score matchng procedure, BI and LS. 295

19 Table 3. Average Productvty Gan of Exporters Matchng Method IB LS s = 0 s = 1 s = 2 s = *** (0.008) *** (0.010) *** (0.011) *** (0.014) No. Treated No. Controls *** (0.008) *** (0.011) *** (0.014) *** (0.019) No. Treated No. Controls Frst and foremost, all estmated coeffcents are postve and hghly sgnfcant, suggestng the exstence of a learnng-by-exportng effect. Ths s qute a surprse fndng consderng the fact that most prevous studes were skeptcal about the exstence of the learnng-by-exportng effect. Second, productvty gan for starters begns to materalze mmedately after enterng the export market, and the productvty gap between the starters and non-exporters 8 wdens further as tme passes, although at a deceleratng pace. Thrd, t seems that the choce of procedures n constructng the control group does not yeld any materal dfferences n the fnal result, not only qualtatvely but also qualtatvely. The estmated coeffcents from BI procedure ndcate that starters become about 4.1 percent more productve n the year of entry. Over the followng years, productvty gan for starters fluctuates between 6.4 and 7.7 percentage ponts. Thus, t s suggested that enterng the export market has a permanent effect on productvty level, especally durng the frst several years after entry. In other words, export market entry has a temporary effect on productvty growth especally durng the frst few years after entry. 8 Non-exporters correspond to the never group n our earler defnton. 296

20 4.2. Sub-group Estmaton: Plant Characterstcs In order to allow for a dfferental treatment effect dependng on plant characterstcs, we dvded our sample nto three sub-groups accordng to varous features such as an exports-producton rato, the skll ntensty, plant sze measured by the number of workers, and R&D-producton rato. Then we apply the matchng estmators dscussed n Secton 3 and estmate the learnng-by-exportng effect separately for each sub-group. Based on BI procedure 9, we report the estmated productvty gans for starters n each sub-group n Table 4. Frst, the estmated coeffcents are generally larger and more sgnfcant for plants wth hgher exports-producton rato. For example, n the group of low export ntensty wth exports-producton rato of less than 10%, starters become more productve, between 2.5 and 4.1 percent durng the three years after the partcpaton. By contrast, n the group of hgh export ntensty wth an exports-producton rato greater than 50%, productvty gans for starters are between 9.5 and 11.4 percent for the same tme span. In the earler secton, we argued that f the estmated effect of learnng-by-exportng ndeed captures the benefcal consequences of learnng actvtes assocated wth exportng, then the effect s lkely to be stronger for plants wth hgher exports-output ratos; f learnng-by-exportng arses from contact wth foregn buyers and foregn markets, whch requre costly resources, then frms for whom exportng s ther major actvty are lkely to be more heavly exposed to foregn contact and experence productvty gan. The results for sub-groups wth dfferent export ntenstes are very consstent wth ths hypothess. 9 Estmaton results based on LS procedure are reported n the appendx. 297

21 Table 4. Average Productvty Gan of Starters by Frm Characterstcs: BI Procedure Frm Characterst cs Export Rato Skll Intensty Plant Sze (Number Of Workers) R&D Intensty Low Medum Hgh Low Medum Hgh Small Medum Large None Low Medum Hgh s = 0 s = 1 s = 2 s = *** 0.041*** ** (0.013) (0.015) (0.018) (0.020) No. Treated No. Controls *** 0.081*** 0.071*** (0.013) (0.015) (0.017) (0.021) No. Treated No. Controls *** 0.112*** 0.114*** 0.095*** (0.014) (0.016) (0.019) (0.021) No. Treated No. Controls (0.020) (0.027) (0.033) (0.046) No. Treated No. Controls *** 0.054*** 0.065*** 0.033** (0.009) (0.010) (0.012) (0.014) No. Treated No. Controls *** 0.065*** 0.068*** 0.072*** (0.017) (0.022) (0.024) (0.027) No. Treated No. Controls *** 0.124*** 0.207*** 0.177*** (0.015) (0.020) (0.027) (0.033) No. Treated No. Controls *** 0.055*** 0.058*** 0.049*** (0.010) (0.011) (0.013) (0.016) No. Treated No. Controls *** (0.020) (0.023) (0.027) (0.028) No. Treated No. Controls *** 0.065*** 0.08*** 0.069*** (0.009) (0.010) (0.012) (0.014) No. Treated No. Controls (0.035) (0.036) (0.042) (0.044) No. Treated No. Controls (0.031) (0.038) (0.046) (0.041) No. Treated No. Controls (0.048) (0.061) (0.077) (0.073) No. Treated No. Controls

22 Second, the learnng-by-dong effect seems to be more pronounced for plants wth hgher skll ntensty 10. For the group of plants wth a skll ntensty of less than 10%, starters became more productve, between 1.5 and 2.6 percentage ponts durng the three years after begnnng to export. For the group of plants wth a skll ntensty greater than 40%, starters became and remaned between 9.5 and 11.4 percentage ponts more productve durng the same perod. These results suggest that domestc absorptve capacty matters for exportng plants to take advantage of the benefts of nternatonal knowledge spllovers. Specfcally, the result on the correlaton between skll ntensty and productvty gan from startng to export n Table 4 s consstent wth the prevous emprcal lterature that emphaszes the role of human captal n facltatng technology adopton (Welch 1975, Bartel and Lchtenberg 1987, Foster and Rosenzweg 1995, Benhabb and Spegel 1994) 11. Thrd, we also examne whether the degree of learnng-by-exportng s related to plant sze, dvdng the entre sample nto three groups: a group of small plants wth the number of workers less than 10, a group of medum-szed plants wth the number of workers between 11 and 49, and a group of large plants wth 50 or more workers. Table 4 suggests that effect of learnng-by-exportng s generally larger and more sgnfcant for smaller plants. As argued by Albornoz and Ercolan (2007), there seems to be no a pror reason to expect larger learnng-by-exportng effects for small exporters. 12 Whle one can argue that large frms are generally more structured and better suted to facltate absorpton and use new knowledge obtaned through 10 Skll ntensty s measured by the share of non-producton workers out of the total of producton and non-producton workers. 11 These studes are emprcal nvestgatons of Nelson-Phelps hypothess whch suggests that the rate at whch the gap between the technology fronter and the current level of productvty s closed depends on the level of human captal. See Benhabb and Spegel (2005) for detaled explanaton. 12 They also fnd that small frms learn more from exportng actvtes usng frm-level panel data on Argentnan manufacturng. 299

23 exportng actvtes, t s also possble to argue that knowledge mght be easer to dssemnate n a small frm due to ts flexblty and smplcty of organzatonal structure and ts decson makng process. Our fndngs n Table 4 seem to suggest that the latter effect domnates. Fnally, we examne whether plants wth hgher R&D nvestment exhbt a larger learnng-by-exportng effect. To do so, we classfy plants nto four sub-groups: a group wth no R&D nvestment, a low R&D group wth a rato of R&D expendture to producton less than 2 percent, a medum R&D group wth a rato from 2 to 10 percent and a hgh R&D group wth a rato hgher than 10 percent. Somewhat surprsngly, the learnng-by-exportng effect s statstcally sgnfcant only n the no R&D group. Although we cannot come up wth a clear explanaton for the results, we can conjecture that R&D ntensty reflects ndustry specfc characterstcs rather than the nnovatveness of frms Sub-group Estmaton: Export Destnatons as an Industry Characterstc As far as we are aware of, lttle s known about ndustry characterstcs that affect the degree of learnng-by-exportng. In ths subsecton, we examne whether the export destnaton of ndustry as an ndustry characterstc affects the strength of learnng-by-exportng of the plants. If the learnng-by-exportng effect found n ths paper captures nternatonal knowledge spllovers from advanced to less advanced countres whch arse through the contact wth foregn buyers n more advanced countres, then we could expect to fnd that the learnng-by-exportng effect s stronger n ndustres that have larger share of ther exports drected to more advanced 13 It s a well known fact that R&D ntensty vares a lot across ndustres 300

24 countres. However, we cannot expect that learnng-by-exportng wll be stronger unambguously n ndustres wth a larger share of exports drected to more advanced countres for many reasons, ncludng the followng. Frst of all, nternatonal knowledge spllovers mght arse not only through drect contact wth foregn buyers n advanced countres but also through ndrect contact wth foregn compettors n the markets of less advanced countres. For example, Korea s car exporters could learn from the busness practces of German car exporters n the Chnese market. Secondly, generally more ntense competton n export markets can exert pressure on frms that start to export to mprove ther productve effcency. Then the degree of competton n an export market could be an mportant factor n determnng the degree of learnng-by-exportng effect. Thrdly, there should be an ndustry-level technology gap between the exportng country and the fronter country n order for the learnngby-exportng effect to take place. That s, there should be some advanced knowledge out there to learn from n the frst place. If ths s the case, then the drecton of exports would be mmateral for an ndustry that s at or close to the world fronter. 14 Fourthly, f exportng s assocated wth fragmentaton of producton by multnatonal frms, then effcency mprovement comng from the fragmentaton of producton whch, n some cases, nvolves exportng to lower ncome countres wthn the producton network mght be captured as learnng-by-exportng effect. Kmura, Hayakawa, and Matsuura (2009) provde a theoretcal explanaton related to ths story. They show that n the case of vertcal FDI, the larger the gap n captal-labor ratos 14 Ths mght be one reason that learnng-by-exportng effect s occasonally reported n studes of developng countres but not n developed countres, such as the U.S. 301

25 between a Northern fragment and a Southern fragment, the greater the total cost reducton n nternatonal fragmentaton. In ths case, exportng to lower ncome countres wthn a producton network mght be assocated wth a greater learnng-byexportng effect. Although explorng all these possbltes s out of the scope of ths paper, we thnk that examnng whether the drecton of exports matters for the strength of learnng-bydong s the frst step toward understandng the exact nature of the learnng-byexportng effect captured n ths paper. As a prelmnary step, we frst examne whether there are cross-ndustry dfferences n productvty gans from becomng exporters. To do so, we dvded our sample nto 10 sub-ndustres 15 and repeated the matchng procedure for each ndustry. Table 5 shows that productvty gans from learnng-by-exportng are vsble n the textle and apparel, chemcal, metal, and transport equpment ndustres. However, we cannot fnd sgnfcant productvty gans n the food, wood and pulp, general machnery, precson nstrument, and electroncs ndustres. Roughly speakng, the former group of ndustres largely concdes wth the area for whch Korea s beleved to have a comparatve advantage. Therefore, the result can be nterpreted as provdng a pece of evdence supportng the hypothess that nvolvement n exportng actvtes results n productvty gans. However, t s somewhat surprsng that we can fnd no sgnfcant evdence for the exstence of a learnng-by-exportng effect n the electroncs ndustry. Although we could conjecture that ths reflects that many Korean producers n the electroncs ndustry are the fronter producers, a more defntve assessment cannot be made untl a more n-depth analyss s carred out. 15 They are food, textle and apparel, wood and pulp, chemcal, metal, general machnery, electroncs, precson nstrument, transport equpment, and others. 302

26 Nevertheless, Table 5 seems to show that there are some ndustry characterstcs that affect the strengths of the learnng-by-exportng effect. Table 5. Average Productvty Gan of Starters by Industry: BI Procedure Industry s=0 s=1 s=2 s= (0.038) (0.042) (0.052) (0.058) Food No. Treated No. Controls Textle and Apparel Wood and Pulp Chemcal Metal General Machnery Electroncs Precson Instrument Transport Equpment Other 0.099*** 0.117*** 0.129*** 0.097*** (0.018) (0.019) (0.021) (0.025) No. Treated No. Controls *** (0.033) (0.039) (0.042) (0.054) No. Treated No. Controls * 0.158*** (0.021) (0.028) (0.033) (0.035) No. Treated No. Controls *** 0.09** (0.029) (0.038) (0.044) (0.045) No. Treated No. Controls (0.015) (0.019) (0.024) (0.026) No. Treated No. Controls (0.026) (0.031) (0.033) (0.033) No. Treated No. Controls (0.048) (0.056) (0.054) (0.074) No. Treated No. Controls ** 0.15*** (0.040) (0.045) (0.052) (0.051) No. Treated No. Controls (0.029) 0.071* (0.040) 0.1** (0.050) 0.183*** (0.055) No. Treated No. Controls

27 We next turn to the export destnatons of ndustres as one possble factor explanng dfferental strengths of the learnng-by-exportng effect estmated at the sub-group level of ndustres. As explaned above and also n Loecker (2007), ths hypothess s based on the presumpton that a learnng-by-exportng effect wll be stronger for plants that start exportng to more advanced countres, where the opportuntes for learnng new knowledge and technology are relatvely abundant. Although Loecker (2007) examned ths ssue usng plant-level nformaton on the destnaton of exports, we do not have such nformaton avalable for Korea. Instead, we examne whether plants n ndustres wth a hgher share of exports to advanced countres exhbt hgher productvty gans. 16 To do so, we frst matched the drecton of exports dataset at SITC 5 dgt level compled from UNComtrade (Rev. 3) wth the Mnng and Manufacturng Survey dataset at KSIC 17 three-dgt level. Then, we classfed Korea s export destnaton countres nto two groups: lower-ncome and hgher-ncome countres. Here, hgher-ncome countres are those wth an average per capta GDP for the perod from 1990 to 1998 larger than that of Korea. The remanng countres are lower-ncome countres. Next, for each of the 58 three-dgt manufacturng ndustres, we calculated ther shares of exports to lower-ncome and hgher-ncome countres averaged over the same perod. Then, we classfed each ndustry nto hgher-ncome or lowerncome group f ts share of exports to hgher-ncome countres s greater or smaller than lower-ncome countres, respectvely. The estmated productvty gan for starters s reported n Table 6 for each sub- 16 In some respect, drecton of exports s more lkely to be an ndustry characterstc rather than plant characterstc. 17 Korean Standard Industral Classfcaton. 304

28 group. At frst glance, the results are not supportve of the hypothess that the learnng-by-exportng effect s more pronounced n ndustres wth more of ther exports drected to more advanced countres. In fact, the result s the other way around: Learnng-by-exportng effect n the lower-ncome group s stronger than that of the hgher-ncome group, although both are hghly sgnfcant. We conjecture that the result s drven by the fact that the gan from partcpatng n export markets depends on many factors convenently branded as the benefts of openness. We beleve that those factors must be nterlnked n a very complcated fashon and a smple approach lke ours cannot gve the defnte answer to ths mportant queston. Table 6. Average Productvty Gan of Starters by Export Destnatons: BI Procedure Hgher-ncome Lower-ncome t = 0 t = 1 t = 2 t = *** (0.011) 0.068*** (0.013) 0.057*** (0.016) 0.036* (0.020) No. Treated No. Controls *** (0.012) 0.079*** (0.013) 0.081*** (0.016) 0.074*** (0.020) No. Treated No. Controls Gven the nadequate control of varous factors that mght be relevant for determnng the degree of learnng-by-exportng effect, the above results should not be taken as a defntve pece of evdence aganst the hypothess that the learnng-byexportng effect s larger n ndustres wth more of ther exports drected to hgherncome countres. We thnk that varous ndustry as well as plant characterstcs mght also play a role here. Further analyss seems to be warranted to shed lght on ths ssue. 305

29 5. Concluson Ths paper examned the presence of a learnng-by-exportng effect utlzng a unque plant level panel data coverng all manufacturng sectors n Korea. Korean experences offer a good wndow of opportunty to analyze ths ssue n the sense that Korea s one of the best known success stores havng acheved fast economc growth drven by outward-orented development strateges. We fnd clear and robust evdences for a learnng-by-export effect. The total factor productvty gap between exporters and ther domestc counterparts s sgnfcant and shows the tendency to wden durng three years after entry nto the export market. We also fnd that the benefcal effect of productvty gan s more pronounced for plants wth a hgher skll-ntensty or hgher share of exports n producton. Although ths paper examned the learnng-by-exportng effect, t should be born n mnd that learnng-by-exportng s just one of many channels through whch the benefts of openness are realzed. That s, the results of ths paper does not at all exclude the possblty that the benefcal effects of openness are realzed through varous other channels, such as ncreases n consumer surpluses and mprovements of allocaton effcency, knowledge spllovers and market-dscplnng effects from mports, and mprovement of scale effcency, among others. One nterestng polcy mplcaton whch arses from ths paper mght be that neoclasscal orthodoxy of prescrbng uncondtonal openness polcy 18 mght not be entrely warranted. If domestc absorptve capacty s complementary to the openness polcy, as suggested by the evdence of larger a learnng-by-exportng effect n skll- 18 See Sachs and Warner (1995), for example. 306

30 ntensve plants, then upgradng the qualty of human captal mght be necessary to more fully utlze the benefts from openness. 307

31 Table A.1. Average Productvty Gan of Starters by Frm Characterstcs: LS Procedure Plant Characterstcs Export Rato Skll Intensty Number Of Workers R&D Low Medum Hgh Low Medum Hgh Low Medum Hgh None Low Medum Hgh s=0 s=1 s=2 s= *** (0.011) (0.016) (0.022) (0.026) No. Treated No. Controls *** 0.052** (0.012) (0.018) (0.024) (0.033) No. Treated No. Controls *** 0.109*** 0.105*** 0.074** (0.013) (0.019) (0.025) (0.035) No. Treated No. Controls ** 0.099** (0.016) (0.026) (0.037) (0.050) No. Treated No. Controls *** 0.046*** 0.043*** (0.009) (0.013) (0.017) (0.025) No. Treated No. Controls *** 0.057** 0.063** 0.104*** (0.017) (0.025) (0.033) (0.041) No. Treated No. Controls *** 0.074*** 0.108*** (0.015) (0.026) (0.042) (0.060) No. Treated No. Controls *** 0.059*** 0.069*** 0.084*** (0.010) (0.014) (0.018) (0.024) No. Treated No. Controls (0.019) (0.024) (0.030) (0.040) No. Treated No. Controls *** 0.041*** 0.055*** 0.039* (0.008) (0.012) (0.015) (0.022) No. Treated No. Controls (0.035) (0.041) (0.049) (0.066) No. Treated No. Controls (0.030) (0.038) (0.056) (0.068) No. Treated No. Controls (0.047) (0.062) (0.086) (0.132) No. Treated No. Controls

32 Table A.2. Productvty Gan of Starters by Industry: LS Procedure Industry s=0 s=1 s=2 s=3 Food Textle and Apparel Wood and Pulp Chemcal Metal General Machnery Electroncs Precson Instrument Transport Equpment Other 0.074** (0.036) (0.052) (0.063) (0.064) No. Treated No. Controls *** (0.016) 0.128*** (0.024) 0.145*** (0.030) 0.113*** (0.042) No. Treated No. Controls (0.036) (0.051) (0.059) (0.097) No. Treated No. Controls ** (0.019) (0.030) 0.086** (0.037) 0.091* (0.047) No. Treated No. Controls * (0.027) (0.040) 0.123** (0.054) (0.064) No. Treated No. Controls (0.014) (0.022) (0.034) (0.043) No. Treated No. Controls (0.023) (0.033) (0.042) (0.048) No. Treated No. Controls (0.043) (0.062) (0.078) (0.091) No. Treated No. Controls (0.038) (0.048) (0.075) 0.124* (0.075) No. Treated No. Controls (0.028) (0.040) 0.108** (0.051) (0.087) No. Treated No. Controls

33 Table A.3. Average Productvty Gan of Starters by Export Destnatons: LS procedure t = 0 t = 1 t = 2 t = 3 Hgher-ncome Lower-ncome Other lowerncome East Asa 0.021** (0.010) (0.015) (0.022) (0.027) No. Treated No. Controls *** (0.011) 0.034** (0.016) 0.078*** (0.021) 0.091*** (0.030) No. Treated No. Controls *** (0.015) 0.068*** (0.022) 0.104*** (0.026) 0.091*** (0.038) No. Treated No. Controls * (0.015) 0.062*** (0.021) (0.032) (0.042) No. Treated No. Controls