Is "Learning-by-Exporting" Important? Micro-dynamic Evidence from Colombia, Mexico, and Morocco

August 5, 1996 Is “Learning-by-Exporting” Important? Micro-dynamic Evidence from Colombia, Mexico and Morocco SofronisClerides Yale University SaulLach The Hebrew University,the Federal Reserve BoardandNBER James Tybout - Georgetown Universityand the World Bank Abstract: Isthere anyempirical evidencethat firms become more efficient after becoming exporters? Do firmsthatbecome exporters generatepositive spilloversfordomestically-oriented producers? Inthis paperwe analyzethe causal linksbehveenexporting and productivityusingfirm-level paneldata from three semi-industrialized countries. Representing export market participation and production costsas jointly dependentautoregressive processes,we lookfor evidencethatfirms’ stochasticcost processes shifiwhen they break intoforeign markets. We findthat relatively efficient firms become exporters, but firms’ unitcostsare notaffected bypreviousexportmarket participation. Sothewell-known efficiency gap between exporters and non-exporters isdueto self-selection ofthemore efficient firms into the exportmarket, ratherthan learning byexporting. Further, we find someevidencethat exportersreduce thecostsofbreaking intoforeign markets fordomestically oriented producers, butthey donotappear to helpthese producersbecome more efficient. Acknowledgements: Wethank Wolfgang Keller, Don Keesing, Yair Mundlak and Dani Rodrik fortheir comments, aswell as seminarparticipantsatJohns Hopkins University,UCLA, theUniversity of Maryland, the World Bank, and theNBER Summer Institute (International Trade and Investment session). Thispaper was funded bythe WorldBank research project “Micro-foundations of Successful ExportPromotion,”RPO 679-20andThe World Bank’s International Trade Division, International Economics Department. Theviewsexpressed herein are thoseofthe authors, and do notnecessarily reflect thoseofthe World Bank ortheFederal Reserve S>’stem.

I. Introduction Many analystsbelievethatexport-led development strategies improvetechnical efficiency. And oneoft-cited reason isthatexporters may benefit fromthetechnical expertiseoftheir buyers! ... agooddealofthe information needed toaugment basiccapabilitieshascome from thebuyersofexportswho freely provided productdesigns andoffered technical assistanceto improveprocess technology inthe contextoftheirsourcingactivities. Somepartofthe efficiency of export-led development musttherefore beattributedto externalities derived from exporting.” (Evenson and Westphal, 1995). Buyerswant low-cost, better-quality productsfrom major suppliers. To obtainthis,they transmit tacitand occasionally proprieta~ knowledge fromtheir other,ofien OECDeconomy, suppliers.(World Bank, 1993,p.320). The importantthing about foreign buyers,many ofwhich have offices inSeoul,isthat they do much morethan buy and specify. .. They come in,too, with modelsandpatterns forKorean engineers to follow, andthey even go outtothe production linetoteach workers howto dothings (Rhee, Ross-Larson, andPursell, 1984,p.41) In supportofthis view, empirical studiesofien findthat exportingplantsaremore efficient than their domestically oriented counterparts (Aw and Hwang, 1995;Bernard andJensen, 1995,Chen and Tang, 1987;Haddad, 1993;Handoussa, Nishimizu and Page, 1986;Tyboutand Westbrook, 1995; Roberts, SullivanandTybout, 1995). But,with theexception of Bernard and Jensen (1995)noneofthese studieshasasked thequestion ofwhether exportingcausesefficiency gains. Plausiblearguments can be made forcausalityto flow inthe opposite direction: relatively more efficient plants self-select intoexport markets becausethereturnsto doing soare relatively highforthem. Inthis paperwe attempt to sortoutthe direction ofcausality, and insodoing,determine whether there isevidencethat firms learnto bemore efficient by becoming exporters. Further,to assessthe case foractiveexport promotion,we testwhether exportersgenerate external benefitsforotherfirms, eitherby acting asaconduit forknowledge thatthey acquirethrough trade, or bycausing improvementsin 1Forarecentcatalogofadditionalreasons,seeWorldBank(1993,pp.316-324). 1

internationaltransportand exportsupportservices.z Ourmethodology fordetecting learningeffects isbasedona simple idea. If exporting indeed generatesefficiency gains,then thatbegintoexportshouldthereafter exhibitachange inthe firms stochasticprocessthatgovernstheirproductivitygrowth. Hencetheirproductivitytrajectories must improve insome senseafterthey enter foreign markets. Similarly,ifthepresence of exportersgenerates positiveexternalities, non-exporters intheaffected industryorregion shouldexhibitchanges intheircost processwhen the numberof exporterschanges. Increases inthe numberof exporters may also make it easierforothersto break intoforeign markets. To keeptrack ofcausal linkageswe beginbyspeci~ing thegeneraloptimization problemthatwe envisionfirmsas solving(section II). Each manager facesstochasticcostand foreign demand processes, andchooseswhich periodsto participate inforeignmarkets. Their decisionsare complicated bythe presenceofsunkstati-up costswhen they first sellabroad, sincemanagers must research foreign demand andcompetition,establishmarketing channels, andadjusttheirproductcharacteristics andpackagingto meet foreigntastes. The basicfeaturesofthis modelaretaken directly fromthe hysteresis literature developedbyBaldwin (1988),Dixit(1989), andKrugman (1989). Ourtwist isto addthepossibility of learning-by-doingor,more precisely, learning-by-exporting,and examinehowthis affectsthe productivitytrajectories ofexporters and plantsthat switch markets, relative tothoseofnon-exporters. Becausetheframework we developdoesnot lenditselfto closed-form solutions,we discuss its implicationsheuristically andusing simulations. Under certain assumptions onthe exogenous shocksto productivityanddemand,the lattersuggestthat: (a)non-exporters experiencing positive productivity zManybelievethatthesespillovereffectsaresignificantindevelopingcountries.Forexample,Aitken, HansonandHarrison(forthcoming)writethat” ..the developmentofgarmetexportersinBangladesh,suggests thatinformationaelxternalitiesarelikelytobeextremelyimportant.TheentryofoneKoreangarmentexporterin Bangladeshledtotheestablishmenotfhundredsofexportingenterprises,allownedbylocalentreprenuers... Spilloversmaytakeavarietyofforms.Thegeographicconcentrationofexportersmaymakeitfeasibletoconstruct specializedtransportationinfrastructures,uchasstoragefacilitiesorraillines,ormayimproveaccesstoinformation aboutwhichgoodsarepopularamongforeignconsumers.” 2

shocksself-select intoforeign markets, (b)exportersexperiencingnegative productivityshocksquit foreign markets, and(c) thepresence of learning-by-exportingeffects impliesthatfirms improvetheir relative productivityafterthey beginexporting. Withtheseresults inhand,we examinethe actualperformance ofColombian,Mexican and Moroccan producers(section III). To familiarize ourselveswiththe data,we beginbycomparing the productivitytrajectories ofproducersthatenter exportmarketswith those of non-exporters,ongoing exporters, and firmsthatexit foreignmarkets. (Productivityisproxied by averagevariable costandby laborproductivity.)Thisexercise reveals patternsthatoursimulations suggestwe shouldfind inthe absenceof learning-by-exporting effects. That is,the plantsthat become exporterstypically havehigh productivity beforethey enter foreign markets, andtheirrelativeefficiency doesnotsystematically increaseafter foreign salesare initiated. In some instancesthe relatively strong performance ofexporters traces to high laborproductivity; inother instances itisdueto relatively lowcostsof intermediategoods. Thisfirst lookatthe actualtrajectories castsdoubtonthe importance of learning-by-exporting. But itdoesnot constituteaformaltestofwhether becominganexporter changes afirm’sproductivity trajectory. Accordingly, insectionIVwe estimate econometrically areduced-form version ofthe theoretical modelwhich takesexplicitaccount ofthetwo alternative, butnot incompatible,explanations forthe positive association between export-participation statusand productivity: self-selection ofthe relatively more efficientplants,and learning-by-exporting. Forthe countries with sufficient datato support inference--Colombia andMorocco--we findthatexportmarket participation generally depends uponpast participation and(weakly) upon pastaverage variablecost(AVC),as impliedbythemodel. However, conditioningon capital stockand pastAVCrealizations,currentAVCdoesnotdepend negatively uponpreviousexportmarket participation, as impliedbythe learning-by-exporting hypothesis. Thus,the conclusion suggestedbyour descriptive analysisisborne out by formal Granger causality

tests.sFinally,extending ourmodelto lookforexternalities, we findsome support forthehypothesisthat afirm ismore likelytoexport ifitbelongsto an export-intensive industryorregion, butwe findlittle evidenceofany associated productivitygains. II. AModel ofExport Participation with Learning Effects Our firsttask isto present amodelthat specifiesendogenousandexogenous sourcesofvariation inthe two producercharacteristics weare interested in: exportingstatusand production costs. This model,which will guideus inourempiricalanalysis, isa simplemodification ofexisting modelsto accommodatpeotential learning effects fromexport particiaption (Baldwin, 1989;Dixit, 1989;Krugman, 1989). Webegin byassuming monopolisticcompetition, sothateach firm faces adownward sloping demand curve inthe foreign market,yetviews itselfastoo smallto strategically influencethe behaviorof otherproducers. Specifically, we write foreign demand qf forthe firm’s product atpricep~ as ~f =zf (P) f -~ 9where the randomvariable z~captures theusualdemand shifiers (foreign income level,exchange rates, and other goods’prices) and of> 1.4Firmsface similar demand conditionsinthe domestic market, q ~ = z h(P)h ‘d , andcan price discriminate between foreign and domesticbuyers. Assuming constantmarginal costs,c,the current periodgrossoperating profits canbeexpressed asafunctionofmarginal costsand demand conditions inbothmarkets: (1) \vherez=(Zf 9z h). The profits fromexporting arethe shadedarea depicted infigure 1below,where d ‘d’(d andPh+Ph‘(99are foreign andhome market marginal revenue, respectively. Werepresentthe s TheseconclusionsareconsistentwithBernardandJensen’s(1995)descriptivefindingsusingU.S.data. 4Thisparticularfictional formforthedemandfunctionisgeneratedbytheDixit-Stiglitzutilityfinction overvarieties. 4

homedemand curveasapproaching thevertical axisabovethe foreigndemand curvebecausetransport costsandtrade barrierseat upafraction ofeach unitofrevenuegenerated inforeign markets. Figure 1:GrossProfits fromExporting Letthe per-period, fixed costsofbeing anexporter (e.g., dealingwith customsand intermediaries) beM. Then,theplant willearn positivenetoperatingprofits from exportingwhenever n~(c,zq > M . Accordingly, ifthere were nostart-up costsassociated with becoming an exporter and no learningeffects, producerswould simply participate inforeign markets--choosing theprofit-maximing levelofexports--whenever this conditionwas satisfied. Asfigure 1demonstrates, givendemand conditions, allfirmswith marginal costsbelow somethresholdwould self-select intoexportactivities. ButasBaldwin (1989), Dixit(1989), andKrugman (1989)haverecently stressed,sunkstrart-up costsmodi~ theproblem inanon-trivialway. Supposethatan entrycost ofF dollars is incurredevery timetheplantdecidesto (re)start exporting. Then,onceexporting, itmay beoptimalto keep exporting even if nf(c,z~ iscurrently lessthan Msince, byremaining inthe exportmarket, theplantavoidsfuture re-entry costs. So,ifsunk costsare important--and micro evidence suggeststhatthey are (e.g., Roberts 5

and Tybout, 1995)--producersface adynamicoptimizationproblemwhere, ineach period,they must choose whether ornottoexportonthe basisofcurrentlyavailable information. Thismakesdecisionmaking forward-looking andopensthepossibilitythat exporttoday inanticipationofcost firms reductions later. Hence, even ifthere are no learning-by-exportingeffects, sunk start-upcostsmay imply thatreductions inthemarginalcostofproductionfollowentry intoforeign markets. Because expectationsare important inthiscontext,we mustbespecific abouttheprocessesthat generatethe statevariablescand To accountforheterogeneity inbehavior we allow idiosyncratic, Z. serially correlated shocksto demand and cost foreach firm. We,therefore, assumethatthedemand shifterz isexogenoustotheplant,and follows someserially-correlated, plant-specific process(for simplicity,the plant-subscript isomitted until sectionIV): z, =fix,, z::;) (2) where X, is avectorofexogenousvariablesthat shifithedemand processes, e.g.,theexchangerateand plant-specific noise,and z!c~= L,.l, Z,-2,Z,-3,. . } denotesthe vectorof previousrealizationsonz,,upto and includingperiod [-1. Marginal costalsodependson itsown historyand, inaddition, ispotentially affected bythe firm’sexporting decisionsifthere are learningeffects: (3) lvhere w isavector ofexogenouscostshifiers,e.g., factorprices andplant-specific noise,and t ~{-) t = {Y,>y,-,, Y,-*>. . }denotesthe history ofthe binaryvariable Ywhich, inturn, indicateswhetherthe plantwas exportingj periodsago flj = 1)orwas not, (Y,j=O). Note that we model leaming-byexporting inavery specificmanner: participationinthe foreignmarket lowers marginalcostsof production, irrespectiveofthevolume ofexports: Iffirms learnfrom other producersthatexport, 5Thisassumptionsimplifiestheempiricalanalysissignificantlybecauseoftheendogeneityofthelevelof exports.Italsoplacesfewerrestrictionsonthedata.Notealsothat,forsimplicity,theper-periodfixedcostsof 6

variablesthat describethis external benefitshouldbe includedinthevector w r“ Atest for learning-by-exporting effects basedonequation(3)mustrecognizethat ~isan endogenousvariable, dependinguponthesame observed andunobserved factorsthataffectthecostand thedemand processes. Thus,to sortoutthe relationship betweenc and Ywe needto developa representation of firms’exportingdecisions. Following thehysteresis literature, weassumethatmanagers takeequations(1)through (3) intoconsideration, and plantheirexportmarket participation patternsto satis~: (4) Here Y!+)={Y,,Y,+,,Y,+2,. . } denotesthe entire futuretrajectory of Yvalues, ~ isanexpectations operatorconditioned onthe setof information available attime [,and 8 isthe one-period discountrate! Domestic profits enterthis expression only becauseexportmarket participation may affectthefuturecost trajectory. We assume that firmsneverwish to liquidate. Equivalently, managers can beviewed aschoosingthecurrent ~ value thatsatisfiesBellman’s equation: [( V, =max nf(ct, Z,f) - M - (1 - Y,-l)F)Y,+ aE,(vf+*IY,)] (5) Y, productionarenotaffectedbytheexportparticipationhistory. ‘Thisformulationimpliesthatproducerswhoexittheexportmarketandre-enterfacethesamestart-up costsasproducerswhoneverexported.Inoureconometricrenderingofthedecisiontoexportwewillallowstartupcoststodependuponpreviousexportingexperience. 7

Thischaracterization ofbehavior impliesthat producersparticipate inexportmarketswhenever ~y(c,>z:) - ~ + 6[E,(V,+,] Y,=l) - EJV,+,] Y,=o)]2 F(1 - Y,-,) (6) That is,incumbentexporters continueexportingwhenever current netoperating profitsfrom exportsplus theexpected discounted futurepayofffromremaining an exporter ispositive, andnon-exporters begin exportingwhenever this sum,netof start-upcosts, ispositive.Expected future pay-offs includethevalue ofavoidingstart-upcostsnextperiodandany positive learningeffects thataccrue from foreign market experience. Without learningeffects, expression(6)hasappeared invarious forms inthe hysteresis literature; itwillproveuseful insectionIV. Ifwe allowformuch generality inthe costprocess (3),and/or forthe dependence ofexport profitsonexpofiing history more than oneperiod ago, itisvery difficultto characterize optimal behavior inthis framework. However, some insightscan begained byassumingthat c followsadiscrete, firstorderMarkov processthatdepends onlyon ~., .7Then learning-by-exporting can berepresented by assigningthistermthe transition matrixPOifthe firm isnot exporting,and some stocastically better matrix, sayP,,ifthe fim isexporting. That is,among exporterstheprobability ofadecrease inc is z oreater,and theprobability ofa increase is less. Theno-learning casesimply assignsPOto allfimlS. Simulationsofcosttrajectories based onthisrelatively restrictive framework and arbitrary parameter valuesarepresented infigure2.(Details areprovided inappendix 1.)Thetrajectories are averages overrepeated simulations forfoursubgroupsoffirms, labelled“non-exporters”(thosethat never export), “exporters”(thosethatalways export),“entrants”(thosethat beginexporting), and “quitters”(thosethat cease exporting). Forallfirmsthat beginorcease exporting,we measure time 7Demandshifierscanbeheldconstantsincetheireffectonprofitsisqualitativelythesameasthatof c. 8

I I relativeto thetransition year (period 0), forexample, period-2 istwo years priorto entryforthe so entrantgroup,and two years priortoexitforthequilter group. Firmsthatexport may ormay notbe subjectto learningeffects. Figure 2a: Entrants and Quitters 9 8 LearningModel 7 %06 ------------ 0 &5 ------tiin - g . Model NobamingModel m 54 < 3 2 1 -3 -2 -1 0 1 2 3 Year YearOis~sition year . Entrant—s Quitters Figure2b:Exporters and Non-exporte~ 9 8 king Model 7 G06 NoLwing Model 0 g5 m g4 a 3 ~amingh!odel 2 . . . . . . . . . .. .. . . . . . . . . . . . . .. . . . . .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . .. .. . . . . . . . . .. .. . . . . .. . . .. . . . . . . . . . . -- . - . -- . - . -- . - . -- . - . -- . --- . - . -- ....... NoLearningModel 1 -3 -2 -1 0 1 2 3 Year YearOismiddleyear — Non-exporters.... Exporters Severalpatterns merit note.Firstand mostobviously,regardless ofwhether learningeffects are present, costsamong the entrantandtheexporterfirmsare lowerthan costs intheothertwo groups.Thisreflects 9

the self-selection ofefficientfirms intoexportmarkets, and demonstrates why wecannotrelyon static costcomparisons betweenexporters andnon-exporterstomake inferencesabout learning. Second,regardlessofwhether learningeffectsare present, firmsthat becomeexportersexhibit costdeclines beforetheyenterthe market. These firmsself-select intoexportingonlywhentheirunit costs fallbelow somethreshold,sothey mustexperienceaperiod offallingcostspriorto entry.Hence selection effects may createthe illustionthat becominganexporter actuallyrefardsproductivitygrowth. Third, onedistinguishingfeature ofthe learningtrajectories isthat exportingfirmsexhibit ongoing cost reductionsafier initiating foreign sales.Onlywhen learningeffects arepresent,dofirms continueto pullaway fromnon-exportersafter foreignmarket entry. Finally, relativetothenolearningcase, learningmakes firmsenter (and stay in)exportmarkets athigher costs.Thisoccursbecause the incentivesto exportare largerwhen learningoccurs.Productivity dispersion may thus behigheramong exporters when learningeffects arepresent, andtheproductivity gap between exportersandnon-exporters may besmaller.s Of course,theseresultsare onlysuggestive,andmore complicated costprocessesmightreverse someofthe patterns. Forexample, ifcostswere to followasecond- orhigher-order autogressiveprocess, they might continuetotrend downward after exportmarketentry even without learningeffects. Accordingly, to lookforlearningeffects and externalities,we rely on econometric estimatesofageneral form ofthe costfunction(3),recognizing that exportmarket participation isgovernedbythebehavioral rule(6). III. Learning Effects:ALook atActual Data Before we reportthe resultsofthat exercise, however, it isinstructivetovisuallyexamine cost data on actual producersfromthree semi-industrialized countries. Though merely suggestive,this will 8Moregenerally,thisresultsuggestscautioninterpretingstudiesthatuseproductivitydispersionasa performancemeasure.(Seeespeciallythe“efficiencyfrontier”literature).Whenlearningeffectsarepresent, exportersmaylookrelativelybad. 10

familiarize thereader with the basicpatternswe aretryingtoexplain,andprovide an informalcheck for the distinguishingpatternsthatoursimulations suggestwe shouldfindwhen learning effects arepresent. A. TheData Ourdata allowusto follow individualproducersthroughtime inColombia, Mexico and Morocco. Inthe caseofColombia, theydescribe virtually allplantswith at least 10workers overthe period 1981-1991;inMexico,they describe 2800ofthe largerfirmsoverthe period 1986-1990,and in Morocco they cover mostall firmswith at least 10workers overtheperiod 1984-1991? Standard informationon inputs,outputsandcosts isprovided ineach ofthese databases, aswell as informationon export levels.To simplifyestimation wethrow out firmsthatdo notreport information fortheentire sampleperiod,creating abalanced panel. The datawere cleaned anddeflated asdescribed inAppendix II. Finally,to sharpenthe analysis,we focusonly ontheexport-oriented industriesineach country. These industriesexported at least 10percent oftheiroutput,andhadat least20 exportingplants!” Table 1providesdescriptive information oneach country. Note thatalthoughmost plantstendto stay inthe exportmarketor stayout,there are substantialtransitions inthedata. 9Althoughsomecountriesreportplant-leveldataandsomereportfirm-leveldata,wewillhereafierusethe term“plant”todescribetheunitofobservation.Insemi-industrializecdountrieswherethecalculationispossible, wehavefoundthat95percentoftheplantsareownedbysingle-plantfirms. IOInafewcases,indus~iesthatexpofledlessthan10percentoftheiroutputwereincludedbecausetheY handmanyexportingplantsandor accountedforasubstantialshareoftotalmanufacturedexports. 11

Table 1: Entry, Exit, Number ofPlants and Export Intensity byCountry (Export-oriented Industries) Ave.Annual Ave.Annual Ave.Numberof AverageExport EntryRate ExitRate Plants Intensity Colombia .027 .017 1,354 .095 1981-1991 Morocco .049 .037 938 .360 1984-1991 Mexico, .048 .015 1,327 .230 1986-1990 12

B. Comparing Productivity Trajectories Wewishto familiarize ourselveswith themarginalcosttrajectories ofplantswith different export market participationpatterns, controllingfor industry-widetime effects,andobservable plantspecific productivity determinants likecapitalstocksand age. We usetwo marginalcostproxies: average variablecost(AV~ and laborproductivity(LAB). The former isdefined asthesumof laborand intermediate inputcostsdivided byreal output,andthe latterisreal outputdividedbynumber ofworkers. Real output isthesumof nominaloutputforthedomesticmarketand nominaloutputforexport, each deflated by itsown product-specific deflator. To purgetheseproductivity measures of industry-widetime effects and observableplant-specific characteristics, each isexpressed inlogarithmsand regressed ontime dummies(specificto year zand the j’h3-digitISIC industry),age oftheplant(A), ageofthe plantsquared, capital stockoftheplant(K), and capital stockofthe plantsquared. Both ageandcapital stocksare measured inlogarithms: JT ln(~~C,,)= ~ ~ y,~j, + ~lln(AiJ+ ~2ln(AiJ2+ ~31n(K,J+ ~41n(K,J2+ Ei, j=l f=l Theresidualsfromthese two regressionsarethen usedas our indicesofdeviationfromtime- and industry-specificproductivitynorms. Note thatthere isnoneedto deflatevariable costsdirectly because industry-specifictime dummies playthe role ofprice deflators(inter alia). Also, because logarithmic variable costs arepurgedof correlation with capital,theresiduals can beviewed asthemeasure of variable factorproductivitythat obtains from atotalcostfunction ofthe form Ci,= ~o(Ki,, WiJ +Al(Ki,, WiJQi, ,where parentheses denotefunctions,Ki,is capital,and W&isthe vector offactorprices. To isolatethe relation between exportmarket participation and productivityperformance, we 13

distinguishthe fivevarieties of plantsdefined inTable 2,and foreach plantwere-define periodzeroto betheyear inwhich achange inexportstatustakes place. Thenwe isolatefive-year blocksoftime, runningfromtwo yearspriortothe statuschange(/ = -2)totwoyearsafter(t =2)~1 Fornon-exporters andexportersthere isnochange instatus,sowetake fiveyears inthe middleofthe sampleperiod(for Colombiawetake sevenyears). Finally, afier re-indexingtime inthis way, weaggregate ourproductivity indicesbyplanttypeand compare them. Because switchingfirms stronglyresemble exporterswe omit them fromourgraphsto reduce clutter. Table2:Firm Varieties Non-exporters: firmsthatneverexportedduringthesampleperiod. Exporters: firms thatalwaysexportedduringthesampleperiod. Entrants: firms thatbegantheperiodasnon-exportersb,utbeganexportingduring thesampleperiodandneverstopped. Quitters: firms thatbegantheperiodasexporters,butceasedduringthesample periodandneverresumed. Switchers: firms thatswitchedexportingstatusmorethanonceduringthesampleperiod. c. BasicTrajectory Patterns AverageVariableCos~s:Unweighed average trajectories foraverage variable costsarepresented byplanttype inFigures3.1through 3.3. We beginbyconsidering Colombia andMexico, sincethese countries showsimilarpatterns. Most strikingly,plantsthat cease exporting getsteadily worse before they dropout offoreign markets, and are substantially lessefficientthan the otherplanttypes. Also, entering plantsandexporting plants sharethe distinctionofhavingthe lowestvariablecosts, and non- 11TheColombianpanelislongenoughtoallowustogofromthreeyearspriortothreeyearsafierthe switch. 14

exportersconsistentlyexhibitcostsslightlyaboveaverage, but lessthan quitters.These patternsare similarto the simulations infigure 1forboththe learningandnon-learningmodels,except inthatthe performance ofquittingplants issomewhat worse intheactualdata. Figure3.1: COLOMBIA --PathofA\erage Cost (purgedoftime,age,andsizeeffects) I-Nonexporters I + Entrants + Quitters 7 .= [ ~ ~ b -2 -1 0 1 2 t> -0.2; -0.4+ Year(Oistransition/middle }~ear) Figure3.2: MEXICO--PathofAverageVariableCost(purgedoftimeandsize effects) !-o-Exporters i !+ Entrants ~ 0.8 1 ~-M-Quitters I J 0.6 0.4 1 5 -1 -0.2 ~ I nA , Year Onemight argueaboutwhether the entrants showevidenceofreducing their costsafier beginningto export, buttheeffect iscertainly not dramatic. The F-statisticforthe nullhypothesisthatthe mean average cost levelamong entrants showsnovariation relativetothe industrynormshasap-value of 15

.08 inColombia, andap-valueof .25 inMexico.lz Further, inMexico,entrants’ average costsare slightlyhigherrelativeto industrynormsafterthreeyearsofexportingthan theywere threeyears before. Figure3.3: MOROCCO--PathofAverageVtiable Cost(purgedoftime,size andageeffects) I 1+ Quitters J 0.41, Year The Moroccan graph ismore complex. Mostobviously,there ismuch lessvariation in trajectories herethan we found inColombia andMexico. Further, what littlevariationthere isdoesnot conform tothe patternsdescribed above.TWOyearsafier exiting,the plantsthat abandonforeign markets emerge astheworst plants,but their average variablecostsarehighlyvolatile, and nohigherthanthe industrynorm oneyearafier exiting. Exporters usuallydobetterthan non-exporters, buteventhis isnot guaranteed. These patternsmay well reflect the factthat,unlikeinMexico and Colombia, mostofthe impetus to become anexporter inMorocco came from firm-specifcdemandsideshocks.Many Moroccan exporters areyoungplantsthatwere foundedwiththeexclusivepurposeof sellingparticularappareland textile products abroad(Sullivan, 1995;World Bank 1994;Roberts, Sullivan,Tybout, 1995). Moroccan policies duringthe sampleperiod also provided varioussubsidiesto exporters, andthesemay have allowed lessefficient plantsto compete. Once againthere islittlestatistical evidencethataveragecosts among plantsthatbegin exporting are anything butflatrelativeto industrynoms (thep-value is.24). IQThesetestsarebasedongeneralizationsoftheregressionmodeltoincludeannualdummiesbyPlant type.Wealsoallowthecoefficientsonage,agesquared,capitalandcapitalsquaredtovaryacrossindsutry,andwe treatthedisturbanceascomposedofafixedplanteffectplusrandomnoise. 16

The resultsdiscussedthus farare based onunweighed averages ofourvariable costmeasures. Butsmallplantsare much morecommon than largeplants, soFigures3.1through 3.3 donotnecessarily describesector-wide performance. To determine whether aggregateperformances lookedsimilar,we therefore redid thecalculations, weighting each plant’saveragecostby itsshare inthetotaloutputamong thoseof itstype. We alsoredidtheplant-specific average costmeasures themselves, leavingcapital stocksand age outoftheregression, becausewe didnotwantthem to bepurged ofcorrelationwithsize forthisexercise. Theweighted average costtrajectories byplanttype are presented inFigures4.1through4.3. ForColombiathepattern looksvery similar, exceptthatexporters’costsare no longerbelowthoseof non-exporters. Sosmallexportersare apparently moreefficientthan largeexporters interms ofvariable costs. InMexico the pattern isalso similar, exceptexporters’costs nowturn up dramatically inthe last period. This ispresumably becauseofoneor severalvery largeplants. Finally, itremainstruethat entrantsexhibitrelatively lowaverage costs,bothbeforeand afierthetransition year. There isno obvioustendency for coststo fallafier exporting operationsare initiated. Figure 4.1: COLOMBIA--PathofOutput-WeightedAverageCost (purgedoftimeeffectsonly) _ Nonexporters I Q-Exporters + Entrants , 1 + Quitters I -1 -0.4 i -o.fj~ I noI 1 Year 17

Figure4.2: ME.XICO--PathofOutput-WeightedAverageVariableCost(purged oftimeeffectsonly’) I + Nonexpotiers I-c-Exporters + Enttants 0.6- + Quitters 7 .= i 0.4- f/ Year Figure4.3: MOROCCO--PathofOutput-Wei-@tedAverageVariableCost (purgedoftimeeffectsonl>’) I - Nonexporters 0.8 ~ i+ Expotiers 0.6- I 0.4 + I 0.2+ -0.2+ , Year In Morocco,theweighted averagetrajectories areas flatandtightly clustered astheir unweighed counterparts. Aswiththe othercountries,there isnoevidencethat costsdrop afier exportsare initiated. However, theweighted figuresdoprovide better supportfortheassertion that exporterstend to be lower costthan non-exporters. Laborproductivity:An alternative measure ofperformance isoutput perworker. Unlikecosts, it doesnot reflectvaluation effects dueto changes infactorprices, asmight happen ifexporters began 18

i Figure5.1: COLOMBIA--PathofAveragebbor productivi~,(purgedoftime, age.andsizeeffects) Year usingimported inputsunder adutydrawback scheme, or ifrealexchange rate fluctuationsaffected producersto theextentthatthey used importedintermediates. Like averagevariable costs, labor productivity isnotatruemeasure oftotalfactorproductivity,butafier ithasbeen purgedofcorrelation with capital stocksitisconceptually closer. Figure5.1presents unweighed averagevaluesofthis productivitymeasure byplanttype for Colombia, Theentrantsonce again performthebest,and incontrasttotheir average costpatterns (Figures 3.1and4.1)do seem to improvewhen they initiateforeign sales!3 Ongoing exporters continue to showthe nextbestperformance, andquitterscontinueto showtheworst performance, particularly aroundthetimeoftheirexit. So inColombia, itappears thatoutputperworker isone importantsourceof variation inaveragevariable costs, andwe havethefirst bitofevidencethatexporting mightimprove performance. Quittersand ongoing exporters inMexico followthe Colombian patiem. ButMexican entrants donot. Their average laborproductivity isvery stablerelativeto industrynorms @-value= .61),and a bitbelowaverage (Figure 5.2). Sothere isnosuggestionofa learningeffect from exporting inthis country,nor,forthatmatter, does itappearthatcross-plant exportingpatterns can betraced to differences ‘sTheF-statisticforthenullhypothesisthat averagelaborproductivityamongtheseplantsdoesn’tchange relativetoindustrynormshasap-valueof.01. Seealsofootnote11. 19

inlaborproductivity. Figure5.2:MEXICO-PathofA\”eragLeaborProductivit(ypurgedoftimeand sizeeffm) 0.4- 0.3- 0.2- ,- Enttants -0.4- ‘+Quitters 1 n= Yeal MOROC-cO Fi~e 5.3: PathofLaborProductivity(purgedoftime,size,and ageeffects) Y I 0.6~ I I Year Finally, interms of laborproductivity, Moroccan patterns appearto resemble Colombia’s(Figure 5.3). Entrants’productivityjumps intheyear thatthey beginexporting (although thep-value is.41),and 20

ongoinge~po~ers exhibithigher laborproductivity than either non-exportersorquitters!4 Interestingly, there ismuch morevariation acrossplanttypes inourMoroccan laborproductivity seriesthanwe found inouraverage costseries,suggestingthat laborcostsco-vary negativelywith intermediategoodscostsin this country. Weightingour laborproductivityseriesbyplant sizegivesabetter picture ofaggregate performance for each oftheplanttypes. We reportthe resultsofthisexercise inFigures6.1through6.3, usingseriesthathavenotbeen purgedofcorrelation with capitalstocksorage. In Colombia,thepattern forentrantsandquittersseems largelyunchanged from Figure5.1,althoughthere isnolongerany recovery in laborproductivityamong the latter intheyears afierthey exit. Also, ongoingexportersno longerconsistentlyoutperformnon-exporters, suggestingthatthis wasmainly a smallplantphenomenon. Thesearethe samecontrastswe identifiedwith ourweighted versus unweighed averagecostseries. Figure6.1: COLOMBIA--PathofOutput-\VeightedLaborProductivity(purged oftimeeffectsonly) -3 --7 -1 0 1 2 3 Year laTheoneexceptionisthat]aborproductivi~ ishighforexitingp]antsduringtheirlastyearinforegin markets. 21

Figure6.3: hfOROCCO--PathofOutput-WeightedLaborProductivity(purgedof timeeffectsonly) 4<. 1 I 4L 3.5 T 3+ 2.5; 1 * Entrants 0.5- I + Quitters n 1 -2 -1 0 I 2 Year Figure6.2: MEXICO--PathofOutput-Weighted LaborProductivity(purgedof timeeffectsonly) -2 -1 0 1 ~ Year The Mexican trajectories donotseem to dependmuch uponwhether we useweighted or unweighed laborproductivity,althoughquitters’performance ismore stablewiththeformer (Figure 6.2). Onthe otherhand, inMorocco ouruseofaweighted series(Figure 6.3)undoesmuchofwhat emerged in theunweighed series. (Figure 5.3). Exporters become theleastproductive whenweweight by size,and thetendency forentrantsto improveafterthey beginexporting isdampened. Theseresults, inconduction with the lowpower oftest statisticsforourunweighed trajectories, suggestthattheunweighed series reflect the influenceof smallplantsthatare outliers. 22

LaborQuaZi&:One fundamentalsourceofvariation inlaborproductivity isassociatedvariation inthe skillmix ofemployees. TOseewhether differences inskillmixesare behindthe laborproductivity differences we analyzed the ratio ofskilledto total labor inthesameway we analyzed ourtwo productivityseries.15The results forColombiaare presented inFigure7.1. Interestingly,entrantsappear tohavehigh laborproductivitypartly becausethey use skilled laborrelatively intensively,and quitters have lowproductivity becausethey use littleskilled labor. There isalsosomeevidencethatrelative labor quality improvesovertime fortheentrants(thep-value is.09). Thehigh laborproductivityofongoing exportersdoesnot,however, appeartocome from unusually highskillintensity. None ofthesepatterns issensitivetowhether we useweighted or unweighed skill intensityseriessowereportonlythe latter. One interpretation isthat breaking intoforeign markets requiresnewproductdesignandotherformsof technical assistance, butestablished exportproduction can beroutinized. Figure7,l: COLOMBIA -- PathofLaborQuality (purgedoftime,age,andsizeeffects) 0.06 T1 + Nonexponers + Exporters -0.04- > I n 1 J Year In Mexico, laborqualitytrajectories match laborproductivitytrajectories forongoingexporting plants(which are skill-intensive) and exitingplants(which are not). However, althoughMexican entrants resemble Colombian entrants interms ofskill intensity,this isnotsufficientto getthem high labor productivity(Figure 7.2versus Figure 5.2). The Mexican resultsare somewhatsensitivetoweighting; when this isdonenon-exporters actuallyexceed exporters interms ofskillintensity. Nonetheless, ‘sNoresultsarepresentedforMoroccobecauseoflackofappropriatedata. 23

entrantsremain aboveaverage, and quittersarewell belowaverage. Fi~e 7.2:MEXICO-PathofLaborQuality(purgedoftimeandsizeeffeets) O.M- I -0.16+ I n.9- Year In summary,ourperformance measures indicatethatentrants generally dobetterthannonexporters andexitingplants. They have higherlaborproductivity,andthis appearsto bepartly dueto their heavierrelianceonskilled labor. Also,despitetheir highqualityworkers, newexportershave relatively lowaveragevariable costs. On theotherhand,we find littleto suggestthatproductivitygains follow entry intoforeignmarkets. Laborproductivityand skill-intensity appearto improvefor Colombian plantsthat beginexporting, butthis phenomenon didnot showup intheothercountries. Interestingly, Bernard andJensen (1995)haverecentlyreported that many ofthese patternsare evident among U.S. manufacturers aswell. Iv. An Econometric Test ofLearning Effects The figureswe report inthe previoussectionare revealing, butthey do notconsituteadirecttest forwhether pastexportmarket participation influencescurrent costs;the F-statisticswe calculated merely indicatewhether costtrajectories forentering firmsdeviate significantlyfrom industrynorms.Hence in this sectionwe developan alternative approach totest forthe presence of learning-by-exporting effects in the data. Specifically,weestimate afairly generalversion ofequations(3)jointly with areduced form of 24

(6), andtestwhether exportinghistory, Y~.) ,enterssignificantly inthecostequation.lGIn other words,weperfo~ Granger causalitytests inanon-linearcontext. A. TheEconometric Model Ourtreatment oftheexportmarket participation decision followsRobertsandTybout(1995). First,we generalizeequation(6) sothat firmswhichexit andre-enter theexportmarket paydifferent start-up coststhan firmsthatneverexported. Specifically,defneFOasthe start-up costforanonexporterwith nopreviousexperience, andFJasthestart-up costfaced byafirmthat lastexportedj-1 years ago(notethat F1=O). Then equation (6)generalizesto implythatthe ~~firm willexport inyeart (i.e., willchoose ~ = 1)whenever ~J(ciz,,;)- ~ +a[Ef(v,I,+Y,i,=l)- E,(vi,+lIY,[=O)] F“ - ~ (F”-F~)Pif-j , (7) 2 j=l where ~i,-j isone if thefirm lastexported in yeart-j andzero otherwise!’Next, we define the ]atent variable ~ ascurrent netoperatingprofits plustheexpected futurereturnto being an exporter inperiod~: E,(Vi,+l Yi;=n~(c z i lj r, - ~ + 6 Iyif=1) - E[(vif+I,Y,,=O)] [ Equation(7)then impliesthat ~,obeysthefollowing dynamic discrete process: 16Ifweallowformuchgeneralityintheprocesses(2)and(3),itbecomesverydifficulttoestimateallthe structuralparametersofthemodel.Yetimposingrestrictionsontheseprocessesmaywellleadustoincorrectly concludethatpastparticipationintheexportmarketinfluencescurrentmarginalcosts. Forexample,misspeci~ing equations(2)and(3)tobefirst-orderprocessesmayforceanydependenceofcurrentcostsonadditionallags(e.g., c,.,,c,-,)tocomethrough~-,,~oivingthefipressionthatexportmarketparticipationaffectscosts.Thus, ifthegoal istolookforlearningeffects,itisdesirabletopursueareduced-formapproach. 17Notethat ~,r-l= Y,,_l and,for~>2, yi,-j= Yi,-j’~(1-Yi,_~). k=l 25

1 if Y,; z F“ - ~ (F”-Fj) P,,-j j=l ~,= (8) { O otherwise Finally,\veexpressthe latentvariable ~, -F asa reduced form in demand shifters, marginal costs, as well as thevariables thathelp predict future marginal costsand demand in each market.18 Operationally, this means includingexogenous plantcharacteristics (xti), timedummies (D),a distributed lag in our marginal costproxy, AVCi,, and a serially correlated disturbance. Writing startup costsat the i’~plant as their mean value plus another disturbance, we have: (9a) F,: . FJ+ (F (9b) it Then substituting(9a) and (9b) into (8)we obtain a representation ofexport market participation decisionsthat can be estimated: (lo) o otherwise where ~i, is a serially correlated compound disturbance based on theerrors inequation (9). Notice that the estimatedcoefficients on our laggedparticipation variables measure the discount on entry coststhat plantswith exporting experience inprevious years enjoy, relative to plantswith no exporting experience. For example, a plant that most recently exportedj-1 years ago pays ody FO - Fj to resume exporting Operations, while aplant that never exported pays start-up costs ‘sThisreducedformapproachisalsousedinSullivan(1995)andRoberts,SullivanandTybout(1995). 26

FO . Further discussion may be found in Roberts and Tybout (1995). As noted in section II, iffirms learn by exporting thestochasticprocess thatgenerates costs also dependsupon the history of ~.,values. Accordingly, z general log-linearized specificationof the marginal cost finction (3) includesnot only capital stocks, aggedcost, and factor prices, but lagged K., valuesaswell: T J J ln(AYCiJ= y. + ykki, +~ y~Di~+~ y~ln(AVCi,-~+) ~ ~Yit-~+ vi, (11) ~=1 j=; j=] Heretime dummies control forchangesinfactor pricesthatarecommon to all plants. Together, equations (1O)and(11)describe exportmarketparticipation patternsandmarginal costsrealizationsforthecase ofno learningexternalities. Estimated asasystem,they shouldreveal whether marginalcosts influencetheexport participation decision,asthemodel implies,andwhether firmstypicallyexperience COSrIeductionsoncethey have begunto service foreign markets, aspositedby the learning-by-exporting hypothesis. Tests onthe costcoefficients inthe participation equation indicate whether firmsrespond to costreductionsby becoming more likelyto export, andtests onthe ~.,j coefficients inthecost function indicatewhether exporting experience leadsto lowercosts. It is,of course,the latterdirection ofcausalitythat isofprimary interestto us. Serialcorrelation is likelyinbothequations becausepersistentunobserved plantcharacteristics makesomefirms consistently lowcostand/or consistently proneto exporting, conditioningonobservable variables. Hence, we modeleach disturbanceascomposed ofan unobserved (random) planteffect,a, and az, plustransitory noise: qi, = ali + cl,, and v 1 , [ = a2i +~2i, . We allowthe planteffects andthe transitory noiseto becorrelated acrossequations. Also, without lossofgenerality, we imposethe normalization var(qi) = I sothatallcoefficients inequation(10)aremeasured relative tototal unexplainedvariation. Unfortunately, the combinationof laggeddependentvariableswith serial correlation creates 27

specialproblems inpaneldata. These laggedvariablesarethemselves functionsoftheunobserved plant effects, sothey arecorrelated with thedisturbance. Butthis dependence isnotcaptured by equation(10) norbyequation(11)forthe firstJyearsofdata. WeadoptHeckman’s (198la, 1981b)solutiontothis “initialconditionsproblem” byaddingextraequationsto thesystemthat represent ~,through Yuas finctions of ~li(interalia),andAV~.lthroughAVCUas functionsof ~zi(inter alia). Theresult isavariant ofKeane, MoffittandRunkle’s(1988)and Sullivan’s(1995) estimator.lgThe system isestimated using maximum likelihood,integratingoutthetwo randomeffects with Gaussian quadrature. Details ofthe likelihoodfunctionmay befound inAppendix III. B. The E}~idenceonLearning Becausewe expectthe cost function,theprofit function,entry costs,and thepotentialforactive learningto differacross industries,we fitoursystemof equationsseparatelyto each industryinwhich we havesufficientobservationsto support inference. We includeindustrieswhich arerelatively intensivein human capital(e.g., Chemicals) aswell asthosethat are not(e.g., Apparel) inorderto lookfor correspondingvariation inactive learningeffects. Although themodelcan befitto ourMexican panel, thetime period spannedbythatdata setprovedto shortto isolaterandom effects fromthehistory ofcost andexport marketpartication.20we therefore focusour attentionontheColombian and Moroccan results reported inTables3.1 and 3.2. The participationequation yieldsresultssimilartothosereported elsewhere (Roberts andTybout, 1984;Roberts, Sullivanand Tybout, 1995). In allcountriesand inall industries,plantswith largecapital 19Sullivan’s(1995)estimatordoesnotdealwithinitialconditionsproblemsintheequationthathasa continuousdependentvariable;andKeane,MoffitandRunkle’s(1988)estimatordoesnotdealwithdynamicsatall. Otherwisethestructureofourestimatoristhesame. 201nMexicoweobserveonlyfiveyearsofdata,andthreeoftheseyearsarelosttolagsonparticipationand averagecosts,leavingtwoin-sampleyears.Estimatesofthemodel(availableuponrequest)attributetoomuch explanatorpyowertolaggedcostandexportparticipationa,ndimplythat var(al) =var(a2)=O . Nonetheless, theresultsconcerningtheeffectoflaggedparticipationoncostrealizationsareconsistentwiththoseobtainedinthe othercountries.TheconvergenceofrandomeffectProbitestimatorstosimpleProbitestimatorswhenTissmallis discussedinGuilkeyandMurphy(1993). 28

stocksaremore likelyto beexporters. One likelyinterpretation isthatthere are fixedcostsassociated with exportshipments,and producers whocan produces largebatchesare betterableto spreadthese costs. Consistentwith ourconceptual framework, plantsthathave lowermarginal costsare more likely to beexporters,otherthings being equal. Butindividuallagsofthisvariable arenever statistically significant,perhapsbecauseofthehighcollinearitybetween them, and,eventhough insome instances the estimated coefficients alternate insign,theirsum isalways negative?i The factthat standarderrors forcostcoefficients arerelatively largeforMorocco isconsistentwiththe lackofcostvariation across typesoffirms observed inthe graphs discussedearlier insectionIII. The effect ofexportexperience ismostdramatic forplantsthatexported lastperiod,and hence face nore-entry costs. Inall industriestheeffect issubstantial. InMorocco,the effect issmaller in textilesthan inchemicals. This resultjibes withourpriorsthat breaking intothe foreigntextilesmarket involveslesssunkcostthan breaking intothechemical markets. Althoughexporting experience acquired morethan oneyear agoproved marginally significantinearlierwork onparticipation (Roberts and Tybout, 1994),itappearsto be unimportantformost industriesinthepresentapplication. Thisfinding isprobablyduetotherelatively smallsampleswe usehere, sincethecoefficients on laggedparticipation variablesthemselves arenot systematically smaller inMorocco. ‘lOnereasontheseresultsarelessdramaticthanourgraphsisthathereweconditiononlagged participation,whichiscolinearwithAVC. Putdifferently,firmswithlowunitcoststendedtobeexportersinthe past,sorecentinnovationsinAVChavelimitedexplanatorypower. 29

Table3.1: Estimation Results ofEquations(10)and(11), Colombia 1983-91 II II Chemicals Textiles Apparel II II Participation Equation intercept -15.06 (5.87)* -4.34 (1.33)* -5.20(1.58)* Zn(exchangerate,) .462(2.27) 3.93(1.14)* 3.60(0.94)* Zn(capitalstoc~,.,) 3.73(0.72)* 1.71(0.34)* 2.03(0.37)* Zn(agei,) 5.94(3.87) -0.12(0.85) 0.64(1.07) l?’l(agei)2 -0.88 (0.58) 0.09(0.14) -0.01(0.18) business~pei, -0.00 (0.28) 0.23(0.15) 0.28(0.14)* ln(AVCi\_J -0.14 (0.28) -0.18(0.18) -0.05(0.11) In(AVCijl- 0.06(0.29) -0.16 (0.18) -0.10 (0.13) Yir., 1.86(0.41)* 2.04(0.27)* 1.03(0.25)* Yir-2 0.18 (0.40) 0.25(0.29) -0.10 (0.25) Yi,-3 0.26 (0.40) -0.57(0.46) -0.20(0.25) II CostFunction II intercept -0.16 (0.13) -0.33 (0.06)* -0.36(0.06)* Zn(capitalstock,) -0.17 (0.11) 0.05(0.04) 0.09(0.06) In(AVCi,-J 0.36 (0.05)* 0.65(0.02)* 0.74(0.02)* [n(AVCi~_J 0.45(0.05)* 0.16(0.03)* 0.12(0.03)* Yit_, 0.09(0.05) 0.01(0.03) 0.08(0.03)* Yif-2 0.22 (0.09)* 0.04(0.06) 0.01 (0.08) Yi,-3 0.11 (0.10) 0.06(0.07) 0.01(0.08) II Variance-Covariance Matrix forDisturbances II var(a’) 0.298 0.167 0.574 var(az) 0.005 0.0001 0.001 corr(a1,a2) -0.106 -0.147 -0.082 var(E’) 0.702 0.823 0.426 var(C2) 0.995 0.106 0.163 corr(E1,62) 0.180 -0.110 -0.026 ~~Noo.bservations 567 1,854 2,547 II ~Log-likelihood -309.61 -938.45 -1,679.24 I 30

Table3.2: Estimation Results ofEquations (10)and (11),Morocco 1984-90 Chemicals Food Textiles ParticipationEquation intercept 17.66(18.18) 16.35(15.04) 16.49(10.05) Zn(exchangerate,) 5.69(3.90) 3.58(3.22) 3.50(2.13) /n(capital stoc~,.l) 2.64(0.67)* 1.03(0.46)* 2.45 (0.43)* /n(ageit) 2.45 (3.23) -1.56 (1.73) -2.09 (1.47) ln(agt?iJ2 -0.39 (0.48) 0.28 (0.26) 0.35(0.24) business~pei, 0.81(0.70) 0.26 (0.15) 0.07(0.16) ln(AVCi,.J -1.16 (0.60) -0.18 (0.42) 0.06 (0.18) ln(AVCijl- -1.05(0.89) 0.92(0.47) -0.24 (0.30) Fi,_, 1.14(0.45)* 1.25(0.54)* 0.91 (0.36)* 0.27(0.36) 1.25(0.51)* 0.50 (0.28) 0.28(0.35) 0.67(0.47) 0.02 (0.28) CostFunction intercept 0.05(0.03) -0.07 (0.03)* -0.12 (0.03)* Zn(capitalstock,) -0.06 (0.04) -0.16 ((0.04)* -0.07 (0.04) ln(AVCil_J 0.39(0.05)* 0.20(0.05)* 0.15(0.04)* ln(AVCi,.J 0.40(0.07)* 0.25(0.05)* 0.07(0.05) Yi,_l -0.02(0.03) 0.12(0.02)* -0.02(0.02) 7i1_2 0.03(0.05) 0.02(0.04) 0.00(0.05) ?i,-3 0.02(0.04) 0.01(0.10) -0.12 (0.06) Variance-Covariance Matrix forDisturbances var(al) 0.546 0.724 0.677 var(a2) 0.0001 0.003 0.001 corr(a’,a2) 0.046 -0.559 -0.590 var(61) 0.454 0.276 0.323 var(E2) 0.024 0.022 0.057 c0rr(6’,E2) -0.019 -0.040 -0.057 No. observations 637 1,169 1,722 Lo~-[ikelihood 69.13 117.79 -517.87 31

Intheory,devaluations should increasetheprobability ofbecomingan exporter, butweonlyfind significanteffectsoftheexchange rate inColombiantextiles and apparel industries. These arethegoods thatColombiasendsnorth,so it isnotsurpisingthatthe realexchange ratevis-a-visthedollar isa strong predictorforboth industries. Chemicals,ontheother hand,are mainlysold inLatin America and producers inthatindustrydonot showasmuchresponsivenessto real exchangerates. Pointestimatesof the responsetodevaluation inMorocco resemblethose forColombia, buthave much larger standard errors. Now considerthecostequation,which isthefocusofour analysis?2 Asexpected, plantswith largercapitalstockstend to have lowermarginalcosts, althoughthere aresome insignificantcoefficients. Also, conditioningon capital stocksand unobservedplanteffects, marginal costsappearto followa 23But critically, expofiinghistorycontributes 1ittle‘0 second-order,orhigher, autorregressive process. > theexplanationofmarginal costsoncewe haveconditionedonthese variables. Indeed, inthefew instanceswhere laggedexperience isstatisticallysignificant,the coefficient suggeststhat exporting itfcreasescosts. Mightthiscost increasebeastatisticalartifact? One possibility isthatwe systematically overestimatethe priceofexported products, sothatwhen producers startexportingthereal valueoftheir output issystematically under-estimated, causingaverage coststo beoverstatedand laborproductivityto 220neestimationdetailworthnotingisthatthevarianceoftherandomeffectinourcostequationisalways veryclosetozero. Thisisapparentlybecausethelaggedcostvariablespickupalloftherelevantheterogeneity. Oneimplicationisthatwereallydidn’tneedtoestimatethecostfinctionsjointlywiththeparticipationequation: thereisnoreasontothinkthatlaggedparticipationiscorrelatedwiththedisturbance. 23Indeed~ Prelfiinq work,ath~dlagofthecostvariablewassometimesStatisticallysignificant. Becausetheresul~andconclusionsarenotsignificantlychangedbytheinclusionofathirdlag,ofcosts,and becauseoftheincreasedcomplexityoftheestimation,wedecidedtoworkwiththemoreparsirnonousspecification. Thesameaplliestotheuseofyeardummiesinsteadoftheexchangerate. 32

beunderstated.2JAnother sourceof biasderives fromthefactthatwe arenotmeasuringtotalcosts. If theproduction technologyforexports ismore labor-intensive(or skilled-labor intensive) thanthe technology for domesticgoodsproduced bythe samefirm,we mightmissoffsetting reductionsincapital costs byfocusing exclusivelyon laborandmaterials. Third,workers may capture efficiencygainsas higherwages, leavingaverage variable costs unaffected. Finally,we may well bepickingup someofthe sunkentry costsassociatedwith becoming an exporter inourvariable costmeasure. ButinColombiathe coefficients on allthree lagsofexportmarket participation arepositive, soforat leastthreeyears, these entry costsare notoffset byproductivity gainsdueto learning-by-exporting effects. In sum, itishardto reconcile ourresultswiththe presenceof strong learning-by-exporting effects. Ifthey arenotdriven by measurement problems,theseestimates suggestthatnegativecost shockscan Granger cause increases intheexpected return fromexporting. Buttheactofexportingdoes not Granger causereductions inmarginal costs. c. The Evidence onExternalities In order totest forregionaland industryspillovereffects, were-specified equations(10)and(11) to includeregionaland industryexport intensityvariables. Export intensitywas measuredasthe percentage of firms inthe industry/regionthatwere exporting inthe particular year. Thesevariableswere added tothe participation equationto testwhethersunk entry costsdepend upon exportingactivi~ by other firms. Theywere also addedto thecost functionto testwhether all firms enjoycostreductions when somefirms export.To control forpermanent, unobservedregional effects likeaccessto ports, regional dummies were also included. ‘4Givenourmethodologyforconstructionofpriceindices,thiscouldhappenifweover-estimatedthe changeintherealexchangeratebetweenthebaseyearandthecurrentyear.Note,thatitisunlikelytoresultfrom changesinproductquality.Ifqualityimprovementsaccompanyforeignmarketentry,theyshouldincreaseboththe numeratorandthedenominatorofAVC,sothisphenomenonneednotbeatendencytomis-interpreqtuality improvementsasadversecostshocks. 33

Adding ourmeasure of industry-specificexportactivitytothe model increased thecollinearityof theexplanatory variable setbecause, likethe exchange rate, itiscommonto allfirms inthe industry. The (absolute)correlation betweenthesetwo variables liesbetween 0.66and0.76 inColombia, andbetween 0.62and 0.91 inMorocco, depending uponthe industry. Thesecorrelations, while high,donotprohibit estimation. However, thecondition numberofthemoment matrix oftheregressors--the squarerootof theratio ofthe largesttothe smallestcharacteristic root--liesbetween 648ad 1018inMorocco, butonly between47 and 56 inColombia.z5Inaddition,theestimated coefficients ofthe exchange rateandthe spillovervariables inMorocco were absurdly highand imprecisely estimated.Forthese reasons,we decidedto focusthe spilloveranalysisonColombiaonly.zb Suppressingthenon-exemality coefficients, which are similartothose inTable 3.1,we report coefficients forthe Colombian externalityvariables inTable4.All butoneofthe externality coefficients intheparticipation question arepositive; however, onlyone issignificant. Nonetheless, thisseemsto providesomeevidence thatthe presence ofmany exporters increasesafirm’s chances ofbeingan exporter itself. These resultsbuttressAitken,Hanson andHarrison’s(forthcoming) conclusions basedon cross-sectional analysisofMexican data. The same cannot besaidoftheeffect onaverage variable costs.Four outof sixcoefficientsare positive,suggesting that highexport intensityactually raisescosts. Thechemical industry isparticularly noteworthy because thetwo coefficients are bothsignificantandthey havecounteracting effects. Unit productioncosts arereduced byexport intensityinthe region,perhaps because ofdemonstration effects, andthedevelopment ofbettertransport services forexporters. On theother hand,the presence ofother ‘sForthepurposesofthesecalculationstheregressorsincludeonlyaconstant,theexchangerate,the proportionofexportersintheindustry,theproportionofexportersintheregion,andtheregionaldummies.A conditionnumberabove30suggestspotentialmulticolinearityproblems(Belsley,KuhandWelsch,1980). *bUsingtheproportionofexportersintheregioninsteadoftheproportionofexportersintheindustryisnot feasibleinMoroccobecausemostfms arelocatedinoneregion.Forexample,90percentofthechemicalfirms and70percentofthetextilefirmsareintheCasablancaarea. 34

chemicals exportersappearsto increaseunitcosts.One interpretationisthatother exporters inone’sown industrybid upthe localcostofspecialized inputs. These possibilitiesare intriguing,butgiventhe inconsistence across industrieswerefrain fromgeneralization. 35

Table4: Coefficients ofSpilloverVariables, Colombia Chemicals Textiles Apparel Participation Equation VOExporters, -2.03(3.07) 0.04(6.07) 1.63(1.30) Industry Exporters, 2.74(5.00) 6.14(10.37) 5.01(0.82)* 0/0 Region CostFunction 0/0Exporters, 2.72(0.65)* 0.24(0.85) -0.43(0.62) Industry Exporters, -3.31(0.94)* 1.02(1.30) 1.76(1.30) YO Region Log-likelihood -289.28 -924.65 -1,603.87 Standarderrors are inparentheses. *significantatthe .05 1evel. V. Summary andConclusions Micro data indeveloping countriesoften showthatexporting firmsare moreefficient thannonexportingfirms. This studyconfirmsthatpattern, andaddsthe findingthatplantswhich cease exporting are typically lessefficient, sometimes dramatically so. Butmore importantly,this studyaddressesthe questionofwhether the association between exporting and efficiency reflects causation flowingfrom exportingexperience to improvements inperformance. Surprisingly,despitemany anecdotes inthe literatureto the contrary, we find scantevidence ofsuch acausality pattern. If learning-by-exporting isimportant,then the stochasticprocessesthatgenerate costand productivitytrajectories should improvewith changes inexporting status. To getsome senseforthe natureofthe respone,we began byplottingcostand exportingtrajectories fromactual plant-levelpanel data fromColombia, Morocco and Mexico. We foundthatplantswhich beginexportingtend to have relatively lowaverage variable cost,and plantsthatcease exporting arebecoming increasinglyhighcost, 36

as impliedbythe model. Similar patternsemerged whenwe used laborproductivityasourperformance measure. However, costand productivity trajectories generally didnot continuetochangeafier entering foreign markets. That is,the patternswe found inthe actualdataresembled ourno-learning-by-exporting scenario,underwhichthepositive association betweenexportstatusand productivityisduesolelytothe self-selection ofrelatively more efficient plantsintoforeignmarkets. To formally testwhether theassociation between exporting and efficiencyreflects morethan self-selection, we simultaneouslyestimated anautoregressivecost function andadynamicdiscrete choice equationthatcharacterized exportmarket participation decisions. Exporting historydidnotsignificantly shifithecostfunction,andtothe extentthat itdid,the shifiwas inthe“wrong” direction. The association between exportingand efficiency isthus mostplausibly explainedas low-costproducers choosingto becomeexporters. Finally, lookingforevidence ofexternalities, we foundthatthe presence ofotherexportersmight make iteasier fordomestically-oriented firmsto break intoforeign markets. Inprinciple,this opensthe possibility thatexportpromotion policiesare welfare improving.On the otherhand,thepresence of exporters doesnotappearto reduce theunit productioncostsofneighboring firms inmost instances. Soif exportersactas conduitsof foreign knowledge to localproducers,this effect isweak, slow,ormasked by other cost-increasing spilloversofexport activity. 37

References Aitken,Brian, Gordon H.Hanson, andAnn Harrison (forthcoming). “Spillovers,ForeignInvestment, and Export Behavior.”Journal ofInternationalEconomics. Aw, Bee-Yan andHwang, A.R. (1995), “Productivityand theExportMarket: AFirm-Level Analysis’’JournalofDevelopment Economics 47:313-332. Baldwin, Richard (1989)“SunkCostHysteresis,”NBER WorkingPaperNo. 2911. Belsley,D.A., E.Kuh and R.E. Welsch(1980).RegressionDiagnostics: ldentl~’ingData andSources of Collinearity.New York: Wiley. Bernard,Andrew B.andJ. Bradford Jensen (1995), “ExceptionalExporter Performance: Cause, Effects, orBoth?”,mimeo. Chen, T.and D.Tang (1987), “ComparingTechnical Efficiency between Impoti-Substitution and Export Oriented Firms inaDeveloping Country”Journal ofDevelopment Economics 26:277-89. Dixit,A. (1989), “Exitand Entry DecisionsUnderUncertainty”Journal ofPoliticalEconomy 97:620-38. Evanson,Robert.andLarry Westphal(1995)“Technological Change and Technology Strategy,” inT.N. Srinivasanand Jere Bherman, eds.,Handbook ofDevelopment Economics. Volume3. Amsterdam: North-Holland. Guilkey,D.And J.L. Murphy (1993).“Estimation and Testing inthe Random EffectsProbitModel,” Journal ofEconometrics 59:301-17. Haddad, M. (1993) “HowTrade Liberalization Affected Productivity inMorocco.” Policy Research Working Paper 1096,The WorldBank. Handoussa, H., M.Nishimizu andJ.Page (1986), “ProductivityChange inEgyptianPublic Sector Industriesafierthe ‘Opening’”Journa[ofDevelopment Economics 20:53-74. Heckman, J.(1981a), “StatisticalModels forDiscrete PanelData” inC. Manski andD.McFadden, eds., Structura[Analysis ofDiscrete Data withEconometricApplications. CambridgeMA: MIT Press. (1981b), “TheIncidentalParameters problem andtheProblem ofInitialConditions in Estimating aDiscrete Time-Discrete Data StochasticProcess” inC. ManskiandD. McFadden, eds.,StructuralAnalysis ofDiscrete Data withEconometric Applications. Cambridge MA: MIT Press. Keane, Micheal, RobertMoffit andDavid Runkle(1988) “RealWages over theBusinessCycle: Estimatingthe Impact ofHeterogeneity with Micro Data,”Journal ofPoZiticaZEconomy 96: 1232-66. 38

Krugman, Paul(1989)Exchange rateZnstabiZi~.Cambridge:MIT press. Rhee,Ross-Larson and Pursell(1984)Korea’s Competitive Edge: Managing the Entry into World Markets. Baltimore: JohnsHopkinsUniversity Press. Roberts,Mark, Theresa SullivanandJames Tybout(1995)“What MakesExports Boom? Evidence from Plant-level PanelData, The WorldBank,processed. Roberts,Mark and James Tybout(1995)“An Empirical Modelof SunkCostsandthe Decisionto Export,” PRD WorkingPaper 1436,The World Bank. Roberts,Mark and James Tybout,eds.(forthcoming)IndustrialEvolution inDeveloping Countries.New York:Oxford U. Press. Rust,J.(1994), “Structuralestimation ofMarkov Decision Processes”,inR.Engle andD.McFadden, eds.,Handbook ofEconometrics, VolumeIV VolIV,Elsevier. Sullivan,Theresa (1995)“Micro-foundations ofExport SupplyinMorocco,” Ph.D. dissertation, Georgetown University. Tybout,J. andM.D. Westbrook (1995)“Trade Liberalizationand Dimensions of Efficiency Change in Mexican Manufacturing Industries.” Journal ofInternationalEconomics31 (August)53-78. World Bank (1993) TheEastAsianMiracle.NewYork: OxfordUniversity Press. WorldBank (1994)Kingdom ofMorocco --Republic of Tunisia,Export Growth:Determinantsand Prospects. ReportNo. 12947-MNA. 39

Appendi.rI: TheDynamicOptimizationProblem Appendix I: TheDynamic Optimization Problem A. TheNumerical Solution Inthetypeofproblempresented insectionII,the optimal participationpolicy usually involves twotrigger levelsdefined,forourpurposes, interms ofmarginal cost levels,CL< Cu,suchthat anonexportingplantwill startexportingwhen itscostsare no largerthan ~, and anexporting plantwill cease exportingwhen itscostsequalorare largerthan Cu. Wewillassumethatthis isindeedthetypeofpolicy followed bythe plantand focus on computingthetrigger points. Forthis pupose,we assumethat c isarandomvariablethattakes anyone ofKdiscretevaluescl,...,c~,andevolvesovertime according to astationaryMarkov processwith transitionprobabilitymatrixP = [Pj],where pijdenotesthe probability ofobservingciattgiventhat~ occurred att-1. The valueofany particularpolicy(CbCu)is (Al) T=o where expectationsaretaken overmarginalcostsconditional onthevalueofthe current marginal cost andtheexport statusatthe beginningofperiodt, andwhere, 1 if c,scL or {if C,<CUand Y,-l=1 } Y, ={ (A2) o if C,>CU or {if c,>c~and Y,-l‘0 } isthe binaryvariablethattakesthe valueofone ifthe plant isexporting and Ootherwise. Wetreat the demand shifier#as fixedandwe ignorelearningeffects ondomestic marketprofits. 40

Appendix I: TheDynamicOptimizationProblem Wenowshowhowto compute V~U(c,,Y,.lf)oran arbitrary policy (C. Cu)andgivenparameters.z’ Define s~= (c~,Y~-l)asthe statevariable inperiodt. Let currentreturns underan arbitrary policy be R~u(s,)which equalsY,(n(c,)-M) ifY,-l= 1orY,(n(’,) - M- F) ifY,.l= O(recallthat, from (A2), Y,isa functionof s,). Then,the policy’svalueequalsthe discountedexpected value ofthe sequenceof returns R~u(s,+~)B. ecause cand Yare discrete, >., takesanyoneof2K possiblevalues. Lets= (Sl=(C,,O),...S,~=(C~,O)s,~.l=(c,,l),..., sz~=(c~,l))’,bethe2Kx1 vector ofstatevalueswithcoordinates q, i=l ,...,2K. Then, forexample,the discountedexpected valueoftomorrow’s returns issimplyan average ofthe2K possiblereturnsweighted bythe onestep-ahead (conditional) probabilities ofbeing nextperiod instatesi,giventhattoday’sstate is~andthatthe plantusesthe policy (C’bCu). Denotingthese probabilities byetithis values is~8 2K Notice thattime per-se, hasnoeffect onthe valueofa policy;onlythevaluetaken bythestate variablematters, i.e.,theproblem isstationary. Also,the vector s follows aMarkov process. Hence, fromthe 2Kx2K one-steptransition matrix forthe statevectors, denoted by ~u, onecaneasily obtain the n-steptransition probabilitymatrix ELUandcomputethe expected discounted valueofapolicyfor any initialstateas, VLU(S)=RLV(S)+6ELURLU(S)+62E~uRLu(s)+... =[1-6ELU]-lRLU(S) (A4) where R~u(s)isthe 2KX1vector of returnsateach possiblevalue ofthe statevariable andVLU(Si)sa 27BasedonRust(1994). 28Althoughitisomittedfornotationalconvenience,itshouldbeunderstoodthate. dependsonthevalues ofthetriggerpointsandshouldthereforebealsoindexedby(L,U). Forexample,ifL=3then,whenC=Cl,theplant willbeexporting.Hence,theprobabilityofbeinginstates,iszerosince(C1,Oc)anneverhappen. 41

Appendi.rI: The@.namicOptimizationProblem 2Kx1 vector givingthe valueofthe policy ateach initialstate. Forexample,the first(last)coordinate of V~u(s)isthe valueofthe policy giventhe initialstate(c,,O)((c~,l)). Hence, ifE~uand R(s)are bown V~u(s) can beeasilycomputed. The E~uprobabilitymatrix isderivedafollows. LetRj=Pr(c,=ci/ c~.l=cj)and letelk=Pr(s,=sl/ s,. l=S~).Letpj,=(pjl>...,pjL),bethejh row ofthe matrix P. Then,foranynumbers Os L s U s K, ELU equals, 1 0 LxK o(U-l)XK PL+l Pu. . P K. E (A5) LU = pl. PI. . I PL. P(u-1). o o (K-L)xK (K-U-1 )xK V~u(c,Y) iscalculated for allpolicies satis~ing 1s L<U s K,atotal of K(K+l)/2 - K alternatives. Foreach state~=(ci,Y)we determinethe (relevant)triggercostdeliveringthe maximum valueamong allpossiblepolicies, i.e.,q for stateswith Y= Oand ~ forstateswith Y=1. Thesetrigger pointsdefinethe optimalparticipation policyofthe plantand itamountsto the sequenceofentry/exit decisions {Y,+r}thatmaximizes the expected discounted valueoftheplant. That is, L(si) =argmaxLVLu(si) i = 1,...,K. (A6) U(si) =argmuxu VLu(si) i =K+1,...,2K. 42

Appendix I: TheDynamic OptimizationProblem Noticethatthetrigger points in(A6) areexpressed interms ofthe indicesofthe costvector. From(A6)wecan inferwhich stateshaveprobabilityzero underthe optimal policy. StateswithY= () andcostslowenough sothat ci<C~(C~i,tand,conversely, stateswith Y= 1and ciz Cu(C,i,,haveprobability zero ofoccurring. Atthispointanumerical examplecan be illustrative. Considerthe casewhere K= 10andtheten possiblerealizationsofthe marginalcostsare (1,2,...,10). Assume alsothat givenacostqtodaythe plantcanmoveonlyto neighboringcostsci.lorci+lwith probability orremain with the same costwith probability I-2P. The optimaltriggerpointsassociatedwith each realization of ~=(ci,Y)are shown in TableAl, forp =0.3. Table Al: Trigger Points State Trigger Decision St=(C[,yt.1) Pointz9 Y, C~f,olrCUf,J (1,0) 3 1 (2,0) 3 1 (3,0) 2 0 (4,0)to(10,0) 3 0 (1,1) to(3,1) 4 1 (4,1)to(10,1) 4 0 ‘9Noticethatwiththecurrentparametrizationofmarginalcoststhetriggercostandthetriggerpointarethe same,i.e.CL=Land Cu=U. 43

Appendi.rI: TheDynamicOptimizationProblem Non-exporters facing costsequal or lowerthan 2willenterthe exportmarket immediately since theirmarginalcostsare lowerorequalthan theirtrigger points. Plantswith c> 3,willremain outsidethe exportmarket. Exporting plantswillremain inthe markettilltheir costraiseto 4 ormore, atwhich level theywillcease exporting. Thus, a non-exportingplantwithcs 2will enterthe exportmarket andan exportingplantwithc>4 will quitexpofiing(CL=2 andCu=4). Table Al fully characterizes the dynamicprocess followedbythe plant. B. The Structure ofour Simulations Our simulationsare based onthe numerical example presented above. The critical issueishowto representthe stochasticcostprocess summarized byequation(3). Our strategy isto treat costsas followingafirst-order Markov processthat, inthe caseof learning, isrelatively favorableto exporting firms. Specifically,we discretize c,intoKpossible realizations(c,,. .., c~)and define atransition matrixPOthat governsmovements between thealternativevaluesfrom oneperiod tothe next. This matrix isassumed to beband-diagonal: givenacostqtodaythe plantcan move onlyto neighboringcosts Ci.,orCi+,with probability orremain withthe samecostwith probability I-2P. Notethat this specificationdoesnotallowusto explorethe effectsofhigherordercost processes. Recall, however, that, oureconometrictests insection IVdonot sufferfromthis limitation. Whenthere isno learning-by-exporting,thePOmatrix described aboveappliesto allfirms. But inthe learningcase,we letthe probabilityof obtaininga lowercostnextperiod berelatively largefor exporters.30Weanalyzethree alternativeways inwhichthematrix POchanges intothe matrix P,,where PIisstochasticallybetterthan PO.Inthe firstcase(learningmodeZI), the probability of movingto a lowercostlevelforexporters isp +A,andthe probabilityofmovingto a higher cost leveltop -A. The 30Durationintheexportmarketorvolumeofcumulativeexportsdonothaveadditionaleffects. 44

Appendi.rI: TheDynamic OptimizationProblem secondcase (learningmodeZII) spreadsthesetransitionprobabilitiesto p +2Aand p -2A, respectively. Finally,the third case (learningmodeZII~ makesc1an absorbing statesothatplantswhich achievethe lowestcostremain atthatcost forever;the othercolumnsofP,are as inthe no-learning case. This isthe purest instanceoftechnologicalcatch-up through exporting: oncethebest-practice technology is achieved,the catching up isaccomplished. Thevaluesweassume for p, A. and othermodel parameters are given inTableA2.31 TableA2 :Parameters Values for Simulations # d M F 6 P A K Cl,...,ck 30 2 2 2.1 0.95 0.3 0.1 10 1,...,10 Withtheoptimalbehavioral rules (A6)andthe parameter values inhand,we simulatethe dynamicprocess followed byaplant. We startform an initialstate inperiod 1, s = (cl, YO)randomly selected amongthe2K possibilities.~zThe optimaltrigger value forthis state isobtained from (A6)and compared to cl. The resultfromthis comparison tells uswhether ornotthe plantchanges itsexportstatus fromthat given byYO.This ishow Y, = Y(sl) isdetermined. Assumingthatallthese decisionsaremade atthe beginningofperiodt,Y, isthen the export status ofthe plant dtiring period 1?3 C2is now randomly drawn according totheprobabilities inthe columnofthetransition matrix Pthat corresponds to q,and Y =(cz,Y1)isthus obtained. Followingthe same stepsas inperiod 1,Y5= Y(sz)isgenerated and soon. Inthis fashion, we obtainforeach plantatime seriesof {q,Y,}. Foreach hypothetical plant,we simulate 3’Becasuecl+andzentertheprofitfinction(1)symmetricallywetreatzasfixedinthesimulationssince itsrandomizationdoesnotchangethequalitativeaspectsoftheexcercise. 32Noticethatsubscriptsnowindicatetimeperiods. 33Alternativelyw, ecanintroduceaone-periodlagandmakeY1theexportstatusduringperiod2. 45

Appendi.rI: TheDynamic OptimizationProblem 61periods. Inordertotakeaccount ofany possibleeffect fromthe initialstate,wediscard the first50 periodsandare leftwith 11observations oncand Y. Foreach ofthe four learningregimes, we repeat thisprocess2,000times. The numberofobservations correspondsapproximatelytothe number ofplants inouractual(non-simulated) dataandtothetypical lengthofthetime-series data availablefor each plant. In orderto lookatthe patternsofassociationbetween exporting statusandcosts underour alternativeassumptionsabout learning-by-exporting,andto facilitate comparison between simulatedand actualdata,wesorttrajectories accordingtothe fivevarieties ofplantsdescribed intable 2 inthe text. Next, foreach plantwere-define time zeroto bethe year inwhich achangeinexportstatustakes place, and isolateseven-yearblocksoftime, runningfromthree years priortothe statuschange (t= -3)tothree years after (t= 3).sdFornon-exporters andfor exportersthere isno change instatus,sowe takethe seven years inthe middleoftheeleven-years period. Finally,after re-indexingtime inthis manner, we average the simulatedmarginal costdata byplanttypeand plotthem againsttime.ss c. The Simulations FigureAl summarizes oursimulationsforthe caseofno learning. 36 Note that non-expofiers andexitingfirmshavehigher productioncoststhan exporters, but exitingfirmsmaintain acostadvantage overperpetualnon-exporters. Without learningeffects, producers exitwhen drawing ac>4. More importantforourpurposes isthe constancyofthemarginal costtrajectory forentrantsafter period O, 34Thusforentrants,periodOindicatesthefirstyearexportingandforexitorsitindicatesthefirstyearnot exporting. 35Sincetheentry/exitdecisionsoccurindifferentperiodsweareaveragingcostsfromdifferentyearsbut equallydistantfromtheperiodinwhichtheplantchangedexportstatus.Entrantswererequiredtohaveatleast4 yearsexportingand3yearsnon-exportingwhileexitorswererequiredtohaveatleast4yearsnon-exportingand3 yearsexporting. 36Switchingfirmsgenerallylooklikeexporters,soweomitthemtoreduceclutter. 46

Appendix I: The@namic OptimizationProblem their entry time. Inorderto becomean exporter, a planthasto drawacost equaltoorbelow2and selectingplantsthateventuallyentertheexportmarket inducesthe negative slope inthe costtrajectory priorto periodO.Afier entry,there doesnotseemto beasignificantchange inproductioncosts. In short,the simulationsinfigureAl reflect the premise ofourno-learning transition matrix:becoming an exporterdoesnothaveany significanteffects onthe stochasticcostprocess?’ Noticethat, givenourtransition matrices, ittypically takesmore than oneperiodofdeclining coststo reachthe entrythreshold. Hence thenegativeslopeofthe costtrajectory forentrantspriorto entry isaconsequenceofourgroupingcriteriaand notof learning-effects orsunk-entrycosts?g FiguresA2.1-A2.3 showthatthemain difference beween thepassive and activemodelofexport participation isthatentrantscontinueto lowertheir costsafter enteringthe exportmarket. Thepattern of costtrajectories fortheothertypes ofplantsdoesnotchangemuch with the introductionof learning effects. In particular,itremainstruethat exporters havesystematically lower costthan non-exporters, andexitorsexperiencecostdeterioration priorto their exit. Interestingly, even with learningeffects,the rateof reductioninmarginal costsslowswhen firms become exporters. This illusorynegativeeffect of becoming anexporter issimplyaconsequence ofthe selectivity effects. To reach marginalcost levels lowenoughtobecomeanexporter, firms musttypically experience several consecutive yearsofcost reductions. TableA3 summarizesthe results ofthe simulations. 37Wecautionthatifcostsfollowedasecond-orderprocesstheymightcontinuetofallafierentrybecause ofthelingeringeffectsofpre-entrycostreductions. 38Thesamereasoningexplainstherisingcostpriortoexittimeforexitingplants,whiletheflatnessofthe costtrajectoryafierentryreflectsthesymmetricfirst-ordermarkovprocessfollowedbycosts. 47

i AppendixI: TheDynamic OptimizationProblem FigureAl: Simulated Pathof Costs No-Learning Model 9 8 7 — — — — — — a) ~5 > ‘4 3 — — 2 — — 1I -3 -2 -1 0 1 2 3 Year(Oistransition/middyleear) + Non-exwder~ Exporters + Entrants + Quitters FigureA2.I: Simulated PathofCosts LearningModelI 10 9 8 -. — — . — — — 7 g 06 a) m $5 2 4 3 — — — z. : . — — 1I .3 -2 -1 0 1 2 3 Year(Oistransition/middle year) + Non-exwfier% Exporters + Entrants + Quitters L 1 48

AppendixI:TheDynamic Optimization FigureA2.2:SimulatedPathofCosts LearningModelII — — — — — — — — — — — . — — — — — — -3 -2 -1 0 1 2 3 Year(Oistransition/middleyear) 7 + Non-expoRer~ @otiers + Entrants + Quitters Figure A2.3: Simulated Path of Costs Learning Model Ill 10 9 8 7 — — — — — — — — ~ 06 a) m $5 2 4 3 2 . — 1 — — — — — — -3 -2 -1 0 1 2 3 Year (Oistransition/middleyear) ~ Non-expo~ Exporters~ Entrants * Quitters 49

Appendi.rI: TheDynamic OptimizationProblem TableA3:Simulation Results Model Entry Exit Entrants’Average Ratio ofEntrants’ Trigger Trigger Change inCosts toNon-exportes’ CL Cu 3 yearsafier costs 3 years after No Learning 2 4 4V0 31 % Learning I 3 6 -16 % 35 Yo Learning II 4 8 -27 % 38% Leamin~ III 2 4 -2R0A 21 Yo Becausethe incentivesto export increasewhen learningoccurs,thetrigger points shifiupwards. That is,itpaysto enteratahigheraverage costlevelinanticipation of lowercostdueto leaming-byexportingand,similarly,itpaysto remain anexporters even at highercost levels. This property ofthe model impliesthatthecross-sectional distributionofcostsfor exportersshifistothe rightwhen leamingby-exportingoccurs. Thismay well result inhigherexpected costs forexporting plantswhen learning effects are present. Putdifferently, less-efficientfirmstend to entertheexport market and remain exporters for longerperiodsoftime when learningeffects prevail . Figure2.2Aprovides5uchanexample. In this case, the entry trigger is ~= 4 andthe exittrigger isCu=8. Afier threeyears inthe exportmarkettheaverage costs is2.93while inthe no-learningmodel it equals2.08. Hence,eventhough costsdeclineby27 percent afier 3years exporting,the averagecostsof exporters relativetotheaverage costofnon-exporters ishigherwhen learning-by-exporting ispresent.39 The messageofthis example issimplythatdrawing conclusionsaboutthe presence of learning effects across industriesfrom comparisonsofentrants’cost levelsvis-a-visthe costsofnon-exporters 39Notealsothattheaveragecostsofnon-exportersishigherwhenlearningoccurs- 50

Appendix I: TheDynamic OptimizationProblem may bemisleading. Hence,cross-industry comparisonsofthe productivitygap be~veenexporters and ‘ non-exporterstellusnothingaboutthe relativestrengthof learningeffects indifferent industries. Put differently, inorderto discriminate betweenexplanations forthe export-productivity relationship,onehas to examinethe dynamictrajecto~ ofcosts;asinglecross-section \villnotdo. 51

AppendixII:DataPreparation APPENDIX II: DataPreparation A. ThePanelData Sets The Colombian datawere obtainedfromthe Departamento AdministrativoNacional de Estadisticaforthe period 1981-1991. They provideannual informationonthe inputs,outputs,exports, andcharacteristics ofallplantswith at least 10workers. Similarannualsurvey datawere obtainedfrom Morocco’sMinistryofCommerce and Industryfortheperiod 1984-1990and fromMexico’s Instituto Nacional deEstadisticaGeografia eInformation fortheperiod 1984-1990,althoughthe Mexican data onlycover3200 ofthe largerfirms. Mostofthe detailsonthe cleaning anddeflating ofeach ofthesedata basesmay befound inthe relevant country studychaptersofRobertsandTybout (forthcoming). However, for ouranalysisofaverage costs itwas necessaryto pay specialattentionto the deflationof exports, soadditional measures were taken. Our procedure issummarized below. B. Construction of Export Unit Values, Price Indicesand Average CostMeasures40 Trade data for the the selected sectors for Colombia, Mexico and Morocco were obtained from the United Nations Trade Database. The data are amual totalvalues and quantitiesfor exports and imports for the years 1980-1992. The values are innominal U.S. dollars, and the data are at themost disaggregate level available within the selected sectors. Disaggregated export unit values were constructed from this data, but could not be used directly in deflatingthe export volumes reported in the country databases for two reasons. First, because the countriesuse industrial classificationsystems that differ from the U.N. system, the classificationsystemshad to be matched. Second, the unit values had to be aggregated up to the level observable in the survey data in each country. 40ThissectionisreproducedfromRoberts,Sullivan,andTybout(1995),whichisbasedonthesamedata. 52

Appendix II:DataPreparation Colombiauses a version ofthe International Standard Industrial Classification System(ISIC), revision2. The ISIC was matched to the U.N. ’SStandard Industrial Trade Classification (SITC) systematthe four digit ISIC levelfor the industriesof interest. Becausea detailed listingofproducts ineach four digit ISIC isavailable (United Nations, 1973),thismatching is quiteaccurate. The ‘ Mexico Industry Classification System(Censo Industrial 1975,referred to here as MICS) was matched tothe SITC for those four digit industrieswithinthe eightindustriesof interest that actually appear in the Mexico data from 1984-90. Thismatching is lessprecise because MICS isonly availableat a higher levelof aggregation than the SITC. The Moroccan Nomenclature of Industrial Activity (NMAE) was matched to the SITC atthe very disaggregated sixdigit level, and the correspondence shouldtherefore be accurate. After the industry classificationsystemswere matched, the disaggregated export unit valueshad to be aggregated. Initially, export unit values were aggregated across allSITC commodities contained ina country-specific disaggregate sector (e.g., Colombia four digit ISIC) using as weights actual export value shares from the country data. Some of the export unit valuesfor the SITC commodities showed an unreasonable amount of variation from year to year, however, and thispersisted in these initialcalculationsof an export unit value for an aggregate of these commodities. Because ofthis, the export unit values were instrumentedusing importunit values. Three methods of using the import unit valueswere considered, two of which use information from all countries for which SITC data was obtained.4*First, a weighted average of import unit values for each commodity was calculated usingas weightsthe actual import value share of each country for each commodity. Second a weighted average importunit value was calculated using the average import value share ofeach country for each ‘i InadditiontodatafromColombia,MexicoandMorocco,Venezuelatradedatawasusedininstrumenting exportunitvalues. 53

Appendix II:DataPreparation commodity, where the average istaken over the timeperiod for which the data was obtained. Third, for each country, the import unit valuesfrom thatparticular country could be used. Ultimately, the secondmethod of aggregating the importunitvalues was chosen. The first method using actual shares wasnot chosen because the importunit values in each year were greatly influenced by commodities entering or exitingtheunit value aggregation. The third method using only country specific informationwasnot used because it ignored valuable information on unit values from other countries. The actualexport unit values were regressed onthe importunit values calculated usingthe secondmethod described above, yielding fittedvalues for each disaggregate SITC commodity. Two methodsof aggregating the export unit valuesfor these SITC commoditiesto the levelobservable in the country databaseswere considered. The fitted export unit values could be aggregated using either country-specificactual export value shares or average shares over time from the SITC data. The latter method was chosen, again to eliminatethe effect on the unit values of changes inthe commodity compositionofthe sectors. These instrumented export unit valueswere in nominal U.S. dollars, and had to be converted to domesticcurrency unitsbefore they could be used for deflating export values reported in the country databases. This required nominal exchange rate indicesand a U.S. price index. The U.S. producer price index for consumer goods reported inInternational Financial Statisticswas used to convert to real U.S. dollars. For each country, the amual average nominal exchange rate between the domestic currency and the USdollar reported invarious issuesof International Financial Statisticswasused to convert the export unit values in real U.S. dollars to domestic currency units. Country specificoutput price indiceswere constructed, where possible, using the survey data, or insome casesusingother information provided by the respective countries. The Colombiadatabase containsdata on both nominal and real totalvalue ofproduction for each plant. The ratio ofnominalto 54

Appendix1[:DataPreparation real value ofproduction isthe implicitoutputprice deflator, and this is sector specificatthe three digit SIC level. There is someunexplainedvariation inthethree digitprice index across plantswithin an industry, sothe mean price indexacross plants in an industryisused as the outputprice index. Several plantsare deleted from the calculationof the mean price index inparticular years because their implicit price index as calculated from thesurvey data isfar from the industry average in aparticular year. A broad based outputprice index for Colombia was constructed as a weighted average ofthe sectoral outputprice indices, where the weightsare sectoral shares oftotalreal value ofproduction. For Morocco, the sectoral domestic outputprice index istaken from various years of the Moroccan StatisticalBulletin. A broad based domestic outputprice index was constructed as a weighted average of the sectoral outputprice index, using sectoral shares oftotalmanufacturing production as weights. 55

Appendi.rIII:TheLikelihoodFunction Appendix III:The LikelihoodFunction Ourestimatorforequations(10)and(11) iscloselyrelated to Keane, Moffit,and Runkle’s (1988),aswell asSullivan’s(1995). To begin,we writethese equations inshortenedform bycollapsing theirright-hand-sidevariablestothevectorsZ,andZz,respectively: b Yi; Zl,p,,+al + ~~ r =J+l9..9T = Cit=z2i,p2 + U2 + 62 t = J+I,..,T where 1 iJ Yi; 20 Yi, ={ o if Yi; < 0 (Recallthat~periods of lagsareneeded.) Thenthe likelihoodfunction, conditionedon~, 22,, al, and az may bewritten as: L(Y,c12,,Zz>al$aze$) = i ; ~ciJYi,=l)Pr(Y,=,l)]r”~ciJY,,=o)Pr(Y=i,o)]’-y” 1=1~s~+l = ; ; ~ciJYi;202)0P])rY( ~”Y ci , JY ; i;<o)Pr(Yi,o<)]’-y” ~eltc~1+ where e=(PlP,2 C , Y0~~, U 2, e , ,,E2.) To simpli~ the conditionaldensity functions,notethat: . fici)Yi;20)Pr(Yi; >o) = j ficit,yi;)dyi; = J fiy,;]ci)flci)dyi; o 0 ‘fici)[l-ifiyi’’ciJd = flcif) ~flYi;[C,JdYi; ‘fic-iG)[l(0)] o 56

Appendix III: TheLikelihood Fltnction where G( ) isthe cumulativedistribution for Yi: [Cif . similarly, flci,lY,;sO)Pr(YirsO)=flciJG(0). Assumingthat (~li,,e,,,) isjointly normal,we have: ‘i;’caifl-i”+N[%[z“)’1(icr-p*ir1l-+7z2ifp2-a sothe conditionallikelihoodfunction can bewritten as where 7 and 0( ) and $( ) arethe standard normal distributionand density functions,respectively. There aretwo complications involved inestimation. The first isthatthe error components a, and al, are unobserved,sothey cannotbeconditionedupon inestimation. The second derives . fromthe factthat laggedendogenous variables appear inbothZ,andZz,. To deal withthe firstproblem, we assumeabivariatenormaldistribution fortheerror components , ~(al, a,) , and integrate them out.qzTo dealwiththe secondproblem, we followHeckman’s (1981)suggestion ofadding equationsto 42Morespecificallyw, efirstexpressthetwocomponentsaslinearcombinationsoftwoorthogonalrandom variablesusingaCholeskydecomposition.Thenweintegrateouttheseorthogonalrandomvariablesusingbivariate Gaussianquadrature.SeeSullivan(1995)fordetails. 57

Appendix III: TheLikelihoodFunction the systemthatrepresentthedependence of (Yi,,Yil,..,YU)on ali andthe dependence of on azi : (Ci,, Ci2, .., cu) ‘i: =‘;fv,+ t = 1,..,J P,a, + ~,i, cIt=‘j[$* +P~‘* + ~~i~ t=l ,.. ,J Here z;, c Zli, isthe vectorofstrictlyexogenousdeterminants of Y,; , Z2~c,Z2i,isthe vectorof strictly exogenousdeterminants of Ci, , and (~li,,~2i) isa serially uncorrelated bivariate normal randomvector. The likelihoodfunction conditionedonlyon observabledatabecomes: co. L= ; iflc,)[l - G(o)]y”G(o)’-y” i flci)[l - G(0)]y”G(0)’-*” h(al,a,) da, da, - [ = / -m j=][l=] 1[t=J+ ] 1 where The nuisanceparametersassociated withtheJpresample years ofdataarenotreported inTable 6. 58