IS LEARNING BY EXPORTING IMPORTANT? MICRO-DYNAMIC EVIDENCE FROM COLOMBIA, MEXICO, AND MOROCCO* SOFRONIS K. CLERIDES SAUL LACH JAMES R.

Transcription

1 IS LEARNING BY EXPORTING IMPORTANT? MICRO-DYNAMIC EVIDENCE FROM COLOMBIA, MEXICO, AND MOROCCO* SOFRONIS K. CLERIDES SAUL LACH JAMES R. TYBOUT Do irms become more eicient ater becoming exporters? Do exporters generate positive externalities or domestically oriented producers? In this paper we tackle these questions by analyzing the causal links between exporting and productivity using plant-level data. We look or evidence that irms cost processes change ater they break into oreign markets. We ind that relatively eicient irms become exporters; however, in most industries, irms costs are not aected by previous exporting activities. So the well-documented positive association between exporting and eiciency is explained by the sel-selection o the more eicient irms into the export market. We also ind some evidence o positive regional externalities. I. INTRODUCTION Many analysts believe that export-led development strategies improve technical eiciency. And one ot-cited reason is that exporters may beneit rom the technical expertise o their buyers: 1 Participating in export markets brings irms into contact with international best practice and osters learning and productivity growth [World Bank 1997].... a good deal o the inormation needed to augment basic capabilities has come rom the buyers o exports who reely provided product designs and oered technical assistance to improve process technology in the context o * We thank Eli Berman, Andrew Bernard, Donald Keesing, Lawrence Katz, Wolgang Keller, Yair Mundlak, Dani Rodrik, and an anonymous reeree or their comments, as well as seminar participants at Johns Hopkins University, the University o Caliornia at Los Angeles, the University o Maryland, the World Bank, Michigan State University, and the NBER Summer Institute (International Trade and Investment session). We are also grateul to Manuel Arellano and Steven Bond or the use o their Dynamic Panel Data program. This paper was unded by the World Bank research project Micro-Foundations o Successul Export Promotion, RPO and the World Bank s International Trade Division, International Economics Department. This paper was written while Clerides was a Summer Intern at the World Bank, Lach was visiting the Industrial Output Section o the Federal Reserve Board, and Tybout was visiting the International Trade Division in the International Economics Department o the World Bank. The views expressed herein are those o the authors, and do not necessarily relect those o the World Bank or the Federal Reserve System. 1. For a recent catalog o additional reasons, see World Bank [1993, pp ] by the President and Fellows o Harvard College and the Massachusetts Institute o Technology. The Quarterly Journal o Economics, August 1998

2 904 QUARTERLY JOURNAL OF ECONOMICS their sourcing activities. Some part o the eiciency o export-led development must thereore be attributed to externalities derived rom exporting [Evenson and Westphal 1995]. Buyers want low-cost, better-quality products rom major suppliers. To obtain this, they transmit tacit and occasionally proprietary knowledge rom their other, oten OECD-economy, suppliers [World Bank 1993, p. 320]. The important thing about oreign buyers, many o which have oices in Seoul, is that they do much more than buy and speciy....they come in, too, with models and patterns or Korean engineers to ollow, and they even go out to the production line to teach workers how to do things [Rhee, Ross-Larson, and Pursell 1984, p. 41]. When local goods are exported the oreign purchasing agents may suggest ways to improve the manuacturing process [Grossman and Helpman 1991, p. 166]. In support o this view, empirical studies oten ind that exporting plants are more eicient than their domestically oriented counterparts [Aw and Hwang 1995; Bernard and Jensen 1995; Chen and Tang 1987; Haddad 1993; Handoussa, Nishimizu, and Page 1986; Tybout and Westbrook 1995; Roberts, Sullivan, and Tybout 1995]. But, with the exception o Bernard and Jensen, none o these studies has asked the question o whether exporting causes eiciency gains. Plausible arguments can be made or causality to low in the opposite direction: relatively more eicient plants sel-select into export markets because the returns to doing so are relatively high or them. In this paper we attempt to sort out the direction o causality, and in so doing, determine whether there is any evidence that irms learn to be more eicient by becoming exporters. Further, to assess the case or active export promotion, we test whether exporters generate external beneits to other irms, either by acting as a conduit or knowledge that they acquire through trade, or by causing improvements in international transport and export support services. 2 Our methodology or detecting learning eects is based on a simple idea. I exporting indeed generates eiciency gains, then irms that begin to export should thereater exhibit a change in 2. Many believe that these spillover eects are signiicant in developing countries. For example, Aitken, Hanson, and Harrison [1997] write that... the development o garment exporters in Bangladesh, suggests that inormational externalities are likely to be extremely important. The entry o one Korean garment exporter in Bangladesh led to the establishment o hundreds o exporting enterprises, all owned by local entrepreneurs....spillovers may take a variety o orms. The geographic concentration o exporters may make it easible to construct specialized transportation inrastructure, such as storage acilities or rail lines, or may improve access to inormation about which goods are popular among oreign consumers.

3 IS LEARNING BY EXPORTING IMPORTANT? 905 the stochastic process that governs their productivity growth. Hence their productivity trajectories must improve in some sense ater they enter oreign markets. Similarly, i the presence o exporters generates positive externalities, nonexporters in the aected industry or region should exhibit changes in their cost process when the number o exporters changes. Increases in the number o exporters may also make it easier or others to break into oreign markets. To keep track o causal linkages, we begin by speciying the general optimization problem that we envision irms as solving (Section II). Each manager aces stochastic cost and oreign demand processes, and chooses which periods to participate in oreign markets. Their decisions are complicated by the presence o sunk start-up costs when they irst sell abroad, since managers must research oreign demand and competition, establish marketing channels, and adjust their product characteristics and packaging to meet oreign tastes. The basic eatures o this model are taken directly rom the hysteresis literature developed by Baldwin [1989], Dixit [1989], and Krugman [1989]. Our twist is to add the possibility o learning-by-doing or, more precisely, learning-byexporting, and examine how this aects the productivity trajectories o exporters and plants that switch markets, relative to those o nonexporters. Because the ramework we develop does not lend itsel to closed-orm solutions, we discuss its implications heuristically and using simulations. Under certain assumptions on the exogenous shocks to productivity and demand, the simulations suggest that (a) nonexporters experiencing positive productivity shocks sel-select into oreign markets, (b) exporters experiencing negative productivity shocks quit oreign markets, and (c) the presence o learning-by-exporting eects implies that irms improve their relative productivity ater they begin exporting. With these results in hand, we examine the actual perormance o Colombian, Mexican, and Moroccan producers (Section III). To amiliarize ourselves with the data, we begin by comparing the productivity trajectories o producers that enter export markets with those o nonexporters, ongoing exporters, and irms that exit oreign markets (productivity is proxied by average variable cost and by labor productivity). This exercise reveals patterns that our simulations suggest we should ind in the absence o learningby-exporting eects. That is, the plants that become exporters typically have high productivity beore they enter oreign mar-

4 906 QUARTERLY JOURNAL OF ECONOMICS kets, and their relative eiciency does not systematically increase ater oreign sales are initiated. In some instances the relatively strong perormance o exporters traces to high labor productivity; in other instances it is due to relatively heavier reliance on skilled labor. This irst look at the actual trajectories casts doubt on the importance o learning-by-exporting. But it does not constitute a ormal test o whether becoming an exporter changes a irm s productivity trajectory. Accordingly, in Section IV we estimate econometrically a reduced-orm version o the theoretical model that takes explicit account o the two alternative, but not incompatible, explanations or the positive association between exportparticipation status and productivity: sel-selection o the relatively more eicient plants, and learning-by-exporting. For the countries with suicient data to support inerence Colombia and Morocco we ind that export market participation generally depends upon past participation and (weakly) upon past average variable cost (AVC), as implied by the model. However, conditioning on capital stock and past AVC realizations, current AVC does not typically depend negatively upon previous export market participation, as implied by the learning-by-exporting hypothesis. Thus, with a ew possible exceptions, the conclusion suggested by our descriptive analysis is borne out by ormal Granger causality tests. 3 Finally, extending our model in order to look or externalities, we ind some support or the hypothesis that a irm is more likely to export i it belongs to an export-intensive industry or region, and that Colombian irms in export-oriented regions enjoy relatively lower production costs, regardless o their own market orientation. 3. These conclusions are also consistent with Bernard and Jensen s [1995] indings using data on U. S. manuacturing plants and irms during Both papers address the same basic question: does exporting cause increases in productivity, or do increases in productivity increase the probability o export participation? Besides dierences in the data, and in that Bernard and Jensen s paper also looks at other measures o perormance, the two papers dier in the methodology o the analysis. In our paper we build an econometric model that is tightly linked to a theoretical model. The econometric model both is dynamic and accounts or unobserved (and observed) sources o heterogeneity across irms. This approach has the advantage that regression results can be interpreted within the context o the model, and that it enables us to deal with the endogeneity o the key variables in the analysis, i.e., export participation and the measure o productivity, in order to obtain consistent estimates o their eects. The disadvantage is the computational complexity o the estimation procedure. Despite these dierences, it is comorting that the empirical indings in both papers are virtually the same.

5 IS LEARNING BY EXPORTING IMPORTANT? 907 II.A MODEL OF EXPORT PARTICIPATION WITH LEARNING EFFECTS Our irst task is to present a model that speciies endogenous and exogenous sources o variation in the two producer characteristics we are interested in: exporting status and production costs. This model, which will guide us in our empirical analysis, is a simple modiication o existing models to accommodate potential learning eects rom export participation [Baldwin 1989; Dixit 1989; Krugman 1989]. We begin by assuming monopolistic competition, so that each irm aces a downward sloping demand curve in the oreign market, yet views itsel as too small to strategically inluence the behavior o other producers. Speciically, we write oreign demand q or the irm s product at price p as q z (p ), where the random variable z captures the usual demand shiters (oreign income level, exchange rates, and other goods prices) and 1. 4 Firms ace similar demand conditions in the domestic market, q h z h (p h ) h, and can price discriminate between oreign and domestic buyers. Assuming that marginal costs (c) do not depend upon output, the current period gross operating proits can be expressed as a unction o marginal costs and demand conditions in both markets: (1) (c,z) c 1 z ( 1) 1 ( ) c 1 h z h ( h 1) h 1 ( h ) h (c,z ) h (c,z h ), where z (z,z h ). The proits rom exporting are the shaded area depicted in Figure I, where (1 (1/ )) p and (1 (1/ h )) p h are oreign and home market marginal revenue, respectively. 5 We represent the home demand curve as approaching the vertical axis above the oreign demand curve because transport costs and trade barriers eat up a raction o each unit o revenue generated in oreign markets. Let the per-period, ixed costs o being an exporter (e.g., dealing with customs and intermediaries) be M. Then, the plant will earn positive net operating proits rom exporting whenever 4. This particular unctional orm or the demand unction is generated by the Dixit-Stiglitz utility unction over varieties. 5. We have linearized the demand and marginal revenue curves to simpliy this igure.

6 908 QUARTERLY JOURNAL OF ECONOMICS FIGURE I Gross Operating Proits rom Exporting (c,z ) M. Accordingly, i there were no start-up costs associated with becoming an exporter and no learning eects, producers would simply participate in oreign markets choosing the proitmaximizing level o exports whenever this condition was satisied. As Figure I demonstrates, given demand conditions, all irms with marginal costs below some threshold would sel-select into export activities. But as Baldwin [1989], Dixit [1989], and Krugman [1989] have stressed, sunk start-up costs modiy the problem in a nontrivial way. Suppose that an entry cost o F dollars is incurred every time the plant decides to enter or reenter oreign markets. Then, once exporting, it may be optimal to keep exporting even i (c,z ) is currently less than M since, by remaining in the export market, the plant avoids uture reentry costs. So, i sunk costs are important and microevidence suggests that they are (e.g., Roberts and Tybout [1997]) producers ace a dynamic optimization problem where, in each period, they must choose whether or not to export on the basis o currently available inormation. This makes decision-making orward-looking and opens the possibility that irms export today in anticipation o cost reductions, or oreign demand increases, later on. Because expectations are important in this context, we must be speciic about the processes that generate the state variables c and z. On the demand side, we assume that z is exogenous to the

7 IS LEARNING BY EXPORTING IMPORTANT? 909 plant and ollows some serially correlated, plant-speciic process (or simplicity, the plant-subscript is omitted until Section IV): (2) z t (x t,z t 1 ). Here x t is a vector o exogenous variables that shit the demand processes, e.g., the exchange rate and plant-speciic noise, and z t 1 (z t 1,z t 2,z t 3,...)denotes the vector o previous realizations o z up to and including period t 1. Marginal cost also ollows a plant-speciic serially correlated process. In addition, it is potentially aected by the irm s exporting decisions i there are learning eects: (3) c t g(w t,c t 1,y t 1 ), where w t is a vector o exogenous cost shiters, e.g., actor prices and plant-speciic noise, c t 1 denotes the vector o previous realizations o c up to and including period t 1, and y t 1 denotes the history o the binary variable y t which, in turn, indicates whether the plant was exporting j periods ago ( y t j 1) or was not ( y t j 0). Note that equation (3) implies learning-by-exporting depends only on a irm s previous participation in oreign markets, and not on the cumulative volume o its exports. This keeps the simulations in this section tractable. However, in Section IV we will econometrically test alternative speciications in which the history o exporting volumes enters the cost unction. In addition, we will add variables that measure the amount o exporting activity in irms industry or geographic region in order to test or spillover eects. A test or learning-by-exporting eects using equation (3) must recognize that y t 1 is a vector o endogenous variables depending upon the same observed and unobserved actors that aect the cost and the demand processes. Thus, we need to model the process by which irms sel-select into export markets i we are to sort out the relationship between c and y. Following the hysteresis literature, we assume that managers take equations (1) through (3) into consideration, and plan their export market participation to satisy (5) V t max E t 5 y t r6 r 0 0 [ y t ( (c t,z t ) M (1 y t 1 )F) h (c t,z h t )],

8 910 QUARTERLY JOURNAL OF ECONOMICS where E t is an expectations operator conditioned on the set o inormation available at time t, and is the one-period discount rate. 6 Domestic proits enter this expression only because, with learning, participation aects the uture cost trajectory. We assume that irms never wish to liquidate. Equivalently, managers can be viewed as choosing the current y t value that satisies Bellman s equation: V t max y t [ y t ( (c t,z t ) M (1 y t 1 )F) h (c t,z t h ) E t (V t 1 0y t )]. This characterization o behavior implies that producers participate in export markets whenever (6) (c t,z t ) M [E t (V t 1 0y t 1) E t (V t 1 0 y t 0)] F(1 y t 1 ). That is, incumbent exporters continue exporting whenever current net operating proits rom exports plus the expected discounted uture payo rom remaining an exporter is positive, and nonexporters begin exporting whenever this sum, net o start-up costs, is positive. Expected uture payos include the value o avoiding start-up costs next period and any positive learning eects that accrue rom oreign market experience. Without learning eects, expression (6) has appeared in various orms in the hysteresis literature; it will prove useul in Section IV. I we allow or much generality in the cost process (3), or we allow start-up costs to depend upon exporting experience more than one period ago, it is very diicult to characterize optimal behavior in this ramework. However, some insights can be gained by assuming that c ollows a discrete, irst-order Markov process that depends only on y t 1. 7 Learning-by-exporting can then be captured by changes in the transition matrix governing the evolution o c over time. Speciically, let P 0 be the transition matrix or c when the irm is not exporting, and let some stochastically better matrix, say P 1, be the transition matrix when 6. This ormulation implies that producers who exit the export market and reenter ace the same start-up costs as producers who never exported. In our econometric rendering o the decision to export in Section IV, we will allow start-up costs to depend upon previous exporting experience. 7. Demand shiters can be held constant since their eect on proits is qualitatively the same as that o c.

9 IS LEARNING BY EXPORTING IMPORTANT? 911 the irm is exporting. That is, when P 1 governs c realizations, the probability o a decrease in c is greater than when P 0 governs c.in the no-learning case P 0 is assigned to all irms. In this type o problem, the optimal participation policy usually involves two threshold levels deined, or our purposes, in terms o marginal cost levels c L c U such that a nonexporting irm will start exporting when its costs all to c L and an exporting irm will cease exporting when its costs rise to c U. We will assume that this is indeed the type o policy ollowed by the irm. Thus, i learning eects are present, the transition matrix governing the Markov process c changes at c L rom P 0 to P 1, and returns to P 0 when c reaches c U. Simulations o cost trajectories based on this relatively restrictive ramework and arbitrary parameter values are presented in Figure II (details are provided in Appendix 1 o the working paper version; see Clerides, Lach, and Tybout [1996]). The trajectories are averages over repeated simulations or our subgroups o irms, labeled nonexporters (those that never export), exporters (those that always export), entrants (those that begin exporting), and quitters (those that cease exporting). For all irms that begin or cease exporting, we measure time relative to the transition year (period 0). For example, period 2 is two years prior to entry or the entrant group, and two years prior to exit or the quitter group. Firms that export may or may not be subject to learning eects. Several patterns merit note. First, and most obviously, regardless o whether learning eects are present, costs among the exporter irms are lower than costs among the nonexporters. This relects the sel-selection o eicient irms into export markets, as we discussed earlier in connection with Figure I. Second, irms that become exporters exhibit cost declines beore they enter the market. This occurs with and without learning eects. These irms sel-select into exporting only when their unit costs all below the entry threshold, c L, so they must experience a period o alling costs prior to entry. Selection eects may thus create the illusion that becoming an exporter actually retards productivity growth. More generally, sel-selection means that cross-sectional dierences in the productivity growth rates o exporters versus nonexporters should not be used to quantiy learning eects.

10 912 QUARTERLY JOURNAL OF ECONOMICS FIGURE IIa Simulated Average Cost Trajectories without Learning Third, regardless o whether learning eects are present, irms that eventually cease exporting exhibit cost increases beore they quit the export market. This is also a relection o the two-threshold policy ollowed by the irm in the presence o sunk start-up costs. Fourth, one distinguishing eature o the learning trajectories is that exporting irms exhibit ongoing cost reductions ater initiating oreign sales as a result o the switch to a better transition matrix. Only when learning eects are present, do irms continue to pull away rom nonexporters ater entering the oreign market. Finally, relative to the no-learning case, learning allows irms to enter (and stay in) export markets with higher production costs. This occurs because the incentives to export are larger when learning occurs. Productivity dispersion may thus be higher among exporters when learning eects are present, and the

11 IS LEARNING BY EXPORTING IMPORTANT? 913 FIGURE IIb Simulated Average Cost Trajectories with Learning productivity gap between exporters and nonexporters may be smaller. More generally, this result suggests caution interpreting studies that use productivity dispersion as a perormance measure (as in the eiciency rontier literature), and that relate these measures to trade orientation. 8 O course, these results are only suggestive, and more complicated cost processes might reverse some o the patterns. For example, i costs were to ollow a second- or higher-order autoregressive process, they might continue to trend downward ater export market entry even without learning eects. Or, i or c c L expected cost increases were relatively large, then learning eects 8. For example, Caves and Barton [1991] regress dispersion-based productivity measures on measures o export intensity and ind that [p]articipation in exporting activities... is negatively related to technical eiciency [p. 93]. They go on to hypothesize that any procompetitive eects o exporting activity may be swamped by the uneven distribution o quasi-rents and apparent technical ineiciency that exporting is likely to produce [p. 94].

12 914 QUARTERLY JOURNAL OF ECONOMICS would not necessarily imply cost reductions ater entry. Learning eects imply only a change or the better in the stochastic process ollowed by the cost variable ater the irm enters the export market. Accordingly, in order to look or learning eects and externalities, we rely on econometric estimates o a general orm o the cost unction (3), recognizing that export market participation is governed by the behavioral rule (6). III. LEARNING EFFECTS: A LOOK AT ACTUAL DATA Beore we report the results o this exercise, however, it is instructive to visually examine cost data on actual producers rom three semi-industrialized countries. Though merely suggestive, this will amiliarize the reader with the basic patterns we are trying to explain, and provide an inormal check or the distinguishing patterns that our simulations suggest we should ind when learning eects are present. A. The Data Our data allow us to ollow individual producers through time in Colombia, Mexico, and Morocco. In the case o Colombia they describe virtually all plants with at least ten workers over the period ; in Mexico they describe 2800 o the larger irms over the period , and in Morocco they cover nearly all irms with at least ten workers over the period Standard inormation on inputs, outputs, and costs is provided in each o these databases, as well as inormation on export levels. To simpliy estimation, we eliminate irms that do not report inormation or the entire sample period, creating a balanced panel. 10 The Appendix provides details on our data cleaning and delation procedures. 9. Although some countries report plant-level data and some report irm-level data, we will hereater use the term plant to describe the unit o observation. In semi-industrialized countries where the calculation is possible, we have ound that 95 percent o the plants are owned by single-plant irms. 10. Excepting the Moroccan apparel and textiles industries, irms that do not report inormation or the entire sample period are mostly very small nonexporters. Because these irms are either new or ailing, they probably have belowaverage perormance, so their elimination rom the sample probably improves the perormance o nonexporters (again, Moroccan apparel and textiles are an exception). This may have biased the estimated eect o past participation on the level o costs against the inding o learning eects, but given their small market share we do not think this bias is likely to be quantitatively large.

13 IS LEARNING BY EXPORTING IMPORTANT? 915 TABLE I ENTRY, EXIT, NUMBER OF PLANTS, AND EXPORT INTENSITY BY COUNTRY (EXPORT-ORIENTED INDUSTRIES) Average annual entry rate Average annual exit rate Average number o plants Average export intensity Colombia Morocco Mexico Finally, to sharpen the analysis, we ocus only on the exportoriented industries in each country. These industries exported at least 10 percent o their output, and had at least twenty exporting plants. 11 Table I provides descriptive inormation on each country. Note that there are substantial transitions in the data even though most plants tend to stay in or out o the export market. B. Comparing Productivity Trajectories We wish to amiliarize ourselves with the marginal cost trajectories o plants with dierent export market participation patterns, controlling or industrywide time eects, and observable plant-speciic productivity determinants like capital stocks and age. We use two marginal cost proxies: average variable cost (AVC) and labor productivity (LAB). The ormer is deined as the sum o real labor and real intermediate input costs divided by real output, and the latter is real output divided by number o workers. Real labor and intermediate costs are reported nominal values o these variables delated by the manuacturing-wide wholesale price index. Real output is the sum o nominal output or the domestic market and nominal output or export, each delated by its own product-speciic delator. To purge these productivity measures o industrywide time eects and observable plant-speciic characteristics, each is expressed in logarithms and regressed on time dummies, D jt (speciic to year t and the jth three-digit ISIC industry), age o the plant (A), age o the plant squared, capital stock o the plant (K), 11. In a ew cases, industries that exported less than 10 percent o their output were included because they had many exporting plants or accounted or a substantial share o total manuactured exports.

14 916 QUARTERLY JOURNAL OF ECONOMICS TABLE II TYPES OF FIRMS Nonexporters: Exporters: Entrants: Quitters: Switchers: and capital stock o the plant squared. Both age and capital stocks are measured in logarithms: J ln (AVC it ) j 1 J ln (LAB it ) j 1 irms that never exported during the sample period. irms that always exported during the sample period. irms that began the period as nonexporters, but began exporting during the sample period and never stopped. irms that began the period as exporters, but ceased during the sample period and never resumed exporting. irms that switched exporting status more than once during the sample period. T t 1 T t 1 jt D jt 1 ln (A it ) 2 ln (A it ) 2 3 ln (K it ) 4 (ln K it ) 2 c it jt L D jt 1 L ln (A it ) 2 L ln (A it ) 2 3 L ln (K it ) 4 L (ln K it ) 2 it L. The residuals rom these two regressions denoted ˆ c and ˆ L, respectively are then used as our indices o deviation rom timeand industry-speciic productivity norms. Note that industryspeciic time dummies control or industrywide variation in relative prices (inter alia). Also, because logarithmic variable costs are purged o correlation with capital, the residuals can be viewed as the measure o variable actor productivity that obtains rom a short-run total cost unction o the orm, TC it 0 (K it,w it ) 1 (K it,w it )Q it, where parentheses denote unctions, K it is capital, and W it is the vector o actor prices. To isolate the relation between export market participation and productivity perormance, we distinguish the ive varieties o plants deined in Table II, and or each plant we redeine period 0 to be the year in which a change in export status takes place. Then we isolate ive-year blocks o time, running rom two years prior to the status change (t 2) to two years ater (t 2). 12 For 12. The Colombian panel is long enough to allow us to go rom three years prior to three years ater the switch.

15 IS LEARNING BY EXPORTING IMPORTANT? 917 nonexporters and exporters there is no change in status, so we take ive years in the middle o the sample period (or Colombia we take seven years). Finally, ater reindexing time in this way, we aggregate our productivity indices by plant type and compare them. Because switching irms strongly resemble exporters, we omit them rom our graphs to reduce clutter. 13 C. Basic Trajectory Patterns Average Variable Costs: Average trajectories or average variable costs are presented by plant type in Figures IIIa through IIIc. We begin by considering Colombia and Mexico, since these countries show similar patterns. Most strikingly, plants that cease exporting get steadily worse beore they drop out o oreign markets, and are substantially less eicient than the other plant types. Also, entering plants and exporting plants have the lowest variable costs, and nonexporters consistently exhibit costs slightly above average, but less than quitters. These patterns are similar to the simulations in Figure II or both the learning and nonlearning models, except in that the perormance o quitting plants is somewhat worse in the actual data. One might argue about whether the entrants show evidence o reducing their costs ater beginning to export, but the eect is certainly not dramatic. The F-statistic or the null hypothesis that the mean average cost level among entrants shows no variation relative to the industry norms and has a p-value o 0.08 in Colombia and a p-value o 0.25 in Mexico. 14 Further, in Mexico, entrants average costs ater two years o exporting are slightly higher relative to industry norms than they were two years beore. The Moroccan graph is more complex. Most obviously, there is much less variation in trajectories here than in Colombia and Mexico. Further, what little variation there is does not conorm to the patterns described above. Two years ater exiting, the plants that abandon oreign markets emerge as the worst plants, but their average variable costs are highly volatile, and no higher 13. A possible explanation or the resemblance is that switching plants are always relatively eicient, but they experience relatively dramatic oreign demand shocks that drive them in and out o the export market. 14. These tests are based on generalizations o the regression model to include annual dummies by plant type. We also allow the coeicients on age, age squared, capital, and capital squared to vary across industry, and we treat the disturbance as composed o a ixed plant eect plus random noise.

16 918 QUARTERLY JOURNAL OF ECONOMICS FIGURE IIIa COLOMBIA Path o Average Variable Cost (purged o time, age, and size eects) FIGURE IIIb MEXICO Path o Average Variable Cost (purged o time and size eects) than the industry norm one year ater exiting. Exporters usually do better than nonexporters, but even this is not guaranteed. These patterns may well relect the act that, unlike in Mexico and Colombia, most o the impetus to become an exporter in Morocco came rom irm-speciic demand side shocks. Many

17 IS LEARNING BY EXPORTING IMPORTANT? 919 FIGURE IIIc MOROCCO Path o Average Variable Cost (purged o time, size, and age eects) Moroccan exporters are young plants that were ounded with the exclusive purpose o selling particular apparel and textile products abroad [Sullivan 1995; World Bank 1994; Roberts, Sullivan and Tybout 1995]. Moroccan policies during the sample period also provided various subsidies to exporters, and these may have allowed less eicient plants to compete. Once again, there is little statistical evidence that average costs among plants that begin exporting are anything but lat relative to industry norms (the p-value is 0.24). Labor productivity: For several reasons (discussed in Section IV below), average variable costs may pick up spurious correlation between exporting status and eiciency. Hence we also examine output per worker. Like average variable costs, labor productivity is not a true measure o total actor productivity, but ater it has been purged o correlation with capital stocks, it is conceptually closer. Figure IVa presents average values o this productivity measure by plant type or Colombia. The entrants once again perorm the best, and in contrast to their average cost patterns, they do seem to improve when they initiate oreign sales. 15 Ongoing exporters continue to show the next best perormance, 15. The F-statistic or the null hypothesis that average labor productivity among these plants does not change relative to industry norms has a p-value o Also see ootnote 13.

18 920 QUARTERLY JOURNAL OF ECONOMICS FIGURE IVa COLOMBIA Path o Average Labor Productivity (purged o time, age, and size eects) FIGURE IVb MEXICO Path o Average Labor Productivity (purged o time and size eects) and quitters continue to show the worst perormance, particularly around the time o their exit. So, in Colombia it appears that output per worker is one important source o variation in average variable costs, and we have the irst bit o evidence that exporting might improve perormance. Quitters and ongoing exporters in Mexico ollow the Colombian pattern. But Mexican entrants do not. Their average labor

19 IS LEARNING BY EXPORTING IMPORTANT? 921 FIGURE IVc MOROCCO Path o Average Labor Productivity (purged o time, size, and age eects) productivity is very stable relative to industry norms ( p- value 0.61), and a bit below average (Figure IVb). So there is no suggestion o a learning eect rom exporting in this country, nor, or that matter, does it appear that cross-plant exporting patterns can be traced to dierences in labor productivity. Finally, in terms o labor productivity, Moroccan patterns appear to resemble Colombia s (Figure IVc). Entrants productivity jumps in the year that they begin exporting (although the p-value is 0.41), and ongoing exporters exhibit higher labor productivity than either nonexporters or quitters. 16 Interestingly, there is much more variation across plant types in our Moroccan labor productivity series than we ound in our average cost series, suggesting that labor costs covary negatively with intermediate goods costs in this country. One undamental source o variation in labor productivity is the associated variation in the skill mix o employees. To see whether dierences in skill mixes are behind the labor productivity dierences, we analyzed the ratio o skilled to total labor in Colombia and Mexico. 17 The results or Colombia indicate that 16. The one exception is that labor productivity is high or exiting plants during their last year in oreign markets. 17. Graphs o this variable are not reported to conserve space. See Clerides, Lach, and Tybout [1996] or the igures tracing the path o labor quality or the dierent varieties o plants. No results are presented or Morocco because o lack o appropriate data.

20 922 QUARTERLY JOURNAL OF ECONOMICS entrants have high labor productivity partly because they use skilled labor relatively intensively, and quitters have low productivity because they use little skilled labor. There is also some evidence that relative skill-intensity increases over time or the Colombian entrants (the p-value is 0.09). The high labor productivity o ongoing exporters does not, however, appear to come rom unusually high skill intensity. One interpretation is that breaking into oreign markets requires new product design and other orms o technical assistance, but established export production can be routinized. In Mexico, skill intensity trajectories match labor productivity trajectories or ongoing exporting plants (which are skillintensive) and exiting plants (which are not). However, although Mexican entrants resemble Colombian entrants in terms o skill intensity, this is not suicient to get them high labor productivity. In sum, our perormance measures indicate that entrants generally do better than nonexporters and exiting plants. They have higher labor productivity, and this appears to be partly due to their heavier reliance on skilled labor. Also, despite their high quality workers, new exporters have relatively low average variable costs. On the other hand, we ind little to suggest that productivity gains ollow entry into oreign markets. Labor productivity and skill intensity appear to improve or Colombian plants that begin exporting, but this phenomenon did not show up in the other countries. 18 IV. AN ECONOMETRIC TEST OF LEARNING EFFECTS The igures we report in the previous section are revealing, but they do not constitute a direct test or whether past export market participation inluences current costs; the F-statistics we calculated merely indicate whether cost trajectories or entering 18. The results discussed thus ar are based on unweighted averages o our variable cost measures. But small plants are much more common than large plants so the igures do not necessarily describe sectorwide perormance. To determine whether aggregate perormances looked similar, we redid the calculations, weighting each plant s average cost by its share in the total output among those o its type. We also redid the plant-speciic average cost measures themselves, leaving capital stocks and age out o the regression, because we did not want them to be purged o correlation with size or this exercise. Overall, the pattern looks very similar to the unweighted ones, although with some exceptions. It remains true, however, that entrants exhibit relatively low average costs, both beore and ater the transition year, and that there is no obvious tendency or costs to all ater exporting operations are initiated. Complete results, including igures, are reported in the working paper version [Clerides, Lach, and Tybout, 1996].

21 IS LEARNING BY EXPORTING IMPORTANT? 923 irms deviate signiicantly rom industry norms. Hence in this section we develop an alternative approach to test or the presence o learning-by-exporting eects in the data. Speciically, we estimate equation (3) and test whether exporting history, y it 1, enters signiicantly in the cost equation. In other words, we perorm a type o Granger causality test. We obtain estimates o (3) in two alternative ways. First, we it (3) jointly with a reduced-orm version o (6) using ull inormation maximum likelihood (FIML). This approach allows us to characterize the sel-selection process, and to explore the relation between the error terms in the two equations. However, the participation equation (6) is nonlinear, dynamic, and has serially correlated errors, so FIML estimation is relatively complex, and the cost unction (3) must be kept airly simple. Thus, as a robustness check, we also it generalized versions o equation (3) using a generalized method o moments (GMM) estimator to deal with endogeneity and serial correlation. 19 A. The Econometric Model For FIML estimation we must develop an estimable version o the participation equation (6). Our approach ollows Roberts and Tybout [1997]. First, we generalize (6) so that irms that exit and reenter the export market pay dierent start-up costs than irms that never exported. Speciically, deine F 0 as the start-up cost or a nonexporter with no previous experience, and F j as the start-up cost aced by a irm that last exported j 1 years ago (note that F 1 0). Then this generalization o equation (6) implies that the ith irm will export in year t (i.e., will choose y it 1) whenever (7) (c it,z it ) M [E t (V it 1 0 y it 1) E t (V it 1 0 y it 0)] J F 0 j 1 (F 0 F j )ỹ it j, where ỹ it j is one i the irm last exported in year t j and zero otherwise. 20 Next, we deine the latent variable y* it as current net 19. I we allow or much generality in the average variable cost process, it becomes very diicult to estimate all o the structural parameters o the model. Yet imposing restrictions on these processes may lead us to incorrectly conclude that past participation inluences current costs. For example, misspeciying the cost unction to be a irst-order process may orce any dependence o current costs on additional lags (e.g., c t 2 or c t 3 ) to induce correlation with y t 1, given that export market participation decisions are based partly on lagged costs. 20. Note that ỹ it 1 y it 1, and or j 1, ỹ it j y it j j 1 k 1 (1 y it k ).

22 924 QUARTERLY JOURNAL OF ECONOMICS operating proits plus the expected uture return rom being an exporter in year t: y* it (c it,z it ) M [E t (V it 1 0 y it 1) E t (V it 1 0 y it 0)]. Equation (7) then implies that y it obeys the ollowing dynamic discrete process: 51 i y* it F 0 j 1 (F 0 F j )ỹ it j (8) y it 0 otherwise. Finally, we express the latent variable y* it F 0 as a reduced orm in demand shiters, marginal costs, and the variables that help predict uture marginal costs and demand in each market. 21 Operationally, this means including exogenous plant characteristics (X it ), the exchange rate (e t ), a distributed lag in our marginal cost proxy, AVC it, and a serially correlated disturbance: (9) y* it F 0 x X it e e t j 1 c j ln (AVC it j ) it. Then substituting (9) into (8), we obtain a representation o export market participation decisions that can be estimated: 51 i 0 xxit eet j 1 c j ln (AVC it j ) J (10) y it j 1 (F 0 F j )ỹ it j it 0 otherwise. Notice that the estimated coeicients on our lagged participation variables measure the discount on entry costs that plants with exporting experience in previous years enjoy, relative to plants with no exporting experience. Further discussion may be ound in Roberts and Tybout [1997]. As noted in Section II, i irms learn by exporting, the stochastic process that generates costs also depends upon the history o y it values. Accordingly, a general log-linearized speciication o the marginal cost unction (3) includes not only capital J J J 21. This reduced-orm approach is also used in Sullivan [1995] and Roberts, Sullivan, and Tybout [1995].

23 IS LEARNING BY EXPORTING IMPORTANT? 925 stocks and lagged average variable costs, but lagged y it values as well: J (11) ln (AVC it ) 0 j 1 k j ln (K it j) e ln (e t ) J C j 1 J Y j c ln (AVC it j) j 1 j y y it j v it. Here the real exchange rate (e) controls or changes in relative prices that are common to all plants, the distributed lags in K, AVC and y it capture systematic plant-speciic deviations rom these industrywide patterns, and v it represents unobserved idiosyncratic shocks. Together, equations (10) and (11) describe export market participation patterns and marginal cost realizations (learning externalities will be added to these equations later). Estimated as a system, industry by industry, they should reveal whether marginal costs inluence the export participation decision as the model implies, and whether irms typically experience cost reductions once they have begun to service oreign markets, as posited by the learning-by-exporting hypothesis. Tests on the cost coeicients in the participation equation indicate whether irms respond to cost reductions by becoming more likely to export, and tests on the y it j coeicients in the cost unction indicate whether exporting experience leads to lower costs. It is, o course, the latter direction o casuality that is o primary interest to us. Serial correlation is likely in both equations because persistent unobserved plant characteristics and demand conditions make some irms consistently low cost or consistently prone to exporting, conditioning on observable variables. Hence, we model the disturbances as composed o unobserved (random) plant eects, 1 and 2, plus transitory noise: it 1i 1it and v it 2i 2it. We allow the plant eects and the transitory noise to be correlated across equations. Also, without loss o generality, we impose the normalization var ( it ) 1 so that all coeicients in equation (10) are measured relative to total unexplained variation. Unortunately, the combination o lagged dependent variables with serial correlation creates special problems in panel data. Equation (10) and equation (11) represent the processes that generates y it and ln (AVC) it only or the years J 1 through T. But realizations on y i1 through y ij and on ln (AVC) i1 through ln (AVC) ij

24 926 QUARTERLY JOURNAL OF ECONOMICS are themselves endogenous, since they are unctions o the unobserved plant eects, 1 and 2. Treating them as orthogonal to the disturbances it and v it or t J would yield inconsistent estimates. We adopt Heckman s [1981a, 1981b] solution to this initial conditions problem. This amounts to adding J equations to the system that represent the probability o becoming an exporter in years 1 through J as unctions o 1i and the other variables in equation (10) except the lagged participation variables. Similarly, we also add J equations representing AVC i1 through AVC ij as unctions o 2i and the other variables in equation (11) except the lagged cost and participation variables. The result is a variant o Keane, Moitt, and Runkle s [1988] and Sullivan s [1997] estimator. 22 The system is estimated using ull inormation maximum likelihood (FIML), integrating out the two random eects with Gaussian quadrature. Details o the likelihood unction may be ound in Appendix II o Clerides, Lach, and Tybout [1996]. The FIML approach yields estimates o the cost unction (11) jointly with the participation equation (10). But the cost unction unlike the participation equation is linear, so it can also be estimated by itsel using the GMM estimator described by Holtz- Eakin, Newey, and Rosen [1988] and Arellano and Bond [1991]. This single-equation approach provides a robustness check in several senses. First, once we abandon our FIML estimator, it is easible to let average variable costs depend upon output levels. 23 Second, we can test whether learning is related to irms cumulative export volumes rather than simply to export participation in the recent past. Finally, i we are not trying to simultaneously identiy the parameters o the participation equation (10), our cost unction estimator demands less o the data in terms o length o time period and number o transitions in exporting status. So we can it the cost equation by itsel to more industries than we are able to treat with our system approach, and we can use more general speciications without overparameterizing the model. 22. Sullivan s [1995] estimator does not deal with initial conditions problems in the equation that has a continuous dependent variable; and Keane, Moit, and Runkle s [1988] estimator does not deal with dynamics at all. Otherwise, the structure o our estimator is the same. 23. To do so in the context o maximum likelihood, we would have needed to add an additional equation representing output choices, with a disturbance that was potentially correlated with those o the cost equation and the participation equation. The resulting system would have involved trivariate integration, and it would have been highly parameterized. For these reasons, we considered it an unattractive option.

25 IS LEARNING BY EXPORTING IMPORTANT? 927 Our generalized cost unction is (11 ) J K T ln (AVC it ) 0 j 1 k j ln (K it j ) 1 d D t J Q 1 q h ln (q t 1 M q t 1 ) m 1 B m m B i 1 A ln (A it ) 2 A ln (A it ) 2 j 1 J Y j 1 j y y it j v it. J C j c ln (AVC it j) This equation is distinguished rom (11) in that (a) the distributed lags on both capital stocks (K) and average variable costs are longer (J K J K,J C J C ); (b) it controls or relative prices using annual time dummies rather than the real exchange rate; and (c) it allows average variable costs to depend upon output levels, business type (B) and age (A). We will also estimate a variant o equation (11 ) in which the distributed lag in oreign market participation dummies, y it, is replaced with a distributed lag in logs o export volumes, ln (q it 1). B. The Evidence on Learning Because we expect the cost unction, the proit unction, entry costs, and the potential or active learning to dier across industries, we it our system o equations separately to each industry in which we have suicient observations to support inerence. We include industries that are relatively intensive in human capital (e.g., Chemicals) as well as those that are not (e.g., Apparel) in order to look or corresponding variation in active learning eects. Although the model can be it to our Mexican panel, the time period spanned by that data set proved too short to isolate random eects rom the lagged endogenous variables. 24 We 24. In Mexico we observe only ive years o data, and three o these years are lost to lags on participation and average costs, leaving two in-sample years. Estimates o the model (available upon request) attribute too much explanatory power to lagged cost and export participation, and imply that var ( 1 ) var ( 2 ) 0. Nonetheless, the results concerning the eect o lagged participation on cost realizations are consistent with those obtained in the other countries. The convergence o random eect Probit estimators to simple Probit estimators when T is small is discussed in Guilkey and Murphy [1993].