Exporting and Plant-Level Efficiency Gains: It’s in the Measure∗ Alvaro Garcia-Marin

Nico Voigtländer

Universidad de Chile

UCLA and NBER

First draft: April 2013 This draft: March 2017 Abstract While there is strong evidence for productivity-driven selection into exporting, the empirical literature has struggled to identify export-related efficiency gains within plants. Previous research typically derived revenue productivity (TFPR), which is downward biased if more efficient producers charge lower prices. Using a census panel of Chilean manufacturing plants, we compute plant-product level marginal cost as an efficiency measure that is not affected by output prices. For export entrant products, we find efficiency gains of 15-25%. Because markups remain relatively stable after export entry, most of these gains are passed on to customers in the form of lower prices, and are thus not reflected by TFPR. These results are confirmed when we use tariffs to predict export entry. We also document very similar results in Colombian and Mexican manufacturing plants. In addition, we find sizeable efficiency gains for tariff-induced export expansions of existing exporters. Only one quarter of these gains are reflected by TFPR, due to a partial rise in markups. Our results thus imply that within-plant gains from trade are substantially larger than previously documented. Evidence suggests that a complementarity between exporting and investment in technology is an important driver behind these gains. JEL: D24, F10, F14, L25, L60 Keywords: International Trade, Gains from Trade, Productivity, Markups ∗

We would like to thank Roberto Álvarez, Johannes van Biesebroeck, Joaquin Blaum, Ariel Burstein, Lorenzo Caliendo, Donald Davis, Jan De Loecker, Gilles Duranton, Eduardo E. Engel, Juan Carlos Hallak, Amit Khandelwal, Edward Leamer, Steve Redding, Ariell Reshef, Andrés Rodríguez-Clare, Veronica Rappoport, John Ries, Chad Syverson, Eric Verhoogen, and Romain Wacziarg, as well as seminar audiences at Chicago Booth, Copenhagen Business School, Columbia, CREI, Dartmouth, the FREIT Conference in Virginia, KU Leuven, Mannheim, Frankfurt, Pontificia Universidad Católica de Chile, Princeton, RMET Conference, TIGN Conference, UCLA, and Universidad de Chile for helpful comments and suggestions. Claudio Bravo-Ortega and Lucas Navarro kindly shared data to complement our Chilean plant-product panel. Henry Alexander Ballesteros and Oscar Fentanes provided excellent research assistance in Bogotá and Mexico City, respectively. We gratefully acknowledge research support from the UCLA Ziman Center’s Rosalinde and Arthur Gilbert Program in Real Estate, Finance and Urban Economics.

1 Introduction A large literature in empirical trade has shown that exporting firms and plants are more productive than their non-exporting counterparts. In principle, this pattern may emerge because exporters have higher productivity to start with, or because they become more efficient after export entry. The former effect – selection across plants – has received strong theoretical and empirical support (c.f. Melitz, 2003; Pavcnik, 2002). On the other hand, evidence for export-related within-plant productivity gains is much more sparse, with the majority of empirical studies finding no effects (for recent reviews of the literature see Syverson, 2011; Bernard, Jensen, Redding, and Schott, 2012). In particular, the productivity trajectory of plants or firms typically look flat around the time of export entry, suggesting that producers do not become more efficient after foreign sales begin.1 This is surprising, given that exporters can learn from international buyers and have access to larger markets to reap the benefits of innovation or investments in productive technology (Bustos, 2011). In other words, there is strong evidence for a complementarity between export expansions and technology upgrading (c.f. Lileeva and Trefler, 2010; Aw, Roberts, and Xu, 2011), and technology upgrading, in turn, should lead to observable efficiency increases. Why has the empirical literature struggled to identify such gains? In this paper, we use rich Chilean, Colombian, and Mexican data to show that flat productivity profiles after export expansions are an artefact of the measure: previous studies have typically used revenue-based productivity, which is affected by changes in prices. If cost savings due to gains in physical productivity are passed on to buyers in the form of lower prices, then revenue-based productivity will be downward biased (Foster, Haltiwanger, and Syverson, 2008).2 Consequently, accounting for pricing behavior (and thus markups) is key when analyzing efficiency trajectories. We show in a simple framework that under a set of non-restrictive assumptions (which hold in our data), marginal costs are directly (inversely) related to physical productivity, while revenue productivity reflects efficiency gains only if markups rise. We begin by using our main dataset – an unusually rich panel of Chilean manufacturing plants between 1996 and 2007 – to analyze the trajectories of marginal cost, markups, and prices around 1

Early contributions that find strong evidence for selection, but none for within-firm efficiency gains, include Clerides, Lach, and Tybout (1998) who use data for Colombian, Mexican, and Moroccan producers, and Bernard and Jensen (1999) who use U.S. data. Most later studies have confirmed this pattern. Among the few studies that document within-plant productivity gains are De Loecker (2007) and Lileeva and Trefler (2010). Further reviews of this ample literature are provided by Wagner (2007, 2012). 2 Recent evidence suggests that this downward bias also affects the link between trade and productivity. Smeets and Warzynski (2013) construct a firm level price index to deflate revenue productivity and show that this correction yields larger international trade premia in a panel of Danish manufacturers. Eslava, Haltiwanger, Kugler, and Kugler (2013) use a similar methodology to show that trade-induced reallocation effects across firms are also stronger for price-adjusted productivity.

1

export entry and export expansions. To derive markups at the plant-product level, we apply the method pioneered by De Loecker and Warzynski (2012), in combination with the uniquely detailed reporting of product-specific input cost shares by Chilean multi-product plants. In addition, our dataset comprises physical units as well as revenues for each plant-product, allowing us to calculate product prices (unit values). Dividing these by the corresponding markups yields marginal costs at the plant-product level (De Loecker, Goldberg, Khandelwal, and Pavcnik, 2016). This procedure is flexible with respect to the underlying price setting model and the functional form of the production function. Importantly, by disentangling the individual components, we directly observe the extent to which efficiency gains (lower marginal costs) are translated into higher revenue productivity (by raising markups), or passed on to customers (by reducing prices). To compare our results with the typically used efficiency measure, we also compute revenue productivity (TFPR) at the plant-product level. Figure 1 presents our main results – within plant-product trajectories for export entrants in Chile. Time on the horizontal axis is normalized so that zero represents the export entry year. The left panel confirms that, in line with most of the previous literature, the trajectory of TFPR is flat around export entry. The right panel disentangles this pattern and shows that (i) marginal costs within plant-products drop by approximately 15-25% during the first three years after export entry; (ii) prices fall by a similar magnitude as marginal costs; (iii) markups do not change significantly during the first years following export entry. Our findings suggest that export entrants do experience efficiency gains, but that these are passed on to their customers. In other words, constant markups and falling prices explain why revenue productivity is flat around export entry. Our results for export entrants are very similar when we use propensity score matching to construct a control group of plant-products that had an a-priori comparable likelihood of entering the export market. In addition, we show that we obtain very similar results when (i) computing physical productivity (TFPQ, which requires stronger assumptions than marginal costs at the plantproduct level, as discussed in Section 2.5), and (ii) when using reported average variable costs at the plant-product level. This suggests that our findings are not an artefact of the methodology used to calculate marginal costs; in fact, the computed marginal costs are strongly correlated with the reported average variable costs. We also discuss that our results are unlikely to be confounded by changes in product quality.3 We then exploit falling tariffs on Chilean products in destination 3

The bias that may result from changes in quality works against finding efficiency gains with our methodology: exported goods from developing countries are typically of higher quality than their domestically sold counterparts (c.f. Verhoogen, 2008) and use more expensive inputs in production (Kugler and Verhoogen, 2012). Thus, exporting should raise marginal costs. This is confirmed by Atkin, Khandelwal, and Osman (2014) who observe that quality upgrading of Egyptian rug exporters is accompanied by higher input prices. Using Mexican data, Iacovone and Javorcik (2012) provide evidence for quality upgrading right before, but not after, export entry.

2

countries to predict the timing of export entry. Due to the limited variation in tariffs, this exercise serves as a check, rather than the core of our analysis. Nevertheless, the combined variation in tariffs over time and across 4-digit sectors is sufficient to yield a strong first stage. We confirm our findings from within-plant trajectories: tariff-induced export entry is associated with marginal costs declining by approximately 25%. In relative terms, this corresponds to approximately onethird of the standard deviation in year-to-year changes in marginal costs across all plant-products in the sample. We provide evidence that technology upgrading is the most likely explanation for declining marginal costs at export entry. Plant-level investment (especially in machinery) spikes right after export entry. In addition, marginal costs drop particularly steeply for plants that are initially less productive. This is in line with Lileeva and Trefler (2010), who point out that, for the case of investment-exporting complementarity, plants that start off from lower productivity levels will only begin exporting if the associated expected productivity gains are large. In addition to export entry, we also analyze export expansions of existing exporters that are induced by falling export tariffs on Chilean products. Over our sample period, these tariff-induced export expansions lead to a decline in marginal costs by approximately 20% among existing exporters. Since export expansions are accompanied by investment in capital, technology upgrading is a likely driver of efficiency gains among existing exporters, as well. We also show that in the case of established exporters, pass-through of efficiency gains to customers is more limited than for new export entrants: about three quarters of the decline in marginal costs translate into lower prices, and the remainder, into higher markups. Consequently, TFPR also increases and reflects about one-fourth of the actual efficiency gains. Thus, while the downward bias of TFPR is less severe for established exporters, it still misses a substantial part of efficiency increases. Why are markups stable around export entry, but increase for established exporters after tariffinduced expansions? This pattern is compatible with a ‘demand accumulation process’ (Foster, Haltiwanger, and Syverson, 2016) – while existing exporters already have a customer base abroad, new entrants may use low prices to attract buyers.4 To support this interpretation, we separately analyze the domestic and export price of the same product in a subset of years with particularly detailed pricing information. We find that for export entrants, the export price drops more than its domestic counterpart (19% vs. 8%). There is also some evidence in our data that markups grow as export entrants become more established.5 4

Foster et al. (2016) provide evidence that supports this mechanism in the domestic market. They show that by selling more today, firms expand buyer-supplier relationships and therefore shift out their future demand. 5 There is a longer delay between export entry and changes in markups in our data as compared to De Loecker and Warzynski (2012), who document increasing markups right after export entry for Slovenian firms. However, our data confirm De Loecker and Warzynski’s cross-sectional finding that exporters charge higher markups.

3

Finally, we examine whether our main findings hold in two additional countries with detailed manufacturing panel data that are suited for our analysis: Colombia (2001-13) and Mexico (19942003). Both datasets have been used extensively in studies of international trade, and we show that they are representative of the stylized facts documented in the literature (c.f. Bernard and Jensen, 1999).6 We find strong evidence for our main results. As shown in Figure 2 for Colombia and in Figure 3 for Mexico, there is no relationship between TFPR and export entry. On the other hand, marginal costs decline strongly after export entry in both countries. Prices fall handin-hand with marginal costs, while markups are relatively stable.7 We also show that investment (especially in machinery and equipment) spikes after export entry in both samples. The fact that our main findings hold for exporting plants in three different countries strongly suggests that our main conclusion is broadly applicable: revenue-based productivity measures miss important exportrelated efficiency gains within manufacturing plants. Our findings relate to a substantial literature on gains from trade. Trade-induced competition can contribute to the reallocation of resources from less to more efficient producers. Bernard, Eaton, Jensen, and Kortum (2003) and Melitz (2003) introduce this reallocation mechanism in trade theory, based on firm-level heterogeneity. The empirical evidence on this mechanism is vast, and summarizing it would go beyond the scope of this paper.8 In contrast, the majority of papers studying productivity within firms or plants have found no or only weak evidence for exportrelated gains. Clerides et al. (1998, for Colombia, Mexico, and Morocco) and Bernard and Jensen (1999, using U.S. data) were the first to analyze the impact of exporting on plant efficiency. Both document no (or quantitatively small) empirical support for this effect, but strong evidence for selection of productive firms into exporting. The same is true for numerous papers that followed: Aw, Chung, and Roberts (2000) for Taiwan and Korea, Alvarez and López (2005) for Chile, and Luong (2013) for Chinese automobile producers.9 The survey article by ISGEP (2008) compiles micro level panels from 14 countries and finds nearly no evidence for within-plant productivity 6

One limitation is that – unlike the Chilean data – the Colombian and Mexican data do not provide product-specific variable costs. We therefore cannot exploit this information to derive product-specific markups and marginal costs in multi-product plants. Consequently, we restrict our analysis to the subset of single-product plants, where all inputs are clearly related to the (single) produced output. 7 We discuss the (quantitatively small) increase of markups after export entry in Colombia in Section 6. 8 Two influential early papers are Bernard and Jensen (1999) and Pavcnik (2002), who analyze U.S. and Chilean plants, respectively. Recent contributions have also drawn attention to the role of imports. Amiti and Konings (2007) show that access to intermediate inputs has stronger effects on productivity than enhanced competition due to lower final good tariffs. Goldberg, Khandelwal, Pavcnik, and Topalova (2010) provide evidence from Indian data that access to new input varieties is an important driver of trade-related productivity gains. 9 Alvarez and López (2005) use an earlier version of our Chilean plant panel. They conclude that "Permanent exporters are more productive than non-exporters, but this is attributable to initial productivity differences, not to productivity gains associated to exporting." [p.1395] We confirm this finding when using revenue-productivity.

4

increases after entry into the export market. The few papers that have found within-plant productivity gains typically analyzed periods of rapid trade liberalization, such as De Loecker (2007) for the case of Slovenia and Lileeva and Trefler (2010) for Canada, or demand shocks due to large (and permanent) exchange rate changes such as Park, Yang, Shi, and Jiang (2010).10 Our results illustrate why it may be more likely to identify within-plant gains in revenue productivity during periods of major tariff reductions: especially for established exporters, declining export tariffs have effects akin to a demand shock, which may lead to rising markups in general demand structures such as Melitz and Ottaviano (2008). Then, TFPR will rise because of its positive relationship with markups.11 The downward bias in TFPR can also be tackled by computing quantity productivity (TFPQ). In a paper that follows ours, Lamorgese, Linarello, and Warzynski (2014) document rising TFPQ for Chilean export entrants.12 Our findings are compatible with Caliendo, Mion, Opromolla, and Rossi-Hansberg (2015) who show that in response to productivity or demand shocks, firms may reorganize their production by adding a management layer. This causes TFPQ to rise, while TFPR falls because the increase in output quantity leads to lower prices. Relative to the existing literature, we make several contributions. To the best of our knowledge, this paper is the first to use marginal cost as a measure of efficiency that is not affected by the pricing behavior of exporters, and to document a strong decline in marginal costs after export entry and tariff-induced export expansions.13 Second, we discuss in detail the conditions under which declining marginal costs reflect gains in physical productivity. Third, we show that disentangling the trajectories of prices and physical productivity is crucial when analyzing export-related efficiency gains: it allows us to quantify the bias of the traditional revenue-based productivity measure. We find that TFPR misses almost all efficiency gains related to export entry, and a substantial share of the gains from tariff-induced export expansions. Consequently, we identify substantial exportrelated efficiency gains that have thus far passed under the radar. This also applies to the few studies that have found export related changes in TFPR within plants: our results suggest that the 10

Van Biesebroeck (2005) also documents efficiency gains after export entry – albeit in a less representative setting: among firms in sub-Saharan Africa. These gains are likely due to economies of scale, because exporting lifts credit constraints and thus allows sub-Saharan African firms to grow. 11 Potentially, markups could rise even if the actual efficiency is unchanged, causing an upward-bias of TFPR. However, our data suggest that changes in markups generally fall short of actual efficiency gains, so that altogether, TFPR is downward biased. 12 We discuss below that marginal costs have an advantage over TFPQ in the context of our study: For multi-product plants, product-level marginal costs can be computed under relatively unrestrictive assumptions. This allows us to analyze efficiency gains by decomposing prices into markups and marginal costs – all variables that naturally vary at the product level. Disentangling these components also has the advantage that we can analyze pass-through of efficiency gains. 13 De Loecker et al. (2016) document a fall in the marginal cost of Indian firms following a decline in input tariffs.

5

actual magnitude of efficiency gains is likely larger. Our study thus complements a substantial literature that argues that within-plant efficiency gains should be expected.14 Fourth, as a corollary contribution, our unique main (Chilean) dataset allows us to verify the methodology for computing marginal costs based on markups (De Loecker et al., 2016): we show that changes in computed plant-product level marginal costs are very similar to those in self-reported average variable costs. Finally, by confirming that our results hold for two additional countries (Colombia and Mexico), we provide strong support for their general validity. The rest of the paper is organized as follows. Section 2 discusses our use of marginal cost as a measure of efficiency and its relationship to revenue productivity; it also illustrates the empirical framework to identify the two measures. Section 3 describes our datasets. Section 4 presents our empirical results for Chilean export entrants and Section 5, for continuing exporters. Section 6 provides evidence for Colombian and Mexican export entrants. Finally, Section 7 discusses our results and draws conclusions.

2 Empirical Framework In this section, we discuss our efficiency measures and explain how we compute them. Our first measure of efficiency is revenue-based total factor productivity (TFPR) – the standard efficiency measure in the literature that analyzes productivity gains from exporting. We discuss why this measure may fail to detect such gains, and show how we calculate TFPR at the plant-product level. Our second measure of efficiency is the marginal cost of production, which can be derived at the plant-product level under a set of non-restrictive assumptions. We also discuss the relationship between the two measures, and under which conditions marginal costs reflect physical productivity. 2.1 Revenue vs. Physical Total Factor Productivity Revenue-based total factor productivity is the most widely used measure of efficiency. It is calculated as the residual between total revenues and the estimated contribution of production factors (labor, capital, and material inputs).15 ++ To illustrate this concept, let’s consider a a standard Hicks-neutral production function relating physical output (Qit ) of a given plant i in period t, to a vector of inputs Xit . Denoting with lowercases the logarithms of the variables, then a log-linear 14

Case studies typically suggest strong export-related efficiency gains within plants. For example, Rhee, RossLarson, and Pursell (1984) surveyed 112 Korean exporters, out of which 40% reported to have learned from buyers in the form of personal interactions, knowledge transfer, or product specifications and quality control. The importance of knowledge transfer from foreign buyers to exporters is also highlighted by the World Bank (1993) and Evenson and Westphal (1995). López (2005) summarizes further case study evidence that points to learning-by-exporting via foreign assistance on product design, factory layout, assembly machinery, etc. 15 Some authors have used labor productivity – i.e., revenues per worker – as a proxy for efficiency. This measure is affected by the use of non-labor inputs and is thus inferior to TFPR (c.f. Syverson, 2011).

6

representation of the production function is given by qit = x′it α +ωit , where ωit is physical productivity (TFPQ).16 When physical output is unobserved, researchers typically rely on plants’ revenues (Rit ) to proxy for physical output, in which case: rit = x′it α + (ωit + pit ) | {z }

(1)

ω bit

ωit ). When revenues Equation (1) highlights an important shortcoming of revenue productivity (b are used as output variable, the residual term ω bit reflects both prices (pit ) and physical productivity (ωit ): ω bit = pit + ωit . Thus, if output prices respond to a producer’s efficiency, TFPR is biased. For

example, when facing downward-sloping demand, firms typically respond to efficiency gains by expanding production and reducing prices. This generates a negative correlation between pit and ωit , so that TFPR will typically underestimate physical productivity.17 Empirical studies attempt

to address this bias by deflating revenues with industry price indexes when computing TFPR. However, the price bias persists within industries, reflecting the difference between individual plants’ prices and the corresponding industry price index. It is important to note that TFPR is not always inferior to TFPQ (or marginal costs); instead, the applicability of the different measures depends on the context. For example, when analyzing misallocation as in Hsieh and Klenow (2009), TFPR is the more appropriate measure. In this framework, with downward-sloping iso-elastic demand and CRS technology, high-TFPQ firms charge lower prices that exactly offset their TFPQ advantage, equalizing TFPR. This provides a useful benchmark: in the absence of distortions, TFPR should be the same across plants in an industry, even if their TFPQ differs. At the same time, the Hsieh-Klenow framework also illustrates the shortcomings of TFPR: in the absence of distortions, plants with higher TFPQ are larger and make higher aggregate profits – these differences are not reflected by TFPR.18 Despite the shortcomings of TFPR, the majority of studies have used this measure to analyze productivity gains from exporting. One practical reason is the lack of information on physical quantities.19 While some corrections to the estimation of production functions have been pro16

++In the discussion that follows, we assume that input quantities are observed by the researcher. When researchers only observe input expenditure, the productivity estimation is also subject to an additional source of bias, the so-called input price bias. We discuss this in more detail below, and in appendix A.3. For more detail, see ?.++ 17 ++In this discussion we assume that the production function coefficients (α) are correctly estimated. However, this is not necessarily the case when physical output is unobserved. Establishments charging higher prices, ceteris paribus, are expected to sell less units of output and use less inputs. Then, pit and xit are more likely to be negatively correlated, leading to a downward bias in the coefficients of the production function and in the returns to scale when they are estimated through OLS (see ?, for more details).++ 18 Foster, Grim, Haltiwanger, and Wolf (2016) point to limitations of the Hsieh-Klenow framework. In particular, they show that under deviations from CRS, the variation in TFPR is also affected by shocks to demand and TFPQ. 19 Data on physical quantities have only recently become available for some countries (c.f. De Loecker et al., 2016;

7

posed, only a few studies have derived ωit directly.20 To circumvent some of the issues related to computing ωit , we propose marginal costs as our main measure of efficiency. Next, we discuss under which conditions declining marginal costs reflect efficiency gains. 2.2 Marginal Cost as a Measure of Efficiency, and its Relationship to TFPR In standard production functions, marginal costs are inversely related to efficiency (physical productivity) A. To illustrate this relationship, we use the generic functional form mc(ωit , wit ), where wit is an input price index, and the subscripts i and t denote plants and years, respectively (for ease of exposition, we do not further differentiate products within plants for now). The derivatives with respect to the two arguments are mc1 < 0 and mc2 > 0. Next, we can use the fact that prices are the product of markups (µit ) and marginal costs to disentangle TFPR (assuming Hicks-neutrality – as is standard in the estimation of productivity): ω bit = pit + ωit = µit + mc(ωit , wit ) + ωit

(2)

△ωit = △b ωit − △µit − △mc(ωit , wit )

(3)

Deriving log-changes (denoted by △) and re-arranging yields a relationship between efficiency gains and changes in TFPR, markups, and marginal costs:

In order to simplify the interpretation of (3) – but not in the actual estimation of mc(·) – we make two assumptions. First, that the underlying production function exhibits constant returns to scale (CRS). This assumption is supported by our data, where the average sum of input shares is very close to one (see Table A.5 in the appendix). This first assumption implies that we can separate △mc(ωit , wit ) = △φ(wit ) − △ωit , where φ(·) is an increasing function of input prices (see the proof in Appendix A.1). Second, we assume that input prices are unaffected by export entry or expansions, i.e., they are constant conditional on controlling for trends and other correlates around the time of export entry: △φ(wit ) = 0. Our dataset allows us to calculate input prices, and we show below in Section 4.5 that these do not change with exporting activity. Kugler and Verhoogen, 2012, for India and Colombia, respectively). 20 Melitz (2000) and De Loecker (2011) discuss corrections to the estimation of the production function to account for cross-sectional price heterogeneity in the context of a CES demand function. Gorodnichenko (2012) proposes an alternative procedure for estimating the production function that models the cost and revenue functions simultaneously, accounting for unobserved heterogeneity in productivity and factor prices. Hsieh and Klenow (2009) recover ωit using a model of monopolistic competition for India, China, and the United States. Foster et al. (2008) obtain ωit using product-level information on physical quantities from U.S. census data for a subset of manufacturing plants that produce homogeneous products. Finally, Eslava et al. (2013) and Lamorgese et al. (2014) compute TFPQ and use it to analyze gains from trade.

8

With constant input prices, we obtain three simple expressions that illustrate the relationship between physical efficiency gains and changes in marginal costs, markups, and TFPR: 1. △ωit = −△mcit , i.e., rising efficiency is fully reflected by declining marginal costs. Note that this is independent of the behavior of markups. Using this equality in (3) also implies: 2. △b ωit = △µit , i.e., revenue productivity rises if and only if markups increase. For example, even if ωit rises (and mcit falls), TFPR will not grow if markups remain unchanged. And vice-versa, if markups rise while ωit stays the same, TFPR will increase. This underlines the shortcomings of TFPR as a measure of efficiency – it can both fail to identify actual efficiency gains but may also reflect spurious gains due to demand-induced increases in markups. 3. △b ωit = △ωit if △µit = −△mcit , i.e., changes in revenue productivity reflect the full efficiency gains if markups rise in the same proportion as marginal costs fall, i.e., if the output price remains constant and pass-through of efficiency gains is zero. We use these insights when interpreting our empirical results below. For young exporters, the evidence points towards constant markups. Thus, all efficiency gains are passed on to customers, so that they are reflected only in marginal costs, but not in TFPR. For more mature exporters there is some evidence for declining marginal costs together with rising markups, meaning that at least a part of the efficiency gains is also reflected in TFPR. 2.3 Estimating Revenue Productivity (TFPR) To compute TFPR, we first have to estimate the revenue production function. We specify a CobbDouglas production function with labor (l), capital (k), and materials (m) as production inputs. We opt for the widely used Cobb-Douglas specification as our baseline because it allows us to use the same production function estimates to derive TFPR and markups (and thus marginal costs). This ensures that differences in the efficiency measures are not driven by different parameter estimates.21 Following De Loecker et al. (2016), we estimate a separate production function for each 2-digit manufacturing sector (s), using the subsample of single product plants.22 The reason for using single-product plants is that one typically does not observe how inputs are allocated to individual outputs within multi-product plants. For the set of single product plants, no assumption on the 21

As discussed below, TFPR needs to be estimated based on output measured in terms of revenues, while deriving markups based on revenues (rather than quantities) can lead to biased results. In our baseline Cobb-Douglas case, this bias does not affect our results because production function coefficients are constant and are therefore absorbed by plant-product fixed effects. Consequently, the Cobb-Douglas specification allows us to use the same production function coefficients to estimate both TFPR and markups (and thus marginal costs). In Appendix C.1 we show that the more flexible translog specification (where fixed effects do not absorb the bias) confirms our baseline results. 22 The 2-digit product categories are: Food and Beverages, Textiles, Apparel, Wood, Paper, Chemicals, Plastic, Non-Metallic Manufactures, Basic and Fabricated Metals, and Machinery and Equipment.

9

allocation of inputs to outputs is needed, and we can estimate the following production function with standard plant-level information: s qit = βls lit + βks kit + βm mit + ωit + εit

(4)

where all lowercase variables are in logs; qit are revenues of single-product plant i in year t, ωit is TFPR, kit denotes the capital stock, mit are material inputs, and εit represents measurement error as well as unanticipated shocks to output. We deflate all nominal variables (revenues, materials, wages) using 4-digit industry specific deflators provided by ENIA.23 Estimating (4) yields the s sector-specific vector of coefficients β s = {βls , βks , βm }. When estimating (4) we follow the methodology by Ackerberg, Caves, and Frazer (2015, henceforth ACF), who extend the framework of Olley and Pakes (1996, henceforth OP) and Levin-

sohn and Petrin (2003, henceforth LP). This methodology controls for the simultaneity bias that arises because input demand and unobserved productivity are positively correlated.24 The key insight of ACF lies in their identification of the labor elasticity, which they show is in most cases unidentified by the two-step procedure of OP and LP.25 We modify the canonical ACF procedure by specifying an endogenous productivity process that can be affected by export status and plant investment. In addition, we include interactions between export status and investment in the productivity process. Thus, the procedure allows exporting to affect current productivity either directly, or through a complementarity with investment in physical capital. This reflects the corrections suggested by De Loecker (2013); if productivity gains from exporting also lead to more investment (and thus a higher capital stock), the standard method would overestimate the capital coefficient in the production function, and thus underestimate productivity (i.e., the residual). Finally, using the set of single-product plants may introduce selection bias because plant switching from single- to 23

++AG: new, check // Checked. ++To keep our baseline estimation comparable to previous studies, we do not deflate material inputs by plant-specific deflators from the Chilean ENIA (which are typically not available). This gives rise to a potential (well-documented) input price bias (see De Loecker et al., 2016). Nevertheless, our baseline estimates with a Cobb-Douglas production function are immune to this bias (see footnote 21). In addition, in Appendix we show alternative results where we: (i) proxy for input prices using output prices and market share as suggested by De Loecker et al. (2016), and (ii) use plant-specific input price deflators to deflate materials’ expenditure. Results in both cases are very similar to our baseline case with revenue production function (see Appendix ). Production function coefficients are comparable to the baseline coefficients, although returns to scale are slightly higher – about 2.5% higher in the case where both input and output price biases are addressed –.++ 24 We follow LP in using material inputs to control for the correlation between input levels and unobserved productivity. 25 The main technical difference is the timing of the choice of labor. While in OP and LP, labor is fully adjustable and chosen in t, ACF assume that labor is chosen at t − b (0 < b < 1), after capital is known in t − 1, but before materials are chosen in t. In this setup, the choice of labor is unaffected by unobserved productivity shocks between t − b and t, but a plant’s use of materials now depends on capital, productivity, and labor. In contrast to the OP and LP method, this implies that the coefficients of capital, materials, and labor are all estimated in the second stage.

10

multi-product may be correlated with productivity. Following De Loecker et al. (2016), we correct for this source of bias by including the predicted probability of remaining single-product, sˆit , in the productivity process as a proxy for the productivity switching threshold.26 Accordingly, the law of motion for productivity is: ωit = g(ωit−1 , dxit−1 , diit−1 , dxit−1 × diit−1 , sˆit−1 ) + ξit

(5)

where dxit is an export dummy, and diit is a dummy for periods in which a plant invests in physical capital (following De Loecker, 2013). In the first stage of the ACF routine, a consistent estimate of expected output φˆt (·) is obtained from the regression qit = φt (lit , kit , mit ; xit ) + εit We use inverse material demand ht (·) to proxy for unobserved productivity, so that expected output s is structurally represented by φt (·) = βls lit + βks kit + βm mit + ht (mit , lit , kit , xit ).27 The vector xit contains other variables that affect material demand (time and product dummies, reflecting aggregate shocks and specific demand components). Next, we use the estimate of expected output together with an initial guess for the coefficient vector β s to compute productivity: for any can  s s s s s ˜ ˜ ˆ ˜ ˜ didate coefficient vector β , productivity is given by ωit (β ) = φt − βl lit + βk kit + β˜m mit . ˜ s : following Finally, we recover the productivity innovation ξit for the given candidate vector β ˜ s ) non-parametrically as a function of its own lag (5), we estimate the productivity process ωit (β ˜ s ), prior exporting and investment status, and the plant-specific probability of remaining ωit−1 (β single-product.28 The residual is ξit . The second stage of the ACF routine uses moment conditions on ξit to iterate over candidate ˜ s . In this stage, all coefficients of the production function are identified through GMM vectors β using the moment conditions E (ξit (β s )Zit ) = 0

(6)

where Zit is a vector of variables that comprises lags of all the variables in the production function, 26

We estimate this probability for single-product plants within each 2-digit sector using a probit model, where the explanatory variables include product fixed effects, labor, capital, material, output price, as well as importing and exporting status. 27 We approximate the function φˆt (·) with a full second-degree polynomial in capital, labor, and materials. 28 Following Levinsohn and Petrin (2003), we approximate the law of motion for productivity (the function g(·) stated in (5)) with a polynomial.

11

as well as the current capital stock. These variables are valid instruments – including capital, which is chosen before the productivity innovation is observed. Equation (6) thus says that for the optimal β s , the productivity innovation is uncorrelated with the instruments Zit . Given the estimated coefficients for each product category s (the vector β s ), TFPR can be calculated both at the plant level and for individual products within plants. For the former, we use the plant-level aggregate labor lit , capital kit , and material inputs mit . We then compute plant-level TFPR, ω ˆ it : s ω ˆ it = qit − (βls lit + βks kit + βm mit )

(7)

where qit are total plant revenues, and the term in parentheses represents the estimated contribution of the production factors to total output in plant i. Note that the estimated production function s allows for returns to scale (βls + βks + βm 6= 1), so that the residual ω ˆ it is not affected by increasing or decreasing returns. When computing plant-level TFPR in multi-product plants, we use the vector of coefficients β s that corresponds to the product category s of the predominant product produced by plant i. Next, we compute our main revenue-based productivity measure – product-level TFPR. To perform this step for multi-product plants, the individual inputs need to be assigned to each product j. Here, our sample provides a unique feature: ENIA reports total variable costs (i.e., for labor and materials) T V Cijt for each product j produced by plant i. We can thus derive the following proxy for product-specific material inputs, assuming that total material is used (approximately) in proportion to the variable cost shares: Mijt = sTijtV C · Mit

where

T V Cijt sTijtV C = P j T V Cijt

(8)

Taking logs, we obtain mijt . We use the same calculation to proxy for lijt and kijt . Given these values, we can derive plant-product level TFPR, using the vector β s that corresponds to product j: s ω ˆ ijt = qijt − (βls lijt + βks kijt + βm mijt )

(9)

where qijt are product-specific (log) revenues. 2.4 Estimating Marginal Cost To construct a measure of marginal production cost, we follow a two-step process. First, we derive the product-level markup for each plant. Second, we divide plant-product output prices (observed in the data) by the calculated markup to obtain marginal cost. The methodology for deriving markups follows the production approach proposed by Hall 12

(1986), recently revisited by De Loecker and Warzynski (2012). This approach computes markups without relying on market-level demand information. The main assumptions are that at least one input is fully flexible and that plants minimize costs for each product j. The first order condition of a plant-product’s cost minimization problem with respect to the flexible input V can be rearranged to obtain the markup of product j produced by plant i at time t:29 Pijt = µijt ≡ |{z} MCijt

Markup

!  V · Vijt ∂Qijt (·) Vijt . Pijt , ∂Vijt Qijt Pijt · Qijt | {z } | {z } 

Output Elasticity

(10)

Expenditure Share

where P (P V ) denotes the price of output Q (input V ), and MC is marginal cost. According to equation (10), the markup can be computed by dividing the output elasticity of product j (with respect to the flexible input) by the expenditure share of the flexible input (relative to the sales of product j). Note that under perfect competition, the output elasticity equals the expenditure share, so that the markup is one (i.e., price equals marginal costs). In our computation of (10) we use materials (M) as the flexible input to compute the output elasticity – based on our estimates of (4).30 Note that in our baseline estimation (due to its use of a Cobb-Douglas production function), the output elasticity with respect to material inputs is given s s by the constant term βm . Ideally, βm should be estimated using physical quantities for inputs and output in (4). However, as discussed above, this would render our results for TFPR and marginal

cost less comparable, since differences could emerge due to the different parameter estimates. s The Cobb-Douglas case allows us to compute markups based on revenue-based estimates of βm , without introducing bias in our within-plant/product analysis (see Section 2.5 for detail). Thus, our baseline results use the same elasticity estimates to compute both TFPR and markups. The second component needed in (10) – the expenditure share for material inputs – is directly observed in our data in the case of single-product plants. For multi-product plants, we use the V · Vijt = Mijt . Since proxy described in equation (8) to obtain the value of material inputs Pijt total product-specific revenues Pijt · Qijt are reported in our data, we can then compute the plantproduct specific expenditure shares needed in (10).31 This procedure yields plant-product-year 29

Note that the derivation of equation (10) essentially considers multi-product plants as a collection of singleproduct producers, each of whom minimizes costs. This setup does not allow for economies of scope in production. To address this concern, we show below that all our results also hold for single-product plants. 30 In principle, labor could be used as an alternative. However, in the case of Chile, labor being a flexible input would be a strong assumption due to its regulated labor market. A discussion of the evolution of job security and firing cost in Chile can be found in Montenegro and Pagés (2004). 31 By using each product’s reported variable cost shares to proxy for product-specific material costs, we avoid shortcomings of a prominent earlier approach: since product-specific cost shares were not available in their dataset, Foster et al. (2008) had to assume that plants allocate their inputs proportionately to the share of each product in total

13

specific markups µijt . Finally, because output prices (unit values) Pijt are also observed at the plant-product-year level, we can derive marginal costs at the same detail, MCijt . To avoid that extreme values drive our results, we only use observations within the percentiles 2 and 98 of the markup distribution. The remaining markup observations vary between (approximately) 0.4 and 5.6. In Table A.10 we show the average and median markup by sector. 2.5 Marginal Cost vs TFPQ In the following, we briefly discuss the advantages and limitations of marginal cost as compared to quantity productivity (TFPQ) as a measure of efficiency in the context of our study. For now, suppose that the corresponding quantity-based input elasticities β s have been estimated correctly. Then, in order to back out TFPQ by using (7), ideally both output and inputs need to be observed in physical quantities. Output quantities are available in some datasets. But for inputs, this information is typically unavailable.32 Thus, researchers have adopted the standard practice of using industry-level price indexes to deflate input expenditures (Foster et al., 2008). This approximation may lead to biased TFPQ estimates if input prices or the user cost of capital vary across firms within the same industry. A further complication arises if one aims to compute product-specific TFPQ for multi-product plants, where physical inputs need to be assigned to individual products. While our dataset has the unique advantage that plants report the expenditure share of each product in total variable costs (which is sufficient to derive the product-specific material expenditure share needed in (10) to compute markups), it does not contain information on how to assign input quantities to individual products. Thus, assigning mit , lit , and kit to individual products is prone to error. This is especially true in the case of capital, which is typically not specific to individual output products. In light of these limitations, most studies compute TFPQ at the plant or firm level.33 An additional complication arises for kit in TFPQ calculations because the capital stock is only available in terms of monetary values and not in physical units. Contrast this with the computation of markups in (10), still assuming that β s has been correctly s estimated. The output elasticity with respect to material inputs is given by βm , and – for singleproduct plants – the expenditure share for material inputs is readily available in the data. For revenues. This is problematic because differential changes in markups across different products will affect revenue shares even if cost shares are unchanged. De Loecker et al. (2016) address this issue by using an elaborate estimation technique to identify product-specific material costs; this is not necessary in our setting because the uniquely detailed Chilean data allow us to directly compute product-specific material costs from reported data. 32 Exceptions, where input quantities are available, include Ornaghi (2006), Davis, Grim, and Haltiwanger (2008), and Lamorgese et al. (2014). 33 A shortcoming of this more aggregate approach is that plant-level output price indexes do not account for differences in product scope (Hottman, Redding, and Weinstein, 2016).

14

multi-product plants, we use the approximation with reported variable cost shares in equation (8) to back out plant-product specific input expenditure shares. Thus, plant-product specific markups can be immediately calculated in our Chilean data.34 We now turn to the estimation of β s , which is challenging and may introduce further error. When using a Cobb-Douglas production function, this issue is less severe for markups than for s TFPQ in the context of our analysis. The computation of markups uses only βm from the vector s β s . Note that measurement error of βm will affect the estimated level of markups, but not our s within-plant results: because we analyze log-changes at the plant-product level, ln(βm ) cancels

out. In other words, the estimated log-changes in markups in (10) are only driven by the observed s 35 material expenditure shares, but not by the estimated output elasticity βm . Contrast this with the computation of TFPQ, which uses all coefficients in β s , multiplying each by the corresponding physical input (or deflated input expenditures) in (7). In this case, analyzing log-changes in TFPQ will not eliminate errors and biases in the level of β s . We discuss further issues related to marginal cost and TFPQ in the appendix. Appendix A.2 discusses the implications of deviations from CRS. We show that in the presence of increasing returns, marginal costs will tend to overestimate actual efficiency gains. In this case, TFPQ is the preferable efficiency measure (subject to the concerns discussed above), since its estimation allows for flexible returns to scale. Throughout the empirical sections, we thus present results based on TFPQ as a robustness check. Appendix ?? discusses the estimation of quantity-based production functions, and Appendix A.4 shows that marginal costs and TFPQ are equally affected by investment in new technology (even if only TFPQ directly takes the capital stock into account).

3 Data Our primary dataset is a Chilean plant panel for the period 1996-2007, the Encuesta Nacional Industrial Anual (Annual National Industrial Survey – ENIA). In addition, we confirm our main results using plant-level panel data from Colombia (for the period 2001-2013) and from Mexico (for 1994-2003). A key advantage of the Chilean data is that multi-product plants are required to report product-specific total variable costs. These are crucial for the calculation of plant-product level markups and marginal costs in multi-product plants, as described in Section 2.4. In the Colombian and Mexican samples, this information is not available. In order to keep the method34

Note that when computing product-level markups for multi-product plants, we only need to proportionately assign the expenditure share of material inputs to individual products. This procedure is not needed for labor or capital. 35 This is also the reason why we can use estimates of β s from the revenue production function, i.e., the same s coefficients used to compute TFPR. Note that for the more flexible translog specification, βm itself depends on the use of inputs by each plant and may thus vary over time. We show in Appendix C.1 that our results are nevertheless robust to this specification.

15

ology consistent, we thus restrict attention to single-product plants in these countries, where all inputs are clearly related to the single output. Correspondingly, the Chilean ENIA is our main dataset, and we describe it in detail below. The Colombian and Mexican datasets are described in Appendix B.3 and B.4, and we compare the three datasets in Appendix B.6. Overall, the sectoral composition of the three datasets is similar, but export orientation is markedly stronger for Mexican manufacturing firms, where almost 40% of all plants are exporters, as compared to 20% and 25% in the Chilean and Colombian samples, respectively. Data for ENIA are collected annually by the Chilean Instituto National de Estadísticas (National Institute of Statistics – INE). ENIA covers the universe of manufacturing plants with 10 or more workers. It contains detailed information on plant characteristics, such as sales, spending on inputs and raw materials, employment, wages, investment, and export status. ENIA contains information for approximately 5,000 manufacturing plants per year with unique identifiers. Out of these, about 20% are exporters, and roughly 70% of exporters are multi-product plants. Within the latter (i.e., conditional on at least one product being exported), exported goods account for 80% of revenues. Therefore, the majority of production in internationally active multi-product plants is related to exported goods. Finally, approximately two third of the plants in ENIA are small (less than 50 workers), while medium-sized (50-150 workers) and large (more than 150 workers) plants represent 20 and 12 percent, respectively. In addition to aggregate plant data, ENIA provides rich information for every good produced by each plant, reporting the value of sales, its total variable cost of production, and the number of units produced and sold. Products are defined according to an ENIA-specific classification of products, the Clasificador Unico de Productos (CUP). This product category is comparable to the 7-digit ISIC code.36 The CUP categories identify 2,242 different products in the sample. These products – in combination with each plant producing them – form our main unit of analysis. 3.1 Sample Selection and Data Consistency In order to ensure consistent plant-product categories in our ENIA panel, we follow three steps. First, we exclude plant-product-year observations that have zero values for total employment, demand for raw materials, sales, or product quantities. Second, whenever our analysis involves quantities of production, we have to carefully account for possible changes in the unit of measurement. For example, wine producers change in some instances from "bottles" to "liters." Total revenue is generally unaffected by these changes, but the derived unit values (prices) have to be corrected. This procedure is needed for about 1% of all plant-product observations; it is explained 36

For example, the wine industry (ISIC 3132) is disaggregated by CUP into 8 different categories, such as "Sparkling wine of fresh grapes," "Cider," "Chicha," and "Mosto."

16

in Appendix B.1. Third, a similar correction is needed because in 2001, ENIA changed the product identifier from CUP to the Central Product Classification (CPC V.1) code. We use a correspondence provided by the Chilean Statistical Institute to match the new product categories to the old ones (see Appendix B.1 for detail). After these adjustments, our sample consists of 118,178 plantproduct-year observations. 3.2 Definition of Export Entry The time of entry into export markets is crucial for our analysis. We impose four conditions for product j, produced by plant i, to be classified as an export entrant in year t: (i) product j is exported for the first time at t in our sample, which avoids that dynamic efficiency gains from previous export experience drive our results, (ii) product j is sold domestically for at least one period before entry into the export market, i.e., we exclude new products that are exported right away, (iii) product j continues to be reported in ENIA for at least two years after export entry, which ensures that we can compute meaningful trajectories, and (iv) product j is the first product exported by plant i. The last requirement is only needed for multi-product plants. It rules out that spillovers from other, previously exported products affect our estimates. Under this definition we find 861 export entries in our ENIA sample (plant-products at the 7-digit level), and approximately 7% of active exporters are new entrants. For our auxiliary Colombian and Mexican data, the construction of export entry is described in detail in Appendix B.5. 3.3 Validity of the Sample Before turning to our empirical results, we check whether our data replicate some well-documented systematic differences between exporters and non-exporters. Following Bernard and Jensen (1999), we run the regression ln(yist ) = αst + δ dexp ist + γ ln(List ) + εist ,

(11)

where yist denotes several characteristics of plant i in sector s and period t, dexp ist is an exporter dummy, List is total plant-level employment, and αst denotes sector-year fixed effects.37 The coefficient δ reports the exporter premium – the percentage-point difference of the dependent variable between exporters and non-exporters. Table 1 reports exporter premia for our main dataset – the Chilean ENIA. We find similar results for both unconditional exporter premia (Panel A) and when controlling for plant-level employment (Panel B): within their respective sectors, exporting plants are larger both in terms of employment and sales, are more productive (measured by revenue productivity), and pay higher wages. This is in line with the exporter characteristics documented by 37

Whenever we use plant-level regressions, we control for sector-year effects at the 2-digit level. When using the more detailed plant-product data, we include a more restrictive set of 4-digit sector-year dummies.

17

Bernard and Jensen (1999) for the United States, Bernard and Wagner (1997) for Germany, and De Loecker (2007) for Slovenia, among others. Using product-level data in column 5, we also find that markups are higher among exporters, confirming the findings in De Loecker and Warzynski (2012). Our Colombian and Mexican data show very similar patterns (see Appendix B.3 and B.4).

4 Efficiency Gains of Export Entrants in Chilean Manufacturing In this section we present our empirical results for new export entrants in Chile. We show the trajectories of revenue productivity, marginal costs, and markups within plant-products around the time of export entry. We verify that our results hold when we use propensity score matching to construct a reference group for export entrants, and when we use tariff changes to predict export entries. We also provide suggestive evidence that the observed efficiency gains are driven by a complementarity between exporting and investment. 4.1 New Export Entrants: Plant-Product Trajectories To analyze trajectories of various plant-product characteristics, we estimate the following regression for each plant i producing product j in period t: yijt = αst + αij +

−1 X

k + Tijt

k=−2

| {z } Pre-Trend

L X

exit l + δijt + εijt , Eijt

(12)

|l=0{z }

Post-Entry Trend

where yijt refers to TFPR, marginal cost, markup, or price; αst are sector-year effects that capture trends at the 4-digit level, and αij are plant-product fixed effects (at the 7-digit level). We include two sets of plant-product-year specific dummy variables to capture the trajectory of each variable k yijt before and after entry into export markets. First, Tijt reflects pre-entry trends in the two periods l before exporting. Second, the post-entry trajectory of the dependent variable is reflected by Eijt , which takes value one if product j is exported l periods after export entry.38 Finally, the dummy exit δijt allows for changes in trajectories when plant-products exit the export market. Table 2 (Panel A) reports the coefficients of estimating (12) for the sub-sample of export entrants (and Figure 1 above visualizes the results). TFPR is virtually unrelated to export entry, with tight confidence intervals around zero. This result is in line with the previous literature: there are no apparent efficiency gains of export entry based on TFPR. The trajectory of marginal costs shows a radically different pattern. After entry into the export market, marginal costs decline markedly. 38

Due to our relatively short sample, we only report the results for l = 0, ..., 3 periods after export entry. However, l all regressions include dummies Eijt for all post-entry periods. Also, in order to make trajectories directly comparable across the different outcomes, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero.

18

According to the point estimates, marginal costs are about 12% lower at the moment of entry, as compared to pre-exporting periods. This difference widens over time: one period after entry it is 20%, and after 3 years, 26%. These differences are not only economically but also statistically highly significant. In relative terms, the observed decline in marginal costs after export entry corresponds to approximately one-third of the standard deviation in year-to-year changes in marginal costs across all plant-products in the sample. The trajectory for prices is very similar to marginal costs. This results because markups remain essentially unchanged after export entry. The pattern in markups coincides with the one in TFPR, in line with our theoretical results in Section 2. Finally, physical quantities sold of the newly exported product increase by approximately 20%. Reported Average Variable Costs and TFPQ One potential concern with respect to our marginal cost results is that they rely on the correct estimation of markups. If we underestimate the true changes in markups after export entry, then the computed marginal cost would follow prices too closely. We can address this concern by using the unique feature that plants covered by ENIA report the variable production cost per product, as well as the number of units produced. The questionnaire defines total variable cost per product as the product-specific sum of raw material costs and direct labor involved in production. It explicitly asks to exclude transportation and distribution costs, as well as potential fixed costs. Consequently, dividing the reported total variable cost by the units produced of a given product yields a reasonable proxy for its average variable cost. Figure 4 plots our computed marginal costs against the reported average variable costs (both in logs), controlling for plant-product fixed effects, as well as 4-digit sector-year fixed effects (that is, the figure plots the within plant-product variation that we exploit empirically). The two measures are very strongly correlated. This lends strong support to the markup-based methodology for backing out marginal costs by De Loecker et al. (2016). Panel B of Table 2 shows that reported average variable costs (AVC) decrease after export entry, closely following the trajectory that we identified for marginal cost. Export entry is followed by a decline in reported AVC by 13% in the period of entry, growing to 18% after one year, and to 25% three periods after entry. These results confirm that the documented efficiency gains after export entry are not an artefact of the estimation procedure for marginal costs. Another concern is that the decline in marginal (and average) costs may be driven by increasing returns to scale in combination with expanded production after export entry. Our production function estimates suggest that this is unlikely; we find approximately constant returns to scale in most sectors – the mean sum of all input shares is 1.023 (and weighted by plants in each sector, the average is 1.009).39 Nevertheless, we also compute TFPQ as an alternative efficiency measure 39

Table A.5 in the appendix reports further details, showing output elasticities and returns to scale for each 2-digit

19

that allows for flexible returns to scale (but is subject to the caveats discussed in Section 2.5).40 The last row of Table 2 shows that the trajectory for TFPQ is very similar to marginal costs. This suggests that our results are not confounded by deviations from CRS. 4.2 Matching Results Our within-plant trajectories in Table 2 showed a slight (statistically insignificant) decline in prices and marginal costs of new exported products before entry occurs (in t = −1). This raises the concern of pre-entry trends, which would affect the interpretation of our results. For example, price and marginal cost could have declined even in the absence of exporting, or export entry could be the result of selection based on pre-existing productivity trajectories. In the following we address this issue by comparing newly exported products with those that had a-priori a similar likelihood of being exported, but that continued to be sold domestically only (De Loecker, 2007). This empirical approach uses propensity score matching (PSM) in the spirit of Rosenbaum and Rubin (1983), and further developed by Heckman, Ichimura, and Todd (1997). Once a control group has been identified, the average effect of treatment on the treated plant-products (ATT) can be obtained by computing the average differences in outcomes between the two groups. All our results are derived using the nearest neighbor matching technique. Accordingly, treatment is defined as export entry of a plant-product (at the 7-digit level), and the control group consists of the plant-products with the closest propensity score to each treated observation. We obtain the control group from the pool of plants that produce similar products as new exporters (within 4-digit categories), but for the domestic market only. To estimate the propensity score, we use a flexible specification that is a function of plant and product characteristics, including the level and trends in product-specific costs before export entry, lagged product-level TFPR, the lagged capital stock of the plant, and a vector of other controls in the pre-entry period, including product sales, number of employees (plant level), and import status of the plant.41 Appendix A.6 sector in our ENIA sample. Table A.5 also shows that returns to scale are very similar when we instead estimate a more flexible translog specification. The translog case allows for interactions between inputs, so that output elasticities depend on the use of inputs. Consequently, if input use changes after export entry, this could affect elasticities and thus returns to scale. To address this possibility, we compute the average elasticities for 2-digit sectors using i) all plants, and ii) using only export entrants in the first three periods after entry. Both imply very similar – approximately constant – returns to scale, as shown in columns 5 and 6 in Table A.5. In addition, Table A.12 splits our Chilean sample into sectors with above- and below-median returns to scale and shows that the decline in marginal costs after export entry are actually somewhat stronger in the subset with below-median returns to scale. Thus, it is unlikely that our main results are driven by increasing returns to scale. 40 The estimation procedure for TFPQ is described in Appendix ??. 41 Following Abadie, Drukker, Herr, and Imbens (2004), we use the 5 nearest neighbors in our baseline specification. The difference in means of treated vs. controls are statistically insignificant for all matching variables in t = −1. We include import status to account for the possibility that input trade liberalization drives export entry as in Bas (2012). As a further check, we also replicated our within-plant trajectories in Table 2, controlling for log imports at the plant level. Results are virtually unchanged (available upon request).

20

provides further detail on the methodology. Once we have determined the control group, we use the difference-in-difference (DID) methodology to examine the impact of export entry on productlevel TFPR, marginal cost, and markups. As Blundell and Dias (2009) suggest, using DID can improve the quality of matching results because initial differences between treated and control units are removed. Table 3 shows the matching estimation results. Since all variables are expressed in logarithms, the DID estimator reflects the difference in the growth of outcomes between newly exported products and their matched controls, relative to the pre-entry period (t = −1).42 When compared to the previously reported within-plant-product trajectories, the PSM results show a slightly smaller decline in marginal costs at export entry (6.5% vs. 12.1%) – which is to be expected if the PSM procedure corrects for pre-trends. However, for later periods, decreases in marginal costs are the same as documented above: the difference in marginal cost relative to the control group grows to 11% in the year after entry, to 20% after two years, and to 27% three periods after entry. Our alternative efficiency measures – reported average variable costs and TFPQ – confirm this pattern. Changes in TFPR after export entry are initially small and statistically insignificant. However, after three periods, TFPR increases by about 9% more for export entrant products than for the matched control products. This suggests that, eventually, efficiency gains are partially reflected in TFPR – we discuss this pattern in more detail below in Section 4.6. 4.3 Robustness and Additional Results In this subsection we check the robustness of our results to alternative specifications and sample selection. Due to space constraints, we present and discuss most tables with robustness checks in Appendix C, and we summarize the main takeaways here. Balanced Sample of Entrants To what extent does unsuccessful export entry drive our results? To answer this question, we construct a balanced sample of export entrants, including only plant-products that are consistently exported for four subsequent years. Table 4 shows the propensity score matching results for this balanced sample. The main pattern is unchanged. TFPR results are quantitatively small and insignificant in the first two years of exporting, but now there is stronger evidence for increases in TFPR in later periods (which coincide with increasing markups). Marginal costs drop markedly after export entry – by approximately 20-30%. The main difference with Table 3 is that marginal costs are now substantially lower already at the time of export entry (t = 0). This makes sense, given that we only focus on ex-post successful export entrants, who will tend to experience larger 42

For example, a value of 0.1 in period t = 2 means that two years after export entry, the variable in question has grown by 10% more for export entrants, as compared to the non-exporting control group.

21

efficiency gains. In addition, in our baseline matching results (Table 3), efficiency continued to increase over time. This may have been driven by less productive products exiting the export market, so that the remaining ones showed larger average differences relative to the control group. In line with this interpretation, the drop in marginal costs is more stable over time in the balanced sample. Our alternative efficiency measures TFPQ and reported AVC show the same pattern (Panel B of Table 4). In sum, the results from the balanced sample confirm our full sample estimates and suggest relatively stable efficiency gains over time. Single-Product Plants In order to estimate product-level TFPR, marginal costs, and markups, we had to assign inputs to individual products in multi-product plants. This is not needed in single-product plants, where all inputs enter in the production of one final good. Table A.11 uses only the subset of single-product plants to estimate the trajectories following equation (12).43 Despite the fact that the sample is smaller, results for single-product plants remain statistically highly significant and quantitatively even larger than for the full sample. Marginal costs fall by 24-40% after export entry, and this magnitude is confirmed by TFPQ and reported average costs. There is also evidence for increases in TFPR and markups in later periods, but these are quantitatively much smaller than the changes in marginal costs. Further Robustness Checks In our baseline matching estimation, we used the 5 nearest neighbors. Table A.14 shows that using either 3 or 10 neighbors instead does not change our results. Next, we investigate to what extent our results change if we deviate from the Cobb-Douglas specification in our baseline productivity estimation. In Table A.15, we present plant-product level estimates based on the more flexible translog production function, which allows for a rich set of interactions between the different inputs. Again, there is no significant change in TFPR after export entry. In Panel B and C of Table A.15 we use the production function coefficients based on the translog specification to compute markups and marginal costs. This has to be interpreted with caution: because the translog production function is estimated based on revenues and allows for varying input shares over time, it gives rise to a potential bias in the coefficient estimates (see Appendix A.5 for further discussion). In contrast to the Cobb-Douglas specification, this bias is not constant over time and thus not absorbed by fixed effects in within-plant/product analyses. Nevertheless, the bias is probably of minor importance: we obtain very similar results for markups and marginal costs as in the baseline specification. In the same table, we also demonstrate that our results are the same as in the baseline 43

For single-product plants, the product index j in yijt is irrelevant in (12). In line with our methodology for plant-level analyses, we include sector-year fixed effects at the 2-digit level (see footnote 37).

22

when we estimate a quantity production function for the Cobb-Douglas case. Finally, Appendix C.4 shows that results are also relatively similar when analyzed at the plant level. Appendix C discusses the additional robustness checks in greater detail. 4.4 Export Entry Predicted by Tariff Changes In the following, we attempt to isolate the variation in export entry that is driven by trade liberalization. This strategy helps to address endogeneity concerns – in particular, that unobservables may drive both export entry and improvements in efficiency. We follow a rich literature in international trade, using tariff changes to predict export entry. Before presenting the results, we discuss the limitations of this analysis in the context of our Chilean data. Limitations of the 2SLS approach Declines in export tariffs during our sample period (1996-2007) are limited because Chile had already undergone extensive trade liberalization starting in the mid-1970s. Nevertheless, there is some meaningful variation that we can exploit: during the second half of the 1990s, Chile ratified a number of trade agreements with neighboring countries, and between 2003 and 2005, with the United States and the European Union. On average across all destinations, export tariffs for manufacturing products fell from 10.1% in 1996 to 4.5% in 2007 (using total sectoral output in 1996 as constant weights). The European Union and the U.S. were the most important destinations, accounting for 24% and 16% of all exports, respectively, on average over the period 1996-2007. The export tariff decline was staggered over time and thus less dramatic than other countries’ rapid trade liberalization (e.g., Slovenian manufacturing export tariffs to the EU fell by 5.7% over a single year in 1996-97). However, we can exploit differential tariff changes across Chilean sectors. These are illustrated in Figure 5 for 2-digit industries. For example, ‘clothes and footwear’ saw a decline by approximately 10 percentage points, while export tariffs for ‘metallic products’ fell by as little as 2 p.p. In addition, there is variation in the timing of tariff declines across sectors, and the plotted average tariff changes at the 2-digit level in Figure 5 hide underlying variation for more detailed industries. We exploit this variation in the following, using 4-digit ISIC tariff data (the most detailed level that can be matched to our panel dataset).44 This leads to the second limitation of our analysis: as in Bustos (2011), we use industry level tariffs, so that the identifying variation is due to changing export behavior on average for plant44

Chilean tariffs are available at the HS-6 level, but a correspondence to the 7-digit ENIA product code does not exist. The most detailed correspondence that is available matches tariff data to 4-digit ISIC – an industry code that is provided for each ENIA plant. When aggregating export tariffs to the 4-digit level, we use total Chilean exports within each detailed category as weights. For multi-product plants, ENIA assigns the 4-digit ISIC code that corresponds to the plant’s principal product. This does not impose an important constraint on our analysis: for the vast majority (85%) of export-entrant multi-product plants in our sample, the principal product (highest revenue) is in the same 4-digit product category as the one that is exported.

23

products within the corresponding 4-digit tariff categories. The third limitation follows from the staggered pattern of (relatively small) tariff declines over time – as opposed to a short period of rapid trade liberalization. In order to obtain sufficiently strong first stage results, we have to exploit the full variation in tariffs over time. In particular, in most specifications, including year effects – or 2-digit sector-year effects – leaves us with a weak first stage. Consequently, we do not include such fixed effects, so that the full variation in tariffs – across sectors and over time – is exploited. This leads to the possibility that other factors that change over time may drive our results. To alleviate this concern, we control for total sales of each plant. Thus, our results are unlikely to be driven by sales expansions over time that happen to coincide with trends in tariffs. We perform a number of checks to underline this argument. Nevertheless, in light of the limitations imposed by the data, our 2SLS results should be interpreted as an exploratory analysis. Empirical setup We continue to exploit within-plant-product variation, using plant-product fixed effects. In the first stage, we predict export entry based on export tariffs: Eijt = αij + β1 τst + γ1 ln(salesijt ) + εijt ,

(13)

where Eijt is a dummy that takes on value one if plant i exports product j in year t, salesijt are total (domestic and exported) sales, and τst are export tariffs in sector s (to which product j belongs) in year t, as described in footnote 44. Correspondingly, all standard errors are clustered at the 4-digit sector level s. Because we use plant-product fixed effects αij , neither established (continuing) exporters nor plant-products that are never exported affect our results. We thus restrict the sample to export entrants as defined in Section 3.2. Note that our analysis is run in levels rather than changes. This allows for tariff declines in different years to affect export behavior – as we discussed above, Chile’s trade liberalization over our sample period was a staggered process, so that we cannot explore before-after variation over a short time window as in Bustos (2011). In addition, running the analysis in levels with fixed effects (rather than, say, annual changes) allows for flexibility in the timing with which tariff declines affect exporting. For example, if the reaction to lower tariffs gains momentum over time (as in the Canadian case documented by Lileeva and Trefler, 2010), annual changes would not properly exploit this variation. Finally, we use OLS to estimate (13); probit estimates would be inconsistent due to the presence of fixed effects. Column 1 in Table 5 presents our first-stage results for export entrant products, showing that declining export tariffs are strongly associated with export entry. The first stage F-statistic is well above the critical value of 16.4 for 10% maximal IV bias. As discussed above, we only exploit the extent to which tariffs predict the timing of export entry, by including plant-product fixed effects 24

and restricting the sample to those plant-products that become export entrants at some point over the period 1996-2007. The highly significant coefficient on export tariffs thus implies that export entry is particularly likely in 4-digit sectors (and years) where export tariffs decline more steeply. In other words, plant-products that eventually become exporters are particularly likely to do so when they face lower export tariffs. The magnitude of the first-stage coefficient (-8.403) implies that an extra one-percentage-point decrease in export tariffs (both over time and across 4-digit sectors) is associated with an increase in the probability of exporting by 8.4% among those plantproducts that become exporters at some point. Our methodology tackles the endogeneity of export entry in two ways: First, we address the possibility that plant-products that ‘react’ to lower tariffs by export entry differ systematically from those that never start exporting – by restricting the sample to the former. Second, by exploiting only the variation in exporting that is predicted by tariffs, we address the possibility that the timing of export entry may be driven by unobserved productivity trends. Next, we proceed with the second stage, where we regress several characteristics yit that include marginal costs, markups, and TFPR on predicted export entry Eˆijt : ln(yijt ) = αij + β2 Eˆijt + γ2 ln(salesijt ) + ϑijt .

(14)

Columns 2-5 in Table 5 report the second-stage results for our main outcome variables. Marginal costs drop by 27.7% after tariff-induced export entry, and this effect is statistically significant with a p-value of 0.03 (we report weak-IV robust Anderson-Rubin p-values in square brackets, based on Andrews and Stock, 2005). This estimate is remarkably similar to those presented above in Tables 2-4. On the other hand, neither markups nor TFPR change upon (predicted) export entry, while output prices drop similar to marginal costs. This also confirms our results for within-plant trajectories. Our alternative efficiency measures in columns 6 and 7 – reported AVC and TFPQ – also show changes that are quantitatively very similar to those based on marginal costs. In the appendix, we present a number of additional checks. Table A.16 shows that the reducedform results of regressing export entry directly on tariffs show the same pattern as the 2SLS estimates. We also show that there is no relationship between export tariffs and domestic sales at the plant level (Table A.17). This makes it unlikely that our results are driven mechanically by falling tariffs that coincide with expanding sales over time. In sum, despite the limited variation in tariffs, there is compelling evidence for within-plant efficiency gains after tariff-induced export entry, and for our argument that these gains are not captured by TFPR.

25

4.5 Interpretation of Export Entry Results and Possible Channels In the following, we discuss possible channels that may drive the observed trajectories of prices and marginal costs for export entrants. We differentiate between demand- and supply-side explanations. Among the latter, export entry can be driven by selection on pre-exporting efficiency (as in Melitz, 2003), or by a complementarity between exporting and investment in new technology (c.f. Constantini and Melitz, 2007; Atkeson and Burstein, 2010; Lileeva and Trefler, 2010; Bustos, 2011). In addition, anticipated learning-by-exporting also provides incentives for export entry. We discuss the extent to which each of these explanations is compatible with the patterns in the data. Demand-driven export entry If demand shocks – rather than changes in production – were responsible for our results, we should see no change in the product-specific marginal costs, while sales would increase and markups would tend to rise. This is not in line with our empirical observation of falling marginal costs and constant markups. Thus, demand shocks are an unlikely driver of the observed pattern. Selection on pre-exporting productivity Firms that are already more productive to start with may enter international markets because of their competitive edge. Consequently, causality could run from initial productivity to export entry, reflecting self-selection. In this case, the data should show efficiency advantages already before export entry occurs. Since we analyze within-plant-product trajectories, such pre-exporting efficiency advantages should either be captured by plant-product fixed effects, or they would show up as declining marginal costs before export entry. There is only a quantitatively small decline in marginal costs in our within- plant/product trajectories, and a much stronger drop in the year of export entry (see Figure 1). In addition, our matching estimation is designed to absorb preentry productivity differences, and our 2SLS results for tariff-induced export entry are unlikely to be affected by selection. In sum, while we cannot fully exclude the possibility of selection into exporting, it is unlikely to be a major driver of our results. Learning-by-exporting Learning-by-exporting (LBE) refers to exporters gaining expertise due to their activity in international markets. LBE is typically characterized as an ongoing process, rather than a one-time event after export entry. Empirically, this would result in continuing efficiency growth after export entry. There is some limited evidence for this effect in our data: Tables 2 and 3 show a downward trend in marginal costs during the first three years after export entry. However, this may be driven by the differential survival of more successful exporters. In fact, the trend in marginal costs is less pronounced in the balanced sample in Table 4. Thus, learning-by-exporting can at best explain

26

parts of our results. Complementarity between Technology and Exporting Finally, we analyze the case where exporting goes hand-in-hand with investment in new technology. As pointed out by Lileeva and Trefler (2010), expanded production due to export entry may render investments in new technology profitable. In this case, a plant will enter the foreign market if the additional profits (due to both a larger market and lower cost of production) outweigh the combined costs of export entry and investment in new technology. This setup implies an asymmetry in efficiency gains across initially more vs. less productive plants (or plant-products in our setting). Intuitively, productive plants are already close to the efficiency threshold required to compete in international markets, while unproductive plants need to see major efficiency increases to render exporting profitable. Thus, we should expect "negative selection" based on initial productivity – plant-products that are initially less productive should experience larger changes in efficiency. This prediction can be tested in the data. Table 6 provides evidence for this effect, reporting the change in marginal costs for plantproducts with low and high pre-exporting productivity.45 We find a steeper decline in marginal costs for plant-products with low pre-exporting productivity, and the difference is particularly pronounced for ‘young’ exporters in the first two years after export entry. This result is in line with a complementarity channel where exporting and investment in technology go hand-in-hand, and where initially less productive plants will only make this joint decision if the efficiency gains are substantial (Lileeva and Trefler, 2010). The complementarity channel is also supported by detailed data on plant investment. ENIA reports annual plant-level investment in several categories, allowing us to analyze the corresponding trends for export entrants. Because investment is lumpy, we examine the trend in the following intervals: the last two years before export entry ("pre-entry"), the entry year and the first two years thereafter ("young exporters"), and three or more years after entry ("old" exporters). In Panel A of Table 7 we present the results. Coefficients are to be interpreted as within-plant changes relative to the industry level (since we control for plant fixed effects and 2-digit sector-year effects). Overall, investment shows a marked upward trend right after export entry. Disentangling this aggregate trend reveals that it is mainly driven by investment in machinery and – to some degree – by investment in vehicles. Investment in structures, on the other hand, is unrelated to export entry. We also 45

Because marginal costs cannot be compared across plant-products, we use pre-exporting TFPR to split them into above- and below median productivity. Also, pre-exporting TFPR can only be computed when the export entry date is known with certainty. Thus, we cannot apply our 2SLS methodology where tariff changes predict the probability of export entry. Consequently, we use propensity score matching, applied to the subsamples of plant-products with high and low pre-exporting TFPR.

27

confirm this pattern in our auxiliary Colombian and Mexican data, where investment spikes after export entry exclusively for machinery, but not for vehicles or structures (see Table A.27 and Table A.28 in the appendix). The observed time trend in investment is in line with the findings in Bustos (2011).46 Overall, our investment data suggest that the observed efficiency gains are driven by a complementarity between investment in new productive technology and export entry. Alternative Interpretations: Input Prices, and Product Quality Could marginal costs fall after export entry simply because exporters purchase inputs at discounted prices? Panel B in Table 7 examines this possibility, reporting trends in the average price of all inputs, as well as for a stable basket of inputs (those that are continuously used for at least two periods before and after export entry). The table shows that input prices remain relatively stable after export entry, making it unlikely that this channel confounds our results. It is also unlikely that quality upgrading of exporters is responsible for our results, since higher product quality is associated with higher output prices and production costs (c.f. Kugler and Verhoogen, 2012; Manova and Zhang, 2012; Atkin et al., 2014; Fan, Li, and Yeaple, 2015). This is not compatible with the observed decline in output prices, marginal costs, and the relatively stable input prices in our data. In addition, the results from a structural model by Hottman et al. (2016) suggest that quality differences are predominantly associated with TFPR differences, rather than differential costs. On balance, our findings point to exporting-technology complementarity as an important driver of efficiency gains among export entrants. Importantly, the main contribution of our findings is independent of which exact channels drive the results: we show that there are substantial efficiency gains associated with entering the export market, and that the standard TFPR measure does not capture these gains because of relatively stable markups during the first years after entry. 4.6 Stable Markups after Export Entry – A Result of ‘Foreign Demand Accumulation’? We observe that, on average, prices of plant-products fall hand-in-hand with marginal costs after export entry. Understanding why prices fall is important for the interpretation of our results; if they did not change, TFPR would reflect all efficiency gains, eliminating the need for alternative measures. We observed that export entrants charge relatively constant markups (at least in the periods immediately following export entry), so that efficiency gains are passed through to customers. One explanation is that new exporters engage in ‘demand accumulation,’ as described by Foster et al. (2016) – charging lower prices abroad in an attempt to attract customers where ‘demand capital’ 46

It is possible that the installation of new equipment began before export entry, but was reported only after its completion. For example, the ENIA investment category allows for "assets measured in terms of their (historical) accounting cost of acquisition."

28

is still low. If this is the case, we should expect a stronger decline in export prices as compared to their domestic counterparts, because export entrants are already established domestically, but still unknown to international customers. In the following, we provide supportive evidence for this assertion. We can disentangle domestic and foreign prices of the same product in a subsample for 1996– 2000. For this period, the ENIA questionnaire asked about separate quantities and revenues for domestic and international sales of each product. Thus, prices (unit values) can be computed separately for exports and domestic sales of a given product. Within this subsample, we define ‘young’ export entrants as plant-products within 2 years after export entry and compare their average domestic and foreign prices. We find that within plant-products of ‘young’ exporters, the price of exported goods is about 22% lower than pre-export entry, while the price of the same good sold domestically falls by 8%.47 Assuming that the marginal cost of production is the same for both markets, the results provide some evidence that efficiency gains are passed on to both domestic and foreign customers – but significantly more so to the latter. While we cannot pin down the exact mechanism that explains the observed price setting, our observations are in line with ‘demand accumulation’ in foreign markets.

5 Export Expansions of Existing Exporters We have shown that marginal costs drop substantially after export entry, while markups and TFPR remain roughly unchanged. We have interpreted this as evidence for quantitatively important efficiency gains within plants that are not captured by standard productivity measures. Does the same pattern hold for existing exporters – that is, do increases in export volume have the same effect as export entry itself? In the following, we examine this question, exploiting export tariff changes. 5.1 Empirical Setup with Existing Exporters When analyzing existing exporters, we have to switch from the plant-product to the plant level. The reason is that export sales – a crucial variable in this analysis – are reported only at the plant level by ENIA (while export status is reported for each product as a dichotomous variable). Before proceeding, we first check whether our previous findings also hold at the plant level. These results are presented in Appendix C.4.48 Table A.20 presents within-plant trends after export entry, 47

To obtain these estimates, we separately regress logged domestic and export prices (at the 7-digit plant-product level) on an exporter dummy, controlling for plant-product fixed effects and 4-digit sector-year effects. Table A.18 in the appendix shows the results. In addition, Table A.19, estimates the effect of export entry on domestic and foreign profit margins after export entry (which is discussed in detail in Appendix C.3). 48 For multi-product plants, TFPR at the plant level can be calculated with the procedure described in Section 2.3, but aggregating markups and marginal costs to the plant level is less straightforward. We employ the following method, which is explained in more detail in Appendix B.2. First, because our analysis includes plant fixed effects, we can

29

showing that TFPR increases only slightly, while marginal costs decline substantially. The fact that plant-level results are similar to those at the plant-product level is not surprising, given that the exported product typically accounts for the majority of output in exporting multi-product plants. We run the following regression at the plant (i) level: \ it ) + γ ln(domsalesit ) + δi + εit , ln(yit ) = β ln(exports

(15)

where yit denotes our standard outcome variables: marginal costs, markups, and TFPR. We use ex\ it ); more precisely, since we include plant port tariffs to predict plant-level export sales ln(exports fixed effects δi , we implicitly use changes in tariffs to predict changes in exports. As discussed in Section 4.4, we exploit the variation in tariffs over time and across 4-digit sectors – the same limitations as discussed above apply here, too. Next, domsalesit denotes total domestic sales. Controlling for domsalesit ensures that our results are not driven by plant size and are instead attributable to expansions of exports relative to domestic sales. Throughout our analysis of existing exporters, we report results for different subsamples of plants, according to their overall export share. We begin with the full sample that includes all exporters (i.e., all those with export shares above zero) and then move to plants with at least 10%, 20%,...,50% export share. This reflects the following tradeoff: On the one hand, plants that export a larger fraction of their output will react more elastically to changes in trade costs than plants that export little. Thus, estimated effects will tend to increase as we raise the export share cutoff. On the other hand, for plants that already have a high export share there is a smaller margin to increase exports relative to total sales. This will attenuate the effect of falling tariffs. In combination, the two opposing forces should lead first to stronger and then to weaker effects as we increase the export share cutoff. Indeed, we find that results are typically strongest for plants with 20-40% export shares. 5.2 Tariff Changes and Within-Plant Efficiency Gains: 2SLS Results We obtain a strong first stage when estimating (15) – the first stage F-statistics typically exceed the critical value for a maximal 10% IV bias (detailed first stage results are shown in Appendix Table A.22). In terms of magnitude, tariff declines over our sample period predict increases in export sales by approximately 20-30% among existing exporters (on average across the different specifications). Table 8 presents the second stage of our 2SLS results. These show that tariff-induced exnormalize plant-level marginal costs and markups to unity in the last year of our sample, 2007 (or the last year in which the plant is observed). We then compute the annual percentage change in marginal cost at the plant-product level. Finally, we compute the average plant-level change, using product revenue shares as weights, and extrapolate the normalized plant-level marginal costs. For markups, we use the same product revenue shares to compute a weighted average plant-level markup.

30

port expansions led to statistically significant efficiency increases, as measured by falling marginal costs (panel A) and rising TFPQ (panel B). To interpret the magnitude of effects, we compute the b change in each outcome due to the overall tariff reduction over the sample period (denoted by △).

For example, in col 3, panel A, the effect size of -0.218 is obtained by multiplying the coefficient b in exports for 1996-2007 from the estimate (-0.845) with the corresponding predicted increase △ first-stage regressions in Appendix Table A.22 (0.258). We find that export tariff declines are asso-

ciated with marginal costs falling by approximately 25% over the sample period; the TFPQ results confirm this magnitude. This is similar to the observed efficiency gains after export entry (15-25% as reported in Table 5). If taken at face value, our results thus suggest that export entry has (on average) a similar effect on productivity as a tariff-induced increase in export volume by 20-30% among existing exporters. Next, we turn to the results for markups and TFPR (panel C and D in Table 8, respectively). Both variables increase statistically significantly with tariff-induced export expansions among firms that export more than 10% of their output (cols 2-6). Nevertheless, TFPR captures only about one quarter of the efficiency gains reflected by marginal costs and TFPQ: tariff declines over our sample period raised TFPR by approximately 5%. The increase in markups is very similar, in line with our result in Section 2. Our results for tariff-induced export expansions thus also imply that about three-quarters of the efficiency gains reflected by lower marginal costs are passed on to customers in the form of lower prices. In Appendix C.2 we present a number of consistency checks. Table A.23 shows the reducedform results corresponding to Table 8. We confirm the 2SLS results: lower tariffs lead to significant declines in marginal costs, and to significant (but relatively smaller) increases in markups and TFPR. Next, Table A.24 shows that falling export tariffs are not associated with changes in domestic sales. This suggests that we identify a pattern that is specific to trade, and not driven by a general expansion of production. In Table A.25 we show that input prices are largely unchanged following tariff-induced export expansions. Finally, Table A.26 shows that tariff-induced export expansions are also associated with increases in capital stock. This is compatible with our interpretation that investment in new technology is responsible for the observed efficiency increases. The fact that for existing exporters some of the increased efficiency is captured by TFPR marks an important difference to the results on export entry, where markups and TFPR remained largely unchanged. The core of the difference is related to pricing behavior: while new export entrants pass efficiency gains on to their international customers, established exporters raise markups. Related to our discussion in Section 4.6, existing exporters may face relatively less elastic demand because they already have an established customer base. This may explain why efficiency increases translate – at least partially – into higher markups for established exporters. This interpretation is also 31

in line with models such as Melitz and Ottaviano (2008), where lower tariffs have an effect akin to a demand shock for existing exporters, inducing them to raise markups.

6 Evidence from other Countries: Colombia and Mexico In this section, we repeat our main empirical analysis for two additional countries: Colombia (2001-13) and Mexico (1994-2003). Both provide datasets with similarly detailed coverage as the Chilean ENIA, and these datasets have been used extensively in studies of international trade.49 Appendices B.3 and B.4, respectively, describe the Colombian and Mexican data in detail and show that the standard stylized facts documented for Chile in Table 1 hold in these samples, as well. Appendix B.5 discusses export entry in the two samples, and Appendix B.6 compares them to the Chilean ENIA, showing that the sectoral composition in all three samples is similar. In terms of export orientation, Chile and Colombia are also comparable, with about 20-25% of all plants being exporters. Mexican manufacturing plants, on the other hand, exports more of their output – about 39% (which may in part be due to larger plants being overrepresented in the Mexican sample). One important limitation is that – unlike the Chilean ENIA – the Colombian and Mexican data do not provide product-specific variable costs. We therefore cannot use equation (8) to compute product-specific material shares in multi-product plants – the basis to derive product-specific markups and marginal costs. We thus restrict our analysis for Colombia and Mexico to the subset of single-product plants, where all inputs are clearly related to the (single) produced output. Fortunately, both datasets include a large number of single-product plants – with almost 20,000 plant-year observations each (as compared to 25,000 for Chile). This allows us to compare the single-product results for Chile (shown in Table A.11) to those obtained for Colombia and Mexico, using exactly the same methodology.50 We begin by describing the within-plant trajectories for Colombia in Figure 2, with the coefficients presented in Table 9. TFPR remains essentially unchanged after export entry. Marginal costs, on the other hand, show a steep and highly significant decline by up to 40% after export entry. Markups increase mildly, by less than 10%.51 TFPQ confirms the magnitude of the marginal 49

For example, Kugler and Verhoogen (2012) and Eslava et al. (2013) use the Colombian firm-level data from the Annual Manufacturing Survey (Encuesta Annual Manufacturera); Iacovone and Javorcik (2010) and Eckel, Iacovone, Javorcik, and Neary (2015) use data from the Mexican Monthly Industrial Survey (Encuesta Industrial Mensual) and from the Annual Industrial Survey (Encuesta Industrial Anual). 50 In all three cases, we estimate (12) for single-product plants, including plant fixed effects. We also include sectoryear fixed effects at the 2-digit level, in line with our methodology for plant-level analyses (see footnote 37). 51 The fact that markups grow somewhat more than TFPR is discussed in Appendix A.2: Colombian manufacturing shows on average (slightly) increasing returns to scale. In this case, fast expansions of volume (which are also observed for Colombia – see Panel B of Table 9) can lead to MC overestimating efficiency gains, and to markup changes

32

cost trajectory. Figure 3 and Table 10 present the within-plant trajectories for Mexican export entrants. There is no change in TFPR or markups. Marginal costs, on the other hand, decline by 15-20% in the three years after export entry. This is quantitatively smaller than in the case of Colombia, but the results remain statistically significant at the 5% level. The results for TFPQ confirm the efficiency gains reflected by marginal costs. One potential reason for the relatively smaller efficiency gains after export entry is that larger plants are overrepresented in the Mexican data (see Appendix B.4). Larger plants are on average more productive (Syverson, 2011), and we know from the Lileeva and Trefler type test in Section 4.5 that more productive plants tend to see smaller efficiency gains after export entry. In fact, when splitting the Chilean sample into plants with above- and below-median employment, we also find smaller productivity gains for larger plants after export entry (see Table A.13). Altogether, the results for Colombia and Mexico strongly confirm our findings for Chile: after export entry, plants experience significant efficiency increases, and these are almost entirely passed on to consumers in the form of lower prices. Thus, TFPR remains almost unchanged, which confirms its inferiority to alternative measures such as marginal costs or TFPQ. In Tables A.27 and A.28 in the appendix we show that investment of Colombian and Mexican export entrants spikes after export entry for "young exporters," and that this is almost entirely driven by increasing investment in machinery (as opposed to structures or vehicles). This confirms our findings for Chile, and suggests that an export-investment complementarity is a likely candidate for explaining the observed efficiency gains in Colombia and Mexico, as well.

7 Discussion and Conclusion Over the last two decades, a substantial literature has argued that exporting induces within-plant efficiency gains. This argument has been made by theoretical contributions in the spirit of Grossman and Helpman (1991) and is supported by a plethora of case studies in the management literature. The finding that exporting induces investment in new technology also suggests that within-plant efficiency gains must exist (Bustos, 2011). A large number of papers has sought to pin down these effects empirically, using firm- and plant-level data from various countries in the developed and developing world. With less than a handful of exceptions, the overwhelming number of studies has failed to identify such gains. We pointed out a reason for this discrepancy, and applied a recently developed empirical methodology to resolve it. Previous studies have typically used revenue-based productivity measures, which are downward biased if higher efficiency is associated with lower output prices. In order to avoid this bias, exceeding TFPR changes.

33

we estimated marginal costs as a productivity measure at the plant-product level, following the approach by De Loecker et al. (2016). We have documented that marginal costs drop significantly after export entry, while markups remain relatively stable. Thus, productivity gains after export entry are largely passed on to customers in the form of lower output prices. We also showed that the typically used revenue-productivity remains largely unchanged after export entry. These results hold in three different countries that provide sufficiently detailed manufacturing data for our analysis: Chile, Colombia, and Mexico. Thus, our results likely reflect a general pattern, implying that a large number of previous studies has underestimated export-related efficiency gains by focusing on revenue-based productivity. To support our argument that the observed efficiency gains are indeed trade-related, we used tariff variations in the particularly rich Chilean manufacturing panel. In this context, we distinguished between tariff-induced export entry and expanding foreign sales by established exporters. We found that both are associated with declining marginal costs (and – as a robustness check – with increasing TFPQ). We also compared these results to those based on the typically used TFPR. For tariff-induced export entry, TFPR fails to identify any gains; for tariff-induced export expansions, TFPR gains are statistically significant, but they reflect only one quarter of the productivity gains captured by marginal costs. These differences arise from the behavior of markups: on average, export entrants pass on almost all efficiency gains to customers – markups are unchanged, and therefore TFPR is unchanged. Established exporters, on the other hand, translate part of the efficiency gains into higher markups. These observation are compatible with ‘demand accumulation’ (Foster et al., 2016): new exporters may charge low prices initially in order to attract customers, while established exporters can rely on their existing customer network, so that lowering prices is less vital. To gauge the quantitative importance of our findings, we compare the observed within-plant efficiency gains after export entry for the different productivity measures. We begin with TFPR. For export entrants, we found no increase in TFPR; and for tariff-induced export expansions of established exporters, the gains over the full sample period are approximately 5% (Table 8). Thus, if we had used the common revenue-based productivity measure, we would have confirmed the predominant finding in the previous literature – little evidence for within-plant efficiency gains. Based on marginal costs, on the other hand, new export entry is accompanied by efficiency increases of 15-25%. In addition, tariff-induced export expansions led to approximately 20% higher efficiency over our sample period – roughly four times the magnitude reflected by TFPR. Compare this to Lileeva and Trefler (2010), who found that labor productivity rose by 15% for Canadian exporters during a major trade liberalization with the US in 1984-96. Since labor productivity is subject to the same (output) price bias as TFPR, the actual efficiency gains may well have been larger – if 34

Canadian exporters, similar to their Chilean counterparts, passed on some of the efficiency gains to their customers in the form of lower prices. Note that TFPR underestimating export-related efficiency gains is not a foregone conclusion: In principle, TFPR could also overestimate actual efficiency gains – if markups rise more than productivity. An extreme example would be exporters that raise their markups when tariffs fall, but do not invest in better technology. While our results suggest that such a strong response of markups is unlikely, we do observe markup increases among existing exporters when tariffs fall. This implies that the output price bias of TFPR is weaker during trade liberalization. One interpretation is that export tariff declines have an effect akin to demand shocks, which creates incentives to raise markups in models with endogenous markups such as Bernard et al. (2003) or Melitz and Ottaviano (2008). Consequently, it is more likely to find TFPR (i.e., markup) increases during periods of falling export tariffs. This may explain why the few studies that have identified exportrelated within-plant efficiency gains exploited periods of rapid trade liberalization (such as De Loecker, 2007 or Lileeva and Trefler, 2010). Our results have two important implications for gains from trade: First, they rectify the balance of within-plant efficiency gains versus reallocation across plants. So far, the main effects have been attributed to the latter. For example, Pavcnik (2002) estimates that reallocation is responsible for approximately 20% productivity gains in export-oriented sectors during the Chilean trade liberalization over the period 1979-86. Using marginal cost as a productivity measure that is more reliable than its revenue-based counterparts, we show that export-related within-plant efficiency gains probably have a similar order of magnitude. Second, our results underline the necessity for future empirical studies to use productivity measures that are not affected by changes in output prices – and to re-examine previous findings that used revenue productivity. In particular, future studies should make further progress where our analysis was mostly exploratory due to the limited variation in Chilean export tariffs. Ideally, more detailed tariff changes at the plant- or disaggregated industry-level should be combined with marginal costs as a more reliable proxy for efficiency gains. Finally, our results imply that relatively stable markups are the reason why efficiency gains are not fully translated into higher revenue productivity. Thus, future research should examine the relationship between exporting and markups in more detail.

35

References Abadie, A., D. Drukker, J. L. Herr, and G. W. Imbens (2004). Implementing Matching Estimators for Average Treatment Effects in Stata. Stata Journal 4, 290–311. Ackerberg, D. A., K. Caves, and G. Frazer (2015). Identification Properties of Recent Production Function Estimators. Econometrica 83(6), 2411–2451. Alvarez, R. and R. López (2005). Exporting and Performance: Evidence from Chilean Plants. Canadian Journal of Economics 38(4), 1384–1400. Amiti, M. and J. Konings (2007). Trade Liberalization, Intermediate Inputs, and Productivity: Evidence from Indonesia. American Economic Review 97(5), 1611–1638. Andrews, D. W. K. and J. H. Stock (2005). Inference with Weak Instruments. NBER Working Paper 313. Atkeson, A. and A. T. Burstein (2010, 06). Innovation, firm dynamics, and international trade. Journal of Political Economy 118(3), 433–484. Atkin, D., A. K. Khandelwal, and A. Osman (2014). Exporting and Firm Performance: Evidence from a Randomized Trial. Working Paper. Aw, B. Y., S. Chung, and M. J. Roberts (2000). Productivity and Turnover in the Export Market: Micro-level Evidence from the Republic of Korea and Taiwan (China). World Bank Economic Review 14(1), 65–90. Aw, B. Y., M. J. Roberts, and D. Y. Xu (2011). R&D Investment, Exporting, and Productivity Dynamics. American Economic Review 101(4), 1312–44. Bas, M. (2012). Input-trade Liberalization and Firm Export Decisions: Evidence from Argentina. Journal of Development Economics 97(2), 481–493. Bernard, A. and J. Wagner (1997). Exports and Success in German Manufacturing. Review of World Economics (Weltwirtschaftliches Archiv) 133(1), 134–157. Bernard, A. B., J. Eaton, J. B. Jensen, and S. Kortum (2003). Plants and Productivity in International Trade. American Economic Review 93(4), 1268–1290. Bernard, A. B. and J. B. Jensen (1999). Exceptional Exporter Performance: Cause, Effect, or Both? Journal of International Economics 47(1), 1–25. Bernard, A. B., J. B. Jensen, S. J. Redding, and P. K. Schott (2012). The Empirics of Firm Heterogeneity and International Trade. Annual Review of Economics 4, 283–313. Blundell, R. and M. C. Dias (2009). Alternative Approaches to Evaluation in Empirical Microeconomics. Journal of Human Resources 44(3), 565–640. Bustos, P. (2011). Trade Liberalization, Exports, and Technology Upgrading: Evidence on the Impact of MERCOSUR on Argentinian Firms. American Economic Review 101(1), 304–40. Caliendo, L., G. Mion, L. Opromolla, and E. Rossi-Hansberg (2015). Productivity and Organization in Portuguese Firms. Clerides, S. K., S. Lach, and J. R. Tybout (1998). Is Learning By Exporting Important? Micro-Dynamic Evidence From Colombia, Mexico, and Morocco. Quarterly Journal of Economics 113(3), 903–947. Constantini, J. A. and M. J. Melitz (2007). The Dynamics of Firm-Level Adjustment to Trade Liberalization. In E. Helpman, D. Marin, and T. Verdier (Eds.), The Organization of Firms in a Global Economy. Cambridge: Harvard University Press.

36

Davis, S., C. Grim, and J. Haltiwanger (2008). Productivity Dispersion and Input Prices: The Case of Electricity. Working Papers 08-33, Center for Economic Studies, U.S. Census Bureau. De Loecker, J. (2007). Do exports generate higher productivity? Evidence from Slovenia. Journal of International Economics 73(1), 69–98. De Loecker, J. (2011). Product Differentiation, Multiproduct Firms, and Estimating the Impact of Trade Liberalization on Productivity. Econometrica 79(5), 1407–1451. De Loecker, J. (2013). A Note on Detecting Learning by Exporting. American Economic Journal: Macroeconomics. forthcoming. De Loecker, J., P. K. Goldberg, A. K. Khandelwal, and N. Pavcnik (2016). Prices, Markups and Trade Reform. Econometrica 84(2), 445–510. De Loecker, J. and F. Warzynski (2012). Markups and Firm-Level Export Status. American Economic Review 102(6), 2437–2471. Eckel, C., L. Iacovone, B. Javorcik, and J. Neary (2015). Multi-Product Firms at Home and Away: Costversus Quality-Based Competence . Journal of International Economics 95(2), 216–232. Eslava, M., J. Haltiwanger, A. Kugler, and M. Kugler (2013). Trade and market selection: Evidence from manufacturing plants in colombia. Review of Economic Dynamics 16(1), 135–158. Evenson, R. E. and L. E. Westphal (1995). Technological Change and Technology Strategy. In H. Chenery and T. Srinivasan (Eds.), Handbook of Development Economics, Volume 3 of Handbook of Development Economics, Chapter 37, pp. 2209–2299. Elsevier. Fan, H., Y. A. Li, and S. Yeaple (2015). Trade Liberalization, Quality and Export Prices. Review of Economics and Statistics 97(5), 1033–1051. Foster, L., C. Grim, J. Haltiwanger, and Z. Wolf (2016). Firm-Level Dispersion in Productivity: Is the Devil in the Details? American Economic Review, Papers & Proceedings 106(5). forthcoming. Foster, L., J. Haltiwanger, and C. Syverson (2008). Reallocation, Firm Turnover, and Efficiency: Selection on Productivity or Profitability? American Economic Review 98(1), 394–425. Foster, L., J. Haltiwanger, and C. Syverson (2016). The Slow Growth of New Plants: Learning about Demand? Economica 83(329), 91–129. Goldberg, P. K., A. K. Khandelwal, N. Pavcnik, and P. Topalova (2010). Imported Intermediate Inputs and Domestic Product Growth: Evidence from India. Quarterly Journal of Economics 125(4), 1727–1767. Gorodnichenko, Y. (2012). Using Firm Optimization to Evaluate and Estimate Productivity and Returns to Scale. Working paper. Grossman, G. M. and E. Helpman (1991). Innovation and Growth in the Global Economy. Cambridge, MA: MIT Press. Hall, R. E. (1986). Market Structure and Macroeconomic Fluctuations. Brookings Papers on Economic Activity 17(2), 285–338. Heckman, J. J., H. Ichimura, and P. E. Todd (1997). Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme. Review of Economic Studies 64(4), 605–54. Hottman, C., S. J. Redding, and D. E. Weinstein (2016). Quantifying the Sources of Firm Heterogeneity. Quarterly Journal of Economics. forthcoming.

37

Hsieh, C.-T. and P. J. Klenow (2009). Misallocation and Manufacturing TFP in China and India. Quarterly Journal of Economics 124(4), 1403–1448. Iacovone, L. and B. Javorcik (2012). Getting Ready: Preparation for Exporting. CEPR Discussion Paper 8926. Iacovone, L. and B. S. Javorcik (2010). Multi-Product Exporters: Product Churning, Uncertainty and Export Discoveries. Economic Journal 120(5), 481–499. ISGEP (2008). Understanding Cross-Country Differences in Exporter Premia: Comparable Evidence for 14 Countries. Review of World Economics (Weltwirtschaftliches Archiv) 144(4), 596–635. [ISGEP: International Study Group on Exports and Productivity]. Klette, T. J. and Z. Griliches (1996). The Inconsistency of Common Scale Estimators when Output Prices are Unobserved and Endogenous. Journal of Applied Econometrics 11(4), 343–61. Kugler, M. and E. Verhoogen (2012). Prices, Plant Size, and Product Quality. Review of Economic Studies 79(1), 307–339. Lamorgese, A. R., A. Linarello, and F. Warzynski (2014). Free Trade Agreements and Firm-Product Markups in Chilean Manufacturing. Working Paper, School of Economics and Management, University of Aarhus. Levinsohn, J. and A. Petrin (2003). Estimating Production Functions Using Inputs to Control for Unobservables. Review of Economic Studies 70(2), 317–341. Lileeva, A. and D. Trefler (2010). Improved Access to Foreign Markets Raises Plant-Level Productivity...for Some Plants. Quarterly Journal of Economics 125(3), 1051–1099. López, R. A. (2005). Trade and growth: Reconciling the macroeconomic and microeconomic evidence. Journal of Economic Surveys 19(4), 623–648. Luong, T. A. (2013). Does Learning by Exporting Happen? Evidence from the China Automobile Industry. Review of Development Economics. forthcoming. Manova, K. and Z. Zhang (2012). Export Prices across Firms and Destinations. Quarterly Journal of Economics 127(1), 379–436. Mas-Colell, A., M. D. Whinston, and J. R. Green (1995). Microeconomic Theory. Oxford University Press, USA. Melitz, M. (2000). Firm Productivity Estimation in Differentiated Product Industries. Working paper. Melitz, M. J. (2003). The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity. Econometrica 71(6), 1695–1725. Melitz, M. J. and G. I. P. Ottaviano (2008). Market Size, Trade, and Productivity. Review of Economic Studies 75(1), 295–316. Montenegro, C. E. and C. Pagés (2004). Who Benefits from Labor Market Regulations? Chile, 1960-1998. In Law and Employment: Lessons from Latin America and the Caribbean, NBER Chapters, pp. 401–434. National Bureau of Economic Research, Inc. Olley, G. S. and A. Pakes (1996). The Dynamics of Productivity in the Telecommunications Equipment Industry. Econometrica 64(6), 1263–1297. Ornaghi, C. (2006). Assessing the effects of measurement errors on the estimation of production functions. Journal of Applied Econometrics 21(6), 879–891.

38

Park, A., D. Yang, X. Shi, and Y. Jiang (2010). Exporting and Firm Performance: Chinese Exporters and the Asian Financial Crisis. Review of Economics and Statistics 92(4), 822–842. Pavcnik, N. (2002). Trade Liberalization, Exit and Productivity Improvements: Evidence from Chilean Plants. Review of Economic Studies 69, 245–76. Rhee, W., B. Ross-Larson, and G. Pursell (1984). Korea’s Competitive Edge: Managing Entry into World Markets. Johns Hopkins University Press, Baltimore. Rosenbaum, P. R. and D. B. Rubin (1983). The central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika 70(1), 41–55. Smeets, V. and F. Warzynski (2013). Estimating productivity with Multi-Product Firms, Pricing Heterogeneity and the Role of International Trade. Journal of International Economics 90(2), 237–244. Syverson, C. (2011). What Determines Productivity? Journal of Economic Literature 49(2), 326–365. Van Biesebroeck, J. (2005). Exporting Raises Productivity in Sub-Saharan African Manufacturing Firms. Journal of International Economics 67, 373–391. Verhoogen, E. (2008). Trade, Quality Upgrading and Wage Inequality in the Mexican Manufacturing Sector. Quarterly Journal of Economics 123(2), 489–530. Wagner, J. (2007). Exports and Productivity: A Survey of the Evidence from Firm-level Data. The World Economy 30(1), 60–82. Wagner, J. (2012). International trade and firm performance: a survey of empirical studies since 2006. Review of World Economics (Weltwirtschaftliches Archiv) 148(2), 235–267. Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press. World Bank (1993). The East Asian Miracle: Economic Growth and Public Policy. Oxford University Press, New York.

39

FIGURES Price, Marginal Cost and Markups .3

.3

Revenue Productivity (TFPR)

Mg.Cost

Markup

2

3

−.3

−.15

0

.15

Price

−.45

−.45

−.3

−.15

0

.15

TFPR

−2

−1

0

1

2

3

−2

Periods before/after export entry

−1

0

1

Periods before/after export entry

Figure 1: Trajectories for Export Entrants in Chile Notes: Data are from the Chilean Annual Industrial Survey (ENIA) for the period 1996-2007. The figure shows the trajectories for our main outcome variables before and after export entry; period t = 0 corresponds to the export entry year. The left panel shows the trajectory for revenue productivity (TFPR); the right panel, for marginal cost, price, and markup. All results are at the plant-product level. A plant-product is defined as an entrant if it is the first product exported by a plant and is sold domestically for at least one period before entry into the export market (see Section 3.2). Coefficient estimates are reported in Table 2. The lines and whiskers represent 90% confidence intervals.

40

Price, Marginal Cost and Markups .3

.3

Revenue Productivity (TFPR)

Mg.Cost

Markup

2

3

−.45

−.3

−.15

0

.15

Price

−.6

−.6

−.45

−.3

−.15

0

.15

TFPR

−2

−1

0

1

2

3

−2

−1

Periods before/after export entry

0

1

Periods before/after export entry

Figure 2: Trajectories for Export Entrants in Colombia Notes: Data are from the Colombian Annual Manufacturing Survey for the period 2001-13 (described in Appendix B.3). The figure shows the trajectories for our main outcome variables before and after export entry; period t = 0 corresponds to the export entry year. The left panel shows the trajectory for revenue productivity (TFPR); the right panel, for marginal cost, price, and markup. All results are for singleproduct plants. The coefficient estimates are reported in Table 9. The lines and whiskers represent 90% confidence intervals.

Price, Marginal Cost and Markups .3

.3

Revenue Productivity (TFPR)

Mg.Cost

Markup

2

3

−.3

−.15

0

.15

Price

−.45

−.45

−.3

−.15

0

.15

TFPR

−2

−1

0

1

2

3

−2

Periods before/after export entry

−1

0

1

Periods before/after export entry

Figure 3: Trajectories for Export Entrants in Mexico Notes: Data are from the Mexican Annual Industrial Survey for the period 1994-2003 (described in Appendix B.4). The figure shows the trajectories for our main outcome variables before and after export entry; period t = 0 corresponds to the export entry year. The left panel shows the trajectory for revenue productivity (TFPR); the right panel, for marginal cost, price, and markup. All results are for singleproduct plants. The coefficient estimates are reported in Table 10. The lines and whiskers represent 90% confidence intervals.

41

10 5 Marginal cost 0 −5 −10

−10

−5 0 Reported average variable cost

5

10

Figure 4: Estimated Marginal Cost and Reported Average Variable Cost

20

Notes: The figure plots plant-product level marginal costs computed using the methodology described in Section 2 against plant-product level average costs reported in the Chilean ENIA panel (see Section 3). The underlying data include both exported and domestically sold products, altogether 109,612 observations. The figure shows the relationship between the two cost measures after controlling for plant-product fixed effects (with products defined at the 7-digit level) and 4-digit sector-year fixed effects. The strong correlation thus indicates that changes in computed marginal cost at the plant-product level are a good proxy for changes in actual variable costs.

Clothes & Footwear Non−Metallic Products

Average Tariff (%) 10

15

Machinery & Equipment Paper & Printing Food & Beverages

Wood & Furniture

5

Chemicals, Rubber & Plastic Other Manufactures

0

Metallic Products

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

Figure 5: Average Chilean Export Tariffs (2-digit industries) Notes: The figure plots the average export tariff for all 2-digit ISIC industries. We first compute average tariffs at the 6-digit HS product level across all destinations of Chilean exports, using destination-specific aggregate export shares as weights. We then derive average tariffs at the more aggregate 2-digit ISIC level.

42

TABLES Table 1: Plant-Level Stylized Facts in Chilean Manufacturing (1)

(2)

Plant Size Dependent Variable

ln(workers)

ln(sales)

(3)

(4)

(5)

Productivity

Wages

Markup

ln(TFPR)

ln(wage)

ln(markup)

Panel A: Unconditional Premia Export dummy

1.402*** (.071)

2.295*** (.170)

.209** (.073)

.463*** (.036)

.0332*** (.010)

Sector-Year FE

X

X

X

X

X

.264 53,536

.317 53,536

.532 53,536

.247 53,536

.062 105,619

R2 Observations

Panel B: Controlling for Employment Export dummy

—

Sector-Year FE R2 Observations

— —

.645*** (.0706)

.186*** (.0295)

.242*** (.0279)

.0320*** (.0108)

X

X

X

X

.715 53,536

.533 53,536

.302 53,536

.062 105,619

Notes: The table reports the percentage-point difference of the dependent variable between exporting plants and non-exporters in a panel of approximately 9,600 (4,500 average per year) Chilean plants over the period 1996-2007. All regressions control for sector-year effects at the 2-digit level; the regressions in Panel B also control for the logarithm of employment. Markups in column 5 are computed at the plant-product level. Standard errors (in parentheses) are clustered at the plant (col 1-4) and plant-product (col 5) level. Key: *** significant at 1%; ** 5%; * 10%.

43

Table 2: Within Plant-Product Trajectories for Export Entrants in Chile Periods After Entry

-2

-1

0

1

2

3

Obs/R2

Panel A: Main Outcomes TFPR

-.0029 (.0193)

.0029 (.0159)

-.0061 (.017)

.0017 (.0212)

.0264 (.0263)

.0159 (.0269)

3,330 .535

Marginal Cost

.0406 (.0651)

-.0406 (.0498)

-.1207** (.0614)

-.1997*** (.0676)

-.2093*** (.0787)

-.2583*** (.0927)

3,330 .792

Markup

-.012 (.0219)

.012 (.0174)

-.0042 (.0189)

.011 (.0233)

.0359 (.0288)

.0189 (.0311)

3,330 .492

Price

.0286 (.0634)

-.0286 (.0491)

-.1248** (.0582)

-.1887*** (.0665)

-.1735** (.0738)

-.2394*** (.0897)

3,330 .804

Physical Quantities

-.0437 (.0913)

.0437 (.0667)

.1899*** (.0719)

.2672*** (.0905)

.1923* (.1045)

.2098* (.1198)

3,330 .822

Panel B: Additional Efficiency Measures Reported AVC

.0297 (.0642)

-.0297 (.0511)

-.1286** (.0600)

-.1838*** (.0672)

-.1904** (.075)

-.2535*** (.0918)

3,330 .795

TFPQ

-.0389 (.0732)

.0389 (.0536)

.118** (.0600)

.1646** (.0683)

.1768** (.0803)

.1937** (.0945)

3,330 .798

Notes: The table reports the coefficient estimates from equation (12). All regressions are run at the plant-product level (with products defined at the 7-digit level); they control for plant-product fixed effects and 4-digit sector-year fixed effects. A plant-product is defined as an export entrant if it is the first product exported by a plant and is sold domestically for at least one period before entry into the export market. Section 4.1 provides further detail. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant-product level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%. TFPR = Revenue productivity; TFPQ = Quantity Productivity; AVC = Average variable cost (self-reported).

44

Table 3: Matching Results: Exported Entry and Efficiency Gains in Chilean Manufacturing Periods After Entry

0

1

2

3

.0152 (.0298)

.0887** (.0396)

Panel A. Main Outcomes TFPR

-.0164 (.0183)

-.0352 (.0236)

Marginal Cost

-.0647* (.0347)

-.110** (.0439)

Markup

.00379 (.0216)

-.0193 (.0246)

Price

-.0609** -.129*** (.0305) (.0420)

-.199*** -.269*** (.0657) (.0882) .0415 (.0300)

.0506 (.0401)

-.158** (.0609)

-.218*** (.0719)

Panel B. Additional Efficiency Measures Reported AVC TFPQ Treated Observations Control Observations

-.0834** -.157*** (.0345) (.0437)

-.153** (.0689)

-.263*** (.0777)

.0470 (.0320)

.0956** (.0429)

.151** (.0667)

.339*** (.0946)

261 1,103

179 752

128 534

75 299

Notes: Period t = 0 corresponds to the export entry year. Coefficients reflect the differential growth of each variable with respect to the pre-entry year (t = −1) between export entrants and controls, all at the plant-product level. The control group is formed by plant-products that had a-priori a similar likelihood (propensity score) of becoming export entrants, but that continued to be sold domestically only. We use the 5 nearest neighbors. Controls are selected from the pool of plant-products in the same 4-digit category (and same year) as the export entrant product. The specification of the propensity score is explained in Section 4.2 and in Appendix A.6. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%. TFPR = Revenue productivity; TFPQ = Quantity Productivity; AVC = Average variable cost (self-reported).

45

Table 4: Matching Results for Chile: Balanced Sample Periods After Entry

0

1

2

3

Panel A. Main Outcomes – Balanced Sample TFPR

.0335 (.0299)

.0421 (.0348)

.112*** (.0355)

.109*** (.0380)

Marginal Cost

-.190** (.0839)

-.234** (.0887)

-.308*** (.0933)

-.225** (.0877)

Markup

.0266 (.0369)

.00565 (.0401)

.110*** (.0382)

.0594 (.0414)

Price

-.151* (.0782)

-.210** (.0795)

-.189** (.0870)

-.152** (.0724)

Panel B. Additional Efficiency Measures – Balanced Sample Reported AVC

-.227** (.0919)

-.268*** (.0843)

-.242** (.0977)

-.220*** (.0813)

TFPQ

.183** (.0831)

.269*** (.0850)

.348*** (.100)

.318*** (.0911)

70 275

71 277

70 276

70 278

Treated Observations Control Observations

Notes: The results replicate Table 3 for the sample of plant-products that are observed in each period t = −2, ..., 3 (balanced panel). See the notes to Table 3 for further detail. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

46

Table 5: Tariff-Induced Export Entry in Chile. Plant-Product Level Analysis First Stage (1)

———————– Second Stage ———————– (2)

(3)

(4)

(5)

Main Outcomes Dependent Variable Export Tariff First Stage F-Statistic Export Dummy Plant-Product FE log Sales Observations

(6)

(7)

Additional Outcomes

Export Dummy

TFPR

MC

Markup

Price

Reported AVC

TFPQ

-8.403*** (1.151) 53.09

—

—

—

—

—

—

–

.0291 [.608]

-.277** [.0338]

.0268 [.702]

-.255* [.0541]

-.312** [.0228]

.259* [.0525]

X X 2,081

X X 2,081

X X 2,081

X X 2,081

X X 2,081

X X 2,081

X X 2,081

Notes: This table examines the effect of tariff-induced export entry on our main outcome variables, as well as on reported average variable costs (AVC) and TFPQ. We report plant-product results, including only plant-products that become new export entrants (see definition in Section 3.2) at some point over the sample period. Export tariffs (at the 4-digit ISIC level) are used to instrument for the timing of export entry. The first stage results of the 2SLS regressions are reported in col 1, together with the (cluster-robust) Kleibergen-Paap rK Wald F-statistic. The corresponding StockYogo value for 10% maximal IV bias is 16.4. Second stage results (cols 2-7) report weak-IV robust Anderson-Rubin p-values in square brackets (see Andrews and Stock, 2005, for a detailed review). All regressions control for the logarithm of plant sales and include plant-product fixed effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to variation in tariffs. Key: *** significant at 1%; ** 5%; * 10%. TFPR = Revenue productivity; TFPQ = Quantity Productivity; AVC = Average variable cost (self-reported).

47

Table 6: Marginal Cost by Initial Productivity of Export Entrants in Chile. Matching Results. Periods After Entry Low Inital Productivity

0

1

-.167*** -.193*** (.0520) (.0649)

2

3

-.148* (.0817)

-.276** (.113)

High Inital Productivity

.0335 (.0449)

-.0331 (.0587)

-.247** (.102)

-.262* (.134)

p-value for difference

[.004]

[.07]

[.45]

[.94]

Treated Observations Control Observations

261 1,103

179 752

128 534

75 299

Notes: The table analyzes heterogenous effects of export entry on marginal costs at the plant-product level, depending on the product-specific initial productivity. Coefficients are estimated using propensity score matching; see the notes to Table 3 for further detail. We use pre-exporting TFPR to create an indicator for plant-products with above- vs. belowmedian productivity and then estimate the average treatment of the treated (ATT) effect separately for the two subsets. Period t = 0 corresponds to the export entry year. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%. The p-value refers to the null hypothesis of equal coefficients for low and high initial productivity.

48

Table 7: Investment and Input Price Trends Before and After Export Entry Period:

Pre-entry

‘Young’ Exp. ‘Old’ Exp. Obs./R2

Panel A. Investment Overall

0.169 (0.269)

0.635** (0.271)

0.337 (0.290)

2,761 0.519

Machinery

0.258 (0.264)

0.737*** -0.277

0.447 (0.294)

2,761 0.521

Vehicles

0.469** (0.232)

0.607** (0.253)

0.267 (0.236)

2,761 0.324

Structures

0.240 (0.249)

-0.147 (0.274)

0.0758 (0.269)

2,761 0.486

Panel B. Input Prices All inputs

-0.0361 (0.155)

-0.0563 (0.163)

-0.0460 (0.195)

7,120 0.368

Stable inputs

-0.0888 (0.152)

0.0284 (0.142)

-0.0946 (0.252)

2,375 0.339

Notes: This table analyzes investment and input prices before and after export entry. All dependent variables are in logs, and all regressions include fixed effects; thus, coefficients reflect the percentage change in investment (panel A) or input prices (panel B) in each respective year relative to the average across all years. ‘Old Exp.’ groups all periods beyond 2 years after export entry; ‘Young Exp.’ comprises export periods within 2 years or less after export entry; and ‘Pre-Entry’ groups the two periods before entry. Regressions in panel A are run at the plant level and control for plant sales, plant fixed effects, and sector-year effects (at the 2-digit level). Regressions in Panel B are run at the 7-digit input-plant level and control for plant-input fixed effects and 4-digit input sector-year effects. In the first row of Panel B (‘All inputs’), we use all inputs observed in the export entry year; in the second row (‘Stable inputs’), we restrict the sample to the set of inputs that are also used at least two periods before and after export entry. The criteria for defining a plant as entrant are described in the notes to Table 2. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

49

Table 8: Tariff-Induced Export Expansions of Exporting Plants in Chile – 2SLS

Export Share

(1)

(2)

(3)

(4)

(5)

(6)

>0%

>10%

>20%

>30%

>40%

>50%

Panel A. log Marginal Cost Index log Exports (predicted) weak-IV robust p-value:

b MC‡ ∆ First Stage F-Statistic Observations

-.692** -.55** -.845*** -.919*** -.879*** -.822*** [.0215] [.0183] [.001] [.0011] [.0017] [.0078] -.119 -.130 -.218 -.242 -.245 -.244 8.92 24.27 21.59 20.56 19.46 11.91 6,996 4,089 3,257 2,815 2,443 2,137 Panel B. log TFPQ

log Exports (predicted) weak-IV robust p-value:

b TFPQ‡ ∆ First Stage F-Statistic Observations

.734** [.0126] .124 8.746 6,988

.52** [.0382] .122 24.12 4,083

.759*** [.0057] .196 21.58 3,256

.728*** [.0089] .192 20.55 2,814

.677** [.0102] .189 19.43 2,442

.627** [.0301] .186 11.91 2,137

.262*** [.0001] .057 22.34 3,974

.223*** [.0001] .052 20.31 3,454

.145*** [.0004] .036 12.87 3,015

Panel C. log Average Markup log Exports (predicted) weak-IV robust p-value:

b Markup‡ ∆ First Stage F-Statistic Observations

.0235 [.78] .003 10.44 9,855

.22*** [.0081] .042 25.19 5,744

.227*** [.0004] .047 24.55 4,570

Panel D. log TFPR log Exports (predicted) weak-IV robust p-value b TFPR‡ ∆ First Stage F-Statistic Observations For all regressions: Plant FE log Domestic Sales

.0461 [.469] .009 10.44 9,855

.182** [.0114] .043 25.19 5,744

.172** [.0134] .044 24.55 4,570

.195*** [.0053] .053 22.34 3,974

.163** [.0115] .047 20.31 3,454

.11 [.195] .034 12.87 3,015

X X

X X

X X

X X

X X

X X

Notes: This table examines the effect of within-plant export expansions due to falling export tariffs on plant-level marginal costs (panel A), TFPQ (panel B), markups (panel C), and TFPR (panel D). The regressions in columns 1-6 are run for different samples, according to the plants’ export shares: col 1 includes all plants with positive exports, col 2 those whose exports account for more than 10% of total sales, col 3, 20%, and so on. The first stage regresses plant-level log exports on sector-specific export tariffs. Export tariffs vary at the 4-digit ISIC level. The first stage regression results are reported in Table A.22 in the appendix. Each panel above reports the second-stage coefficients for the respective outcome variable, together with the weak-IV robust Anderson-Rubin p-values in square brackets (see Andrews and Stock, 2005, for a detailed review). We also report the (cluster-robust) Kleibergen-Paap rK Wald F-statistic for the first stage. The corresponding Stock-Yogo value for 10% maximal IV bias is 16.4. For multiproduct plants, the dependent variables in panels A, B, and C reflect the product-sales-weighted average, as described in Appendix B.2. All regressions control for the logarithm of plant-level domestic sales and include plant fixed effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to the level at which tariffs are observed. Key: *** significant at 1%; ** 5%; * 10%. ‡ b denotes the predicted change in the corresponding dependent variable due to export In each panel of the table, △ tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

50

Table 9: Colombia: Within Plant-Product Trajectories for Export Entrants Periods After Entry

-2

-1

0

1

2

3

Obs/R2

Panel A: Main Outcomes TFPR

.0124 (.0347)

-.0124 (.0260)

.0317 (.0281)

.0344 (.0333)

.0172 (.0393)

.0105 (.0453)

1,056 .616

Marginal Cost

.0143 (.103)

-.0143 (.0862)

-.1128 (.0862)

-.346*** (.113)

-.397*** (.127)

-.393*** (.152)

1,056 .940

Markup

.0172 (.0437)

-.0172 (.0352)

.0508 (.0418)

.0784* (.0443)

.0904* (.0531)

.0684 (.0546)

1,056 .660

Price

.0314 (.0857)

-.0314 (.0708)

-.0624 (.0701)

-.267*** (.0955)

-.306*** (.107)

-.324** (.135)

1,056 .956

Panel B: Additional Outcomes Physical Quantities

-.0355 (.0968)

.0355 (.0777)

.213*** (.0782)

.424*** (.101)

.577*** (.113)

.541*** (.141)

1,056 .945

TFPQ

-.0166 (.0933)

.0166 (.0773)

.0859 (.0732)

.291*** (.104)

.325*** (.115)

.349** (.143)

1,056 .946

Notes: The table reports the coefficient estimates from equation (12), using Colombian manufacturing data. All regressions are run for single-product plants; they control for plant-product fixed effects and for 2-digit sector-year fixed effects. Export entry is defined in Section 3.2, and more specifically for single-product plants, in Appendix B.5. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant-product level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

51

Table 10: Mexico: Within Plant-Product Trajectories for Export Entrants Periods After Entry

-2

-1

0

1

2

3

Obs/R2

Panel A: Main Outcomes TFPR

.0018 (.0205)

-.0018 (.0229)

.0094 (.0225)

-.007 (.0259)

.0101 (.0242)

-.0189 (.0294)

2,036 .720

Marginal Cost

.0112 (.0505)

-.0112 (.0584)

-.0787 (.0678)

-.140** (.0703)

-.174** (.0786)

-.199** (.0904)

2,036 .959

Markup

-.0002 (.0221)

.0002 (.0239)

-.0023 (.0255)

-.0072 (.0272)

.0112 (.0253)

.0115 (.0313)

2,036 .795

Price

.011 (.0453)

-.011 (.0528)

-.0811 (.0615)

-.1471** (.0656)

-.1621** (.0741)

-.1881** (.0807)

2,036 .962

Panel B: Additional Outcomes Physical Quantities

.0002 (.0694)

-.0001 (.0782)

.066 (.0878)

.1362 (.0962)

.1994** (.0975)

.117 (.111)

2,036 .947

TFPQ

-0.013 (0.0535)

0.013 (0.0613)

0.026 (0.0714)

0.129** (0.0746)

0.181** (0.0793)

0.154** (0.0932)

2,036 0.955

Notes: The table reports the coefficient estimates from equation (12), using Mexican manufacturing data. All regressions are run for single-product plants; they control for plant-product fixed effects and for 2-digit sector-year fixed effects. Export entry is defined in Section 3.2, and more specifically for single-product plants, in Appendix B.5. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant-product level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

52

Online Appendix Exporting and Plant-Level Efficiency Gains: It’s in the Measure

A

Alvaro Garcia-Marin

Nico Voigtländer

Universidad de Chile

UCLA, NBER and CEPR

Technical Results and Methodology

A.1 Separability of Marginal Costs into A and Input Prices This appendix complements Section 2.2 in the paper; it proves the log-separability of marginal costs MC(A, w) into physical productivity (A) and a function of input prices φ(w) under the following assumptions: i) the production function is Hicks-neutral and exhibits constant return to scale (i.e., it is homogeneous of degree 1), and ii) input prices are given. Suppose a general production function, Y = A · F (X), where X is a vector of inputs, and Y is measured in physical units of output. The first order condition of the cost minimization problem for any two inputs X i , X j with the corresponding input prices w i and w j , is ψ



Xj Xi



wi = j ⇒ Xj = ζ w



wi wj



Xi

(A.1)

where ψ(·) is an increasing function. The first equality follows from the fact that all homogeneous functions are homothethic, and in the second we replace ζ(·) = ψ −1 (·). Conditional input demand follows from replacing (A.1) for each input j in the production function and rearranging: Xi =

Y i h (w) A

(A.2)

where hi (·) is a function of input prices. Importantly, to obtain (A.2), we use the assumption that the production function is homogeneous of degree 1. Replacing individual factor demands in the objective function yields the cost function: C(Y, w) =

Y φ(w) A

Appendix p.1

(A.3)

where φ(·) is a function of input prices which is homogeneous of degree one and increasing in input prices w.1 Finally, we obtain the marginal cost function from MC(w) =

1 φ(w) A

(A.4)

A.2 The Role of (non-constant) Returns to Scale In this section, we explore how non-constant returns to scale affect the relationship between TFPR, TFPQ (A), and marginal cost (MC). For simplicity, we use a Cobb-Douglas production function, where γ = αL + αM + αK denotes the degree of returns to scale, with the subscripts L, M, and K denoting labor, material inputs, and capital, respectively. Total and marginal costs are then given by:   γ1    αL αM αK Q 1 γ γ γ TC = γ · wL wM wK αK αM αL A αK αM αL  1−γ  γ1  αL αM αK   1 Q γ γ γ wL wM wK MC = αK αM αL A αK αM αL

(A.5) (A.6)

where wi denotes the price of input i and Q is physical output volume. Percentage changes in MC are then the (negative) convex combination of percentage changes in volume and TFPQ (A):2 ∆MC = −



γ−1 γ



 1 ∆Q + ∆A γ

(A.7)

Note that with constant returns to scale (γ = 1), we obtain our result ∆MC = −∆A irrespective of changes in quantity produced (as in Section 2.2). On the other hand, with increasing returns (γ > 1), increases in quantity will also translate into lower marginal costs. Consequently, in the presence of increasing returns to scale, decreases in MC will (in absolute value) overestimate changes in physical efficiency. However, the bias stemming from the increase in quantity has an upper bound that is related to the magnitude of γ. For example, in the Colombian manufacturing data, we have (approximately) γ = 1.05. At the same time, Table 9 in the paper shows that ∆Q ≤ 60% after export entry in the Colombia data. This implies an upper bound for the quantitydriven bias of approximately 4.8% (0.05/1.05 × 0.6).3 1

See for example Mas-Colell, Whinston, and Green (1995, ch. 5) for a proof of these properties of the cost function. All properties of the cost function related to input prices are passed on to the term φ(·). 2 Here we assume that changes in factor prices are zero. 3 Note, however, that at the same time, the increase in ∆A (the second term in equation (A.7)) is scaled down in the case of increasing returns by the factor 1/γ. With γ = 1.05, only about 95% of the productivity gains ∆A will show up as decreases in MC. Thus, if increases in quantity ∆Q are small, while increases in productivity ∆A are large, then

Appendix p.2

Next, we examine to what extent increasing returns affect the relationship between TFPR and markup. Plugging (A.7) into equation (3) in the paper yields: ∆T F P R − ∆µ =



 γ−1 (∆A − ∆Q) γ

(A.8)

With constant returns, this simplifies to our baseline case ∆T F P R = ∆µ, as derived in Section 2.2. With non-constant returns to scale, any differences between ∆T F P R and ∆µ reflect deviations in the growth of TFPQ and output Q. In particular, when quantity increases more rapidly than TFPQ (∆Q > ∆A), increasing returns imply that ∆T F P R < ∆µ.4 Our Colombian data show a (slight) tendency towards increasing returns (as shown in Table A.6) in combination with substantial increases in output and somewhat smaller changes in TFPQ (i.e., the Colombian data feature ∆Q > ∆A). In line with equation (A.8), we observe ∆µ > ∆T F P R in the Colombian trajectories shown in Table 9. The Mexican data also exhibit a slight tendency towards increasing returns (Table A.7). However, here the percentage change in quantities is much smaller and similar in size to TFPQ changes (see Table 10 in the paper). Correspondingly, changes in markups and TFPR are very similar. A.3 Estimating Quantity-Based Production Functions and TFPQ Estimation For completeness, Table A.15 also presents results for quantity-based estimation of our baseline (Cobb-Douglas) production function. There are two main technical differences when the production function is estimated using quantities as opposed to revenues (the latter is described in Section 2 in the paper). First, the output variable in equation (4) is physical quantities instead of revenues. Second, since we do not observe physical inputs in a consistent way for the full sample (see below), we implement the correction suggested by De Loecker, Goldberg, Khandelwal, and Pavcnik (2016) to control for the plant-specific variation in input prices.5 In practice, we modify the first stage of the ACF procedure (given by equation (6) in the paper), by including a vector of variables to proxy for input prices,6 and we modify the second stage by adding lags of these variables as instruments to identify the additional parameters. Again, we confirm our baseline results for both ∆M C will tend to underestimate true productivity increases. 4 To provide intuition for this result, suppose that i) physical efficiency does not change (∆A = 0), ii) output price is unchanged, so that ∆T F P R = ∆A = 0, while iii) quantity rises (∆Q > 0). Then, due to increasing returns, marginal costs must drop (see equation (A.7)), and thus, markups must rise (∆µ > 0). Consequently, increasing returns can drive a wedge between ∆T F P R and ∆µ. 5 This source of bias appears to be less problematic when plant revenues are used as output variable (see De Loecker et al., 2016). Under quality considerations, plants charge higher prices for their outputs and pay more for their inputs (Kugler and Verhoogen, 2012), implying that the input price bias tends to be compensated by the output price variation. 6 Following De Loecker et al. (2016), we include output prices and plant-product sales relative to the overall sales of the same product, as well as the interaction of these variables with capital and materials.

Appendix p.3

markups and marginal costs when estimating these with a Cobb-Douglas production function in quantities (see Table A.15). Given the quantity-based estimation of the production function (i.e., the vector of quantitybased elasticities β s ), we can back out physical productivity TFPQ, using the quantity-equivalent of equation (9). On the output side, physical quantities are directly observed at the plant-product level in the Chilean data. Similarly, physical output is reported in the Colombian and Mexican samples. ++AG: I changed the following please check // NV: Checked.++ As for inputs, we deflate plant-level input expenditures by 4-digit industry deflators for the input-producing industries, following Foster, Haltiwanger, and Syverson (2008). Next, for the Chilean data, we assign these deflated input expenditures to individual products within plants using the reported expenditure shares from ENIA (as calculated in (8)). ++[NV: new from here] This gives rise to the so-called input price bias (see De Loecker et al., 2016). Since more expensive inputs lead to higher materials expenditure with no changes in physical output, the estimated coefficients of the production function will be downward biased in absence of corrections.7 To deal with this source of bias, we apply the correction suggested by De Loecker et al. (2016). Their solution is to include output price and market share in the control function to control for input price variation across producers. With this information, we estimate new production function coefficients and then back out TFPQ at the plant-product-year level. In the cases of our auxiliary datasets for Colombia and Mexico, productspecific input cost shares are not available. We therefore compute TFPQ only for single-product plants in the spirit of Foster et al. (2008).++ Plant-Level Input Price Deflators [++AG: i made substantial changes here, please check // NV: Checked. I made substantial changes too!++] The rich Chilean ENIA data enable us to apply an alternative approach to address the potential input price bias. Rather than using input price deflators at the 4-digit industry level, we can construct a plant-specific input price index. However, this price index can only be constructed for about two-thirds of the observations in our sample, implying a substantially smaller sample. For this reason (and to allow direct comparability of our results to other studies that cannot rely on plant-specific input price deflators), we use the standard industry specific deflators in our main results. In the following, we present the alternative results based on plant-level input price deflators. As we discuss in section 3, the Chilean ENIA provides information on plant-level input prices. This allows us to construct unit-free Tornqvist input-price indexes. ++[NV: New/revised from here] For each plant i in period t, we first define its average input price change as ∆Xit = 7

++? discuss conditions under which this bias is attenuated. In particular, the input price bias may be offset by output price variation when estimating a revenue production function (see section 2.3).++

Appendix p.4

P

κ∈Ji

αiκ ∆ ln xiκt , where αiκ is the average share of input κ in plant i total materials’ expenditure,

xijt is the price of input j, and ∆ ln xijt = ln xijt − ln xij,t−1 . After ∆Xit is computed, the entire sequence of plant-level input prices can be recovered through the recursion Xit = Xi,t−1 + ∆Xit . An important aspect in the calculation of Tornqvist input-price index relates to the construction of the initial average input price Xi0 . For the initial period, we define the normalized initial input price for each input j consumed by plant i (e xij0 ) as the log difference of input price x with its average across all plants s ∈ I that use the same input (measured in the same unit) in the first period P during which input j is used by plant i (i.e., x eij0 = ln xij0 − (1/I) s∈I ln xsj0 ). The advantage of

this normalization is that, provided that the input is used by a sufficiently large number of plants, its price in any period can be interpreted as the log-deviation from the average price computed over all plants using the same input. [++AG: We still need to explain how we get to the plant-level input deflator // Done, see previous paragraph // (and also why we don’t do this at the plant-product level // NV: I don’t really get what you mean here. You mean constructing an input price index for each output?? Because that would be impossible since we have a single input price for each input used by the plant// (in a footnote – i guess noting that since coefficients are estimated using single-product plants, this doesn’t matter, right?)).++]

Once the plant-level input price deflator is obtained, we deflate material expenditures, and then re-estimate the production function with the standard ACF methodology presented in section 2.3. Unlike the baseline revenue productivity case, we express both output and inputs in terms of physical volumes.8 Finally, physical productivity is recovered as the log difference between physical output and the contribution of material inputs (see equation (7)). ++[New from here] Figure A.1 illustrates the trajectories resulting from estimating TFPQ for the various methodologies discussed in this section. All three TFPQ measures in the figure consider output in physical terms. The measures differ in the methodologies used for estimating the production function coefficients, and in the deflators used for materials’ expenditure. TFPQ1 uses production function estimates derived with ? input price’s correction (see section ), and deflates materials’ expenditure with 4-digit price indexes. TFPQ2 uses the same data than TFPQ1, but considers production function coefficients estimated with materials deflated with plant-level input price indexes (see section ). Finally, TFPQ3 considers the same production function coefficients than TFPQ2, but deflates materials with plant-level input prices when computing physical productivity. As it can be seen, all three TFPQ measures are very consistent in showing and increasing 8

[++AG: the following sentence is not clear // Check if now is better // ++] Note that our estimates may still be subject to input price bias coming from variation in capital stock prices. We control for this by including the same vector of variables we used above to proxy for input prices.

Appendix p.5

trajectory after export entry. For the reasons discussed at the beginning of this section, the trajectory for TFPQ3 – which deflates materials with the Tornqvist input price index – displays larger standard errors, resulting in non-significant coefficients for periods zero and one. Nevertheless, the point estimates for all three measures suggests efficiency gains of about 20-30% three periods after export entry, supporting our main results based on marginal cost.++ TFPQ2

TFPQ3

−.3

−.15

0

.15

.3

.45

TFPQ1

−2

−1

0

1

2

3

Periods before/after export entry

Figure A.1: Physical Productivity Trajectories for Export Entrants in Chile Notes: The figure shows the trajectory for various measures of physical productivity (TFPQ) before and after export entry. Data are from the Chilean Annual Industrial Survey (ENIA) for the period 1996-2007. All three TFPQ measures consider output in physical terms; the measures differ in the methodologies used for estimating the production function coefficients, and in the deflators used for materials’ expenditure. TFPQ1 uses production function estimates derived with ? input price’s correction (see section ), and deflates materials’ expenditure with 4-digit price indexes. TFPQ2 uses the same data than TFPQ1, but considers production function coefficients estimated with materials deflated with plant-level input price indexes (see section ). Finally, TFPQ3 considers the same production function coefficients than TFPQ2, but deflates materials with plant-level input prices when computing physical productivity. Period t = 0 corresponds to the export entry year. All results are at the plant-product level. A plant-product is defined as an entrant if it is the first product exported by a plant and is sold domestically for at least one period before entry into the export market (see Section 3.2). The lines and whiskers represent 90% confidence intervals.

A.4 The Effect of Investment on Marginal Costs and TFPQ Since we study efficiency gains in the context of investment-exporting complementarity, it is also worthwhile to discuss how investment in new technology affects TFPQ and marginal cost. In particular, one may worry that while TFPQ explicitly accounts for the effect of fixed-cost investment Appendix p.6

in capital equipment, our estimation may (wrongly) identify declining marginal cost even if the technology itself does not change. We show that this is not the case if plants minimize costs, and under our assumptions from Section 2.2 that i) input prices do not change with export activity and ii) constant returns to scale. In the following discussion, we assume that the input elasticities β s have been correctly estimated for the quantity production function, so that changes in physical output can be readily computed using the quantity-equivalent of (4). Suppose that a plant raises its capital stock by △k (in log changes), adding the same type of machines, so that true efficiency is unchanged (△ω = 0). Because it is minimizing costs, the plant will maintain its expenditure shares for all other inputs (material m and labor l) proportional to the respective input elasticities. Under constant input prices, this implies that △k = △m = △l. Thus, due to constant returns, total output increases by △q = △k, and (7) correctly implies that TFPQ is unchanged. Next, we turn to marginal costs. Recall that we use materials as the variable input V . Also, for the moment, hold output prices P fixed. The first term of (10) – the material input elasticity – is unchanged. In the second term, the quantity of the flexible input V has increased by △m log points, and physical quantity Q has increased by △q log points. Because △q = △k = △m, markups are unchanged for given output prices P . However, the latter may have changed during the plant’s investmentdriven expansion. Suppose that output prices fell by △p log points (e.g., because the plant had to charge lower prices in order to sell its increased output volume). Then (10) implies that the total effect of the investment-driven expansion is a decline in markup by △µ = △p log points. Log-changes in marginal cost can then be computed as △mc = △p − △µ = 0. Consequently, marginal costs correctly reflect that efficiency has not changed. Finally, the same calculation can be made for investment-driven expansions that raise efficiency by △ω > 0 (e.g., by adding new, more efficient machines). Provided that the new technology uses all inputs in the same proportions as before (Hicks-neutrality – a standard assumption in the productivity literature), both TFPQ and marginal costs will drop by △ω. A.5 Discussion of Potential Bias when using Revenue Production Functions Our baseline specification estimates a revenue production function. This has the advantage that we estimate TFPR as well as markups and marginal costs based on the same production function. However, the downside is that for marginal costs and markups, revenue-based output elasticities are potentially affected by an output price bias (see Klette and Griliches, 1996; De Loecker et al., 2016). More precisely, the computation of markups relies on output elasticities, which in turn may be biased. Output elasticities may be biased for two reasons when estimated based on (deflated) revenues. First, if output prices are negatively correlated with input purchases, the coefficients of

Appendix p.7

the production function are biased downward. This correlation could arise if plants cut prices as they add inputs to expand their production. The use of proxy estimators – such as ACF, which we apply – corrects for this bias, since the productivity proxy controls for price variation that is correlated with productivity (c.f. De Loecker and Warzynski, 2012). However, the use of proxy estimators does not correct the second source of bias, which arises if prices are correlated with unobserved demand shocks (that are in turn uncorrelated with expected output). As discussed in Section 2.5, this bias does not affect our within-plant results in the presence of constant output elasticities (i.e., for Cobb-Douglas production functions), because it is absorbed by plant fixed effects. However, the more flexible translog production function allows for varying output elasticities over time and thus potentially introduces price bias in our estimation. We check whether this bias is potentially important below in Table A.15. We find that the within-plant trajectories of marginal costs and markups are practically identical for the translog and the CobbDouglas cases. This suggests that – even when using a translog production function – our results are unlikely to be subject to a quantitatively important output price bias. A.6 Propensity Score Matching Estimation In this appendix we provide technical details on the implementation of the matching estimator outlined in Section 4.2. The specification of the propensity score is P r(ENTRYij,t = 1) = Φ{f (∆acij,t−1 , acij,t−1 , T F P Rij,t−1, ki,t−1 , Zij,t−1 )}

(A.9)

where ENTRYij,t is a categorical variable equal to one if product j produced by plant i enters the export market in period t and Φ(·) is the normal cumulative distribution function.9 As dependent variables, we include a polynomial in the elements of f (·) listed in equation (A.9).10 As Wooldridge (2002) suggests, this could improve the resulting matching as less structure is imposed. Importantly, in our specification we include the lagged and differential average cost (∆acij,t−1 , acij,t−1) to control for pre-trends.11 Following De Loecker (2007), we also include lagged TFPR and lagged capital stock. Both summarize the state of the plant at the pre-entry period. Finally, we include additional product- and plant-level variables in the vector Zij,t−1. These include the number of employees and product sales to control for the size and scale of production, and the import-status of the plant to control for potential differences in efficiency arising from the use of more advanced technology embodied in the use of foreign goods. 9

The definition of entry is provided in section 3.2. We include the level and square of all variables, and all variables are interacted with size and product-level sales. 11 We use average costs because these are directly reported at the product level by each plant. Using instead marginal costs to control for product-specific pre-trends yields similar results. 10

Appendix p.8

All our results are derived using the nearest neighbor matching technique. Accordingly, the group of controls are the plant-products with a propensity score that is closest to that of the new exported product. In our benchmark analysis we use the five nearest-neighbors. As we show in Table A.14, our results are almost unchanged if three or the ten nearest neighbors are used instead. We perform this matching procedure within 4-digit product categories and years. To minimize the presence of low-quality matches, we trim the resulting distribution of controls – we drop the 1st and 99th percentiles – before computing the average effect of entry. In our benchmark results we replace the dropped ‘outlier’ controls with the next best match.12

B Data Appendix In this appendix we provide additional detail on our Chilean dataset, and we describe the Colombian and Mexican data. The Colombian data are very similar to the Chilean ENIA (in particular, both are censuses of manufacturing firms with more than 10 employees). The Mexican data covers 85% of total output in each of 205 detailed manufacturing sectors. In the following, we briefly discuss how we deal with inconsistent product categories, units of output, and other issues of sample selection. We also explain the construction of the plant-level marginal cost index that we use in Section 5. B.1

Additional Detail and Data Cleaning for Chilean ENIA Sample

In this section we describe in detail the procedures used to clean the Chilean ENIA data, and to generate a consistent plant-product dataset (our main dataset). In order to ensure consistent plant-product categories in our panel, we follow four steps. First, we drop plant-year observations whenever there are signs of unreliable reporting. In particular, we exclude plants that have missing or zero values for total employment, investment, use of raw materials, sales, and product quantities. Second, given that we use unit values to proxy for prices, we restrict our sample to the set of plantproduct-year observations with strictly positive sales and quantities. In addition, to reduce noise we drop observations where price or quantities jump by a factor of 10 or more relative to the previous or following year. Third, whenever our analysis involves quantities of production, we have to carefully account for possible changes in the unit of measurement. For example, wine production changes from "bottles" to "liters." Total revenue is generally unaffected by these changes, but the derived unit values (prices) have to be corrected. We correct the derived unit values as follows: Suppose that the unit of measurement changed in year t. We assume that total quantity (measured in the ‘old’ 12

In the (very rare) cases where the next best match is also an outlier, we do not search for ‘second next best’. Also, main results remain qualitatively unchanged when not replacing by the next best match and instead using a smaller control group. These results are available upon request.

Appendix p.9

unit) grew at the same rate as total revenue between t − 1 and t. This allows us to derive quantity measured in the ‘old’ unit for period t, Qold t . Consequently, we can derive the price in terms of the old old old and the new unit: Pt = Rt /Qt ; Ptnew = Rt /Qnew , where R denotes revenue. This implies t the conversion rate X = Ptold /Ptnew that we use for all periods from t onwards – which allows us to measure the good in the old unit throughout the sample period. This procedure is needed for 501 cases, less than 1% of all plant-product observations. Fourth, a similar correction is needed because the product identifier in our sample changes in the year 2001. Before that date, the Chilean Statistical Institute (INE) used the Clasificador Unico de Productos (CUP); thereafter, it adopted the Central Product Classification (CPC) code, V.1.13 This classification was compiled by the U.N.; it contains more than 5,000 products at the 7-digit level and is comparable to the 7-digit ISIC code. The INE provides a correspondence between CUP and CPC. However, this crosswalk does not allow to establish a one-to-one match for all observations. For a subset of the sample (5% of the original sample) no correspondence is provided, and only the new product category is available. We drop all plant-product pairs for which no correspondence is available. We also use an additional procedure to link old and new product identifiers in 2001: We chain products within plants when there is reasonable evidence that they represent the same product. In particular, we assign a common product category for (i) singleproduct plants producing products in the same 4-digit category (1,296 changes), (ii) multipleproduct plants with no adding or dropping of products and with exactly one product changing classification per year (538 changes), (iii) multiple-product plants with at most one product being exported, and with the exported product in two consecutive years changing the product category (167 changes). For (i)-(iii), we require potential candidates to stay within the same 4-digit category (in CUP) before and after the change.14 In addition, whenever the chained products are recorded in different units we apply the procedure outlined in the previous paragraph to homogenize the unit of measurement. This methodology expands the sample by 2,001 plant-product observations – approximately 2% of the overall sample size. After these adjustments, our sample consists of 118,178 plant-product-year observations. Fifth, starting with its latest release that contains the data for 2006-07, INE no longer provides consistent product identifiers because of confidentiality concerns. Instead, randomly generated numbers are reported, which cannot be used to identify product categories. Nevertheless, we managed to match plant-products, using the fact that the latest wave covers the years 2001-07, while an earlier release covers the period 1996-2005 (for which the CPC identifier was still included). Our procedure uses the following steps: i) we match plants based on the ENIA-specific plant iden13 14

This information is available until 2005 – see the explanation in the next paragraph below. The categories we use – which we define in terms of the CUP – are comparable to 4-digit ISIC categories.

Appendix p.10

tifier (the plant identifier is consistently available for the earlier and the later waves), ii) we match products from the two releases within plants in the year 2005, using the fact that product-level expenditures and quantities are available. For example, for a given plant that produced five products in 2005, we matched each of the five products to the later ENIA release based on the productspecific expenditure and quantity produced. Using this procedure, we managed to find a match for more than 90% of the plant-products reported in the earlier 1996-2005 release of ENIA. Consequently, we can track existing plant-products until 2007 and compute all relevant outcome variables for these. However, one caveat remains: new products that were introduced by multi-product plants after 2005 cannot be tracked consistently over our sample.15 Thus, the product categories are unknown for new products after 2005, so that for these, we cannot compute markups or marginal costs. Therefore, our plant-product-level analyses include i) all products produced by single-product plants between 1996 and 2007, and ii) plant-products of multi-product plants that were introduced by 2005 (but potentially produced until 2007).16 Similarly, our plantlevel analysis uses i) all single-product plants between 1996 and 2007, ii) all multi-product plants between 1996 and 2005, and iii) multi-product plants that did not introduce new products in 2006 or 2007. In other words, our plant-level analysis excludes multi-product plants that introduced new products in 2006 or 2007 because for the products introduced after 2005, the plant-level marginal cost index (described in Appendix B.2) cannot be computed. Table A.1 summarizes the number of plants as well as the number of exporting plants in the Chilean ENIA sample. It also lists the number of single-product plants and the exporters among them. B.2

Plant-Level Marginal Cost Index in the Chilean Data

While we can estimate TFPR both at the plant- and at the plant-product level, marginal costs and markups are product-specific. Consequently, we have to aggregate markups and marginal costs for all our analyses at the plant level (such as in Table 8). In the following we explain how we perform this aggregation. Marginal costs are product-specific. Since different product units are typically not comparable, marginal costs cannot be aggregated up to the plant level directly. For this reason, we apply the 15

Since these products did not yet exist in 2005, we cannot use our sample-overlap procedure to match them to the earlier release of ENIA that includes the product-level CPC category. Note that for single-product plants, this problem is less important, because they add new products less frequently. They would need to simultaneously drop another product to keep their single-product status; otherwise, they would become multi-product plants. 16 Since our definition of export entry requires products to be sold domestically for at least one year prior to export entry, we only lose very few "export entries" (i.e., those that were new products – sold only domestically – in 2006 and were then exported in 2007). Our results are robust to restricting the plant-product level analysis to the period 1996-2005.

Appendix p.11

Table A.1: Plants in the Chilean Sample All Plants

Single-Product

# Plants # Exporters # Plants

# Exporters

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

5,263 5,069 4,698 4,548 4,347 4,461 4,766 4,745 4,962 4,815 4,924 4,678

1,164 1,137 1,053 954 882 900 979 1,039 1,088 1,118 1,111 1,062

2,383 2,362 2,256 2,063 2,138 1,978 2,036 1,942 2,049 1,964 2,116 1,744

432 459 438 376 377 328 337 351 387 389 369 337

Total

57,276

12,487

25,031

4,580

Notes: The table shows the number of plants in the Chilean sample, for the both the full sample and for the sub-sample of single-product plants. The table also reports the number of exporting plants in each subset.

Appendix p.12

following procedure to compute a plant-level marginal cost index. First, we compute the change in marginal cost in two consecutive periods for each product j produced by plant i in period t, ∆MCijt (where percentage changes are denoted by the operator ∆). Next, we calculate product j’s share in the sales of plant i in year t: Pijt Qijt , r∈Jit Pirt Qirt

wijt = P

(A.10)

where Jit is the set of products produced by plant i. Then, we compute for each plant the saleweighted average change in marginal cost as g it = ∆MC

X

wijt ∆MCijt

(A.11)

r∈Jit

Finally, we compute the marginal cost index for each plant i in period t recursively starting from the last period Ti for which the plant is observed and marginal costs can be computed for all products (in most cases, Ti = 2007):17 g it = MC

  1, 

g M C i,t+1 , g ∆M C i,t+1

if t = Ti

(A.12)

if t < Ti

If a plant-product has a missing value after its first and before its last year in the sample, this procedure will not be able to compute an MC index for all years. To by-pass this issue, we complete the recursion in these discontinuities using changes in MC over longer time windows (2 or maximum 3 years). B.3

Colombian Data

Colombian plant-level data – Encuesta Annual Manufacturera (EAM) – are collected annually by the National Statistical Administrative Department – Departamento Administrativo Nacional Estadístico (DANE). EAM covers the universe of manufacturing plants with 10 or more workers. It contains detailed information on plant characteristics, such as sales, spending on inputs and raw materials, employment, wages, investment, and plant-product-level export status. We focus on the period 2001-2013 – the longest coherent period over which plant-product level export status is available. Over this period, EAM contains information for more than 7,000 manufacturing plants per year with positive sales and employment information. Out of these, about 25% are exporters. 17

Because ENIA does not report consistent product categories after 2005, we cannot compute marginal costs for newly introduced products in multi-product plants in 2006 and 2007. See the discussion in Appendix B.1 for detail.

Appendix p.13

The number of plants covered in each year is shown in Table A.2. Approximately 70 percent of the plants in EAM are small (less than 50 workers), while medium-sized (50-150 workers) and large (more than 150 workers) plants represent 19 and 11 percent, respectively. Table A.2 also shows that EAM contains about 23,000 observations for single-product plants over the period 2001-13, and among these, more than 4,000 are exporters. Products are defined according to the Central Product Classification V.1 (CPC), adapted to Colombia. This classification contains more than 5,000 products at the 7-digit level (comparable to the 7-digit ISIC code). In order to ensure consistent plant-product categories in the Colombian data, we follow the same steps as described for Chile in Appendix B.1.18 After these adjustments, the Colombian sample consists of 348,465 plant-product-year observations. Unfortunately, in contrast to the Chilean ENIA, the Colombian EAM does not provide data on product-specific total variable costs. Thus, we cannot assign material input costs to individual products in multi-product plants – as we do in the case of Chile in equation (8). Therefore, we restrict the sample to single-product plants. These account for about 12 percent of all plants (see Table A.2). For the subset of single-product plants, all relevant information to compute our main outcome variables is available. Our sample contains 22,869 (single-product) plant-year observations. B.4

Mexican Data

Our Mexican data combines two surveys: the Monthly Industrial Survey – Encuesta Industrial Mensual (EIM) and the Annual Industrial Survey – Encuesta Industrial Anual (EIA). Both are administered by the Instituto Nacional de Estadística, Geografía e Informática (INEGI) in Mexico and cover the same sample of establishments. EIM provides monthly data on production output (quantity and value), but not on inputs (which are available annually from the EIA). We thus compute annual output of each plant-product by summing up the monthly EIM data. We use the longest sample period over which a consistent set of plants is covered by the survey: 1994-2003. Each establishment is classified based on its principal product – its "class of activity" (clase). The clase corresponds to the 6-digit Mexican System of Classification of Productive Activities (CMAP). Each clase, in turn, contains a list of (sub-)products. This list was developed in 1993 and remained unchanged during the period 1994-2003. The list includes 4,085 possible products; of these, 3,183 were produced during our sample period. The sampling methodology has been designed based on the Mexican 1993 industrial census. In each of the 205 clases, the EIM and EIA surveys sample the largest firms until the coverage reaches 18

In the case of the Colombian data, the third step mentioned in Appendix B.1 is not necessary because the product identifier did not change.

Appendix p.14

Table A.2: Plants in the Colombian Sample All Plants # Plants 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013

7,182 7,082 7,076 7,015 7,282 7,022 6,922 7,684 8,730 9,363 9,255 8,943 8,906

Total 102,462

Single-Product

# Exporters # Plants # Exporters 2,034 2,107 2,202 1,561 1,477 1,574 1,354 1,938 2,010 2,051 2,048 2,033 1,954

1,603 1,531 1,473 1,446 1,510 1,400 1,364 1,610 2,008 2,318 2,246 2,161 2,199

319 317 324 225 216 241 201 322 374 407 416 422 404

24,343

22,869

4,188

Notes: The table shows the number of plants in the Colombian sample, for the both the full sample and as for single-product plants. The table also reports the number of exporting plants in each subset.

Appendix p.15

85% of the sectoral output in the 1993 industrial census. In sectors with fewer than 20 plants, all establishments are covered. Large establishments with more than 100 employees are automatically included in the sample. Maquiladoras (import-processing plants) are not surveyed by either EIM or EIA. The surveys cover about 85% of Mexican industrial output (produced by non-maquiladora establishments). Table A.3 shows the number of plants covered by the EIM sample over this period. This number falls from 6,242 plants in 1994 to 4,591 in 2003. This is due to the sampling procedure employed by INEGI, as explained by Iacovone and Javorcik (2010, p.484): "The size of the sample decreases over time due to some establishments exiting the market. No systematic effort is made to replace exiting producers. This means that we observe exit but no entry in our data. The sample covers on average over 50% of total Mexican exports with the percentage covered increasing in the later years." In order to ensure consistent plant-product categories in the Mexican data, we follow the same steps as described for Chile in Appendix B.1.19 After these adjustments, the Mexican sample consists of 173,115 plant-product-year observations. The combined EIM and EIA samples provide most of the variables that are discussed in the context of the Chilean dataset. One important difference is that INEGI does not ask plant managers to report product-specific total variable costs. Thus, we cannot assign material input costs to individual products in multi-product plants – as we do in the Chilean data in equation (8). Consequently, we restrict the sample to single-product plants. These account for about one-third of the Mexican sample (see Table A.3). For the subset of single-product plants (comprising 18,069 plantyear observations), all relevant information to compute our main outcome variables is available. Importantly, plant-level export status is reported annually. B.5

Export Entry in the Colombian and Mexican Sample

The time of entry into export markets is crucial for our analysis. For the case of the Chilean ENIA, our definition of export entry has been described in Section 3.2. Here, we describe the corresponding procedure for Colombia and Mexico. Because for these countries, information on product-specific inputs is not reported for multi-product plants (see Appendix B.3 and B.4), the definition of export entry refers to single-product plants only. We impose three requirements for single-product plant i to be classified as an export entrant in year t: (i) plant i is exporting for the first time at t in our sample, (ii) plant-product i is sold domestically for at least one period before entry into the export market, i.e., we exclude new products that are exported from the very beginning, and (iii) plant-product i continues to exist in the sample for at least two years after export entry, which ensures that we can compute meaningful trajectories. Based on this definition 19

As was the case for Colombia, in the Mexican data the third step mentioned in Appendix B.1 is not necessary because the product identifier did not change.

Appendix p.16

Table A.3: Plants in the Mexican Sample All Plants

Single-Product

# Plants # Exporters # Plants

# Exporters

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003

6,242 5,990 5,722 5,526 5,358 5,215 5,086 4,914 4,752 4,591

1,708 2,077 2,257 2,359 2,238 2,282 2,237 2,095 1,931 1,712

2,089 1,915 1,893 1,843 1,811 1,762 1,753 1,690 1,672 1,641

560 643 744 777 743 764 764 724 681 616

Total

53,396

20,896

18,069

7,016

Notes: The table shows the number of plants in the Mexican sample, for the both the full sample and as for single-product plants. The table also reports the number of exporting plants in each subset.

we find the following numbers for the Colombian and Mexican samples: • Colombia: 264 export entrants (single-plant-products at the 7-digit level). • Mexico: 363 export entrants (single-plant-products at the 7-digit level). Note that our Chilean sample includes 314 are single-product plants. Thus, the number of singleproduct export entrants is of comparable magnitude for all three samples. B.6

Comparison and Validity of the Three Samples

In this section, we compare the composition of the three samples that we use in our study, and we compare some key statistics to stylized facts in the international trade literature. We also report further detail, such as returns to scale in each country’s production function estimates. Sectoral Composition and Export Orientation The left part of Table A.4 reports the sectoral composition of each sample. While there are some differences, the sectoral composition is overall comparable. In particular, all countries have a large share of firms in food and beverages (Chile being the largest with almost one-third of all plants in this sector). Also, apparel, basic chemicals, and machinery & equipment are important sectors in most countries (with Mexico showing a particularly strong concentration in machinery & equipment). In the right part of Table A.4, we examine the export orientation within sectors Appendix p.17

for each country, i.e., the share of plants that were exporters in a given year in the respective sector. Chile and Colombia have similar export orientation, with about 20-25% of all plants being exporters. Mexico, on the other hand, exports much more of its manufacturing output – about 39%. In particular, metallic manufactures and machinery & equipment are significantly more export-oriented in Mexico than in Chile or Colombia. Table A.4: Sectoral Composition and Export Orientation in the Three Samples Sectoral Composition

Export Orientation

(% of plant-year obs. in each sector)

(% of plant-year obs. exporting)

Chile

Colombia

Mexico

Chile

Colombia

Mexico

32.6

20.9

18.5

20.9

12.7

27.4

Textiles

3.5

4.0

7.9

23.3

33.7

44.2

Apparel

10.9

14.2

8.4

14.1

31.1

30.9

Wood and Furniture

11.2

9.6

3.9

19.6

17.2

27.8

Paper

7.3

7.9

8.2

16.9

23.4

25.1

Basic Chemicals

6.0

9.7

11.1

33.1

30.4

49.6

Plastic and Rubber

6.2

8.2

9.2

22.7

28.7

41.9

Non-Metallic Manufactures

4.4

8.6

7.3

12.7

23.3

26.6

Metallic Manufactures

9.4

5.6

2.5

21.5

18.9

56.1

Machinery and Equipment

8.6

11.5

22.8

18.1

30.4

52.8

Sector Food and beverages

Notes: The left part of the table shows the sectoral composition of the three samples used in our study. The numbers represent the percentage of plant-year observations in each of the sectors. The right part of the table shows the export orientation within each sector, i.e., the percentage of plants that were exporters in a given year in the respective sector.

Returns to Scale As discussed in Section 2.2, constant returns to scale simplifies the interpretation of marginal costs as a measure of productivity. In the following, we discuss the degrees of returns to scale implied by the production function estimates for our Chilean, Colombian, and Mexican samples. Beginning with the Chilean ENIA, Table A.5 shows output elasticities and returns to scale for each 2-digit sector. The sum of output elasticities corresponds to returns to scale. We find that on average, the estimated coefficients point to constant returns to scale. In particular, when weighted by the number of plants in each sector, the sum of output elasticities is 1.009 in our baseline CobbDouglas case. Table A.5 also shows that returns to scale are very similar when we instead estimate a more flexible translog specification.The translog case allows for interactions between inputs, so Appendix p.18

that output elasticities depend on the use of inputs. Consequently, if input use changes after export entry, this could affect elasticities and thus returns to scale. To address this possibility, we compute the average elasticities for 2-digit sectors using i) all plants, and ii) using only export entrants in the first 4 periods after entry. Both imply very similar – approximately constant – returns to scale, as shown in cols 5 and 6 of Table A.5. Tables A.6 and A.7 show the output elasticities for Colombia and Mexico, respectively. In both samples, the sum of output elasticities are very close to 1 in most sectors. Overall, there is some evidence for (very mildly) increasing returns to scale, with an average sum of output coefficients of about 1.04 in both samples.

Appendix p.19

Table A.5: Output Elasticities and Returns to Scale in the Chilean Data Elasticities

Returns

Labor

Capital

All

Entrants

(1)

(2)

(3)

(4)

(5)

(6)

Food and Beverages

0.146 (0.032)

0.139 (0.007)

0.637 (0.017)

0.922 (0.043)

1.078 (0.043)

1.066 (0.052)

Textiles

0.201 (0.177)

0.064 (0.018)

0.598 (0.116)

0.863 (0.165)

1.037 (0.154)

1.000 (0.131)

Apparel

0.313 (0.101)

0.068 (0.035)

0.711 (0.097)

1.092 (0.045)

1.14 (0.188)

1.089 (0.209)

Wood and Furniture

0.381 (0.07)

0.065 (0.017)

0.628 (0.059)

1.075 (0.06)

0.968 (0.127)

0.989 (0.128)

Paper

0.338 (0.053)

0.071 (0.025)

0.65 (0.042)

1.06 (0.024)

0.69 (0.189)

0.63 (0.189)

Basic Chemicals

0.168 (0.07)

0.181 (0.027)

0.754 (0.056)

1.103 (0.046)

0.721 (0.171)

0.712 (0.185)

Plastic and Rubber

0.231 (0.025)

0.116 (0.011)

0.766 (0.023)

1.112 (0.038)

1.037 (0.017)

1.037 (0.019)

Non-Metallic Manufactures

0.141 (0.057)

0.214 (0.035)

0.574 (0.04)

0.929 (0.036)

1.079 (0.148)

1.067 (0.174)

Metallic Manufactures

0.287 (0.045)

0.098 (0.016)

0.655 (0.031)

1.04 (0.061)

1.142 (0.059)

1.115 (0.055)

Machinery and Equipment

0.347 (0.182)

0.085 (0.019)

0.603 (0.062)

1.035 (0.116)

1.069 (0.11)

1.015 (0.087)

0.255 0.243

0.110 0.111

0.658 0.655

1.023 1.009

1.028 1.014

0.988 0.981

Average (Unweighted) Average (Weighted)

Materials to Scale

Translog: RTS

Notes: The table reports the estimated output elasticities for the baseline Cobb-Douglas production function and for a translog production function, estimated for single-product plants within 2-digit sectors. Columns 1-3 display the Cobb-Douglas elasticities with respect to each production factor (labor, capital, and material inputs). In column 4, we report the implied returns to scale, which are equal to the sum of the coefficients in columns 1-3. Columns 5-6 show the average returns to scale for the translog production function. In column 5 we compute the translog returns to scale considering the whole sample of plant-products within each 2-digit sector (trimming percentiles 1 and 99 of the computed elasticities). In column 6 we use the same methodology, but consider only newly exported plant-products in the first four years after export entry. All variables are computed at the plant level. Standard errors are in parenthesis. The weighted average at the bottom of the table uses the number of plants in each sector as weights.

Appendix p.20

Table A.6: Output Elasticities and Returns to Scale in the Colombian Data Elasticities Returns Labor Capital Materials to Scale Food and Beverages

0.209 0.195 (0.091) (0.023)

0.626 (0.084)

1.03 (0.031)

Textiles

0.217 0.141 (0.211) (0.031)

0.62 (0.14)

0.977 (0.086)

Apparel

0.445 0.023 (0.127) (0.022)

0.604 (0.083)

1.071 (0.043)

Wood and Furniture

0.32 0.104 (0.111) (0.034)

0.59 (0.103)

1.014 (0.033)

Paper

0.201 0.197 (0.375) (1.013)

0.585 (0.115)

0.983 (0.93)

Basic Chemicals

0.311 0.141 (0.052) (0.026)

0.597 (0.04)

1.049 (0.032)

Plastic and Rubber

0.245 0.07 (0.075) (0.011)

0.702 (0.047)

1.017 (0.027)

Non Metallic Manufactures

0.337 0.074 (0.598) (0.024)

0.68 (0.252)

1.091 (0.364)

Metallic Manufactures

0.413 0.155 (0.209) (0.077)

0.531 (0.21)

1.099 (0.081)

Machinery and Equipment

0.259 0.053 (2.507) (0.074)

0.718 (1.147)

1.031 (1.288)

0.625 0.629

1.036 1.038

Average (Unweighted) Average (Weighted)

0.296 0.294

0.115 0.115

Notes: The table reports the estimated output elasticities in the Colombian sample. All elasticities are estimated using a Cobb-Douglas production function, estimated for single-product plants within 2-digit sectors. Columns 1-3 display the Cobb-Douglas elasticities with respect to each production factor (labor, capital, and material inputs). In column 4, we report the implied returns to scale, which are equal to the sum of the coefficients in columns 1-3. Standard errors are in parenthesis. The weighted average at the bottom of the table uses the number of plants in each sector as weights.

Appendix p.21

Table A.7: Output Elasticities and Returns to Scale in the Mexican Data Elasticities Returns Labor Capital Materials to Scale Food and Beverages

0.247 0.129 (0.025) (0.016)

0.719 (0.035)

1.095 (0.033)

Textiles

0.423 0.058 (0.061) (0.017)

0.587 (0.069)

1.067 (0.023)

Apparel

0.144 0.000 (0.032) (0.006)

0.862 (0.047)

1.006 (0.02)

Wood and Furniture

0.259 0.007 (0.128) (0.024)

0.779 (0.197)

1.045 (0.098)

Paper

0.317 0.055 (0.018) (0.005)

0.641 (0.026)

1.013 (0.016)

Basic Chemicals

0.259 0.142 (0.026) (0.013)

0.654 (0.031)

1.054 (0.02)

Plastic and Rubber

0.202 (0.032)

0.054 (0.02)

0.749 (0.047)

1.005 (0.036)

Non Metallic Manufactures

0.362 (0.066)

0.079 (0.04)

0.602 (0.081)

1.044 (0.041)

Metallic Manufactures

0.219 0.025 (0.181) (0.034)

0.823 (0.137)

1.068 (0.123)

Machinery and Equipment

0.277 0.069 (0.009) (0.005)

0.685 (0.01)

1.031 (0.006)

0.710 0.698

1.043 1.045

Average (Unweighted) Average (Weighted)

0.271 0.270

0.062 0.076

Notes: The table reports the estimated output elasticities in the Mexican sample. All elasticities are estimated using a Cobb-Douglas production function, estimated for single-product plants within 2-digit sectors. Columns 1-3 display the Cobb-Douglas elasticities with respect to each production factor (labor, capital, and material inputs). In column 4, we report the implied returns to scale, which are equal to the sum of the coefficients in columns 1-3. Standard errors are in parenthesis. The weighted average at the bottom of the table uses the number of plants in each sector as weights.

Appendix p.22

Validity of the Samples Next, we check the validity of the Colombian and Mexican samples, comparing key trade statistics to well-documented stylized facts in the literature. We implement the exercise described in Section 3.3, running regression (11). In other words, we replicate Table 1 from the paper for Colombia (Table A.8) and Mexico (Table A.9). For both countries, we focus on single product plants – the relevant subset of the data for which our results are derived (due to the absence of information on product-specific variable costs in multi-product plants). We focus on the exporter premium – the percentage-point difference of the dependent variable between exporters and non-exporters.20 In each table, Panel A reports the unconditional exporter premium, while Panel B controls for plantlevel employment. The results are similar for both specifications: within their respective sectors, exporting plants are larger both in terms of employment and sales, are more productive (measured by revenue productivity), pay higher wages, and charger higher markups. This is in line with the literature discussed in Section 3.3 in the paper, and thus confirms that the Colombian and Mexican data – just like the Chilean ENIA – provide externally valid samples for our analysis.

20

The underlying regressions control for sector-year effects at the 2-digit level. Thus, the observed coefficients represent differences for exporting plants within their respective sectors.

Appendix p.23

Table A.8: Plant-Level Stylized Facts in Colombian Manufacturing (1)

(2)

Plant Size Dependent Variable

ln(workers)

ln(sales)

(3)

(4)

(5)

Productivity

Wages

Markup

ln(TFPR)

ln(wage)

ln(markup)

Panel A: Unconditional Premia Export dummy

1.121*** (.069)

1.547*** (.132)

.120* (.055)

.467*** (.060)

.0856*** (.025)

Sector-Year FE R2 Observations

X .193 18,178

X .254 18,178

X .398 18,178

X .157 18,178

X .054 17,429

Panel B: Controlling for Employment Export dummy

—

.350*** (.0838)

.146 (.0856)

.110** (.0428)

.0597** (.0261)

Sector-Year FE R2 Observations

— — —

X .698 18,178

X .399 18,178

X .276 18,178

X .056 17,429

Notes: The table reports, for each dependent variable, the percentage-point difference between exporting plants and non-exporters in a panel of 2,250 (1,750 average per year) single-product Colombian plants over the period 2001-2013. All regressions control for sector-year effects at the 2-digit level; the regressions in Panel B also control for the logarithm of employment. Standard errors (in parentheses) are clustered at the plant level. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.24

Table A.9: Plant-Level Stylized Facts in Mexican Manufacturing (1)

(2)

Plant Size Dependent Variable

ln(workers)

ln(sales)

(3)

(4)

(5)

Productivity

Wages

Markup

ln(TFPR)

ln(wage)

ln(markup)

Panel A: Unconditional Premia Export dummy

.776*** (.097)

1.188*** (.172)

.0897*** (.022)

0.273*** (.045)

.0439* (.026)

Sector-Year FE R2 Observations

X .163 16,869

X .203 16,869

X .548 16,869

X 0.253 16,869

X .060 16,056

Panel B: Controlling for Employment Export dummy

—

.332*** (.0926)

.0968** (.0314)

0.166*** (0.0479)

.0572** (.0265)

Sector-Year FE R2 Observations

— — —

X .637 16,869

X .548 16,869

X 0.293 16,869

X .061 16,056

Notes: The table reports, for each dependent variable, the percentage-point difference between exporting plants and non-exporters in a panel of 3,118 (1,825 average per year) single-product Mexican plants over the period 1994-2003. All regressions control for sector-year effects at the 2-digit level; the regressions in Panel B also control for the logarithm of employment. Standard errors (in parentheses) are clustered at the plant level. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.25

C

Additional Results for the Chilean ENIA

This appendix performs a number of additional robustness checks for our main sample – the Chilean ENIA. C.1 Additional Statistics and Plant-Product Results Table A.10 presents the estimated markups at the level of 2-digit industries. For example, a markup of 1.5 implies that prices are 50% above marginal costs. Markups are relatively low for homogenous sectors like ‘food and beverages’ or ‘textile,’ and they are higher for sectors that produce more differentiated products such as ‘plastic and rubber.’ Overall, markups are roughly similar to those estimated by De Loecker et al. (2016).

Table A.10: Estimated Markups in the Chilean Sample Mean Food and Beverages Textiles Apparel Wood and Furniture Paper Basic Chemicals Plastic and Rubber Non-Metallic Manufactures Metallic Manufactures Machinery and Equipment Average

Unweighted

Weighted

Median

1.264 (0.576) 1.398 (0.845) 1.551 (0.786) 1.349 (0.699) 1.737 (0.779) 1.697 (0.916) 1.718 (0.813) 1.433 (0.82) 1.568 (0.878) 1.462 (0.822) 1.454 (0.763)

1.275 (0.608) 1.335 (0.664) 1.575 (0.698) 1.31 (0.555) 1.66 (0.683) 1.434 (0.784) 1.479 (0.582) 1.961 (1.105) 1.684 (1.123) 1.169 (0.655) 1.486 (0.866)

1.131 1.150 1.341 1.173 1.578 1.437 1.509 1.187 1.311 1.240 1.248

Notes: This table reports the estimated markup by aggregate sector for the period 1996-2007 (see Appendix 2.4 for details on the computation). Columns 1 and 2 display the unweighted and sales-weighted average markup, respectively. Standard errors are in parenthesis.

Appendix p.26

Table A.11 reports within-plant trajectories for single-product plants in the Chilean ENIA. We confirm all our main results from Table 2 in the paper. If anything, efficiency increases after export entry are even stronger in single-product plants. Table A.11: Trajectories for Single-Product Plants – Chile Periods After Entry

-2

-1

2

3

Obs/R2

0.0669** (0.0337)

0.1081*** (0.0375)

1,203 0.58

-0.2639** -0.4174*** -0.4253*** (0.1162) (0.1373) (0.1503)

1,203 0.735

0

1

Panel A: Main Outcomes TFPR

-0.0134 0.0134 (0.0271) (0.0257)

0.0497* (0.0258)

Marginal Cost

0.0591 -0.0591 (0.1298) (0.1012)

-0.2403** (0.103)

Markup

-0.017 0.017 (0.5847) (0.5427)

0.0346 (0.242)

Price

0.0421 -0.0421 (0.1289) (0.0989)

-0.2057** (0.0962)

0.0508 (0.0341)

0.028 (0.4629)

0.0482 (0.2871)

-0.2359** -0.3693*** (0.1109) (0.1295)

0.093* (0.0732)

1,203 0.493

-0.3323** (0.1426)

1,203 0.744

Panel B: Additional Efficiency Measures Reported AVC

0.0554 -0.0554 (0.1322) (0.1002)

-0.227** (0.0974)

TFPQ

-0.0559 0.0559 0.2772*** (0.1385) (0.1037) (0.0991)

-0.2409** (0.1141)

-0.387*** (0.1349)

-0.3489** (0.1514)

1,203 0.745

0.2854** (0.1228)

0.413*** (0.1443)

0.4021** (0.1608)

1,203 0.754

Notes: The table reports the coefficient estimates from equation (12), using only single-product plants from the Chilean ENIA manufacturing data. All regressions control for plant-product fixed effects and for 2-digit sector-year fixed effects. Export entry is defined in Section 3.2. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

Table A.12 splits our Chilean sample into sectors with above- and below-median returns to scale (those are reported in Table A.5). Panel A of Table A.12 shows the within-plant trajectories for plant-products in 2-digit sectors with below-median returns to scale, and Panel B, for plants with above-median returns to scale. The trends in marginal costs are not significantly different in the two panels, and declines in marginal costs are actually somewhat stronger in the subset with below-median returns to scale. This makes it very unlikely that our main results are driven by increasing returns to scale. Table A.13 reports within-plant trajectories for the Chilean ENIA by plant size. Panel A restricts the sample to plant-products produced by plants with below-median employment, and Panel B, to plants with above-median employment. We find that the trend of falling marginal costs after export entry is somewhat stronger in smaller plants. Because smaller plants tend to be less producAppendix p.27

Table A.12: Trajectories for Sectors with Above- and Below-Median Returns to Scale – Chile Periods After Entry

-2

-1

0

2

3

Obs/R2

-.3176** (.1374)

-.286** (.14)

1,489 .775

-.1941* (.1177)

1,842 .808

1

Panel A: Below-Median Returns to Scale Marginal Costs

.055 (.1049)

-.055 (.0739)

-.2249** (.0967)

-.2243** (.1113)

Panel B: Above-Median Returns to Scale Marginal Costs

.0139 (.0803)

-.0139 (.0677)

-.0469 (.078)

-.1918** (.0811)

-.1326 (.0896)

Notes: The table reports the coefficient estimates from equation (12) for plant-products from the Chilean ENIA manufacturing data. Panel A restricts the sample to plant-products in sectors with below-median returns to scale, and Panel B, to sectors with above-median returns to scale. Sector-specific estimates for returns to scale in the Chilean sample are listed in Table A.5; the median sectors has returns to scale (sum of output elasticities) of 1.035. All regressions control for plant-product fixed effects and for 4-digit sector-year fixed effects. Export entry is defined in Section 3.2. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

tive,21 this finding is compatible with our results in Table 6 that initially less productive plants see larger efficiency gains after export entry. In other words, in line with Lileeva and Trefler (2010), export-related efficiency gains are larger in initially smaller, less productive plants. Table A.14 reports robustness checks for our propensity score matching exercise. In Section 4.2 in the paper, we used the 5 nearest neighbors. Here, we use the 3 nearest neighbors (Panel A) and the 10 nearest neighbors (Panel B). We find results that are very similar – both in terms of magnitude and significance – to those presented in Table 3 in the paper. As a final check of our plant-product level results, Table A.15 reports the trajectories in TFPR (Panel A), marginal costs (Panel B), and markups (Panel C) for i) our baseline estimation of the production function (Cobb-Douglas in revenues), ii) a translog production function in revenues, and iii) a Cobb-Douglas production function in quantities.22 For comparability, we report our baseline results ("CD Revenue") using only those observations for which TFPR based on translog is between the 5th and the 95th percentile. We trim the distribution in this way to avoid that outliers drive the results in the more flexible translog estimation with plant-product specific output elasticities. The baseline ("CD Revenue") results in this reduced sample confirm our main results from Table 2 both in terms of magnitude and significance. Results are also very similar when we 21

The coefficient of a regression of TFPR on log employment (including 4-digit sector-year fixed effects as in Table 1) is 0.050 with a standard error of 0.012. As discussed in the paper, a comparison of productivity across plants has to rely on TFPR, because output quantities are not comparable across products. 22 For TFPR (Panel A), case iii) is not reported, because by construction, TFPR is based on a production function estimation in revenues.

Appendix p.28

Table A.13: Trajectories for Plants with Above- and Below-Median Employment – Chile Periods After Entry

-2

-1

0

2

3

Obs/R2

-.4202*** (.145)

-.5334*** (.173)

1,652 .726

-.167 (.135)

1,678 .807

1

Panel A. Below-Median Employment Marginal Costs

.0676 (.0798)

-.0676 (.0661)

-.2153** (.101)

-.3313*** (.1175)

Panel B. Above-Median Employment Marginal Costs

.0102 (.0874)

-.0102 (.0575)

-.0488 (.0683)

-.1678* (.0889)

-.1313 (.1126)

Notes: The table reports the coefficient estimates from equation (12) for plant-products from the Chilean ENIA manufacturing data. Panel A restricts the sample to plant-products produced by plants with below-median employment, and Panel B, to plants with above-median employment. The median plant size in the Chilean sample of export entrants is 62 employees. All regressions control for plant-product fixed effects and for 4-digit sector-year fixed effects. Export entry is defined in Section 3.2. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

derive them based on the translog production function. Note that the trajectories for i) and iii) are identical in Panels B and C because in the Cobb-Douglas case, difference in production function coefficients are absorbed by plant-product fixed effects (see Section 2).

Appendix p.29

Table A.14: Matching Results for Chile: Robustness Periods After Entry TFPR

0

1

Panel A: 3 Neighbors -.0149 -.0337 (.0194) (.0239)

2

3

.00322 (.0301)

.104** (.0406)

Marginal Cost

-.0382 (.0341)

-.0828* (.0456)

-.203*** (.0687)

-.309*** (.104)

Markup

.000794 (.0229)

-.0174 (.0264)

.0260 (.0315)

.0728* (.0416)

Price

-.0374 (.0286)

-.100** (.0411)

-.177*** (.0637)

-.237** (.0920)

259 697

178 473

129 342

75 193

.00989 (.0288)

.0854** (.0396)

Treated Observations Control Observations TFPR

Panel B: 10 Neighbors -.0190 -.0233 (.0179) (.0222)

Marginal Cost

-.0618* (.0346)

-.0908** (.0436)

-.174*** (.0654)

-.256*** (.0909)

Markup

.000483 (.0210)

-.00771 (.0232)

.0299 (.0296)

.0481 (.0404)

Price

-.0613** (.0298)

-.0985** (.0408)

-.144** (.0594)

-.208*** (.0734)

260 1,981

180 1,365

130 979

73 495

Treated Observations Control Observations

Notes: This table documents the robustness of the results in Table 3 when changing the number of neighbors in the matching procedure. The benchmark number of neighbors used in Table 3 in the paper is 5. Here, we use 3 neighbors in panel A and 10 neighbors in panel B. Coefficients correspond to the differential growth of each variable with respect to the pre-entry year (t = −1) between entrants and controls. Period t = 0 corresponds to the entry year. See notes to Table 3 in the paper for further detail. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.30

Table A.15: Within Plant-Product Results: Baseline vs. Alternative Production Functions Periods After Entry

-2

-1

0

1

2

3

Obs/R2

Panel A: Revenue Productivity Baseline: - CD Revenue

-.0041 (.0198)

.0041 (.0175)

-.0078 (.019)

.0024 (.0237)

.0306 (.0298)

.0186 (.031)

2,842 .525

Translog

.0173 (.1094)

-.0173 (.0869)

.0519 (.0985)

.0318 (.1311)

.0823 (.1579)

.0108 (.1662)

2,842 .753

Panel B: Marginal Cost Baseline: - CD Revenue

.0388 (.0729)

-.0388 (.0565)

-.1296* (.0705)

-.2014*** (.0775)

-.2424*** (.0922)

-.2902*** (.1078)

2,842 .770

Translog

.0435 (.0756)

-.0435 (.0594)

-.1274* (.0728)

-.1921** (.0789)

-.2058** (.0964)

-.2983*** (.1107)

2,652 .776

CD Quantity

.0388 (.0729)

-.0388 (.0565)

-.1296* (.0705)

-.2014*** (.0775)

-.2424*** (.0922)

-.2902*** (.1078)

2,842 .770

Panel C: Markup Baseline: - CD Revenue

-.0117 (.0233)

.0117 (.019)

-.011 (.0213)

.0038 (.0259)

.0423 (.0322)

.0126 (.0361)

2,842 .459

Translog

-.0103 (.0246)

.0103 (.0174)

-.0072 (.0195)

.0038 (.0234)

.0146 (.0293)

.0159 (.0335)

2,652 .527

CD Quantity

-.0117 (.0233)

.0117 (.019)

-.011 (.0213)

.0038 (.0259)

.0423 (.0322)

.0126 (.0361)

2,842 .459

Notes: This table compares our baseline results (derived from estimating a revenue-based Cobb-Douglas production function) with the cases of a (revenue-based) translog production function and a quantity-based Cobb-Douglas (CD) production function (described in Appendix ??). Panel A shows within-plant-product trajectories for TFPR, Panel B for marginal costs, and Panel C for markups. For comparability, we report our baseline results ("CD Revenue") using only those observations for which TFPR based on translog is between the 5th and the 95th percentile. We trim the distribution in this way to avoid that outliers drive the results in the more flexible translog estimation with plantproduct specific output elasticities. For marginal costs and markups in the translog case, the number of observations is somewhat lower than in the baseline case because in addition to trimming the distribution, we also lose some observations for which the material input expenditure shares are not available. In each panel, the first row replicates the estimates from Table 2 in the paper. Period t = 0 corresponds to the entry year. See notes to Table 2 for further details on the underlying regression and controls. Standard errors (clustered at the plant-product level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.31

C.2 Tariff-Induced Export Entry – Additional Results The results presented in this appendix complement Section 4.4 in the paper. Table A.16 presents the reduced-form results corresponding to our analysis of tariff-induced export entry in Table 5 in the paper. We confirm the pattern documented in our 2SLS analysis: there is a statistically (marginally) significant relationship between export tariffs and export entrants’ efficiency gains captured by marginal costs, reported AVC, and TFPQ. TFPR and markups, on the other hand, are not related to export tariffs. To interpret the magnitude of effects, we report b Outcome", which corresponds to the predicated change in the respective outcome due to the "∆

average decline in export tariffs across all sectors over the sample period. Chilean export tariffs

declined by 5.6% between 1996 and 2007. The estimate in column 2 in Table A.16 shows that his translated into efficiency gains of 13.1% among those plant-products that eventually entered the export market (i.e., 2.333× -5.6%). Table A.16: Tariff-Induced Entry and Plant-Product Outcomes – Reduced Form (1)

(2)

(3)

Main Outcomes MC Markup

(4)

(6)

Additional Outcomes Reported AVC TFPQ

Dependent Variable

TFPR

Export Tariff

-.245 (.547)

2.333* (1.260)

-.186 (.659)

2.147* (1.277)

2.624* (1.321)

-2.183* (1.290)

.014

-.131

.010

-.120

-.147

.122

X X 2,081 .941

X X 2,081 .962

X X 2,081 .650

X X 2,081 .968

X X 2,081 .962

X X 2,081 .961

b Outcome‡ ∆ Controls: Plant-Product FE log Sales Observations R-squared

Price

(5)

Notes: This table presents the reduced-form relationship between export tariffs and the various outcome variables at the plant-product level. The sample includes only plant-products that become new export entrants at some point between 1997 and 2007. Tariffs are observed at the 4-digit ISIC level. All regressions include plant-product fixed-effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to the variation in tariffs. Key: *** significant at 1%; ** 5%; * 10%. ‡ b denotes the predicted change in the corresponding dependent variIn each column of the table, △ able due to export tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

Table A.17 performs a placebo exercise. It first shows that there is no (reduced-form) relationship between export tariffs and domestic sales (col 1). Then, column 2 shows that applying our 2SLS estimation to domestics sales yields a small and insignificant coefficient on (predicted) export entry. This makes it unlikely that our results from Section 4.4 are confounded by broader Appendix p.32

expansions in the domestic market that happen to coincide with both export entry and tariff declines. Table A.17: Tariff-Induced Export Entry and Domestic Sales Dependent Variable: log domestic sales Dependent Variable

(1)

(2)

Reduced Form

2SLS

–

.0484 [.781]

-.469 (1.268)

–

.0262 –

.0176 40.67

X 1,892

X 1,892

Export Dummy weak-IV robust p-value: Export Tariff b Dom.Sales‡ ∆ First Stage F-Statistic Plant FE Observations

Notes: This table examines the relationship between export entry and domestic sales. Dependent variable is the logarithm of plant-level domestic sales. The sample includes only plants that become new export entrants at some point between 1997 and 2007. Column 1 shows the reduced form regression between the logarithm of domestic sales and export tariffs (which vary at the 4-digit ISIC level). The corresponding 2SLS regression is shown in column 2, using export tariffs to instrument for the timing of export entry. Col 2 also reports the (cluster-robust) Kleibergen-Paap rK Wald F-statistic. The corresponding Stock-Yogo value for 10% maximal IV bias is 16.4. Second stage results (col 2) report weak-IV robust Anderson-Rubin p-values in square brackets (see Andrews and Stock, 2005, for a detailed review). Standard errors are clustered at the 4-digit ISIC level, corresponding to the variation in tariffs. Key: *** significant at 1%; ** 5%; * 10%. ‡ b △ denotes the predicted change in domestic sales due to export tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

C.3 Prices and Profit Margins for Domestic and Exported Sales of the Same Product For the period 1996-2000, ENIA reports, for each plant-product, both the domestic price and the export (FOB) price. This enables us to examine the trends in prices separately after export entry.23 Table A.18 presents the results. We find that domestic prices fall mildly, by 7.7% relative to the preentry period(s). On the other hand, the price of the exported fraction of the export-entry product declines significantly, by 19%. Thus, export entrants lower their prices particularly strongly for international customers. This is in line with tougher international competition, but also with the channel of foreign demand building discussed in Section 4.6. Next, we relate the observed pricing behavior of export entrants to the profitability of export entry. Table A.19 sheds light on the relationship between export entry and the profitability of domestic 23

This analysis is subject to the caveat that the detailed information is only available for a short sample, so that we have relatively few export entries, and relatively short periods before and after export entry.

Appendix p.33

Table A.18: Change in Domestic and Exported Prices after Export Entry Dependent Variable: Price of Product when: Exported

Domestic

Export Dummy

-.190** (.0833)

-.0771 (.0985)

R-squared Obs.

.793 1,754

.797 1,754

Notes: The regressions report the average change in exported and domestic prices of new exported products after export entry. The analysis includes only plant-products that are export entrants between 1996 and 2000; for this period, our data contain separate information for exported and domestic product prices. The excluded category in both regressions are plant-products before export entry, or after exit from the export market. All regressions control for plant-product fixed effects (at the 7-digit level), and for product-year effects at the 4-digit level. A product is defined as an entrant if it is the first product exported by a plant and is sold domestically for at least one period before entry into the export market. Standard errors (clustered at the plant-product level) in parentheses. Key: ** significant at 1%; ** 5%; * 10%.

and international sales. In particular, one may wonder whether export entrant products are initially sold at a loss, given the steep decline in export prices documented in Table A.18. To examine this issue, we compute the operating profit margin, which is defined as follows:24 OpMarginijt =

T Rijt − T V Cijt Pijt − AV Cijt = T Rijt Pijt

(A.13)

where T Rijt is total revenue and T V Cijt is total variable cost of product j produced by plant i in year t. In words, the operating profit margin reflects the proportion of the plant-product’s revenue that is left after paying for variable costs of production (labor and material expenditures). Under the assumption that average variable costs are the same for domestic and exported production of the same product, we can compute OpMarginijt separately for the two markets (but only over the period 1996-2000, when ENIA reports the domestic price and the export price separately). Table A.19 first reports the (average) OpMarginijt for the period(s) before plant-product ij is being exported, i.e., when only domestic sales occur. On average, the the operating profit margin is 64.1%, varying between 9.4% in the lowest decile and 157% in the highest decile. Importantly, positive operating profit margins do not necessarily mean that the product is sold profitably after accounting for fixed costs (e.g., interest on capital and depreciation). Column 2 reports the operating margin after export entry, for the domestically sold fraction of the product. We observe a significant shift to the right of the profit margin distribution – margins increase in every decile, and 24

We are grateful to Marcus Asplund for suggesting this analysis.

Appendix p.34

the average rises to 77.9%. This is not surprising and can be illustrated with the findings in Table 2 and Table A.18.25 In the year of export entry, reported AVC fall by about 13% for the plant-product overall (Table 2 for t = 0), while domestic prices fall only by 7.7% (Table A.18). If price falls by less than AV C in equation (A.13), then the operating profit margin increases. Finally, column 3 in Table A.19 shows that after export entry, the operating profit margin declines for the exported fraction of the product. This is in line with the observations that reported AVC fall by about 13%, while export prices fall by 19% (Table A.18). For the lowest decile, the margin even becomes negative, suggesting that some export entrants incur (short-run) operational losses for the newly exported product – although these operational losses are at least in part compensated by operational profits from the domestic sales of the same product. Altogether, our findings suggest that after export entry, plants lower the operational profit margin on the exported product, while the margin increases for domestic sales of the exported product. In other words, price falls by more then AV C for exports, and it falls by less than AV C for domestic sales. This is in line with markups increasing domestically, but falling in the export market (relative to the markup that was charged for domestic sales prior to export entry). This finding is compatible with a linear demand system such as in Melitz and Ottaviano (2008) where i) markups grow if marginal costs fall, and ii) markups decrease in the competitiveness of a market. In this setting, export entry, accompanied by falling marginal costs, can lead to declining markups in the international market and to increasing markups domestically. C.4 Plant-Level Results for the Chilean ENIA In this subsection, we present additional tables that show plant-level results for exporters in the Chilean ENIA. For single-product plants, we use all outcome variables as derived in our standard analysis. For multi-product plants, TFPR can also be estimated directly at the plant level, while we need to construct plant-level indexes for output prices, MC, and markups as described in Appendix B.2. For plant-level TFPQ, we deflate plant-level revenues with the plant-level price index. Calculating the plant-level indexes requires coherent observations in consecutive years for all products in multi-product plants. This restriction implies that the number of observations in the plant-level sample is significantly smaller than in our plant-product level analysis in Table 2 in the paper. For comparability, we compute TFPR in the same sample for which the various indexes are available. Table A.20 presents the plant-level results for our main outcome variables. Despite the smaller sample we confirm our main results – there are only minor increases in TFPR after export entry, but large and significant efficiency gains (in the order of 20%) as reflected by MC and TFPQ. 25

The horizon in this analysis is very short, with export entries occurring between 1997 and 2000 (1996 does not qualify for entries as by the conditions stated in Section 3.2). Thus, we compare the magnitudes in Table A.18 to those immediately after export entry in Table 2.

Appendix p.35

Table A.19: Operating Profit Margin by Export Status of Plant-Products: Chilean Data (1)

(2)

Before Export Entry

(3)

After Export Entry Domestic Sales Exported Sales

Percentile

(domestic sales only)

10th pctile 25th pctile 50th pctile 75th pctile 90th pctile

9.4% 21.7% 30.4% 60.0% 157.1%

16.5% 30.0% 43.6% 89.5% 177.1%

-25.0% 10.8% 32.0% 67.2% 118.2%

Mean

64.1%

77.9%

62.0%

Notes: The table reports the operating profit margin for plant-products that are export entrants and for which prices are reported separately for domestic and international sales. This information is only available over the period 1996 – 2000. The operating margin is the ratio of operating profits (total revenue net of variable costs) over total revenue. Thus, the operating margin reflects the proportion of the plant-product’s revenue that is left after paying for variable costs of production. The operating profit margin reported in the table is computed for a plant-product ij in year t as OpM arginijt = (Pijt − AV Cijt )/Pijt , where AV Cijt is reported plant-product specific average variable cost. "Before export entry" (col 1) reflects the average over all consecutive periods before a plant-product’s export entry (which has to occur between 1996 and 2000), and "After export entry" (cols 2 and 3), reflects the post-export-entry average (again confined to the period 1996-2000). A plant-product is defined as an entrant if it is the first product exported by a plant and is sold domestically for at least one period before entry into the export market.

Next, in Table A.21, we compare plant-level TFPR from our baseline estimation (Cobb-Douglas in revenues) with TFPR based on the more flexible Translog production function. The TFPR trajectories after export entry look very similar for the two approaches. This suggests that the small and mostly insignificant changes in TFPR are not an artefact of using a Cobb-Douglas production function in the baseline.26

26

For some plants with unusual combinations of inputs, the estimated coefficients in the Translog production functions imply negative material elasticities. We exclude these from the analysis. For comparability, we use the same sample for the Cobb-Douglas case.

Appendix p.36

Table A.20: Plant-Level Trajectories for Export Entrant Plants in Chile -2

-1

0

1

2

3

Obs/R2

TFPR

-.0132 (.0216)

.0132 (.019)

.0312 (.0212)

.0313 (.0258)

.0499* (.0285)

.0764** (.0378)

2,267 .703

Marginal Cost

.1024 (.0724)

-.1024* (.0594)

-.2035** (.0786)

-.2377*** (.09)

-.1576 (.1048)

-.2414** (.1198)

2,276 .744

Markup

-.0052 (.0231)

.0052 (.0202)

.0022 (.0244)

.0048 (.0307)

.0264 (.0355)

.0308 (.0425)

2,276 .566

Price

.0856 (.0699)

-.0856 (.0576)

-.1898** (.0735)

-.2269*** (.0846)

-.1284 (.1029)

-.1999* (.1165)

2,276 .779

TFPQ

-.0839 (.0783)

.0839 (.0625)

.2252*** (.078)

.2268** (.0963)

.1382 (.1133)

.2444* (.1298)

2,276 .761

Periods After Entry

Notes: The table reports the coefficient estimates from equation (12). All regressions are run at the plant level; they control for plant fixed effects and 2-digit sector-year fixed effects. A plant is defined as an export entrant if it sells its product(s) domestically for at least one period before entry into the export market, and if it has not exported before during the years observed in our sample. Section 4.1 provides further detail. For comparability, we normalize all coefficients so that the average across the two pre-entry periods (-1 and -2) equals zero. Standard errors (clustered at the plant-product level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%. TFPR = Revenue productivity; TFPQ = Quantity Productivity.

Table A.21: Within-Plant Results for TFPR, and TFPR based on Translog Production Function -2

-1

0

1

2

3

Obs/R2

Baseline (Cobb-Douglas)

-.0121 (.0213)

.0121 (.0189)

.0177 (.0208)

.0262 (.0258)

.0464* (.0275)

.0608* (.0365)

2,143 .714

Translog

-.0401 (.0505)

.0401 (.04)

.0019 (.048)

.0497 (.06)

.0503 (.0704)

.0893* (.0535)

2,143 .801

Periods After Entry

Notes: This table compares the benchmark within-plant results for revenue productivity (derived from a revenue based Cobb-Douglas production function) with the case of a translog production functions. The first row replicates the estimates in Table 2 in the paper for plant-level productivity. Period t = 0 corresponds to the entry year. See notes to Table 2 for further details on the underlying regression and controls. Standard errors (clustered at the plant level) in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.37

C.5 Tariff-Induced Export Expansions of Existing Exporters – Additional Results In this section, we provide additional results for our analysis of tariff-induced export expansions among existing exporters. This complements the results shown in Section 5 in the paper. Table A.22 presents the first-stage results corresponding to the 2SLS analysis in Table 8 in the paper. Table A.22: Tariff-Induced Export Expansions of Exporting Plants in Chile – First Stage

Export Share

(1)

(2)

(3)

(4)

(5)

(6)

>0%

>10%

>20%

>30%

>40%

>50%

Panel A. log Marginal Cost Index Export Tariff b Exports‡ ∆ First Stage F-Statistic Observations

-3.067*** -4.218*** -4.599*** -4.707*** -4.979*** -5.293*** (1.027) (.856) (.990) (1.038) (1.129) (1.534) .172 .236 .258 .264 .279 .296 8.921 24.27 21.59 20.56 19.46 11.91 6,996 4,089 3,257 2,815 2,443 2,137 Panel B. log TFPQ

Export Tariff b Exports‡ ∆ First Stage F-Statistic Observations

-3.021*** -4.208*** -4.600*** -4.709*** -4.983*** -5.293*** (1.022) (.857) (.990) (1.039) (1.130) (1.534) .169 .236 .258 .264 .279 .296 8.746 24.12 21.58 20.55 19.43 11.91 6,988 4,083 3,256 2,814 2,442 2,137 Panel C. log Average Markup

Export Tariff b Exports‡ ∆ First Stage F-Statistic Observations

-3.296*** -4.228*** -4.609*** -4.811*** -5.174*** -5.479*** (1.020) (.842) (.930) (1.018) (1.148) (1.528) .148 .190 .207 .217 .233 .247 10.44 25.19 24.55 22.34 20.31 12.87 9,855 5,744 4,570 3,974 3,454 3,015 Panel D. log TFPR

Export Tariff b Exports‡ ∆ First Stage F-Statistic Observations For all regressions: Plant FE log Domestic Sales

-3.296*** -4.228*** -4.609*** -4.811*** -5.174*** -5.479*** (1.020) (.842) (.930) (1.018) (1.148) (1.528) .185 .237 .258 .269 .290 .307 10.44 25.19 24.55 22.34 20.31 12.87 9,855 5,744 4,570 3,974 3,454 3,015 X X

X X

X X

X X

X X

X X

Notes: This table presents the first stage results corresponding to the 2-SLS regressions in Table 8 in the paper. Controls are the same as in Table 8. Standard errors are clustered at the 4-digit ISIC level, corresponding to the level at which tariffs are observed. Key: *** significant at 1%; ** 5%; * 10%. ‡ b Exports denotes the predicted (percentage) change in plant-level export volume due to In each panel of the table, ∆ export tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

Table A.23 presents the reduced form results corresponding to Table 8, regressing the various Appendix p.38

outcomes directly on export tariffs. The results are very similar to those in Table 8, both in terms b as the predicted change of the sample of statistical significance and magnitude (represented by ∆ period 1996-2007).

Table A.23: Tariff-Induced Export Expansions and Plant-Level Outcomes – Reduced Form

Export Share

(1)

(2)

(3)

(4)

(5)

(6)

>0%

>10%

>20%

>30%

>40%

>50%

4.324*** (1.516) -.294 2,815 .850

4.378*** (1.601) -.282 2,443 .847

4.353** (1.888) -.265 2,137 .850

-3.428** (1.504) .248 2,814 .870

-3.374** (1.510) .240 2,442 .867

-3.319* (1.768) .231 2,137 .866

Panel A. log Marginal Cost Index Export Tariff b MC‡ ∆ Observations R-Squared

2.122** (1.047) -.129 6,996 .841

2.319** (1.133) -.151 4,089 .850

-2.217** (1.008) .134 6,988 .870

-2.187* (1.217) .144 4,083 .878

3.885*** (1.353) -.258 3,257 .858

Panel B. TFPQ Export Tariff b TFPQ‡ ∆ Observations R-Squared

-3.491** (1.453) .242 3,256 .876

Panel C. log Average Markup Export Tariff b Markup‡ ∆ Observations R-Squared

-.0775 (.319) .003 9,855 .645

-.929** (.409) .052 5,744 .630

-1.048*** -1.261*** -1.154*** -.792*** (.342) (.305) (.267) (.263) .062 .080 .072 .051 4,570 3,974 3,454 3,015 .626 .628 .628 .628

Panel D. TFPR Export Tariff b TFPR‡ ∆ Observations R-Squared

For all regressions: Plant FE log Domestic Sales

-.152 (.241) .013 9,855

-.769** (.355) .050 5,744

-.794** (.373) .055 4,570

-.939** (.391) .070 3,974

-.844** (.390) .069 3,454

-.603 (.544) .055 3,015

.944

.934

.922

.919

.909

.903

X X

X X

X X

X X

X X

X X

Notes: This table examines the reduced-form relationship between tariffs and marginal costs (panel A), TFPQ (panel B), markups (panel C), and TFPR (panel D). The corresponding 2SLS results are shown in Table 8. The regressions in columns 1-6 are run for different samples, according to the plants’ export shares: col 1 includes all plants with positive exports, col 2 those whose exports account for more than 10% of total sales, col 3, 20%, and so on. Export tariffs vary at the 4-digit ISIC level. For multi-product plants, the dependent variables in panels B, C, and D reflect the product-sales-weighted average, as described in Appendix B.2. All regressions control for the logarithm of plant-level domestic sales and include plant fixed effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to the level at which tariffs are observed. Key: *** significant at 1%; ** 5%; * 10%. ‡ b denotes the predicted change in the corresponding dependent variable due to export In each panel of the table, △ tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

Appendix p.39

Table A.24 applies our tariff-based 2SLS analysis to domestic sales. The table shows that there is no relationship between predicted export expansions and domestic sales. This makes it unlikely that our results are confounded by tariff declines and increasing exports that coincide with domestic sales expansions. Table A.24: Existing Exporters: Impact of Tariff Declines on Domestic Sales, 2SLS (1)

(2)

(3)

(4)

(5)

(6)

Export Share

>0%

>10%

>20%

>30%

>40%

>50%

Export Tariff

-3.270** (1.266) .183 .916

-4.307*** (1.126) .241 .950

-4.982*** (1.197) .279 .954

-5.483*** (1.273) .307 .957

-5.821*** (1.601) .326 .959

Panel A. log Exports

b Exports‡ ∆ R-Squared Export Tariff b Domestic Sales‡ ∆ R-Squared

For all regressions: Plant FE Observations

-4.849*** (1.115) .272 .951

Panel B. log Domestic Sales .320 (2.461) -.018 .799

.941 (5.220) -.053 .774

-.170 (5.212) .010 .769

-.481 (6.085) .027 .767

-.838 (6.917) .047 .765

-3.570 (6.944) .200 .766

X 6,719

X 3,997

X 3,225

X 2,820

X 2,475

X 2,196

Notes: This table examines the reduced-form relationship between export tariffs and exports (panel A) and domestic sales (panel B). Panel A shows that the strong negative relationship between export tariffs and log exports also holds when not controlling for domestic sales (which is included in our baseline analysis in Section 5. Panel B shows that, on the other hand, there is no relationship between export tariffs and domestic sales. Tariffs are observed at the 4-digit ISIC level. The regressions in columns 1-6 are run for different samples, according to the plants’ export shares: col 1 includes all plants with positive exports, col 2 those whose exports account for more than 10% of total sales, col 3, 20%, and so on. All regressions include plant fixed effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to the variation in tariffs. Key: *** significant at 1%; ** 5%; * 10%. ‡ b denotes the predicted change in the corresponding dependent variable due to export In each panel of the table, △ tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

Table A.25 examines the relationship between tariff-induced export expansions and input prices. For each plant, we compute the input-cost-weighted average input price index, adapting the methodology described in Appendix B.2 to the various plant-specific inputs.27 The table shows that there is no relationship between predicted export expansions and the price index of inputs used by ex27

When computing the input price index, we use deflators corresponding to input-producing industries (at the 4digit level) to deflate input prices. This avoids that differential price trends across industries, or price trends over time, affect our results. ENIA reports deflators corresponding to each input for the period 1996-2005; but for 2006-07, these are not available. Thus, the analysis in Table A.25 is restricted to the period 1996-2005. Nevertheless, a plant-specific deflator for the compound of all material inputs is available for 2006-07 as well – this allows us to deflate plant-level input use for the computation of TFPR and TFPQ over the whole sample period 1996-2007 (but it does not allow us to construct the input price index following the methodology from Appendix B.2).

Appendix p.40

porting plants. This suggests that the observed decline in marginal costs in Table 8 in the paper are unlikely to be driven by decreasing input prices. Table A.25: Existing Chilean Exporters: Tariff-Induced Export Expansions and Input Prices, 2SLS

Export Share

(1)

(2)

(3)

(4)

(5)

(6)

>0%

>10%

>20%

>30%

>40%

>50%

-4.109*** (1.199) .189 11.75

-4.504*** (1.292) .207 12.15

-4.695*** (1.371) .216 11.73

Panel A: First Stage Export Tariff b Exports‡ ∆ First Stage F-Statistic log Exports (predicted) b Input Price Index‡ ∆

For all regressions: Plant FE Observations

-1.598 (1.275) .074 1.571

-3.123** (1.216) .144 6.598

-4.097*** (1.091) .188 14.09

Panel B: Second Stage, Input Price Index 1.801 [.001] .132

.306 [.292] .044

.253 [.34] .048

.122 [.668] .023

.0718 [.787] .015

.127 [.628] .027

X 4,005

X 2,254

X 1,787

X 1,547

X 1,339

X 1,160

Notes: This table examines the effect of within-plant export expansions due to falling export tariffs on input prices (panel B). The dependent variable in panel B reflects the input-cost-weighted average input price index, adapting the methodology described in Appendix B.2 to the various plant-specific inputs. The first stage results of these 2SLS regressions are reported in panel A, together with the (cluster-robust) Kleibergen-Paap rK Wald F-statistic. The corresponding Stock-Yogo value for 10% maximal IV bias is 16.4. Second stage results report weak-IV robust AndersonRubin p-values in square brackets (see Andrews and Stock, 2005, for a detailed review). Tariffs are observed at the 4-digit ISIC level. The regressions in columns 1-6 are run for different samples, according to the plants’ export shares: col 1 includes all plants with positive exports, col 2 those whose exports account for more than 10% of total sales, col 3, 20%, and so on. All regressions are run at the plant level and include plant fixed effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to the variation in tariffs. Key: *** significant at 1%; ** 5%; * 10%. ‡ b denotes the predicted change in the corresponding dependent variable due to export In each panel of the table, △ tariff reductions over the sample period for which the input price index can be computed: 1996-2005 (see footnote 27 in the appendix). Tariffs declined by 4.6 p.p. on average (sales-weighted) over this period.

Appendix p.41

Finally, Table A.26 shows that there is a strong relationship between tariff-induced export expansions and the plant’s capital stock. This is in line with the exporting-investment complementarity mechanism discussed in Section 4.5 in the paper. Table A.26: Existing Exporters: Tariff-Induced Export Expansions and Capital Stock, 2SLS Panel A. First Stage Export Tariff b Exports‡ ∆ First Stage F-Statistic Export Tariff b Capital Stock‡ ∆

For all regressions: Plant FE Observations

-3.271*** (1.114) .183 8.626

-4.305*** (.974) .241 19.55

-4.845*** (.968) .271 25.06

-4.976*** (1.040) .279 22.87

-5.472*** (1.108) .306 24.41

-5.809*** (1.387) .325 17.55

Panel B. Second Stage, log Capital Stock .538* [.0729] .099

.843** [.007] .203

.979*** [.0003] .266

.990*** [.0004] .276

.975*** [.0001] .299

1.171*** [.0001] .381

X 6,722

X 4,000

X 3,228

X 2,823

X 2,478

X 2,199

Notes: This table examines the effect of within-plant export expansions due to falling export tariffs on plant-level capital stock (panel B). The first stage results of these 2SLS regressions are reported in panel A, together with the (cluster-robust) Kleibergen-Paap rK Wald F-statistic. The corresponding Stock-Yogo value for 10% maximal IV bias is 16.4. Second stage results report weak-IV robust Anderson-Rubin p-values in square brackets (see Andrews and Stock, 2005, for a detailed review). Tariffs are observed at the 4-digit ISIC level. The regressions in columns 1-6 are run for different samples, according to the plants’ export shares: col 1 includes all plants with positive exports, col 2 those whose exports account for more than 10% of total sales, col 3, 20%, and so on. All regressions are run at the plant level and include plant fixed effects. Standard errors are clustered at the 4-digit ISIC level, corresponding to the variation in tariffs. Key: *** significant at 1%; ** 5%; * 10%. ‡ b denotes the predicted change in the corresponding dependent variable due to export In each panel of the table, △ tariff reductions over the sample period (tariffs declined by 5.6 p.p. on average (sales-weighted) in 1996-2007).

Appendix p.42

D

Additional Results for the Colombian and Mexican Samples

This appendix presents the investment trends around export entry in the Colombian and the Mexican samples in Tables A.27 and A.28, respectively. In both cases we find trends that are very similar to those in for Chilean export entrants: investment spikes for ‘young’ exporters, and this pattern is driven largely by investment in machinery. This supports our interpretation that an investmentexporting complementarity is a likely driver of the observed efficiency gains after export entry. Table A.27: Colombia: Investment Trends Before and After Export Entry ‘Young’ Exp. ‘Old’ Exp. Obs./R2

Period:

Pre-entry

Overall

-.230 (.343)

.545* (.325)

.278 (.439)

1,609 .417

Machinery

-.00482 (.325)

.766** (.353)

.0875 (.495)

1,609 .396

Vehicles

-.489* (.290)

-.197 (.337)

-.453 (.399)

1,609 .243

Structures

.0254 (.319)

.0956 (.361)

-.428 (.466)

1,609 .339

Notes: This table analyzes investment trends before and after export entry in Colombian single-product plants (investment trends are very similar for the full sample that includes multi-product plants – available upon request). All dependent variables are in logs, and all regressions include fixed effects; thus, coefficients reflect the percentage change in investment in each respective period relative to the average across all periods. ‘Old Exp.’ groups all periods beyond 2 years after export entry; ‘Young Exp.’ comprises export periods within 2 years or less after export entry; and ‘Pre-Entry’ groups the two periods before entry. Regressions are run for single-product plants and control for plant sales, plant fixed effects, and sector-year effects (at the 2-digit level). The criteria for defining a plant as export entrant are described in Appendix B.5. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.43

Table A.28: Mexico: Investment Trends Before and After Export Entry ‘Young’ Exp. ‘Old’ Exp. Obs./R2

Period:

Pre-entry

Overall

.0842 (.172)

.447** (.176)

.158 (.199)

3,089 .488

Machinery

.0173 (.185)

.373* (.198)

.256 (.227)

3,089 .446

Vehicles

-.176 (.148)

-.0775 (.161)

-.120 (.198)

3,089 .342

Structures

-.0706 (.143)

.200 (.154)

-.0765 (.166)

3,089 .398

Notes: This table analyzes investment trends before and after export entry in Mexican single-product plants (investment trends are very similar for the full sample that includes multi-product plants – available upon request). All dependent variables are in logs, and all regressions include fixed effects; thus, coefficients reflect the percentage change in investment in each respective period relative to the average across all periods. ‘Old Exp.’ groups all periods beyond 2 years after export entry; ‘Young Exp.’ comprises export periods within 2 years or less after export entry; and ‘Pre-Entry’ groups the two periods before entry. Regressions are run for single-product plants and control for plant sales, plant fixed effects, and sector-year effects (at the 2-digit level). The criteria for defining a plant as export entrant are described in Appendix B.5. Robust standard errors in parentheses. Key: *** significant at 1%; ** 5%; * 10%.

Appendix p.44

References Andrews, D. W. K. and J. H. Stock (2005). Inference with Weak Instruments. NBER Working Paper 313. De Loecker, J. (2007). Do exports generate higher productivity? Evidence from Slovenia. Journal of International Economics 73(1), 69–98. De Loecker, J., P. K. Goldberg, A. K. Khandelwal, and N. Pavcnik (2016). Prices, Markups and Trade Reform. Econometrica 84(2), 445–510. De Loecker, J. and F. Warzynski (2012). Markups and Firm-Level Export Status. American Economic Review 102(6), 2437–2471. Foster, L., J. Haltiwanger, and C. Syverson (2008). Reallocation, Firm Turnover, and Efficiency: Selection on Productivity or Profitability? American Economic Review 98(1), 394–425. Iacovone, L. and B. S. Javorcik (2010). Multi-Product Exporters: Product Churning, Uncertainty and Export Discoveries. Economic Journal 120(5), 481–499. Klette, T. J. and Z. Griliches (1996). The Inconsistency of Common Scale Estimators when Output Prices are Unobserved and Endogenous. Journal of Applied Econometrics 11(4), 343–61. Kugler, M. and E. Verhoogen (2012). Prices, Plant Size, and Product Quality. Review of Economic Studies 79(1), 307–339. Lileeva, A. and D. Trefler (2010). Improved Access to Foreign Markets Raises Plant-Level Productivity...for Some Plants. Quarterly Journal of Economics 125(3), 1051–1099. Mas-Colell, A., M. D. Whinston, and J. R. Green (1995). Microeconomic Theory. Oxford University Press, USA. Melitz, M. J. and G. I. P. Ottaviano (2008). Market Size, Trade, and Productivity. Review of Economic Studies 75(1), 295–316. Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press.

Appendix p.45