Endogenous markups, international trade and the product mix

Transcription

1 Endogenous markups, international trade and the product mix Carlo Altomonte y Università Bocconi Alessandro Barattieri Boston College August 7, 2007 Abstract We investigate the e ects of import penetration on the estimated price-cost margins of more than 28,000 rms operating in the Italian manufacturing sector, controlling for the potential endogeneity of the productivity shock. In the period considered ( ), we nd broad evidence of pro-competitive gains from trade at the aggregate level. However, when performing the same analysis at a more detailed industry level, we nd a substantial heterogeneity of responses: in some industries the increased exposure to international trade is associated with higher, rather than lower, markups, while in others the relationship is not signi cant. In particular, the industries in which we nd a positive impact of import penetration on markups exhibit, on average, a larger variation in the composition of their product-mix, and decreasing productivity. We show how these features can be nested into existing theoretical models of rms heterogeneity. JEL classi cation: F15, L11 Keywords: price cost margins, trade openness, rms heterogeneity Corresponding author: IEP-Università Bocconi, Via Gobbi, 5, I Milan. (T) ; (F) ; carlo.altomonte@unibocconi.it y Acknowledgements: We wish to thank seminar participants at the 8th ETSG meeting (Vienna), the 2007 EIIE Conference (Ljubljana), the 2007 CNR-CEIS meeting (Turin), LICOS (KU Leuven), CEU (Budapest), CESPRI (Bocconi University) and Bologna University for very helpful comments and suggestions. The usual disclaimer applies. 1

2 1 Introduction "More than half of the surveyed manufacturing rms have changed strategy in the last ve years. The 12 per cent of rms who have switched products in new industries have generated pro ts higher than the average." [translation from Mario Draghi, Bank of Italy Governor, Annual Relation, 31 May 2007, p. 8] The latter quotation is just another example of the nowadays widespread debate on the reactions of rms s strategies to changes in the business environment, especially as the result of an increased exposure to international competition. In particular, a survey on a balanced sample of some 3,100 manufacturing rms undertaken by Bank of Italy (2007) 1 shows that only 27 per cent of the surveyed rms see themselves as having an advantage with respect to competitors. More speci cally, international competition especially from China or other low-cost countries (South East Asia or Central and Eastern Europe) is seen as a major source of potential weakness for the rm. In this context, within the period more than 50 per cent of the surveyed rms have changed strategies: 7 per cent of rms have internationalized their activities, 15 per cent have increased investment in core products, while 31 per cent of the surveyed rms have changed the range of products produced. Within the latter group, 88 per cent of rms have changed products within the same sub-sector, while the remaining 12 per cent have switched production to contiguous (10 per cent) or totally di erent (2 per cent) industries, experiencing as a results higher than average pro ts. Most of these results are hardly surprising to the international economic literature (see for example the survey by Tybout, 2003), with the e ects of trade liberalization on average price-cost margins, exports, rm sizes, productivity and net entry dynamics having been thoroughly explored across industries and countries. More recently, starting from the seminal works of Bernard, Eaton, Jensen and Kortum (2003) and Melitz (2003), new-new trade models have also explicitly taken into account rm heterogeneity 2. Melitz and Ottaviano (2005) have combined the supply-side features of the Melitz s (2003) model of rm heterogeneity with a demand system di erent than the traditional CES demand function, thus adding the dimension of heterogeneous markups in imperfect competition models of trade 3. The further dimension of product heterogeneity, which seems relevant in the above quoted example, has also started to be explored: based on US data, Bernard et al. (2006a,b) show in fact that some rms might react to international competition endogenously self-selecting into the production of a di erent product mix composed of asymmetric products, each one developed according to a di erent technology. Capitalizing on these ndings, the research has now started to analyze the interactions between the established results of the literature and the new features of heterogeneity discovered. 1 Every year Bank of Italy, the Italian central bank, runs a survey (Indagine sulle imprese industriali e dei servizi) based on questionnaires erogated to a balanced sample of more than 4,000 manufacturing and services enterprises. The 2007 issue included speci c questions on the international positioning of rms and their eventual changes in strategy in the period A survey of this emerging literature is provided by Baldwin (2005). 3 In their framework, monopolistically competitive rms produce one variety of a single product with heterogeneous productivity levels; markups, rather than being driven exogenously by the distribution of rms marginal costs, are endogenous over the di erent product varieties, depending among others on the toughness of competition across countries or industries and hence on the exposure to international trade. 2

3 In particular, in the previously quoted survey some rms are actually able to increase pro ts via product reallocation, notwithstanding an increase of competitive pressures. The channels through which these e ects might take place are however relatively unexplored. In this paper we will thus try to combine the feature of endogenous markups with the dimension of product heterogeneity, in order to explore whether the interaction of these two elements, in a context of trade liberalization, might yield a richer set of results on the competitive e ects of trade which is consistent with the detected empirical evidence. From a methodological point of view, apart from the previously quoted literature, the analysis capitalizes on Konings et al. (2005a and 2005b), who have re ned an algorithm allowing to consistently measure average price-cost margins starting from balance-sheet, rm-level observations. The algorithm overcomes the traditional critique of the Hall (1988) type of approach for estimating markups, i.e. a potential simultaneity bias between output growth and the growth in the input factors 4. It also avoids relying on imperfect measures of rms marginal costs in order to observe rms markups, since price-cost margins can be estimated consistently starting from nominal balance sheets data on sales and input factors 5. Following the latter procedure, we have estimated price-cost margins for a sample of roughly 28,000 rms operating in the Italian manufacturing sector over the period In line with the standard results of the literature, we have found broad evidence of pro-competitive gains from trade at the aggregate level, i.e. rms markups tend to be negatively associated with an increase of import penetration indexes. However, when performing the same exercise at a more disaggregated level, the analysis has revealed a huge variation of responses: in some industries (among which textiles), the standard pro-competitive result is reverted, with an increased exposure to international trade associated with higher price-cost margins. For a third group of industries, the e ect does not seem to be signi cant. We have then tried to relate this evidence to some structural characteristics of industries. No signi cant di erences which can explain the anti-competitive e ect of trade have emerged across the concerned industry groups, with one exception: sectors displaying a positive impact of import penetration on price-cost margins exhibit, on average for the industry, a di erent variation in the composition of their product-mix (dispersion rather than concentration). Moreover, we also nd that the same sectors are characterized by a relatively decreasing productivity. These results allow us to contribute to the existing literature in a number of ways. First, we provide some evidence consistent with the theoretical results of Bernard et al. (2006b) and Eckel and Neary (2006), showing that in the majority of Italian manufacturing industries the impact of an increasing import penetration has led over time to lower markups and a tendency to concentrate the range of products produced. Second, we provide evidence, for given industries, of a positive correlation between import penetration and markups associated to product dispersion, rather than concentration. We link 4 The re nement is originally due to Roeger (1995), who overcomes the problem by subtracting the dual Solow residual from the primal, thus being able to eliminate the unobserved productivity shock, source of the bias, from the estimating equation. Konings et al. (2005a and 2005b) exploit the algorithm in order to estimate, respectively, the e ects of anti-dumping protection and of changes in the corporate governance on domestic rms markups. 5 A common approach is to use the observed rm level price-cost margin, de ned as sales net of labor and material costs over sales. The latter implies that labor and material costs are good proxies of a rms short-term marginal costs. We will use the latter measure as a robustness check of our result. See Tybout (2003) for an overview of markup estimation with rm-level data. 3

4 this result to another phenomenon detected at the empirical level by Bernard et al. (2006a) for the US: rms might react to increased international competitive pressures by endogenously switching their product mix towards products characterized by lower elasticity of demand. The evidence for Italy shows that such an endogenous switch does not necessarily lead to a concentration of products, but rather, in given industries, to a higher product dispersion. The novel implication of such a defensive strategy is that changes in product mix induced by increased trade pressures might have a positive impact on rms markups, thus reverting in speci c industries the traditional nding of pro-competitive e ects of trade. Third, since in these industries we detect a negative impact on productivity, we discuss how the implications for productivity of product mix choices can be nested within the predictions of the Melitz and Ottaviano (2005) model: in their setup, markups and productivity both depend on the cuto cost level of the rm who is just indi erent about remaining in the industry, and the parameter measuring the dispersion of the underlying Pareto distribution of cost draws. If, ceteris paribus, the new product mix entails a higher dispersion of costs, it is relatively straightforward to show that the latter e ect leads to a higher cuto, higher average markups and lower average productivity for the considered industry 6. The structure of the paper is as follows: Section 2 presents in detail our methodological framework, relating a proper estimation of markups to the e ects of trade penetration. Section 3 discusses the dataset and its validation with respect to o cial data, while Section 4 presents our industry-speci c results on the relation between import penetration and price-cost margins, and the relative robustness checks. Section 5 discusses in some more detail our product mix hypothesis, nesting it into the existing literature, while Section 6 concludes. 2 Methodological framework Our methodology is the same introduced by Roeger (1995), who built on the work of Hall (1988). More recently, the methodology has been used also by Gorg and Warzynski (2003) and Konings et al. (2005a and 2005b). All these authors start from a standard production function: Q it = A it F (N it; M it; K it ) (1) where Q it is the output of rm i at time t, N it; M it and K it are the labour, material and capital inputs and A it is the rm s productivity. Starting from an expression for the marginal cost in presence of technical progress analogous to the speci cation used by Hall (1988), it is possible to express the output growth rate as follows: dq it Q it = P nn it dn it + P kk it dk it + P mm it dm it + g it (2) c it Q it N it c it Q it K it c it Q it M it where g it is the productivity growth, c it is the marginal cost and P J (with J = N; M; K) is the unit cost of input factor J. The weights, hence, are the shares of each input in total costs. Since under constant return to scale (CRS) the cost shares sum to one, it is possible to rewrite 6 Bernard et al. (2005) also show that a change in the product mix can be associated to a decrease in rm-level productivity. 4

5 equation (2) as: dq it Q it dk it K it = P nn it dnit c it Q it N it dk it + P mm it dmit K it c it Q it M it dk it + g it (3) K it If now we introduce imperfect competition, with a mark-up over marginal cost it = Pit c it, equation (3) may be written as: dq it Q it dk it K it = it [ Nit ( dn it N it dk it K it ) + Mit ( dm it M it dk it K it )] + g it (4) where now the are shares in the value of production. If we now divide both sides by it = 1 1 it and rearrange, we get: dq it Q it Nit dn it N it dm it Mit (1 Nit Mit ) dk it dqit = M it K it it Q it with the expression now written in terms of it = Pit (PCM) of rm i at time t 7. cit P it dk it + (1 K it )g it (5) it, the Lerner Index or price-cost margin Equation (5) thus decomposes the Solow residual into two terms: a pure technology component g it and a markup factor (1 it ). The problem in estimating equations (3) or (5) as in Levinsohn (1993) is that unobserved productivity shocks g it may be correlated with the input factors. The latter is the traditional critique to the Hall s (1998) approach for estimating markups, which is di cult to overcome since instrumental variables are hard to nd at the rm-level. However, the potential endogeneity of the error term can be overcome following Roeger (1995), who is able to decompose the price-based (or dual) Solow residual according to the following expression, comparable to equation (5): Nit dp Nit P Nit + Mit dp Mit P Mit +(1 Nit Mit ) dp Kit P Kit dp it dpit = P it it P it dp Kit +(1 P it )g it (6) Kit where P J (with J = N; M; K) is again the unit cost of input factor J and the are again shares in value of production. Konings et al. (2005a and 2005b) subtract eq. (6) from eq. (5), ending up with: dqit Q it + dp it P it Nit dnit N it + dp Nit P Nit Mit dmit M it + dp Mit P Mit dkit (1 Nit Mit ) + dp Kit K it P Kit = it dqit Q it + dp it P it dkit K it + dp Kit P Kit (7) In equation (7) the unobserved productivity shock g it is canceled out and therefore the simultaneity bias previously discussed disappear. The Lerner index can thus be estimated consistently. Moreover, equation (7) implies that estimating the price-cost margin requires information about the growth rates of production value, wage bill, material costs and the value of capital. Since no 7 In the remaining of the paper we will use indi erently the terms markup and price-cost margin, although, given our econometric speci cation, we will be referring to the latter. 5

6 de ation is required, also the omitted price variable bias is not a source of trouble 8. As for the rental price of capital P Kit, following Hsieh (2002) and Konings (2005b), it can be computed as P Kit = P I (r it + it ), where P I is an investment good price index retrieved from the EU AMECO database, it is a rm-level depreciation rate computed as depreciation over net tangible xed asset in the previous year and r it is the rm-level real interest rate, an information retrieved from our dataset 9. Konings et al. (2005a and 2005b) label the LHS of eq. (7) as DY and the RHS as DX, and thus obtain a very simple testable equation for estimating the price-cost margins: DY it = it DX it + it (8) A potential shortcoming of this approach is that, in order to estimate eq.(8), one has to assume constant markups over the group of rms considered, since without this assumption it would not be possible to have enough degrees of freedom for the regressions. Although the assumption is rather common in the literature (e.g. Levinsohn, 1993 or Konings et al., 2005b), we will in any case check the robustness of this assumption by comparing the markups so obtained with the "observed" rm-speci c markups inferred from balance sheet data. Since we are interested in assessing the evolution over time of the price cost margins in order to gauge the impact of trade openness, we have modi ed eq. (8) as follows: DY ijt = 1 DX ijt + t DX ijt T t + 2 DX ijt IMP jt + i + ijt (9) In equation (9) the dimension j represents the industry to which the rm i belongs at time t, T t is a set of time dummies which allow us to control for cyclical demand e ects 10, while IMP jt measures the import penetration index in industry j at time t calculated as: IMP jt = IMP ORT S jt IMP ORT S jt + P RODUCT ION jt EXP ORT S jt (10) i.e. the total imports, exports and production value of industry j at time t. Finally, i stands for an unobservable rm-speci c xed e ect. The coe cient ^ 2 in equation (9) therefore captures the marginal impact of import penetration on the PCM s estimates, with the latter retrievable through the sequence of the coe cients (^ 1 +^ t ), where ^ 1 is the estimated PCM in the rst period for the group of rms considered and ^ t are the coe cient of the time dummies 11. Before turning to the description of the dataset and the results obtained, it is necessary to stress some caveats that should be taken in mind when considering this analysis. The rst one is related to eq.(8), in which, in principle, the error term should not appear. However, Roeger 8 Tybout (2003) points out that, lacking speci c information on rms prices, which is common, it could be the case that rms that rapidly increase their inputs tend to drive down their output prices more rapidly than the industry averages, yielding a downward bias in the estimated markups. Klette and Griliches (1996) have been the rst to discuss a similar omitted price variable bias in production functions estimation. De Loecker (2007) discusses the problem within semi-parametric estimations of TFP. 9 The rm-level real interest rate is retrieved by subtracting the CPI variation from the reported balance sheet item "interest rate paid". 10 Shapiro (1987) argues that the primal and the dual Solow residual might be a ected di erently by the state of demand, yielding a non-zero error term. The introduction of time dummies controls for this potential source of inconsistency in our estimates. 11 The PCM in each year t is thus retrieved as ^ 1 + ^ t + ^ 2 IMP _P EN t, the average of the import penetration index across the considered industries. 6

7 (1995) clari es that although a variety of reasons justify the presence of it, in particular possible measurement errors in the variables employed, the latter should not a ect the consistency of the estimates, thus allowing for the implementation of the model. The second criticism that may arise is related to the maintained assumption of constant returns to scale. As discussed by Konings et al. (2005b), not allowing for varying returns to scale may generate an upward or downward bias in the markup levels, depending on whether returns to scale are respectively decreasing or increasing. In addition, in order to estimate eq.(8) one has to assume constant markups over the group of rms considered, an assumption typical of all applications of this type (Levinsohn, 1993 or Konings et al., 2005b), since without this assumption it would not be possible to have enough degrees of freedom in the estimating equation. However, to the extent that returns to scale do not change dramatically over the sample period, the latter bias, if present, can be considered as relatively constant, and thus should not a ect the validity of our results, since we are interested in the variation over time of the markup, rather than its point estimate. Moreover, as a robustness check, we will compare our estimated time-varying markups common for all industries with weighted industry averages of time-varying, rm speci c markups, inferred from individual balance sheet data. Finally, market-power might be product rather than rm-speci c, while we base our estimates on rm-level data. Taking into account this potential criticism implies interpreting our results as the impact of import penetration on the average PCM of the group of rms considered. 3 Data description 3.1 Import penetration indexes for the Italian manufacturing industries In order to compute import penetration indexes according to eq.(10), we need information on trade ows and production at the industry level. As for imports and exports, the Italian National Institute for Statistics (ISTAT) provides the value of import and export at detailed industry level according to the NACE Rev 1.1 classi cation for several years. Data on production are instead collected from EUROSTAT, whose detailed industrial statistics database reports several variables (such as value of production, value added, employment) for the same industries, with a year coverage ranging from 1996 to 2003, which therefore constitutes our period of reference. Table 1 reports some descriptive statistics on the calculated import penetration ratios at the NACE2-digit level of aggregation, while Graph 1 displays the dynamics in the di erent industries 12. [Table 1 and Graph 1 about here] The analysis reveals an ample heterogeneity in the exposure of each industry to international trade ows, with average import penetration ratios ranging from 57.2% in sector 34 to 3.8% in sector 22. Even within each NACE2 manufacturing industry the import penetration ratios might di er a lot when calculated at the NACE3 level of aggregation, as it can be seen looking at the standard deviation of the indexes. 12 From here on we will present the results at this level of aggregation, although in the regressions we use the more disaggregated NACE3 level data. Descriptive statistics for more detailed levels of disaggregation are available upon request. 7

8 As for the evolution over time of the import penetration ratios, we nd a general upward trend, from 19% in 1996 to 24% in 2003, in line with the increasing exposure of the county to international trade ows. At the industry level, however, the growth rate of import penetration displays some heterogeneity, with clearly upward trends in some industries vs. a more or less constant exposure in others. 3.2 The sample of italian manufacturing rms A commercial dataset called AIDA, collected by the Bureau van Dijk, was used in order to retrieve rm level information about production value, material costs, cost of employees, value added, tangible xed asset, depreciation, interest paid over debt and employment. The total sample was made up by 61,335 rms, classi ed in each NACE3 or 4 industry 13. Taking 2001 as reference year and comparing the sample data with the 2001 Industrial Census, these rms accounted for the 73% of total manufacturing value added and the 54% of manufacturing employment. However, due to the quality of data, extensive data cleaning has been necessary in order to apply the methodology previously introduced. We adopted a multi-stage data cleaning procedure. First of all, we concentrated on those rms for which information was available for every variable of interest in at least one year. After having calculated the growth rate of each input variable, we controlled for possible outliers by dropping all those rms for which any percentage variation was larger than 200%. We then computed the cost shares Nit, Mit and (by di erence) Kit and dropped from the analysis those rms with shares belonging to the 1 st or to the 99 th percentile of the relevant distributions. After these steps, the resulting sample was almost halved to 28,076 rms, which are those employed in the analysis. As for the validation of the cleaned sample, these rms account for 34.6% of total Italian manufacturing value added and for 25.8% of total manufacturing employment. We then checked the representativeness of the sample along three dimensions: geographical location, industrial activity and rms size. Table 2 reports the distribution across regions of the rms included in the sample. The number of rms in each Region ranges from 33 (in Valle d Aosta) to 8,128 (in Lombardy). Comparing this distribution with the distribution registered during the 2001 Industrial Census the correlation obtained is 0.96, signi cant at the 1 per cent level. [Table 2, 3 and 4 about here] Table 3 shows instead the distribution of the cleaned sample across the NACE2 industries. Due to lack of su cient observations, we had to drop NACE2 industries 16 (Tobacco), 23 (Petroleum) and 30 (O ce machinery). As Table 3 shows, the number of sample rms in each industry ranges from 379 in sector 35 (Other transport equipment) to 4,259 in sector 29 (Machinery and equipment). The correlation between the sample s distribution and the one of the Census (compared at the more detailed NACE3 level) is 0.71, always signi cant at the 1 per cent level. Finally, in terms of rms size, Table 4 shows the distribution of our sample rms across the size classes adopted by the Italian National Institute of Statistics and measured in terms of employment. 13 For some rms the NACE4 code is available (e.g. 1512), for others only the NACE3 classi cation is reported (e.g. 1510). 8

9 Looking at data in 2001, in order to have a comparison with the Italian Census, in our sample there is a fair number of micro rms (11.4%), although the latter in Italy account for more than 80% of total rms. The relative over-representation of large rms in our dataset is clearly a drawback that must be taken in mind when discussing our results. However, since we are interested in the e ects of an increased international trade exposure on rms pricing behavior, the under-representation of micro rms in our sample should not bias our main ndings. Before turning to present the results of the econometric analysis, Table 5 shows some descriptive statistics of the variables used for our estimation of markups as from eq. (7). From panel A we can see that the average turnover of the rms in the sample is 13.7 million of euros, and the average employment is 63. The huge heterogeneity among rms is witnessed by the high standard deviations reported, together with the values of the 1st and 99th percentiles. Panel B reports instead the percentage variations over time of these variables. [Table 5 about here] 4 Results 4.1 Aggregate results We start by presenting the results obtained at the aggregate level, pooling all our rms observations. Table 6 reports the results for the baseline model of eq.(9). In particular, the rst column reports the estimates using rm- xed e ects and time dummies to control for a possible time trend, and clustering the standard errors at rm level to avoid their possible downward bias induced by regressing rm-level observations on industry-speci c import penetration ratios 14. The main effect of the estimated markup for the baseline year (1998) 15 is 34.7% and statistically signi cant, while the interaction term capturing the marginal impact of the import penetration index on the markup (DX_IM P ) displays a negative and statistically signi cant coe cient, in line with the pro-competitive e ect of trade postulated by the theory. An average import penetration index in 1998 of 0.21 then leads to an actual estimated average Lerner index of around 33%, i.e. around 1.7% lower. The results obtained are not a ected by alternative treatments of the panel dimension (random e ects) and provide broad evidence of pro-competitive gains from trade at the aggregate level in line with the standard results of the literature. [Table 6 about here] The rst robustness check deals with the assumption of constant returns to scale implicit in our algorithm for estimating the markup. Since the latter assumption may generate an upward or downward bias in the estimates, we have compared our estimated markup with an approximation of the price-cost inferred directly from the available balance sheet data. To this extent, a possible 14 We have employed throughout the analysis the import penetration indexes calculated at the NACE3 level. 15 Since the methodology exploits growth rates, the 1996 data are used to build the rm-speci c variables for 1997, which then yield 1998 as the baseline year in which we observe the Lerner index. The time span goes until 2003, the last year in which trade data are available. 9

10 approximation of the PCM can be obtained at the rm level taking the di erence between production value and total variable costs (employment plus material costs) divided by production value (Tybout, 2003): c it P CM it = P it ' P itx it c it X it (11) P it P it X it Graph 2 presents a comparison between the markups estimated for each NACE2 industry in each year, and the ones calculated as in (11) 16. Both PCM measures are increasing in the considered time period and, as the graph shows, are highly correlated (0.85), with no signi cant biases emerging from our estimated measure. [Graph 2 about here] As further robustness checks, column 2 of Table 6 performs the same analysis of column 1, but restricted to the balanced panel resulting from dropping all those rms that did not have data for every time period. As it can be seen, the negative impact of import penetration on the average price-cost margins still holds. We have then tested whether our results are sensitive to the methodology employed in the calculation of the import penetration ratio. In general the NACE classi cation at 3-digits is the level used by the European Commission to distinguish between intermediate and nal products (EC Regulation No.586/2001). However, due to product di erentiation it could be the case that, for some rms, products classi ed within the same NACE-3 category are actually used as inputs, i.e. they complement rather than substitute for domestic production 17. To control for this potential bias, Column 3 of Table 6 reports the results obtained using as a proxy for import penetration the same index calculated at the narrower 4-digits level of details, rather than at NACE-3. Although the number of available rm-level observations is slightly lower, because not every rm is associated to a NACE-4 category of economic activity, the point estimate of the import penetration coe cient is still negative and signi cant. The speci cation in the fourth column of Table 6 exploits a di erent indicator of import penetration, obtained as the ratio of total import over the sum of import and production, thus bounding the index between 0 and 1 avoiding to subtract exports. The impact of import penetration on price-cost margins is larger than before, but sign and signi cance are unchanged. In order to rule out the potential endogeneity of the import penetration index, we have followed Konings et al. (2005b) employing a lagged value of this measure. The results, reported in Column 5 of Table 6, are again entirely similar to our baseline speci cation. Finally, in Column 6 we have looked at the interaction between the Her ndahl index and import penetration, in order to assess whether import penetration has a di erent e ect on price-cost margins in more or less concentrated sectors: the latter characteristic does not seem to be, on average, a signi cant driver of markups. Not surprising, our general results thus point to a pro-competitive e ect of trade in Italian industries. However, given the great deal of heterogeneity present across industries and rms, it is interesting to perform the same analysis at a more detailed industry level. 16 In order to get a meaningful value of (11) at the aggregate level, we have calculated weighted averages at industry level using as weights the rms sales shares 17 In this case the pro-competitive e ect of trade would take place only if the lower costs for rms are transferred to consumers, i.e. one would expect some role for the degree of competition in the industry. 10

11 4.2 Industry level results We have estimated eq. (9) for each NACE2 industry, always using rm- xed e ects and time dummies to control for a possible time trend, and clustering the standard errors to avoid their possible downward bias. Table 7 presents the results of this estimation reporting the estimated Lerner index (DX) for the baseline year (1998), as well as the coe cient of the interaction term with the import penetration index (DX_IM P ). To check the robustness of these estimates, an alternative speci cation has been employed, where random e ects have been used where appropriate according to the Hausman test, as reported in Table 7. Moreover, Table 7 also reports the results obtained running each industry-speci c estimation only on the balanced sample of rms, again in order to exclude that our results are driven by entry and exit dynamics of rms. Based on these results, robust across the di erent speci cations, it is immediate to see that the estimated PCM are always signi cant and vary across industries, which is expected. However, it is quite striking to notice that the sign and signi cance of the interaction terms with the import penetration index display a huge degree of heterogeneity. In particular, three di erent groups of industries are present. In a rst group the impact of import penetration on the price-cost margin appears always negative and statistically signi cant across all speci cations, in line with the standard results of the literature (the group is labelled "Weakened", in accordance with the impact of import penetration). A second group ("Neutral") is characterized by industries in which the impact of import penetration on the price cost margin is not signi cant. Finally, in a third group, which we refer to as "Strengthened" industries, a higher import penetration is always signi cantly associated to a higher price cost margin. [Table 7 about here] Since the latter result is quite controversial, the next section explores its possible determinants. 5 Industry markups, product mix and productivity We have explored whether some structural characteristics at the industry level might explain the di erentiated response of industries markups to trade penetration. No structural di erences seem to exist across the industry groupings in terms of import penetration, which is always displaying an average upward trend (see Table 1 and Graph 1). We have thus analyzed the dynamics of net entry, observed price-cost margins (retrieved as before from individual rm balance sheet data) and the Her ndahl index, aggregating rm-level observations and pooling industries together into the three identi ed groups via weighted averages 18. The results are reported in Table 8a. [Tables 8a and 8b about here] As it can be seen, the most controversial of our industry grouping (the "Strengthened" indus- 18 We have calculated sales weighted averages for each NACE2 industry. When calculating the values for each industry group (W, S or N), the weight attached to each NACE2 industry is the share of the industry in the total sales of the relevant industry group. 11

12 tries) displays increasing price-cost margins also when the latter are observed at the rm-speci c level, thus conferming the trend of our estimated markups reported in Table 7. The increase in markups does not seem to derive from an increase in the concentration of rms, since the dynamics of the Her ndahl index are decreasing, a nding common to all the industry groups (Table 8a). Moreover, the "Strengthened" industry group is also the one displaying the largest negative net entry dynamics over time, a nding which does not support a compositional explanation of higher markups 19. Another possible explanation for our nding relies on the previously discussed hypothesis that, in some industries, imports registered at the relatively aggregated NACE-3 level include not only substitute products directly competing with local ones, but also imports of (cheaper) intermediates. If this is the case, rms operating in industries were the relative presence of imported intermediates is higher should experience a reduction of their costs rather than prices, thus possibly leading to higher markups. The e ect would then be stronger in relatively more concentrated industries, where the pressure to transfer these gains onto lower prices is smaller. To test for this channel, we have introduced in our econometric speci cation an interaction with a dummy S equal to 1 if the industry j belongs to the "Strenghtened" group. The evidence is however not consistent with an e ect on markups induced by a reduction of marginal costs, for a number of reasons: rst of all, comparing Column 1 and 2 of Table 8b we nd that the positive e ect of import penetration on markups in the considered industries is robust to a measure of import competition at the NACE-4 digit, a re nement which, as previously argued, should allow us to better focus on substitute products, thus limiting the e ect of imported products acting as intermediates for the considered rms. Second, the interaction between the Her ndahl index and import penetration in the "Strengthened" industries, reported in Column 3 of Table 8b, is signi cant but always with a negative sign, i.e. a higher import competition tends to lower markups in relatively more concentrated (and hence more protected) industries, thus acting through prices, rather than costs 20. In Column 4 of Table 8b we have also run our speci cation using the lagged import penetration index, to rule out the potential endogeneity of the latter, nding virtually identical results 21. Moreover, when looking at the componentes of the rm-speci c observed PCM calculated as in (11), in the considered period the "Strengthened" industries display the highest yearly average growth rate of both total variable costs (1.7 per cent, vs. 1.4 and 0.7 for the Neutral and Weakened industries, respectively) and value of production (2 per cent, vs. 1.6 and 0.6). All these ndings thus point to an increase in markups deriving mainly from the e ect of higher prices, rather than lower costs, for the relevant industry group. We have then tried to link the positive correlation between import penetration and average industry markups with the recent evidence that rms adjust their product mix in response to trade pressures (Bernard et al., 2006a). In particular our hypothesis is that, in certain industries, 19 Melitz and Ottaviano (2005) show that, in a monopolistically competitive framework with rm heterogeneity and endogenous markups, increased import competition leads to rm exit and lower average markups, as the direct upward e ect on residual demand price elasticities (i.e. lower markups for all rms) outweighs the selection e ect, according to which the least productive rms exit and only rms with lower cost (and relatively higher markups) remain in the market. 20 Similar e ects are detected by Konings et al. (2005b) in their analysis of the e ects of import competition on the average markup of Romanian rms. 21 Note that the general argument for a potential endogeneity of the import variable is not straightforward in our case, given the di erent sign that import penetration generates on the markups of di erent industries. Nevertheless, throughout the paper we always control for the robustness of the results to lagged imports. 12

13 rms are more likely to contrast an increase of foreign competition with a switch of their product mix towards products characterized by lower elasticities of demand, and thus end up with higher average price-costs margins as a result of an increase in import penetration. Unfortunately, we do not have rm-speci c data on individual product choices, but we can rely on industry-speci c measure of product heterogeneity. In particular, in order to test our hypothesis we have used the Eurostat PRODCOM database, which collects data in time series on production at the nest possible level of detail (8-digits), for every EU country 22. We have then calculated a time-varying index capturing the dispersion/concentration of the product mix for our NACE2 industries, using as observational units the share of each product code in each industry and year. In particular, we have used an entropy index calculated over the individual product shares for eache NACE-2 industry and year 23. With respect to traditional measures of dispersion (e.g. the standard deviation of the product shares), entropy has the advantage of being less a ected by the shape of the distribution of the observational units. In our case, in fact, our observational units (the product shares) have an heterogeneous support in each comparison group (a NACE-2 sector), because di erent sectors are partitioned in a di erent number of product codes. Moreover, also the ordering of product codes, which have no particular ranking but the one coded by the statistical o ces, might a ect traditional measures of dispersion 24. Graph 4 reports the evolution over time of the entropy index, normalised to 1 in 1998 for each of the three industry groupings previously identi ed 25. Consistently with our hypothesis, the "Strengthened" industries, i.e. those where we nd a positive correlation between import penetration and average industry markups, are the only ones displaying a clear dynamics in the evolution of their product mix, which tends to become more dispersed over time. [Graph 3 about here] We have then tried to assess the signi cance of this nding within our econometric model, modifying our eq. (9) as follows: DY ijt = 1 DX ijt + t DX ijt T t + 2 DX ijt IMP jt + 3 DX ijt ENT ROP Y jt + i + ijt (12) As for the previous estimating equation, in Eq. (12) the dimension j represents the industry 22 The PRODCOM list consists of about 4500 headings relating to manufactured products. Products are detailed on an 8-digit level; 1 to 4 digits refer to the NACE classi cation in which producing enterprise is normally classi ed. Most headings correspond to one or more Combined Nomenclature (CN) codes used for trade data. For example, NACE4 code refers to "Knitted pullovers or similar products"; within this category, the Prodcom list distinguishes 10 di erent products, e.g. the code "Men s or boys jerseys, pullovers, sweatshirts, waistcoats and cardigans, of wool or ne animal hair (excluding jerseys and pullovers containing 50% of wool and weighing 600g)", the code , which refers to females models for the same product, or the code "Lightweight ne knit roll, polo or turtle neck jumpers and pullovers, of cotton". 23 Entropy is de ned as E jt = P i [prod it ln(prod it )] = ln(k), where prod it is the share of each product code in each NACE-2 industry j in year t, and k is the number of (8 digit) product codes in which each NACE-2 industry has been partitioned by the PRODCOM dataset. The index is bound between 0, corresponding to perfect concentration, and 1, indicating equal dispersion (i.e. a uniform distribution of product shares). 24 In economic geography, where observational units typically are spatial zones, a similar problem is known as the MAUP (Modi able Areal Unit Problem): these zones are often arbitrary in nature, and di erent areal units can be just as meaningful in displaying the same base level data. See Unwin (1996) for a discussion. 25 We have calculated sales weighted averages of each industry measure in order to retrieve the group value. The normalisation allows to capture the relative dynamics of each industrial group. For the "Strenghtened" industries the entropy index was in 1998, then raising by 10% to in , i.e. signalling an increase in the dispersion of the product mix in these industries. 13

14 to which the rm i belongs at time t, T t is a set of time dummies which allow us to control for cyclical demand e ects, i stands for an unobservable rm-speci c xed e ect, and ^ 2 captures the marginal impact of the import penetration on the PCM s estimates. The speci cation is then augmented with another interaction term (DX ijt ENT ROP Y jt ), capturing the marginal impact of each industry s distribution of the product mix on the average markup. Table 9 reports the result of the estimation across all industries, always using rm- xed e ects and time dummies to control for a possible time trend, and clustering the standard errors. While the overall impact of product heterogeneity on the markup is not signi cant for the average industry (Column 1), the latter terms becomes signi cant when the dummy S for our "Strenghtened" industries is considered (Column 2). The negative sign is consistent with our hypothesis that, for these industries, a wider range of products (i.e. a more dispersed product mix) is associated on average to higher rms price-costs margins 26. Columns 3 and 4 explore whether the change in the distribution of the product mix (i.e. the interaction term between DX ijt and the rst di erence ENT ROP Y jt ), rather then the level, is associated to an increase in the PCM of the considered industries. The idea is that in the previous speci cations we might have been picking up just a product life-cycle e ect: typically, mature sectors like textiles are partitioned in a ner number of product codes with respect to more innovative sectors, where product di erentiation is not yet pervasive. As a result, these industries would be characterized by a structurally higher number of products, and hence a possibly higher product dispersion, yielding a spurious correlation between markups and distribution of the product mix. However, our results remain consistent. On average we nd that, always controlling for the e ect of the import penetration, rms tend to concentrate their product composition to maintain a positive marginal e ect on the PCM (Column 3). The result is in line with recent theoretical models of multi-product rms (e.g. Eckel and Neary, 2006; Bernard et al., 2006b) predicting that greater market integration leads rms to focus on "core competencies", thus dropping their lowestexpertise products 27. We nd however that such an outcome is not homogeneous across industry groups: in our "Strenghtened" group of industries the interaction term enters with a negative sign, i.e. a higher dispersion of products (a negative value of the term ENT ROP Y ) is associated with a higher price-cost margin. The result is again in line with the intuition that in the latter industries the product mix has been (partially) switched towards products characterized by a lower elasticity of substitution, thus generating the detected positive correlation between higher product dispersion and markups. [Table 9 about here] Finally, as a robustness check, we have recalculated Table 9 controlling for the lagged import penetration, again to rule out the potential endogeneity of the import penetration index. The results, available on request, do not change. 26 Recall that entropy is bounded between 0 and 1, with the latter indicating a higher (at the limit uniform) dispersion in the distribution of the product shares. 27 From an empirical point of view, Del Gatto et al. (2007), working on a similar sample of Italian rms albeit observed only up to 1999, nd that on average more open industries are characterized by a higher concentration of costs across active rms. 14

15 5.1 Firm-speci c robustness Insofar, we have shown results where average markups have been endogenously retrieved from rm level observations, but have been assumed constant across the group of considered rms within each industry. It is however interesting to check whether our ndings hold also when considering the entire distribution of the observed, rather then estimated, rm-speci c price-cost margins (the term P CM it previously calculated), knowing that the latter are correlated in an unbiased way with the estimated markups (see Graph 1). To this extent, Graph 4 shows the distribution by industry groupings of the rm-speci c markups. It is interesting to note that, while the distribution of markups is by and large normal, the "Strenghtened" group of industries displays a higher average and a larger dispersion of the PCM. Moreover, the latter industry displays a standard deviation of markups increasing at a faster pace than other industry groupings. The nding is con rmed by an industry-speci c regression, whose results are reported at the bottom of Graph 4, in which the standard deviation of the industry-speci c (NACE3) markups are regreessed against the (lagged) index of import penetration (calculated over the same NACE3 cathegories) for each available year. As it can be seen, lagged import penetration is signi cantly associated to a higher dispersion of the markups, with the e ect being larger by a factor of 10 in our S-industries 28. To explore more speci cally the dynamics of the product mix and trade openness using the rm-speci c observed PCM, we have then estimated the following rm-speci c regression: P CM it = IMP jt + P ROD jt + T t + I j + it (13) Variables are taken in rst-di erences in order to wipe-out unobserved individual rm e ects which might potentially a ect the correlation between the price cost margin and our variable of interest, while we control for time and industry (at the NACE3 level) xed e ects. Since the regression is performed on micro units using mainly aggregate variables as covariates (recall that we do not have rm-speci c information on product choices), we also control for the potential downward bias in the estimated errors by clustering the standard errors. The results are reported in Columns 1 and 2 of Table 10, and con rm our main ndings: changes in import penetration are associated on average to negative and signi cant e ects on the price-cost margin of each rm (Column 1). However, the e ect of import penetration is not signi cant for our group of "Strenghtened" industries (Column 2). The PCM of rms operating in these industries also reacts di erently to changes in the product mix. Controlling for the import penetration, for the majority of rms positive changes in rms PCM are signi cantly associated with an increase in product concentration, while for our "Strenghtened" group of industries the product mix of rms becomes more dispersed towards products characterized by a lower elasticity of substitution (a negative term ENT ROP Y ) and thus higher markups (Column 2). In Columns 3 and 4 we interact entropy with the index of import penetration, in order to explore whether the e ects on the markup induced by a change in the product mix work through the changes in the import penetration. Our intuition seems to be supported by the data: while for the average of rms the 28 The theoretical framework of Melitz and Ottaviano (2005) predicts that the dispersion of markups is negatively associated to market size or increased trade ows. Here, quite consistently, we nd a large and positive impact in the industries in which the endogenous changes in the product mix are such to revert the standard pro-competitive e ects of trade. 15