Local product space and firm-level churning in exported products

Abstract

1. Introduction

Recent empirical evidence suggests high levels of churning in exported products at firm level, and highlights the contribution of these changes to growth of aggregate output. Iacovone and Javorcik (2010) report that in Mexico, over the period 1994–2003, export variety creation represented 19% of annual exports of all varieties while export variety destruction accounted for 11%. Bernard et al. (2010) in a study of US manufacturing firms over the period 1987–1997 found that the gross contribution of product churning within firms to aggregate manufacturing output is as large as that of firm entry and exit. These changes in firms’ product ranges are also key to understand the dynamics of country level comparative advantage. Hanson et al. (2015) studied 135 industries in 90 countries between 1962 and 2007, and found that 60% of the products that account for the top 5% of a country’s absolute-advantage industries in a given year, were not in this 5% two decades earlier. Hence, knowing what determines changes in the range of products that a firm exports and whether these changes show some sort of pattern are important issues, worthy of investigation.

Nevertheless, the economic literature on international trade has overlooked firm-level churning in exported products for a long time. The traditional theory of comparative advantage relies on a competition framework in which firms are intrinsically monoproduct. This also holds for the new trade theory although it does introduce the notions of product differentiation (Krugman, 1979; Ethier, 1982) and heterogeneous firm efficiency (Melitz, 2003). Hence, in these theoretical models churning in exported products is reduced to the decision to start or to stop exporting. Multiproduct exporters have attracted some recent research attention but existing theoretical work explains how economies respond to trade liberalization shocks through within firm adjustment margins (Eckel and Neary, 2010; Mayer et al., 2014, 2016; Lopresti, 2016). Whereas, empirical work focuses on growth in intensive margin or the survival of new products (Poncet and Starosta de Waldemar, 2015; Goya and Zahler, 2017). As a result, the determinants and direction of churning in exported products at firm level is still an issue at stake in this literature.

The aim of this article is to address this open question empirically, building on the theoretical and empirical findings in three different literature strands: industry dynamics, theory of the firm, and economic geography. The industry dynamics literature proposes some explanations for why firms change the range of products produced (not the range exported). The seminal models in this stream of literature (Jovanovic, 1982; Hopenhayn, 1992; Ericson and Pakes, 1995) are based on the monoproduct firm assumption but the extension proposed by Bernard et al. (2010) provides a general equilibrium framework in which firm entry and exit can be distinguished from product-market entry and exit.¹ In their extension, firms’ product-market entry decisions are explained as resulting from evolving firm and firm product-specific factors namely productivity and consumer tastes. While this simple model setting proposes explanations of the key drivers of product churning, it does not perform well for explaining the tendency of firms to produce certain products simultaneously (Bernard et al., 2010).

On the other hand, the theory of the firm helps to explain this tendency for coproducing certain products. In particular, the theory of multiproduct firm (Teece, 1982) builds on the resource-based perspective (Penrose, 1959), which states that the firm’s production resources can provide a number of different production services. At a given point in time, the firm exploits a particular subset of these services and its product portfolio represents the way it uses these resources. Due to the indivisibility of resources or learning, the firm can have excess resources and it diversifies into related markets when this excess cannot be used more effectively by selling/leasing them (Teece, 1982). Hence, the tendency for firms to produce multiple products is driven by commonalities in their underlying resources or the services enabled by these resources or their” relatedness” more generally.

A number of applied studies investigate this related diversification argument based on different connotations of relatedness and diversification. One strand of this work focuses on diversification of firms into new industries or technologies. For example, following the resource-based view, defining diversification as entry into new industries, and relatedness as belonging of industries to same (sub)classes in standard industry classifications; Chatterjee and Wernerfelt (1991) show that depending on the type of excess resources firms can diversify in the related or unrelated direction. Neffke and Henning (2013) measure skill-relatedness among industries using labor flows, and find that firms are more likely to diversify into skill-related industries. Whereas, Breschi et al. (2003) focus on technology resources and conclude that technology relatedness has an impact on the diversification of innovative activities. Another strand of work considers export diversification at different scales ranging from countries to firms (Hidalgo et al., 2007; Hausmann and Hidalgo, 2011; Boschma et al., 2012; Goya and Zahler, 2017). At the firm level and using Chilean data Goya and Zahler (2017) show that for any given firm a new-product line is more likely to survive market competition if it is closely related to the firm’s existing products. In other words, they underline the role played by within firm complementarities and commonalities among capabilities to explain churning in exported products.

However, the economic geography literature provides further arguments and evidence that relatedness can operate beyond firm boundaries and motivates considering the local economic structure as a factor shaping firm-level churning in exported products. To illustrate, Poncet and Starosta de Waldemar (2015) focus on product relatedness to the local product space and explore its impact on firms’ export performance rather than their product-market entry decisions. Based on data on Chinese manufacturing firms, they show that firms producing goods that are closely related to the goods in which their locality is specialized, enjoy higher levels of export revenues in the next period. Based on data on Turkish manufacturing firms Lo Turco and Maggioni (2016) also show that firms are more likely to introduce new products that are more closely related to both within firm and local product-specific capabilities.

Building on this body of work, in this paper we study the main drivers of churning in exported products at firm level, and try to elucidate the particular role played by the locality via product relatedness. We start with a simple model suited to capturing the essence of the relationship between the characteristics of local product space and the dynamics of firm-level product churning, and test the model empirically. The simple model we work with is an extension to the model proposed by Bernard et al. (2010) on multiple product firms and product switching. We extend it to a multiregional setting and introduce the local dimension via urbanization economies common to all firms in the locality, and demand and export externalities the firms enjoy depending on the degree of relatedness of their products to the products exported by the locality.

Empirically, we base our investigations on a rich dataset on French monoregional manufacturing firms that export both in 2002 and 2007. We use the density measure proposed by Hidalgo et al. (2007) to quantify product relatedness and investigate two main questions. The former inquires the determinants of decisions about the entry (exit) to (from) product markets, especially the impact of the structure of the local product space on these choices. The latter investigates given that a firm starts exporting new products, whether each has an equal share in terms of revenue generation or the structure of the local product space affects growth in each product market.

Doing that the study mainly contributes to the literature on multiproduct exporters by addressing the open question on determinants of churning in exported products but its findings can be articulated to two broader discussions. The first one refers to whether and how spatial distribution of economic (export) activity affects firms’ export behavior. Existing work on this relationship includes mostly empirical papers, addressing export performance in terms of the intensive and extensive margins, or the triggers of export activity (Greenaway and Kneller, 2008; Koenig, 2009; Koenig et al., 2010; Cadot et al., 2013; Poncet and Starosta de Waldemar, 2015). Our findings extends this stream of work by showing that space matters also for export diversification. The second question refers to the sources of regional diversification. Existing evidence highlights a tendency at the country/regional level to move in the product space step-by-step by jumping to related goods (Hidalgo et al., 2007; Hausmann and Hidalgo, 2011; Boschma et al., 2012) yet this is the average behavior, not a rule. Different actors (monoregional or multiregional incumbents/export starters) can drive abrupt changes to comparative advantage under different conditions (Boschma et al., 2017; Neffke et al., 2018) and hence microdynamics underlying behind these aggregate changes are not straightforward. The findings from the present study add to this work by suggesting that monoregional incumbents tend to diversify in the direction implied by regional capabilities.

The rest of the article is organized in five sections. The next section presents the theoretical framework. Afterwards, in Section 3, the data sources, sample selection, and measurement of product relatedness will be explained. In Section 4, the econometric specifications will be provided. In Section 5 the estimation results will be presented, and their robustness and implications will be discussed. The final section offers some conclusions.

2. A simple theoretical framework

Our study exploits the model proposed by Bernard et al. (2010), which considers firm-level product churning at the steady state. In this framework, which we call the BRS framework onwards, firms add-drop products in the steady state due to idiosyncratic shocks to productivity and consumer taste. In what follows, we introduce and describe the BRS framework in more detail and show how it can be extended to address the role of local interdependencies in changes in the range of exported products at firm level. The BRS model envisaged through this spatial lens is what allows us to derive the key testable propositions of our empirical analysis.

2.1 The BRS framework

Bernard et al. (2010) extend existing theories in industrial dynamics by proposing a model that explains how a firm chooses in which product markets it will be active. In former models, this question has been the same as firm entry and exit as these models assume a single-product firm. In the BRS model (Bernard et al., 2010), firms choose to produce an endogenous range of products based on the evolution of the firm and firm-product level characteristics.

The BRS model assumes that firm entry entails a sunk entry cost. Once this cost is incurred, firms observe their initial productivity, which is specific to the firm and common to all of its products, and the consumer taste for each of their products. Consumer taste can be interpreted as the product quality as well as some subjective product characteristics that affect consumers’ demand for that product. There is also a fixed production cost for each product. Productivity and consumer tastes are assumed to evolve stochastically over time and to be serially autocorrelated but uncorrelated to each other. Finally, labor is assumed to be the only production factor and in inelastic supply.

According to equation (1) firms with higher productivity $(ϕ)$ face a lower zero-profit consumer taste cutoff $(λk*(ϕ))$⁠. Then, the probability that these firms have a consumer taste for their variety of product k higher than $λk*(ϕ)$ is higher. This means that it is more likely that producing product k is profitable for them. As productivity increases the probability of producing a product, more productive firms can be expected to have a larger product range.

However, this proposition of the model relies on the assumption that consumer tastes are independent within the firm, and across its products. Bernard et al. (2010) recognize that products might have some common features that are valued by customers in different product markets. Therefore, they suggest demand and production interdependencies across products within each firm as a possible extension of the model. However, these interdependencies are not necessarily confined within firm boundaries.

2.2 The BRS framework through a spatial lens

The BRS framework captures the essence of firms’ product selection but it is totally silent on the role of firm interactions and agglomeration economies. To investigate whether spatial factors affect how firms determine their product switching strategies, the BRS framework needs to be extended to a multiregional setting, and relaxation of the simplifying assumption that there are no spatial interdependencies across products and firms. In this article, we propose to model the simplest spatial interdependencies required for a useful data analysis while retaining as many as possible of the simplifying features in the original BRS framework.

In extending the framework to a multiregional setting, we consider that the profitability of a(n) (exported) product for a given firm has not only a firm and a firm-product component as in the original BRS framework but also a location and a product-location component. In introducing location in the analysis, we make the simplifying assumption that firm location is exogenous and fixed over time. Hence, we study only how a firm once it is born can optimize its product mix in a context where the profitability of each product line is determined jointly by firm, firm product, location and location-product-specific random shocks.³ Compared with the BRS, our framework is more complex because it includes the impact of the spatial interdependencies across firms on the firm’s choice.

Formally, we assume that new firms are established randomly in a given location.⁴ We follow BRS by assuming that each of those randomly and locally born firms incurs a (common to all locations) sunk entry cost of $fe>0$ units of labor in order to observe its intrinsic productivity $ϕ∈[ϕ̲,ϕ¯]$ and its intrinsic product qualities for the characteristics embodied in its blueprint for every product, $λk∈[λ̲,λ¯]$⁠.

We then depart from BRS by assuming that agglomeration economies prevail in our economy. Those agglomeration economies take two different forms: urbanization economies, which are available to all firms established in a given locality, and localization economies, which are available only to firms that produce similar or related products in a given locality. In the first case, all firms in a given locality enjoy some level of competitive advantage regardless of their product portfolio. This might stem from a common (transportation) infrastructure, social interactions and institutions facilitating exports, or knowledge externalities in general. In the BRS framework, we model this form of agglomeration economies as a shifter that lowers the fixed cost of production/export for all products in a given locality. Specifically f_pkl denotes the fixed cost of production of product k in location l. We assume that f_pkl has two distinct components: a product-specific component f_pk as in the original BRS model, and a location specific component f_pl, which depends only on location characteristics not product characteristics. We then assume that f_pl is lower in a more urbanized locality.

The second type of agglomeration economies depends on the specificity of the locality’s product space. Basically, we assume that the benefits derived from knowledge externalities, cheaper intermediate goods, common export facilities, or social interactions will be stronger if the products exported from a given region are closely related. This type of agglomeration economies arises from colocation of a specific set of industries and is described in the literature as Marshall-Arrow-Romer (MAR) and Jacobian externalities. Specifically, MAR externalities include knowledge externalities fostered by specialization, labor market pooling, and savings on transportation costs (Glaeser et al., 1992) while Jacobs externalities derive from knowledge complementarities and cross-fertilization opportunities across a diverse spectrum of knowledge (Jacobs, 1969). Frenken et al. (2007) suggest also that diversity can take two different forms, that is related or unrelated variety, and argue that related industries have correlated demand shocks and that related variety fosters Jacobs externalities. In contrast to urbanization externalities, in the case of both MAR and Jacobs externalities the benefits the firm can obtain depend on the degree of the firm’s similarity to the other firms in its location.

Those agglomeration economies can work through rich and diverse channels and impact the firm’s marginal costs on the supply side, and the firm’s intrinsic product characteristics (technological or subjective) on the demand side. In our setting, agglomeration economies are introduced via product-specific consumer taste distributions (the distribution of λ_k).⁵ More precisely, we assume that the probability of firm i to draw a high consumer taste for a given product k is positively correlated not only to its past draws for that product but also to the past draws of neighboring firms for product k or some other products closely related to product k. In addition, we assume that those agglomeration economies are stronger for products that are exported with comparative advantage from the locality. In other words, where core product refers to products exported with comparative advantage by the locality, local externalities will prevail primarily across the “core products” in any given locality.

Recalling from the previous section that consumer taste embraces both the quality and subjective aspects of demand, agglomeration economies can operate on consumer taste via technological complementarities/commonalities affecting quality and/or via subjective consumer preferences. For the first case, one may think of some high-quality inputs or production capabilities associated to the locality’s core products. Local knowledge spillovers or local labor mobility can foster the quality of goods grounding on the same/complementary inputs or capabilities. In the second case, one may consider firms benefiting from foreign consumers’ familiarity with product k originating from location l, and these consumers’ willingness to buy product k with this provenance.

Therefore, in our framework and in contrast to the BRS framework, the firm’s intrinsic product characteristics are not the only determinant of the (foreign) consumer tastes for its variety of products. Also, the characteristics of the local product space matter because of the prevalence of demand complementarities in exporting activities. Based on these assumptions, the effective consumer taste for a given firm product is no longer independent of firm location. In other words, two firms that would draw the same intrinsic quality for their varieties of a given product k, might nonetheless face very different effective consumer tastes for those varieties depending on how product k is embedded in the firm’s local environment. At one extreme, a firm located in a region where product k or products closely related to product k, are already exported with comparative advantage by many other local firms would benefit from a much higher effective consumer taste than its intrinsic consumer taste. At the other extreme, a firm in a location where product k or products closely related to product k have not previously been exported would be disadvantaged by being unable to rely on existing exporting facilities and knowledge flows, and its effective consumer taste for this product could be lower.

Formally, the assumption that demand complementarities arise across firms producing closely related products in a given locality, can be introduced in the BRS framework by making the λ_ks a combination of two components: a purely random component whose distribution has the same property as the distribution modeled in the original BRS, and a deterministic spatially dependent component, which depends on the congruency of product k to the product space in the locality of the firm (Ω_l).

On its own equation (2) cannot fully describe the new steady state in our theoretical section. It also cannot substitute for a full analysis of the transition path toward this new steady state. However, this consideration is beyond the scope of our article. For the purposes of our empirical investigation, equation (2) aims only at making it clear how the equilibrium profit cutoff condition for any individual firm would be impacted by its local environment under the assumption of location and product-specific complementarities across firms.

From equation (2), we can derive several testable propositions related to both the spatial determinants of firm-product churning strategies, and the relationship between those microeconomic determinants and the dynamics of regional (local) competitiveness. In the case of firms’ churning strategies, equation (2) has direct implications for both the probability that the firm will switch products and the direction of the portfolio change. More precisely, it can be expected that all else being equal, for a given idiosyncratic productivity, a firm will be able to produce a larger set of products in a more urbanized area because of the fixed production cost advantages available. Also, the firm will be more likely to add new products that are more related to the core products of the locality because of the demand complementarities that prevail across neighboring firms. Finally, all else being equal, the firm will be more likely to drop a product k if this product is less congruent with the core products of the locality. These implications allow us to derive our first testable proposition about the direction of firm-level product churning:

Testable proposition 1: For a given idiosyncratic productivity and a given location, firm i is more likely to add (drop) product k to (from) its export portfolio if product k is more (less) congruent with locality l’s product space.

The second set of implications from our extended BRS framework is related to firm sales dynamics. Basically, as equilibrium profits for related products are higher in our framework, those products can be expected to be the new paths through which firms grow. From this implication, we can derive a second testable proposition.

Testable proposition 2: Given its location and product-market entry decisions, among the new-product markets that the firm starts to serve, the share of export revenues from product k will be larger the greater the congruency of product k to the local product space.

3. Data, sample, and measurement of product relatedness

3.1 Data sources

For the empirical analysis, we combine information from three different sources to build an original dataset for the period 2002–2007, comprising French manufacturing exporters, their characteristics, the products they export, and the extent that their products are related to the core competencies of their locality. The first data source is the French Customs Statistical Office, which provided us with information on the quantities and values exported by each firm at product level for each destination. The second source is the French National Institute of Statistics and Economic Analyses (INSEE). Via INSEE, we accessed FICUS database, which provides balance sheet and financial information on individual enterprises. Finally, we use the BACI⁶ database maintained by CEPII⁷ research center. BACI (Gaulier and Zignago, 2010) provides information on values and quantities of bilateral world trade flows at the Harmonized System (HS) six-digit product disaggregation, for more than 240 countries and 5039 products since 1994.

We use the first two data sources to determine the sample of firms, their characteristics (such as productivity, location, export revenues, etc.), and products. Whereas CEPII database allows us to quantify the extent that a product is related to the core products in a locality. The next two subsections in the sequel explain the use of these data sources in greater detail.

3.2 Sample selection

To study firm-level churning in exported products during 2002–2007, first we used French Customs data to determine the set of exporters existing in both 2002 and 2007. Then, we used FICUS data to geolocalize these firms. FICUS includes two types of information on firm location: the region where the firm’s headquarters is located, and the extent of spatial concentration of the firm’s employees. Regarding spatial concentration of employees, it classifies firms into three categories: multiregional enterprises (not more than 80% of their employees located in the same region), quasi monoregional enterprises (80–100% of their employees located in the same region), monoregional enterprises (100% of their employees located in the same region). For multiregional and quasi monoregional enterprises, FICUS enables us to geolocalize only the headquarters although they may have multiple establishments in different regions. This means that for these firms it is not possible to geolocalize precisely which products are produced and exported at which production site, and thus in which region. Hence, in our analysis, we focus on monoregional exporters and exclude those who changed location between 2002 and 2007. Finally, using FICUS, we also identify the firms’ principal economic activity. We restrict our sample to exporters whose principal economic activity is manufacturing. Again, we exclude firms that experienced a change in their primary economic activity between 2002 and 2007.

Table 1 summarizes the number of all exporters and their export revenues in 2002 and 2007, and provides the same information for manufacturing firms, monoregional firms, and monoregional manufacturing firms. Table 1 summarizes that the number of all firms that we can trace in both 2002 and 2007 and characterize their location and principal economic activity is around 50,000. Firms engaged primarily in manufacturing comprise 43% of these firms and generate 73% of export revenues. Firms established in a single region (which remained in that region) constitute 82% of these firms and account for 41% of export revenues. Table 1 summarizes also that most manufacturing firms are monoregional. However, the export revenue generated by monoregional manufacturing firms is smaller than that generated by multiregional manufacturing firms.

In the econometric analysis, we focus on the subsample of monoregional manufacturing firms that exported continuously between 2002 and 2007. This involves several selection issues. First, this focus does not allow a comprehensive explanation of the microdynamics of regional diversification. For instance, the impact of relatedness on firm entry/exit shown in Klepper (2007) and Neffke et al. (2012) is beyond the scope of our study. Second, multiregional firms are not the majority of all firms but constitute a large portion of export revenues. By excluding them, we might be underestimating the core products of any given region. Third, our strict focus on manufacturing firms is motivated by the fact that consumer taste in our theoretical framework could also be interpreted in terms of product quality, and a region-specific component of consumer taste makes more sense for producers than for trader firms. However, firms are engaged in manufacturing to different degrees whereas we made a binary classification, that is those firms whose primary economic activity is manufacturing are labeled manufacturing firms. Therefore, the conclusions that can be derived from this sample on diversification behavior will be ambiguous regarding the source of relatedness, that is it might be due to complementarities/commonalities arising from purely demand or purely knowledge basis, and skills, etc., or a combination of both.

Given these selection issues, over the 5 years analyzed Table 1 shows that the aggregate export revenues generated by these firms increased by 34%.⁸ However, this aggregate growth is created via the different processes that take place within firms: revenue losses due to dropping some products from the export basket, increase or decrease in export revenues from goods preserved in the export basket, generation of additional export revenues from adding goods to the export basket. Table 2 breaks down our subsample of firms by the type of changes to their export basket between 2002 and 2007. We distinguish between firms that both added and dropped products from firms that only added or only dropped products or experienced no change to their export basket over the period.

According to Table 2, in the period 2002–2007, only 14% of monoregional manufacturing exporters maintained the range of products exported unchanged. Note that during 2002–2007, 15% of monoregional manufacturing exporters added at least one new product to their export basket while the share of those that dropped at least one existing product from their portfolio was 12%. Finally, more than 58% of mono-regional manufacturing exporters modified their export basket by adding at least one product and dropping at least one product. The econometric analysis presented in this article tries to explain the determinants of these changes.

3.3 Quantifying product relatedness

Product relatedness or proximity⁹ can stem from several dimensions since they may require similar or complementary sets of resources, skills, knowledge bases, or institutions. Hidalgo et al. (2007) argue that as a consequence of relatedness, countries with a competitive edge in one good may have or may develop advantage in some other good. Hence, they propose measuring the similarity between two products as the conditional probability of having revealed comparative advantage (RCA) in one of them given that the country has a comparative advantage in the other. Their measure has been used widely to investigate structural transformation and economic development in a number of countries.¹⁰

In fact, the co-occurrence of two products in a country’s export basket (with RCA) could stem from not only overlaps in the underlying production processes such as common production factors, input–output relations, common skills, common knowledge base, etc. but also from overlaps in institutions or social and business networks facilitating the export process. Therefore, here the term relatedness goes beyond mere connotation of similarity in terms of sophistication of goods or production features, and embraces also the material and immaterial settings of the production and export processes.

To estimate the proximity between each product pair, we calculate $ϕ$ making use of BACI database provided by CEPII. BACI harmonizes importer and exporter information, and it is available with versions 1992, 1996, 2002, and 2007 of the HS classification and is updated annually. In this study, we work with version 1992 of the HS product classification. To compute $ϕkn$⁠, we retain only manufacturing products at HS 4-digit¹¹ and 217 countries.¹² To illustrate our computations, Table 3 reports some bilateral proximity indexes between selected HS-4 digit products. For comparability, we use the same set of products as Poncet and Starosta de Waldemar (2015),¹³ and for instance, consistently find that Computers and TVs have a high proximity index while Rice and Cars have a low proximity index.

Next, we make use of bilateral product proximity to express the extent to which a product is linked to the local product structure. To do that we use the density measure proposed by Hidalgo et al. (2007), which has been employed in a number of empirical studies (Poncet and Starosta de Waldemar, 2015; Boschma et al., 2012). The density measure focuses on the products exported by the locality with comparative advantage. If a product is proximate to these core products, it is considered densely connected to the local product structure.

In this analysis, we define the locality at the NUTS¹⁴—2 level, that is French region level. Since France has a large base of manufacturing exporters, and the spatial scale of our analysis is quite broad, individual exporters represent a small fraction of regional exports, meaning that individual exporters have little influence on the set of core products in a region. Indeed, in only 1.9% of cases do the export revenues of a firm in a given market provide the region with a RCA. However, note that the density variable depends on the region’s set of core products (reference set), and the bilateral product proximity. Even if a firm has a non-negligible impact on the reference set, bilateral proximity values are determined exogenously using international trade flows, therefore the reference set does not imply an increase or a decrease in product density.¹⁵

As shown in Table 4, the density of a given product might take different values in different regions since each region has a different set of core products. For instance, Table 4 shows that the congruence of the capabilities required to produce and export computers to core local capabilities is the highest in Rhône Alpes and the lowest in Corse.

In the econometric specifications, which are presented in the next section, we use the $Densitykl$ variable to test the propositions about the effect of product relatedness to the local product space on individual firms’ product-market entry/exit decisions, and those firms’ relative share of exports in each product market.

4. Econometric specifications

We observe the firm’s export basket at the beginning and at the end of a 5-year time window. Provided that product k does not exist in the initial export basket of firm i, the dependent variable (⁠$yik1a$⁠) gets the value 1 if product k is exported by firm i at the end of the time window (t), and 0 otherwise. Here, we assume a 5-year time window to account for the fact that entering a new-product market may involve idea generation, product development, tests and trials, adjustments to production lines, administrative activity related to starting exporting, etc. Also, in earlier studies a 5-year lag is assumed to allow for these preparatory phases (Poncet and Starosta de Waldemar, 2015; Boschma and Capone, 2016).

On the right-hand side of equation (6), β₀ is the constant term while α_im represents the effect of the initial product mix m of firm i (in t – 5). This effect comprises existing resources, skills, knowledge bases, or institutions in the firm, and captures the average relatedness of these products to all potential products that the firm might diversify in.¹⁷

Productivity_i refers to the initial labor productivity of firm i, and is calculated by dividing value added (before taxes) by the average effective number of employees in t – 5 and expressed in logarithms.¹⁸β₁ is the coefficient of the productivity variable and is expected to be positive since all else being equal, in our theoretical framework more productive exporters are more able to bear the fixed costs of diversifying.

Our main explanatory variables come next in equation (6). First, $Densitykl$ denotes the extent that product k is related to the products in which firm i’s locality (l) has a comparative advantage in t – 5. The complete definition of this variable was presented in the previous section. It is introduced in equation (6) after taking the logarithm. β₂ is the coefficient of the density variable, which we expect to have a positive value.

We also introduce $RCAkl$⁠, a dummy variable that controls for products already exported by the region with a comparative advantage in t – 5 with β₃ the related coefficient. β₃ is expected to be positive since the consumer taste that the firm observes for a product that is already being exported by the locality with comparative advantage would be higher in our framework. Finally, in equation (6), δ_l denotes locality fixed effects and captures urbanization economies common to all the firms in locality l regardless of what they produce, and ϵ_ik is the error term.

In this equation, ψ₀ is the constant variable. α_im, $Densitykl$⁠, and $RCAkl$ are defined as above. ψ₁, the coefficient associated to $Densitykl$⁠, is expected to take a positive value since firms are more likely to enjoy a high level of consumer taste for products that are related to the core products of the locality due to demand interdependencies. In turn, this would cause equilibrium profits to be higher for such products, leading the firm to grow more in that product line compared with other products introduced during the same period. Along similar lines, ψ₂, which is the coefficient associated with $RCAkl$⁠, is also expected to be positive.

In equation (8), we introduce two additional explanatory variables to control for differences in market entry mode across firms. First, firms can decide to start exporting product k uniquely, or simultaneously with other new products. Second, firms can differ in the chosen timing of their entry over the 5-year time window. Both variations could affect the revenues that can be earned by firm i from a given product k.

To control for the first, we introduce the variable OTH_ik, which is defined as the sum of the densities of other new products introduced by the firm during the 5-year time window. This variable allows us to control for two main mechanisms. First, it controls for the mechanical relationship existing between the number of new products and the average value of $yik2$⁠: the higher the former, the lower will be the latter. Second, it controls for how much the firm can grow in a given product market depending on which other product markets it has entered, and the level of coherence of these products with the local product space. The coefficient of the variable ψ₃ is expected to have a negative sign, indicating that the relative growth in kth product market will be lower if the firm enters multiple product markets that are related strongly to the local product space.

To control for the second, we introduce the variable Past_ik, which is defined as the time elapsed since firm i introduced new product k. This variable can take the values 1 to 5 over our 5-year time window. We expect that the earlier the firm enters a product market, the more time it will have to develop distribution channels and penetrate the market. Hence, ψ₄ is expected to have a positive sign.

During estimations, we also extended the above specifications by including other firm-level controls, that is firm age, belonging to a group. Age_i refers to the logarithm of the age of firm i in t – 5. Group_i is a dummy variable, which takes the value 1 if firm i is affiliated to a foreign or French corporate group at time t – 5. This binary variable can be considered to assess approximately the impact of organizational ties, which might facilitate the transfer of knowledge (within or across regions), decrease fixed production costs, or affect consumer tastes. $Market_grk$ refers to the average annual growth of the world market for product k in the 5 years previous to the start of our time window. Table 5 presents summary statistics for firm productivity and some other control variables.

5. Results and discussion

5.1 Estimation results

The results of our baseline estimation for the first logistic regression model, that is equation (6), are presented in column (a) in Table 6. The $Densitykl$ and $RCAkl$ variables have the expected positive and statistically significant coefficient estimates corroborating our working hypothesis that specialization of the region and the extent of product-market relatedness to the region’s specialization have an impact on the firm’s product-market entry decision. However, although the theory in Section 2 explains that higher firm productivity decreases the zero-profit consumer taste cutoff and increases the likelihood that the firm will enter a given product market, this model does not provide statistical evidence on such an effect. The estimates in column (b) show that in addition to controlling for productivity, controlling for other firm characteristics such as firm age and group belonging yields the same results as presented in column (a).

We diagnose that this finding is driven by the product-mix controls. These controls were added to the model to account for whether the firm’s current exporting activity might be affecting its product-market entry decision since there may be some commonalities or complementarity in terms of resources, skills, knowledge bases, or institutions. The advantage of using a dummy for each product mix is that it takes account of any observable or unobservable effects that might arise from the initial configuration of the trade basket. However, in practice, the use of product-mix dummies causes omissions of substantial numbers of observations from the estimations and causes bias. In our sample, a large number of exporters (about 37%) have at most one counterpart with the same product mix, and these exporters are excluded from the estimations since we require at least three observations to properly estimate the effect of each product mix. When we compare the excluded exporters to those included in the estimations (see Table 7), we find that average labor productivity, average tendency to diversify, and average size of the initial product mix are statistically different between the two samples, and are higher for those exporters excluded from the estimations.

To correct for this, we replace the product-mix dummies with two proxies to try to capture the effect of the initial configuration of the trade basket. One of these proxies is $Num_Prodi$⁠, which refers to the size of the initial trade basket. It is defined as the logarithm of the number of products that firm i exported in 2002. The other refers to industry dummies, which control for the firm’s main industry activity. Column (c) in Table 6 presents the estimation results for this revised version of equation (6). Finally column (d) extends the model in column (c) by controlling for firm age, belonging to a group, and growth in the product market.

Substituting the proxies for the product-mix dummies means omitting some part of the effect of the initial configuration of the trade basket, which is reflected by Tukey’s Link Test. On the other hand, the use of proxies increases the number of observations used in the estimation and also improves the goodness-of-fit as demonstrated by the Hosmer–Lemeshow test statistic. Furthermore, the model including the proxies not only confirms the conclusions of the initial model about the key variables of interest (⁠$Densitykl$ and $RCAkl$⁠) but also the theoretical reasoning about the role of productivity.

Observation of the coefficient estimates in columns (c) and (d) allows us to conclude that for a given product market, the exporters that enter that product market are those with higher productivity as predicted by the theory. In addition to productivity, firm age is found to be positively related to the probability to add product k to export range, all else being equal. One reason for this positive relationship might be that new-born firms may need time to mature on their core product before they consider diversifying. Also, affiliation to a group is found to have a positive effect on the probability to enter a new-product market. This binary variable can be considered as a very rough assessment of the impact of organizational ties, which may function as a conveyor of knowledge (within or across regions) and decrease fixed production costs or have an impact on consumer tastes. Finally, product-market growth appears to increase the probability to enter the market. Expansion of the product market in the world can allow firms to produce in larger quantities in order to meet the high demand, and to enjoy scale economies, which implies lower production fixed costs.

In the case of the coefficient estimates of $Densitykl$ our key variable of interest, the results imply that firms tend to enter into product markets that have more common and complementary capabilities with products in which their locality has RCA. This result is in line with the findings in Lo Turco and Maggioni (2016) for Turkish manufacturing firms that the introduction of new products depends on the availability of product-specific competencies in the locality. Recall that the $Densitykl$ variable is introduced in the regression equation in logarithms; thus, a percentage increase in the density of product k in locality l would lead all else being equal to a 6% increase in the odds of entering that product market for firms located in l [exp (5.840 × log (1.01)) = 1.059]. The positive and statistically significant coefficient estimate for $RCAkl$ reveals further that the probability that a firm diversifies into products that are already one of the core products of the locality is even higher. Odds of entering a product market for core goods is 126% higher than that for noncore goods [exp (0.815) = 2.259]. The coefficient estimates for $Densitykl$ and $RCAkl$ taken together suggest that among a locality’s core goods the one that is the most strongly linked to other core goods is the most likely to be added to the product range.

While Table 6 helps to explain product-market entry decisions, Table 8 presents the results for the determinants of product-market exit decisions formulated in equation (7). As in the previous table column (a) presents the estimations of the original equation, column (b) extends the specification in (a) by including additional firm-level controls, column (c) re-estimates the specification in (a) replacing the product-mix dummies with proxies, and finally column (d) extends the specification in (c) by adding further firm-level controls. The table shows that product-market exit decisions similar to product-market entry decisions are also shaped by productivity and product relatedness. More precisely, it shows that less productive exporters tend to drop products, and the choice of which product to drop is affected by the extent of local externalities. Products that require very different capabilities to what are already available and abundant in the locality are more likely to be dropped. A percentage increase in the density of product k in locality l leads to a 1.2% decrease in the odds of exiting from that product market for firms located in l all else being equal [exp (−1.255 × log (1.01)) = 0.988]. Whereas, the odds of dropping a core good is 38% lower than for a noncore good [exp (−0.480) = 0.619]. Therefore, firms are more likely to drop products that are not among the locality’s core products and are less related to the locality’s core goods.

The next set of results is based on Equation 8 and shifts the focus from churning decisions to the share of export revenues for a given product range. Hence, it complements the first set of results by investigating whether related diversification is sustained by related growth, which is also a conclusion that can be derived from the theoretical framework since it predicts higher operating profits for related products. Addressing this issue is essential also to understand consequences of these firm-level choices at the aggregate level for local comparative advantage.

Table 9 column (a) presents the results for equation (8), column (b) presents the results derived from replacing the product-mix dummies with proxies. Both models, despite differences in the goodness-of-fit and number of observations lead to the same conclusions. The positive and statistically significant coefficient estimates for $Densitykl$ and $RCAkl$ confirm the proposition about related growth. Among new-product markets, the growth path is the product that is densely linked to the locality’s core products. The growth increases if the product is already a core product of the locality. Inter-firm differences in growth in a new-product market (relative to the firm’s other new-product markets) stem not only from differences in coherence of that product with the local product space but depend also on how many other new-product markets the firm has entered, and how these products are related to the local product space (OTH_ik). The higher the number of product markets that are related highly to the local product space, the less the given product market will grow everything else being equal. Finally, the positive and statistically significant coefficient estimate Past_ik tells us that among new-product markets the ones that have a larger share of the export revenues are those that are introduced earlier.