Abstract

Belgium was one of the first continental countries to undergo industrialization and develop an extensive transportation infrastructure during the nineteenth century, completing the integration of its internal market by the early twentieth century. As such, the country is an ideal case study of the driving forces behind the decisions that industries made about where to locate. An analysis of factors embedded in both the Heckscher-Ohlin model and the new economic geography indicates that the main determinant of Belgium’s industrial locational pattern between 1896 and 1961 was proximity to regions with a high market potential.

In twentieth-century Belgium, the location of industrial activity shifted drastically from the southern part to the northern part of the country. Scholars have mainly studied this shift in a descriptive manner, following a standard historical approach. This article, however, proposes a quantitative analysis of the possible causes of this shift, using the Middelfart-Knarvik (mk) model of the location of economic activity, the potential of which economic historians have begun to explore recently. The main attractiveness of the mk model is its nesting within one framework the two main theories that offer an explanation for industrial location (agglomeration forces and endowments). In addition, by applying and estimating the mk model for industrial location in twentieth-century Belgium, we can check its usefulness in cases outside the Anglo-Saxon and Spanish countries, to which its application in historical analysis mainly has been limited.

Our results indicate that the shift in Belgium’s industrial pattern is explicable primarily by market potential (agglomeration effects). We found no indication that endowments had a significant impact. The literature contains similar results, though without consensus regarding the respective effect of the two prominent determinants of economic location. The capacity of the mk model to study this question may recommend it as a paradigm for new cases along the same lines, permitting more general conclusions about the circumstances under which industrial location finds a better explanation through neoclassical theory (endowments) or through the new economic geography (neg) (agglomeration forces).

Current Belgian political and public discourse is dominated by the differences between the wealthier Dutch-speaking Flemish region in the north and the poorer French-speaking Walloon region in the south. This linguistic division, a defining characteristic of Belgium from the last quarter of the nineteenth through the twentieth century, was reinforced by ideological differences—the north being mostly rural and Catholic and the south more liberal socially. Hence, Belgium evolved from a unitary state into a more regionalized federal conglomeration, although regionalization did not become legally established until the 1960s.1

According to Vogl and Hüning, however, language and culture did not constitute the only differences; economic origin mattered as well. Based on gdp estimations, Buyst recently concluded that a century ago, Wallonia was richer than Flanders. The reversal can be attributed to industrial location, as the work of Olyslager, Vanderrmotten, and De Brabander indicates. At the end of the nineteenth century, an industrial belt spanned the southern region from Mons to Liège. During the twentieth century, this belt lost its importance in favor of a northern counterpart, ranging from Brussels (in the center of the country) to Antwerp (in the northwest). Lefebvre and Buyst claimed that this shift could be explained via an agglomeration mechanism, like that embedded in the neg. They viewed the Brussels-Charleroi canal as a main determinant, but despite their focus on Brussels, they also underscored the port of Antwerp as an important factor. At that time, the disadvantages of the neg model were reputed to be its abstract and mathematical nature, which made it problematical for empirical research, and its apparent inability to trace long-term evolutions. This predicament seemed to preclude a precise reading of the shifting industrial pattern in twentieth-century Belgium and a definitive insight into its underlying dynamics.2

Around the same time, in the field of economic geography, the mk model was gaining in popularity. Although it emerged in 2001, some years passed before economic historians used it to examine long-term trends. The model nests both Hecksher-Ohlin (ho, the main example of traditional trade theory) and neg factors, quantifing the determinants of industrial locations. Because the robustness of the model has increased in endogeneity and heterogeneity in recent years, it can be used to study the shifting industrial pattern of twentieth-century Belgium in a quantitative way.3

This trade theory holds that assuming the absence of transportation costs and constant returns to scale, choices for industrial locations are based on factor availability and a region’s natural endowments. As initially theorized by Krugman in the 1990s, it holds that the synergy between transportation costs, returns to scale, and market potential are the main determinants of industrial locational decisions. In this model, industries respond to a decrease in transportation costs by migrating to economic centers. Given their high market potential, these agglomerations profit from productivity benefits and greater returns to scale.4

The contention herein is that for twentieth-century Belgium, market potential became more important because of the reduction of transportation costs that occurred during the nineteenth century. Belgium benefited from the industrial waves that swept over the continent, resulting in the development of manufacturing processes as well as an extensive railway and canal network, which forged better connections between the ports and the hinterland. Following the industrialization process, the country also benefited from the first wave of globalization during the latter half of the nineteenth century. The upshot was a decrease in transportation costs that permitted commodities to be imported from overseas through the Belgian ports; it was no longer necessary for industries to locate near natural endowments. Since transport along highways did not become commonplace until the 1960s, the research for this study concludes with the census of 1961.

The transportation methods in use at the time, railroads and water, did not undergo major transformations during the first half of the twentieth century. Although the railway network grew from close to 3,000 km in 1870 to as much as 5,000 km in 1912 (followed by a phase of stagnation), the main economic centers were already connected by rail before 1880. Transport by railroad democratized largely during the nineteenth century, when the government sought to increase the productive capacity of the country. At the end of the nineteenth century, the average distance from a village to the nearest station was not more than 6.43 km. The important cities were also connected by a network of canals and rivers; Belgium’s waterway infrastructure had been completed by the end of nineteenth century. Van der Herten reported that government investment in both waterways and railways enabled people and goods to move on a much larger scale than ever before.5

We also opt to end our research in 1961 because of the re-formation of the internal boundaries and the onset of the Expansion Laws during the 1960s. As a result of the guarantees and subsidies made possible by the Expansion Laws, the flow of foreign direct investment (fdi) into the northern part of the country increased, causing an artificial growth of market potential there. Hence, our research is based on three data points (1896, 1937, and 1961) for which industrial, agricultural, and population censuses are available, allowing a thorough analysis of the general evolution in industrial trends without the complications of the 1960s.6

This article contributes to the literature by examining a region that has heretofore been neglected. Until now, research has focused primarily on Spain and the Anglo-Saxon world; no research tests whether similar effects occur in different contexts. An overview of this literature is in the following section. As mentioned above, Belgium’s history offers an interesting natural experiment because of its robust transportation network. It also provides an opportunity to examine the formation of a new manufacturing belt, specifically to test whether the interaction between market potential and plant size is indeed crucial for the settlement of industrial concentrations; the largest and most productive firms were the first to move into the new markets. Another advantage of Belgium as a case study is that the historical-statistical data are detailed. Hence, we can examine location decisions with a nuts 3-level disaggregation from the late nineteenth century onward. We inject a further methodological wrinkle, proposing an extra robustness test to check for unobserved heterogeneity. Much of the preceding literature consults the oldest-available industrial input–output tables to extrapolate for earlier years, which can lead to estimation problems due to endogeneity; industries might have adapted to the abundance or scarcity of natural resources. The methodological section explains how we handle this matter.7

Literature and Theory

Two theories are important for our research—the traditional trade theory and neg. In the trade theory, geographical entities specialize in particular goods following a comparative advantage. Heckscher and Ohlin assumed that in a two-region, two-industry sector with a two-production factor set-up, efficiency would be similar in both regions, which implies that differences would be determined by natural resources. The industries in each country would specialize in a production that requires the inputs of its abundant natural resources. neg models hold that the interaction between transport costs, increasing returns to scale, and market size can lead to agglomerative effects. In a first step, provided that transportation costs are sufficiently low, firms locate near their largest markets, since they offer forward and backward linkages. The implication is that final products are cheaper to purchase and that more demand exists for intermediate goods. But transport costs must be sufficiently low; relocating is possible only when it is not too costly to transport the needed resources. These centripetal forces are strengthened by the incentives of industries that sell heavily to other industries or that use many intermediate goods to locate to these economic centers.8

As is common in the literature, we opt for a variant of the mk model, an enhancement of the Kim model, to test whether industrial locations are determined by the proximity to natural endowments or by market potential. The advantages of the mk model over the Kim model are its interactions between explanatory variables that allow analysis at the industrial level rather than the aggregated level. We follow the suggestions of Wolf and of Klein and Crafts, who dealt with the econometric problems of unobserved heterogeneity and endogeneity, respectively. More specifications for the model can be found in the methodology section (below).9

The conclusions of earlier research involving the United States, the United Kingdom, Spain, Poland, and the European Union are mixed. Kim concluded that natural resources were the most important determinants of the long-term evolution of industrial locations in the United States between 1860 and 1987. Midelfart-Knarvik et al.’s study of industrial locations in the European Union from 1980 to 1997 concluded that both theories were useful in explaining why and where industry located. In contrast, Klein and Crafts highlighted the importance of forward and backward linkages in 2012, stating that Kim’s results can be explained by the exclusion of these linkages, which lead to an overestimation of the Heckscher-Ohlin-variables. The Spanish authors concluded that neg forces became more important throughout the study period. Crafts and Mulatu performed a similar study about the United Kingdom and concluded that neoclassical and neg determinants were important between 1856 and 1929. Wolf reached the same conclusion for twentieth-century Poland.10

Lefebvre and Buyst first explored the ideas of the neg in relation to the industrial pattern of Flemish Brabant in 2007, finding it more useful than the Krugman model, which was too abstract to test empirically. Before their publication, most research about industrial locations lacked an empirical framework. Olyslager was the first to study the geographical locations of industries systematically. He was followed by Vandermotten and De Brabander. Although their works are still valuable sources for economic historians, they failed to quantify and explain the mechanism behind the shift that they described. In 2010, Buyst used the Geary-Stark method to estimate historical Belgian gdp amounts. Although this article does not discuss the shifting industrial pattern, it provides insight into the outcomes of this process. As a result of these advances, we now have a way to test the shifting industrial pattern more statistically and to unravel the dynamics in greater detail.11

The Model

To estimate the relative importance of neg versus neoclassical determinants in the location decisions of industries, Midelfart-Knarvik et al. used the following model:
formula
where is the share of industry k in region i at a certain time t. The variables popi and mani, which stand for the size of the population and the size of manufacturing employment in region i, are included to control for size effects; they control for the possibility that larger regions have stronger industries. The expression y[j]i is the strength of a regional characteristic j in region i—for example, the amount of mines in a certain region—and z[j]k is the value of an industrial characteristic—for example, the intensity of mine inputs in a certain industry j. The coefficients to be estimated are α, β, and K. To estimate this model, Midelfart-Knarvik et al. suggested the following regression:12
formula
β[j]y[j]iz[j]k captures the interaction between industrial and regional characteristics. β[j]γ[j]z[j]k controls for the impact of industrial and β[j]K[j]y[j]i of regional characteristics. If, say, β[j]γ[j]z[j]k measures a certain industry’s demand for agricultural land and β[j]K[j]y[j]i is a region’s endowment of agricultural land, β[j]y[j]iz[j]k represents the interaction between the demand for, and endowment of, land.
As in the work of Klein and Crafts, our dependent variable is measured in terms of employment, implying that we must control for changes in productivity across regions. As a solution, we estimate the equation using regional and industrial dummies, as suggested by Wolf, so that we can control for productivity changes at the regional and industrial levels. This strategy is acceptable because we are interested only in the coefficients of the interaction terms. More specifically, the following model is estimated using:
formula
with [j]i,t as the interaction between industrial and regional characteristics, Di,t and Dk,t as regional and industrial dummies, and c as a constant. measures the impact of the interaction terms that are of interest herein.13

Table 1 gives an overview of the industrial and regional characteristics (see Appendix 1 for a detailed description of these variables). The studies covered in the preceding section also discussed most of the regional characteristics. We included the mining industry, however, because we want to check whether its disappearance played an important role in the shift of economic power from the south to the north. Similarly, we added the use of intermediate goods from the mining industry to the list of industrial characteristics. One of the benefits of the Belgian case study is that its eighteenth-century statistical tradition results in data of high quality. Table 2 provides an overview of the interactions between regional and industrial characteristics, which are separated into neoclassical (column 1) and neg interaction terms (column 2).

Table 1

Industrial and Regional Characteristics

regional characteristicsindustrial characteristics
Agricultural surface Agricultural input 
Mines Input of mines 
Active population Labor intensity 
Educated population White-collar intensity 
Market potential Use of intermediate goods in % of output 
Sales to industry in % of output 
Size 
regional characteristicsindustrial characteristics
Agricultural surface Agricultural input 
Mines Input of mines 
Active population Labor intensity 
Educated population White-collar intensity 
Market potential Use of intermediate goods in % of output 
Sales to industry in % of output 
Size 

source Economic and population censuses of 1896, 1910, 1937, and 1961; the input–output table of 1959; the Annuaire statistique/statistische jaarboeken (statistical yearbooks) of the corresponding years.

Table 2

Interactions between Industrial and Regional Characteristics

neoclassicalneg
Agricultural surface × agricultural input (AS × AI) Market potential × intermediate goods (MP × IG) 
Mines × input of mines (M × IM) Market potential × sales to industry (MP × SI) 
Active population × labor intensity (AP × LI) Market potential × size (MP × S) 
Educated population × white-collar intensity (EP × WI)  
neoclassicalneg
Agricultural surface × agricultural input (AS × AI) Market potential × intermediate goods (MP × IG) 
Mines × input of mines (M × IM) Market potential × sales to industry (MP × SI) 
Active population × labor intensity (AP × LI) Market potential × size (MP × S) 
Educated population × white-collar intensity (EP × WI)  

sources Authors’ calculations based on the sources cited in Table 1 (abbreviations in parentheses).

Estimation Issues

As discussed by Klein and Crafts, a correct estimation of the model brings up three important issues—heteroskedasticity, endogeneity, and the use of panel data. We also have to pay attention to the availability of data.14

First of all, working in three dimensions (time, industry, and region) raises the risk of unincorporated cluster effects in the standard errors, invalidating the assumption of homoscedasticity. By using Miller, Gelbach, and Miller’s multi-way non-nested clustering, however, we can correct the standard errors for clustering on both the regional and industrial levels.15

Secondly, market potential, active craftsmen, and an educated population carry possible endogeneity issues. The basic estimation attempts to measure whether an increase in market potential will influence the location decision of firms. In other words, do firms locate near regions that offer forward and backward linkages? But including market potential as an explanatory variable could be problematical, because as the number and size of firms in a region increase, so do the regional gdp and the market potential. Thus, the dependent variable could also exert an influence on the explanatory variables. To correct for this tendency, we adopted an instrumental-variable approach. The literature largely relies on distance to a main economic center, but as explained by Mayer and Head, this approach could raise an endogeneity issue, since the choice of cities depends on economic potential. We follow the lead of these authors and use centrality, measured as , as an instrumental variable. The same endogeneity problem arises for the educated and active population, since firms can locate near a market that offers these advantages but also cause an inflow of population groups that manifest them. Therefore, we use the lagged values of the educated and active population as instrumental variables (1890 for 1896, 1896 for 1937, and 1937 for 1961).16

The third estimation issue is that the panel dataset is not continuous, covering only the years 1896, 1937, and 1961, which could cause difficulties if the characteristics of the Belgian economy changed drastically during the intervening years. A Chow test confirms a structural break in the parameters: The null hypothesis of constant parameters was rejected at 1 percent for the years 1896 and 1937 and at 5 percent for 1937 and 1961. The resulting fact that the panel aspect of the dataset cannot be exploited in the estimations is not a concern, however, since our interest lies in the changed coefficients of the three studied years. For this reason, our cross-sectional dataset allows us to provide an answer to the questions raised.

In addition to these methodological issues related to the specific model, we must also reckon with the availability of data. Because the input and output tables that the Midelfart-Knarvik model demands to estimate the industrial characteristics were not available in Belgium before 1959, we follow Martinez-Galarraga and use the aforementioned procentual inputs and outputs for the earlier years. Although this is a common procedure in the literature, it is a major disadvantage of the model when it is used to analyze historical data. Extrapolating the input–output tables means that the industrial characteristics are not absolutely exogenous, since industries probably adapted to the abundance or scarcity of natural resources. To be sure that the results remain similar, we run single-year regressions without interaction terms, following Kim, as a test of robustness. The model for this test then becomes:
formula
This approach, however, is missing one of the main advantages of the Midelfart-Knarvik model, into which all of the different industries enter individually because interaction terms are used. Because the Kim model simply offers a view of the determinants on an aggregated level, we use it only as a robustness test.17

Another issue relating to the availability of data concerns the industrial branches. In the censuses, the category “other industries” contains a collection of production units that could not be assigned to extractive industries, food, textiles, clothing, metal, or building. Hence, we were not able to find input–output figures for these industries. As a solution, we took averages for all of the other industries. But to check this strategy, we estimated all of the regressions twice, once with “other industries” included and once with it excluded. Since the results remain the same, this article incorporates only the estimations with all of the data included (the other regressions are available upon request).

The Data

This study discerns three benchmark years (1896, 1937, and 1961) for forty-one nuts-3 level districts (arrondissements) and seven industrial sectors—extractive, food, textile, clothing, metal, building, and other (all of the productive units not captured by the preceding categories). The data derive from the statistical yearbooks and agricultural, social, and economic censuses conducted by the National Institute of Statistics from the end of the nineteenth century onward. We needed data from all of these sources to estimate the outcomes for each of the three specified years. The years of the industrial censuses were 1896, 1937, and 1961; the agricultural censuses 1895, 1929, and 1959; and the population censuses 1900 and 1961. We also consulted a dataset that contains yearly population figures from 1880 to 1976.18

The first variable that can be calculated based on employment figures in the industrial censuses is the size of an industry k in the total manufacturing activity of region i: . This endogenous variable is defined as:
formula
with as the level of industrial activity, measured as employment per sector per region, of industry k in region i at time t. These figures pertain to the districts where people worked, not where they lived. Figure 1 plots the industrial centers of gravity, measured as employment in a region versus total employment, clearly showing the shift in industrial location from the south to the north. In 1896, economic activity was concentrated at the industrial centers of Brussels, Charleroi, and Liège in the south of the country. In 1937, activity shifted toward the north. Although Brussels remained important, Antwerp started to gain influence at its expense, and Liège fell back. The northeast also saw an increase in industrial concentration, primarily due to the quarrying of the mines in Limburg that began during the early twentieth century. Although this mining galvanized other industries to some extent, Antwerp and Brussels were much stronger attractions.
Fig. 1

Industrial Concentration in 1896, 1937, and 1961

Fig. 1

Industrial Concentration in 1896, 1937, and 1961

The second manufacturing belt continued to increase in importance until 1961. Similar concentration patterns are evident in the work of De Brabander and Olyslager. Appendix 2 provides an overview of the absolute number of employees in each industrial sector. The metal, food, and building industries show a strong increase throughout the twentieth century. Employment in the extractive industry drastically decreased, but employment in the textile and other industries remained stable.19

As Crafts stated, there are many options for calculating market potential, but deciding on the right one depends on its use. In this article, market potential is used to capture the physical proximity to consumer markets and industrial centers. We follow Crafts’ suggestion and employ the definition of economic potential proposed by Harris:
formula
with Mj as a measure of the size of region j and Dij as the distance between regions i and j. We use relative regional gdp to measure the size of a region. The gdp data for 1896, 1937, and 1970 come from Buyst, who studied the changing gdp figures in twentieth-century Belgium, comparing the total industrial gdp with the Belgian average. On this basis, we can find the relative market potential for each benchmark year, as Martinez-Galaragga determined. For 1961, we extrapolate the 1970 data. Based on these figures, we measure the market potential per district based on population figures, taking the provincial data to estimate the market potential of the district. These data come from the censuses and statistical yearbooks. The literature gives the internal expanse of the region as 0.333√ (surfacei/pi). Since we work on a nuts-3 level for a small country, the variance of the surfaces after using the Keeble formulae has no impact on the estimation results. This article incorporates the estimations without the Keeble formulae (the other results are available on request).20

The ranking of the different regions remained relatively stable. Figure 2 shows that the internal market potential per province changed little during the seventy-year period in our sample. Nevertheless, the market potential of the northern and central provinces consistently increased, whereas that of the southern regions deteriorated (the first four in Figure 2 from the left are the northern provinces; Brabant is in the middle; and the last four are the southern provinces).

Fig. 2

Development of the Relative Market Potential in Belgian Provinces—1896, 1937, and 1970

Fig. 2

Development of the Relative Market Potential in Belgian Provinces—1896, 1937, and 1970

As mentioned earlier, the regional characteristics comprise the active and educated population, agricultural area, and the number of mines, all of which were constructed by means of the industrial, agricultural, and population censuses. The selected variables were the number of mines, the agricultural area, the total active population, and the educated population. The data are in the logarithmic form. The data concerning the agricultural surface and number of mines were registered in the censuses. The active population is defined as men and women between the ages of fifteen and fifty-five (a category in the censuses). The educated population is measured differently due to data availability in 1890, 1896, and 1937/1961. For the instrumental data of 1890 and 1896, we selected the number of literates. Since the censuses of 1937 and 1961 no longer included literacy as a category, we re-defined education as schooling until age fourteen. According to these definitions, the number of active people and educated people remained relatively constant throughout the studied time period—as did the agricultural surface area—although the center of Belgium shows an increased concentration of people from 1937 onward. The mines fell primarily in the southern manufacturing belt in 1896 and 1937. In 1961, Limburg had a higher concentration of them.

The industrial characteristics are demand for an active and educated population, demand for coal and agricultural inputs, use of intermediate goods, number of sales to other industries, and the size of the production units (averages per industry). We constructed them from the earliest input–output table in Belgium (1959) and the annual statistical yearbooks of 1896, 1931, and 1961. Since Belgium had no input–output tables before 1959, we followed the approach of Martinez-Galarraga, selecting the oldest one and adopting the aforementioned procentual input and output for the earlier years. As a proxy of the demand for educated people, we selected the demand for white-collar workers, as is common in the literature. The resulting dataset contains information about the use of agricultural input (percentage of total inputs), the use of mines (percentage of total inputs), the demand for workers (total number of craftsmen), the demand for white-collar workers (total number of white-collar workers), the demand for intermediate goods (percentage of total inputs), sales to other industries (percentage of total sales), and the size of industries (number of employees).

The extractive industries required the most craftsmen; the food/textiles/clothing industries had the greatest need for white-collar workers; only the food industry required agricultural inputs; and the metal industry depended most heavily on mines. Four of the seven industries relied on sales to other industries—the extractive, metal, food, and building industries, as well as the “other industries” category. These industries also needed considerable input from other industries (more detailed overview of these industrial and regional characteristics available upon request).21

Empirical Results

This section presents the estimations and the robustness tests, followed by the interpretation of the results. Table 3 shows the results of the ordinary least square (ols) regression of the employment shares on the neoclassical and neg explanatory variables, with fixed effects on a regional and industrial level, as justified by the intuition that the regression does not capture certain geographical and industrial characteristics. For 1896, two coefficients are significant and have the expected sign—mines × input of mines and market potential × size. For 1937, as in 1961, the interaction between industrial size and market potential remains significant, but the interaction between mine endowment and intensity of usage loses its significance; market potential × size is the only coefficient that is significant, at the 1 percent level. The ols estimates indicate that market potential determined the shift in industrial locations during the twentieth century. In the next step, we test whether this finding holds after the suggested robustness tests.

Table 3

ols Using Regional and Industrial Dummies

189619371961
neoclassical variables 
AP × LI −0.0131 (0.0125) 2.33e-07 (1.34e-06) 1.16e-09 (2.60e-08) 
EP × WI 1.71e-09 (8.82e-08) 1.02e-06 (3.60e-06) 4.29e-07 (6.57e-07) 
AS × AI −4.10e-07 (3.11e-07) −4.76e-07 (5.82e-07) −5.70e-08 (6.17e-07) 
M × IM 6.75e-06** (3.29e-06) 4.37e-06 (4.52e-06) 0.00230 (0.00173) 
neg variables 
MP × IG 0.000333 (0.00101) −0.000271 (0.000954) −0.000718 (0.000632) 
MP × SI 0.000138 (0.000469) −0.000316 (0.000493) −2.07e-05 (0.000323) 
MP × S 0.000628*** (0.000156) 0.000495*** (−0.000271) 0.000127*** (2.23e-05) 
Constant 7.096 (9.808) −4.669 (9.518) −7.454*** (0.904) 
Regional dummy Yes Yes Yes 
Industry dummy Yes Yes Yes 
Observations 287 287 287 
R-squared 0.535 0.606 0.649 
189619371961
neoclassical variables 
AP × LI −0.0131 (0.0125) 2.33e-07 (1.34e-06) 1.16e-09 (2.60e-08) 
EP × WI 1.71e-09 (8.82e-08) 1.02e-06 (3.60e-06) 4.29e-07 (6.57e-07) 
AS × AI −4.10e-07 (3.11e-07) −4.76e-07 (5.82e-07) −5.70e-08 (6.17e-07) 
M × IM 6.75e-06** (3.29e-06) 4.37e-06 (4.52e-06) 0.00230 (0.00173) 
neg variables 
MP × IG 0.000333 (0.00101) −0.000271 (0.000954) −0.000718 (0.000632) 
MP × SI 0.000138 (0.000469) −0.000316 (0.000493) −2.07e-05 (0.000323) 
MP × S 0.000628*** (0.000156) 0.000495*** (−0.000271) 0.000127*** (2.23e-05) 
Constant 7.096 (9.808) −4.669 (9.518) −7.454*** (0.904) 
Regional dummy Yes Yes Yes 
Industry dummy Yes Yes Yes 
Observations 287 287 287 
R-squared 0.535 0.606 0.649 

***significant at 1%.

**significant at 5%.

*significant at 10%.

note Standard errors in parentheses.

As a first robustness test, we control for heteroskedasticity using clusters on the regional and industrial levels (Table 4). The use of multi-way clustering changes the results from Table 3. While the market potential × size interaction remains significant at the 1% level, for two out of the three years, certain Heckscher-Ohlin variables also become significant. The interaction term agricultural area × agricultural input becomes negative and significant in 1961. Since the use of agricultural inputs is less than 1 percent for almost all of the industrial branches, we run the same multi-way clustering regression but exclude the food industry, the only industrial sector highly dependent on agricultural inputs (23.74 percent). As expected, only market potential × size and mines × input of mines remain significant (see Appendix 3).

Table 4

ols Regressions Using Multi-Way Clustering

189619371961
neoclassical variables 
AP × LI −0.0131 (0.0116) 2.33e-07 (9.48e-07) 1.16e-09 (2.31e-08) 
EP × WI 1.71e-09 (8.33e-08) 1.02e-06 (2.40e-06) 4.29e-07 (4.92e-07) 
AS × AI −4.10e-07** (1.67e-07) −4.76e-07 (3.05e-07) −5.70e-08 (3.61e-07) 
M × IM 6.75e-06*** (2.07e-06) 4.37e-06** (1.97e-06) 0.00230** (0.00107) 
neg variables 
MP × IG 0.000333 (0.00105) −0.000271 (0.000659) −0.000718 (0.000554) 
MP × SI 0.000138 (0.000346) −0.000316 (0.000305) −2.07e-05 (0.000253) 
MP × S 0.000628* (0.000336) 0.000495*** (0.000170) 0.000127*** (4.54e-05) 
Constant 7.096 (9.082) −4.669 (6.665) −0.000718 (0.000554) 
Observations 280 287 287 
R-squared 0.535 0.606 0.649 
189619371961
neoclassical variables 
AP × LI −0.0131 (0.0116) 2.33e-07 (9.48e-07) 1.16e-09 (2.31e-08) 
EP × WI 1.71e-09 (8.33e-08) 1.02e-06 (2.40e-06) 4.29e-07 (4.92e-07) 
AS × AI −4.10e-07** (1.67e-07) −4.76e-07 (3.05e-07) −5.70e-08 (3.61e-07) 
M × IM 6.75e-06*** (2.07e-06) 4.37e-06** (1.97e-06) 0.00230** (0.00107) 
neg variables 
MP × IG 0.000333 (0.00105) −0.000271 (0.000659) −0.000718 (0.000554) 
MP × SI 0.000138 (0.000346) −0.000316 (0.000305) −2.07e-05 (0.000253) 
MP × S 0.000628* (0.000336) 0.000495*** (0.000170) 0.000127*** (4.54e-05) 
Constant 7.096 (9.082) −4.669 (6.665) −0.000718 (0.000554) 
Observations 280 287 287 
R-squared 0.535 0.606 0.649 

***significant at 1%.

**significant at 5%.

*significant at 10%.

note Standard errors in parentheses.

Finally, we check for potential endogeneity. According to neg, locations with a high market potential attract new firms, which in turn results in a higher market potential. In addition, the number of educated and active people in an area can also be affected by a higher market potential. To see whether this condition influences the results, we apply the Durbin-Wu-Hausman test with the concentration indices proposed by Head and Mayer and the lagged values of the educated and active population as instrumental variables (1890 for 1896, 1896 for 1937, and 1937 for 1961). The test rejects the null hypothesis of no selection bias in the results, undermining the estimations presented above. The instrumental variable estimations (Table 5) show that active population × craft intensity becomes significant instead of agricultural area × agricultural inputs. The general conclusion remains the same, however: In 1896, the neoclassical determinants were important, whereas from 1937 onward, market potential became the most important determinant of industrial location.22

Table 5

Instrumental Variables Distance

189619371961
neoclassical variables 
AP × LI 0.0165*** (0.00453) −2.93e-06 (1.78e-06) −4.83e-07 (2.04e-06) 
EP × WI 1.20e-07 (7.33e-08) −6.58e-06 (4.81e-06) −8.66e-07 (5.25e-06) 
AS × AI 9.80e-08 (1.92e-07) 6.18e-07 (5.14e-07) −7.93e-06 (2.03e-05) 
M × IM 1.02e-05*** (3.79e-06) 3.61e-06 (6.28e-06) −4.74e-06 (5.48e-06) 
neg variables 
MP × IG −0.000162 (0.00108) −0.000685 (0.000604) 8.41e-05 (0.00163) 
MP × SI 0.000165 (0.000844) 0.000449 (0.000362) −0.000536 (0.000730) 
MP × S 0.000454 (6.51e-05) 0.000117*** (2.00e-05) 0.00462*** (0.000988) 
Constant −5.225*** (0.473) −4.452*** (1.576) −6.484*** (0.376) 
Observations 287 287 287 
189619371961
neoclassical variables 
AP × LI 0.0165*** (0.00453) −2.93e-06 (1.78e-06) −4.83e-07 (2.04e-06) 
EP × WI 1.20e-07 (7.33e-08) −6.58e-06 (4.81e-06) −8.66e-07 (5.25e-06) 
AS × AI 9.80e-08 (1.92e-07) 6.18e-07 (5.14e-07) −7.93e-06 (2.03e-05) 
M × IM 1.02e-05*** (3.79e-06) 3.61e-06 (6.28e-06) −4.74e-06 (5.48e-06) 
neg variables 
MP × IG −0.000162 (0.00108) −0.000685 (0.000604) 8.41e-05 (0.00163) 
MP × SI 0.000165 (0.000844) 0.000449 (0.000362) −0.000536 (0.000730) 
MP × S 0.000454 (6.51e-05) 0.000117*** (2.00e-05) 0.00462*** (0.000988) 
Constant −5.225*** (0.473) −4.452*** (1.576) −6.484*** (0.376) 
Observations 287 287 287 

***significant at 1%.

**significant at 5%.

*significant at 10%.

note Standard errors in parentheses.

As mentioned earlier, we checked our results using Kim’s methodology to ensure that the coefficients are not biased due to the extrapolation of an input–output table from 1959. The results in Appendix 4 show the regressions with the inclusion of regional and industrial dummies. In 1896, only the active population is significant at the 10 percent level. In 1937 and 1961, the positive coefficient of market potential becomes significant at the 1 percent level, but no other coefficients are significant. Yet this approach is missing one of the main advantages of the mk model over the Kim model, namely, that all different industries enter the model individually because of the interaction terms. For this reason, it finds use in this case solely as a robustness test. All of the robustness tests indicate that market potential indeed became a more important determinant of industrial location during the twentieth century.23

Thus far, we have examined only the significance of the variables. Given our interest in which of the two theories better explains Belgium’s industrial location patterns, we must also examine the relative size coefficients. Table 6 shows the beta coefficients of the interaction terms for the fixed-effects regression, allowing a direct comparison of the size of the coefficients: The higher the beta-coefficient, the larger is the impact on the dependent variable. The upshot is that the interaction between market potential and size is not only the most significant; it is also the most important explanation in terms of the size of the effects. The relationship of the coefficients remains stable with the robustness checks.

Table 6

Beta Coefficients for the Interaction Variables

189619371961
AP × LI −1.54 0.87 0.01 
EP × WI 0.01 0.15 0.07 
AS × AI −0.14 −0.09 −0.01 
M × IM 0.11 0.05 0.07 
MP × IG 0.08 −0.05 −0.19 
MP × SI 0.04 −0.07 −0.01 
MP × S 0.24 0.34 0.26 
189619371961
AP × LI −1.54 0.87 0.01 
EP × WI 0.01 0.15 0.07 
AS × AI −0.14 −0.09 −0.01 
M × IM 0.11 0.05 0.07 
MP × IG 0.08 −0.05 −0.19 
MP × SI 0.04 −0.07 −0.01 
MP × S 0.24 0.34 0.26 

note Significant variables are indicated in bold.

Regardless of the estimation technique, the coefficients all indicate that in 1896, the variable mines × input of mines was most important in explaining industrial locations. In 1937 and 1961, market potential × size was the only significant variable. Although the other market-potential variables are not significant, it does hold that bigger firms tended to locate near regions with a high market potential. As shown in Figure 2, the provinces with the highest market potential were those of Brabant and Antwerp. These regions also attracted most of the large industries, thus forming the northern manufacturing belt. The question remains, however, why the market potential of these regions became increasingly more important than that of the southern regions. The answer lies in Buyst’s research, which highlighted the importance of the coastal regions in the twentieth century.24

At the end of the nineteenth century, industry was still concentrated in the southern industrial belt. The region was rich in coal mines that attracted many industries. As such, it was one of the economic powerhouses of the time, which explains the significance of the interaction mines × input of mines. From the beginning of the twentieth century onward, however, the northern regions became the industrial heavyweights of Belgium. The reduction of transportation costs, as anticipated by the neg, led to an alternate industrial pattern. The Belgian harbors profited optimally from the first wave of globalization during the second half of the nineteenth century, and links with the hinterland by road and canal emerged, creating an attractive region where industries could settle. The zone around the Brussels–Antwerp canal also began to industrialize because of its proximity to sea access. Because of the infrastructural work that allowed ocean ships into Brussels, the Antwerp–Brussels region could benefit from the drawing forces of market potential that eventually formed the industrial belt of the north.25

The reductions in transportation costs led industries to locate away from regions with natural resources and near regions with high market potential. In the mid-1950s, this development, which intensified with the importation of cheap oil, brought the unprecedented opportunity to transport massive amounts of bulk cargo. More industries arrived to exploit the burgeoning transport by sea, the low wages, and the populated hinterland. The importation of both coal and oil from overseas had a negative impact on the Walloon industrial regions. The coal industry could not cope with the international competition. Industries had an alternative place from which to import their commodities; no longer obliged to locate in the southern part of Belgium, they could go north, near regions with a high market potential and import what they needed. The regressions indicate that, indeed, the largest firms went to regions with a high market potential during all three of the benchmark years.

In 1896, when the southern provinces had the highest market potential, the biggest firms, which located near regions with mines, were automatically in an area of high market potential. In the next two benchmark years, the interaction mines × input of mines lost its significance, whereas market potential × size remained significant. Hence, the shifting industrial pattern appears to be explicable by the changing pattern of market potential. The locational decisions of the largest firms, attracted by access to big markets, were an important determinant in the establishment of a new manufacturing belt. The other market-potential variables might not be in evidence at this point, since they were expected to have an impact later in the twentieth century, but checking this assumption against an alternative for the extrapolated input–output tables could be a useful topic for future research.

The results herein fall in line with studies about locational patterns in Spain and the United States where neg-type mechanisms were identified as the driving forces. It is striking that of all these mechanisms, the interaction between market potential and size has proven to be the most relevant one so far. We cannot state that the same mechanism holds true for every country examined, but other researchers—Crafts and Mulatu in work about Britain and Wolf in work about Poland, to name two—have emphasized the importance of traditional factor endowments.26

The new econometric models, which are much more advanced than the pioneering ones advanced by such scholars as Buyst, and Olyslager and De Brabander, provide an opportunity to test the causes of the industrial shift from the south to the north of Belgium empirically.

We answered the central question of this article—what determined the location of industry in Belgium between 1896 and 1961?—by gathering data on a nuts-3 level and employing a variation of the mk model, which permitted us to assess the importance of neoclassical and neg forces. In other words, we examined whether industries tended to locate near natural endowments or near economic agglomerations. Our use of a nested multi-way clustering method, instrumental variable-regressions, and the Kim model compensated for potential effects of heteroskedasticity and endogeneity, and our use of industrial and regional dummies helped to account for differences in productivity.

Regardless of estimation technique, we found that both neg and neoclassical factors were at work, though neg factors played a significantly more important role from the twentieth century onward. Because industries at the end of the nineteenth century located near endowments—coal in particular—the Walloon region, which was rich in mines, thrived. At the beginning of the twentieth century, however, when importing coal from overseas became cost-effective, heavy industries began to locate near harbors and coastal areas, and production units positioned themselves close to other industries, highly populated markets, export possibilities, and coal imports. By the 1950s, the northern regions had become the economic powerhouse of Belgium, whereas the Walloon region continued on a downward spiral. As expected, the interaction between market potential and plant size was crucial in establishing the northern manufacturing belt.

APPENDIX 1: OVERVIEW OF REGIONAL AND INDUSTRIAL CHARACTERISTICS

7

regional characteristics 
Agricultural surface Agricultural land per district in square km 
Mines Absolute number of mines per region 
Active population Absolute number of population between 15 and 55 years old 
Educated population For 1896: literates; for 1937/1961: absolute number of persons who went to school until the age of 14 
Market potential Physical proximity to regions with a high gdp 
industrial characteristics 
Agricultural input The number of inputs from agriculture needed by a certain industry as a % of total inputs 
Input of mines The number of inputs from the mines needed by a certain industry as a % of total inputs 
Labor intensity The number of workers needed by a certain industry 
White-collar intensity The number of white-collar workers needed by a certain industry 
Use of intermediate goods, % of output The number of intermediate goods needed by a certain industry as a % of total inputs 
Sales to industry, % of output The number of goods sold to other industries by a certain industry as a percentage of total sales 
Size The number of employees by industry 
regional characteristics 
Agricultural surface Agricultural land per district in square km 
Mines Absolute number of mines per region 
Active population Absolute number of population between 15 and 55 years old 
Educated population For 1896: literates; for 1937/1961: absolute number of persons who went to school until the age of 14 
Market potential Physical proximity to regions with a high gdp 
industrial characteristics 
Agricultural input The number of inputs from agriculture needed by a certain industry as a % of total inputs 
Input of mines The number of inputs from the mines needed by a certain industry as a % of total inputs 
Labor intensity The number of workers needed by a certain industry 
White-collar intensity The number of white-collar workers needed by a certain industry 
Use of intermediate goods, % of output The number of intermediate goods needed by a certain industry as a % of total inputs 
Sales to industry, % of output The number of goods sold to other industries by a certain industry as a percentage of total sales 
Size The number of employees by industry 

APPENDIX 2: NUMBER OF EMPLOYEES PER INDUSTRY

8

189619371961
Extractive 167,319 174,964 104,115 
Food 90,436 85,394 132,158 
Textile 168,491 162,138 142,591 
Clothing 137,963 53,915 99,949 
Metal 132,163 217,713 456,938 
Building 93,576 97,036 249,071 
Other 308,512 351,864 351,562 
Total 1,098,460 1,143,024 1,536,384 
189619371961
Extractive 167,319 174,964 104,115 
Food 90,436 85,394 132,158 
Textile 168,491 162,138 142,591 
Clothing 137,963 53,915 99,949 
Metal 132,163 217,713 456,938 
Building 93,576 97,036 249,071 
Other 308,512 351,864 351,562 
Total 1,098,460 1,143,024 1,536,384 

APPENDIX 3: REGRESSION TO CHECK THE IMPORTANCE OF MINES AND AGRICULTURAL LAND ENDOWMENTS

9

variables/lnski (food excluded)lnski (metal excluded)
neoclassical forces 
AP × LI 6.22e-05 (4.42e-05) 8.66e-05** (3.44e-05) 
EP × WI −1.15e-07 (1.64e-07) −1.53e-07 (1.15e-07) 
AS × AI 1.88e-05 (1.96e-05) −1.32e-06** (6.46e-07) 
M × IM 0.00497*** (0.00186) −0.0321 (0.0420) 
neg variables 
MP × IG 0.000502 (0.000658) −0.000208 (0.000693) 
MP × SI 2.66e-05 (0.000860) 0.000461 (0.000662) 
MP × S 8.06e-05*** (2.04e-05) 8.16e-05*** (2.03e-05) 
AP × LI 
Constant −6.146*** (0.310) −5.964*** (0.401) 
Observations 232 232 
variables/lnski (food excluded)lnski (metal excluded)
neoclassical forces 
AP × LI 6.22e-05 (4.42e-05) 8.66e-05** (3.44e-05) 
EP × WI −1.15e-07 (1.64e-07) −1.53e-07 (1.15e-07) 
AS × AI 1.88e-05 (1.96e-05) −1.32e-06** (6.46e-07) 
M × IM 0.00497*** (0.00186) −0.0321 (0.0420) 
neg variables 
MP × IG 0.000502 (0.000658) −0.000208 (0.000693) 
MP × SI 2.66e-05 (0.000860) 0.000461 (0.000662) 
MP × S 8.06e-05*** (2.04e-05) 8.16e-05*** (2.03e-05) 
AP × LI 
Constant −6.146*** (0.310) −5.964*** (0.401) 
Observations 232 232 

***significant at 1%.

**significant at 5%.

*significant at 10%.

note Standard errors in parentheses.

APPENDIX 4: ESTIMATIONS FOLLOWING THE KIM METHODOLOGY (1999)

10

189619371961
neoclassical variables 
Active population 6.730* (4.039) −0.0717 (0.599) 1.000 (0.657) 
Educated population −4.729 (4.486) 1.744 (1.166) −0.856 (1.273) 
Agricultural areal 0.131 (0.719) 0.759 (0.630) −0.0990 (0.694) 
Mines 0.0552 (0.0654) 0.0238 (0.0724) −0.0315 (0.0778) 
Market potential 0.00227 (0.0118) 0.760*** (0.223) 1.185*** (0.248) 
Constant −14.00*** (2.564) −16.47*** (4.343) −15.81*** (4.841) 
Observations 287 287 273 
R-squared 0.267 0.284 0.225 
189619371961
neoclassical variables 
Active population 6.730* (4.039) −0.0717 (0.599) 1.000 (0.657) 
Educated population −4.729 (4.486) 1.744 (1.166) −0.856 (1.273) 
Agricultural areal 0.131 (0.719) 0.759 (0.630) −0.0990 (0.694) 
Mines 0.0552 (0.0654) 0.0238 (0.0724) −0.0315 (0.0778) 
Market potential 0.00227 (0.0118) 0.760*** (0.223) 1.185*** (0.248) 
Constant −14.00*** (2.564) −16.47*** (4.343) −15.81*** (4.841) 
Observations 287 287 273 
R-squared 0.267 0.284 0.225 

***significant at 1%.

**significant at 5%.

*significant at 10%.

note Standard errors in parentheses.

Notes

1 

A third, much smaller, linguistic zone—located on the eastern border of the Liege province—not as prominent in the political debate, is predominantly German-speaking. It was appended to Belgium following World War I. Lode Wils, Political Development in Belgium 1894–1914 (Haarlem, 1978).

2 

Guido de Brabander, Regional Specialization, Employment and Economic Growth in Belgium between 1846 and 1970 (New York, 1981); Paul Olyslager, De localisering der Belgische nijverheid (the location of Belgian industry) (Antwerp, 1947); Christian Vandermotten, “Les structures spatiales de l’économie belge et de leur évolution de la période fordiste à ajourd’hui” (the spatial structure of the Belgian economy), Hommes et Terres du Nord, LVII (1985), 186–197; Wim Lefebvre and Erik Buyst, “Enkele verklaringsfactoren voor de regionaal gedifferentieerde industriële ontwikkeling van Vlaams-Brabant tussen 1896 en 1961” (some regional determinants of the differentiated regional development of Flemish Brabant between 1896 and 1961), Belgisch tijdschrift voor nieuwste geschiedenis, XXXVII (2007), 41–77.

3 

Karen Midelfart-Knarvik et al., “The Location of European Industry,” in European Economy, Special Report No. 2 (2002), 213–269.

4 

Eli Heckscher and Bertil Ohlin, Heckscher-Ohlin Trade Theory (New York, 1991); Paul Krugman, “Increasing Returns and Economic Geography,” Journal of Political Economy, XCIX (1991), 483–499; Yuko Aoyama, J. T. Murphy, and S. Hason, “Key Concepts in Economic Geography,” Economic Geography, LXXXVIII (2012), 225–226; Donald Davis and D. E. Weinstein, “Market Access, Economic Geography and Comparative Advantage: An Empirical Test,” Journal of International Economics, LIX (2003), 1–23.

5 

Between 1840 and 1905, revenue per traveler dropped from 1.84 to 0.55 Belgian francs, and the number of travelers rose from 4,046,950 to 77,394,090. The transportation of goods witnessed a similar evolution as the prices decreased from 19 to 2.9 Belgian francs per ton during the second half of the nineteenth century. Bart van der Herten, M. van Meerten, and G. Verbeurgt, Sporen in België: 175 jaar spoorwegen 75 jaar NMBS (railroads in Belgium: 175 years of railroads 75 years NMBS) (Leuven, 2001); P. Ton, De Belgische waterwegen (Delft, 1983); van der Herten, België onder stoom: transport en communicatie tijdens de 19de eeuw (Leuven, 2004).

6 

The Expansion Laws, an attempt to attract more investment after the economic crisis of 1959, managed only to increase the ongoing industrialization of Flandres—the sole region attractive to foreign investors. Van der Herten, van Meerten, and Verbeurgt, Sporen in België.

7 

“Nuts 3” refers to the geographical level of forty-four arrondissementen, or “districts.” Julio Martinez-Galarraga, “The Determinants of Industrial Location in Spain, 1856–1929,” Explorations in Economic History, XLIX (2010), 255–275; Joan Roses, “Why Isn’t the Whole of Spain Industrialized? New Economic Geography and Early Industrialization, 1797–1910,” Journal of Economic History, LXIII (2003), 995–1022; Richard Baldwin and Toshihiro Okubo, “Heterogneous Firms, Agglomerations and New Economic Geography: Spatial Selection and Sorting,” Journal of Economic Geography, VI (2005), 323–346.

8 

Martinez-Galarraga, “Determinants of Industrial Location.”

9 

Midelfart-Knarvik, Overman, and Venables, Comparative Advantage and Economic Geography: Estimating the Location of Production in the EU, Centre for Economic Policy Research discussion paper 2618 (London School of Economics, 2001); Sukkoo Kim, “Expansion of Markets and the Geographic Distribution of Economic Activities: The Trends in U.S. Regional Manufacturing Structure, 1860–1987,” Quarterly Journal of Economics, CX (1995), 881–908; idem, “Regions, Resources and Economic Geography: Sources of U.S. Regional Manufacturing Structure, 1880–1987,” Regional Science and Urban Economics, XXIX (1999), 1–32; Alexander Klein and Nicholas Crafts, “Making Sense of the Manufacturing Belt: Determinants of U.S. Industrial Location, 1880–1920,” Journal of Economic Geography, XII (2012), 775–809.

10 

Kim, “Expansion of Markets”; Midelfart-Knarvik et al., “Comparative Advantage”; Klein and Crafts, “Making Sense of the Manufacturing Belt”; Perez Betran, “Industria y crecimiento económico en el primer tercio del siglio XX. España, 1913–1929,” unpub. Ph.D. diss. (Univ. of Valencia, 1995); Daniel Tirado, E. Paluzie, and J. Pons, “Economic Integration and Industrial Location: The Case of Spain before World War I,” Journal of Economic Geography, II (2001), 343–363; Martinez-Galarraga, “Determinants of Industrial Location”; Crafts and Abay Mulatu, “What Explains the Location of Industries in Britain, 1871–1931?” Journal of Economic Geography, V (2005), 499–518; Nikolaus Wolf, “Endowments vs. Market Potential: What Explains the Relocation of Industry after the Polish Reunification in 1918?” Explorations in Economic History, XLIV (2007), 22–42.

11 

Lefebvre and Buyst, “Enkele verklaringsfactoren”; de Brabander, “Regional Specialization”; Olyslager, De localisering der Belgische nijverheid; Vandermotten, “Les structures spatiales”; Buyst, “Reversal of Fortune in a Small, Open Economy: Regional GDP in Belgium, 1896–2000,” Rivista di Storia Economica, XXVI (2010), 75–92.

12 

Midelfart-Knarvik, Overman, and Venables, Comparative Advantage; Midelfart-Knarvik et al., “Location of European Industry.”

13 

Klein and Crafts, “Making Sense of the Manufacturing Belt”; Wolf, “Endowments vs. Market Potential.”

14 

Klein and Crafts, “Making Sense of the Manufacturing Belt”; Jeffrey Woolridge, Econometric Analysis of Cross Section and Panel Data (New York, 2002).

15 

Colin Cameron, Jonah Gelbach, and Douglas Miller, “Bootstrap-Based Improvements for Inference with Clustered Errors,” Review of Economics and Statistics, XC (2008), 414–427.

16 

Keith Head and Thierry Mayer, “Regional Wage and Employment Responses to Market Potential in the EU,” Regional Science and Urban Economics, XXXVI (2006), 573–594.

17 

Martinez-Galarraga, “Determinants of Industrial Location”; Kim, “Regions, Resources and Economic Geography: Sources of U.S. Regional Manufacturing Structure, 1880–1987,” Regional Science and Urban Economics, XXIX (1999), 1–32.

18 

For the agricultural censuses of 1895 and 1959; the industrial censuses of 1896, 1910, 1937, and 1961; population censuses of 1900, 1910, and1961, see Ministère de l’Intérieur, Annuaire statistique/statistisch jaarboek (statistical yearbook) (Brussels, 1896, 1937, and 1961). Recensement général des industries et des métiers (31 octobre 1896): Dénombrement A. Répartition par commune des industries et des métiers (Brussels, 1900), I–II,; economic and industrial censuses of 1937 (Brussels), IX; industrial and trade census of 1961 (Brussels, 1965–1967), IV; agricultural censuses of 1895 and 1959 (Brussels), IV, VII; population censuses of 1900 and 1963 (Brussels), II; Mouvement de la Population et de l’Etat Civil (population movements and marital-status changes), the registers of population and death and the population and socio-economic censuses. These sources are all unpublished. More information can be obtained from LOKSTAT.

19 

De Brabander, Regional Specialization; Olyslager, De localisering der Belgische nijverheid.

20 

Crafts, “Market Potential in British Regions, 1871–1931,” Regional Studies, XXXIX (2005), 1159–1166; Chauncy Harris, “The Market as a Factor in the Localization of Industry in the United States,” Annals of the Association of American Geographers, XLIV (1954), 315–348; Martinez-Galarraga, “Determinants of Industrial Location”; Buyst, “Reversal of Fortune in a Small, Open Economy.” This approach was suggested by David Keeble, P. L. Owens, and C. Thompson (1982), “Regional Accessibility and Economic Potential in the European Community,” Regional Studies, XVI (1982), 419–432, in which the formula 0.333√ (surfacei/pi) appears. One problem with defining market potential in this manner is that it does not take foreign trading partners, whether in neighboring countries or overseas, into account. One possible way to correct for the presence of the international harbor in Antwerp is to estimate the impact of incoming and outgoing flows on market potential (as in Klein and Crafts, “Making Sense of the Manufacturing Belt”); however, we lack the data to make this estimation for each year. Moreover, this effect is already partly captured because proximity to the harbor is incorporated by the local gdp. The influence of neighboring countries is also taken into account through the use of the external market potential, which is calculated by replacing the Belgian districts in equation 4 with the following neighboring districts outside Belgium—North Rhine, Rhineland, Lorraine, Champagne, Nord-Pas-de-Calais, Northern-Brabant, Limburg, and Sealand. The data come from Brian R. Mitchell, International Historical Statistics: Europe 1750–2000 (New York, 2003) and J. Bolt and Jan Luiten Van Zanden, “The First Update of the Maddison Project; Re-Estimating Growth Before 1820,” Maddison Project Working Paper 4 (2013). Unfortunately, we had to omit Luxembourg because historical gdp data for this region is unavailable. The smallest correlation between internal and external market potential over time is 0.75, which is likely explained by the geographical nature of Belgium. Because it is only a small country, the centrality of a district is more important than its GDP. We perform our regressions with the internal market potential, since these data are more reliable. As a control, we included the total market potential (these results are available upon request).

21 

Martinez-Galarraga, “Determinants of Industrial Location”; R. A. Straelen and P. H. Virenique, De input-output analyse: een methode voor het kwantitatief onderzoek der economische structuur (the input–output analysis: a method for quantitative research on economic structure) (Leuven, 1961).

22 

Head and Mayer, “Regional Wage and Employment Responses,” 573–594.

23 

Kim, “Regions, Resources and Economic Geography.”

24 

Buyst, “Reversal of Fortune in a Small, Open Economy.”

25 

The canal was not deep enough for large ocean ships to travel from Brussels to Charleroi. Not until the 1960s, with the installation of the Ronquières ship lift, could heavier ships sail to Charleroi. See Lefebvre and Buyst, “Enkele verklaringsfactoren.” Buyst, “Reversal of Fortune in a Small, Open Economy.”

26 

Klein and Crafts, “Making Sense of the Manufacturing Belt”; Martinez-Galarraga, “Determinants of Industrial Location”; Wolf, “Endowments vs. Market Potential”; Betran, “Industria y crecimiento económico”; Tirado, Paluzie, and Pons, “Economic Integration and Industrial Location”; Crafts and Mulatu, “What Explains the Location of Industries in Britain,” 499–518.

Author notes

Stijn Ronsse is Research Assistant, Faculty of Economics and Business Administration, Ghent University. He is the author of “De consumptierevolutie in het achttiende- en vroeg negentiende-eeuwse Gentse advertentiewezen,” Handelingen der Maatschappij Voor Geschiedenis en Oudheidkunde te Gent, LXII (2008), 161–184.

Glenn Rayp is Professor of Economics, Ghent University. He is the editor, with Michel Dumont, of International Business, Not As Usual (Philadelphia, 2011); the author, with Ilse Ruyssen, of “Determinants of Intraregional Migration in Sub-Saharan Africa 1980–2000,” Journal of Development Studies, L (2014), 426–443.