Spatial concentration of manufacturing industries in the United States: re-examination of long-run trends

Abstract

1. Introduction

It is well known that patterns of regional specialization and the spatial concentration of American manufacturing industries have changed markedly over time. A standard narrative concerns the rise and fall of the manufacturing belt in the mid-19th century and the second half of the 20th century, respectively. This is seen as a key aspect of a pattern of divergence followed by convergence of US regions.

Kim (1995) provided a much-cited quantitative account of these trends. He calculated Hoover’s coefficient of localization for two-digit industries through time and found that the weighted average rose from 0.243 in 1880 to 0.316 in 1927 before falling to 0.197 in 1987. Krugman (2009) saw this experience in terms of new economic geography with the success of the manufacturing belt based on a phase of increasing returns in manufacturing but with the applicability of this model evaporating in the late 20th century.

In this paper we seek to re-examine and improve on these accounts. First, we take advantage of improved measurement techniques to estimate the extent of spatial concentration allowing for industrial structure and the checkerboard problem. To do this, we use an approximation to a spatially weighted Ellison–Glaeser index (henceforth EG index). Second, we highlight the importance of changing locations patterns within the manufacturing belt, and the propensity of manufacturing to move outside the manufacturing belt already before World War II. We describe a clear tendency to spatial dispersion even during the heyday of a rising size of plants.

In order to re-examine long-run trends in the spatial concentration of US manufacturing industries we construct a new dataset which permits the calculation of a spatially adjusted version of the EG index at both SIC2 and SIC3 levels for selected census years from 1880 through 2007. To circumvent data limitations we use the spatially weighted version of the Maurel and Sedillot (1999) adaptation of the EG index, which does not require plant-level employment data. Construction of the index required assignment of industries into SIC categories and a procedure to deal with problems posed by withholding of data to prevent identification of individual firms.

Our main findings are as follows. First, the weighted average of the spatially weighted EG index for SIC3 industries is at its maximum in 1880 at 0.223 after which it declines slowly to 0.183 in 1940 and then more rapidly to 0.098 in 2007. Unlike Kim (1995), we do not find an episode of increasing spatial concentration in the early 20th century. Spatial weighting is important in arriving at this conclusion. Second, increasing spatial dispersion over time is a general experience across American manufacturing industries over the long run and especially after 1940. At SIC2 level, all but one sectors have lower spatial concentration in 2007 than either in 1880 or in 1940 while 17 of 20 industries were already more dispersed in 1940 than in 1880. At SIC3 level, in 12/20 SIC2 categories at least two thirds of the constituent SIC3 industries were more dispersed in 2007 than in 1880 and in 12/20 SIC2 categories the same was true for 1940 compared with 1880.

Third, even so, it is important to recognize that almost all SIC3 industries always exhibit spatial concentration in the sense that their spatially weighted EG index score is positive and significantly different from zero. This is the case even at the end of the period when spatial concentration has generally declined. In fact, all 20 exceptions of 1300 observations occur before 1947. The average of 0.098 in 2007 is at a level where it can be thought of as economically significant according to the criterion proposed by Ellison and Glaeser (1997). It would be incorrect to suppose that spatial concentration of manufacturing industry was no longer an important phenomenon in the early 21st century.

2. Literature Review

The relative decline of the manufacturing belt in the second half of the 20th century is well known and features prominently in economists’ reviews of the evolution of American industrial geography. Krugman (2009) in his Nobel Prize Lecture highlights that the manufacturing belt began to dissolve after World War II, while Holmes and Stevens (2004) in their handbook chapter stress that as late as the 1950s manufacturing activity was still heavily concentrated in the North East and Upper Midwest around the Great Lakes in the manufacturing belt after which time it moved out and into other parts of the country. The data reported in table 1 are consistent with these accounts in that they show 72.5 percent of manufacturing employment was in the manufacturing belt in 1947, but this share fell to only 42.9 percent in 2007.

That said, the economic geography literature has always recognized that the spatial distribution of manufacturing had evolved considerably before World War II. Already in the 1930s and 1940s geographers were discussing the “decentralization of industrial activity.” Smith (1947) in a quantitative analysis of manufacturing employment commented on a steady movement in the direction of decentralization. Hoover (1948) noted a trend toward more equal inter-regional distribution of manufacturing for many decades prior to 1940 and pointed out that the locational histories of many industries involved an early stage of increasing concentration followed by a later stage of re-dispersion. Easterlin (1960) found that there had been a substantial shift in the location of manufacturing between 1869 and 1947 and calculated that a minimum of 30 percent of wage earners in 1947 would need to be relocated to restore the 1869 percentage distribution by state (specifically, of the 30 percent, 8.3 percentage points accrued between 1889 and 1909, 10.8 percentage points between 1909 and 1929 and 5.3 percentage points between 1929 and 1947.). Eriksson et al. (2019) documented the spread of manufacturing between 1910 and 1940 noting the decline of New England and the Northern Great Lakes region and the expansion of the Southern Great Lakes region and most of the Appalachians.

This also is reflected in table 1 where it is seen that New England declined from 24.0 percent of manufacturing employment in 1880 to 12.5 percent in 1940 while the East North Central region rose from 19.1 percent to 27.9 percent. Table 1 also shows that there had been a notable decrease in the share of the manufacturing belt between these two dates from 87.2 percent to 73.6 percent. The point to note is that while the manufacturing belt still accounted for much of American manufacturing employment in the 1940s it was already a good deal less dominant than in the 1880s. We should also note that the overall share of manufacturing employment shows a declining trend since the WWII as noted in Fort et al. (2018).

These developments in shares of manufacturing employment were related to the pattern of trade within the United States. By 1949, the earliest date for which railroad freight data are available, as is reported in table 2, the East North Central region was responsible for more inter-state trade than New England and Middle Atlantic combined while West North Central and West South Central together exceeded Middle Atlantic while accounting for 21.7 percent of trade despite having only 9.4 percent of employment, and California was the source of nearly as much inter-state trade (3.3%) as New England (3.5%).

A staple finding of the literature on the location of manufacturing is that industries with larger plant sizes tend to have higher levels of geographic concentration (Holmes and Stevens, 2004). The basic new economic geography model reviewed by Krugman (2009) predicts that industry will concentrate in the core region with the best market access if economies of scale are large relative to transport costs. Kim (1995) pointed to a rise in the scale of production as reflected by the size of plants measured in terms of employment as a key factor in first rising and then declining spatial concentration over the course of the 20th century. Table 3 reports average plant size at the SIC two-digit level, and this confirms that plant sizes were generally rising until at least the 1940s but were generally falling in the decades towards the end of the century. However, a quite important point to note is that even prior to World War II an increasing number of locations had market sizes that could support large-scale production. For example, Rhode (2001) stresses that this was true of California by the 1920s and 1930s where the automobile and tire industries constructed plants at that time. In 1949, shipments of cars from California to Oregon and Washington were 20 times those from Michigan while within California shipments by rail were 10 times those from Michigan to California (Interstate Commerce Commission, 1951).

Nevertheless, Kim (1995) found that spatial concentration was increasing in the early decades of the 20th century. He calculated Hoover’s coefficient of localization for two-digit industries and found that the unweighted and weighted average figures rose from 0.243 in 1880 to 0.327 in 1947 and from 0.242 in 1900 to 0.316 in 1927, respectively, before subsequently declining. However, since Kim wrote his paper, which has become the standard reference on the topic, there have been important developments in the measurement of spatial concentration, which suggest that a new look is required.

Ellison and Glaeser (1997) explained that it is important to control for differences in the size distribution of plants to obtain a meaningful measure of spatial concentration. They showed that an industry in which production comes from very few plants can appear as spatially concentrated even if it is randomly located. To remedy this, they developed an index in which raw geographic concentration is modified by taking account of the plant Herfindahl index. An important refinement to the basic EG index is to take account of the geographical position of regions through allowing for “neighbourhood effects.” This leads to the spatially weighted version of the EG index proposed by Guimarães et al. (2011), which represents a significant advance on Hoover’s localization coefficient.

3. Methodology

One of the pitfalls of spatial indices, including the EG index, is that they do not take into account the relative geographical position of regions, known as the “checkerboard problem.” To illustrate the problem, we follow an example from Guimarães et al. (2011) who used a diagram reproduced here as figure 1. It is intuitively obvious that spatial concentration is greater in figure 1a than in figure 1b—spillovers across regional boundaries would seem much more likely in the former case. The EG index, however, would not be able to distinguish 1a from 1b and would give them the same levels of spatial concentration, which is an example of the “checkerboard problem” mentioned above.

Figure 1

The checkerboard problem panel a: hypothetical distribution of firms and panel b: hypothetical distribution of firms

Open in new tabDownload slide

In fact, the checkerboard problem appears to be important in the early decades of the 20th century notably in the context of movement within the manufacturing belt from the north-east to the mid-west. For example, in the case of SIC 364 (electric lighting and wiring equipment), Map 1 shows that, in 1900, there were two disjointed clusters of a few states with employment in electric lighting concentrated around Illinois and the state of New York, respectively. By 1920, as Map 2 shows, this feature becomes even more pronounced and employment concentrates mostly in New York, Ohio, and Illinois—states without adjacent borders. As Map 3 shows, by 1940, the checkerboard problem is maintained by the rise of California. Since the EG index does not take the geographical position of individual states into account, it misinterprets the concentration of employment into fewer states as a sign of higher geographical concentration even though those states are geographically disjointed, an error that is corrected by spatial weighting.

A solution is to find a measure that takes the relative geographical position of regions into account. Before that, as a useful first step, we formally test for spatial correlation, hence whether the geographical position of individual states matters or not. We use Moran’s I index of spatial autocorrelation, which allows us to diagnose the presence of spatial correlation among US states. It would, for example, correctly diagnose the unsuitability of the EG index for SIC364 as it shows that this industry was highly spatially autocorrelated with the statistically significant values of the statistics of 0.043, 0.039, and 0.058 in 1900, 1920, and 1940, respectively.

Maps 1–3

SIC364: Electric Lighting & Wiring Equipment. Note: All maps plot geographical index |$ G_{i}^{S} $| as defined in Section 3.

Open in new tabDownload slide

Moran’s I is, however, a statistic designed to test for spatial correlation and is not a measure of spatial concentration per se. Guimarães et al. (2011) addressed this challenge and developed the spatially adjusted version of the EG index that takes neighbourhood effects into account.

We implement variants of this approach. Our main results are derived using a first-order contiguity matrix W defined such that each element takes one for contiguous US states and zero otherwise. As a robustness check, we also use an alternative spatial weighting also suggested by Guimarães et al. (2011). In particular, we consider spatial matrices in which neighbours are identified using a pre-defined bandwidth: a spatial unit j is considered a neighbour of a spatial unit i if the distance between their centroids is less than the pre-defined bandwidth b. We discuss this in detail later.

4. Data Sources

We analyse the evolution of the spatial concentration of SIC two- and SIC three-digit level industries across 48 US states in every decade between 1880 and 1997, specifically for the following years: 1880, 1890, 1900, 1910, 1920, 1930, 1940, 1947, 1958, 1967, 1977, 1987, 1997, and 2007. The construction of the indices requires data on employment and on the number of plants by US states at SIC two- and SIC three-digit level industries, and also a spatial weight matrix. The spatial weight matrix for 48 US contiguous states was obtained from the REPEC data repository. Following Guimarães et al. (2011) we used the usswm package developed by Scott Merryman; the original spatial weight matrix was created by Luc Anselin.

The data on US state-industry employment and number of plants were collected from the US Census of Manufactures for the periods 1880–1967, from the Bureau of Labor Statistics for the years 1977–1997, and from the US Economic Census for 2007.

The construction of the EG index over the period of 120 years presents three challenges. First, we need to harmonize SIC two- and SIC three-digit level industries across time. Harmonization of the data for the post World War II period is straightforward as the Census of Manufactures reports the SIC industrial categories and a great deal of information was published about changes in SIC classifications between 1947 and 1997. There are no SIC codes reported in the Censuses before 1947. Here we use the assignment of industries into SIC two- and three-digit categories created by Klein and Crafts (2012) and by Klein and Crafts (2020) for the years 1880, 1890, 1900, 1910, 1920, 1930, and 1940. Details of the harmonization of SIC three-digit industries are in online Appendix 3. Second, construction of the Herfindahl index requires data on employment in plants. Ellison and Glaeser (1997) used data from the 1987 Census of Manufactures, which reports employment in plants belonging to 10 employment size categories. Unfortunately, the Census of Manufactures does not report plant employment data before 1947. Therefore, we use the MS index and the spatially adjusted version of it (SMS), which require only the number of plants, making it feasible to construct the indices all the way back to 1880. Third, when there are issues about disclosure of information on individual companies, the Census either withholds the data or reports the data in employment classes. Similarly, the Bureau of Labor withholds information in order to protect the identity or identifiable information of individual firms. Hence, we have incomplete state-industry employment and plant data. Fortunately, the data are in the form of matrices with rows being totals for US states and columns totals for US industries. This allowed us to take advantage of a methodology developed in Golan et al. (1994). They use a maximum entropy procedure to recover missing data in multi-sectoral matrices with information about row and column sums as well as information contained in the multi-sectoral matrices. In our case, we used across-state and across-industry adding-up constraints to recover missing information on state-industry employment and plant data.

5. Results

The methodology we use to re-examine the long-run patterns in spatial concentration in manufacturing was set out in section 3. The first step is to use Moran’s I spatial autocorrelation index as a diagnostic tool to detect whether the geographical position of individual states matters or not. We calculate the Moran’s I statistic for each SIC three-digit industry in all years under study. Table 4 presents a summary of the results by year and SIC two-digit category. We see in Panel A that in a large percentage of SIC 3 industries Moran’s I is statistically significant, starting at over 80 percent in 1880 and 1890. The percentage declines over time but even in 2007 about 40 percent of industries still exhibit significant spatial autocorrelation. In Panel B, we report that all SIC 2 sectors exhibit a large share of SIC 3 subcategories with significant spatial autocorrelation except for printing and publishing. This confirms that spatial correlation across individual states mattered for the entire period 1880–2007 and is not confined to a few industries. Therefore, we use the SMS index, which addresses this substantial spatial autocorrelation. As for the sign of statistically significant Moran’s I, almost all had a positive sign with only 11 year-industry pairs showing negative signs. Figures A1–A4 in online Appendix 4 depict histograms and kernel distributions of statistically significant Moran’s I for the periods 1880–2007, 1880–1920, 1920–1958, and 1958–2007, respectively. They confirm that almost all were positive with fewer very high values in the later periods.

We first report the results of the weighted average SMS index for all SIC3 industries over the long run where the weights are the shares of employment in SIC3 industry, robustness checks with respect to the spatial matrix, and a comparison with the original, spatially EG unweighted index respectively. The weighted average SMS Index is reported in table 5, column I, and we plot it in figure 2 as well. The highlight of this longer-term account is that the levels of spatial concentration were considerably higher (almost twice as large as in 2007) in the early decades of the 20th century through to 1940 and then fell quite rapidly after World War II. Furthermore, mean spatial concentration for SIC3 industries was distinctly lower in 1930 and 1940 than in 1880. Although the rate of decrease of mean SMS accelerated after 1940, about a third of the total fall between 1880 and 2007 had already occurred by 1940. Overall, our estimates show that spatial concentration of industries was much more prevalent in the late 19th than in the late 20th century. We also explored alternative methods of spatial weighting as a robustness test, see online Appendix 1.

Figure 2

SMS and MS index, 1880–2007

Open in new tabDownload slide

It is interesting to compare these results with the (spatially unweighted) MS index, other EG-type indices in the literature, and Kim (1995). The MS index is presented in table 5, column II, and in figure 2: this shows a similar proportionate decline in geographical dispersion between 1910 and 2007. Unlike the SMS index, however, the MS index shows an increase in spatial concentration in the early 20th century. A comparison with the EG averages reported by Dumais et al. (2002) for the years 1972 to 1992 reveals that our estimates are somewhat larger but show a similar decrease in this period. Contrary to Kim (1995), who reported the weighted average of Hoover’s coefficient of localization for SIC2 industries that is presented in figure 3, we do not find an episode of increasing spatial concentration in the 1910s and 1920s when looking at the SMS index.

Figure 3

MS index 1880–2007, Kim (1995) index 1880–1987

Open in new tabDownload slide

To examine this in more detail, we look at which industries drove the increase in MS index. We calculate the ratio of the MS and SMS indices to their 1900 values respectively for each decade between 1910 and 1940. The results are summarized in table 6 and discussed further in online Appendix 4. Table 6 shows the percentage of SIC 3 industries for which the MS ratio is greater than the SMS ratio for each decade between 1910 and 1940. A clear pattern emerges: a majority of SIC 3 industries have an MS ratio larger than the SMS ratio. Even the SIC 2 industries with the lowest percentage, such as SIC 29 petroleum and coal products, or SIC 32 stone, clay, and glass products, have more than 50 percent of their SIC 3 subcategories in which this is the case. This, similarly to Moran’s I index, confirms that the checkerboard problem affects the entire spectrum of the manufacturing sector, leading to bias in the MS index as well as Kim’s (1995) index. Accordingly, the spatial concentration of manufacturing sector in the first four decades of the 20th century was not an inverted U-shape as suggested by Kim (1995); actually, it was declining slowly.

SMS estimates for all SIC2 industries are reported in table 7. A general tendency to greatly increased spatial dispersion over time is clear; in every case except one, namely, SIC 21 tobacco and tobacco products, the SMS index was lower in 2007 than in either 1880 or 1940 and in all but one sector the reduction was at least 40 percent. The highest SMS score in 1997 (0.17) would have been the second lowest in 1880. In the vast majority of sectors (17/20), there was already dispersion between 1880 and 1940. The largest reductions in the SMS index between 1880 and 2007 are in SIC 30, rubber and plastic products, SIC 35, machinery, SIC 36, electrical equipment, and SIC 37, transportation equipment.

The experience of changing spatial concentration at SIC3 level is summarized in table 8. In 12/20 SIC2 categories at least 67 percent of the constituent SIC3 industries were more dispersed in 2007 than in 1880 and in 12/20 SIC2 categories the same was true for 1940 compared with 1880. So, there was quite a high incidence of spatial dispersion but it was by no means universal (our results do not lend support to the hypothesis of stability in geographic concentration advanced by Dumais et al. (2002) and discussed in online Appendix 2).

The evolution of spatial concentration in three groups of industries, those whose origins were in the first industrial revolution, those from the second industrial revolution, and those from the ICT revolution, is displayed in figure 4. The list of industries belonging to the first, second, and ICT revolutions, respectively, is in online Appendix, tables A3–A5 and is based on Mowery and Rosenberg (1999). The category of first industrial revolution contains traditional industries that began before 1870, the category of second industrial revolution includes industries includes industries that emerged in the period 1870–1914 and are based on the technologies of that era. ICT revolution industries are the ones emerging in the second half of the 20th century and would be generally regarded as part of the late 20th century general purpose technology.

Figure 4

Spatially weighted SMS index by industries, 1880–2007

Open in new tabDownload slide

In each case, spatial concentration starts out quite high and then decreases much as Hoover (1948) suggested. Interestingly, the second industrial revolution industries are dispersing continuously from 1910 onwards and the ICT industries are the least spatially concentrated of the three groups in the late 20th century. This is because ICT industries were developing in three geographically disjoined states of Texas, California, and later Washington, which is controlled for in our spatially weighted version of Ellison and Glaeser index.

Although we have stressed that there was a strong tendency for spatial concentration of industries to decline over time, especially after 1940, it is important to recognize that even at the end of our period there was a very high incidence of localization at the SIC3 level. Spatial concentration was almost always present to an extent which was both statistically and economically significant. We have tested the statistical significance using equation 4 under the null hypothesis that SMS index is equal to zero. We can reject the null hypothesis at the 1 percent significance level in all but 20 instances (none after 1940). Table 9 lists all the cases where the SMS index is not statistically significantly above zero.

Furthermore, figure 5 displays kernel distributions for SMS for selected years with the charts on the right truncated at zero for 1880 and 1940. It is apparent that, with spatial weighting, there are very few observations below 0.05, the conventional level described as “highly concentrated” and, as we saw in table 5, the mean SMS at SIC3 level is way above 0.05 throughout the period. The criterion of 0.05 was originally chosen by Ellison and Glaeser (1997) because it is consistent with the existence of substantial local cost advantages. Therefore, our results imply that economically significant spatial concentration was the norm across industry continuously from 1880 through 2007.

Figure 5

Kernel density of SMS index

Open in new tabDownload slide

6. Discussion

A notable implication of our results is that forces promoting the spatial dispersion of American manufacturing were present throughout the 20th century. The most important of these was surely the continuing long-run decline of transport costs first in the railroad era and then sustained by trucking. Lower shipping costs for goods meant that manufacturing could move out of the large industrial cities in which it concentrated at the start of the 20th century (Glaeser and Kohlhase 2004). Market potential would matter less and high wage costs in production would matter more, and this eroded the advantages of the manufacturing belt. The ratio of the average wage in states in the manufacturing belt compared with other nearby states followed an inverted-U shape with its peak in 1940. Over the long run, industrial location continually evolved as fundamentals changed.

Glaeser and Kohlhase (2004) noted that the costs of moving manufactured goods declined by over 90 percent in real terms between 1890 and 2000 from 18.5 cents per ton-mile to 2.3 cents (at 2001 prices). In fact, much of this decrease occurred by 1967 when the cost was only 5.6 cents (at 2001 prices) and by 1891 the railroad revolution had cut transport costs to about 10 percent of the 1820s’ level (these estimates of transport costs are based on Carter et al. (2006), volume 4, pages 781 and 932–934). We calculate that the ratio of the average wage in manufacturing in East North Central and Mid-Atlantic states relative to East and West South Central states rose from 1.22 in 1890 to 1.52 in 1940 before falling to 1.15 in 1987 (the average wage rates are obtained by dividing the wage bill by the number of workers in the Census of Manufactures).

Average plant size according to our estimates from the Census of Manufactures rose from 11.0 in 1880 to 60.6 in 1947, after which it stayed on a plateau until 1977 when it was 62.7 before falling to 41.8 in 2007. As many writers including Kim (1995) have noted, the decrease in plant size in the later 20th century was conducive to lower spatial concentration. In the period of rising plant size combined with spatial dispersion prior to World War II, the point to note is that the rise of the mid-west relative to the north-east, which tended to lower SMS scores was associated with establishment of larger plants. By 1940, 14 SIC two-digit industries of 20 had a larger average plant size in the mid-west than in New York whereas in 1880 that was true of only 3 of the 20.

So, in the long run, the locational advantages of agglomeration in the manufacturing belt were undermined by rising wage costs, falling transport costs and a reduction in average plant size. In some respects, this combination of changes over time is reminiscent of the later phase of the stylized core-periphery model presented by Krugman and Venables (1995). This model would see a move from very high to intermediate to very low transport costs driving a move from dispersed to spatially concentrated then back to dispersed locations for manufacturing. In the spatially concentrated (manufacturing belt) phase the core benefits from economies of scale and proximity to markets and suppliers raise productivity but also tend to raise wages; subsequently, however, in the context of much lower transport costs, the wage gap becomes too high and moves to the periphery promote a convergence of wage rates.

Recent research has produced empirical results which are broadly consistent with a core-periphery model. Klein and Crafts (2012) found that the location of manufacturing in the early 20th century was strongly influenced by the attraction of market potential to industries with large plants and strong linkages with industrial customers and suppliers. This pattern underpinned the existence of the manufacturing belt. Crafts and Klein (2015) found that home bias in US domestic trade was much lower in 1949 than in 2007. In 1949, some commodities actually exhibited negative home bias at a time when the ratio of inter- to intra-state trade was much higher and much production in the manufacturing belt was still exported to the rest of the United States. They showed that in 1949 home bias was inversely correlated with geographic concentration of industries. This configuration had, however, evaporated by 2007.

Reality was often more complex but reflected similar issues. An excellent example of this is Motor Vehicles and Equipment (SIC 371) where overall geographic concentration fell in the second half of the 20th century but where significant localization persisted in a new configuration. The SMS index for SIC 371 was 0.191 in 1940, 0.120 in 1958, 0.106 in 1977, and 0.094 in 1997. Maps 4 to 7 show an evolving pattern of its spatial concentration over time such that by 1997 the move away from the 1940 situation of a dominant position for Michigan and an east–west corridor in the southern Great Lakes region has been superseded by one in which Michigan is still a major centre but clusters within “Auto Alley” extend as far south as Alabama (Klier and Rubenstein, 2008). Two key developments that underlay these changes were the switch of assembly plants in the 1960s away from the coasts to central areas to reduce the costs of transporting cars to customers once these plants became specialized in models for sale throughout the United States and the advent of Japanese producers in the 1980s and 1990s who chose to locate further south—initially Kentucky and Tennessee and then in the deep south. Throughout, parts suppliers wanted to locate close to auto producers. Transport costs were instrumental in some of these decisions, but the move to the south by the Japanese was encouraged by a quest for lower labour costs.

It is interesting to view changes in the location of manufacturing together with the evolution of spatial concentration in knowledge-intensive business services (knowledge-intensive business services comprise finance, insurance and real estate, business services, and professional services). These activities that are typically supplying intermediates, often to other business services providers, appear to benefit strongly from economies of agglomeration, which stem from thick markets in human capital, advantages of proximity to users and suppliers and knowledge spillovers. Since 1980, their geographical concentration has been increasing and these activities are now strongly localized in densely populated metropolitan areas such that they have become more agglomerated than manufacturing whereas the opposite was very much the case in 1930. It appears that the attraction of market potential in the context of linkage effects has started to matter a lot for knowledge-intensive business services while at the same time it has lost its attraction for manufacturing (Cermeno, 2019).

Maps 4–7

SIC 371 – Motor Vehicles & Motor Vehicle Equipment. Note: All maps plot geographical index |$ G_{i}^{S} $| as defined in Section 3.

Open in new tabDownload slide

There is a marked contrast between employment patterns in large metropolitan areas at the beginning and end of the 20th century. In 1910, 35.1 percent of employment in the largest MSAs was in manufacturing and 6.2 percent was in business services compared with 25.1 percent and 4.4 percent, respectively, in non-MSA locations. In 1995, 14.3 percent of employment in the largest MSAs was in manufacturing compared with 21.3 percent in business services compared with 26.9 percent and 9.1 percent, respectively, in non-MSA locations (Kolko, 1999; Kolko notes that in 1995 business service occupations accounted for 41.8 percent of employment in the largest MSAs.) By 1995 in an era of much lower transport costs for goods, relatively land-intensive manufacturing had relocated to areas where real estate was cheaper and, at least in the cities that had successfully regenerated, been replaced by human-capital intensive business services (Desmet and Fafchamps, 2005). This is the story for Boston though not for Detroit (Glaeser, 2005). A subset, but only a subset, of traditional manufacturing cities was able to make the transition to becoming a successful services-based agglomeration.

Accounting for the checkerboard problem corrects the long-run pattern of spatial concentration: the inverted U-shape pattern, driven by an increase of spatial concentration in the first half of the 20th century, does not hold anymore and instead we find stability followed by a slow decline before WWII. This is consistent with the core-periphery pattern analysed in Klein and Crafts (2012). They show that the manufacturing belt was the main location of industrial sector as early as the beginning of the 20th century. Industries continued to locate there, but that does not necessarily imply an increase in spatial concentration if the checkerboard problem was present, as indeed it was. Maps 1–3 illustrate this clearly. We see that between 1900 and 1920, even though the employment in this industry increased in the manufacturing belt, it increased in the states are geographically disjointed: employment concentrates mostly in New York, Ohio, and Illinois—states that are not adjacent. As for 1920–1940, while the increase in employment created a contiguous area of Middle Atlantic and Midwest regions, we see a substantial increase of employment in California. This reinstates the checkerboard problem because these two dominant regions of employment are geographically disjointed: one in the East and Midwest, the other in the West.

Besides contributing to the checkerboard problem, the ascent of California as a manufacturing location adds to the richness of the historical picture. Initially, Californian manufacturing was based mainly on resource-processing industries but already by the late-1930s it was developing a significant presence in knowledge-based industries and a comparative advantage based on human capital and localized technological spillovers, first in aircraft followed by electronics and information technology (Rhode, 2001). A good post-war example can be found in the semi-conductor industry where spatial dispersion took place over the long run in the context of a reconfiguration of the sector driven by technological change. The key development was the advent of the integrated circuit in 1959, which was discovered in California and Texas. This triggered a long-term move to those states and away from Massachusetts and New York where, in the 1950s, semiconductors were produced by vacuum tube manufacturers. Nevertheless, the industry continued to experience a significant level of localization in which knowledge spillovers and proximity to buyers played a big part (Ketelhӧhn, 2006).

In the context of a general move towards greater spatial dispersion, it is noteworthy how weak correlations of localization at the industry level were over time, as discussed in online Appendix 2. Even so, it is striking how pervasive significant excess spatial concentration has been throughout our period. As the manufacturing belt lost its manufacturing dominance the new locational patterns saw new pockets of spatial concentration emerge rather than a scattering of plants across the rest of the country. Nevertheless, it seems quite possible that the underlying reasons for concentration have changed over time and that individual-industry experiences provide many variations on this theme. These are important topics for future research (for example, as one of the founding fathers of the “new economic geography” reflected, its models may have more salience to the era of the manufacturing belt than the present day (Krugman, 2011)).

7. Conclusions

We have constructed spatially weighted indices of geographic concentration of SIC2 and SIC3 manufacturing industries in the United States over the period 1880 to 2007 and have shown that this is possible notwithstanding data constraints. These estimates embody recent methodological innovations. We offer a new and improved perspective on long-run trends in spatial concentration of American manufacturing. We show that it is very important to use spatial weighting in order to achieve this. This leads us to a very different picture of long-run trends in spatial concentration from that which was found by Kim (1995); we do not find an inverted-U shape.

The first striking feature of our estimates is that by the end of the 20th century average levels of spatial concentration in manufacturing were much lower than in the late 19th century. The weighted average for SIC3 industries for the SMS index was 0.098 in 2007 compared with 0.223 in 1880. Although spatial concentration fell more rapidly after World War II, a significant decrease had already taken place by 1940 in the context of an early decline in the importance of the manufacturing belt and a switch towards the mid-west within the manufacturing belt. A second important point is that this experience is characteristic of the vast majority of SIC2 industries. It is also notable that correlations over time of our index of geographic concentration are quite low, as seen in online Appendix 2. The third major finding that comes from our estimates is that ‘excess’ spatial concentration is pervasive at the SIC3 level throughout the whole period. Across almost all industries and all years, spatial concentration is significant both statistically and economically.

Acknowledgements

We are grateful to the editor, Joan Roses, and three anonymous referees for helpful comments and Sylvain Barde for his help with the estimation procedure to recover missing observations.

References