تمرکز استخدام در سراسر شهرهای ایالات متحده

This paper examines the spatial distribution of jobs across U.S. counties between 1970 and 2000, and investigates whether sectoral employment is becoming more or less concentrated. The existing literature has found deconcentration (convergence) of employment across urban areas. Cities only cover a small part of the U.S. though. Using county data, our results indicate that deconcentration is limited to the upper tail of the distribution. The overall picture is one of increasing concentration (divergence). While this seemingly contradicts the well documented deconcentration in manufacturing, we show that these aggregate employment dynamics are driven by services. Non-service sectors – such as manufacturing and farming – are indeed becoming more equally spread across space, but services are becoming increasingly concentrated.

Economic activity is unevenly distributed across space. The interaction of positive and negative externalities creates intricate geographical patterns of city clusters and rural hinterland (Henderson, 1988 and Fujita et al., 1999). Over time, these patterns evolve because of changes in preferences, production technologies and transport costs. As a result, the spatial distribution of employment adjusts as jobs are created in certain locations, and destroyed elsewhere. Understanding how economic activity is likely to be distributed through space in the future is important for policy makers at the national and local level. This paper describes the geographical evolution of jobs in the U.S. between 1970 and 2000, with the goal of understanding what the future spatial distribution of employment would look like if current tendencies were to continue. We use county-level employment in 13 different sectors – ranging from farming to manufacturing and services – and focus on the ergodic distribution of jobs. Our work differs from the existing literature in a number of respects. First, rather than looking at income per capita or population, we are interested in employment. Many authors have studied whether standards of living in the U.S. are becoming more similar over time. For instance, Higgins et al. (2003) find a strong evidence of income convergence across counties. This is not entirely surprising, given the high degree of labor mobility in the U.S. (Blanchard and Katz, 1991). However, income convergence does not tell us anything about where economic activity is locating. Is the U.S. moving towards a situation with more or with less large- and medium-sized metropolitan counties? Are rural counties losing or gaining jobs? These are the kinds of questions we address in our paper.1 This is similar to studying whether population is becoming more or less concentrated in space. In this respect, Beeson and DeJong (2002) are of particular interest. They find population divergence across counties, especially in the post-WWII period. Our work is complementary to theirs. By looking at employment, rather than population, we get additional insights from sectoral disaggregation. Second, we examine the country as a whole, not just metropolitan areas. Most of the literature on the spatial organization of economic activity in the U.S. has focused on cities. One central finding of that line of research is that city growth is independent of city size, a phenomenon known as Gibrat's Law (Sutton, 1997). However, as pointed out by Beeson et al. (2001), limiting the analysis to urban areas introduces a selection bias, since cities are those areas which experienced high growth in the past. A recent paper by Eeckhout (2004) addresses this issue by revisiting Gibrat's Law using Census ‘places’. In contrast to metropolitan areas, these data cover the entire size distribution, including small towns and villages. He confirms that growth is independent of size. However, ‘places’ still do not cover the entire U.S. In the 2000 Census they accounted for 74% of the population. Our third point of departure with the existing literature is our methodology. Instead of relying on a single method – whether β-convergence, σ-convergence, or ergodic distributions – we develop a methodology that encompasses them all. Much of the existing work comparing geographical units is couched in terms of Barro's β-convergence: the underlying model is deterministic in nature ( Barro, 1991 and Mankiw et al., 1992). As first emphasized by Quah, evidence of β-convergence can yield a misleading picture, because it can arise even when countries or regions are getting further apart, and vice versa ( Quah, 1993, Quah, 1996a and Durlauf and Quah, 1999). As a solution, Sala-i-Martin (1996) suggests studying distributions by looking at the evolution of the variance over time, a concept known as σ-convergence. Quah, 1996b and Quah, 1997 goes one step further by focusing on the ergodic distribution. This refers to the long-term spatial distribution of economic activity that would arise if current transition probabilities would remain constant. The ergodic distribution is the distributional equivalent of the β coefficient in a standard Barro model: it predicts in which direction the process goes, should current structural factors remain unchanged. Of course, structural parameters may change, in which case the direction of the process would change as well. The ergodic distribution is thus but a way of describing the current trend of the distribution. In this paper we start by computing parametric and non-parametric versions of the unconditional and conditional β-convergence tests and explain how they can be understood as describing the expectation of the transition probability. We then compute two versions of the ergodic distribution. The first version is the stochastic equivalent of unconditional β-convergence: it assumes that all counties are inherently equivalent and could switch places with each other over time. The second version is the stochastic equivalent of conditional β-convergence: it conditions on county characteristics that are constant over time. It is our best estimate of how economic activity would be distributed across U.S. counties should current tendencies remain unchanged. In addition, we also introduce a number of practical innovations when deriving the ergodic distribution. In particular, by computing the transition matrix from the smoothed conditional distribution rather than directly from the data, we get a better approximation of the ergodic distribution. This makes the results both more detailed and more accurate.2 The methodology is easy to implement, and can be applied to any empirical study involving distribution dynamics. We now turn to describing our main findings. Whereas recent work on metropolitan areas shows a tendency towards deconcentration, with total employment becoming more equally spread across cities ( Chatterjee and Carlino, 2001 and Carlino and Chatterjee, 2002), standard β-convergence tests using U.S. county data suggest the contrary, with jobs becoming more concentrated over recent decades (Desmet and Fafchamps, 2005). The analysis presented here resolves this apparent puzzle. Results show that, compared to the current distribution of total employment across counties, the ergodic distribution is a lot flatter, with the middle group thinning out. The overall picture that emerges is thus one of concentration (divergence), with lots of small- and medium-sized counties losing jobs to the more urban ones. At the upper tail of the distribution, however, the opposite is true, with large metro counties losing jobs in favor of intermediate-sized urban counties. In other words, there is deconcentration (convergence) in the upper part of the distribution, and concentration (divergence) in the distribution at large. This explains the opposing results of Chatterjee and Carlino (2001) and Desmet and Fafchamps (2005). Our findings confirm the results of Beeson and DeJong (2002) for population growth: for the post-WWII period they report divergence across most of the distribution, but convergence in the upper decile. The increased concentration evident in total employment stands in contrast to what happened within the manufacturing sector. It is by now a stylized fact that since World War II manufacturing employment has become less concentrated, albeit at a slow pace (Dumais et al., 2002 and Kim, 1995). Our data confirm this empirical regularity. Although manufacturing cannot account for the spatial dynamics of aggregate employment, services can. The main service sectors – ‘retail’, ‘finance, insurance and real estate’ and ‘other services’ – exhibit concentration (divergence) in the middle part of the distribution and deconcentration (convergence) in the upper tail. This is most patent in the case of ‘other services’, where we get ‘twin peaks’ — a bimodal ergodic distribution. That overall trends in the economy are driven by services should not come as a surprise, given their weight. However, the fact that services behave differently from the rest of the economy is interesting, because empirical work in economic geography has mainly focused on manufacturing. Our findings confirm that the much heralded demise of cities, epitomized by manufacturing jobs moving to less dense areas, is not occurring. The reason is the rise of the service industry (Kolko, 1999). Though not the subject of this paper, our results have implications for the spatial dynamics of productivity and wages. Following Ciccone and Hall (1996), our findings suggest that sectors that have been deconcentrating, such as manufacturing, may have experienced spatial convergence in wages and productivity. In contrast, we would expect aggregate employment and services, which have been concentrating, to have seen increasing spatial divergence in wages and productivity.

In this paper we examined how the distribution of employment across U.S. counties is likely to evolve if current concentration and deconcentration forces remain unchanged. To do so, we developed a methodology borrowing from the work of Quah and building upon the literature on β- and σ-convergence. We computed non-parametric β-convergence regressions, conditional and unconditional. Using non-parametric methods, we also computed detailed ergodic distributions for total employment and sectoral employment across U.S. counties. Our results suggest that employment is becoming increasingly concentrated across counties. Although very large metro counties may be losing jobs, the proportion of counties with modal employment is decreasing in favor of medium to high employment counties. More specifically, the 8% largest counties exhibit deconcentration; the remaining 82% exhibit concentration. This result is consistent with deconcentration across urban areas (Chatterjee and Carlino, 2001) and concentration across U.S. counties (Desmet and Fafchamps, 2005). It also confirms the results of Beeson and DeJong (2002) of population divergence across counties. Whether the overall picture is one of concentration or deconcentration is not entirely obvious. In terms of the number of counties, concentration holds the upper hand. However, in terms of the number of people, deconcentration dominates, since the 8% highest employment counties accounted for nearly two thirds of total employment in 1970. There are important differences across sectors. As in the rest of the literature, we find deconcentration in manufacturing. Deconcentration is also the norm in other non-service sectors. However, service activities are becoming more concentrated, in particular ‘retail’ trade, ‘finance, insurance and real estate’, and ‘other services’. Given the importance of these sectors, they drive the evolution of the spatial distribution of total employment. Limiting the focus of analysis to manufacturing is misleading. The U.S. is a service economy, and services are behaving very differently from the other sectors. Although we have limited our analysis to employment, our findings may shed light on the spatial dynamics of productivity and wages across the United States. Using county-level data, Ciccone and Hall (1996) conclude that doubling employment density leads to a 6% increase in productivity. Similar numbers have been found in subsequent studies by Harris and Ioannides (2000) for U.S. metropolitan areas and by Ciccone (2002) for European regions. Here we have looked at employment levels rather than at employment density. But our qualitative results remain basically unchanged if density is taken as the dependent variable. Following the insights of Ciccone and Hall (1996), we would expect 23 sectors that have been deconcentrating – such as manufacturing – to have experienced a fall in spatial productivity (and wage) differences. The opposite should have occurred for aggregate employment and services. These predictions about productivity and wages are speculative, and warrant further investigation. They are based on a world in which employment dynamics are driven by changing agglomeration and congestion effects on the production side. An example of such approach can be found in Chatterjee and Carlino (2001) who argue that rising aggregate employment causes congestion costs to rise faster in more dense areas, leading to deconcentration of jobs. However, other forces – such as congestion on the consumption side or a change in people's preferences – may also be at work. In that case, the picture may be more complex. For instance, soaring house prices in urban areas could be consistent with densely populated areas losing employment but experiencing rising wages. Similarly, if people have an increasing preference to live in low density areas (Beale, 1977), this may lead to deconcentration of employment but increasing wage differentials. On the methodological side, our research shows the importance of using non-parametric methods and of looking at the entire distribution, not just at cities. It also demonstrates that β-convergence tests, even when done non-parametrically, are not sufficiently informative. Computing the ergodic distribution associated with a given set of transition probabilities is more useful to understand spatial trends. Moreover, our approach is able to condition on time-invariant characteristics in a way that is fully consistent with standard analysis of conditional β-convergence. The methodology developed here can easily be applied to the study of any distributional dynamics. This paper leaves a number of other questions unanswered. First, it is unclear whether the forces identified here operate in a similar manner in other time periods and other parts of the world. Applying the same approach to other data sets is necessary before we can conclude that the process described here generalizes beyond the confines of this study. Second, the methodology presented here does not (yet) allow statistical inference in the normal sense. Statistical tests are reported for some of the statistics presented here, such as confidence intervals for kernel regressions. But we do not present a ‘test’ of (conditional or unconditional) convergence based on estimated ergodic distributions. In principle, such a test could be developed provided an intuitively satisfying counter-factual distribution could be devised. It should also be possible to use bootstrapping to test whether the mode of the ergodic distribution has shifted to the left or the right relative to the current distribution (Kremer et al., 2000). Developing such tests is left for future research.