Road to Despair and the Geography of the America Left Behind Mark Partridge1,2,3 and Alexandra Tsvetkova1* 1 The Ohio State University 2 Jinan, University, Guangzhou China 3 Urban Studies and Regional Science, Gran Sasso Science Institute, L’Aquila, Italy December 15, 2017 Abstract: President Trump’s election highlights US economic disparities, especially in rural America. This study assesses 21st century economic conditions to identify broad forces underlying the uneven economic performance of US counties, stressing factors that may be important for lagging regions. We examine the effects of three groups of variables (economic, social/demographic, and geography) on job growth, poverty, and median income. To this end, we split the time period before and after the Great Recession and use standard regression analysis augmented by quantile regressions to assess the heterogeneity in economic performance. The results suggest an increasing role played by economic factors including the benefits of having a fast-growing industry structure. Perhaps more importantly, measures of economic dynamics—the ability of a local economy to “rewire” by reallocating resources in response to economic shocks—emerge as important predictors of performance. Acknowledgements: The authors appreciate the partial support of USDA AFRI grant #11400612 “Maximizing the Gains of Old and New Energy Development for America`s Rural Communities” and US Economic Development Administration grant #ED17HDQ3120040 “Regional Economic Development as a New Theoretical Framework: Integrating Data Science, Complex Adaptive Systems and Social Sciences for Precision Policies” * Corresponding author 2120 Fyffe Rd, Columbus, OH 43210; [email protected] 2 The United States has always experienced spatial differentials in economic activity and wellbeing. Yet, structural changes such as deindustrialization, technological change, and globalization has led to a perceived widening of spatial income differentials such as declines in Rustbelt, coal country, and much of rural America. The uneven recovery from the Great Recession has further fueled perceptions that large regions are being left behind. Meanwhile, the US has come under grips of a wrenching opioid crisis and related “deaths of despair” that are often associated with the lack of economic opportunity (Betz and Jones 2018; Case and Deaton 2015; Goetz and Davlasheridze 2018). In conjunction, rising income inequality has led many Americans to question whether they will be as well off as their parents (Chetty et al. 2014). The growing angst in stagnating regions is often credited as a key reason for President Trump’s surprising 2016 victory (Goetz et al. Forthcoming). As the country is deeply divided economically and culturally, it is important to identify the general processes that underlie the recent trajectories of US regional development. Understanding the underlying forces would help in shaping policy responses to bring the localities left behind back to the table of economic opportunity and growth. Indeed, this is consistent with calls to depart from development strategies that were ineffective pre-recession and to identify new approaches to be used after the Great Recession (Fodor 2012). To this end, we examine the effects of three main socioeconomic groupings of factors that reflect different aspects of a locality’s economic structure, social/demographic attributes, and natural amenities, as well as position within the urban- rural hierarchy. The selection of the three general variable groupings follows from the economic development literature (Beyers 2013; Partridge 2010; Rupasingha, Goetz and Freshwater 2002). To address whether the period after the Great Recession represents a structural shift in regional dynamics, we further divide the data into pre-recession (2000- 2007), recession (2007-2010) and post-recession (2010-2015) periods. We then 2 3 investigate the changing importance of each factor in explaining employment growth, change in poverty rates, and median household income growth for US counties. In developing our models, we supplement more traditional (mainly) “static” economic measures with novel measures that approximate the dynamism of a local economy and the ability of a county to rewire by reallocating employees from shrinking to expanding sectors. We use cross-sectional regression analysis augmented by first- difference analysis to understand contemporary determinants of local economic well- being and whether the Great Recession altered this relationship. Finally, we use quantile regressions to assess the presence of heterogeneity in the economic relationships between the most and least prosperous locales. Our results suggest some changing structural relationships between the three explanatory groupings and economic outcomes. We find that the role played by county industrial composition (if it is fast- or slow-growing) is of increasing importance. Also, another increasingly important factor is the local labor market’s ability to “rewire” by facilitating the movement of workers across industries and occupations in response to changing economic conditions. Interestingly, after the dynamism of a local economy is accounted for, industrial diversity is insignificant, suggesting that diversity’s role in stabilizing and promoting growth in local communities (Hammond and Thompson 2004; Watson and Deller 2017) may be working more through labor-market flexibility. Some of our most policy-relevant finding comes from the quantile analysis of differenced job growth. For counties that are lower at the distribution of the response function, the labor- market measures of flexibility emerge as important predictors of growth, suggesting that removing barriers to flow of resources within lagging economies might be a viable policy option. In what follows we start with a brief descriptive analysis to ascertain that economic well-being is diverging geographically. Concluding that there are good reasons to believe so, we follow with a short literature review. We next describe the data and 3 4 empirical specifications followed by the empirical results. We separately discuss the results for poverty (and median household income growth to a lesser degree) and job growth. The paper finishes with our concluding thoughts and policy suggestions. Is Basic Economic Well-Being Diverging? Since the Great Recession, there is a growing sense that some places are being left behind. Yet, this flies against conventional economic wisdom from neoclassical growth theory that regional incomes have been converging since the Civil War (Barro and Sala-i- Martin 1990). To investigate, using US Bureau of Economic Analysis (BEA) data, we calculate the average standard deviation in per-capita income for each year between 1969-2016 (standardized by national per-capita income) for US states and counties. Specifically, we calculate unweighted standard deviations that reflect differences across space and standard deviations weighted by population to show spatial differences for the average person (which is national income inequality minus the within-state/county component of inequality). The results are plotted in figure 1 where the left panel shows unweighted standard deviations for states and counties, and the right panel shows corresponding standard deviations weighted by population. Figure 1 shows that analysis at the state level masks considerable within-state inequality. Turning to the county-level results, the unweighted standard deviations show a slight downward trend until 1994, falling to about 0.17 before rising almost 50% to 0.25 in 2014, then falling back to 0.23 in 2016. The population-weighted standard deviations illustrate an even stronger upward trend. After falling slightly to about 0.20 in 1976, the weighted standard deviation steadily increases to about 0.32 in 2016, or a rise of about 60%. The analysis was repeated by removing transfer payments and the divergence pattern for the resulting “market per-capita” personal income is even more striking, further suggesting that economic opportunities are increasingly geographically unequal (not shown). We also did the same using the unweighted and weighted standard deviations of annual wage and salary job growth. There, the trend is steady convergence 4 5 of job growth rates until 2010. After which, there has been about a one-third increase in the unweighted variation between 2010-2016, though the weighted standard deviation had a more modest increase (not shown). Overall, in terms of income, there has been a steady increase in divergence for nearly 25 years. The divergence in job growth is much more modest, though the post-recession period represents a departure from the pre- recession trend of convergence. The implication is that there are reasons to believe that some regions are increasingly lagging.i Figure 1. Average standard deviations in per-capita income (unweighted on the left, weighted on the right) Literature review The literature investigating the determinants of regional economic growth is enormous; it highlights many factors that are important for local socioeconomic wellbeing across space and time. Such factors can be generally grouped into several broad categoriesii related to the presence of certain industries and related structural metrics such as (1) industry diversity (Watson and Deller 2017), (2) human capital and innovation (Faggian and McCann 2008; Fallah, Partridge and Rickman 2014; Goetz and Hu 1996), (3) population demographics (Stephans and Deskin 2018; Amcoff and Westholm 2007), (4) culture, social capital and related factors (Akçomak and Ter Weel 2009; Rupasingha, Goetz and Freshwater 2002; Rupasingha, Goetz and Freshwater 2000), and (5) amenities (Deller, Lledo and Marcouiller 2008; Deller et al. 2001) among others. The performance 5 6 of rural and remote regions has been further defined by remoteness and access to agglomeration (Andersson and Lööf 2011; Partridge et al. 2007; Partridge et al. 2009). The Great Recession threw the US economy from its long-term growth trend and further intensified scholarly debates on the determinants of regional economic growth. The central topic has increasingly moved to the notion of resilience—i.e., the ability of regions to withstand and recover from shocks. Aside from a concerted effort to operationalize and measure resilience, the discussion focuses on the same broad categories described above (Martin, Sunley and Tyler 2015). The economic resilience literature suggests that the Great Recession revealed many underlying discrepancies in regional economic fundamentals, speeding up the process of divergence in economic fortunes that could be undetectable during prosperous times (Lagravinese 2015). Some researchers note that the Great Recession weakened the regions that lacked strong engines of growth (Martin, Sunley and Tyler 2015) and exacerbated the long-simmering economic and social problems in rural and lagging communities. Others believe that the Great Recession was a watershed for the US economy (Florida 2009; Gore 2010), implying that the nation will need the new ways of resource allocation to respond to a rapidly changing world. When one thinks about economic resilience as an adjustment process to a shock, the economic variables currently used in the literature may be insufficient, as they focus on a structure of a local economy (e.g. Lagravinese 2015) and generally ignore the dynamics of how a local economy readjusts and rewires. Thus, a key goal of our study is to develop new dynamic measures of local economic adjustment and to assess their effects on economic outcomes. The literature also points to an important role played by various social/demographic factors in defining regional performance. For instance, the importance of human capital in affecting economic growth is well established (Lucas, 1988; Nelson & Phelps, 1966). Other research points to the local racial and ethnic 6 7 composition as important for social and economic wellbeing. For example, Easterly (2001) and Partridge and Rickman (2005) find that high-poverty places in the US tend have greater minority populations. Putnam, Leonardi and Nanetti (1994) stress the role of social capital in regional socioeconomic outcomes. The level of social capital in a community is generally related to participation in associational activities and trust. Several empirical studies find a positive effect of social capital on a range of economic growth indicators in the US (Rupasingha, Goetz and Freshwater, 2000; 2002). Finally, amenity-led economic development has received significant scholarly attention (Green et al. 2005). Many high-amenity places have been able to capitalize and attract in-migration, even to rural areas (Partridge 2010), although it is unclear how the Great Recession and housing bust affected the long-run prospects of high-amenity locales. Empirical implementation, data and variables We start our analysis with a descriptive look at changes in poverty and job growth pre- and post-recession followed by cross-sectional regressions for the post-recession period (2010-2015). This represents our initial exploration of the key factors driving county- level job growth and poverty rates. Of course, cross-sectional approaches can suffer from omitted variable bias. Thus, in the next step we repeat the analysis using a differencing strategy in order to account for time-invariant unobservable factors and to benchmark the post-recession dynamics against the pre-recession period (2010-2015 minus 2000-2007). This allows us to appraise whether the Great Recession led to structural change. We then estimate corresponding models by differencing out the recession years (2010-2015 minus 2007-2010) to isolate changes that occurred since the recession. Finally, to assess heterogeneity among fast- and slow-growing locals, quantile regression of the differenced models is used to estimate changes at the 10th, 50th and 90th percentiles of the conditional distribution of the dependent variable. The analyses are performed using data for over 7 8 3,000 counties in the continental US (1,986 nonmetro and 1,052 metro). All models are estimated separately for nonmetro and metro counties to avoid aggregation bias and to account for differing levels of agglomeration. Cross-Section “Level” Equations for 2010-2015 The cross-sectional model for the 2010-2015 period is shown in (1): 𝑌 = 𝛽 +𝛽 𝑬𝑪𝑶𝑵𝟏 +𝛽 𝑬𝑪𝑶𝑵𝟐 +𝛽 𝑺𝑶𝑪 +𝛽 𝑮𝑬𝑶𝑮 +𝑿𝜷+𝜃 +𝜀 (1) !! ! ! !! ! !" ! !! ! ! ! !! where c denotes county, 𝜏 is a time period from time t to time t1, and subscript s indicates state. The error terms 𝜀 are clustered by BEA economic areas to account for spatial !! autocorrelation. Our discussion focuses on the 2010-2015 results, though we briefly review corresponding models for the 2000-2007 pre-recession and the 2007-2010 recession periods (the results are in the Appendix). The two dependent variables are the 2010-2015 annual (average) change in the poverty rate and the 2010-2015 annualized job growth. Since our sample periods have different durations, we use annualized and average measures to maintain comparability. The vectors ECON1, ECON2, SOC, and GEOG refer to economic indicators measured over the period under consideration, initial-period economic indicators (measured at the beginning of the period), initial-period social indicators and the county’s geographical attributes, respectively. Using explanatory variables at their beginning levels should alleviate reverse causality concerns, though omitted variable bias may still exist. To be sure, our key economic variables should be exogenous as described below. The vector X comprises a set of controls and 𝜃 are state dummies to capture the role of state-specific ! policies on growth and other factors fixed for each state. The average annual change in the poverty rate is calculated by dividing the change in poverty over the whole period by the number of years, whereas annualized job growth is calculated using the compound annual growth rate formulaiii. The poverty data are from the Small Area Income and Poverty Estimates (SAIPE) program and 8 9 employment is from US Census Bureau County Business Patterns (CBP). Note that CBP data do not include government employment, which means that our results are most applicable to the private sector. In addition to several traditional economic measures used in the literature, we include a set of relatively novel variables that approximate the degree of rewiring of the local economy, which, taken together, constitute the ECON1 and ECON2 vectors in Equation (1). Starting with ECON1, the industry mix variable, IndMix, is the predicted growth rate of county employment if all its industries grow at corresponding national growth rates. This measure is sometimes called the Bartik instrument (Bartik, 1991) and is routinely used as an exogenous instrument for employment growth. Rather, we are using it as an exogenous measure of demand shocks that arise from each local area having different industry compositions (Betz and Partridge 2013; Tsvetkova, Partridge and Betz 2017). Equation (2) shows how IndMix is calculated: 𝐼𝑛𝑑𝑀𝑖𝑥 = ! 𝑆ℎ 𝑁𝑎𝑡𝐺𝑟 (2) !! !!! !"# !! where all subscripts are identical to above with subscript i indicating industry at the 4- digit NAICS level and there are N industries. 𝑆ℎ is the share of industry i’s !"# employment in county c at the beginning of the period τ and 𝑁𝑎𝑡𝐺𝑟 is the annualized !! national industry growth rate over the period. Because national growth rates and initial industry shares are used, industry mix is typically assumed to be exogenous. This condition is true as long as there are no labor supply responses associated with lagged industry composition aside from labor supply variables we already control for (reducing any labor supply factors in the residual correlated with lagged industry composition). One limitation of the CBP is that it has numerous data suppressions when the Census Bureau is concerned that individual firms can be identified in the data. Generally, suppressed values are predominantly found in the information for smaller rural counties. Thus, we use CBP four-digit level data after a linear programming algorithm estimates 9 10 the suppressed values. The source for these data is the Upjohn Institute for Employment Research that uses the Isserman and Westervelt (2006) algorithm in constructing the data.iv The JobsFlow variable is a measure that approximates the expected ease of finding a job in a different industry if one is displaced from work. The variable takes into account job-to-job flow information at the 2-digit NAICS level from the US Census Bureau Longitudinal Employer-Household Dynamics (LEHD) program and industrial composition of a county at the beginning of a period as reflected in the CBP. It is calculated as follows. 𝐽𝑜𝑏𝑠𝐹𝑙𝑜𝑤 = 𝑆ℎ 𝑆ℎ 𝐹𝑙𝑜𝑤 (3) !! ! ! !"# !"# !" where 𝑆ℎ is county c’s share of employment in the origin sector i at time t, the !"# beginning of a period under consideration; 𝑆ℎ is county c’s share of employment in the !"# destination sector j at time t and 𝐹𝑙𝑜𝑤 is the percent of total employment leaving sector !" i that ends up in sector j as reflected in the LEHD. Thus, for each industry × industry pair, the larger the size of the job flow 𝐹𝑙𝑜𝑤 from industry i to industry j, the easier it should !" be to move between the two sectors if there are job losses or growth in either sector. The sectors are defined at the 2-digit NAICS level and circular flows within a sector are excluded, i.e. when calculating (3), i ≠ j. Because the job flow data is at the national level, like the industry mix term, it should be exogenous. The CBP is the data source for employment shares used in calculations. The next two measures, OccEmpMobility and IndEmpMobility, approximate the dynamics (changes) in a local economy over period 𝜏 as evidenced by moves of employees across industries and occupations during the period (Levernier, Partridge and Rickman 2000). It follows the logic of dissimilarity index used in research on racial segregation and diversity (Ellis Wright and Parks 2004) but instead of differences in a locality’s racial composition, it captures dissimilarity in employment distribution at the 10