The Spatial Distribution of Poverty and the Long Reach of the Industrial Makeup of Places: New Evidence on Spatial and Temporal Regimes

Abstract

A wealth of research identifies industrial structure as a central correlate of place-level poverty and suggests that changes in and the clustering of industry contribute to the spatial clustering of poverty over time. However, few studies have investigated the spatial and temporal dimensions simultaneously, and none have effectively examined spatio-temporal interactions. Consequently, a core tenet of theory on poverty in place has not been adequately examined. To address this limitation, we explicitly test hypotheses about systematic variation in the poverty-industry relationship over time and across space using a new method to quantify dynamic associations by simultaneously accounting for spatial and temporal autocorrelation and relationship heterogeneity. The Upper Midwest is our study site given dramatic regional changes in dominant industries (i.e., manufacturing, services, and agriculture) and poverty during the past several decades. We find that the specific character of the poverty-industry relationship systematically varies along both the temporal and spatial dimensions: industry is more protective in certain periods than in others according to sector trends, and is more protective in certain places than others conditional on sector dependence. Our approach yields a more precise and reliable understanding of the long reach of local industrial structure on the spatial clustering of poverty.

Poverty in the United States is spatially clustered. Places with high rates of poverty tend to neighbor other places reporting high poverty rates (Glasmeier 2006) and, for the most part, high poverty counties have been impoverished for decades (Beale and Gibbs 2006). Despite a vast body of research on poverty in place, scholars do not fully understand why poverty is spatially clustered. One central idea points to industry, which is spatially clustered and relatively slow-changing (Brady and Wallace 2001; Friedman and Lichter 1998, see also Voss et al. 2006; Green and Sanchez 2007; Kodras 1997; Lichter and McLaughlin 1995; Lobao and Schulman 1991; Tickamyer and Tickamyer 1988; Weinberg 1987). However, few large-scale studies of spatial inequality have explicitly and simultaneously investigated how temporal trends in and the spatial clustering of industry impact poverty. Our study addresses this limitation and, in doing so, examines a core tenet of theory on poverty in place: the role of industry. Our approach is novel in its dynamic quality, conceptualizing the temporal dimension in terms of industry protectiveness and the spatial dimension as industry dependence.

Recent research has used advanced statistics to begin to address the interconnectedness of space and time to understand what generates, perpetuates, and changes the spatial pattern of poverty (Chokie and Partridge 2008; Curtis et al. 2013; Jha 2000). However, no study to date has investigated the ways in which space and time operate to affect the relationships between county poverty and its hypothesized drivers (i.e., industry). There are two central limitations to existing work. First, either serial or spatial autocorrelation is addressed, but not both. Consequently, estimates are potentially inaccurate and, in turn, conclusions are potentially incorrect since addressing one form of autocorrelation does not necessarily address the other. Second, research has not adequately estimated the potentially varying influence of industry between time periods and across specific places. Research generally has adopted an analytical approach that quantifies an average association that is applied to all places and/or in all periods. This strategy is problematic because research demonstrates and theory asserts variation in the relationship between poverty and its drivers over time and across places. Industry, while slow-changing, nonetheless is dynamic and is spatially clustered. An analytical strategy that explicitly addresses the dynamic qualities of poverty in place is necessary.

We ask a central question about spatial inequality: Does the spatial clustering of industry underlie the spatial pattern of poverty over time? In our study, we apply advanced spatial statistical models to determine whether industrial clustering explains the spatial patterning of poverty over time. Specifically, we develop and test hypotheses about spatial and temporal variation in the poverty-industry relationship using a modified regime approach (Curtis et al. 2012; O’Loughlin et al. 1994). We choose as our site six states in the Upper Midwest that have experienced marked changes in poverty and industry over the past 50 years to examine how industrial shifts affect the spatial distribution of poverty. The study region is ideal for our research question given the wide range and dynamic nature of poverty and industrial concentration.

Our study shows that fortunes in place depend on the rise and fall of industrial sectors. With the demise of high-paying manufacturing jobs, places within spatial clusters of manufacturing jobs are at less of an historic advantage now more than ever, especially places that experienced the most dramatic change in manufacturing employment. Places within clusters of service sector jobs have a similar pattern to that found for manufacturing, but there is no statistical evidence of an association after considering spatial and serial autocorrelation, patterns in other industries and unemployment, and the metropolitan status of the area. In contrast, places with relatively large shares of agricultural employment are less likely to be in high poverty areas over time, most especially in the recent recession period, suggesting that a small and dwindling farm population benefitted from recent increases in crop prices and longer-term income subsidy programs. Our results suggest that part of the clustering of poverty is due to the long reach of the industrial makeup of places. Ultimately, the precise nature of the poverty-industry relationship varies along both temporal and spatial dimensions, with industry more versus less protective in certain periods that reflect sector trends and more versus less protective in certain places contingent on sector dependence.

SPATIAL CLUSTERING IN POVERTY AND INDUSTRY TRENDS IN THE MIDWEST

Poverty in the United States has generally declined since the 1960s, but the local manifestation of that trend has varied tremendously across place. High-poverty counties have been impoverished for decades (Beale and Gibbs 2006) and tend to neighbor one another, creating spatial clusters of concentrated poverty (Adams and Duncan 1992; Lobao and Saenz 2002; Lobao 2004; Glasmeier 2006; Partridge and Rickman 2006). Concentrated and persistent poverty is historically highest in rural areas (Tickamyer and Duncan 1990; Weinberg 1987), places that tend to have populations subjected to lower wages, geographic isolation and limited public transportation, lower quality schools, restricted access to basic public services and institutional support, and greater exposure to environmental toxins and work-place injuries (Brown and Swanson 2003; Jensen et al. 2003). However, research has shown that some spatial clusters of poverty may be drying up and, relatedly, spatial inequality declined during the 1990s when economic conditions broadly improved throughout most of the United States (Lichter and Johnson 2007), and tax and welfare policy reforms were enacted (Lichter and Jensen 2002; Lichter et al. 2005). Critically, research has shown that forces promoting poverty are unevenly distributed across space, most centrally industry and accompanying industrial shifts (Curtis et al. 2012; Friedman and Lichter 1998; McCall 2001; McLaughlin and Perman 1991). Examining the spatial and temporal dimensions of the relationship between industry and poverty is essential to understanding the spatial distribution of poverty.

The Midwest is not known as an area of high poverty concentration. Indeed, the Midwest has generally fared better than the nation overall. However, the regional advantage has narrowed in recent years with reported poverty rates of 18% in 1959 (versus 22% for the nation) compared to 10% in 2000 (versus 12%) and 14% in 2015 (versus 15%) (US Census Bureau 2015, 2016). In addition, the region is home to counties with some of the lowest poverty rates in the nation (e.g., Ozaukee County, Wisconsin) but also some of the highest (e.g., Wayne County, Michigan). The region’s temporal trends in poverty, wide range of poverty rates, concentration of rural areas, and unique mix of industry make it a compelling case to examine the poverty-industry relationship.

Manufacturing, services, and agricultural industries have experienced dramatic shifts especially since the 1970s and, given the region’s dependence on each of these sectors, so too has the Midwest’s economic vitality undergone marked change. In the 1960s, the Midwest and Northeast housed two-thirds of the nation’s manufacturing jobs, most of which were located around central cities (Kasarda 1976, 1978). Such strong dependence made the Midwest especially sensitive to the contraction of the manufacturing industry that ensued in the 1970s. Total manufacturing earnings declined in the Midwest since 1970 even though the region still contributed one-third of the nation’s total manufacturing earnings through 1990 (Kasarda 1995). Between 1980 and 1990, for example, manufacturing earnings fell by 25% in Cook County, IL (Chicago), 32% in Wayne County, MI (Detroit), and 33% in Milwaukee County, WI (Milwaukee) (Kasarda 1995:262).

At the same time, the Midwest, like other regions, grew in service sector jobs (Kasarda 1989; 1995) but with differential consequences for the region (Bluestone 1990). This industrial transition – deindustrialization – was marked by declines in once well-paying manufacturing jobs, the emergence of low-wage manufacturing jobs, and concurrent increases in service jobs, many of which were low-paying and some of which were high-paying although reserved for more educated laborers. In addition to the direct impact of changes in the industrial structure, institutional factors such as the decline of unionization promoted greater economic vulnerability among places relying on manufacturing and service jobs (Bluestone 1990; McCall 2001). The Midwest was particularly affected by deindustrialization given a strong dependence on manufacturing.

The Midwest also is home to one of the most productive areas of agriculture in the world, generating a wide range of goods including corn, soybeans, livestock, vegetables, fruits, tree nuts, berries, and nursery and greenhouse plants (Hatfield 2012:3). As with manufacturing and services, the agriculture sector has experienced a remarkable transformation. Trends in agriculture in the Midwest have mirrored those of the nation. Farm mechanization occurred well before the 1970s, yet the technological advancements of the late 20^th century had significant impacts for human labor and, in turn, affected earnings on the farm and in related support sectors (e.g., seed and parts distributors, financial institutions) (see Lobao and Meyer 2001). As farm labor has declined, off-farm work in other sectors has increased (Dimitri et al. 2005). Thus, an increasing share of the farm population has turned to other economic sectors. Within this context of industrial restructuring and shrinking farm employment, farm policies have evolved to stabilize farm income, ranging from loan programs to direct payments (e.g., Goodwin and Smith 2013).

Recent industrial trends center on the Great Recession, which differentially impacted particular sectors as broad and deep contractions emerged in 2007 and again in 2009. Construction, manufacturing, and finances (representing the higher-wage service jobs) were most hard-hit whereas the energy and agriculture sector (except forestry given the connection to construction) were less impacted (Kusmin 2009; Uphoff 2010). Indeed, farmers and ranchers earned the highest real income since 1973 in the four years following the Great Recession, and farm income in the Midwest had been waning until the recent boom in crop prices generated a boon for producers (Oppedahl 2015). It follows that places more dependent on contracting industries (i.e., manufacturing and services) were likely differentially and adversely impacted by the Great Recession as compared to places more dependent on less-affected industries (i.e., agriculture). Still, as a whole, places with high concentrations of employment in agriculture might not have benefited given the farm population’s relatively small share of an area’s total population.

A central thesis in research on the poverty-industry link is that spatial clusters of higher and lower poverty emerge because specific industries locate in particular places. The recession and earlier dramatic industrial shifts, such as deindustrialization and agricultural mechanization, combined with institutional forces (i.e., deunionization and income subsidies) to reorient local and regional economic development (Kasarda 1989, 1995; McCall 2001). Consequently, temporal trends in and spatial patterns of poverty might reasonably trace trends in and patterns of industry in a way that produces systematic variation in the nature of the relationship (the extent to which an industry is protective against poverty) and the strength of the relationship (the extent to which an area is dependent upon an industry).

Although scholars have demonstrated that industrial structure is related to the spatial distribution of economic outcomes (e.g., Friedman and Lichter 1999; McCall 2001; Parks 2012; Voss et al. 2006), there are few examples of research that have adequately examined whether industry’s relationship with economic outcomes varies across space and over time. Indeed, while there is a voluminous literature on the basic connection between industry and poverty, there is scant research that investigates the way in which industry affects economic outcomes differently across places or types of places. Some research has explored variation in industry effects between non-metropolitan and metropolitan areas (Gibbs et al. 2004; Goe 2002), but most research focuses on variation in economic outcomes only (e.g., unemployment rates, income inequality, median earnings) (Lobao 2004; McCall 1998). The primary exception is Green and Sanchez (2007), who investigate industry’s relationship to underemployment and find that places with an historical dependence on manufacturing benefit less from manufacturing as compared to areas that have only recently shifted economic activity into the manufacturing sector. This research suggests potentially meaningful spatial differences in economic fortunes that are driven by trends in industries upon which places depend.

A larger body of work has examined the effect of industrial changes over time on economic outcomes. In their analysis of manufacturing and welfare receipt between 1964 and 1993, Brady and Wallace (2001) find that loss of manufacturing jobs in previous years is associated with increases in welfare receipt in subsequent years and that gains in service jobs – in effect, jobs that replaced manufacturing jobs – did not reduce welfare receipt (see also Lichter and McLaughlin 1995). In their study of the agricultural industry, Lobao and Schulman (1991) find that shifts within the industry (i.e., mechanization) since the 1960s have concentrated the economic benefits to a smaller portion of the population, and thereby limit the potential for agriculture to promote economic development in the aggregate. Together, previous research suggests that industrial shifts have promoted increasing economic vulnerability over time among places relying on manufacturing, services, and agriculture.

A DYNAMIC APPROACH: INDUSTRY PROTECTIVENESS AND DEPENDENCE

Prior work addresses how industries have changed and the implications for economic outcomes, yet it does not address how the relationship between industry and economic outcomes changes over time or the associated implications for spatial patterns of inequality. This subtle but important distinction is central to our study. We are concerned with whether the spatial clustering of industry underlies the spatial pattern of poverty over time, understanding that while industry is slow-changing and poverty is generally persistent, both structures are temporally dynamic. We investigate the implications of the dynamic qualities of industry and poverty to explain the spatial clustering of poverty in the Midwest by conceptualizing temporal and spatial variation in terms of industry protectiveness and dependence.

Within the spatial literature, dynamic relationships are considered a form of spatial heterogeneity (LeSage 1999) and are evidenced by differences in the magnitude and nature of relationships across a spatial region. This perspective directly challenges the “constancy assumption” (Freedman et al. 1991:678) that underlies most statistical approaches, which summarize all observations to yield a single “average” estimate. In our study, the constancy assumption, and associated standard analytical approaches, asserts that industry has the same association with poverty in all places and periods. In contrast, a spatial heterogeneity approach permits spatially-specific circumstances to influence structural relationships (O’Loughlin et al. 1994:359). The unique combination of social forces (e.g., cultural and/or historical processes) in a place may generate effects not found in places with other, unique combinations of social forces (Massey 1994:120). That is, for our study, the relationship between industry and poverty is expected to vary across spatial units. Extended to the temporal dimension, the approach also permits industry to affect poverty differently over time.

Adopting a heterogeneity perspective, we develop and test two hypotheses. First, in terms of temporal patterns, we hypothesize the relationship between each of the dominant industries – manufacturing, services, and agriculture – and poverty will become increasingly positive. Specifically, we hypothesize that the industries will grow less protective over time and, thus, more positively associated with poverty. This hypothesis is consistent with previous research examining the temporal dimension of how industry relates to poverty (e.g., Brady and Wallace 2001), but our analytical approach provides a more robust assessment than earlier studies by accounting for the impact of spatial processes also at play. We refer to this as the “industry protectiveness” hypothesis.

Second, in terms of spatial clustering, we elaborate on our temporal hypothesis by investigating whether the way in which industry relates to poverty and how that relationship changes over time will depend on the spatial context. Specifically, we hypothesize that manufacturing and service jobs will be more positively associated with poverty among counties in states that are historically dependent on manufacturing (i.e., Illinois) compared to other Upper Midwestern states. Similarly, we hypothesize agriculture will be more strongly related to poverty in historically agriculture-dependent states (i.e., Iowa) relative to other Upper Midwestern states. Examining how the temporal dimension of the poverty-industry relationship differs across spatial regimes explicitly acknowledges the intersection between temporal and spatial dynamics underlying the distribution of poverty in the Upper Midwest. We refer to this as the “industry dependence” hypothesis.

In the following section, we describe our data, measures, and analytical strategy for assessing the potentially dynamic contribution of industry to poverty while accounting for serial and spatial autocorrelation structures underlying the data, which can compromise statistical estimates and subsequent conclusions. Our strategy permits us to generate unbiased estimates that test long-standing tenets about the temporal trends in and the spatial clustering of poverty in relation to industry.

DATA AND MEASURES

We examine county panel data for six states in the US Upper Midwest between 1970 and 2010 drawn from decennial censuses and the 2006-10 American Community Survey 5-year Estimates. The area’s overall trends and wide range in poverty plus the industrial mix and dramatic industrial restructuring since the mid-20^th century make it an ideal site for our research on heterogeneity in the poverty-industry association. We examine all counties within the states of Illinois, Indiana, Iowa, Michigan, Minnesota, and Wisconsin.¹ Descriptive statistics on central variables are included in Tables A1 through A4 in the accompanying appendix.

Poverty

Our dependent variable is the reported poverty rate, based on the proportion of the county population living below the poverty threshold (logit transformed), which is ideal for our purposes since it is comparable across time.² We analyze county-level poverty for two reasons. First, we are interested in investigating the association between county-level factors and county poverty given the political and administrative role of the county. This study does not address individual-level processes that contribute to poverty. Rather, we are interested in understanding the spatial inequality in economic vulnerability among counties. Second, the county is a logical unit of interest because it embodies structural factors that produce economic vulnerability. Poverty happens to communities and to places. From a place-based perspective, counties are geographic units that represent socially constructed yet physically bounded areas with characteristics that interact with location to create spatially divergent social, economic, and political outcomes.

We report poverty rates from 1970 through 2006-10 for the six states in our study region with a comparison to the national average in Table 1. The estimates reveal a wide range in poverty across states and the deteriorating regional advantage over the nation. Minnesota and Iowa reported the most stability in poverty over the period, whereas Michigan and Indiana show the most variable poverty rates. For Iowa, poverty rates ranged from a low of 9.1% in 2000 to a high of 12.2% in 2006-10, a 3.1 percentage-point difference. Similarly, Minnesota had a 3.3 percentage-point range. In contrast, Michigan’s range was 6.9 percentage-points, with a period low of 9.4% in 1970 and high of 16.3% in 2006-10. Not as severe, but still substantial, poverty in Indiana ranged by 5.3 percentage-points. When comparing the bookends of the study period (1970 and 2006-10), the range of change in poverty rates is considerable, with a 0.5 percentage-point change in Minnesota and a 6.9 percentage-point change in Michigan. Consistent across all states, the change was positive – poverty was highest for all states in the most recent period under analysis.

Table 1.

Total (Individual) Poverty Rate (%) in the Upper Midwest by State with National Comparisons, 1970 to 2006-10

	1970	1980	1990	2000	2006-10	Δ 1970-2010
United States	13.7	12.4	13.1	12.4	13.8	0.1
Illinois	10.2	11.0	11.9	10.7	13.7	3.5
Indiana	9.7	9.7	10.7	9.5	14.7	5.0
Iowa	11.6	10.1	11.5	9.1	12.2	0.6
Michigan	9.4	10.4	13.1	10.5	16.3	6.9
Minnesota	10.7	9.5	10.2	7.9	11.2	0.5
Wisconsin	9.8	8.7	10.7	8.7	12.5	2.7

The trend in increasing poverty underlies the deteriorating regional advantage over the national average. Indeed, the range of percent change over the 40-year period for the nation was less than 1%. Before 1990, no state within the study region exceeded or even matched the national poverty rate. However, in 1990, Michigan was at the national average and by 2006-10, Illinois, Indiana, and Michigan had similar or higher poverty rates than the nation overall.

Figure 1 shows poverty rates by county in the six Midwestern states from 1970 through 2006-10. The severity of poverty was not stable over the study period. Poverty was highest in 1970, with a large majority of counties in the Upper Midwest reporting poverty rates at 10% or higher. Poverty rates decreased during the 1970s, grew during the 1980s, declined again in the 1990s, and increased once more in the 2000s. Although the extent of poverty varied over this period, the spatial distribution of poverty was relatively stable. The value of the highest poverty rate was not the same across the period, yet the highest rates were consistently concentrated in the northernmost and southernmost counties of the Upper Midwest. The lowest levels of poverty were located within the mid-section of the region (where low-poverty Ozaukee County, WI is located). Overall, as poverty generally declined, pockets of concentrated poverty contracted, leaving increasingly fewer counties at a growing disadvantage relative to other counties within the region. As poverty increased, spatial clusters of poverty expanded, creating larger clusters of disadvantaged counties.

Figure 1.

Industry

Local-area industrial structure is represented by variables reflecting the proportion of the civilian labor force 16 years-old and older in each county that is employed in manufacturing, services, and agriculture. We focus on these three industries since they are central to the region and have undergone significant changes since 1970. We report the proportion of the workforce employed in each of the industries for the region as a whole and for each state by decade in Table 2. For the region overall, the proportion employed in manufacturing and agriculture decreased between 1970 and 2006-10, whereas employment in services increased over the period.

Table 2.

Percent Employed in Manufacturing, Service, and Agriculture in the Upper Midwest by State, 1970 to 2006-10

	1970	1980	1990	2000	2006-10	Δ1970-2010
Manufacturing
Region	24.6	21.4	20.1	19.6	17.5	−7.1
Illinois	24.6	21.0	18.2	16.8	14.7	−9.9
Indiana	35.8	29.8	27.1	26.0	22.9	−12.9
Iowa	15.8	16.7	16.9	18.4	17.1	1.3
Michigan	29.7	22.5	20.4	19.3	16.1	−13.6
Minnesota	16.5	15.6	16.0	16.6	15.3	−1.2
Wisconsin	26.4	23.5	22.7	21.4	19.5	−6.9
Service
Region	23.4	23.7	25.7	25.4	27.4	4.0
Illinois	22.9	22.9	24.9	25.1	28.0	5.1
Indiana	21.7	22.4	24.7	24.6	26.7	5.0
Iowa	25.1	25.6	25.8	24.2	26.2	1.1
Michigan	23.3	23.0	27.2	27.0	28.9	5.6
Minnesota	24.4	24.9	26.0	25.7	26.6	2.2
Wisconsin	23.1	23.3	25.7	26.1	28.1	5.0
Agriculture
Region	12.3	9.6	7.7	4.6	4.5	−7.8
Illinois	10.3	7.8	6.7	3.9	3.6	−6.7
Indiana	6.7	5.3	4.3	2.3	2.5	−4.2
Iowa	21.0	16.3	12.9	7.5	7.1	−13.9
Michigan	4.3	3.5	3.2	2.3	2.6	−1.7
Minnesota	17.8	13.8	10.7	6.7	6.3	−11.5
Wisconsin	12.7	10.4	8.3	5.2	4.6	−8.1

At the beginning of the period, manufacturing was largely concentrated among counties in the eastern-most states within the Upper Midwest, especially in counties comprising the central part of the region (Figure 2). Correspondingly, industry contractions were concentrated in these areas, as formerly secure and well-paying manufacturing jobs moved southward within and beyond the US borders (Kasarda 1989; 1995). Indiana reported the highest employment in manufacturing throughout the study period, followed by Wisconsin (Table 2). Michigan also reported relatively high manufacturing employment and experienced the largest decline over the period, dropping by 13.6 percentage-points. Similarly, Indiana’s rate of manufacturing employment fell by 13 percentage-points followed by Illinois’ 10 percentage-point drop. In contrast, Wisconsin’s rate declined by 7 percentage-points between 1970 and 2006-10 (the regional average).

Figure 2.

The service sector was emerging in the Upper Midwest as manufacturing was declining and restructuring. Growth in this sector, however, might not necessarily reduce economic vulnerability but may have promoted poverty given the less secure qualities that tend to characterize a growing share of service jobs (e.g., low pay, low benefits, low hours) (Collins and Mayer 2010). Services have increased modestly throughout the entire Midwest (Figure 3), with 23% employed in 1970 versus more than 27% in 2006-10. In the recent period, services have concentrated somewhat in the northern parts of the region and among a stretch of counties in northern Illinois. Despite the recent contraction for specific counties, taken as a whole, Michigan (29%), Wisconsin (28%), and Illinois (28%) report the highest rates of service employment in 2006-10. The rate of growth in the service sector was highest for these three states and Indiana, each increasing by around 5 percentage-points.

Figure 3.

The agricultural sector experienced its share of economic contraction through mechanization and, in recent years, economic expansion during and on the tails of the recession. Changes in this sector might have produced increased economic vulnerability over the study period as a growing share of the total and farm population relied on non-farm sectors (Dimitri et al. 2005; Lobao and Schulman 1991). The overall proportion of agricultural workers has continued to decline in the region, falling from 12% in 1970 to 4.5% in 2006-10. However, agricultural employment has remained concentrated in the more westward states and counties comprising the northern-most and southern-most parts of the region since 1970 (Figure 4). At the end of the period, Iowa (7%), Minnesota (6%), and Wisconsin (5%) reported the highest rates of employment in agriculture (Table 2). Iowa and Minnesota also experienced the largest losses since 1970, declining by 14 and 11 percentage-points, respectively. Despite these losses, Iowa and Minnesota reported the highest rates of agriculture employment in the region at 7% and 6%, respectively.

Figure 4.

We are motivated to present the most parsimonious model given the statistical complexity and computational demands of integrating the spatio-temporal autocorrelation structure. Thus, our current analytical attention is limited to the key confounders of industry, unemployment, and metropolitan status. Industry confounders include all other sectors – mining, FIRE (finance, insurance, and real estate), and other professional positions (e.g., science, technology, education). We also control for the county unemployment rate to distinguish the impact of industry contraction on poverty through unemployment from the influence of changes in the structure of employment within the industry; we are interested in employment in specific industries, not an absence of employment. Finally, we account for the influence of metropolitan status since both poverty and industry systematically vary between non-metropolitan and metropolitan counties.

ANALYTICAL STRATEGY

Addressing our question requires attention to spatial and temporal patterns, both of which are embedded in concerns regarding “autocorrelation.” Spatial and serial autocorrelation refers to the similarity of variable values between observations due to the proximity of those observations, either in terms of space or time. In addition to the need to address these elements of the data structure to generate robust and unbiased estimates, we have theoretical motivations to model spatial and temporal dynamics simultaneously. Previous work offers an analytical approach to estimate both the spatial and temporal autocorrelation in a single model (Curtis et al. 2013). However, this approach does not address the interaction between temporal changes and how temporal dynamics differ across space. This is central to our research question and a key contribution of our study. To assess the spatio-temporal dynamics characteristic of our data structure and of substantive relevance to our research question, we develop a new data analytical method that accounts for both aspects of autocorrelation and is feasible and practical to accommodate the added complexity of a space-time interaction. In doing so, our study contributes to this journal’s tradition of developing and applying advanced spatial regression models to sociological and demographic research (e.g., Brasier 2005; Brown et al. 2011; Chi 2010; Deller and Deller 2012; Lobao et al. 2016; Peters 2012).

We adopt an approach akin to a “regime” analysis while simultaneously accounting for underlying spatial and serial autocorrelation. Space and time are treated as endogenous effects through a spatial error regression and, concurrently, as exogenous effects through a regime approach. By incorporating both endogenous and exogenous effects, we address both dependence and heterogeneity processes, respectively. In our study, we are primarily concerned with the heterogeneity process (i.e., identifying the relationship between industry and poverty) net of any underlying dependence process, which is likely an artifact of the spatio-temporal panel data structure. Standard approaches to spatio-temporal models do not accommodate the complexity of this space-time interaction. Thus, identifying spatial and temporal relationship heterogeneity simultaneously requires the use of new, advanced data analytical methods.³

Temporal Regime Autoregressive Model: Industry Protectiveness

For spatial unit i=1,…, N and time point t=1,…, T, let Y_it denote the response variable at the i th spatial unit and the t th time point. Our data consists of 533 counties (N = 533) in 6 states (S = IL, IN, IA, MI, MN, and WI) for 5 census years (T = 1970T = 1980T = 1990T = 2000, and 2006-10); Y_it is the logit poverty rate at (i, t), t=1 is census year 1970, t=2 is census year 1980, and so on. To account for spatial and temporal autocorrelation we use a Besag model with a second-order queen neighborhood structure for spatial dependence, and an autoregressive AR(1) model with time lag 1 (10 years) for temporal dependence. Let x_it denote the intercept and the industry variables and β_t the corresponding regression coefficients at time t. Let τ_t denote the AR (1) temporal autoregressive effect, ε_i the Besag spatial autoregressive effect, and γ_it white noise. We use a spatio-temporal model as follows.

$$Y_{it}=x_{it}^T{\beta }_t+{\tau }_t+{\epsilon }_i+{\gamma }_{it}$$ (2)

We use the model represented in equation (1) to test our first hypothesis on the poverty-industry relationship, what we call industry protectiveness. We derive estimates of how the key industries (i.e., manufacturing, services, and agriculture) relate to poverty for the whole region and assess whether the relationships are increasingly or decreasingly protective over time. We employ pairwise comparison tests that are adjusted for multiple comparisons to determine the extent to which an industry’s relationship with poverty differs across the observed years. Further, the residuals of equation (1) did not provide evidence of a non-Gaussian distribution; thus, an assumption of Gaussian distribution of errors is reasonable.

Spatial Regime Autoregressive Model: Industry Dependence

To test our second hypothesis, what we call industry dependence, we derive state-specific estimates of how the temporal trend(s) differs spatially, according to state historical dependence on industry.

Let ${\beta }_t^S$ denote the regression coefficients for spatial regime S at time t. This is done for each spatial regime (i.e., IL, IN, IA, MI, MN, and WI), which is identified using the indicator function $𝟙\{\cdot \}$.

$$Y_{it}=x_{it}^T{\beta }_t^S\dot{}𝟙\left\{i\in S\right\}+{\tau }_t+{\epsilon }_i+{\gamma }_{it}$$ (2)

For example, to estimate the poverty-industry relationship for counties in Wisconsin at t = 1 (year 1970), we draw from the following model

$$Y_{i1}=x_{i1}^T{\beta }_1^{WI}+{\tau }_1+{\epsilon }_i+{\gamma }_{i1}$$

The state-specific regime analysis will enable us to report coefficient estimates for the focal industry variables (i.e., manufacturing, services, and agriculture). Just as for hypothesis 1, we will assess the temporal trend using pairwise comparison tests that are adjusted for multiple comparisons, but now we also include a test for the extent to which the coefficients differ across states in the same year. Our strategy generates accurate estimates that test long-standing tenets about the temporal trends in and spatial clustering of poverty. In addition, the residuals of equation (2) did not provide evidence of a non-Gaussian distribution; thus, the model assumption of a Gaussian distribution is reasonable for the errors.

The autocorrelation structure in equations (1) and (2) is spatio-temporal separable and adequate because the regression coefficients vary in space and time. For estimation of the parameters in (1) and (2), we approximate the posterior means by an integrated nested Laplace approximation (INLA; Rue et al. 2009), which enables the fitting of a wide class of models ranging from (generalized) linear mixed models to spatio-temporal models (Blangiardo et al. 2013) and is readily available for use in R software (R-INLA). R-INLA provides the posterior mean and the standard deviation for each regression coefficient in the Bayesian framework as well as the fitted value for each observation. The maximum likelihood estimate (MLE) is traditionally used for estimating autocorrelation models, however, the computational time required to solve a complex autocorrelation model using MLE is a major limitation as the necessary time can easily turn into days. Incorporating INLA into our estimation provides a significant improvement over previous efforts to draw inference about complex space-time models because its computational efficiency allows us to empirically assess how spatial and temporal differences interact to affect the poverty-industry relationship.

As with all modeling strategies, INLA has strengths and limitations. A central limitation is that while the approach can deal with various kinds of models, some important spatial models are not yet available in R-INLA (e.g., SAR). Still, the most relevant models can be implemented via R-INLA, including multiple likelihood and various latent models. Moreover, INLA is readily implemented for users through the R program (R-INLA package http://www.r-inla.org/ (Rue et al. 2009).

RESULTS

Temporal Variation: Industry Protectiveness

The temporal regime analysis shows some support for our first hypothesis on industry protectiveness. The relationship between industry and poverty changes over time such that industry prevalence grows less protective over time, although distinct trajectories are found for each industry (Table 3). Counties with a higher concentration of manufacturing employment are increasingly likely to be within spatial clusters of poverty (less protective), whereas the opposite pattern is found for places with higher concentrations in agriculture employment (more protective). Counties with a higher concentration of service employment share a similar pattern to manufacturing (less protective), although it is not statistically significant.

Table 3.

Regression results testing for temporal variation in poverty-industry relationship, all Upper Midwest counties by year (N=533)

		β		SE
1970	Intercept	−1.45	***	0.42
	Manufacturing (%)	−1.63	***	0.42
	Service (%)	−0.58	ns	0.53
	Agriculture (%)	0.69	ns	0.43
1980	Intercept	−2.30	***	0.46
	Manufacturing	−0.89	*	0.45
	Service	0.36	ns	0.56
	Agriculture	1.40	**	0.47
1990	Intercept	−2.07	***	0.44
	Manufacturing	−1.12	**	0.44
	Service	0.59	ns	0.53
	Agriculture	1.01	*	0.48
2000	Intercept	−2.08	***	0.43
	Manufacturing	−0.99	*	0.43
	Service	0.36	ns	0.52
	Agriculture	0.26	ns	0.54
2000-10	Intercept	−1.71	***	0.43
	Manufacturing	−0.87	*	0.43
	Service	0.43	ns	0.51
	Agriculture	−1.07	ns	0.55
Autocorrelation	rho	0.00	ns	0.70
	nu	1.40	***	0.10
Residual autocorrelation (MI)	1970	0.10
	1980	0.03
	1990	0.00
	2000	0.02
	2006-10	0.00
Model fit	DIC	−2838.45

Note: Models include controls for extractive, FIRE, and professional industry sectors, unemployment rate, and metropolitan status (coefficients not reported).

^***p <.001,

^**p <.01,

^*p <.05, ns non-significant

The spatio-temporal model parameterizes the spatial and serial autocorrelation structures. The rho reflects the serial autocorrelation structure (in the tau of equations 1 and 2), where rho is the 10-year temporal lag for all time points. The estimate is close to zero, suggesting there is no significant endogenous effect in poverty rates between years. The nu corresponds with the spatial autocorrelation structure (in the epsilon of equations 1 and 2) and is positive for all counties, evidencing spatial clustering and indicating that counties tend to have poverty rates that are similar to those of their neighbors. We report the deviance information criterion (DIC), which can be interpreted similarly to AIC or BIC. Model diagnostics show that residual autocorrelation is near zero, suggesting no remaining effects of serial or spatial autocorrelation.

For ease of interpretation, we visualize the focal industry coefficients generated in the multivariate temporal regime analysis (Figure 5). The figure highlights the nature of change in the relationship to poverty for each industry. The coefficients for manufacturing and service employment trend in a direction that is consistent with our hypothesis that industry grows less protective over time, whereas agriculture employment shows an increasingly protective association. Manufacturing is negatively associated with poverty throughout the period, indicating an overall protective association. However, the magnitude of the association weakens over the period; places with high concentrations of employment in manufacturing tend to have lower poverty rates in 1970 than in any other decade, especially compared to 1980 and 2006-10. Taking the period as a whole, manufacturing becomes increasingly less protective. The temporal regime analysis illuminates the dynamic nature of the poverty-manufacturing relationship, and demonstrates that the relationship changed over time in a way that is consistent with industry trends and shifts.

Figure 5.

Similar to manufacturing, service grows increasingly disadvantageous over the study period. However, the poverty-service relationship is not statistically significant, at least not for the region as a whole and within a multivariate analysis, as discussed further in the next section. Employment in service jobs is not statistically associated with poverty in any year for all counties combined – it is neither protective nor detrimental for the Upper Midwest as a whole.

The relationship between agriculture and poverty is dynamic over time, and in an entirely different way than manufacturing and service. The poverty-agriculture relationship has grown more protective over time, however, it is statistically significant for all counties combined in only 1980 and 1990.

Results from the temporal regime analysis confirm significant temporal variation in the relationship between poverty and industry in ways that are sometimes consistent with expectations. We find support for an increasingly less protective impact of manufacturing on poverty, suggesting that the industry has grown less economically secure over time. A similar, but statistically non-significant temporal pattern is found for services, and the opposite pattern is found for agriculture.

These estimates have the advantage of taking into account the serial and spatial autocorrelation structures, thereby generating coefficients robust against the endogenous effects of time and space. They apply to the entire study region and, thus, reflect for each point in time the average poverty-industry relationship for counties in all six states. Our central question concerns the spatial pattern of poverty over time, and whether the persistence or change corresponds with industrial shifts. Thus, with the temporal pattern of the poverty-industry relationship in mind, we now turn to our spatial regime results to identify whether industry dynamics underlie patterns in the spatial clustering of poverty.

Spatial Variation: Industry Dependence

Consistent with the industry dependence hypothesis, we find evidence that the relationship between industry and poverty is spatially clustered and differs across states. However, the thrust of the results suggests that the spatial pattern is complex and can be understood only in part by, or a change in, dependence on industry. Indeed, there is substantial evidence of deviation from the industry dependence hypothesis as originally stated.

The spatio-temporal model parameterizes the spatial and serial autocorrelation structures underlying the spatial panel data (Table 4). The nu is positive, showing that counties tend to have poverty rates similar to neighboring counties, whereas the rho is negative, indicating that poverty rates differ between time periods. Model diagnostics show that residual autocorrelation is negative, suggesting a slight overcorrection that is common when accounting for serial and spatial autocorrelation.

Table 4.

Regression results testing for spatial variation in poverty-industry relationship, all Upper Midwest counties by state and year (N=533)

		1970			1980			1990			2000			2006–10
		β		SE	β		SE	β		SE	β		SE	β		SE
Illinois	Intercept	−1.77	ns	0.91	−3.23	***	0.80	−2.88	***	0.71	−3.20	***	0.70	−4.44	***	0.83
	Manufacturing (%)	−1.58	ns	0.91	0.09	ns	0.78	0.13	ns	0.67	0.32	ns	0.69	1.77	*	0.81
	Service (%)	−1.64	ns	1.14	0.20	ns	1.09	1.63	ns	1.00	1.10	ns	0.99	3.34	***	1.03
	Agriculture (%)	1.59	ns	1.00	1.09	ns	0.85	1.73	*	0.87	1.15	ns	1.09	2.00	ns	1.22
Indiana	Intercept	−1.70	**	0.66	−1.83	ns	0.99	−1.24	ns	1.04	−0.53	ns	1.22	−1.04	ns	1.12
	Manufacturing	−1.19	ns	0.65	−1.21	ns	1.00	−1.94	ns	1.04	−2.68	*	1.22	−1.95	ns	1.15
	Service	−0.18	ns	1.11	−0.59	ns	1.28	−2.20	ns	1.38	−1.19	ns	1.57	−1.98	ns	1.39
	Agriculture	1.18	ns	0.79	0.94	ns	1.14	−0.20	ns	1.36	−2.27	ns	1.89	−1.41	ns	1.87
Iowa	Intercept	−1.25	ns	1.45	−0.57	ns	1.40	0.28	ns	1.50	−0.09	ns	1.22	0.07	ns	1.13
	Manufacturing	−1.54	ns	1.50	−2.22	ns	1.43	−3.23	*	1.56	−3.06	*	1.27	−2.24	ns	1.15
	Service	−0.30	ns	1.71	−2.00	ns	1.57	−2.56	ns	1.73	−1.02	ns	1.48	−2.38	ns	1.35
	Agriculture	0.41	ns	1.50	0.59	ns	1.41	−0.88	ns	1.59	−1.65	ns	1.36	−4.60	***	1.34
Michigan	Intercept	−0.95	ns	0.90	−2.80	**	0.91	−1.79	*	0.88	−1.80	*	0.88	−2.12	**	0.81
	Manufacturing	−2.08	*	0.89	−0.31	ns	0.88	−1.02	ns	0.92	−0.94	ns	0.91	−0.30	ns	0.91
	Service	−0.74	ns	1.26	1.42	ns	1.38	0.95	ns	1.27	1.20	ns	1.22	1.85	ns	1.13
	Agriculture	−0.52	ns	0.98	1.37	ns	1.11	0.40	ns	1.14	−0.50	ns	1.61	0.56	ns	1.31
Minnesota	Intercept	−2.29	ns	1.43	−1.84	ns	1.46	−1.67	ns	1.53	−1.99	ns	1.42	−2.97	*	1.28
	Manufacturing	−0.59	ns	1.43	−1.20	ns	1.47	−1.60	ns	1.53	−1.52	ns	1.44	0.31	ns	1.30
	Service	1.77	ns	1.74	0.61	ns	1.75	0.09	ns	1.76	0.83	ns	1.63	2.68	ns	1.49
	Agriculture	1.91	ns	1.44	1.72	ns	1.50	0.28	ns	1.61	0.25	ns	1.60	0.16	ns	1.51
Wisconsin	Intercept	−1.70	ns	1.49	−0.94	ns	1.38	−0.94	ns	1.47	−1.03	ns	1.45	0.25	ns	1.34
	Manufacturing	−1.71	ns	1.47	−2.58	ns	1.32	−2.58	ns	1.46	−2.62	ns	1.45	−3.07	*	1.34
	Service	1.16	ns	1.89	−0.55	ns	1.77	−0.55	ns	1.63	−0.99	ns	1.58	−2.22	ns	1.55
	Agriculture	1.46	ns	1.49	0.57	ns	1.41	0.57	ns	1.50	1.38	ns	1.67	−1.71	ns	1.79
Autocorrelation	rho	−0.01	ns	0.70
	nu	1.53	***	0.12
Residual autocorrelation (MI)	1970	−0.01
	1980	−0.02
	1990	−0.02
	2000	−0.01
	2006-10	−0.01
Model fit	DIC	−3079.34

Note: Models include controls for extractive, FIRE, and professional industry sectors, unemployment, and metropolitan status (coefficients not reported).

^***p < .001,

^**p < .01,

^*p < .05, ns non-significant

We visualize the coefficients derived from the spatial regime analysis in Figure 6. Like the temporal regime results, estimates from the spatial regime analysis are robust against the endogenous effects of serial and spatial autocorrelation. Manufacturing is negatively associated with poverty at some point in all states to varying degrees, with the exception of Minnesota (Figure 6a). Recall that for the region as a whole, higher employment in manufacturing corresponds with lower county poverty rates throughout the period, although to a lesser degree over time. The spatial regime analysis shows a positive poverty-manufacturing relationship in Illinois in 2006-10, suggesting a shift from manufacturing as protective against poverty to manufacturing as poverty promoting. For counties in other states, the coefficients remain negative throughout the study period.

Figure 6.

Of all the states, Illinois has the lowest share of its population employed in manufacturing by 2006-10 (14.7 percent), having experienced the second largest loss in employment over the study period (−9.9 percentage-points, behind Michigan’s −13.6 percentage-points). Indiana and Wisconsin, states with the largest dependence on manufacturing throughout much of the period maintain a negative, though largely statistically non-significant association throughout the period. Minnesota, which has the least dependence on manufacturing during the majority of the period, registered no statistically significant association after accounting for other industries, unemployment, and metropolitan status. We report bivariate associations as a point of comparison, and to highlight that there are surface-level poverty-industry associations to understand that vary temporally and spatially (Table 5).

Table 5.

Pearson’s correlation coefficients for bivariate poverty-industry association, all Upper Midwest counties by state and year

		1970		1980		1990		2000		2006-10
		r		r		r		r		r
Illinois	Manufacturing (%)	−0.68	***	−0.63	***	−0.48	***	−0.31	**	−0.22	*
	Service (%)	0.32	**	0.14	ns	0.12	ns	0.10	ns	0.11	ns
	Agriculture (%)	0.55	***	0.40	***	0.40	***	0.24	*	0.19	ns
Indiana	Manufacturing	−0.34	***	−0.13	ns	−0.07	ns	−0.14	ns	0.01	ns
	Service	0.03	ns	−0.06	ns	0.14	ns	0.36	***	0.25	*
	Agriculture	0.48	***	0.26	*	0.10	ns	0.05	ns	0.01	ns
Iowa	Manufacturing	−0.47	***	−0.49	***	−0.16	ns	0.02	ns	0.09	ns
	Service	−0.07	ns	−0.25	*	−0.09	ns	0.14	ns	0.29	**
	Agriculture	0.55	***	0.57	***	0.24	*	−0.02	ns	−0.29	**
Michigan	Manufacturing	−0.52	***	−0.43	***	−0.33	**	−0.25	*	−0.13	ns
	Service	0.52	***	0.34	**	0.37	***	0.35	**	0.33	**
	Agriculture	0.34	**	0.31	**	0.23	*	0.21	ns	0.16	ns
Minnesota	Manufacturing	−0.61	***	−0.65	***	0.55	***	−0.41	***	−0.22	*
	Service	0.19	ns	0.08	ns	0.23	*	0.35	***	0.36	***
	Agriculture	0.66	***	0.59	***	0.42	***	0.39	***	0.02	ns
Wisconsin	Manufacturing	−0.71	***	−0.69	***	−0.52	***	−0.47	***	−0.29	*
	Service	0.38	**	0.13	ns	0.29	*	0.35	**	0.32	**
	Agriculture	0.55	***	0.48	***	0.27	*	0.35	**	0.15	ns

^***p <.001,

^**p <.01,

^*p <.05, ns non-significant

Taken together, results from the spatio-temporal analysis (Table 4 and Figure 6a) show the hypothesized shift in association – from protective to poverty promoting – for counties within a state that experienced the largest shift in dependence on manufacturing, not continual high dependence. The pattern of spatial variation in the relationship between poverty and manufacturing can be understood in terms of change in industry dependence.

We find similar results for the service sector (Figure 6b). Illinois, again, shows a shift in association over time, moving from protective to poverty promoting. Counties within this state underwent the second largest change in the share of the population employed in services (again, behind Michigan), and ended the period with the second largest proportion employed in the service sector. Net of controls and spatio-temporal endogeneity, no statistically significant poverty-service association is found in any other state, including Michigan and Wisconsin, states with slightly higher dependence in recent years, or in Iowa, a state with higher dependence in earlier years. As with manufacturing, these results suggest it is a change in industry dependence that matters for poverty, not historical or necessarily current dependence.

Agriculture promoted poverty, but in contrast to expectations (Figure 6c). Places with an historically higher share of employment in agriculture have significantly lower poverty rates. Despite suffering the largest losses in agricultural jobs over the period (−13.9 percentage-points), Iowa has the highest dependence on agriculture throughout the study period (see Table 2 and Figure 4) and the most poverty protective association. Iowa’s trend appears to be driving the regional trend – shifting from a positive association in the earlier years to a negative one in the most recent year (Figure 5). Wisconsin has a similar although statistically non-significant pattern, at least in the multivariate analysis. Indeed, no statistical association is found for any other state in any other year in the multivariate analysis.

Taken together, results from the spatial regime analysis confirm significant spatial variation in the relationship between poverty and industry in ways that are both consistent and inconsistent with expectations. Stronger dependence does not always translate into a stronger association, or one characterized as poverty promoting. For manufacturing and services, large shifts in dependence are associated with shifts from protective to poverty promoting. For agriculture, in contrast, historical dependence is associated with shifts, although from poverty promoting to protective. Clearly, industry is related to poverty in ways that are distinct across the Upper Midwest. Yet the pattern of spatial variation is more complex than our initial expectation that dependence on an industry generates a more positive and, thus, detrimental poverty-industry association.

Discussion

Despite widespread agreement that poverty is spatially clustered, scholars do not fully understand why. A central tenet is that the pattern of poverty can be explained by industry. Although it is well-documented that both poverty and industry are spatially clustered and slow changing, research to-date has not formally tested whether industry patterns and shifts underlie the spatial clustering of poverty over time. We set out to address this fundamental limitation of poverty research and, indeed, find evidence that the economic vitality of places is intertwined with the rise and fall of industries.

We use a novel analytical strategy to advance the understanding of the links between poverty and industry. By adopting a regime approach, we directly analyze hypothesized variation in the poverty-industry relationship over time and across space, conceptualized in terms of industry protectiveness (against poverty) and industry dependence. Moreover, we simultaneously incorporate the serial and spatial autocorrelation structures to yield coefficients robust to endogenous effects that are typical of spatial panel data.

Our results suggest that industry underlies changes in the spatial clustering of poverty. Consistent with our first hypothesis, the nature of the association changes over time and in a way that is concordant with industry trends with some important caveats. As an industry loses economic vitality, it loses its protective power against poverty and, instead, becomes a poverty-promoting force; places dependent on distressed sectors become distressed places. Our data suggest a dynamic poverty-industry association that is sometimes consistent with our industry protectiveness hypothesis. Manufacturing is most protective against poverty in 1970, and has a weaker protective influence in later years. High rates of employment in the service sector share a similar pattern, but are only suggestive given the lack of statistical significance. Agriculture, in contrast to our hypothesis, grows increasingly protective against poverty. These dynamic temporal patterns are further illuminated in our analysis that tests for systematic spatial differences in the poverty-industry relationship.

We find mixed support for our industry dependence hypothesis. When examined separately by state, we find evidence that the spatial pattern of county poverty is informed by forces that include and extend beyond industry dependence. For manufacturing, counties within Illinois show the strongest pattern; manufacturing employment shifts from protective in 1970 to poverty promoting in 2006-10. Illinois has an historical dependence on manufacturing, but so too, and to an even greater degree, does Indiana, Michigan, and Wisconsin. Statistically weaker and less pronounced associations are found in these three states, suggesting that the industry dependence hypothesis needs refining. Indiana is set apart from other states in the magnitude of change in the share of its population employed in manufacturing, perhaps indicating that it is change in dependence as opposed to level of dependence that matters for the poverty-industry association. Yet the findings for the service sector pose a challenge to this possibility. Only for counties in Illinois does service employment shift from protective in 1970 to poverty promoting in 2006-10. However, the change in service employment among these counties is on par with those in Indiana and Wisconsin, and lower than in Michigan, and the patterns in these counties are unlike those observed in Illinois.

To further complicate, and possibility elucidate, the industry dependence hypothesis as originally stated, Iowa has the strongest dependence on agriculture and the strongest poverty-agriculture association, yet the pattern is not as expected: the agriculture sector grew more protective against poverty during the study period in this state; counties more dependent on agriculture were increasingly more likely to be located in clusters of lower poverty. Although employment in this sector has declined over time, and most dramatically for Iowa, perhaps the recent recessionary boom in crop prices, and the on-going income subsidies and other social welfare programs benefiting those within the industry are having an aggregate impact (Oppedahl 2015; e.g., Goodwin and Smith 2013). With this in mind, perhaps the spatial differences in the poverty-industry relationship for manufacturing and service might be due to state differences in deunionization and worker protection during the study period (e.g., McCall 2001).

Our analysis represents a critical first step in simultaneously incorporating the temporal and spatial dynamics of the poverty-industry relationship, and provides a much needed foundation for future studies to interrogate other institutional forces affecting industries and, in turn, poverty, including state legislation on worker protection and sector-specific income subsidies. Our study also provides empirical evidence for simultaneous spatial and temporal effects typical of spatial panel data, and offers a conceptually straightforward analytical strategy applicable to myriad studies using similarly structured data. We show that counties tend to neighbor counties with similar poverty rates, and counties tend to report a similar level of poverty over time. Findings are consistent with previous observations about the temporal persistence of poverty and the spatial concentration of poverty (e.g., Jargowsky, 1997; Glasmeier, 2006; Lichter and Johnson, 2007).

A significant difference between our approach and those taken in previous studies is our ability to simultaneously parameterize spatial and temporal effects. The result is more precise and reliable coefficients that characterize the spatially and temporally dynamic poverty-industry relationship as well as a more comprehensive understanding of the spatial and temporal dimensions of poverty exogenous to industry trends. There are considerable computing costs associated with more complex models such as these, mainly time, that must be resolved. Our approach, using R-INLA, offers faster computing times relative to alternative strategies including Markov Chain Monte Carlo (MCMC) and MLE for spatio-temporal models. Indeed, our models have successfully run on a simple laptop, a Windows 10 Dell with four cores (Intel Core i5) and 8 GB RAM, taking 1.5 minutes for equation (1) and less than four minutes for equation (2). Thus, our approach shows promise for future work investigating larger geographic regions, longer time periods, and additional covariates.

We focus on the Upper Midwest, yet our strategy can be applied to the entire United States or generalized to any other geographic context. It also can be extended to test theoretically-based assertions of racialized poverty-industry associations that are rooted in longstanding and well documented racialized patterns of settlement and employment across industries within the United States (e.g., McCall 2001). Future research also could investigate the potential influence of unionization and other institutional forces, as already discussed, and examine possible differences in the poverty-industry association by educational levels of the local-area workforce (e.g., Bluestone 1990). Indeed, the spatial pattern of poverty over time is an area still ripe for exploration, and new tools and theories directly addressing the spatial and temporal dynamics of poverty can bring new evidence to this old topic.