The US Economy as a Network: A Comparison across Economic and Environmental Metrics

Abstract

Environmental-economic analysis is an evolving field that seeks to situate the human economy within environmental systems through its consumption of environmental resources and cycling of resources and waste products back into the environment. Environmental accounting has seen increased focus in recent years as national and regional governments look to better track environmental flows to aid in policy development and evaluation. This study outlines a conceptual environmental-economic framework founded on network science principles. An empirical study operationalizes portions of the framework and highlights the need for further research in this area to develop new data sources and analytic methods. We demonstrate a spatial mismatch between the location of water-intensive industries and the natural location of water resources (i.e., lakes, rivers, and precipitation), which climate change is likely to exacerbate. We use eigenvector centrality to measure differences in the US economy according to economic trade flow and five associated environmental flow accounts (land use, water consumption, energy use, mineral metal use, and greenhouse gas production). Population normalization helps to identify low-population counties that play a central role in the environmental-economic system as a function of their natural resources.

1. Introduction

As a society, we face challenges from a variety of directions: health risks due to disease and poor nutrition, infrastructure deterioration due to inadequate funding, and economic inequality, to name just a few. The breadth of these challenges necessitates a systems perspective, one that moves beyond traditional framings of society as a system of monetary transactions between households, firms, and governments. In the neoclassical economic framing, environmental impacts are externalities (i.e., unpriced impacts) solvable by associating a price with them. This framing has come to permeate both economic and engineering analyses. For example, in the transportation engineering field, travel time savings are monetized to provide a numerical benefit measure [1]. However, recent discourses across academic disciplines have challenged this framing and its focus on monetary growth (i.e., GDP) as a unifying goal [2,3,4]. We can contrast the neoclassical economic model with a more holistic framing of the economy as one component of larger societal and earth systems (see Figure 1). The perspective depicted in Figure 1 has parallels with the circular economy principles that motivate much work in ecological economics and industrial ecology. The challenge for the researcher is how to represent this conceptual model via mathematics and data. We propose a network science approach, as outlined below.

Network science is the study of the features and behavior of networks [5]. Many systems can be abstracted to network representations, denoted by nodes and connecting edges. A social network comprises persons connected by friendships. Network science defines measures of node centrality, network resilience, and redundancy. In a social network, a person with many friends would be considered a central node. Their friends then having large friend groups further increases that person’s network centrality. In a transportation network composed of road edges and destination nodes, network resilience is improved by the availability of redundant road links. Such redundancies become critical at times of network disruption—e.g., during a vehicular crash or evacuation event. More broadly, systems interact as symbiotic networks wherein the transportation network might be affected by a disruption in the water or fuel distribution networks—e.g., a burst water or oil pipeline, respectively.

In the same way, the economy can be represented as a network. Input-output (IO) tables produced by federal statistical agencies summarize monetary flows between industries—an industry produces a good or service, which is then used by other sectors, and receives monetary compensation in return. Introducing a spatial dimension, trade between nations, states, or counties is simply a network of economic nodes transferring goods and services along supply chain edges (e.g., roads, pipelines, railways, shipping lanes, and flight paths). One can then measure the strength of economic connections by the magnitude of monetary flows between any two nodes. Implicit in these economic flows is a corresponding set of embodied environmental resource flows. Environmental analysts have recognized this connection through the development of environmentally-extended input-output (EEIO) analysis [6,7]. In simple terms, EEIO starts from economic IO tables and a set of industry-level resource production and consumption vectors—e.g., water consumption, greenhouse gas (GHG) emissions production, and land consumption. The power of this approach arises from the standard Leontief inverse, which provides a mechanism for quantifying the additional resource requirements across the supply chain for an incremental change in industry production [8]. For example, a bank might consume electricity in its operations from which one can associate GHG emissions based on the electricity generation mix in its region. Returning to the network analogy, economic flows can be translated into equivalent environmental flows following the EEIO framework. This translation will shift the centrality of a node depending on its environmental resource requirements per dollar of production.

In this paper, we first provide a conceptual environmental-economic network framework. The following section describes the network science metrics used in our empirical case study. We then detail the datasets and their processing into a network structure. Empirical results and discussion are provided for a multi-layer network representation of the United States economy. We conclude with future research directions based on our empirical case study.

2. Conceptual Framework

2.1. Signed Multi-Layered Network

Complex systems can be thought to encompass multiple entities (e.g., economic sectors, social networks, infrastructure systems) interacting as multi-layered networks [9]. Figure 2 illustrates the proposed model based on this concept. Beginning with the center column, a multi-regional economy can be depicted as a multi-layer network wherein each layer represents a sector and each node represents a region. Intra-layer edges represent economic trade flows within a sector between regions. Inter-layer edges represent inter-sector trade flows within a region and between regions.

The shift in perspective proposed herein away from a neoclassical framing becomes apparent with discussion of the left- and right-hand columns. These columns show example environmental networks that parallel the economic network. Critically, however, they situate the economy within a larger environmental system through the addition of an environmental layer in each network. The water network shows flows between economic sectors, with nodes given negative signs to represent the fact they are water consumers or sinks. The environmental layer includes water sources by region, with each region having a finite water stock. In theory, one could introduce additional complexity to the model by representing water flows between regions (e.g., stream flows and precipitation); however, this hydrologic modeling would be a significant research task in its own right. While water is principally an input to the economic production layer, $\mathrm{C}{\mathrm{O}}_2$ flows are outputs resulting from economic activity. Again, the environmental layers are critical to understanding the model system. The environment provides a carbon sink, which is a function of available resource stocks. The network abstraction provides a means to represent environmental-economic interactions without the need to monetize environmental flows—i.e., they are internalized through their connecting edges to the economic network. Environmental changes can be mapped onto economic impacts through network inter-dependencies; an economic activity in Region 1 that removes forest cover will reduce the carbon sink, thus impacting the carbon balance for that region. The balance may be maintained through a change in energy inputs to less carbon-intensive feedstocks.

Caschili et al. consider an interdependent multi-layer network model for international trade [10]. Rather than physical flow layers, their network model distinguishes between economic, socio-cultural, and physical infrastructure layers. They forecast bilateral trade, GDP, and migration through time to understand the effects of system shocks. They find that a 2% decrease in US GDP shows an interdependence by Germany, France, and Japan, while China takes advantage of the economic crisis to increase its market share. Alves et al. provide another international trade example, wherein they distinguish between nations and economic sectors [11], similar to our approach herein. Gomez et al. apply a similar approach to the US to study shock propagation, using 115 regions (65 major cities and 50 remaining states, or remainders thereof) [12]. Similar approaches can be found across diverse fields, including biology and medicine, where it has been applied to problems of protein interactions, cell networks, gene expression, and neural networks [13].

Sources and sinks can be considered positive and negative network nodes, respectively. This abstraction is similar to the social network research by Kaur and Singh, wherein they define negative social ties between individuals with hostile social relationships [14]. The research challenge in the proposed case is that definitive source and sink data are not consistently available for all environmental metrics. Land and water sources are readily estimable; however, biogenic CO₂ sinks are uncertain. Some limited estimates are available from the United States Geological Survey (USGS) and other sources for natural resource sources and sinks. Similarly, while fossil fuel resources are known to be finite, total resources are unknown.

Our approach closely aligns with two studies in China using multi-regional input-output analysis and network science metrics. Zhang et al. focus on the embodied CO₂ in inter-city trade [15]. They find that the electricity and hot water production and supply sectors in Beijing have a strong influence on embodied CO₂ transfer through their demand across many sectors. Ninety percent of embodied CO₂ transfer occurs within cities, but the major inter-city transfers are from resource-rich cities. They use eigenvector centrality as their main network science metric. Zhang et al. take a similar approach to assess energy flows between seven regions in China [16]. They find the primary energy flow directions are north-to-south and east-to-west. They use center-of-gravity network science metrics to compare results for direct and indirect energy flows. Chen and Chen provide a global counterpart to this research stream [17]. They find that 70% of global direct energy inputs are for resource, heavy manufacturing, and transportation sectors. However, these sectors represent only 30% of the embodied energy to satisfy final demand (with the remainder of their energy going toward satisfying intermediate production). A main implication of this finding is that the service-based economies of North America and the European Union have high embodied energy associated with their production, whereas Russia is a net exporter of energy through its focus on resource products. Zhang et al. support this finding of high embodied emissions for service sectors in a second global MRIO study [18].

2.2. Ecological Economics Parallels

Within the economics field, one can distinguish between environmental and ecological economic perspectives on how natural resources should be incorporated into valuations. Environmental economics focuses on allocating resources based on notions of utility maximization and ecosystem service monetization. Ecological economics begins from the premise, as outlined by Georgescu-Roegen (Interestingly, Samuelson (referenced above) was a close friend of Georgescu-Roegen and thought highly of his work [19].), that thermodynamic laws impose limits on the economy and that it comprises a subsystem of a larger global ecosystem [20]. These two perspectives are often associated with the terms “weak sustainability” and “strong sustainability”, respectively. To a certain extent, the difference stems from the treatment of natural resources (capital). That is, whether it is assumed that natural capital may be reduced so long as human and manufactured capital replace it in sufficient quantity (weak sustainability) or natural capital is non-substitutable (strong sustainability). The reality is somewhere between these two extremes. Different forms of natural capital are substitutable—i.e., natural gas can replace coal as the combustion fuel in electricity generation. However, natural capital is principally a complement to manufactured capital—i.e., fertile soil, global climate regulation, and the storage and recycling of nutrients have no manufactured substitutes [21]. As noted by Costanza and Daly, if manufactured and natural capital were always substitutable, there would be no need for manufacturing, as an equivalent service would already exist from nature [22].

Much research on the above-mentioned subjects has taken the form of intellectual discourse rather than empirical study. Industrial ecology provides a parallel (and motivating) empirical basis but largely does not engage with economic theory. Common and Perrings use a theoretical model to analytically illustrate that the environmental-economic system is not completely observable and/or controllable by price signals alone [23]. In this paper, we contribute to the literature by exploring the use of network science as a system-agnostic framework for capturing interdependency. By system-agnostic we mean that the framework does not rely on a field(system)-specific theory about component interactions. Environmental economics might translate all environmental systems into equivalent monetary units. Transportation engineering might consider economic and resource flows in terms of vehicular flow and congestion. Climate science might translate economic and resource flows into atmospheric fluxes. Our network science framework is based on the underlying structural relationships between these sub-systems.

3. Materials and Methods

3.1. Methods of Analysis

The above conceptual framework provides a direction for research, from which we focus on an initial empirical study in this paper. We first describe two network science metrics that were found to define the environmental-economic system well. We then outline the data processing steps to combine economic flow and environmental accounts into a multi-layer network dataset.

3.1.1. Center of Gravity

A useful first step to understanding a spatial network is to calculate the center of gravity (COG). We can define the COG as the point that minimizes the sum of the weighted distances between all points in a network. In Equation (1), $d_{ij}$ is the geographic distance between points $i$ and $j$ in the network and $W_i$ is a weight for node $i$ (or node $j$). This approach facilitates a visual comparison of how the COG varies as a function of the chosen weight variable. The simplest application is to use the straight-line distance between spatial unit (e.g., county) centroids, giving the geographic center of the network. Note that this COG will be influenced by the spatial distribution of centroids—counties tend to be smaller in the eastern US than the west, producing an eastward bias in the geographic COG. Geographic distance can be adjusted by introducing a weight, $W_i$, variable that represents a size term for each node. For example, larger area counties will have larger populations than smaller area counties (holding population density fixed). We compare the origin (production) and destination (consumption) economic-weighted COG to the unweighted COG. We then compare these economic-weighted COG results against analogous environmental-weighted COG results:

$$COG={\displaystyle \frac{{\sum }_i{\sum }_j{d}_{ij}{W}_i}{{\sum }_i{W}_i}}$$

(1)

3.1.2. Node Centrality

Large networks require metrics that characterize their features. The most common metric is node centrality, which measures node importance within the overall network structure. The simplest centrality metric is node degree, which is simply a count of the edges connected to a given node (see Figure 3). In a social network, this metric might represent the number of friendships for each person in a given population. Taking the process a step further, a node may be important because it is connected to other nodes that are themselves important. For example, a person may have only one friend but be important because their friend is the president of the country. Eigenvector centrality captures this effect by assigning a centrality score that is a function of the centrality score of adjacency nodes, as defined in Equation (2) [5]:

$$\mathit{A}\mathit{x}=\lambda \mathit{x}$$

(2)

where the adjacency matrix $\mathit{A}$ contains binary indicators of whether two nodes are connected by an edge. Additionally, we can define $\mathit{x}$ as an eigenvector of the adjacency matrix and $\lambda$ as the largest eigenvalue. By the Perron-Frobenius theorem, $\mathit{x}$ must be the leading eigenvector to ensure non-negative centrality measures [5]. Further, the economic system can be described as a directed graph, meaning that edges have a flow direction (i.e., direction of monetary exchange).

We apply the eigenvector centrality measure to a weighted graph representing the US economy (in monetary and various environmental resource units). A highly connected graph will exhibit similar centralities for all nodes, absent the inclusion of edge weights that capture the magnitude of connections. For example, a strong connection to Los Angeles County should make a county more central than an otherwise similar county with a weaker connection to this economically dominant county. A review by An et al. finds that eigenvector centrality is among the most common network science metrics used in the input-output literature to understand the role of sectors or regions in physical flow networks [24].

We are interested in whether a county becomes more (or less) central when moving between environmental units. Defining an appropriate comparison metric can be a challenge. Eigenvector centrality values are not directly comparable between accounts. A simple ranking is a first approximation but does not account for the non-uniform differences between centrality levels—i.e., the first and second ranked counties may be similar in their centrality while there is a large difference between the centrality of the second and third ranked counties. To address this issue, we use a normalization of the centrality values. Specifically, each centrality value is differenced against the lowest value in its corresponding category and divided by the difference between the maximum and minimum centrality values:

$${\displaystyle \frac{{E{C}_i-EC}_{min}}{{E{C}_{max}-EC}_{min}}}\forall county$$

(3)

where $E{C}_i$ is the eigenvector centrality value for county $i$. This normalization means that the most central county is given a value of 1.0 and the least centrality a value of 0.0. Comparing across metrics, differences relative to the normalized trade flow centrality has an intuitive interpretation. Consider a county with (normalized) trade flow and water centrality values of 0.5 and 0.65, respectively. The difference will be 0.65 − 0.5 = 0.15, suggesting that the county is more central in the water network by about 15 percentage points. In the same way, a negative difference means that the county is less central in a given environmental network. We can also compare destination against origin trade flow centrality in a similar manner.

3.2. Data

Our main datasets are economic input-output data purchased from IMPLAN Group LLC. IMPLAN is the most widely used source of regionalized economic input-output data for the United States. These data, sourced from federal government agencies with additional processing by IMPLAN, are as follows (in 2019 dollars):

• Commodity production totals 546 classification codes and county
• Industry output totals 546 classification codes and county
• Proportional industry output by commodity for US average
• Domestic commodity trade flow data by origin and destination county

The Bureau of Economic Analysis (BEA) Benchmark IO tables are the primary data source used to construct IMPLAN data [25]. Domestic trade flows are estimated by IMPLAN using a doubly constrained gravity model, first developed in collaboration with the US Forest Service in 2005 [26]. The gravity model has been a standard trade flow prediction approach for over 50 years, since its original application by Leontief and Strout [27]. Inputs to the trade flow model include transportation cost estimates by Oak Ridge National Laboratory and Census Bureau Commodity Flow Survey data [26].

In addition to economic data from IMPLAN, we use the latest 389 industry environmental impact coefficients contained in the US Environmentally-Extended Input-Output (USEEIO) model [28]. The USEEIO model provides coefficients for land, water, energy, mineral, and metal use, air pollution, nutrients, and toxins. Environmental flow accounts for the USEEIO model are sourced from inventory, survey, and reporting data collected by the US Geological Survey (USGS), US Environmental Protection Agency (USEPA), Bureau of Land Management (BLM), Energy Information Administration (EIA), and US Department of Agriculture (USDA). The USEEIO 2.0 model is validated through its reproduction of national control totals and comparison against the prior model version.

Merging these datasets into a consistent network representation requires a series of data transformations according to Algorithm 1. Further, there are several assumptions necessary to reconcile differences between datasets. The IMPLAN data are available for 546 industries, while the USEEIO model uses only 389 industries. Lenzen argues that the disaggregation of economic accounts is preferable to the aggregation of environmental accounts [29]. In this case, we have the opposite pattern and choose to aggregate the economic sectors because the environmental accounts represent sectors identified as dissimilar in their environmental multipliers (i.e., impacts per dollar of production) by the USEPA. We build a correspondence between these industry classification systems and assume the same environmental impacts for IMPLAN industries with the same USEEIO industry code. It is also assumed that environmental multipliers are constant across the US.

Algorithm 1 Commodity-to-industry trade flow translation algorithm

For each county: Distribute commodity totals among industries according to the US-level industry-commodity proportional output table Perform iterative proportional fitting (IPF) against county-level industry output and commodity production totals. The IPF algorithm, or RAS method, is widely used to balance matrices against known control totals, including by IMPLAN Group LLC in their data construction process [30] Append county-level results to industry trade flow table. Result: industry trade flow

The overall data preparation procedure is illustrated in Figure 4.

The resulting industry trade flow data are combined with USEEIO environmental impact coefficients. We aggregate the trade flow data for all industries, producing a dataset representing county-level origin-destination trade flows by economic and environmental impact total. This aggregation was necessary to reduce the computing load when performing the network science analysis. Economic trade flows alone represent a table with over 400 million rows after removing entries with less than $100 of output. Each environmental impact will require similar computer storage. Further, this aggregation simplifies the interpretation of county variation rather than needing to consider the additional dimension of industry variation. The final dataset is validated by ensuring all marginal totals are within 1% of the input control totals, similar to the approach taken in the construction of the USEEIO [28].

4. Results

4.1. Center of Gravity

Figure 5 provides a first approximation of the role of COG weight definition in a network. The unweighted (geographic) COG for the lower 48 states, measured as the point that minimizes the distance to each county centroid, is in Missouri. By weighing each county by its economic input (i.e., destination center) or output (i.e., origin center), the COG shifts west into Kansas. This geographic shift is a result of western counties tending to have larger populations (due to their larger geographic areas), and therefore larger economic inputs and outputs. Further, we see a shift south due to Texas having a relatively large population and economy relative to northern plains states (North Dakota, South Dakota, Nebraska, Iowa, and Minnesota) and mountain states (Montana and Wyoming).

Figure 6 extends the COG notion to environmental flow metrics. Translating trade flows into GHG emissions shifts the COG southeast, likely as a function of the more GHG-intensive industries and electricity generation composition in the south and the east. The largest shift is to the west for the water consumption COG. Water consumption at the origin represents industries that require water resources as inputs—primarily the agricultural sector. Similarly, agriculture requires land, as illustrated by the westward shift of land use at the origin. Taken together, these results provide an interesting picture of agricultural requirements. Southwest agriculture in California has a significant effect on the water-origin COG. Northwest agriculture benefits from a comparatively higher level of precipitation, but the region shifts the land use COG northwards, likely due to its forest product sector.

4.2. Eigenvector Centrality

Figure 7 plots eigenvector centrality by county for four flow categories (economic, water consumption, land use, and GHG emissions). While it is difficult to discern clear differences across the four plots at the national scale, there is a clear pattern across the four categories of low centrality by the Central Plains states. The northeast tends to produce land-intensive commodities, while the southwest produces water-intensive products.

Figure 8 focuses on trade and water flows in one state, Nebraska. The eastern counties containing the two major cities of Lincoln and Omaha tend to be central to trade flows. The southwest counties become more central when considering water, due to their stronger focus on agricultural production and processing.

Mapping centrality measures is a useful exercise for understanding high-level spatial distributions. However, quantitative analysis is necessary to draw out county-level insights. We accomplish this purpose by focusing on changes in eigenvector centrality when translating from trade flow to environmental units. As outlined above, we use a normalization by the maximum eigenvector centrality difference for a given metric, producing a ratio in the range [0, 1], and compare it against the analogous ratio for origin (destination) trade flow. This approach allows us to compare continuous measures between economic and environmental metrics.

Table 1. Top/bottom changes in eigenvector centrality (relative to trade flow centrality).

Metric	Direction	County	State	Change
Energy flow	O	Harris	TX	0.98
Energy flow	D	Harris	TX	0.99
Water flow	O	Maricopa	AZ	0.51
Mineral metals flow	O	Harris	TX	0.98
Mineral metals flow	O	Los Angeles	CA	−0.94
Mineral metals flow	D	Harris	TX	0.99
Mineral metals flow	D	Los Angeles	CA	−1.00

Notes: O—Origin county and D—Destination county.

provides a summary of the differences greater than 0.5. Starting with energy flows, Harris County, TX (Houston) becomes the most central county, surpassing Los Angeles for both origin and destination flows. This result highlights the central role of the county in the energy economy. As identified in the COG analysis, western states have an outsize role in water consumption. This finding is supported by the increased centrality of Maricopa County, AZ (Phoenix) when considered as a water consumption origin. In contrast, Los Angeles County, CA becomes considerably less central when viewed through the lens of mineral metal consumption and production.

Recognizing that counties are not uniform in their populations provides another perspective, wherein we normalize the eigenvector centrality by county population (

Table 2. Top/bottom changes in population-normalized eigenvector centrality (relative to trade flow centrality).

Metric	Direction	County	State	Change
Energy flow	O	Jefferson	TX	0.95
Energy flow	O	Galveston	TX	0.88
Energy flow	O	Los Angeles	CA	−1.00
Energy flow	D	Matagorda	TX	0.99
Energy flow	D	Los Angeles	CA	−1.00
Water flow	O	Greenlee	AZ	0.80
Water flow	O	Los Angeles	CA	−0.95
Water flow	D	La Paz	AZ	0.95
Water flow	D	Imperial	CA	0.55
Water flow	D	Hayes	NE	0.50
Water flow	D	Los Angeles	AC	−0.95
GHG flow	O	Greenlee	AZ	0.80
GHG flow	D	Midland	TX	0.55
Land flow	O	Dallas	AR	0.98
Land flow	O	Sabine	TX	0.93
Land flow	O	Polk	TX	0.70
Land flow	O	Columbia	AR	0.65
Land flow	O	Pike	AR	0.62
Land flow	D	Hayes	NE	0.93
Land flow	D	Perkins	NE	0.88
Land flow	D	Harding	SD	0.84
Land flow	D	Chase	NE	0.69
Land flow	D	Faulk	SD	0.65
Land flow	D	Los Angeles	CA	−0.94
Mineral metals flow	O	Cameron	LS	0.76
Mineral metals flow	O	Los Angeles	CA	−1.00
Mineral metals flow	D	Livingston	KT	0.98
Mineral metals flow	D	Los Angeles	CA	−1.00

Notes: O—Origin county and D—Destination county.

). With this normalization, we see a significant increase in the centrality of Texas counties when measured in terms of energy rather than economic trade flows. Further, Los Angeles drops to nearly the bottom of the list as a function of the normalization of its large population and smaller role in the energy economy relative to other sectors. Similarly, Arizona counties rise to the top when considering shifts in centrality measured by water consumption. These counties require significant water inputs in order to function, lacking natural supply via precipitation. GHG flows are also relatively high in Arizona and Texas counties, likely as a function of the air conditioning load and its associated electricity requirements.

Focusing on low-population counties in the results highlights the power of centrality analysis. Greenlee County, AZ appears in our list for both water consumption and GHG production, exhibiting a significant move up the rankings; as Arizona’s least populous county [31], it begs the question of how this county becomes central based on these environmental metrics. The answer lies in Greenlee County being home to the Morenci Mine, the largest copper mine in North America [32]. This result demonstrates how a single large industrial facility can shift the centrality of a county when framed in terms of environmental flows. Further, it highlights the spatial mismatch between locations of environmental impacts and final goods consumption. The copper produced in Greenlee County is consumed across the nation, predominantly in larger population centers.

Two interesting cases for mineral metal flow are Cameron Parish, LS as an origin, and Livingston County, KT as a destination. Both counties have low populations (less than 10,000 inhabitants, [33]). “Stone mining and quarrying” is the largest sectoral source of mineral metals. However, the results are initially counterintuitive. Cameron Parish has an economy dominated by petrochemical manufacturing, a sector that consumes no mineral metals according to the EPA USEEIO accounts. Livingston County is a major stone mining source but is considerably less central as an origin for mineral metal flows. Examining these results in greater detail, Cameron Parish has five times the output of Livingston County, despite having a smaller population. In addition, Cameron is more strongly connected to the major mineral metal origins of Harris County, TX and Miami-Dade County, FL. The large mineral metal flows for Harris and Miami-Dade counties highlight another nuance in environmental-economic analysis. These are urban counties that do not contain major mines; rather, their large flows stem from their economic role in the natural resource sector rather than the physical flow of natural resources from these counties.

5. Discussion

Our analysis provides a US counterpart to the work of Zhang et al. [15] and Zhang et al. [16] in China. We extend their work by increasing the number of economic sectors and physical flow accounts considered in the analysis. Similar to Caschili et al. [10] and Zhang et al. [15], we use eigenvector centrality as our network science metric. Applying centrality metrics to a global MRIO network, Cerina et al. find that industries are highly but asymmetrically connected at a global scale [34]. They find that most production still operates at the national scale. Our more spatially disaggregated county-level data suggests that the US economy is highly interconnected and does not exhibit the same local production network structure. This result is partially a function of the lower trade barriers within a single country. Network centrality varies as a function of whether one considers physical flows from a production or consumption perspective. As noted by Lenzen et al., GHG (and environmental impact) responsibility is strongly influenced by the reliance of a nation (or county) on interregional trade [35]—e.g., Midland, TX as a destination for GHG flows due to its role in energy production.

We use center of gravity (COG) to understand how the unit of analysis shifts the network structure. We confirm the results of Chen and Chen for international energy flows [17], showing that they hold true within the US. Resource-rich regions of the country tend to come to prominence when considered through a physical flow lens, with high-income service sector-focused cities having high embodied demands for resources. Prell et al. describe this process as “telecoupling” in reference to land consumption potentially being driven by spatially distant economic consumption [36]. We see this pattern in water consumption by Western states being driven by embodied water demand from Eastern states associated with food consumption. Sheng et al. use an MRIO approach based on IMPLAN county-level data for the 10-state Southeast US region [37]. They find that agriculture and energy production have the largest water multipliers (i.e., water relative to economic centrality), though their analysis aggregates the IMPLAN data into 21 economic sectors. They confirm our finding that services, government, and manufacturing sectors indicative of urban regions are major sources of final demand for water and other physical flows. Garcia et al. also provide a water application for the US using a 41-sector and 115-region multi-layer MRIO network [38]. They identify five node clusters (communities) in the network.

6. Conclusions

In this paper, we present a conceptual framework for representing the environmental-economic system as a multi-layered network. This abstraction has the appeal of representing environmental interactions with the human economy without the need for explicit monetization, as traditionally required by the environmental economics framework. Rather, the economy is one component of a larger environmental system. While conceptually attractive, there are both computational and data challenges associated with operationalizing the framework. First, computational challenges arise from the combinatorial explosion of county × industry × environmental metric combinations. The analysis could be simplified by using a different spatial unit—e.g., major cities and 50 state remainders [12]. Recent years have seen advancements in the availability of environmental flow data, but there remains a paucity of environmental stock data. County-level environmental stock data would be necessary to fully operationalize the conceptual framework, with node signs capturing the source/sink role of industries and the environment.

Our empirical study takes several first steps toward the above conceptual model. We construct a county-level environmental-economic multi-layered network that includes five environmental flow accounts, in addition to economic trade flows. Our use of multiple indicators allows us to improve upon past work in the area by drawing comparisons between the relative roles of regions when examined through different environmental lenses, as well as in comparison to their economic roles. We reconcile industry-specific environmental accounts with commodity trade flow data by leveraging data from an IMPLAN and an IPF approach. We apply network science metrics to analyze the network structure. A limitation of this study is its reliance on proprietary MRIO data and methods by IMPLAN Group LLC. The USEPA has recently developed an open-source alternative for the 50 US states, which decomposes the economy into state-of-interest and rest-of-the-US [39]. The model is integrated with the USEEIO model system, facilitating a simple application for environmental analysis.

Network metrics highlight the shifts in centrality between regions of the US when translating trade flows into various environmental metrics. The COG shifts to the west when considered from a water consumption perspective, as a function of the spatial mismatch between water-intensive economic activities and natural water sources. Texas is central to the energy network, while Los Angeles County becomes less central from this perspective. These mismatches are likely to be exacerbated by climate change, with more extreme precipitation and drought events [40].

Population-normalization of centrality metrics provides additional insights into the role played by low-population resource-extracting counties. In many cases, these counties become among the most central in environmental flow networks. However, results are strongly influenced by the relative connection to high centrality urban counties. These results highlight the need for further work disentangling the spatial locations of environmental flows and economic value. For example, Miami-Dade County, FL and Cook County, IL are listed as among the largest origins for “stone”, despite them not having any major operating mines. Such spatial questions become pertinent when environmentally extended IO analysis shifts from an aspatial perspective where the focus is marginal changes in environmental impact. Such a shift to explicitly spatial IO analysis is necessary to situate the human economy within the environmental system in order to understand the economic-resource mismatch.