Holistic fairness considerations in facility placement decisions Authors: James Crabb, Stephanie L Kane, Assefaw H Gebremedhin Published: 2026 ABSTRACT AbstractNetwork science metrics such as centrality play a crucial role in solving facility placement problems, which in turn have important applications in guiding decision processes for municipalities or regions. Decision makers and planners are increasingly interested in addressing situations where demographics and socioeconomic status are distributed in a non-uniform way over a region. In those cases, optimizing centrality alone may result in disparate access. In this paper, as an illustrative case study, we use cluster analysis and publicly available demographic data to characterize census tracts in Spokane County, WA, USA across many socioeconomic variables. We then compare placement of existing facilities in Spokane using a novel fairness metric and network centrality, first separately and then in combination, to determine optimal facility placement. Our analysis found that undesirable facilities are not fairly placed, but some key desirable facilities such as hospitals and parks are ideally located based on both fairness and centrality. Our approach can easily be applied to any region where similar geographic and demographic data are available, and analyses are performed to inform facility siting, upgrades, or relocation.Subject terms: Mathematics and computing, Geography, Complex networks, Economics, Social policy Introduction City and regional planners often need to make decisions on resource allocation and placement of public facilities and services, which can have far-reaching consequences. In the case of services such as fire stations or emergency medical services, placement decisions may make a difference in whether help arrives in time or not. In other cases, facilities could house resources, services, or recreational opportunities for the public, such as libraries, schools, or parks. These types of facilities are all desirable, but the criteria for siting may vary and may be complicated by rapidly changing demographics in smaller and mid-size cities and exurban areas1. On the other hand, undesirable facilities are those that, while necessary, are potentially hazardous or noxious, such as wastewater treatment plants, landfills, and industrial facilities such as factories or power plants. In some cases, siting decisions for undesirable facilities are quite complicated, and numerous methods have been proposed for optimization2. Classical considerations of facility placement tend to focus on distance as the main factor to be optimized, although specific facilities will often have their own requirements, such as a nuclear power plant needing to be near a large source of water. In the context of facility placement, much of the research to date that has centered on fair placement of desirable facilities has defined fairness to mean that all individuals or households in a region of interest have approximately equal access to services, an objective often referred to as “accessibility.” Accessibility is generally measured in the form of distance or time, using a variety of methods to minimize either the maximum or total travel distance/time, and may involve simultaneously optimizing with respect to facility load3–8. Fairness in accessibility may also be described by hospital admit rate, while also considering hospital capacity and income constraints9. Because many of the computational model problems arising in these circumstances are NP-hard, researchers have developed approximation algorithms to solve fair access to services problems10. These algorithms may be graph (or network) based or not. Figure 1 shows how our work compares with other works referenced in this introduction in terms of modeling distance and consideration of socioeconomic variables. Methods used in this work and previous work. Numbers indicate references used in this manuscript. Minimizing travel distance or time among households is an important measure of accessibility, but this does not consider heterogeneity among households within the service area. In fact, demographic and economic characteristics of households are not randomly distributed in a city or region. In this context, the definition of fairness could be expanded to mean that not only are facilities sited to minimize travel distances to the service for all users, but that proximity to facilities is not correlated with socioeconomic or demographic factors. In other words, those neighborhoods that are farthest from the facilities are not disproportionately members of a particular demographic, even if travel distance or time is minimized for the overall population. This distinction was succinctly described by Chang et al.31 as a shift from examining “spatial efficiency” to “spatial equity”. Perhaps more importantly, facilities should meet the changing needs of the community they serve. Over time, neighborhoods can experience shifts in demographic makeup, requiring new facilities or amenities1. Fairness, particularly with respect to healthcare and emergency services, is especially critical as extreme environmental conditions become more common due to climate change. Natural disasters such as floods, storms, fires, or extreme heat often disproportionately impact low-income households or underrepresented populations11, as well as their access to health care or social services12,13. New diseases may also have spatially disparate impacts. During the COVID-19 pandemic, vaccination rates were significantly lower in low-income and majority-Black neighborhoods that continue to see impacts of historic redlining practices14. Historically marginalized groups are often disproportionately impacted by noxious, polluting, or otherwise dangerous facilities; mitigation of this practice is referred to as environmental justice15. A common method for aiding in facility placement is to model the facility’s intended service area as a network with nodes representing locations and edges representing connections between locations, such as roads. Then, network science can be applied to that network and inform facility placement. When optimizing distances in a network, centrality is the most common metric. There are different methods for evaluating centrality, two of which are eccentricity and closeness16. Eccentricity has often been considered in the case of placing emergency facilities, in which the ability to quickly deliver services (e.g., fire stations) is of high importance. Eccentricity is a min-max optimization criteria, where the distance from the facility to all points in the service area is calculated, and the facility is placed to minimize the maximum distance or response time. Closeness is useful as a measure of centrality in locating service facilities where it is desirable to make it easy for the target population to reach the facility, such as in the case of recreational facilities. Closeness minimizes the combined total travel distance of everyone in the service area. To consider fairness holistically, taking into consideration demographic and socioeconomic heterogeneity, one must add an additional constraint to the objective function. Specifically, the algorithm should account for geographic heterogeneity of populations within the service area to ensure that no group experiences disproportionately adverse impacts and the facility type in question meets the needs of the community it serves. Applying fairness constraints to more traditional facility location problems bring up a variety of interesting research questions, such as the closure of a hospital or health care facility, and a variety of fairness measures could be applied depending on the context17. Accounting for fairness in this manner—specifically by ensuring that historically underrepresented or disadvantaged groups are not disproportionately impacted—offers an extension to work in fair facility placement that considers the utilitarian preference of the users4,5,18. The analysis of fairness or other metrics to assess demographic parity in facility or amenity placement has been explored in many studies17,19–24. To date, fairness has generally only been assessed with respect to one characteristic of the population at a time, generally income or age. For example, Horner et al.22 used weighted average transit times (as a measure of accessibility) by age group for several different types of trips, such as shopping, social events, or work, and they found significant variation among age brackets. Other authors have considered multiple characteristics of the population, but often in these studies, each demographic variable is analyzed independently of other variables. Gupta et al.17 examined emergency room placement, comparing a base model optimizing only distance to one that optimized distance, load, and either race, income, or health insurance status, resulting in 18 models that were useful for comparisons. While any consideration of fairness based on socioeconomic information is useful, cities and neighborhoods are complicated. Not all minority-majority neighborhoods are low-income, and not all predominantly white neighborhoods are high-income. Neighborhoods may also differ in the proportion of households with families, retirees, and access to independent transportation, impacting the need for schools, social services, or public transportation options. These characteristics interact with the built environment, resulting in what has been termed the social-spatial structure for a city or region, and for which the framework of space syntax can be applied to study and assess this interdependence24. Previous work incorporating multiple demographic variables in assessing fair access to facilities or amenities has been done using an index (such as the Multidimensional Poverty Index) to classify neighborhoods or municipal units19. This technique has also been applied in related work on assessing community vulnerability to natural disasters11. Another novel approach to this issue was implemented by ref. 23, who used factor analysis to distill thirteen different socioeconomic variables into three statistically independent factors and examined their impact on late-stage cancer diagnosis and treatment facility access. That technique worked well in the studied situation, as the dependent variable was a health outcome, where facility accessibility and socioeconomic variables were the independent variables. However, this technique does not directly optimize both accessibility and demographic parity simultaneously; rather, it assesses their joint impact on health outcomes while considering the complexities of the population served. In this paper, we build upon prior studies that understand communities to be multifaceted, and that people and their environment interact in a dynamic manner as described in the framework of space syntax24,25. We examine the interaction of many different socioeconomic variables to determine fair facility placement using graphs constructed out of geographic information as starting points for centrality analysis. To accomplish this goal, we holistically characterize neighborhoods (with census tract as a proxy) using K-means clustering of socioeconomic variables rather than basing fairness on a single metric (e.g., racial composition or income) or on a single multidimensional index. We propose this novel method as a simple, adaptable, and effective method of applying a fairness metric in analyses that are used to inform facility siting, upgrades, or relocation. To demonstrate this technique, we examined fair facility placement using the county and city of Spokane, Washington, in the United States as a case study. Located on the Columbia Plateau, the County of Spokane’s population increased 34% between the 2000 and 2020 census. This growth rate parallels the overall rate of growth for the Columbia Plateau1 and can be contrasted to the population growth of the United States during the same period, which increased by 21%. Spokane County’s growth was highest for young adults (age 20–34), whose population grew 42%, and retirees (over age 60), whose population experienced 103% growth. These growth rates are considerably higher than the same demographic populations nationwide, which experienced a 19% and 76% increase, respectively, over the same period, according to the U.S. Census Bureau26. Growth continued during the COVID-19 pandemic27, and these shifts in population create unique challenges. With older adults and retirees more than doubling and a very large increase in the number of young working-age adults, Spokane is facing different demands on city services and infrastructure than in previous decades, which in turn raises the question of how facility placement might need to change. Using data obtained from the American Community Survey28 and the City of Spokane, we determine optimal facility placement accounting for not only distance, but also the demographic make-up of neighborhoods. These data are available throughout the U.S. (the frequency of updates varies by population density), making this a useful dataset in many contexts. In addition to using the measures of centrality described above (eccentricity and closeness), we also explore a metric that includes population density19,29,30. In this study, we use Census Bureau tracts to approximate neighborhoods. Each county in the U.S. is uniquely partitioned into tracts, and each tract lies entirely within the boundaries of a single county. Tracts are designed to be relatively stable over time and roughly similar in terms of population numbers, but they may be split or combined if their populations grow or contract above or below thresholds. This information can easily be obtained by city planners in any community, and new information can be incorporated into models as it becomes available. This allows planners to apply our same framework over time to adapt to rapidly changing neighborhood and population structures. Summary of Contributions. In summary, this work makes the following contributions: We analyze facility placement problems using centrality metrics applied on graphs constructed from geographical information. We analyze facility placement problems using centrality metrics applied on graphs constructed from geographical information. We propose a holistic fairness metric for use in facility placement problems based on graph analysis. We propose a holistic fairness metric for use in facility placement problems based on graph analysis. We incorporate a suite of socioeconomic variables via K-means clustering into our evaluation of fairness in contrast to prior work, which has been based solely on distance and single or no demographic categories. We incorporate a suite of socioeconomic variables via K-means clustering into our evaluation of fairness in contrast to prior work, which has been based solely on distance and single or no demographic categories. We apply centrality and fairness metrics to analyze the placement of desirable and undesirable facilities in Spokane, Washington, to demonstrate our metrics, which can be applied to any community across the U.S. We apply centrality and fairness metrics to analyze the placement of desirable and undesirable facilities in Spokane, Washington, to demonstrate our metrics, which can be applied to any community across the U.S. Through our analysis, we find that several desirable facilities in Spokane are optimally placed based on centrality and fairness, but that some desirable facilities and all undesirable facilities are not optimally placed based on these criteria. Through our analysis, we find that several desirable facilities in Spokane are optimally placed based on centrality and fairness, but that some desirable facilities and all undesirable facilities are not optimally placed based on these criteria. Results The results of the fairness metric for siting undesirable facilities are shown in Fig. 2, and the results for desirable facilities are shown in Fig. 3. For undesirable facilities, we use our novel metric, “population-weighted average cluster distance,” or ACD. For desirable facilities, we use our novel metric, “population-weighted average cluster distance across clusters,” or ACDC. K-means clustering of census data for the considered region resulted in four clusters, described in Table S2 in the Supplementary Information. They are Low Income, Low Education (LILE); Low Income, Moderate Education (LIME); High Income, Low Education (HILE); and High Income, High Education (HIHE). Heatmap of population-weighted average cluster distance for undesirable facilities. LILE Low Income, Low Education, LIME Low Income, Moderate Education, HILE High Income, Low Education, HIHE High Income, High Education. For the facility in each row, green values represent the cluster(s) further from the facility (good), and red values represent the cluster(s) closer to the facility (bad). Values of population-weighted average cluster distance across clusters (ACDC) for desirable facilities. Lower values, such as with Fire Station 6, indicate a “more fair” placement of the facility with respect to ACDC. Higher values, such as with Fire Station 8, indicate a “less fair” placement. Undesirable facilities The overall ACD for all nodes is shown in the first column of Fig. 2, followed by the ACD of nodes within each of the four cluster types. We can see that two clusters, LILE and HILE, are closer to all the undesirable facilities than the overall average, and the other two clusters, LIME and HIHE, are farther away than the overall average. This result indicates a disparity in fairness between these two groups of clusters with respect to undesirable facility placement. In both cases, the undesirable facilities tend to be closer to the two clusters with lower overall education levels. Interestingly, placement is less disparate with respect to income. Desirable facilities With desirable facilities, our analysis went more in-depth because we can make recommendations for placing them that will be unique to Spokane. First, we can look at graph centrality alone to see where the best locations would be for desirable facilities. CE (eccentricity centrality) is typically used for placing emergency facilities such as hospitals and fire stations, and CC (closeness centrality) is used for recreational facilities such as shopping malls or parks. The nodes with highest CC and CE (CT 19 and CT 35, respectively) are shown in Fig. 4 as square nodes. City of Spokane graph. Squares have highest centrality (CC: CT 19; CE: CT 35), downward triangles are the five least fair nodes, upward triangles are the five most fair nodes. Nodes outlined in dashed red optimize fairness and CC, nodes outlined in solid green optimize fairness and CE. CT is Census Tract. CC is closeness centrality, and CE is eccentricity centrality. Alternatively, we can consider holistic fairness (as described in the “Methods” section) and identify the nodes with the best fairness. Figure 4 also shows the five most fair nodes (upward triangles) and the five least fair nodes (downward triangles). Finally, we can optimize for both centrality and fairness, and this is also shown in Fig. 4. If we were to recommend desirable facility placement in Spokane city based only on centrality, then emergency facilities should be placed in Census Tract 35, and recreational facilities should be placed in Census Tract 19. However, we also want to consider fairness in our recommendations. ACDC for existing desirable facility placements is shown in Fig. 3, listed in order from most fair at the top (Fire Station 6) to least fair at the bottom (Fire Station 8). Sacred Heart Hospital and Riverfront Park are close to ideally placed for fairness because there are only a few other nodes (four for Sacred Heart Hospital and eight for Riverfront Park, out of a total of 62) representing areas within the city that have better fairness. Recommending desirable facility placement by fairness alone says that facilities should be located within Census Tracts 30, 32, 36.02, 38, or 40.02. When optimizing for both fairness and centrality, Fig. 4 shows the top five nodes when considering CE (solid green outline, recommendations for emergency facility placement) or CC (dashed red outline, recommendations for recreational facility placement). We can see that several of Spokane’s existing desirable facilities are located at nodes in these regions. For example, Sacred Heart Hospital is at CT 32, which is one of the top five nodes for fairness, CE and CC. Additionally, Riverfront Park and three fire stations fall in these regions. Discussion Optimizing for centrality, holistic fairness, or both can lead to different facility placements. In this paper, we compared using raw eccentricity or closeness alone for facility placement versus considering fairness based on distance to tracts of different socioeconomic characteristics. We also looked at current placements of various facilities to rate how fairly they appear to have been placed. To accomplish these goals, we expanded existing research by applying novel approaches. First, within the context of holistic fairness, we used a K-means cluster analysis to characterize census tracts across a suite of demographic variables, rather than classifying tracts with respect to one variable at a time. Secondly, we propose the inclusion of new holistic fairness metrics, the population-weighted Average Cluster Distance (ACD) and ACD across Clusters (ACDC), in objective functions when considering fairness alongside other metrics such as centrality. This strategy differs from more traditional definitions of fairness that do not consider heterogeneity among households or neighborhoods, nor tailor analyses to the specific demand needed of the population based on demographic or socioeconomic characteristics. Third, our network-based approach, especially with respect to our analyses of desirable facilities within city limits, allowed a realistic and computationally fast mechanism of determining the optimal placement of facilities. To address heterogeneity among households, we use population weighting in ACD and ACDC. Census tracts, which we use as the basis for our graph nodes, are intentionally similar in population size but are not necessarily standardized to account for any other data when they are designated by the government. Additionally, census tracts are updated at most every ten years, during which time populations can change drastically. Therefore, we use population weighting according to the demographic categories determined by our K-means analysis in order to evaluate fairness. Without using population-weighted values, it would be unreasonable to make any claims around fairness of facility placement. As a result of our analysis, we found that many of the facilities within the City of Spokane are well-placed for both centrality and fairness. Sacred Heart Hospital, for example, is within our recommended placement region for optimizing fairness and centrality. By considering centrality and holistic fairness separately and then together, we can see that facility placement recommendations changed, although not dramatically in this case. It is worth noting that Sacred Heart Medical Center is located within a large cluster of health care facilities and hospitals, which also offers other synergistic benefits (such as a cluster of specialist health care clinics in a central location). For our graph of the City of Spokane, the fairest nodes also happened to be near the center of the graph, especially when using eccentricity. Our analysis for undesirable facilities was very straight-forward because placement of undesirable facilities will always have the same goals of keeping them as far away from population centers as possible while considering any other real-world requirements. We found the placement of undesirable facilities in Spokane County is less equitable: populations with less education live closer to Superfund sites than more educated populations. The upside is that these undesirable facilities are Superfund sites, which means they are getting federal support for mitigation. For example, we considered a wastewater treatment facility, which is located on a river inside the city limits. Living close to it is undesirable due to odors and the potential for accidental leakage. The fairness of its placement would be increased by moving it out of the City of Spokane, but the cost of this would likely be prohibitive, and there would be additional constraints, such the requirement of being adjacent to a river and having to pump sewage to the facility (which can become very expensive if it is too far from the city). As such, there are limits to new recommendations we can make for placing undesirable facilities, though municipalities might be interested in equitable remediation of polluted areas. If we were interested in placing a facility that is meant to serve not just city residents but also those in the suburban and exurban areas (such as Spokane Valley, in this context), we could extend the graph we used to include additional census tracts. We would expect both fairness and centrality to change due to the increased travel distances to the new nodes, as well as due to different demographics and socioeconomic structure. Future work could also examine facility location in the context of other frameworks, such as increasing access to public transportation options with the intent of minimizing distance traveled by personal vehicles to reduce pollution. The methods presented here can be used for performing similar analyses in other cities or regions in the U.S. In this paper, we used a single clustering algorithm to characterize census tracts into four distinct clusters. In other settings, researchers or planners using this technique will want to carefully consider the suite of variables used within the cluster analysis, as well as different hyperparameter tuning options (such as number of clusters) to fit the context of the facility placement question and the geography of the city or region. For example, if the facilities in question are schools, the variables considered for inclusion may differ slightly from a situation where the facilities are healthcare clinics. Planners may have additional data available they wish to incorporate into the clustering; the technique presented in this paper demonstrates that an equity assessment for facility placement can be conducted using publicly available data at no cost of data collection to the city. Another factor that is important to consider is the service radius of the facility/amenity. This technique could be adapted to use smaller geographic areas, such as census blocks22 or larger areas such as counties, depending on the research question addressed. Additionally, it is worth noting that our analyses only consider holistic fairness and centrality. Other considerations may also be important, such as land costs, zoning regulations, land availability, and other facility-specific requirements. Our technique could be used in combination with additional constraints. In some situations, the fairest location could be impractical or difficult when optimizing for many criteria, but in that case, planners could consider using the second-best location. Other constraints, such as zoning laws, could be adjusted to take holistic fairness into consideration. Researchers must also consider the geographic size and shape of the census tracts (or units used), which can vary considerably. Within cities, tract sizes are relatively similar (balancing on number of residents), but if the facilities were to be placed in an exurban area or rural setting, it will be important to consider the population distribution of the tract rather than using the center of the tract as the point from which distances are calculated (as described in the Methods section). Work by Kim and Horner30 compared accessibility measures using the geometric centroid (GC, used in this paper) to that obtained by population weighted centroids (PWC) and population referenced centroids (PRC), both of which take the distribution of persons within the geographic unit into account. Their analysis found that in more sparsely populated areas, the use of geometric centroids tends to underestimate accessibility. The authors also recognize that the use of PWC and PRC requires additional data and greater computational burden. Recognizing this limitation, our estimates of centrality and fairness likely perform less well in the larger, less-densely populated census tracts outside of the city limits of Spokane, and future work in exurban areas should take population density into account. One major limitation to our study is that we did not examine facility load in our analysis6,30–32. Facility load would be more important in some contexts than others, but is particularly important in the case of facilities like schools or hospitals. One method of incorporating load is to use the Gaussian two-step floating catchment area, which considers both the supply side of facility capacity and the demand side of facility/amenity accessibility. This model has been used in the context of hospitals30,31 and greenspace32 and uses either population densities or facility capacity (e.g., hospital beds) on the supply side of the equation, and travel time or distance with an attenuation factor as the edge of the catchment is approached. In each of these cases, the spatial equity of facilities is then tested against one or more demographic characteristics of neighborhoods, but again in isolation. In reality, demographic characteristics interact with each other and with accessibility. For example, low-income households with children may not have access to independent transportation and therefore would be restricted to green spaces within walking distance or ones that are available via public transportation, whereas more affluent households would be willing to travel farther. Clustering neighborhoods as we have in this paper and considering a modification of the attenuation factor based on neighborhood characteristics might better account for these types of interaction. Another limitation of our study is that some facilities, such as fire stations, need to be considered in conjunction with each other. One method of handling load is to use the maximal covering location problem (MCLP) to address this situation, which incorporates a demand constraint directly into the objective function6. Alternatively, another strategy that has been proposed for locating facilities that work in this fashion is to use dominating set theory within the graph framework to determine optimal facility placement in a way that minimizes total travel distance33. A “dominating set” for a graph is the subset of nodes, D, within G = (V, E) such that every node not in D is adjacent to at least one node within D. This method can then be applied, using adjacency matrices, to ensure that some specific type of facility (e.g., bus stop, school) must be part of the dominating set to set a maximum travel distance for each node within the network. To use the dominating set theory strategy, the graph would need to be set up in such a way that nodes were roughly geographically of equal size (e.g., same area). Census tracts are established so that they are roughly equal in population size, not land area. Within city limits, tracts are geographically quite small, whereas, tracts outside the city are much larger. Even within the city, tracts are not quite equally sized or shaped, and geographic features such as rivers do not always allow direct travel even between those that share a border. Graphs used in this type of study could be easily adapted to use this technique, and distances could be minimized to all nodes regardless of demographic characteristics. Another small limitation of our study is the exclusion of the node that had missing median household income data (Census Tract 25.03, see “Methods”). In future work, we could impute income data for that node or use a different time window for the ACS data. In this study, we wanted to use the most recent data available, but moving the window back slightly or waiting until the next window opens could resolve the issue. That five-year period (2017–2021) may have been disproportionately impacted by the COVID-19 pandemic and had lower response rates for the ACS. Cities are not built overnight, and in some cases, it might be a bit of a “chicken and the egg” problem. Are undesirable facilities built near low-income or disadvantaged populations, or do more affluent households tend to move away from the area or fight zoning laws if an undesirable facility is proposed? Recognizing that facility relocation is challenging, our methods are ideally used in the context of new facilities, or (as described above), for ensuring that remediation efforts or other types of services are provided to the communities situated near those sites to mitigate any potential negative health or safety implications, such as ensuring that communities experiencing adverse health impacts from a polluted site have easy access to health care facilities. This method could also be highly useful in regions or municipalities that are undergoing rapid population expansion or contraction to ensure that resources for public services are allocated in a manner that is both fair and efficient. Replacement of old undesirable facilities is not out of the question, however. As new technologies are developed, particularly with respect to clean energy, replacement of pollution-heavy facilities may be a priority. A recent paper examining equity and disparate impacts of pollution from Peaker plants in Washington state noted that replacement of these high-particulate emitting plants with energy storage systems could reduce emissions and improve health3,34. The data used in this study are easily available within the United States, but the approach could be used elsewhere. A significant amount of current research and interest focuses on equitable placement of health care facilities, and studies have shown that location itself (or location as a proxy for other variables) can impact health and morbidity35,36. The study19 examined travel distances to health care facilities in Uganda. In that paper, the authors were primarily interested in travel distances (by foot, bicycle, and vehicle) to health care facilities, and whether distances were similar across regions of different income quintiles. An approach like the one we applied could be used to determine where the government could construct additional health care facilities to reduce inequalities in health care access to some rural and/or low-income regions where access to motorized vehicles is limited. Similarly, the public health community has raised concerns about “health care deserts”37,38. Health care deserts are defined as areas that have limited or no access to health care providers or facilities, or areas where the cost of health care is high or involves long waits38. Closure of small rural health care facilities and hospitals accelerated during the COVID-19 pandemic within the U.S., and now residents of many regions of the country must travel long distances for basic services, including primary health care and obstetrics37. These factors not only limit health care access but have implications for climate change, as patients must now drive long distances to access basic services. Our method could be useful when placing new health care facilities based on demographics (for example, obstetrical facilities placed near areas where women of childbearing age are a major demographic group). In those cases, using a PWC or PRC rather than a GC for calculating distance may be advantageous given heterogenous tract sizes. Methods such as these have been increasingly considered within the context of climate change. For instance, researchers have long known about disparate mortality rates for low socioeconomic or historically disadvantaged populations during heat waves12 and tied mitigation to policy changes. A recent study in Richmond, VA, found that historical redlining in neighborhoods is associated with differences in the location of extreme heat within the city. Specifically, areas that were minority-majority or had lower socioeconomic status were disproportionately hot, and the authors suggested that equitable city planning could include building and maintaining additional green space in these areas39 and this technique could inform placement quite easily. Along these lines, the work of Kimpton20 used a k-means cluster analysis to categorize green spaces within the Brisbane statistical division in Queensland, Australia, to determine if there was a difference in the quality of green spaces in affluent vs lower income households and found gaps in social equity. The technique presented in this paper could be applied within this context to assist with optimal green space placement, where both the green spaces and the neighborhoods are categorized by cluster analysis. In conclusion, this work has presented a novel technique for equitable facility placement using an objective function that takes both distance and demographic heterogeneity of a region into account. An additional strength of this method is that, using the same methodological approach and source of data, the analysis can be modified to fit the specific context and type of facility under consideration by adjusting the factors for inclusion in the cluster model, along with hyperparameter tuning. Finally, we have given examples of how this technique may be used with modification or in combination with other techniques in the literature to adapt to other communities and geographical considerations. Given rapid demographic change in many regions, coupled with impacts from climate change, natural disasters, and emergent diseases, as well as a concern to ensure equitable access to facilities, this approach is computationally feasible and broadly adaptable in a variety of contexts. Methods Definitions We collect formal definitions of terminologies used throughout this paper in this section. The section begins with general graph terminology followed by definitions of two methods for quantifying node centrality, and ends with an example to demonstrate these concepts. Most of the definitions for graph concepts given below are adopted from Easley and Kleinberg40. A graph, G, is a relation-based model for some system consisting of two sets: V is a set of vertices, V=v1,….vn, and E is a set of edges, E=e1,….em. Vertices represent the objects in the system being modeled, and edges are the relationships between the vertices in V. In our context, edges have at least two attributes specified: weight and direction. Edge weight assigns a value to the relationship being considered. In cases where this is unimportant to the model or doesn’t make sense to assign unique values, all edges will be weighted equally with a value of 1, in which case the edges and graph are referred to as unweighted. An edge may be directed in the case where the relationship being considered is not always symmetric, or undirected in the case where the relationship is always symmetric, or its direction is unimportant to the model. When E is a set of directed edges, G is said to be a directed graph or digraph. When E is a set of undirected edges, G is said to be an undirected graph. If any edge e∈E is directed, all edges e∈E must be directed. Edges and vertices may have other arbitrary data types associated with them as needed for the model. Two vertices are said to be neighbors if they are connected by an edge. A path p=v1,e1,v2,e2,…,ek−1,vk is an ordered set of alternating neighbors and their connecting edges. A path’s annotation may be simplified to only include vertices in cases where edge identity is unimportant, for example in an unweighted, undirected graph. A path contains a cycle if any specific vertex appears more than once in that path. Paths containing cycles are excluded from consideration in our analysis because they can be arbitrarily long, and this is not useful or logical in the context of facility placement. All following mentions of paths therefore refer only to paths without cycles. The length of path p, l(p), is the sum of weights for all edges e∈p. When G is unweighted, l(p) = k − 1, where k is the number of vertices in the path. Let P (u, v) denote the set of all unique paths between two vertices (u, v) ∈ V. Then, a path q ∈ P (u, v) is said to be a shortest path from u to v if and only if it has the minimum length of all paths p ∈ P (u, v). The length of the shortest path from u to v is also considered to be their distance, d (u, v). Therefore, The following definitions for vertex centrality are based on those provided in Koschützki et al.41. One method for evaluating the centrality of a vertex is to use its eccentricity. The eccentricity, e(u), of a vertex u∈V is the maximum distance between u and any other vertex v∈V. That is, Then, eccentricity centrality can be calculated as CE=1eu such that a vertex with lower eccentricity will have higher centrality (importance). Another method for evaluating centrality is closeness, c(u), which is the total distance from u to all other vertices v∈V, or The term “closeness” is a bit unintuitive in this case, as larger values indicate that a vertex is actually less close, collectively, to the other vertices in the graph and lower values indicate it is closer. Closeness centrality is calculated as CC=1c(u) so that a vertex with lower closeness will have higher centrality. To compare CE and CC, when considering distances between vertices in a graph, CE addresses the worst case, while CC addresses the average case. This can help determine which centrality metric to use in a given scenario. To demonstrate, let us consider the unweighted, undirected graph, G1, shown in Fig. 5. From a cursory visual inspection, we would intuitively expect vertices A and H to have the lowest centrality values, and maybe vertex D to have the highest centrality values. The set of all unique paths between A and H is P(A, H) = {p1, p2, p3, p4}, where p1 = {A, B, D, E, G, H}, p2 = {A, B, D, F, G, H}, p3 = {A, C, D, E, G, H}, and p4 = {A, C, D, F, G, H}. Example graph on eight vertices. Nodes are named A-H, and the graph is used as an example in the text, referred to as G1. In this case, each path is a shortest path because they all have the minimum length of 5, and thus d(A, H) = 5. From this example, we see that distance is a symmetric relation in an undirected graph, so d(A, H) = d(H, A) = 5. It is important to note that this is not the case for digraphs. We can verify our visual intuition of G1 by quickly checking that d(A, H) ≥ d(A, v) : v ∈ V1, therefore e(A) = e(H) = 5 and CEA=CEH=15. Looking at vertex D next, we see that the maximum length for all its shortest paths is d(D, H) = e(D) = 3 and CED=13. However, CEF and CEF are also both 1/3, so in terms of CE, vertices D, E and F are equally central. For the rest of the vertices, CEB=CEC=CEG=14 . Next, let’s look at CC, closeness centrality, in graph G1. We can see that CCA=119, CCB=115, CCC=115, CCD=111, CCE=113, CCF=113, CCG=115, and CCH=121. The important thing to note here is that vertices may or may not have the same CE and different CC, and vice versa. When considering facility placement problems, the choice of centrality measure is important. Besides CE and CC, there are several other centrality measures that can be found in the literature, for example, Koschützki et al.41 but for this work, we focus on these two measures. Data processing This section describes the methods used to find and prepare census data for determining socioeconomic status of areas in Spokane and construction of a graph incorporating this data for use in our analysis. For our analysis, we accessed socioeconomic and geographic data from the Census Bureau28. The most recent detailed tables available for Spokane County, WA, at the level of census tract were the 2021 5-Year American Community Survey (ACS) aggregate data (data estimated from sampling over 2017–2021). These data include some information collected during the pandemic, and thus at least partially reflect changes in household wealth or demography that may have occurred during that period. Using an API key provided by the Census Bureau, we initially imported 55 variables related to a suite of demographic, economic, and household characteristics for each of the 130 census tracts in Spokane County. We examined the variables to determine if there were any data missing across tracts and if sufficient variation existed among tracts to be useful in further analyses. Examples of variables examined for inclusion include total population, median household income, population counts by racial or ethnic group, education level, household access to internet, poverty status, and public transportation use. To partition census tracts into clusters, we applied a K-means clustering algorithm. Of the 55 variables originally examined, 45 were selected for the cluster analysis (see Table S1 in the Supplementary Information). While selecting variables for inclusion, we were careful not to select any combination of variables that would be direct linear combinations of each other. For example, the reference group for household income was the count of households with an annual income greater than $135,000 per year, and the education reference group was households where the primary householder had an education less than high school. One variable that we wished to include (12-month median household income in dollars) was missing for one census tract (Tract 25.03). Because of that, we excluded Census Tract 25.03 from further analyses. Prior to clustering, variables were standardized by dividing each of the variables by their variance using the scale function in R. For clustering, we used the “kmeans” procedure within the R stats package. The “kmeans” algorithm minimizes the sum of squared error to the cluster center. We tuned hyperparameters by varying the number of centers and starting random sets and then examining the clusters for interpretability. We decided on a four-cluster partition as optimally creating clusters which had lower within-cluster variation than between-cluster variation (formally tested using within cluster sums of squares and gap statistics). The four-cluster set was also highly interpretable and avoided overfitting; changing from three clusters to four simply resulted in the largest cluster being split, while moving to five clusters began to repartition some clusters in less easily interpreted ways. The four clusters selected are roughly characterized as: Low Income, Low Education (LILE) Low Income, Low Education (LILE) Low Income, Moderate Education (LIME) Low Income, Moderate Education (LIME) High Income, Low Education (HILE) High Income, Low Education (HILE) High Income, High Education (HIHE) High Income, High Education (HIHE) Table S2 in the Supplementary Information provides a more detailed description of the clusters. In addition to demographic variation, clusters also differed a bit in their geography; for example, HILE clusters tended to be located more on the perimeter of both the city and county of Spokane, while LILE clusters tended to be concentrated within the city center (see Fig. 7). Map of census tracts by cluster type for the city of Spokane. Desirable facilities are shown with black dots and undesirable facilities with red dots. LILE Low Income, Low Education, LIME Low Income, Moderate Education, HILE High Income, Low Education, HIHE High Income, High Education. Facilities chosen for this analysis include five undesirable facilities, four of which are Superfund sites located within Spokane County and one of which is a wastewater treatment plant. We also selected 22 desirable facilities, 16 of which are fire stations (note: according to the City of Spokane website, fire stations with the numbers 10 and 12 do not exist), five are parks including one sports complex with a play field, and one is the Sacred Heart Medical Center, part of a large hospital complex on the southeastern edge of Spokane (see Table 2 in Supplementary Information). These facilities were selected in part due to the specific demographic make-up of the county, with the assumption that the influx of younger professionals may seek access to parks and recreation, while older adults would need easy access to health care. Fire stations were selected because a clear correlation exists between travel distance and service response time in an emergency. For each of these facilities, geographic coordinates were obtained using addresses and Google Maps. Facilities are plotted on the Spokane County map in Fig. 6 and on the Spokane City map in Fig. 7. In both figures, red dots are undesirable facilities, and black dots are desirable facilities. Map of Spokane County by cluster type and facility location. Desirable facilities are shown with black dots and undesirable facilities with red dots. LILE Low Income, Low Education, LIME Low Income, Moderate Education, HILE High Income, Low Education, HIHE High Income, High Education. Graph construction Two types of analyses were performed using the Spokane County tract data. First, undesirable facilities were analyzed by using latitude and longitude to calculate Euclidean distance from the facility under consideration to each tract based on the coordinates provided in the Spokane County data. In this case, the primary concern is simply proximity and not travel time, so there is no practical need to construct a graph for undesirable facilities, but conceptually, it can be thought of as a “star graph” with the facility as the center node and each peripheral node representing a tract. This method provides a simple yet realistic approach to modeling distance to undesirable facilities and is computationally faster than models which consider a radius-based approach10. The second analysis used a subset of 62 tracts located within the City of Spokane proper as nodes to construct a graph for analyzing desirable facility placement. In this case, the primary consideration is travel-time over roads, so nodes were connected using reasonable travel routes between neighbors to incorporate existing roads and barriers (such as the Spokane River, which limits connectivity between North and South Spokane). Assigning edges was done manually by connecting nodes with a direct route between them (no other nodes in the network fall between them on a direct route) and giving an edge weight whose value reflects travel distance from Google Maps, in kilometers (Fig. 8). An issue worth noting here is that the latitude and longitude for each tract give the location of a point within that tract. Our assumption is that these points provided by the census data are near the geographic center of each tract, making the edge weights on our graph representative of an average travel time between any two points in those tracts. City of Spokane graph. Node colors indicate cluster type as in Figs. 5 and 6. Edge thickness indicates edge weight (distance). Fairness-aware metrics We define and explain in this section our proposed metrics for incorporating fairness into facility placement decisions. We have named the fairness metric developed for this work population-weighted Average Cluster Distance, or ACD. Given a graph G with edges and vertices E and V, respectively, clusters Q, and target vertex u ∈ V representing the location of the facility under question within G: The function d(u, v) is the distance function described in Eq. 1, and the function w(v, q) is the population weight function, which returns the percentage value (as a decimal) of the total population under consideration that is associated with cluster q at vertex v. 1 is the indicator function, which in this case returns a value of 1 when the value returned by w is a positive real number and 0 otherwise. This results in a metric incorporating both distance and population distribution that can be compared between clusters to see how they differ in terms of travel distance to the facility or location being analyzed. Smaller values indicate a cluster’s population has a shorter travel distance, collectively and on average, than the population of a cluster with a larger value. For undesirable facilities, we stopped here and compared ACD between clusters, whereas for desirable facilities, we went a step further by taking the difference between the maximum and minimum of the ACD across Clusters, or ACDC, expressed as If the maximum and minimum distances are very close to each other, then the average distances from the facility/location to the populations of each cluster type are very similar, and ACDC will be small. If, instead, the maximum and minimum distances are very different, then ACDC becomes large. Therefore, a smaller ACDC represents a fairer placement than a larger ACDC. For undesirable facilities, fairness calculations didn’t require any external libraries. ACD was calculated for each facility and compared against the overall average to evaluate fairness of facility placement. It was not necessary to consider centrality metrics for undesirable facilities because proximity alone is the relevant attribute, as noxious smells, pollutants, or sounds could impact residents in any direction. The undesirable facilities we considered for this project were four Superfund sites (Colbert Landfill, Fairchild Air Force Base, the Kaiser Aluminum facility, General Electric Service Shop) and the Spokane Wastewater Treatment facility. For desirable facilities, the Spokane city graph was constructed using Igraph42 in Python. Vertices were census tracts, and edges were done manually as described above. Centrality metrics used were weighted closeness centrality (CC) and weighted eccentricity centrality (CE). CC could be done directly in Igraph, but we had to calculate CE ourselves because Igraph only provides unweighted CE. The fairness metric used for desirable facilities, ACDC, was described above. To apply it, we calculated the ACDC of all nodes in the graph and then associated each desirable facility with its closest node based on geographical distance using latitude and longitude. Then, the fairness of a desirable facility is the ACDC of its associated node in the graph. Optimizing for both holistic fairness and centrality involved a few additional steps, building upon the metrics we had already calculated for each node. Because we want to optimize for both fairness and centrality, and we prefer smaller fairness values rather than using CC and CE for this, we used eccentricity and closeness themselves, since smaller values for these metrics indicate higher centrality. Since we are doing a simple summation for our objective function, these values were scaled to be between 0 and 1 by dividing each node’s fairness/eccentricity/closeness value by the largest value that occurred in the graph for each respective metric. This makes sure we are giving the two parameters (fairness and eccentricity or fairness and closeness) equal weight in the optimization. Alternatively, if we wanted to weight one or the other more strongly, we could have introduced coefficients for each term in the objective function. Supplementary information Supplementary information Acknowledgements This work was supported in part by cooperative agreement CDC-RFA-FT-23-0069 from the CDC’s Center for Forecasting and Outbreak Analytics. CDC played no role in the design of the study, the collection, analysis, and interpretation of data or in writing the manuscript. Author contributions S.K. analyzed and interpreted US census data and wrote all R code. J.C. performed graph construction and centrality measurements and wrote all Python code. All authors collaborated on fairness metrics. S.K. and J.C. were the major contributors in writing the manuscript, and A.G. was also a contributor in writing the manuscript. All authors read and approved the final manuscript. Data availability The data that support the findings of this study are available from the US Census Bureau (for census data) [https://data.census.gov/] and Google Maps (for Spokane city layout) [google.com/maps]. Code availability The underlying code for this study is available in Open Science Framework and can be accessed via this link: https://osf.io/2ktg6/?view_only=5d75a8ca72274c34a39f999fa19c804e. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s42949-026-00414-0. References Associated Data Supplementary Materials Supplementary information Data Availability Statement The data that support the findings of this study are available from the US Census Bureau (for census data) [https://data.census.gov/] and Google Maps (for Spokane city layout) [google.com/maps]. The underlying code for this study is available in Open Science Framework and can be accessed via this link: https://osf.io/2ktg6/?view_only=5d75a8ca72274c34a39f999fa19c804e.