A Study of Housing Insecurity Risk Factors Across the Rural-Urban Continuum

1. Introduction

The rural-urban continuum is an important sociological and human geographic phenomenon that has become a core variable in social sciences over recent years. While historically, the rural-urban continuum was meant as a dichotomous concept, it is more accurate and useful to conceptually describe it as a progression [1,2]. Studies of the rural-urban continuum can be seen across the social sciences, with applications to psychology [3], geographic change [4,5], demography [6], political science [7,8,9,10], sociology [11], economics [12,13], and criminology [14,15]. In short, while we as a society tend to acknowledge urban and rural areas, we operationalize these concepts in overly simplistic ways; rural areas are often defined as simply non-urban despite a wide variation geographical and social factors among areas classified in this way [16]. This leads to significant challenges in understanding the natural progression of socio-economic and/or socio-cultural phenomena. And as Lichter and Brown [17] note, rural and urban areas, while distinct across many metrics, are also interdependent and face a range of similar social problems.

In the United States, a series of codes were developed by the U.S. Department of Agriculture [18] in the 1990s and were reevaluated after the decennial census to characterize this rural-urban continuum (Table 1). This process allowed for the designations of areas to change with their present reality through time, representing temporal evolution. A map of current RUCA codes for US census tracts is shown in Figure 1.

In research on social issues, urbanness is often the dominant paradigm. As a result those with lived experiences sometimes express self-identities that are at odds with the RUCA designations [19]. Part of this issue is the fundamental way in which RUCA codes are inferred from census and American Community Survey data. Overall, there are a lot of areas that when you change how they are modelled, a different outcome arises. However, tract level analysis shows about 97% convergence between types [20]. This fundamental assertion makes understanding how social phenomena occur within these groups even more tantamount. Acknowledging the full continuum allows for creativity in policymaking by parsing more nuanced perspectives to match solutions to concepts of underlying reality, impacting areas across the continuum more appropriately to their own context.

When we think about the RUCA code designations, it is important to recognize that RUCA codes are not about population density; rather, they are about the flow of people to work and centralized services. A couple of prime examples in Table 1 are Portland, ME (1), Standish, ME (2), Augusta, ME (4), Presque Isle, ME (7), and Farmington, ME (10). All of these locations house a college or university. Portland itself is a sufficiently large urban city to be defined as a metropolitan core. Residents of Standish typically commute to Portland. Augusta represents a micropolitan area that people commute to. Presque Isle represents a small town that people commute to. Farmington represents a rural enough area that any commuting does not enter a micropolitan area. The population is not the main focus for those areas that give flow, but rather for those that receive it. For example, Livermore Falls (3) has less than half of the population of Farmington (10), yet it has a more urban-like RUCA code. The difference is that Livermore Falls flows into Lewiston-Auburn (1), whereas Farmington pretty much sustains itself and is not incorporated as a city.

Metropolitan, micropolitan, and small-town areas serve as hubs of activity, and because of this, they are more likely to have some semblance of centralized resources that surrounding communities will leverage. This is not universally restrictive. For example, Skowhegan, ME has a functioning hospital and medical offices, despite being classified as a 10. In Skowhegan is some necessity to having these functions due to its location in Somerset County with respect to other communities which cannot on their own support that type of resource. This can lead to fiscal inefficiencies in the provision of resources in rural areas that are otherwise straightforward in urban ones, a concept that is critical to understand in housing insecurity research.

Housing insecurity is a critical concept in understanding public health, where housing is considered an important social determinant of health [21]. Coleman-Jensen, Steffen, and Whitley [22] describe housing as a “somewhat lumpy bundle of services” that consists of “a package of basic shelter, neighborhood security and amenities, access to transportation and social networks, personal consumption, productivity potential, and financial investment potential” (270). Just as housing itself is complex, so are the broad range of social outcomes with which it is associated. While housing insecurity is often most strongly associated with homelessness among the general public [23], it is a far more structurally complex issue,

Swope and Hernandez [21] have convincingly documented the health impacts of insecurity housing using the four C’s framework: Cost, Condition, Consistency, and Context. The four C’s framework gives some structure to organizing the “somewhat lumpy” [22] aspects of housing. The first C is cost, which draws attention to the importance of the cost-burden of housing. If community members pay most of their income to housing (e.g., rent, utilities, maintenance costs), then very little is left to cover other costs (e.g., food, medicine, transportation). Condition examines the quality of the housing, where overcrowded or dilapidated housing can negatively impact health and wellbeing. Consistency draws attention to the instability of involuntary moves (e.g., evictions, couch surfing) and the negative impacts this can have on health. Finally, the fourth C is context, which focuses on the community or neighborhood context of housing. If a community is unsafe, unwalkable, or lacks adequate sources of healthy food, this can be detrimental to health [21]. Swope and Hernandez make a strong argument for broadening research and understanding on housing insecurity to conceptualize it as a continuum of housing struggles, where homelessness is included as an unsheltered expression of this continuum [21].

For the most part, the overwhelming majority of housing insecurity research is anchored in urban areas [16], whereas there is very minimal research into the effects of housing insecurity in rural communities, although the volume of it is increasing [24,25,26,27,28,29]. The existing body of literature on rural housing insecurity indicates that assumptions derived from research on urban homelessness cannot be applied wholesale to rural areas [16,24,25,26,27]. For example, methodologies of measuring rates of homelessness typically applied in urban areas do not adequately capture the extent of the problem in rural areas [25]. Unsheltered and other types of housing insecurity tend to be more hidden [26] and less publicly recognized [27] in rural areas. Gleason et al. [26] have also stressed the need to appreciate the unique housing challenges of different rural areas. They used key informant interviews to explore housing insecurity across four different rural areas in Maine and found that the differing histories and geographies of these areas likely shaped differences in housing issues across the areas. Because rural housing insecurity is under researched, there are still large gaps in our understanding of the diverse risks, contours, and needed interventions in rural areas [26,27]. Montañez’s [30] ethnographic study of Taos is an example of the complex ways in which housing insecurity reflects many aspects of the “four C’s,” and how these struggles are shaped by the specificity of the particular ruralness of Taos. In this location, a gentrifying rural community faced housing shortages due to residential mobility and in-migration, rising costs, and increased incidence of AirBnBs. Montañez noted distinct patterns in how different ethnic groups navigated housing insecurity, with long-term Hispanic residents being more likely to rely on kinship networks and strategies such as renting rooms or trailers on family land, which can provide barriers to applying for aid, and using robust social networks to learn of resources. White newcomers, however, were more likely to engage with formal services and live in vehicles or shelters. Coleman-Jensen, Steffen and Whitley [22] also cite survey data that suggest the tendency to rely on family is more common among the rural housing insecure, whereas urban individuals are more likely to appeal to friends, emphasizing how different social networks and family structures are more or less salient in different spatial contexts.

Speaking more broadly, there are a range of reasons why we might expect rural and urban areas to have different levels of risks across the four C’s of housing insecurity. Some rural areas, for example, have been experiencing population loss (e.g., outmigration) for decades, while others have started to see marked increases in population due to in-migration of retirees and tourist populations [31]. Rural areas also have higher rates of homeownership than the national average, and generally have lower home prices, on average [32]. However, despite this increased affordability and availability, rural areas have higher rates of poverty than the national average [33]. These trends also vary significantly by region as well as rural-urban status [22].

Because of the need to better understand the spectrum of housing insecurity risks in and its intersection with rural-urban dynamics, we propose the following research questions: does the rural-urban continuum relate to shifts in rates of risk across the four C’s of housing insecurity? Do RUCA codes 1, 4, and 7 (anchoring hubs) behave similarly to one another (or more similarly to one another) than do other codes? Would that behavior pass down to the percentile exchange groups (2, 5, 8 and 3, 6, 9 respectively)? How does code 10, the most rural code and one outside of an explicit commuting area, relate to any of these? Prior research has suggested that the more rural an area is, the more likely that the housing insecurity strain is targeted to one or two of the 4 C’s as opposed to systemic failure [26].

In this paper, we tried to predict risk of housing insecurity based on publicly available data regarding demographic information across the United States that reflect the 4 C’s of housing insecurity. To accomplish this, we used several data mining techniques to create an ensemble approach to risk assessment. For each dimension, we asserted a risk for the individual tract. From this, we constructed intersectional profiles of the 4 C’s and compare incidence rates of risk based on the RUCA code designations. The goal of the study was to understand how our risk assessment measure correlates with the structure of the rural-urban continuum. Furthermore, we endeavored to consider how urban areas might be able to learn from the profile of rural areas and similarly how an area going through a transition in the rural-urban continuum might project to appear within those dimensions.

2. Materials and Methods

Using the tidycensus package from R, a total of 85 2020-2024 ACS 5-year variables were extracted at the census tract level to capture the four C’s of housing insecurity. A series of eight concepts were used to characterize areas motivated by extant literature [16]: economic diversity, demographic diversity, housing cost, housing quality, residential mobility, household factors, and housing type. Of particular note in this dataset is that a small set of census tracts from Brentwood, NY and Wyandanch, NY are missing due to an interviewing error from the US Census Bureau. These datapoints have been removed.

A series of measures were constructed to address each of these areas by combining relevant ACS measurements:

• Economic diversity
  • o Sullivan index of economic diversity, GINI index, proportion of workers in below average economic wage categories
• Demographic diversity
  • o Proportions of white, women, and children, Sullivan index of racial diversity
• Housing cost (Cost)
  • o Mortgage, owner non-mortgage, and renters paying over 30% of their income (burden)
• Housing quality (Condition)
  • o Occupied incomplete kitchen or plumbing
  • o All incomplete kitchen or plumbing
• Residential mobility (Consistency)
  • o Same house, moved within county, moved within state, moved out of state
• Household factors
  • o Households with 3 or more workers, households without investment income, households with no other income, households with no vehicle, households with no wages, households with public assistance, households with Social Security Insurance, households with workers with no vehicle
• Housing type
  • o Renter/owner ratio, renters broken down by single, 2-4, 5+ units, mobile homes, and unconventional housing.

The USDA RUCA codes table linking census tracts to the rural-urban continuum for the 2020 decennial US Census were joined to the remaining data columns from the ACS data.

In each of the various groups, k-medoids clustering was employed to determine clusters of high-, medium-, and low-risk of housing insecurity based on how the literature characterizes housing insecurity challenges using silhouette metrics [34]. For example, an area with high economic diversity (i.e., a varied job profile) has better economic resilience as compared to an area with a lower economic diversity is susceptible to employment and income challenges, thus the cluster with the most economic diversity is labelled lowest risk and the one with the least is labelled as highest risk.

Association rules mining was conducted on these associated risk levels to determine if there are patterns within the risk levels. Association rules are formed when there are several comorbidities that tend to be found in concert, increasing the likelihood of the occurrence of the other from the base proportion, evaluated by measures such as support, confidence, lift, gain, and conviction [35]. It is important to consider these concepts to see if the nature of comorbidities is changing along the rural-urban continuum.

To achieve an understanding of the impact of the rural-urban continuum on housing insecurity, these approaches are further analyzed with respect to the RUCA codes. We approach this problem from two different perspectives:

• Central locations (1,4,7) compared to various feeding levels (2,5,8), (3,6,9)
• Size of the centralized hub (1,2,3), (4,5,6), (7,8,9), (10)

3. Results

Utilizing the RUCA codes and various ACS data columns, we present a multi-pronged analysis of the Rural-Urban Continuum across housing insecurity dimensions. In Section 3.1, we create clusters of risk of housing insecurity across seven characteristics. In Section 3.2, we look at how the Rural-Urban Continuum is distributed across the various risk factors. In Section 3.3, we look for comorbidities of housing insecurity risk. In Section 3.4, we look for changes in those comorbidities based on RUCA codes.

3.1. Clusters of Risk

Using the cluster library in R, the ACS data of each census tract with a complete record was processed using the clara function for each of the seven identified dimensions of potential housing insecurity. The clara function attempts k-medoids partitioning by taking samples of larger datasets and then propagating those results across the rest of the dataset a specified number of times. The model was run under an assumption of self-standardization (to account for the diversity of scales of measures compared to one another) and was based on 20,000 samples (~25% of the dataset). Because demographic data is inherently messy in terms of similarity, it is expected that there will be overlap between clusters. As such, we pay closer attention to the silhouette scores than the isolation scores [34]. Silhouette scores can range from -1 (totally random cluster) to 1 (totally isolated clusters). Silhouette scores above 0.7 are considered strong; above 0.5 are considered moderate, and above 0.25 are considered weak. The results of the clustering protocol are enumerated in Appendix A in Table A1, Table A2, Table A3, Table A4, Table A5, Table A6 and Table A7. Silhouette graphs are presented in Appendix A in the odd numbered Figure A1Figures –Figure A14. PCA plots are presented in Appendix A in the even numbered Figure A1, Figure A2, Figure A3, Figure A4, Figure A5, Figure A6, Figure A7, Figure A8, Figure A9, Figure A10, Figure A11, Figure A12, Figure A13 and Figure A14. Table 2 summarizes the silhouette scores for the various clusterings from Table A1, Table A2, Table A3, Table A4, Table A5, Table A6 and Table A7.

Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8 show the spatial distribution of each of the clustering outputs within the coterminous United States. Each of the figures shows substantive variation, suggesting that a typological approach is likely to better serve practitioners.

3.2. Distribution of Risk Across RUCA Codes

Areas with different RUCA codes are fundamentally different by their geographic and communal nature. It is thus important to understand how the classifications break down across RUCA codes. We examined percentages to address this phenomenon as it is disproportionately likely to see census tracts at levels 1, 4, 7, and 10. The resulting tables are found in Appendix B. As a means of summarizing, Table 3 records the proportion of high risk tracts in each risk category.

3.3. Comorbidities in Housing Insecurity Risk Factors

An association rules mining procedure [35] was applied to the dataset to determine if there are risk factors that seem to predict one another at high levels. In Table 16, the list of association rules between high risk classifiers that have at least a confidence of 0.3 are presented. A lift exceeding 1 suggests that the rule predicts a stronger affinity with the antecedent (lhs) than the overall affinity of consequent (rhs) in the dataset at large. To interpret the following tables, we consider the first row of Table 4. The first row of Table 4 has LHS of ROC=H, RHS of HHFC=H, Support of 0.003, Confidence of 0.634, Coverage of 0.005, Lift of 2.478, and Count of 279. In an association rule, the LHS (left-hand side) is called the antecedent. The aim of the analysis is to see how the presence of the antecedent influences some other piece of information. The RHS (right-hand side) is called the consequent is that second piece of information. Support reflects the proportion of the overall dataset that has the two concepts together. Count reflects the raw number within the dataset. Coverage reflects the proportion of rows in the dataset that include LHS, and must be at least as large as Support because Support adds further restriction. Confidence is the conditional probability that a row that contains LHS also contains RHS, computed by dividing Support by Coverage. Lift represents the strength of the rule from the perspective of how much of a gain is produced from the information, namely Confidence/Coverage for RHS. In the case of the first row of Table 4, knowing that ROC is high risk scales the support for HHFC being high by a factor of almost 2.5, suggesting there is a strong connection between these concepts in that particular order. It is possible for Lift to be less than 1, suggesting that knowledge of LHS makes RHS less likely. There are a few instances of this within Section 3.3 and Section 3.4. In total, this line of data suggests that ROC = H is not very prevalent (this is also observed in Table A7 and B7), but when it is, it provides evidence that HHFC = H will appear disproportionately relative to its normal rate.

To place the rules in Table 4 in context, there are 80,855 records for which every measure can be computed. These rules are the only ones in which the antecedent produces the consequent at least 30% of the time (confidence). The largest number of an occurrence of this is less than 1 out of every 9 times. These rules also demonstrate that in four cases, the antecedent more than doubles the rate of occurrence of the actual phenomenon.

Similar to an association rule, it is also important to get sheer counts of the number of times a census tract has a certain number of comorbidities. That insight is shown in Table 5.

Table 5 demonstrates clearly that the majority of tracts have few high risk classifications, but a select few have a very large number of them. Figure 9 shows how these counts distribute through space.

Both Table 4 and Table 5 however treat all census tracts as equal; RUCA codes suggest there are fundamental differences between census tracts that center on surrounding context. In Section 3.4, we break this down further.

3.4. RUCA Codes and the Impact on Comorbidities

In order to get a better understanding of the impact of the RUCA codes on comorbidities, we repeat the analysis from Section 3.3 on each RUCA code subset of the data. The association rules results are shown in Appendix C and summarized in Table 6. Tables C1-C10 do not immediately show patterns in their relative size or their components. To get a better understanding of their similarity, we have compiled a compendium table (Table 6) that shows where the various combinations occur.

Table 6 shows that RUCA code 1 is the dominant feature in producing association rules. This is unsurprising as it has 68% of the dataset. When we remove the effect of RUCA code 1 by letting each RUCA code be its own set, it becomes easier to see that some associations are hidden by the most urban census tracts. The association rules analysis provides a glimpse into which factors occur together in only urban environments, only rural environments, or in any combination. It also reveals those that occur in many conditions.

It is also important to see where there are comorbidities. Table 7 and Table 8 show the raw counts (Table 7) and probabilities (Table 8) of a tract having a certain number of comorbidities.

4. Discussion

There are numerous points of interest to discuss relative to the analyses in Section 3. Section 4 is thus broken down based on the dimension of the analysis in question.

4.1. Clustering Procedure

Table 2 and Appendix A highlight the difficulties in trying to predict risk factors on large scales. The ability to find segregated categorizations in multi-dimensional factors is very difficult [16]. One concept did however occur over and over again: the lowest risk category is the one that is the most isolated, having larger silhouette scores. This is a strong indicator that risk factors are diverse and localized, as shown by Gleason et al. [26].

4.2. Clustering Procedure Across RUCA Codes

The most striking finding in Section 3 is the parallels between the codes, based on their status as “core” or commuting outflow areas as exhibited in Appendix B and Table 3 as well as in Table 7 and Table 8. The data did not cluster in risk factors at ends of the spectrum as one might expect with a clear “rural” vs. “urban” dichotomy. Instead, similar types of places (e.g., flow centers) looked more like their functional counterparts than their spatial (e.g., rural or urban) counterparts. This indicates another example of rural-urban alignment and interdependency [5]. Another noteworthy finding was the comorbidity count for RUCA code 10. From the perspective of having any morbidity with respect to housing insecurity or comorbidities, RUCA code 10 seemed to behave more like the rural codes that flow into RUCA code 7 than the RUCA codes 1, 4, and 7. This indicates that the most rural places in the U.S., especially those that are not within easy commuting distance to employment centers face a range of social challenges, from concentrated poverty and racial and ethnic segregation to becoming home to LULUs (locally undesirable land uses) such as prisons. This is the type of place that Eason [36] dubbed “rural ghettos.” These communities, with little economic opportunity and few ties to regional economies become stigmatized and economically depressed, presenting them with challenges somewhat similar to disadvantaged urban neighborhoods, but without the benefits of population density, economies of scale, or public transportation to help mediate these conditions. This stands in contrast to those areas that flow into metropolitan (RUCA code 2-3) and micropolitan (RUCA code 5-6) areas which have the lowest rates of both morbidity and comorbidity. We might classify these communities conceptually as suburbs. In this sense, trends in the areas surrounding rural and urban centers seem to come together.

Appendix B and Table 3 show interesting patterns that relate to RUCA code designations. Economic diversity (Table B1) and household factors (Table B5) show a similar pattern where the rate of occurrence of a morbidity is strongest in flow centers and the rate of morbidity increases as the population of those flow centers decreases. In other words, high risk in economic diversity was highest in flow centers, with higher rates of high risk in economic diversity in smaller flow centers. We interpret this as inequality in wages and wealth being most prominent in flow centers, especially those with smaller populations (e.g., small towns). This reflects the reality that less than thirty percent of rural counties are classified as economically diverse, while a majority of metropolitan counties are [37]. Household factors, which also capture economic vulnerability (e.g., not having transportation or wages and using government benefits) are also more prominent in flow centers and especially prominent in smaller flow centers. What is notable is that this seems to demonstrate low population flow centers are likely to be areas of high service need with relatively few service resources [38]. For example, few of these areas are likely to have emergency shelters for housing [25].

Demographic diversity (Table B2) and renter/owner characteristics (Table B7) are strongest in flow centers and dissipate rapidly both with flow rate and as the size of the flow center declines. These concepts suggest a phenomenon predicted by population and centrality. Lee and Sharp [39] suggest that at the aggregate level this is true for demographic diversity, but is potentially dissipating and lacks spatial context; similarly, urban environments have far more reliance on rentals than do rural settings [40].

Housing costs (Table B3) and residential mobility (Table B6) represent problems that are strongest in flow centers, but the size of the flow center is largely irrelevant. Housing costs tend to relate to being in desirable locations, closer to amenities and infrastructure [41], and centralized locations fit this bill [42]. Residential mobility is of course predicated by being forced to move and/or having stock available with which to move. This requires sufficient units, again, likely in centralized places.

Finally, housing quality (Table B4) seems to be a morbidity characteristic of dispersal, with the rate of occurrence increasing as the flow center gets smaller but also having stronger signals in places that flow toward these centers. This fits in with trends related to the core-periphery model, where preference is given to centralized locations for value, and thus translates into lower quality moving away from the center (in this case, places with decreased amount of flow) [43].

4.3. Association Rules

Interesting patterns seem to arise within the association rules protocol across the identified factors of housing insecurity from Section 3.3 and Section 3.4 that can be seen by navigating Table 6 with an eye toward the structure of the RUCA codes. We emphasize that the interpretation of association rules can be confusing, so we reiterate that association rules relate categories to categories: an association rule exists when knowledge of one category meaningfully alters the rate of occurrence of another category (i.e., conditional probability). This is not the same as the rate of a measure being positively correlated with the rate of another measure (i.e., regression). It is important to note that often times regression arguments can be helpful in providing support for why a category might spike in conditional probability. As such, we endeavor to provide instances where such spikes are supported by phenomenon found in the literature.

Renter/Owner characteristics (ROC) never occurs as a consequent in an association rule, either within the whole dataset or within the various subsets by RUCA code. This is not necessarily due to it not being a legitimate comorbidity with other concepts, but rather a consequence of the overall size of the high risk cluster, seen in Table A7 and B7. It does, however, serve as an antecedent, in each case predicting a rule that is relevant to flow centers. Its relationship to residential mobility (RMC) is relevant to all flow centers. In all flow centers (RUCA 1, 4, and 7) the high risk group of renter/owner characteristics substantively increases the rate of high risk in the residential mobility dimension. This makes sense as flow centers have higher rental rates [40] and eviction is a much larger problem in these areas than it is in rural environments [44]. ROC’s relationship to economic diversity (EDC) is relevant to metropolitan and micropolitan flow centers, with high risk renter/owner characteristics predicting high risk economic diversity at a stronger rate than normal; this might be caused by the relationship between the GINI index and rental affordability in metropolitan areas [42] combined with higher rates of renting in urban environments [40]. ROC’s relationships with demographic diversity (DDC), housing costs (HCC), and household factors (HHFC) are all relevant to metropolitan centers. This is a clear indication that the overall size of the flow center changes how ROC is found in situ. ROC and housing quality (HQC) do not seem to influence one another notably, which is likely a side effect of the sheer lack of volume of ROC risk in rural environments within the clustering outcome.

Flow centers seem to hold a key place in the risk factor comorbidities. Several pairs of morbidities occur only between flow centers.

High risk in the housing quality dimension appears to predict high risk in the economic diversity dimension. This suggests that substandard housing seems to place itself around areas with lesser economic diversity, higher incidence rates of poverty, and general income inequality. This would make sense across the spectrum. Areas that are characteristically central are likely to have income inequality issues, and housing quality can afflict all areas (e.g., run-down rural communities that have experienced significant economic shifts like in West Virginia [45] or cities with historical development changes such as Detroit and Flint, Michigan [46]).

High risk in special household factors appears to predict high risk in housing costs in all flow centers. This is a difficult area to discuss because many of these factors are about transportation, income, or benefit utilization. An example of this concept is found in practices by landlords to offset benefits for housing with overpriced rentals [47].

In metropolitan and micropolitan flow centers, high risk in special household factors appear to predict high risk in residential mobility. It is likely that areas in the high risk category for special household factors wind up with issues that can lead to eviction, predicating movement.

In metropolitan flow centers, high risk in the residential mobility dimension appears to predict high risk in the demographic diversity dimension. Because rents are more expensive and prevalent in urban areas [40,44] and urban areas are much more associated with demographic diversity [39], such an association makes sense. This is not, however, universal. Certain ethno-racial groups are actually highly immobile [48], however, eviction serves as a harrowing contrast to this, where those same immobile groups have stunning eviction rates [49]. Similarly, high risk in residential mobility appears to predict high risk in housing costs in metropolitan flow centers. Urban costs are skyrocketing, and this has had a substantive impact on evictions amongst urban poor [44].

In metropolitan flow centers, high risk in the economic diversity dimension appears to predict high risk in the demographic diversity dimension. This is a fairly clear connection to the concept of redlining [50], using economic means related to income inequality to segregate communities, which disproportionately impacts minority populations in the urban centers of the US. Similarly, economic diversity appears to predict high risk in the housing cost dimension, suggesting that income inequality is higher where housing is more expensive [51].

In metropolitan flow centers, high risk in the housing quality dimension appears to predict high risk in housing costs and also high risk in the demographic diversity dimension. Housing quality and cost seem to reflect a surprising juxtaposition, but Desmond [52] argues that renters in lower-income urban areas often pay disproportionate to unit worth as landlords jack prices to offload risk of renting in that environment. Demographic diversity is not nearly as surprising as with the sheer amount of income inequality in minority populations combined with the redlining effect would lead to these risk factors aligning in urban environments.

In metropolitan flow centers, high risk in special household factors appears to predict high risk in demographic diversity. One of the components of the special household factors concerned lack of access to personal transportation. There are significant examples of disparities in public transportation access in minority communities [53].

Two different high risk categories appear to predict high risk in residential mobility in micropolitan areas: high risk in economic diversity and high risk in housing costs. Micropolitan areas (like rural areas) are often not the primary target of studies [54], so exploring this linkage is a suggestion for future study.

Residential mobility and economic diversity have an interesting relationship in that they are the only pair which combines the flow centers with the areas that flow highly to them. This is a pairing that we would expect to see in flow centers due to the economic issues and eviction factors [44]. However, the addition of the high flow communities is insightful. Research suggests that suburban areas are shrinking [55]. Also, research suggests that poverty is on the rise in suburbs [56], which could also contribute to this linkage.

High risk in residential mobility appears to predict high risk in housing quality in sub-micropolitan flow providers as well as purely rural tracts. This suggests something of importance: the reason for movement in flow centers is likely centered around cost concerns, whereas in lesser populated areas, the issue is about the quality of the house itself, emblematic of the core-periphery model. Recent studies show that rural areas suffer from housing quality issues more so than their urban counterparts [57].

In both flow centers and rural areas, high risk in residential mobility and high risk in housing quality appear to predict high risk in special household factors. This would suggest that the necessity of moving and challenges with respect to housing quality impact households with specific needs across the board with respect to places that are their own emphasis. This is not surprising as one of the factors in the list is about lack of transportation, meaning that that family would be dependent upon where they live to sustain them economically.

In both flow centers and rural areas, high risk in housing cost predicts high risk in economic diversity. While costs in urban areas are spiraling out of control, the same is true in rural areas [58]. This fits in nicely with the concept of rural gentrification [59].

In both flow centers and rural areas, high risk in special household factors appears to predict both high risk in economic diversity and housing quality. Special household factors are often associated with poverty which knows no real spatial bound, and housing quality can be compromised in several locations.

In sub-micropolitan areas, high risk in demographic diversity appears to predict high risk in both housing quality and special household factors. This would suggest that as population areas get smaller, diverse populations wind up in more distressed communities [37].

In rural areas, high risk in housing cost appears to predict high risk in housing quality. This appears consistent with the idea of the core-periphery model [43], but also consistent with the idea of the growth of the AirBnB market [30].

There is one virtually universal rule (other than RUCA code 2): high risk in housing cost predicts high risk in special household factors. RUCA code 2 is significant in that it is the most prototypically suburban area.

Several rules are hidden when looking at the entire dataset. Table 6 has 12 combinations that are undiscovered until filtering the dataset into subsets by RUCA codes, clear evidence for nuance in factor combinations.

The intricacies of the relationships between various factors have an important role to play in the construction of policies geared at remediation of housing insecurity challenges. The lack of universality of the association rules shows that we must be cognizant of the types of comorbidities that are mutually predicted in areas and use that knowledge to better situate services to meet the needs of communities, be they rural, suburban, or urban. Moreover, these complicated association rules suggest that we should pay attention to both population size and flow related to economic centers of various sizes when parsing need and policy solutions.

5. Conclusions

The overall picture of the results in Section 3 presents a pattern across the rural-urban continuum in the form of the USDA RUCA codes. Flow centers and flow percentages tend to drive the pattern in the dataset for the existence of risks for housing insecurity. Only in isolated categories do we find a general decline or spike based on the relative size. This means that RUCA codes 1, 4, and 7 (core areas) behave more similarly to one another than they do to other codes. RUCA codes 2, 5, and 8 (high commute into core areas) are similarly related; as are RUCA codes 3, 6, and 9 (low commute into core areas). When searching for comorbidities, after getting to rural RUCA codes, the comorbidity profile more closely resembled flow centers. This is not all that surprising as rural communities often have to fend for themselves. Associations between specific comorbidities are dictated largely by the set of urban census tracts, but patterns exist outside of this. Particular comorbidities only come into focus when the effects of the urban tracts are removed as they are disproportionately represented as efforts are made to try to control the populations of census geometries at the tract level to be relatively equal in population where possible. To our knowledge, only Byrne et al. [60] have attempted to compare community level risk factors for housing insecurity (specific to homelessness) nationwide. Their study used U.S. Department of Agriculture’s (USDA) Economic Research Service definitions of rurality, comparing metropolitan and non-metropolitan areas on factors associated Housing and Urban Development Continuum of Care homeless counts. They did find some differences in morbidity factors between rural and urban areas. For example, rent levels and unemployment were stronger predictors of homelessness in non-metro areas. However, the findings described in the present study call into question the validity of approaching rural and urban housing risks in a dichotomous way. Our findings suggest that risk factors for housing insecurity might be patterned along dimensions more related to proximity to an economic core than strictly along metro/non-metro lines.

Interestingly, it seems that the difference between rural, suburban, and urban arise when we address comorbidity prediction. Some characteristics are predicted well as comorbidities in urban settings; others have nuanced predictions that do better in rural, suburban, or urban (and in some cases combinations). While flow appears to be the determining factor for prevalence of high risk measures for housing insecurity, the differences across the continuum that relate to size tend to have a greater effect on prediction of associated problems. This seems to indicate that risk of housing insecurity may occur along multiple dimensions related to both economic flow to centers and the size of the communities in question, with the intersection of these factors appearing to pattern risk in complicated ways.

Gleason et al. [26] have recommended that research on rural housing move towards developing a typology of patterns for rural housing insecurity. Overall, the findings of the present study illustrated clear support for the argument that rural-urban differences exist along a continuum and not a binary. As demonstrated consistently throughout the results, RUCA codes 1, 4, and 7 (core areas) are remarkably consistent in terms of their housing insecurity risk factors, despite being in very different areas of the rural-urban population size continuum. This same pattern plays out for codes 2, 5, and 8 (high commute to core areas) as well as 3, 6, and 9 (low commute to core areas). We also found notable housing insecurity risks for code 10, indicating that the very rural areas in the U.S. that are not within easy commuting distance to employment centers face higher risks for a range of social challenges. These findings highlight the importance of thinking beyond “rural” or “urban” labels and simple population thresholds and thinking instead about the social and economic dimensions of places and how those characteristics may provide a more useful, place-based typology for understanding important social problems such as housing insecurity.

This study overall highlights that understanding housing insecurity at the typological level is more than just an exercise in understanding demography; a typology of housing insecurity should also include the role of the rural-urban continuum, specifically aspects of the place’s regional role and size. Housing insecurity challenges present differently under different conditions. The risk profiles of areas based on their rural-urban continuum position can differ greatly. Understanding a place at the community-level with all of its human interactions will always be the best way to construct effective policy [61].

Future work should be devoted to addressing the diversity of housing insecurity beyond the scope of this paper in two ways: (1) using a regionalized approach to determining risk, thereby narrowing the geographic window to areas with higher similarity [62], and (2) addressing risk factors not by physical proximity, but rather by their role in population dynamics. This work presents strong evidence that the second approach could bear relevant fruit. As with any study anchored in data mining practices, ground truthing risk is an essential next step [63]. This is especially true in places where the social sciences are involved due to the interdisciplinary aspects of the problems faced by human communities. These efforts are costly, but are essential for solidifying risk factors as well as motivating the right policies and services for the population, the ultimate goal of these types of work.