More Powerful Couples: Urban Residence Choice for Advanced Degree Holders Across Demographic Characteristics

Abstract

Past studies of urbanization in the United States have found that college-educated (power) couples migrate to metropolitan areas at higher rates than both educated single persons and other couples with less education. A common explanation is that power couples receive disproportionately high co-location benefits from cities and therefore migrate to urban areas more than other groups. This study leverages the Public Use Microdata Sets to understand urban residence rates for couples between 2007 and 2022, with a focus on couples attaining degrees beyond the bachelor’s level. The data allows couples to be classified into nine different power status categories based on joint educational attainment, along with categories for their sexual orientation and racial composition. Results from fixed-effects logit models reinforce previous work showing that couples at the highest levels of educational attainment reside in urban areas at higher rates than their otherwise equal peers. Urban residence rates are even stronger for more advanced degree holders generally, though the effect is disproportionately higher for black and Asian couples, while it is disproportionately weaker for some same-sex couples. The added nuance of these results provides important insight into the relationship between urban residency, educational attainment, and demographic characteristics.

1. Introduction

The structure and dynamics of urban residence patterns have been popular topics of study in recent decades. Like other countries, many have found that the population in the United States demonstrates an increasing trend towards urban environments. Often termed “rural brain drain”, past work investigated the phenomenon of highly educated individuals migrating to urban areas and away from rural ones ([1,2,3,4,5,6,7,8], for some examples). This pattern of migration engenders concern. Highly educated individuals tend to earn more income than those with less education, which means that migration of the former implies a movement in earnings (and tax revenue) away from departed regions. In the case of prevalent urbanization, the risk is that rural areas, trapped in a vicious cycle of brain drain, will be unable to attract or retain labor market talent and therefore increasingly lag behind urban areas.

Alongside discussions of rural brain drain, Costa and Kahn [9], among others, provide evidence that at least some of these urbanization phenomena can be explained by the co-location desires of highly educated couples. They used U.S. Census data from the decades between 1940 and 1990 to show that the percentage of power couples—partnerships wherein both parties have at least bachelor’s degrees—increased over these years at a rate higher than would be expected from reference groups of highly educated single people. In other words, and as a simplification, if many choose to migrate to urban areas for economic and social benefits, there is something unique about being in a couple that provides additional benefits to settling in an urban environment. They term these additional benefits to couples co-location benefits, which are non-labor-market benefits that accrue to couples that are able to exist in the same location. They estimate the effect of co-location benefits on the probability of urbanization by using a triple difference-in-difference approach and, based on their theoretical model, argue that these couple-specific co-location benefits explain a sizable portion of urbanization patterns for college-educated couples. In other words, the benefits of co-location for highly educated couples surpass co-location benefits of couples with lower educational attainment. While not necessarily their conclusion, the results of Costa and Kahn [9] suggest that one way to prevent increased urban–rural inequality would be to increase work opportunities and amenities that educated couples enjoy in rural areas because this would increase the co-location benefits of locating in these regions.

As the percentage of the population that earns a bachelor’s degree (and beyond) in the United States has grown since the Costa and Kahn [9] study, the worry is that migration of the most educated at higher rates might accentuate the effects of brain drain. Given the granularity of recently available data, our primary research question is, do the previously found trends persist when looking specifically at couples that earn even more advanced degrees? To answer this question, we utilize the Public Use Microdata Sets (PUMS) for 2007, 2012, 2017, and 2022 to analyze the probability of urbanization based on educational attainment, but with more finite definitions of power status. The assembled data also allows us to simultaneously investigate the extent to which educationally based migration patterns may differ across demographic groups (e.g., race and sexual orientation). Finally, an auxiliary contribution to our work is that PUMS data are available annually and include a wide range of demographic, economic, and location information for individuals across the country. The hope is that this study models how use of this freely available data can lead to increased dissemination of timely information for policymakers grappling with urbanization trends.

After assembling the data, married individuals are compared with their spouses and categorized by their joint educational attainment (which we term powerStatus), their sexual orientation (copType), racial composition (raceStatus), and participation in the labor force (laborStatus, which is used to verify results from previous studies). Assigning married individuals to these categories using their individual characteristics allows for a difference-in-difference approach because we use otherwise equal single individuals as points of comparison. To estimate the difference-in-difference models, we employ both fixed-effects logit estimation techniques alongside weighted least squares with fixed effects.

Regardless of specification or estimation technique, the results are consistent and robust, with several key conclusions. First, even before the COVID-19 pandemic, the overall rate of urban residence has been decreasing since 2007. At first glance, this may seem to counter many concerns regarding rural brain drain as “work-from-home” options expanded. However, our results also show that the most highly educated couples continue to move to cities at disproportionately higher rates despite this overall trend, which suggests that concerns about rural brain drain’s effects are quite valid. A second key result is that while all couples with high power classification are more likely to live in cities, separating advanced degree holders demonstrates that the rate of urban residence escalates as joint educational attainment rises. This result is stronger when both partners participate in the labor force and is particularly strong for highly educated black couples. On the other hand, highly educated same-sex (male–male) couples are less likely to live in cities, particularly if only one partner has a doctorate or professional degree.

Collectively, this preview of results suggests that co-location benefits do differ across racial and sexual orientation categories. From the perspective of a particular urban area, this demographically driven result will be important to consider as policymakers look to evaluate local programs and plan for the future demographics of their region. With studies continuing to show the strong, positive link between diversity (both ethnic and intellectual) and economic growth, such an understanding is imperative for good regional policy [10,11,12,13].

The remaining sections of the paper are as follows. Section 2 introduces the most relevant literature for the present analysis, with special attention paid to the literature on migration patterns and urban residence across educational categories. Section 3 describes how our data source differs from those in previous work and provides some insights into the depth of prospective questions that PUMS could address in future work. Section 4 discusses our methodology and econometric approach, given the structure of the PUMS data. Section 5 discusses the results of our empirical estimations, with a focus on power classification even in discussions of the other couple classifications. Section 6 finishes with over-arching conclusions and possibilities for future work.

2. Relevant Studies on Migration Patterns and Co-Location Effects

Urbanization and migration are longstanding concerns in economic geography, with particular attention paid to the relationship between human capital and spatial mobility [14,15]. A prominent body of work explores how individuals with higher education disproportionately move toward larger urban centers—attracted by stronger labor markets, better cultural amenities, and improved public services [16,17,18,19]. This research builds on foundational models of regional growth and human capital spillovers [20]. Indeed, there is evidence that educated individuals tend to move to areas that offer higher returns to education, which are often found in urban settings [21,22,23]. Chen and Rosenthal [24] ask whether migration is more often motivated by job prospects or amenities—an important question when considering highly educated couples.

Within this broader literature, a growing number of studies are examining the spatial dynamics of highly educated couples. Costa and Kahn [9] identify “power couples”—dual-career households where both partners have at least bachelor’s degrees—who are more likely to reside in metropolitan areas than both educated singles and less-educated couples. They explain this via co-location benefits, whereby partners aim to match with two desirable jobs. Cities may disproportionately advantage couples that require joint job-matching [25,26]. Simon [27] further demonstrates how city size and human capital composition shape the career trajectories of power couples.

Additionally, recent scholarship has highlighted the rise of commuter couples—partners living apart due to employment constraints—as a complicating factor in residential decision-making [28]. A robust body of literature has emerged on dual-earner migration, highlighting the trade-offs couples face in co-locating while optimizing both careers [29,30,31,32]. Rusconi [33] provides a systematic framework for understanding these dynamics, emphasizing the interdependence of labor market constraints and gender norms. The interplay between gender and race in dual-earning households’ locational choices has been widely studied [34,35,36].

Other studies emphasize that individuals’ occupational aspirations influence their residential choices, especially in creative and highly skilled professions. Waddell [37] employs discrete-choice models to show how career goals shape housing decisions, and Woldoff et al. [38] highlight urban residential preferences among college students in creative majors.

At the same time, research on rural brain drain shows how out-migration of educated young adults can drain human capital and slow long-run economic development. Carr and Kefalas [3] document sustained talent loss from rural areas, and de Haas [39] argues that migration driven by professional aspirations deepens regional inequality. While many studies adopt a national or regional lens, there is growing interest in the demographic heterogeneity of these patterns. For instance, research finds that racial minorities and same-sex couples often face distinct considerations in migration and residential sorting due to historical, social, or legal constraints [40]. These studies are important because several find significant macroeconomic effects of population flows away from rural areas [7,8,41].

Finally, while prior work on migration often emphasizes brain drain and the dynamics of population flows, our analysis does not rely on one-year migration indicators such as MIGPUMA available in PUMS. As an important point of clarity, rather than tracking active migration events, we examine the residential patterns of highly educated couples as observed at the time of the survey. This work is most closely related to that of Costa and Kahn [9] and Compton and Pollak [42], which analyze urban residence patterns of college-educated (power) couples. Compton and Pollak [42] extend the work of Costa and Kahn [9] to add new couple delineations (e.g., couples where one person has a bachelor’s degree and the other does not) but find that the gender of the college-educated spouse matters. The explicit inclusion of gender, race, and variants of power status they include serve as inspiration for this work.

Our aim is to identify how characteristics such as joint educational attainment, race, and couple type correlate with a household’s current urban or rural residence. In this way, we contribute to the literature by offering a post hoc view of locational outcomes for educationally advantaged households, complementing studies that directly model migration behavior.

3. Data Sources and Construction

To examine how urban residence varies by demographic group, we primarily utilize the Public Use Microdata Sets (PUMS), published annually from the American Community Survey (ACS) of the U.S. Census Bureau. These publicly available data provide detailed variables on individual and household characteristics, including educational attainment, demographic composition, and geographic location. Each year’s data are weighted to ensure national representativeness and permit construction of robust population-level estimates.

We utilize four years of PUMS data—2007, 2012, 2017, and 2022—which coincide with the quinquennial Economic Census. These years were chosen to span a wide enough window to analyze urban residence patterns while remaining manageable within our computational limits. All years except 2007 are drawn from the 5-year PUMS; 2007 uses a 3-year sample, as that was the most granular version available. Data preparation, visualization, and regression analysis required over 200 GB of RAM, which constrained our temporal scope.

Our analysis centers on the person-level PUMS files, allowing us to construct household-level classifications based on joint characteristics. Variables used are listed in

Table 1. Primary variables used from the PUMS data.

Variable	Description	Use
MAR	Marital Status	Used to identify individuals as married or unmarried
PUMA	FIPS PUMA Code	Used to identify PUMA (location) of individual
RAC1P	Racial Category	Used to identify self-reported race of individual and partner
SCHL	Educational Attainment	Used to identify highest degree earned for person and their partner
SERIALNO	Household Identifier	Used to pair potential partners living in the same home
SEX	Sex of Person	Used to identify couple type (e.g., heterosexual, same sex)
ST	FIPS State Code	Used to identify state (location) of individual

, including those needed to define educational, racial, labor force, and couple-type groupings.

We primarily utilize person records from PUMS rather than household records, as this allows us to build customized household-level categories based on individual attributes. Our first step uses the entire person-level file for each year to sort households into categories that define their marital status, racial composition, and educational classification. For clarity, we assign each of these constructed variables an italicized label. These labels are referenced frequently throughout the statistical analysis and are defined in terms of (1) joint educational attainment (powerStatus); (2) couple type (copType); and (3) racial composition (raceStatus).

Marital status is determined using the MAR variable. We classify individuals as married only if there is exactly one other married individual in the same household (as defined by SERIALNO). All other individuals are assigned the category single. This is because some households contain more than two married individuals (e.g., multi-family homes) or report married status while living alone. In the former case, the PUMS data do not allow reliable identification of pairings. In the latter case, the assumption is that a co-location decision is unlikely to apply when partners are not cohabitating. Individuals coded as single are not assigned household-level demographic classifications, as they are not assumed to have made joint location decisions.

For married couples, we then use educational attainment (SCHL) to classify power status. Following conventions in the literature, couples where neither partner completed education beyond high school are labeled powerLeast. Those where both partners completed some college but have no bachelor’s degree are labeled powerLow. For couples where both partners earned a bachelor’s degree or higher, we distinguish between power, powerMore, and powerMost, corresponding to bachelor’s, master’s, and doctoral/professional degrees, respectively.

Mixed-education couples are also categorized based on the highest education attained by either partner. These categories include partPowerLow, partPower, partPowerMore, and partPowerMost, depending on the level completed by the more educated partner.

Couple type (copType) is defined using MAR and SEX to distinguish between female–female (FF), female–male (FM), and male–male (MM) pairings. Racial composition is derived from the RAC1P variable for each partner and used to assign couples to the categories white, black, asian, mixed, or other, as described in

Table 2. Possible categories for individuals and couples.

Characteristic	Single Individuals		Couples
	Variable(s) Used	Possible Values	Variable(s) Used	Possible Values
Power Designation (powerStatus)	SCHL	Varies by PUMS year	SCHL MAR	powerLeast partPowerLow partPower partPowerMore partPowerMost powerLow power powerMore powerMost
Couple Type (copType)	MAR	single	MAR SEX	female–female (FF) female–male (FM) male–male (MM)
Race (raceStatus)	RAC1P	White Black/African American American Indian Alaska Native American Indian/Alaska Native Asian Native Hawaiian/Pacific Islander Some other race Two or more major races	RAC1P MAR	asian black mixed other white
Labor Force Status (laborStatus)	ESR WORKSTAT OCCP	Codes indicating labor force status Code for individual’s occupation (jobTag)	MAR FES OCCP	bothLF oneLF noLF

Italicized values represent how variables are stored, and these terms are used frequently throughout the rest of the paper. Power designations are ordered by educational attainment, from least to most. powerLeast is assigned to couples for which both partners have attained no more than a high school education; partPowerLow is assigned to couples for which one partner has attended some college, but not yet earned a bachelor’s degree; powerLow is analogous, but assigned when both partners have completed some college, and so on. For multi-family and solo married households, we argue that the co-location effect is relatively weak as compared to single-family households. For a married person living alone (without a partner in the house), the rationale behind omitting these groupings is strong. Presumably, any co-location effect would be weaker for such couples as compared to those that cohabitate. Couples are defined as asian, black, or white if both members of the couple fit that racial category. mixed is assigned to couples that self-identify as split race couples (e.g., one individual is black and one is white) and those in which both self-identify as “Two or more major races" in the PUMS survey. other represents all other couples that do not fit in the other categories and is saved for prospective future work. Labor force status assigns a code to each married couple indicating whether both were labor force participants (bothLF), only one was a labor force participant (oneLF), or neither was a labor force participant (noLF). This allows this study to match previous work that focuses on dual-income-earning households while investigating whether the results hold for others.

. We restrict the number of categories analyzed, particularly for racial composition of the couple, to preserve statistical power and interpretability.

Finally, labor force status (laborStatus) is assigned using the household records (FES, OCCP). Couples are categorized as bothLF, oneLF, or noLF, depending on whether both, one, or neither partner is in the labor force. These distinctions allow our analysis to investigate how urban residence patterns differ across couples by employment structure. This addition of laborStatus is motivated by prior work, such as Costa and Kahn [9], which found that co-location benefits tend to be stronger among dual-income households. When focusing on such households, we further incorporate occupational information using the OCCP variable from PUMS. While these occupation codes are numerous and detailed, we simplify them by tagging each occupation with a three-letter prefix—such as ‘MED’ for medical fields or ‘BUS’ for business-related occupations—to facilitate analysis. These tags are stored in a derived variable, jobTag, and later used to control for occupation-specific location patterns in couples where both partners are in the labor force.

Table 2 visually summarizes these classification schemes. A simplified diagrammatic version is also provided in Figure 1.

After constructing the full set of demographic identifiers for individuals and couples, the final step is to determine whether a given respondent lives in an urban environment. Each PUMS record includes both state (ST) and Public Use Microdata Area (PUMA) codes. Because PUMAs are not inherently labeled as urban or rural and may span multiple community types, we use a crosswalk from the Missouri Census Data Center’s Geocorr tool to link PUMAs to Metropolitan Statistical Areas (MSAs) [43].

This crosswalk provides two useful metrics: a proportion-based measure of urban classification for each PUMA (termed urbPro) and a binary indicator of whether a PUMA is at least 70% urban (termed urb). For example, DeKalb County, Georgia, a largely urban area, has an urbPro of 0.989, while Newton County, Georgia—more suburban in character—has an urbPro of 0.763. These measures allow us to distinguish not only whether an individual lives in a city but also the degree of urbanicity of their surrounding area.

We argue that urbPro, the continuous measure, offers a more nuanced depiction of urban residence patterns, capturing gradations that are particularly important near metropolitan boundaries. However, we also employ the binary urb measure for robustness checks. For robustness, we tried alternative definitions that defined urban areas using 75% and 80% as the delineation, but neither meaningfully changed the results. Across specifications, our findings are generally consistent, though the magnitude of coefficients can vary slightly depending on the definition used.

Figure 2 presents maps of both the proportion-based and binary urban classification measures for 2022. These visualizations underscore the regional variation in urbanicity and demonstrate the advantage of using urbPro to account for partial urban exposure, especially along the Eastern Seaboard and in fast-growing areas.

Figure 2a reveals that urbanicity levels vary considerably across the United States, with the most urban areas concentrated along the coasts and in parts of Texas, Florida, and the Great Lakes region. Meanwhile, large portions of the Midwest and Mountain West remain more rural or suburban. By contrast, Figure 2b uses a binary threshold (70% or more urban) and results in a more fragmented representation, highlighting isolated “urban” pockets even in predominantly rural areas. In Figure 2b, a value of 1 denotes an urban area, and 0 denotes a rural area.

This comparison illustrates a key limitation of binary urban–rural classifications: they can obscure meaningful differences within and across geographic units. For instance, communities on the fringe of large metropolitan areas may fall just below the threshold yet share many characteristics with their urban neighbors. In contrast, using the continuous measure urbPro preserves this gradient and more accurately captures exposure to urban environments. We therefore rely on urbPro for our main empirical analysis while reporting robustness checks using urb.

Based on these definitions of urban areas,

Table 3. Select urban residence rate summary statistics for 2007 to 2022—grouped by power category.

Category	Variable	2007		2012		2017		2022
		n	Mean (sd)	n	Mean (sd)	n	Mean (sd)	n	Mean (sd)
powerLeast	urb	1,032,208	48.96 (49.99)	295,536	51.96 (49.96)	1,351,934	52.78 (49.92)	157,426	49.55 (50.00)
	urbPro	1,032,208	68.19 (27.31)	295,536	70.01 (27.61)	1,351,934	70.49 (27.53)	157,426	68.80 (27.62)
partPower	urb	204,982	60.64 (48.85)	66,604	62.33 (48.46)	354,552	63.47 (48.15)	51,330	57.11 (49.49)
	urbPro	204,982	74.90 (25.43)	66,604	75.93 (25.75)	354,552	76.58 (25.51)	51,330	73.41 (26.55)
power	urb	285,918	72.33 (44.73)	98,076	75.24 (43.16)	529,668	75.79 (42.84)	75,542	69.33 (46.11)
	urbPro	285,918	81.46 (22.51)	98,076	83.32 (22.42)	529,668	83.57 (22.17)	75,542	80.47 (24.19)
powerMost	urb	34,418	82.18 (38.27)	12,926	84.22 (36.46)	75,952	84.11 (36.56)	10,938	79.30 (40.52)
	urbPro	34,418	87.34 (19.43)	12,926	88.45 (19.45)	75,952	88.56 (19.26)	10,938	86.25 (21.40)

n represents the number of observations (people that are part of a couple) that match this categorical definition in the sample. mean and standard deviation (sd) represent summary statistics for urban residence rates for each demographic, based on how urban status is defined.

,

Table 4. Select urban residence rate summary statistics for 2007 to 2022—grouped by couple sexual orientation.

Category	Variable	2007		2012		2017		2022
		n	Mean (sd)	n	Mean (sd)	n	Mean (sd)	n	Mean (sd)
Heterosexual (FM)	urb	3,720,536	58.89 (49.20)	1,219,180	61.81 (48.59)	6,133,590	62.82 (48.33)	800,502	58.14 (49.33)
	urbPro	3,720,536	73.88 (25.90)	1,219,180	75.62 (26.08)	6,133,590	76.18 (25.86)	800,502	73.88 (26.61)
Same-sex (FF)	urb	2078	75.26 (43.16)	736	78.26 (41.28)	28,362	71.55 (45.12)	5158	68.24 (46.56)
	urbPro	2078	83.26 (23.65)	736	85.85 (21.49)	28,362	81.47 (24.29)	5158	79.94 (24.98)
Same-sex (MM)	urb	4340	76.54 (42.38)	926	80.56 (39.59)	27,654	77.12 (42.01)	4940	76.23 (42.57)
	urbPro	4340	84.49 (23.39)	926	86.59 (22.11)	27,654	84.73 (22.55)	4940	84.69 (22.84)

n represents the number of observations (people that are a part of a couple) that match this categorical definition in the sample. mean and standard deviation (sd) represent summary statistics for urban residence rates for each demographic, based on how urban status is defined.

and

Table 5. Select urban residence rate summary statistics for 2007 to 2022—grouped by couple racial status.

Category	Variable	2007		2012		2017		2022
		n	Mean (sd)	n	Mean (sd)	n	Mean (sd)	n	Mean (sd)
asian	urb	125,332	93.37 (24.87)	48,506	95.28 (21.20)	264,544	95.53 (20.67)	37,940	93.78 (24.15)
	urbPro	125,332	94.32 (12.55)	48,506	95.71 (11.46)	264,544	95.80 (10.96)	37,940	94.86 (12.57)
black	urb	189,864	73.52 (44.12)	65,810	77.41 (41.82)	318,678	77.75 (41.59)	34,068	71.50 (45.14)
	urbPro	189,864	82.46 (24.45)	65,810	84.35 (23.72)	318,678	84.52 (23.42)	34,068	81.48 (24.87)
mixed	urb	201,134	71.38 (45.20)	70,124	74.46 (43.61)	388,494	75.46 (43.03)	97,254	69.85 (45.89)
	urbPro	201,134	81.08 (23.27)	70,124	82.89 (23.22)	388,494	83.45 (22.95)	97,254	80.85 (24.08)
white	urb	3,081,674	54.87 (49.76)	1,001,136	57.76 (49.39)	5,035,466	58.71 (49.23)	568,834	50.42 (50.00)
	urbPro	3,081,674	71.54 (26.03)	1,001,136	73.24 (26.31)	5,035,466	73.77 (26.14)	568,834	69.34 (26.87)

n represents the number of observations (people that are a part of a couple) that match this categorical definition in the sample. mean and standard deviation (sd) represent summary statistics for urban residence rates for each demographic, based on how urban status is defined. A person receives the tag asian only if their spouse also has the ‘Asian’ code for RAC1P in the PUMS data. The same goes for the other displayed categories except for mixed—which is a tag assigned to an individual that is married to a person with a different racial code.

provide select summary statistics across power status, sexual orientation, and couple racial composition, respectively. For each table, we utilize the psych package in R to group summary statistics by category [44]. Observed differences in mean urban residence patterns across categories help motivate their inclusion in the analysis.

In each table, the variable urb represents a binary definition of urban location based on the 70% threshold from the MCDC, while urbPro reflects the continuous urban residence proportion from Geocorr. As expected, urb produces higher standard deviations since it is a dichotomous measure, whereas urbPro allows for variation across partially urban areas.

Several elements from Table 3 are noteworthy. For brevity, we report only a subset of powerStatus categories here. A full set of powerStatus summary statistics can be found in Appendix A. The summary confirms a pattern where more educated couples are more likely to reside in urban areas. For instance, in 2007, approximately 87% of couples where both partners held doctoral or professional degrees lived in urban areas, compared to roughly 68% of couples with no more than a high school diploma. This gradient is consistent across years and definitions of urban status.

That said, Table 3 also shows a slight decline in urban residence rates between 2017 and 2022 across most education groups. One possible explanation is the increased prevalence of remote work, particularly after 2020 and the impacts of the pandemic. For context, 15.2% of all workers in the PUMS data reported working from home in 2022, up from 5.4% in 2017 [45]. However, we refrain from interpreting this decline causally, as identifying the drivers of such shifts is beyond the scope of this analysis.

Turning to Table 4, we observe that same-sex couples (both male–male and female–female) consistently report higher urban residence patterns than heterosexual couples across all survey years. While the gap persists over time, we again see a modest decline in average urban residence for all groups in the 2022 sample.

A similar pattern holds in Table 5. Couples where both partners identify as Asian are much more likely to reside in urban areas, while white couples are the least urbanized on average. These differences are comparable in magnitude to those seen between low- and high-education couples, suggesting that racial composition plays a key role in geographic sorting.

Across all tables, observed differences in mean urban residence patterns by education, couple type, and race provide support for including each dimension in our empirical models. Results from one-way ANOVA tests (available upon request) confirm that these group differences are statistically significant.

4. Methods Used to Analyze Urban Residence Patterns

This study leverages individual- and couple-level information from PUMS to estimate a model of urban residence patterns using a difference-in-difference framework. This approach builds on methodologies employed in related work, particularly Costa and Kahn [9] and Compton and Pollak [42].

Our basic setup considers individual i in geographic location j and year t. Let $U_{i,j,t}$ denote the urban residence outcome, measured either as the binary indicator urb or as the continuous urban share urbPro, both defined earlier. Because urban status is inherently tied to geography, U is largely a function of j. While geographic boundaries may evolve over time, these changes are minimal across the four cross-sections used in this study. Previous research suggests that the choice to live in an urban environment could stem from some combination of economic opportunity, social prospects, urban amenities available, and/or co-location benefits—each of which could accrue to any individual i, though only co-location benefits accrue to those in a couple. The individual’s choice of location j and, by extension, the degree of urban environment could then be expressed in the baseline model of Equation (1).

$$U_{i,j,t}=\alpha +{\beta }_1\ast I_{i,j,t}+{\beta }_2\ast C_{i,j,t}+{\varphi }_j+{\omega }_t+{ϵ}_{i,j,t}$$

(1)

In this specification, $I_{i,j,t}$ is a vector of individual-level controls (e.g., age, education, sex, race), and $C_{i,j,t}$ includes indicators for couple-level constructed categories: powerStatus, copType, and raceStatus. For unmarried individuals, we assign each of these variables a baseline value of single, providing a common reference group. The coefficient vector ${\beta }_2$ therefore captures the differential association between couple-level characteristics and urban residence relative to demographically similar singles.

To account for contextual variation, the model includes state and PUMA fixed effects (${\varphi }_j$) and year fixed effects (${\omega }_t$). Because individuals may also be attracted to a region for its size, we include the log of total population for each individual’s PUMA. This inclusion in $I_{i,j,t}$ accounts for population growth to a region and therefore proxies for how dynamic the region is. The term ${ϵ}_{i,j,t}$ captures individual-level unobserved heterogeneity. With this setup, ${\beta }_2$ can be interpreted as a difference-in-difference estimator, quantifying how the probability of urban residence differs between single individuals and couples with equivalent individual characteristics.

Equation (1) represents the primary estimating function, but we also estimate models that explicitly include laborStatus in order to see if dual-income couples urbanize differently. This alternative specification includes occupational fixed effects and restricts the sample to working individuals, as shown in Equation (2). In this version, we include only those with reported occupations, comparing singles with couples where both partners participate in the labor force (laborStatus = bothLF).

$$U_{i,j,t}=\alpha +{\beta }_1\ast D_{i,j,t}+{\beta }_2\ast C_{i,j,t}+{\varphi }_j+{\omega }_t+{\Phi }_o+{ϵ}_{i,j,t}$$

(2)

In Equation (2), ${\Phi }_o$ denotes occupation-specific fixed effects based on the jobTag grouping described earlier. This specification allows us to explore whether co-location benefits vary across occupational contexts. In estimations of both Equations (1) and (2), fixed effects play a crucial role in isolating within-location and within-year comparisons. We return to this point later, using visualization to demonstrate the explanatory power of these fixed effects.

We estimate Equations (1) and (2) using the fixest package in R, which provides a robust and flexible framework for fixed-effects modeling [46]. This package supports the inclusion of high-dimensional fixed effects, clustered standard errors, and analytic weights—features necessary for working with Public Use Microdata Sample (PUMS) data. All models apply person-level weights provided in the PUMS dataset, and standard errors are clustered by powerStatus to account for within-group correlation. Importantly, robustness checks using alternative clustering levels yielded qualitatively similar results.

Since the dependent variable is bound between 0 (completely rural) and 1 (completely urban), even in cases when urbPro is used, binomial logit models are our preferred estimates. To verify robustness, we also report weighted least-squares estimates alongside the initial models. These results are largely consistent with the logit specifications, so subsequent analyses focus exclusively on the latter.

Because the variables powerStatus, copType, and raceStatus are constructed jointly, including them simultaneously in a single model would create interpretation challenges due to overlapping subgroup definitions. Therefore, our modeling proceeds in a staged fashion. Initial models estimate differences in urban residence patterns across powerStatus categories while controlling for basic individual characteristics, including age, sex, and race (as described by Equation (1)). We then refine the analysis by restricting the sample to the employed population—specifically, singles with reported occupations and couples where both partners are in the labor force (laborStatus = bothLF). This approach centers on co-location decisions among dual-earner households and their single peers, reducing potential confounding from job-seeking behavior alone (as described by Equation (2)).

To assess whether urban residence patterns vary across sexual orientation and racial composition, we stratify the data by copType and raceStatus, respectively. In these stratified models, we continue to include single individuals as a comparison group, allowing us to interpret the effects of powerStatus within each subgroup context. For instance, one model compares urban residence patterns of FM couples to their single peers, another does the same for FF couples, and so on. The same structure is applied for each racial composition group.

The next section presents estimation results for these models, along with commentary on statistical significance and patterns of interest.

5. Fixed-Effects Results from Logit and Weighted Least Squares Models

To facilitate the dissemination of results, this section is divided into subsections that focus on urban residence outcomes by powerStatus, copType, and raceStatus, in that order.

5.1. Results for powerStatus Designations

Table 6. Fixed-effects regression results for differential urban residence rates by power status.

Dependent Variables:	urbPro		urb		urbPro	urb
Model:	(1)	(2)	(3)	(4)	(5)	(6)
	WLS	Logit	WLS	Logit	Logit	Logit
Variables
powerLeast	−0.025 ***	−0.157 ***	−0.041 ***	−0.368 ***	−0.113 ***	−0.273 ***
	(0.001)	(0.011)	(0.002)	(0.025)	(0.009)	(0.024)
partPower	−0.011 ***	−0.087 ***	−0.017 ***	−0.190 ***	−0.071 ***	−0.167 ***
	(0.001)	(0.006)	(0.002)	(0.010)	(0.023)	(0.052)
partPowerLow	−0.023 ***	−0.155 ***	−0.038 ***	−0.348 ***	−0.136 ***	−0.325 ***
	(0.001)	(0.002)	(0.001)	(0.005)	(0.017)	(0.041)
partPowerMore	−0.015 ***	−0.123 ***	−0.022 ***	−0.258 ***	−0.109 ***	−0.240 ***
	(0.001)	(0.008)	(0.002)	(0.014)	(0.025)	(0.057)
partPowerMost	−0.009 ***	−0.082 ***	−0.012 ***	−0.151 ***	−0.088 ***	−0.178 ***
	(0.002)	(0.012)	(0.002)	(0.022)	(0.028)	(0.063)
power	0.003	0.018	0.009 **	0.067 *	0.025	0.074
	(0.003)	(0.022)	(0.004)	(0.039)	(0.038)	(0.078)
powerLow	−0.007 **	−0.071 ***	−0.010 **	−0.140 ***	−0.050	−0.104
	(0.002)	(0.015)	(0.003)	(0.027)	(0.030)	(0.066)
powerMore	0.005	0.043	0.013 **	0.113 **	0.072	0.167 *
	(0.003)	(0.029)	(0.005)	(0.051)	(0.047)	(0.093)
powerMost	0.010 **	0.116 ***	0.022 ***	0.272 ***	0.144 ***	0.342 ***
	(0.004)	(0.035)	(0.006)	(0.063)	(0.052)	(0.102)
Fixed Effects and Controls
State	Yes	Yes	Yes	Yes	Yes	Yes
PUMA	Yes	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes	Yes
PUMA Population Control	Yes	Yes	Yes	Yes	Yes	Yes
Occupational Code					Yes	Yes
Fit Statistics
Observations	23,280,874	20,789,576	23,280,874	14,047,749	12,204,388	8,167,911
Squared Correlation	0.67630	0.66293	0.60322	0.42453	0.65613	0.42383
Pseudo R2	0.36805	0.78461	0.98535	0.34284	0.78486	0.34120

Coefficients are interpreted as percentages. For the coefficient on powerMost from the first column, the coefficient would imply that two married partners with doctoral or professional degrees are 1% more likely to live in urban environments than their otherwise equal peers. Controls included are age, sex, educational attainment, and racial code, though those coefficients are not included in the table. Standard errors are clustered by power status. Significance Codes are ***: 0.01; **: 0.05; *: 0.1.

provides results for urban residence probability with respect to powerStatus. The coefficients on controls (age, individual race, etc.) are not reported to remain focused on the primary results. They are available upon request for this and future regression tables.

Columns (1) through (4) in Table 6 present results from both weighted least squares (WLS) and binomial logit (Logit) models across the two definitions of urban status described earlier. Column (2), which uses the proportional urban classification with a Logit estimator, serves as our preferred specification. Here, we find that couples jointly holding bachelor’s degrees (power category) are not statistically more likely to reside in an urban area compared to otherwise observationally equivalent single individuals. Interestingly, under a binary definition of urban-ness, power couples are more likely to reside in urban areas, as shown in Column (3) and Column (4). This highlights the skew that such a binary definition might have but matches previous work on couples with this educational background.

The main research question guiding this paper is whether urban residence probabilities increase for couples with even higher educational attainment. Crucially, the results provide some weak evidence that the likelihood of living in an urban area increases with even higher levels of joint educational attainment. Couples classified as powerMost—where both individuals hold doctoral or professional degrees—are associated with an 11.6% increased probability of urban residency.

This steep educational gradient in urban living supports the idea that advanced degree holders receive amplified benefits from cities, whether due to labor market matching, professional opportunities, or cultural amenities. These findings offer quantitative support for the anecdotal and theoretical concern over location patterns of the most educated by showing that the most educated couples continue to concentrate in urban areas. The resulting spatial concentration of human capital may contribute to widening gaps in opportunity and growth between urban and rural areas.

A particularly noteworthy finding is that all forms of partPower couples (e.g., partPowerLow, partPowerMore) are significantly less likely to live in urban areas compared to their otherwise equal single peers. This contrasts with prior assumptions in the literature that suggested mixed-education couples were more urbanized than non-coupled individuals [42]. While these couples do exhibit higher (relative) urban residence patterns than powerLeast couples (in which both individuals have only a high school degree), their lower rates relative to singles indicate that prior work may overstate the urban concentration of partially educated couples.

Columns (5) and (6) estimate the same models on a restricted sample of working individuals—defined as single persons with a reported occupation and couples in which both members are in the labor force (bothLF). These models may better isolate non-economic motivations for location choice by holding employment constant. In the preferred logit specification in column (5), powerMost couples that are dual-income earners are about 14.4% more likely to live in urban environments than their single peers. Interestingly, the estimated coefficients for all mid-level power categories (power, powerLow, powerMore) are statistically insignificant in this subsample. This suggests that the earlier effects observed in column (2) may have been primarily driven by employment opportunity, not co-location benefits alone.

5.2. Results for powerStatus Designations, Stratified by Sexual Orientation (copType)

With a baseline understanding of results stemming from powerStatus, we next stratify the sample by copType to assess whether couples of different sexual orientations exhibit different urban residence patterns. These models compare couples with the same educational attainment but different sexual orientations against otherwise equal single peers. As with prior specifications, columns (2) and (5) from Table 6 serve as the reference point because these specifications utilize logit models with urbPro as the dependent variable. The corresponding stratified results are displayed in

Table 7. Fixed-effects regression results for differential urban residence rates by sexual orientation (copType).

Dependent Variable:	urbPro
Model:	(1)	(2)	(3)	(4)	(5)	(6)
	FF	FM	MM	FF	FM	MM
Variables
powerLeast	−0.161 ***	−0.004 ***	0.141 ***	−0.165 ***	0.010 ***	0.195 ***
	(0.013)	(0.000)	(0.000)	(0.011)	(0.000)	(0.000)
partPowerLow	−0.158 ***	0.002 ***	0.028 ***	−0.178 ***	−0.019 ***	0.002 ***
	(0.004)	(0.000)	(0.000)	(0.003)	(0.000)	(0.000)
partPower	−0.091 ***	0.004 ***	0.113 ***	−0.122 ***	−0.053 ***	0.027 ***
	(0.004)	(0.000)	(0.000)	(0.004)	(0.000)	(0.000)
partPowerMore	−0.127 ***	−0.131 ***	0.113 ***	−0.160 ***	−0.140 ***	0.064 ***
	(0.006)	(0.000)	(0.000)	(0.007)	(0.000)	(0.000)
partPowerMost	−0.085 ***	0.305 ***	−0.293 ***	−0.139 ***	0.306 ***	−0.322 ***
	(0.010)	(0.000)	(0.000)	(0.010)	(0.000)	(0.000)
power	0.015	−0.022 ***	−0.005 ***	−0.026	−0.074 ***	−0.067 ***
	(0.019)	(0.000)	(0.000)	(0.020)	(0.000)	(0.000)
powerLow	−0.074 ***	−0.067 ***	0.016 ***	−0.102 ***	−0.081 ***	−0.014 ***
	(0.012)	(0.000)	(0.000)	(0.012)	(0.000)	(0.000)
powerMore	0.040	−0.048 ***	−0.055 ***	0.021	−0.077 ***	−0.131 ***
	(0.026)	(0.000)	(0.000)	(0.029)	(0.000)	(0.000)
powerMost	0.114 ***	−0.067 ***	−0.186 ***	0.094 ***	−0.057 ***	−0.215 ***
	(0.033)	(0.000)	(0.000)	(0.034)	(0.000)	(0.001)
Fixed Effects and Controls
State	Yes	Yes	Yes	Yes	Yes	Yes
PUMA	Yes	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes	Yes
PUMA Population Control	Yes	Yes	Yes	Yes	Yes	Yes
Occupational Code				Yes	Yes	Yes
Fit Statistics
Observations	20,728,285	9,908,982	9,908,851	12,164,483	6,849,260	6,849,167
Squared Correlation	0.66277	0.66111	0.66116	0.65599	0.65371	0.65377
Pseudo R2	0.78462	0.78389	0.78390	0.78487	0.78378	0.78380

Coefficients are interpreted as percentages. All models reported are estimated using binomial logit regressions. For the coefficient on powerMost from the first column, the coefficient would imply that two female married partners with doctoral or professional degrees are 11.4% more likely to live in urban environments than their otherwise equal peers. Controls included are age, sex, educational attainment, and racial code, though those coefficients are not included in the table. The last three columns isolate couples wherein both partners reported being in the labor force and compares them with their otherwise equal single peers. In these specifications, controls for the occupation are included. Standard errors are clustered by power status. Significance Codes are ***: 0.01; **: 0.05; *: 0.1.

.

Looking at the first three columns of Table 7, a number of findings from the baseline powerStatus results carry over. Most notably, powerMost couples remain more likely to live in urban areas than their single peers, but for the majority of couple definitions, other powerStatus was associated with lower rates of urban residence. In addition, new differences emerge when results are disaggregated by sexual orientation, including notable variation in the magnitude of key coefficients.

For instance, while powerLeast couples were uniformly less likely to live in urban environments in Table 6, this is no longer the case. FF and FM couples with this status still urbanize at lower rates, but MM couples are more likely than their single counterparts to reside in urban areas. Similar divergence appears across the partPower designations, with no consistent pattern emerging across sexual orientations. Without deeper qualitative or survey-based insight, we cannot fully explain these patterns, but they underscore the need to account for couple-level heterogeneity. Put simply, otherwise equal couples may make different location decisions based on their sexual orientation.

Further evidence for this comes from columns (4) through (6), which restrict the sample to individuals in the labor force and include occupational fixed effects. Relative to the first three columns, coefficients generally increase in magnitude. This suggests that employment opportunities, rather than co-location preferences alone, may explain a significant portion of urban residence. Still, differences persist: for example, MM, partPower couples remain more likely to live in urban areas, while FM and FF couples are less likely to do so. Among same-sex couples with partPowerMost status, urban residence rates are lower even after accounting for occupation.

An additional takeaway from Table 7 is the consistently high $R^2$ values and strong statistical significance of most coefficients. The high $R^2$ suggests that our models can reasonably predict urban residence for each group. Likewise, the quantity of statistically significant coefficients reinforces the conclusion that copType is meaningfully linked with the probability of urban residence. Future research would benefit from unpacking these dynamics more fully, especially the mechanisms that differentiate same-sex and different-sex couple decisions conditional on employment status.

5.3. Results for powerStatus Designations, Stratified by Couple Racial Status (raceStatus)

The final set of regression results stratifies the sample based on couple racial composition. This allows us to investigate whether couples with similar educational attainment but different racial pairings exhibit distinctive urban residence patterns. As with prior analyses, we compare each couple type with otherwise similar single individuals. The results are shown in

Table 8. Fixed-effects regression results for differential urban residence rates by race status (raceStatus).

Dependent Variable:	urbPro
Model:	asian	black	mixed	white	asian	black	mixed	white
Variables
powerLeast	0.430 ***	0.095 ***	−0.133 ***	−0.199 ***	0.393 ***	0.109 ***	−0.102 ***	−0.198 ***
	(0.003)	(0.007)	(0.010)	(0.010)	(0.003)	(0.006)	(0.011)	(0.010)
partPowerLow	0.313 ***	0.174 ***	−0.112 ***	−0.186 ***	0.299 ***	0.161 ***	−0.098 ***	−0.211 ***
	(0.001)	(0.006)	(0.009)	(0.007)	(0.001)	(0.006)	(0.012)	(0.005)
partPower	0.365 ***	0.223 ***	−0.062 ***	−0.110 ***	0.329 ***	0.194 ***	−0.091 ***	−0.154 ***
	(0.002)	(0.005)	(0.009)	(0.005)	(0.002)	(0.006)	(0.012)	(0.002)
partPowerMore	0.324 ***	0.176 ***	−0.087 ***	−0.148 ***	0.261 ***	0.168 ***	−0.120 ***	−0.200 ***
	(0.003)	(0.006)	(0.009)	(0.005)	(0.003)	(0.007)	(0.012)	(0.004)
partPowerMost	0.229 ***	0.178 ***	0.002	−0.095 ***	0.059 ***	0.167 ***	−0.079 ***	−0.165 ***
	(0.001)	(0.006)	(0.009)	(0.005)	(0.000)	(0.008)	(0.012)	(0.007)
powerLow	0.262 ***	0.237 ***	−0.049 ***	−0.080 ***	0.178 ***	0.195 ***	−0.077 ***	−0.130 ***
	(0.001)	(0.005)	(0.009)	(0.003)	(0.000)	(0.007)	(0.013)	(0.008)
power	0.323 ***	0.279 ***	0.038 ***	0.016 ***	0.216 ***	0.212 ***	−0.031 **	−0.050 ***
	(0.002)	(0.004)	(0.009)	(0.005)	(0.001)	(0.007)	(0.014)	(0.015)
powerMore	0.257 ***	0.264 ***	0.071 ***	0.044 ***	0.123 ***	0.188 ***	−0.010	−0.006
	(0.001)	(0.004)	(0.009)	(0.007)	(0.002)	(0.008)	(0.014)	(0.024)
powerMost	0.147 ***	0.240 ***	0.122 ***	0.146 ***	0.039 ***	0.177 ***	0.033 **	0.086 ***
	(0.000)	(0.003)	(0.009)	(0.010)	(0.001)	(0.005)	(0.015)	(0.030)
Fixed Effects and Controls
State	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
PUMA	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
PUMA Population Control	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Occupational Code					Yes	Yes	Yes	Yes
Fit Statistics
Observations	15,214,227	15,369,645	15,509,493	23,885,857	7,316,751	7,389,236	7,498,483	11,525,695
Squared Correlation	0.66111	0.65798	0.65836	0.65848	0.65607	0.65313	0.65344	0.65308
Pseudo R2	0.78541	0.78475	0.78491	0.78561	0.78410	0.78362	0.78361	0.78521

Coefficients are interpreted as percentages. For the coefficient on powerMost from the first column, the coefficient would imply that two married partners with doctoral or professional degrees are 0.16% more likely to live in urban environments than their otherwise equal peers. Controls included are age, sex, educational attainment, and racial code, though those coefficients are not included in the table. The last for columns restrict the sample to those that are employed. In these specifications, controls for the occupation are included. Standard errors are clustered by power status. Significance Codes are ***: 0.01; **: 0.05; *: 0.1.

. The consistently high $R^2$ values suggest a strong ability to predict urban residency rates in these specifications.

Across all models, a consistent theme emerges: couples’ racial composition plays a substantial role in explaining variation in urban residence probability. In general, asian and black couples are significantly more likely to live in urban areas than otherwise equal single individuals, while white and mixed couples are significantly less likely to do so. These patterns persist even when the sample is restricted to those in the labor force and when occupation is controlled for. This suggests that racial composition is related to residential choice beyond job-related considerations.

Unlike previous results, the baseline coefficient on powerMost couples is no longer consistently highest for all stratified samples. Instead, powerMost status is only associated with the highest probability of urban residency for mixed and white couples, suggesting an amplified urban residence probability at the top end of the education spectrum only for these groups. Unlike asian and black couples, mixed and white couples display statistically significant relationships between partPower and power status that are mirrored in Table 6. Specifically, partPower status was linked with lower urban residency probabilities, while power status and above were more likely to have higher probabilities.

When the sample is restricted to couples in which both members are in the labor force and compared against single working individuals, many of the same patterns persist, although the magnitude of differences is somewhat reduced. This again supports the idea that urban residence among highly educated couples is partly, but not exclusively, driven by employment opportunities. Cultural, social, and institutional factors may also shape location preferences in ways that vary by couple composition.

Overall, the findings in Table 8 align with the summary statistics reported earlier: racial identity of couples is a non-negligible determinant of urban residence, even after adjusting for education and labor market participation. These results underscore the importance of explicitly including racial composition in empirical models of residential choice, rather than treating it as an individual-level control or omitting it altogether.

6. Discussion

Understanding migration patterns remains central to regional development. While past work has documented the disproportionate movement of highly educated individuals to urban areas, this study contributes new insights by focusing on couples—and particularly those with advanced degrees—across racial and sexual orientation categories. We asked two core questions: (1) Do previously observed urban residence patterns among college-educated couples persist or intensify for more highly educated pairings? (2) Do these patterns vary by couple-level characteristics? Our findings affirm both.

Using fixed-effects logit models and a uniquely constructed dataset from the PUMS, we show that power couples continue to urbanize at higher rates than otherwise equal singles, with the effect strongest among powerMost couples. When we control for labor force status and occupation, many coefficients shrink, suggesting that job location continues to drive urban concentration. Yet, the persistence of some differences, even among dual-earner couples, points to lifestyle preferences and co-location advantages.

Stratified results further reveal that highly educated black couples are disproportionately urbanized, while highly educated same-sex male couples appear less so. These patterns suggest that urban areas may offer amenities, social networks, or perceived inclusivity that matter differently across groups [47]. This complexity cannot be fully captured by education alone.

The patterns observed for highly educated couples have meaningful implications for regional economic development and workforce planning. If the most educated dual-earner households consistently gravitate toward urban areas, rural and mid-sized communities may face ongoing challenges in retaining or attracting high-skill labor. Jeworrek and Brachert [48] find that rural firms often struggle to attract high-skilled workers, even when wage offers are competitive. These trends point to deeper concerns about the perceived job attractiveness and infrastructure gaps in less urbanized areas [48,49].

From a policy standpoint, this underscores the need for place-based strategies that go beyond job creation alone. Investments in digital infrastructure, cultural amenities, and dual-career job-matching programs may be required to make non-urban locations viable for highly educated couples. Moreover, educational and workforce development programs in rural regions may need to be paired with retention initiatives to ensure that talent developed locally is not inevitably lost to urban migration.

This framework offers a starting point for future analyses of demographic-specific urban residence patterns. Extensions could further disaggregate racial and ethnic categories, consider immigrants and language minorities, or explore how preferences shift over time. In all cases, our results underscore the importance of incorporating couple-level heterogeneity into locational choice research and policy design.