The persistence of behavioral disparities post-pandemic: Insights from activity time series data

Abstract

The COVID-19 pandemic exposed and exacerbated deep-rooted socioeconomic disparities, however there is a gap in our understanding of the persistence and magnitudes of these disparities over time. In efforts to understand how diverse populations adapted their behavior throughout the COVID-19 pandemic, we leveraged large-scale mobile phone data at the census tract level between January 2020 – April 2022 to analyze temporal patterns in trips made to specific destination types. We employ time series clustering to the origin-destination trip data, which consistently identifies two distinct population groups that vary in both their respective behaviors and sociodemographic profiles. Furthermore, each group’s behavioral patterns in response to the local infection dynamics were observed to change over time, indicating a shift in relative risk-responsive behavior across groups. Specifically, in the early stages of COVID-19, the more socio-economically (SES) vulnerable population group was more active, likely capturing this group’s role in essential, lower-wage, in-person jobs and higher levels of social engagement, in contrast to the less vulnerable group, which was more likely to work from home and socially isolate. However, after restrictions lifted and vaccines became widely available, this trend reversed - the more SES vulnerable groups appear less active, especially in times of increased infection activity (e.g., new variant-driven waves), indicating less tolerance towards infection risk. The findings were consistent across all ten states included in the study. This study is the first to characterize and quantify evolving behavioral disparities through population-level activity patterns as captured through mobile phone data.

Introduction

The emergence of the COVID-19 pandemic in early 2020 caused a global health emergency and a profound societal disruption that fundamentally reshaped human behavior. To mitigate the spread of the SARS-CoV-2 virus, governments worldwide implemented a range of public health interventions¹, which transformed how individuals and communities interacted with their surroundings². Combined with the rapid emergence of new viral variants, they created a large-scale, real-world experiment in behavioral adaptation, forcing populations to adjust dynamically to rapidly changing environment. Studying how behaviors evolved and populations adapted provides essential insights for improving infectious disease modeling and strengthening future public health strategies.

The pandemic’s impact and the subsequent behavioral adaptations are far from uniform across populations. Societies entered the pandemic with existing inequalities in health, wealth, access to resources, and risk perceptions, all of which shaped both vulnerability to disease and the ability to adapt to pandemic-induced disruptions³. These disparities are observed not only in unequal health outcomes, with vulnerable populations experiencing disproportionately higher rates of infection, hospitalization, and mortality^4–6, but also in behavioral responses to the evolving crisis^7,8. Characterizing these behavioral disparities and their evolution across different outbreak phases is critical for addressing pandemic-induced inequities, informing equitable strategies, and supporting preparedness in future crises.

One key way to measure population behavior during a public health crisis is through the analysis of mobility data. In the US, a growing body of research has documented widespread reductions in mobility across various metrics and geographic contexts during the initial lockdown^9,10. Disparities in behavioral adaptations were also evident from the onset of the pandemic. Studies consistently found that disadvantaged groups, often characterized by lower socioeconomic status (SES), exhibited relatively higher levels of mobility during lockdown periods than higher-SES groups^11–13. Lower-SES populations were also less able to engage in protective behaviors¹⁴, such as limiting non-essential travel¹⁵. At the same time, higher-SES groups reduced their mobility more sharply and sustained these changes over longer periods¹⁶. Additional work has revealed disparities in policy responses¹⁷, segregation patterns¹⁸, and spatial-temporal mobility patterns^19,20, reflecting inequalities in the ability to adapt to these challenges. A dominant explanation attributes these patterns to the greater access of advantaged populations to flexible work arrangements, financial security, and other resources that enabled behavioral and lifestyle adaptations^21–23. Although some studies have begun to examine behavioral disparities during the recovery phases²⁴, the existing literature has primarily focused on the initial lockdown period, and it is still unclear whether and how these disparities persisted or reversed during later phases of the pandemic. A critical gap remains: How behavioral adaptations and disparities observed during the initial lockdown evolved as the pandemic progressed through subsequent policy shifts and variant waves?

In this study, we investigate how behavioral adaptations and disparities evolved throughout the COVID-19 pandemic, spanning diverse policy interventions and variant-driven waves. We leverage large-scale, anonymized mobile phone data to track census tract-level human activity patterns across ten U.S. states, evaluating activity levels at five types of destinations: essential, healthcare, accommodations, office buildings, and entertainment locations. We then apply timeseries clustering to identify groups of census tracts exhibiting distinct patterns of behavioral adaptation and link these clusters to detailed sociodemographic characteristics. Furthermore, we quantify the time-varying disparities in behavioral responses across identified population groups. This study offers insights into the evolution of behavioral shifts, adaptations and disparities, highlighting how population-level inequalities have been impacted by successive public health interventions and emerging epidemiological threats.

Results

Tracking the behavior variation using activity-derived clusters

Our analysis spans from January 2020 to April 2022, covering the initial lockdowns, reopening phases, and the periods dominated by three major SARS-CoV-2 variants: Alpha, Delta, and Omicron. The dominant period for each variant is determined based on genomic surveillance data from GISAID, where a variant is considered dominant once its relative frequency exceeds 50% among collected samples²⁵. We leveraged large-scale anonymized mobile phone data collected by SafeGraph’s Weekly Patterns dataset²⁶ to analyze census tract-level activity patterns across ten representative U.S. states: California, Colorado, Florida, Illinois, Kansas, Louisiana, Maine, Maryland, New York, and Texas. This dataset captures aggregated counts of visits to various points of interest (POIs) and has been widely utilized for studying mobility patterns during the COVID-19 pandemic^27–29. To characterize behavioral variation over time, we extracted weekly visit counts to POIs categorized into five destinations based on their North American Industry Classification System (NAICS) codes³⁰. The five destination categories include essential (such as grocery stores), healthcare (such as hospitals and clinics), accommodation (such as hotels and restaurants), office buildings (such as real estate and banking offices), and entertainment (such as cinemas and amusement parks). Additional details on data processing and destination classification are provided in the Materials and Methods.

For each census tract within each state, we constructed a weekly timeseries of destination-specific activity levels for each of the five destination categories for trips originating in the tract. Activity levels were defined as the number of weekly visits relative to a pre-pandemic baseline, calculated as the average weekly visits in 2019 (see formal definition in Materials and Methods). Behavioral similarity between census tracts was quantified by calculating the Pearson’s correlation coefficient for each of the five destination-specific activity levels. To obtain an overall measure of behavioral similarity, we generated an average similarity matrix for each state by calculating the arithmetic mean of the five destination-specific correlation matrices. This averaged similarity matrix thus provides a measure reflecting the overall similarity in the temporal dynamics of activity level between any two census tracts. A higher average correlation indicates that two census tracts exhibited similar temporal activity patterns over the study period.

We then applied the K-means clustering algorithm^12,31,32 to extract groups of census tracts that mirror similar activity patterns (for details see Materials and Methods). Since the similarity metric is based on Pearson’s correlation between standardized timeseries, this clustering procedure is mathematically equivalent to grouping tracts using Euclidean distance in the z-score standardized timeseries space³³. This approach groups census tracts according to the shape and synchronicity of their temporal dynamics across the average of the five destination categories, effectively capturing how activity levels co-varied over time. This clustering analysis was applied to each state independently, yet consistently identified two distinct groups of census tracts with similar behavioral patterns. The robustness of this two-cluster structure was confirmed by multiple evaluation metrics, details of which are provided in the SI Appendix Figure S2.

Demographic disparities between identified clusters

Leveraging the census-tract resolution of our activity-based clustering, we characterized each identified behavioral cluster with detailed sociodemographic profiles³⁴. This analysis relies on the assumption of a homogeneous spatial distribution of residents within each census tract, allowing the use of census tract level data as a representative proxy for resident sociodemographics. To illustrate the key differences in sociodemographic profiles between the two identified clusters, Fig. 1 presents boxplots of four selected key sociodemographic variables for the two clusters across the ten study states. The four variables depicted in Fig. 1 are: (a) Median household income (in thousands of US dollars), (b) Percentage of White population (%), (c) Percentage of population with a Bachelor’s degree or higher (%), and (d) the overall Social Vulnerability Index (SVI). The overall SVI is a composite metric that captures vulnerability across four dimensions: Socioeconomic Status, Household Composition & Disability, Minority Status & Language, and Housing Type & Transportation (ranging from 0 to 1, with higher values indicating greater vulnerability). Comprehensive details regarding the sources, definitions, and distributions of all sociodemographic variables utilized, including results for additional demographic variables, can be found in the SI Appendix.

Fig. 1.

Sociodemographics Profiles for Identified Clusters. The figure presents boxplots illustrating the distribution of four selected key sociodemographic variables for the two identified clusters within each state. Variables are shown in separate panels, from top to bottom: (a) Median household income (in thousands of US dollars), (b) Percentage of White population (0–100%), (c) Percentage of population with a Bachelor’s degree or higher (0–100%), and (d) Overall Social Vulnerability Index (SVI, ranging from 0 to 1, where higher values indicate greater vulnerability). For each boxplot, the central line denotes the median, the box encompasses the interquartile range (IQR, 25th to 75th percentile), and the whiskers extend to the most extreme data points within 1.5 times the IQR from the box edges. Outliers beyond the whiskers are not displayed. States are ordered along the x-axis by increasing difference in the median Overall SVI between the two clusters. Cluster 1, subsequently characterized as “vulnerable,” is represented by red boxplots, while Cluster 2, characterized as “baseline,” is shown in blue. Sample sizes (number of census tracts analyzed) are: California (N=7,689), Colorado (N=1,213), Florida (N=4,119), Illinois (N=3,004), Kansas (N=747), Louisiana (N=1,063), Maine (N=311), Maryland (N=1,354), New York (N=4,531), and Texas (N=5,141). Sources for all sociodemographic data are detailed in the Materials and Methods section. Boxplots for additional sociodemographic variables are provided in the SI Appendix Figures S8 and S9.

As is evident from Fig. 1, clear sociodemographic disparities emerged between the two activity-determined clusters. Based on the patterns observed across these variables, we characterize one cluster as “vulnerable” (represented in red) and the other cluster as “baseline” (represented in blue). The vulnerable cluster predominantly comprises census tracts characterized by lower median household incomes, a lower proportion of White residents, lower educational levels, and higher overall SVIs. Conversely, the baseline cluster exhibited the opposite characteristics.

The x-axis in Fig. 1 orders the states by increasing differences in median overall SVI between the two clusters. Consistent with this ordering, the vulnerable cluster consistently demonstrates a higher median overall SVI across all states, with the largest disparity observed in Louisiana (difference of 0.29). Similar patterns of disadvantage for the vulnerable cluster are evident in economic and educational indicators. Across all states, the vulnerable cluster consistently presents lower median household income and a lower proportion of residents with a Bachelor’s degree or higher across the 25th, 50th (median), and 75th percentiles relative to the baseline cluster. The only exception to this pattern is for the educational variable in Maine, where the red cluster has a relatively higher median Bachelor’s degree rate. Quantifying these disparities, we find that the baseline cluster exhibits, on average, a 33% higher median household income (ranging from 16% higher in Colorado to 65% higher in Louisiana) and a 21% higher average proportion of residents with a Bachelor’s degree or higher (with the largest difference of 40% higher observed in Maryland) compared to the vulnerable cluster. We also observed racial disparities among the identified clusters. Particularly in states such as Illinois, Maryland, and Louisiana, vulnerable clusters exhibited a significantly higher proportion of the White population compared to baseline clusters.

We conducted Mann-Whitney U tests³⁵ to formally evaluate the statistical significance of the observed differences in sociodemographic variable distributions between the two identified clusters. These non-parametric tests were performed independently for each sociodemographic variable and state. At a significance level of p < 0.05, the results indicated that the distribution of median household income was statistically significantly different between the two clusters across all ten states. The overall SVI distribution showed a statistically significant difference in all states except Maine and Colorado. Educational level showed a significant difference in all states except Colorado, as did the proportion of the White population. Full test statistics and p-values for all comparisons are provided in the SI Appendix, Table S4.

It is important to acknowledge, however, the inherent heterogeneity within each group. Not all census tracts within the vulnerable cluster face the same degree of socioeconomic advantages, and, conversely, the baseline cluster also includes some census tracts with low socioeconomic status.

Heterogeneity in activity dynamics by destination type

Similarity of activity level and recovery (defined as the highest post-lockdown activity level) varied significantly across the five destination categories. Visits to entertainment-related locations exhibited the highest temporal similarity, with an average Pearson’s correlation coefficient of 0.81 across states (see SI Appendix Figure S1). This high correlation suggests that entertainment-related mobility patterns changed in a consistent and synchronized manner across different states over time. Visits to entertainment-related locations also demonstrated the most robust recovery during the study period, reaching peak median activity levels exceeding 1.5 (150% of the 2019 average) in the summer of 2021 (as shown in Fig. 2, the top row of the panels a to c) for both clusters. In contrast, visits to healthcare facilities showed the lowest temporal similarity, with an average correlation coefficient of 0.46. This destination category exhibited the weakest recovery, with the median activity level for the vulnerable cluster never rising above 1 (100% of the 2019 average) throughout the study period. Aside from these two categories, visits to office buildings, essential locations, and accommodation sites displayed similar, moderate levels of temporal similarity (around 0.7 average correlation coefficient) and exhibited intermediate levels of activity recovery.

Fig. 2.

Temporal dynamics of activity levels for identified clusters across different policy and outbreak phases. This figure contains three panels. Panel (a) shows the overall temporal trends of the two identified clusters across all ten states. Panels (b) and (c) display corresponding results for New York and Texas, respectively. Each panel includes two rows of plots: the top row shows weekly activity levels across five destination types for the vulnerable (red) and baseline (blue) clusters. Solid lines indicate the median, and shaded bands represent the interquartile range (50% interval). The bottom row shows the median behavioral gap between the two clusters (purple line), overlaid with the logarithm of reported COVID-19 cases (yellow line) in the corresponding region. Background colors indicate key phases of the pandemic, including the lockdown period and the dominance of the Alpha, Delta, and Omicron variants. Variant dominance was determined using genomic surveillance data from GISAID²⁵, defined as the week in which a given variant accounted for more than 50% of sequenced cases in the corresponding state. Equivalent results for additional states are presented in SI Appendix Figures S10 to S11.

Temporal disparities in activity dynamics evolved with time

The temporal evolution of activity levels throughout the study period reveals dynamic disparities in behavior between the two identified clusters, which vary across destination categories and over time (Fig. 2, the bottom row of the panels a to c). During the initial Stay-at-Home period in early 2020, both clusters exhibited a sharp, universal decline in activity across all destination categories, with the most significant reduction observed for visits to entertainment locations (with median activity level below 0.5). Despite this overall sharp decline, the vulnerable cluster consistently maintained higher median activity levels than the baseline cluster across all five destination categories during this initial lockdown phase, a finding consistent with observations from multiple studies examining behavioral patterns during the early period of the pandemic^21–23.

Following the lifting of Stay-at-Home orders, a gradual recovery of visits began across all destination types. Notably, the baseline cluster rebounded faster than the vulnerable cluster. Within approximately three months of the initial recovery phase, the baseline cluster’s median activity levels surpassed those of the vulnerable cluster across all destination categories and states, representing a significant reversal of the activity disparity observed during the lockdown period. This shift in disparity in relative activity levels between the two clusters continued to widen over time (the purple curve in Fig. 2). The divergence between the clusters, with the baseline group maintaining relatively higher activity levels, persisted throughout the subsequent periods of Alpha, Delta, and Omicron variant dominance. The largest median activity gaps between the two clusters are consistently observed during the Delta or Omicron dominance periods. The sustained decreases in activity levels among the vulnerable group likely reflect shifts in risk-responsive behavior, underscoring a key behavioral dynamic revealed by our analysis.

To further explore the relationship between the median activity levels by cluster and the COVID-19 risk, we performed a correlation analysis between the weekly case growth rate (calculated as the log difference in weekly cases) and the fluctuation in median behavioral gap for each state and activity. Specifically, the responsive change in activity-specific behavior is defined as the detrended median behavioral gap obtained by removing linear trends (see details in the SI Appendix). We found a consistent and statistically significant positive correlation (p < 0.05) across all states between increasing median activity gaps and rising case growth rates in essential, office buildings, and accommodation-related destinations since the Alpha variant became dominant. These results indicate that wider activity gaps between vulnerable and baseline clusters were associated with higher rates of case growth. No significant association was observed for healthcare or entertainment visits. Detailed results are provided in the SI Appendix Figures S12 to S21. To clarify the temporal alignment of these associations, we conducted a lagged effect analysis (lag weeks; see SI Appendix Appendix Figures 22–24 and Appendix Table 1). This lagged analysis shows that the strongest associations for essential, office buildings, and accommodation activities are concentrated at short lags, with peak correlations most often occurring at or across states, indicating that activity-gap fluctuations tend to coincide with or slightly precede changes in case growth on a weekly timescale. This analysis is correlational and is intended to quantify co-movement, not causal effects.

Spatial distribution of identified clusters

The identified vulnerable cluster represents a minority of both the total population and the total number of census tracts across the study states. Across the ten states, the vulnerable cluster comprises, on average, 12.2% of the total population (ranging from 8.2% in Texas to 18.6% in California). In terms of geographic units, the vulnerable cluster encompasses, on average, 16.1% of the total census tracts (ranging from 11.4% in Texas to 22.1% in California). Detailed proportions of vulnerable census tracts and vulnerable population by state are documented in the Table 1 compares the proportion of urban housing units between vulnerable and baseline groups, expressed as a ratio relative to the state’s overall urban housing percentage. Across most states, the ratios for the two groups are close to 1, indicating similar urban–rural distributions. Notable exceptions are Louisiana (vulnerable = 1.35; baseline = 0.94) and Maine (vulnerable = 1.63; baseline = 0.96), where vulnerable groups are disproportionately concentrated in urban housing. Census tract–level maps, including zoomed-in visualizations for the most populous counties, are shown in SI Appendix, Figures S3–S7. Consistent with the quantitative results, no clear geographic pattern is observed across states.

Table 1.

Relative urban housing ratio by identified groups. Ratio between the percentage of a group’s urban housing units and the state’s overall urban housing percentage.

State	Vulnerable groups	Baseline groups
California	0.96	1.01
Colorado	1.04	1.00
Florida	1.06	0.99
Illinois	1.11	0.98
Kansas	1.13	0.98
Louisiana	1.35	0.94
Maryland	1.13	0.98
Maine	1.63	0.96
New York	1.01	1.00
Texas	1.16	0.99

Discussion

Our analysis demonstrates that clustering populations based solely on the temporal dynamics of their activity patterns results in groups with distinct sociodemographic profiles. The consistency of two SES distinct clusters across ten diverse U.S. states underscores that collective behavioral responses to COVID-19 risks are not uniform across the population. This phenomenon reflects how sociodemographic factors profoundly shape and constrain individual behavior, influencing daily routines, mobility choices, and access to different destinations^14,36,37. Consequently, areas with similar demographics tend to exhibit similar temporal patterns in their aggregated activity patterns, creating distinct behavioral patterns for different sociodemographic groups.

The temporal similarity of activity patterns varied significantly as a function of the destination types. Entertainment visits show the highest temporal similarity across the two identified clusters. This pattern likely reflects the dual influence of policy interventions and shared sociological responses across SES groups. First, entertainment-related businesses were disproportionately affected by lockdown mandates and reopening policies (with the lowest activity level during lockdown and the highest level since reopening, see Fig. 2), driving uniform fluctuations in access. Second, the resurgence of visits following reopening may reflect a common behavioral response across sociodemographic groups, as individuals sought stress relief and a desire to return to normal³⁸. In contrast, healthcare visits show the lowest temporal similarity. This divergence likely reflects the more individualized and localized nature of healthcare-related activity. Healthcare visits are primarily driven by personal health needs, such as existing medical conditions³⁹, and are shaped by the availability and capacity of local healthcare resources⁴⁰. As a result, patterns of healthcare facility visits are less synchronized across regions.

Most critically, our findings reveal that each group’s behavioral patterns in response to the local infection dynamics were observed to change over time, indicating a shift in relative risk-responsive behavior across groups. Initially, the vulnerable cluster maintained higher activity levels, likely due to a higher proportion of essential in-person jobs with limited remote work options and resources^21–23. However, after restrictions lifted and vaccines became widely available, this trend reversed - the more vulnerable groups appear less active, especially in times of increased infection activity (e.g., new variant-driven waves), indicating less tolerance towards infection risk. The findings were consistent across all ten states included in the study. This trend remained consistent over the study period, with the largest behavioral gap observed during the Delta and Omicron waves (Fig. 2), suggesting an increasing fear of infection among the more vulnerable populations relative to the less vulnerable. The increased sensitivity to infection risk among the vulnerable population may reflect persistent financial strain, job insecurity, and ongoing concerns over workplace and public safety. Supporting these explanations, survey data from April 2022 indicated that high-income workers felt significantly safer than low-income workers (60% vs. 41%)⁴¹. Furthermore, racial disparities in safety concerns were pronounced, with Black workers reporting significantly higher levels of concern about workplace safety (42%) compared to White workers (14%)⁴¹, aligning well with the demographic composition of the identified cluster. Economic constraints in 2021–2022 may also have contributed to reduced activity among vulnerable groups, and our analyses cannot disentangle economic mechanisms from risk-responsive behavior. Nevertheless, this relationship is correlational and can be bidirectional: rising local incidence can lead to voluntary or policy-driven behavior change, while changes in mobility patterns can in turn influence transmission². This long-term behavioral divergence also relates to broader economic anxieties, as lower-income adults were more pessimistic about the timeline for economic recovery (46%) than higher-income adults (31%)⁴². The observed changes in behavior in response to infection risk as a function of socioeconomic status highlight the pandemic impacted behavior disproportionately across SES groups, which persisted well beyond the acute phase of the pandemic, and after mitigation policies were lifted. These findings underscore how public health crises and the policies implemented to mitigate them can have different effects across socioeconomic groups, with inequalities persisting well beyond the acute phase of the pandemic. Although we used a two-cluster solution for a stable and interpretable cross-state summary, we emphasize that socioeconomic status is continuous and heterogeneous; additional clusters may reveal subgroups such as middle-income tracts, but do not change the central reversal pattern reported here.

The observed behavioral disparities underscore that even without formal public health mandates, different populations perceive and respond to risk in markedly different ways. Well beyond the acute phase of COVID-19, differences in perceived risk and lifestyle adaptation persisted, with more vulnerable SES groups enduring greater burdens that negatively impacted their quality of life. While a one-size-fits-all public health approach risks overlooking these disparities, our findings suggest that more customized strategies are needed to address the specific needs and constraints of vulnerable populations, both during a crisis and in its aftermath. These could include targeted financial assistance or unemployment support for those unable to work remotely, provision of paid sick leave, and job protection to enable workers to stay home when needed^43,44. Additionally, outreach and resources, such as testing, vaccination, and healthcare access, should be concentrated in lower-income and high-risk communities⁴⁵. Overall, our findings emphasize that equitable public health requires moving beyond universal mandates to a more nuanced approach that protects vulnerable populations during the crisis and supports them in recovery, ultimately closing the gap in how different groups experience a pandemic and other public health risks.

This study has several limitations. First, the analysis relies on mobile phone data from SafeGraph, which comes with known biases in coverage and representativeness, which may affect the magnitude of comparisons between different SES groups. If device coverage differs systematically by socioeconomic status within or across census tracts, estimated activity levels may be biased and could either attenuate or amplify observed disparities^46,47. Our activity metric is defined relative to each tract’s own pre-pandemic baseline, which reduces sensitivity to time-invariant coverage differences, but potential SES-related underrepresentation remains an important caveat. The dataset only captures a fraction of the population (approximately 7–8% on average) and is unevenly distributed across space⁴⁸. To mitigate this, we excluded census tracts with a daily data coverage below 2%, but it is important to note that low-income households and those with lower education levels, and some minority groups still exhibit underrepresentation. Second, our analysis assumes homogeneity of behavior within each census tract, treating all census tract residents as a single unit with the same mobility characteristics. In reality, this is a simplification that may obscure important intra-tract variability. In addition, this clustering approach does not consider the spatial distribution of the census tract, which may potentially overlook the impact of local resource availability on clustering outcomes⁴⁹. Meanwhile, we quantify activity using visit intensity relative to a pre-pandemic baseline. This measure does not directly reflect trip length or destination proximity, which can change differently from visit frequency and may vary with local accessibility and urban–rural structure. Finally, our clustering approach to classify mobility behaviors does not distinguish between work-related and non-work-related trips. All visits are treated uniformly, regardless of whether they are associated with employment-related travel or non-work activities. However, we did include an office-related destination category, which partially captures the patterns of white-collar work. Future studies incorporating workplace location data or occupation-based segmentation could help refine these distinctions and provide a more precise understanding of activity patterns.

Methods

Estimating purpose-specific activity patterns

To analyze the behavioral response changes since the pandemic, we analyzed anonymized mobile phone data from the SafeGraph weekly patterns dataset²⁶, spanning January 2019 to March 2022. The dataset contains no personal identifiers or information that can be used to ascertain the identity of any individual. SafeGraph requires its mobile application partners to obtain user permission for data collection and for the use of those data in commercial analytics and research, and individuals may opt out of data sharing at any point⁵⁰. All methods were carried out in accordance with relevant guidelines and regulations. Because the research involved only de-identified, non-individualized secondary data and did not involve human participants as defined under Johns Hopkins University Homewood IRB policy, the study did not constitute human subjects research and therefore did not require institutional review board approval or informed consent.

This dataset provides weekly visit counts from anonymized device home locations to various points of interest (POIs), providing a large-scale representation of population movement. Each POI in the SafeGraph dataset is associated with a North American Industry Classification System (NAICS) code, enabling the categorization of destinations by type. For this study, we aggregated visits within each census tract to five distinct destination types: essential services, accommodations, healthcare facilities, offices, and entertainment venues. These destination types were defined using NAICS codes, see Table 2 for details. To mitigate potential biases arising from uneven data coverage, we first aggregated the raw visit counts from the census block group level to the census tract level, which offers a balance between reduced sampling bias⁴⁸ and meaningful spatial granularity linked to local demographics.

Table 2.

Classification of destinations by NAICS codes. This table presents the categories used in this analysis (Essential, Healthcare, Accommodation, Office Buildings, and Entertainment) along with their corresponding NAICS codes and examples of establishments in each category.

Destination	NAICS Code	Description
Essential	42, 44, 45	Wholesale and Retail trade (e.g., grocery stores, pharmacies, general merchandise stores)
Healthcare	62	Healthcare and social assistance (e.g., hospitals, clinics, medical facilities)
Accommodation	72	Accommodation and food services (e.g., hotels, restaurants, bars)
Office Buildings	51-56	Professional, financial, and administrative services (e.g., offices and corporate buildings)
Entertainment	71	Arts, entertainment, and recreation (e.g., theater companies and dinner theaters)

We first estimated the weekly number of visits () from each census tract (i) to each destination category (d) using the following formulation:

1

where is the recorded number of visits from census block group j to destination category d during week t, is the population of census tract i, and is the number of devices residing in given census block group j during week t.

We then proposed a weekly destination-specific activity level metric (). This metric quantifies the relative activity level compared to the 2019 average, serving as a proxy for behavioral shifts since the onset of the COVID-19 pandemic. The formal definition of is presented below:

2

where is the average number of visits from census tract i to destination type d during year 2019. A value of greater than 1 indicates increased visits compared to the 2019 average, while a value less than 1 indicates decreased visits. Census tracts with less than 2% data coverage were excluded from the analysis.

In summary, we estimated for over 30,000 census tracts across ten geographically diverse U.S. states: New York, California, Texas, Florida, Maryland, Louisiana, Kansas, Colorado, Maine, and Illinois. For each census tract, we generated five time series, one for each destination type.

Sociodemographic data

To investigate socioeconomic disparities across different clusters, we incorporated multiple reliable demographic and socioeconomic datasets at the census tract level. Our data selection spans four critical domains of community characteristics known to be associated with disparate disease outcomes:

1. Racial-ethnic Composition, including White, Black or African American, American Indian and Alaska Native, and Asian.

2. Educational Attainment, encompassing high school completion, college degree, and bachelor’s degree.

3. Economic Status, represented by median household income and individual income.

4. Social Vulnerability Index (SVI), which assesses vulnerability across four dimensions: Socioeconomic Status, Household Composition & Disability, Minority Status & Language, and Housing Type & Transportation, along with an overall SVI score.

The majority of the data is sourced from 2019, while the SVI data is from 2018, ensuring consistency across census tracts. We intentionally excluded 2020 datasets to avoid geographic discontinuities introduced by recent census reconfigurations. In the main text, we focused on four representative indicators for each category, White (%), Bachelor’s degree (%), Median household income (k), and overall SVI. The Table 3 provides descriptions for each selected variable. The description and analysis of additional variables are included in the SI appendix for reference.

Table 3.

This table lists key demographic and socioeconomic variables used in this study, along with brief descriptions and their respective data sources.

Variable	Description	Source
White (%)	This variable represents the proportion of individuals identifying as White American within each census tract, as reported in the 2015-2019 American Community Survey (ACS) 5-Year Estimates.	51
Bachelor’s Degree (%)	This variable quantifies the educational attainment within a census tract by measuring the percentage of the population holding a bachelor’s degree, based on data from the 2019 ACS 5-Year Estimates.	51
Median household income (k)	This variable represents the median household income (in thousands of dollars) within a census tract, derived from the 2015-2019 ACS data. It provides insights into the economic status of the population.	52
Overall SVI	A composite index (0-1 scale) measuring social vulnerability at the census tract level, incorporating multiple dimensions such as socioeconomic status, demographic composition, minority status, housing conditions, and transportation access. Higher values indicate greater vulnerability. Data are sourced from the 2018 CDC/ATSDR Social Vulnerability Index (SVI).	53

Timeseries clustering analysis

Given the five time series representing visits to different destination types for each census tract, our goal is to cluster census tracts based on the similarity of their activity patterns (). To achieve this, we employ a systematic approach involving standardization, similarity computation via correlation, and K-means clustering analysis.

To ensure comparability across different scales, each time series is standardized, transforming it to have a mean of zero and a standard deviation of one (z-normalized). This step ensures that correlation-based similarity measures remain meaningful. The standardized time series is denoted as .

Within each state, we quantify the similarity of activity trends between different census tracts for the same destination type by computing the Pearson’s correlation coefficient on their standardized time series. This approach, using Pearson’s correlation on standardized time series, is mathematically equivalent to calculating the Euclidean distance between these standardized time series⁵⁴, a well-established measure of similarity. For two standardized time series, and , corresponding to census tracts i and l for a given destination type d, the Pearson’s correlation coefficient () is defined as:

3

where T is the total number of time points.

To obtain an overall similarity score, we average the correlation coefficients across the five destination types:

4

We apply the K-Means clustering algorithm to the similarity matrix derived from correlation coefficients, where . n is the total number of census tracts for the given states. We repeated the k-means clustering from 30 different random initial centroid initializations and allowed up to 1,000 iterations per run to reduce sensitivity to starting conditions and ensure convergence. To determine the optimal number of clusters K, we evaluate the clustering results across multiple values of K using three evaluation metrics, namely Silhouette Score⁵⁵, Calinski-Harabasz Index⁵⁶, and Davies-Bouldin Index⁵⁷. Details regarding the formulation of the evaluation matrix and the assessment of the clustering results are provided in the SI Appendix.

Author contributions

H.D. and L.M.G. conceptualized and designed the study. H.D. and S.X. collected data and processed the data. H.D. and S.X. performed experiments. H.D. and S.X. prepared the figures. H.D. and S.X. analyzed results. H.D. and S.X. wrote the initial draft. L.M.G., H.D., N.R., and H.F.Y provided guidance and feedback for the study. L.M.G. H.D., S.X., and N.R. revised the manuscript. L.M.G. acquired the funding. All authors prepared the final version of the manuscript.

Funding

This work was supported by NSF Award ID 2229996 (L.G.) and Merck KGaA Future Insight Prize (L.G.). N.R. acknowledges financial support from Johns Hopkins University through the Vivien Thomas Scholars Initiative. The contents are solely the responsibility of the authors and do not necessarily represent the official views of the funding agencies.

Data availability

The raw mobility data used in this study were obtained from SafeGraph during the COVID-19 pandemic. Due to licensing restrictions, these raw data cannot be shared publicly, but they are available from the authors upon reasonable request and with approval from SafeGraph. To facilitate reproducibility, we share the derived similarity metrics between census tracts, which were used as inputs to generate the clustering results. All other datasets analyzed in this study are publicly available and are described in the Methods section.

Code availability

The code used to process the raw data and generate the results is available at https://github.com/hongru94/TS-Clustering.

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-41553-9.