Figure 2 summarizes our methodology for identifying different clusters of counties using Google mobility indicators. Clustering is an unsupervised learning technique that partitions a data set into groups or clusters based on similarity measures. This study leveraged partitioning-based algorithms, which divided the data set into partitions, where each partition was a cluster. For each county included in this study, data were clustered based on a combination of the daily values of the 4 Google mobility indicators. To identify the different clusters of counties, we performed 2 steps [28]:

To compress the multidimensional time series, we implemented the variational autoencoder (VAE) architecture based on long short-term memory (LSTM) [29-31]. The principal concept of this generative approach is to project high-dimensional data into latent variables. Our model comprised 4 blocks [32]:

Identifying the true posterior distribution is intractable [33]. Therefore, to construct the original data, the probabilistic encoder model was approximated by normal distribution p(z|x)N(0,1) and used as a probability decoder [30,33]. Hence, the reconstruction of input was defined by sampling from the distribution of latent variables (z).

To evaluate the performance of the model, the loss function was defined as follows:

The model was trained in Python 3.6 using the Keras library [34] with the Adam optimizer. The batch size and number of the epochs were set to 10 and 100, respectively. The number of nodes for encoder and decoder hidden layers was set to 500. The dimensionality of latent variables was set to 3. We also implemented the L1 and L2 regularizers to avoid overfitting. To evaluate the performance of the model, the VAE total loss was used to identify the reconstruction error between encoder input and decoder output.

Once the model was trained and the encoder, decoder, and VAE were constructed, the output of the encoder model was selected as the representation of the multidimensional patterns of each county. K-means clustering was used to identify the similar segmentation of the counties. To identify the optimum number of clusters as well as the homogeneity of data points within each cluster, the elbow method [35] and the silhouette score [36] were used.

To compare the socioeconomic characteristics of the counties in each cluster, the 2016 MIT election data were used as input, while the classes were the cluster labels. The data were divided into training and testing sets with a 70:30 split, respectively. The random forest classifier [37] with 10 k-fold cross-validations was used to build the predictive models. The area under the curve (AUC) of the model was calculated, and the most important features associated with the cluster numbers were defined as the parameters describing the characteristics of counties in each cluster. Feature scores of different census variables for the clusters were computed, which yielded an idea of the relative importance of different socioeconomic and demographic factors for explaining the different clusters. Figure 3 summarizes our approach.

This study leveraged a partitioning-based deep learning model to cluster counties based on similarities in social mobility. For each county included in this study, data were clustered based on a combination of the daily values of the 4 Google mobility indicators (retail, grocery and pharmacy, workplace, and residence). The multidimensional time series of Google social mobility indicators from 1089 counties was divided into training and testing sets and fed into the VAE model. The result demonstrated a loss of 0.08. The latent variables were extracted as the output of the encoder. The K-means clustering algorithm identified 4 social mobility clusters. The number of counties in these clusters, which were termed as 0, 1, 2, and 3, were 215, 338, 473, and 59, respectively. Figure 4 gives the distortion scores of the K-means clustering.

Across all clusters, visits to retail stores fell significantly after the start of the pandemic until around mid-April, followed by a steady increase and plateauing in early July (Figure 5). Visits to retail outlets began to decline again in late September but then began an upward trend starting on Thanksgiving weekend until the end of December. Retail social mobility values were the highest for cluster 0, followed by clusters 2 and 1, with cluster 3 having the lowest social mobility. Grocery and pharmacy mobility trends reflected those seen for retail social movements but were less pronounced (Figure 6). Cluster 0 had the highest values of grocery mobility, followed by clusters 2, 1, and 3. Workplace mobility showed an initial decline at the start of the pandemic, followed by a steady increase from early May onward (Figure 7). Spikes in mobility were observed during the weekend, which did not significantly decline relative to prepandemic observations. County clusters followed the same order, with cluster 0 having the greatest mobility, followed by clusters 2, 1, and 3.

Finally, residential mobility followed a reverse pattern relative to the other indicators, with cluster 3 having the highest mobility, followed by clusters 1, 2, and 0 (Figure 8). Residential mobility was highest during the onset of the pandemic, followed by a decreasing trend during spring and summer. From late September onward, residential mobility began to increase, and this trend continued until the end of the sample period. The spikes in mobility captured the weekend effects. Our social mobility data indicated differences in mobility between clusters, with counties in cluster 0 having the highest retail, grocery, and workplace mobility and the lowest residential mobility. In contrast, counties in cluster 3 had the lowest social mobility and the highest residential mobility.

To determine whether county characteristics are correlated with differences in social mobility between the clusters, we obtained socioeconomic, demographic, and political data from each county from 2016 census data [18]. These data included 2016 election returns, race, median income, total population, percentage of rural areas, and education level of the population for age and race. These data were supplemented by county variables collected by other studies [23,25].

A random forest classifier was used to generate feature scores of different socioeconomic and demographic characteristics of the counties included in each cluster, across all 4 clusters (mean receiver operating characteristic [ROC] AUC 0.871). Table 2 contains the feature scores of all county-level variables.

The top 10 variables in terms of feature scores were percentage of the population aged 65 years and over (0.41715), percentage of females (0.08784), percentage of Whites (0.03869), percentage of Whites with less than college education (0.03772), percentage of Hispanics (0.03369), percentage of Whites with less than high school education (0.03178), percentage of the population using public transit (0.02967), county unemployment rate (0.02759), proportion of voters for Clinton in 2016 (0.02737), and percentage of the population with less than high school population (0.02719). Hence, although political preference and population composition were important, it is important to note the significance of 3 educational variables among the top 10, with the percentage of the population with less than college education being the 11th variable in terms of feature score.

To explore the top 11 socioeconomic, demographic, and political variables impacting social mobility further, we determined the mean population percentage for each county-level variable across clusters (Table 3). The table also contains results of statistical tests of significance of sample means between clusters. The Z test of sample means was performed to compare the significance of different county-level variables for different clusters. Results demonstrated several variable similarities for clusters with the highest social mobility. The percentage of the population aged 65 years and over, Whites, the percentage of whites with less than high school and college education, and the percentage of the overall population with less than college education were higher in counties defined by clusters 0 and 2. Tests of equality of sample proportions and means confirmed that there was a statistically significant difference between clusters 0 and 2 versus clusters 1 and 3 for these population variables. In contrast, the percentage of Hispanics, percentage of the population using public transit for work, and percentage voting for Clinton in 2016 were lower in clusters 0 and 2 relative to clusters 1 and 3. There was no consistent, significant difference across clusters for the percentage of females, population with less than high school education, and unemployment rates.

Given that policies restricting population mobility were established to curb the spread of COVID-19, we sought to determine whether county clusters with higher social mobility indicators (clusters 0 and 2) reported elevated viral cases and deaths. The daily number of confirmed cases and deaths due to COVID-19 at the county level was obtained from the CSSE at the JHU. We determined the median daily per capita cases (Figure 9) and deaths (Figure 10) by cluster. During the first months of the pandemic, per capita daily cases were quite comparable across clusters (Figure 9). There was a visible divergence that occurred at the beginning of October (onset of the second pandemic wave), with daily cases rising sharply in clusters 0, 1, and 2 relative to cluster 3. For the remainder of the period examined, cluster 0 had the highest number of daily cases, followed by clusters 2 and 1. Cluster 3 retained relatively lower daily cases. Interestingly, clusters 0 and 2 had lower daily deaths until the beginning of September (Figure 10). Daily deaths in these clusters then increased rapidly, and by the beginning of October, per capita deaths in clusters 0, 1, and 2 were higher than in cluster 3.

This study aimed to assess the effect of county-level characteristics on population mobility and the consequences of this mobility on the spread of COVID-19. To the best of our knowledge, this is the first study that has used unsupervised machine learning to understand differences in population mobility across US counties during the first and second waves of the pandemic and determine the relative importance of a wide array of socioeconomic, demographic, and political variables in defining different mobility-based clusters.

Our results demonstrate that of the 4 clusters defined by Google social mobility indicators, the clusters with higher retail, grocery, and work mobility (and lower residential mobility) had several similar population characteristics. Specifically, counties with greater social mobility also had a higher percentage of the population aged 65 years and over, Whites with less than high school and college education, and overall population with less than college education. Counties in these 2 clusters also had a lower share of the population that is Hispanic, the percentage of the population using public transit to work, and the share of voters who voted for Clinton during the 2016 presidential election. Research does suggest that Whites with less than college education constituted a significant voting block for Trump during the 2016 election [38]. In line with this, the 2 clusters with the greatest social mobility also experienced higher per capita COVID-19 case and death rates during most of November and December 2020. These results are consistent with Xie and Li [39], who also used county-level data during the early days of the pandemic and found lower education levels to be correlated with higher infection rates.

The significant increase in COVID-19 cases and deaths in clusters 0 and 2 during November and December 2020 could be a consequence of public rallies and general disregard for social distancing and safety protocols by pro-Trump voters [40]. Although we cannot prove this, the majority of counties in these clusters were Republican leaning during the 2016 presidential election. Moreover, our finding of higher per capita daily COVID-19 cases and deaths in such counties is consistent with other studies. Desmet and Wacziarg [41] found that early on during the pandemic, Republican counties actually experienced lower COVID-19 cases and therefore had lax attitudes toward mask wearing, social distancing, and lockdown measures. However, as the pandemic spread to Trump-leaning counties, population preferences for less stringent social distancing policies had already been formed, making it difficult for policymakers to implement stricter restrictions on social mobility. As a result, this led to greater disease severity in Trump-leaning counties. In a similar vein, Allcott et al [42] found that areas with more Republicans engaged in less social distancing, controlling for other factors, including public policies. In summary, these findings corroborate our own results. Social mobility in the aftermath of the first wave of the pandemic was much higher in Republican counties, which ultimately resulted in higher COVID-19 cases and associated deaths relative to other counties that were Democrat leaning.

Social media is increasingly being used to capture population movements and understand their corresponding impacts on COVID-19 incidence. Social media–based data, including those presented here, have some limitations. Specifically, there is the possibility of sample selection bias if Google Maps users have specific demographic characteristics and are not distributed uniformly across the population. However, data from Statista indicate that in the U.S., Google Maps had 154 million users in April 2018 [43]. Further, published research has done a comparison of Google mobility data against corresponding cellular-generated information by other providers and has found a close correspondence. Specifically, Szocska et al [44] constructed a mobility index and an SAH/resting index based on data on almost all phone subscribers in Hungary and found a close correlation with corresponding Google mobility indices at the national level. There are also a significant number of published studies that have used Google mobility data to capture population movements for different countries and have found them to be important in predicting movements in COVID-19 (Bryant and Elofsson [11], Askitas et al [45], and Stevens et al [46]). For these reasons, we think there is a high likelihood that Google mobility data do reflect population movements. However, Google mobility data do not include information on certain types of public movements, such as election rallies or community gatherings.

Our research demonstrates the usefulness of publicly available Google mobility data and unsupervised machine learning methods in establishing relationships between county-level characteristics, mobility decisions, and COVID-19 incidence. These findings have important implications for policymakers and public health officials in understanding the effects of NPIs, as the efficacy of such measures on mobility is influenced by underlying socioeconomic, demographic, and political ideology characteristics. The use of Google data enables researchers to assess the types of public movements that are most contributory to COVID-19 spread.

The results of this study provide a unique lens on the potential of machine learning to understand social mobility behaviors. These findings are critical for public health organizations trying to understand the levels of mobility in their counties, in addition to providing insights into some of the underlying factors (ie, social determinants of health) contributing to regional differences in COVID-19 caseloads.

Our results emphasize a role for machine learning methods in public health. Publicly available Google data, in conjunction with census data, can be used to understand the socioeconomic, demographic, and political determinants driving population mobility choices across US counties. This knowledge can assist policymakers in developing NPIs to restrict viral spread during the COVID-19 pandemic.