For >2 years, the COVID-19 pandemic has been afflicting global populations in a variety of ways. Although the physical impact of COVID-19 is evident and has been well documented in the literature [1-3], more research is needed to assess the mental impact of the COVID-19 pandemic. Many studies have assessed short-term impacts of the pandemic on mental health. For instance, a meta-analysis by Talevi et al [4] revealed impacts on psychological well-being, including increased anxiety, depression, stress, and insomnia during the first few weeks of the pandemic in March 2020 and April 2020. These associations were detected in a variety of populations, including college students, health care workers, and patients with COVID-19 [4]. Other works have unveiled similar mental impacts because of extended periods of quarantine [5,6]. In light of these findings, fewer works to date have studied the long-term impacts of the COVID-19 pandemic on mental health. In a meta-analysis, Bourmistrova et al [7] found that patients with COVID-19 had mild to no anxiety and depression symptoms for at least 1 month following diagnosis, but these trends were not investigated in the general population (eg, including patients without COVID-19). Conversely, Veldhuis et al [8] found that survey participants (N=1567) had increased depressive symptoms and suicidal thoughts after a 5-month follow-up at the end of summer 2020. Taken together, more research is still needed to assess the long-term mental effects of COVID-19, the factors contributing to these effects, and the populations most at risk of experiencing these effects. Such research will help accentuate the need to institute policy measures that increase access to and strengthen mental health services worldwide as well as increase awareness of the specific mental risk factors that exist among high-risk populations [5]. In this light, this study aimed to contribute to the limited literature on the long-term mental health impacts of COVID-19 through investigating trends in mental health activity along with their possible contributing factors over the first year of the pandemic (March 2020-March 2021) in a large representative national sample in the United States.

Importantly, a data stream that has shown promise for operationalizing mental health is internet search term activity [9,10], with dynamic tools such as Google Trends providing a publicly available platform with which to investigate internet search behavior. Most people aged <25 years perform internet searches as their primary means of seeking mental health help [11]; therefore, internet search activity may be a potent indicator of changes in mental health. Various studies have documented the use of internet search engines for mental health self-diagnosis [9,12]. In addition, myriad works have connected internet searches for psychopathologies with actual experience of them. For instance, a study found that increased self-reported anxiety was associated with increased Google search rates for anxiety [13], whereas another study unveiled a positive relationship between depressive symptom search volume and same-month suicide rates [14]. In the context of COVID-19, Ayers et al [15] discovered that there were sudden spikes in anxiety-related search activity in March 2020 to April 2020 in the United States. Furthermore, analyses by Gianfredi et al [16] revealed significant positive associations between mental health–related search volume (eg, insomnia, suicide, and depression) and COVID-19 deaths. Although it is important to acknowledge that mental health symptoms differ greatly across individuals, analyzing trends in population-level internet search behavior can reveal associations with potentially important implications for public health and policy [14].

Leveraging internet search behavior as an indicator of mental health, this work aimed to specifically disentangle some of the factors that may influence changes in mental health in relation to the COVID-19 pandemic. In particular, this research brought to attention the ubiquity of state-enforced mask mandates throughout the pandemic and their potential to affect mental health. A total of 39 out of 50 states in the United States implemented statewide mask-wearing mandates as a means of limiting the transmission of the novel coronavirus. Although mask mandates may have provided some with a reduction in anxiety because of added protection from the coronavirus or mitigating sources of social anxiety [17], wearing a face mask may be perceived as restrictive or uncomfortable [18] as well as something that may threaten personal freedom [19]; therefore, it is reasonable to believe that mental reactions to the requirement of masking were heterogeneous and often adverse both across the country and throughout this phase of the pandemic. Reasons for an adverse response are manifold and need not be restricted to groups rather than to masking. For example, individuals in favor of masking may respond negatively to the imposed changes as a result of interactions with those that are less in favor of masking. A study revealed that there was an increase in anger present in tweets that were promasking in states both with and without mask mandates [19]. In light of the associations between mask mandates and mental health, no studies have examined the broader, long-term mental impact of mask mandates. In addition, there is evidence suggesting that masking behaviors, including mental responses to masking mandates as well as mental health constructs such as anxiety and fear, differ across political parties in relation to the COVID-19 pandemic [20,21]. Accordingly, this study sought to investigate the associations among state-level mask mandates, political affiliations, and mental health during the COVID-19 pandemic by way of internet search behavior. Such an investigation may provide leaders with important considerations regarding the long-term mental implications of implementing large-scale interventions such as mask mandates and how the political leanings of their populations may contribute to these mental health responses.

The aim of this study was to evaluate the long-term influences of mask mandate implementation and political party affiliation on mental health search term volume between March 24, 2020, and March 29, 2021, in the 50 states in the United States. Changes in mental health were modeled using generalized additive mixed models (GAMMs) on data collected through Google Trends. In particular, this study serves to complement the work by Jacobson et al [22], where an attenuation of mental health search term activity was found in the United States after stay-at-home orders were implemented. Consequently, this work was interested in examining whether statewide mask mandates and political party affiliation, quantified as political party elected during the 2020 presidential election (see the Mask Mandates and Political Party Elected section), had similar significant associations with mental health activity. In addition, physical health symptom search terms both related and unrelated to COVID-19 were analyzed as a means of comparison with the associations between mental health search term volume and mask mandates. In this data-driven exploratory analysis, the authors investigated the research questions outlined in Textbox 1.

Given the prominence of Google as a search engine (eg, Google accounts for roughly 88% of the search engine market in the United States [23]), these analyses leveraged Google Trends, which allows for public access to search term volume for a given time and location. Several studies have leveraged Google Trends data to examine how mental health information is sought out [22,24,25]. As can be found in the documentation for Google Trends, the raw counts for a given search term are normalized by location and time of search and then scaled to a number from 0 to 100 representing the proportion of searches on all topics that the given term constitutes. Such normalization allows for easier comparison across geographic regions, where population may play a significant role in relative search term popularity. This analysis downloaded data from Google Trends using the gtrendsR package in R (version 1.4.8; R Foundation for Statistical Computing) [26]. To obtain data with the most granularity, hourly trend data were queried. The areas of interest were the 50 states constituting the United States; thus, search terms were normalized across states. As Google Trends only stores hourly data for up to 7 days, hourly data were programmatically pulled each Monday from March 23, 2020, to March 29, 2021. This period of 372 days spans the early days of the pandemic to the widespread availability of the COVID-19 vaccine in the United States. During this time frame, 39 states implemented statewide mask mandates, most of which went into effect before or during the summer of 2020. Thus, the given time frame allowed for careful introspection into the short- and long-term effects of mask mandate implementation on mental health search term activity.

The following 19 mental health search terms were queried from Google Trends, as described previously: “anxiety,” “depression,” “ocd” (obsessive-compulsive disorder), “hopeless,” “angry,” “afraid,” “apathy,” “worthless,” “worried,” “restless,” “irritable,” “tense,” “scattered,” “tired,” “avoiding,” “procrastinate,” “insomnia,” “suicidal,” and “suicide.” Aligning with previous work by the authors [10,22], these terms were validated from previous research on using Google Trends to assess mental health [27], as well as from previous research assessing rapid affective symptom changes as defined by the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition [28,29]. In addition to these terms, physical health search terms, both without known associations to COVID-19 (“abrasion,” “allergic,” “angina,” “apnea,” “bleeding,” “blister,” “bruising,” “conjunctivitis,” “constipation,” “discharge,” “earache,” “flatulence,” “fracture,” “hemorrhage,” “incontinence,” “inflammation,” “itching,” “lesions,” “rash,” “spasms,” “swelling,” and “syncope”; 22 terms) and with known associations to COVID-19 (“bloating,” “blurry,” “congestion,” “cough,” “coughing,” “croup,” “diarrhea,” “dizzy,” “fainting,” “fever,” “pain,” “sneezing,” “strep,” “stuffy,” and “vomiting”; 15 terms) were queried to ascertain whether any significantly detected patterns in mental health search term activity were unique to and distinct from those pertaining to physical health. Note that each mental and physical health search term was considered independently in this study; in other words, composite scores aggregating the individual search term counts to create a composite score capturing total mental and physical health activity were not created. This decision was made because combining individual search terms with differential trends throughout the pandemic may attenuate these individual trends in the composite score such that the composite score may not be reflective of changes in specific mental or physical health symptoms, therefore making it uninformative.

This study leveraged GAMMs to examine changes in mental health search trends over the first year of the COVID-19 pandemic. GAMMs are a powerful modeling approach as they can account for nonlinear trends as well as disentangle the interdependence of observations. This analysis used data in which observations (ie, daily search term counts) were nested within groups (ie, states); therefore, incorporation of a mixed-effects framework was suitable to estimate trajectories of search term activity over time. In addition, GAMMs allow for estimations of variable-specific trend comparisons (smooths) via incorporation of smoothing functions (smoothers) [39]. Specifically, this study used splines as the smoothing function as splines account for nonlinearity in the data but only model nonlinearity in the predictor-outcome relationship if nonlinearity provides the best fit for the data [22]. To implement GAMMs, the bam function within the mgcv package was used. The bam function allows for the specification of GAMMs, is optimized for faster run times on larger data sets (eg, each search term had data for 371 days and 50 states, or 18,850 individual observations), and allows for direct hypothesis testing of the estimated smooths [40].

The selected model incorporates 8 smooth terms, denoted by “s,” as well as 6 linear terms pertaining to state-level mask mandate, political party elected, urbanization, per capita income, and COVID-19 cases and deaths. These linear terms were included to isolate the main effects of the binary terms of interest, allowing the smooth terms to capture any nonlinear relationships in the data. Importantly, the negative binomial distribution was chosen for the model because of the use of (nonnormally distributed) count data and overdispersion (ie, variance exceeds the mean) present in the outcome variables. Textbox 2 describes the complete model architecture with associated variable definitions. Each model was independently run on 12 mental health and 26 physical health search terms.

All s terms indicate that smoothing was used to estimate the relationship between the predictor and outcome of interest. Term s₂ represents the random smooth slopes of state j at each time i, allowing a linear or nonlinear random effect to account for changes in the outcome. Terms s₃ to s₈ represent varying coefficient smooths specified by the tensor product interaction smoothing function, which, in addition, accounts for the presence of main effects in the interacting variables [40]. For model-specific hyperparameter specification, “fREML” (“fast REML” computation) was used for smoothing parameter estimation, which, when used in conjunction with setting discrete=TRUE, discretizes covariates into smaller bins and substantially reduces computation time [41]. Furthermore, the negative binomial distribution was used in the model by selecting family=“nb.”

In terms of smooth-level hyperparameter specification, all smooth terms, aside from s₂, were modeled with a thin plate regression spline (TPRS) basis function, which allows for linear or nonlinear relationships between the predictor and the outcome variable while also penalizing nonlinearity such that the data will only be modeled as nonlinear if the model fit is substantially better. In addition, modeling with a TPRS circumvents issues with knot placement [42]. Term s₂ uses a factor smooth interaction basis function that fully accounts for random effects; provides a better fit when the number of grouping variables is high [39], as in these analyses (ie, 50 states); and draws group-level smooth terms toward zero by only estimating 1 smoothing parameter across groups [41]. Along with this, the group-level smoothers, s₃ to s₈, use m=1, which penalizes the TPRS with the squared first derivative of the function, whereas s₁ and s₂ use m=2, penalizing the spline with the squared second derivative of the function. Using m=1 in the group-level smoothers reduces collinearity with the global smoother and group-specific terms (ie, s₁ and s₂) [41]. In addition, the smooth terms s₃ to s₈ use the default k value (k=5) in mgcv, which balances smooth fitting with computational time [40]. However, k is set to 3 for the smooth terms s₁ and s₂ to allow for up to cubic trends to be modeled by the time and state random effects of time variables. Although it is optimal to choose a high k when computationally feasible, limiting k in this context restricts the nonlinear deviations in the time trends that can be explained by the state random effects such that these complex state-specific nonlinearities can be potentially accounted for by other sources.

This study is interested in examining the associations among mask mandates, political party elected, and search term trajectory; therefore, emphasis is placed on the s₃ and s₄ terms. Term s₃ estimates the observed associations of the statewide mask mandate as a deviation from the state-specific counterfactual trend that would have occurred had there not been a statewide mask mandate implemented, and s₄ estimates the observed associations of Republican states with a statewide mask mandate as a deviation from the state-specific counterfactual trend that would have occurred in Democratic states with a statewide mask mandate. Accordingly, the model output from s₃ and s₄ is of particular interest.

These analyses consisted of 38 negative binomial GAMMs, with each model corresponding to the volume of activity for 1 search term over time as the outcome. The results of the smoothing term estimation were of particular interest to assess any nonlinear associations captured by the predictor variables. As smooth terms estimate nonlinear trends in the data, the results of these models are not equivalent to a regression slope; rather, estimated df (EDF) are reported, where EDF=1 corresponds to a linear relationship between the predictor and outcome, 1<EDF≤2 corresponds to a weak nonlinear relationship, and EDF>2 corresponds to a strong nonlinear relationship [43]. The bam function in mgcv performs a significance test and reports the P value for the EDF corresponding to each smooth term. Importantly, as this analysis consisted of 12 independent models for mental health search terms, results of variable significance needed correction for multiple comparisons. Accordingly, P values were adjusted by controlling the false discovery rate using the Benjamini-Hochberg procedure. The same procedure was used for the 17 independent not COVID-19–related and 9 independent COVID-19–related physical health search term models.

To assess how well the models explained the variation in daily search term counts, the R² value was reported. This R² value corresponds to the conditional R², or the total variance in the outcome explained by both the fixed and random effects. Thus, to assess whether the random effects are driving model prediction as well as determine if the fixed effects (ie, mask mandate and political party elected) explain a considerable amount of the variation, the marginal R²—the total variance explained by the fixed effects—was also reported. Given limitations in the ability to directly estimate the marginal R² from models constructed using the mgcv package, the marginal R² was estimated manually by rerunning each model without the s₂ random effect of time and reporting its R² value.

Note that, as MaskMandate_i,j was coded as a “1” when a state had a mask mandate in effect and “0” when the mandate was not in effect or the state did not have a mandate, these analyses only make interpretations of the associations between mask mandates and search term trajectory after the implementation of mask mandates. In this manner, this study does not draw conclusions regarding changes in search behavior leading up to the mandate.

This study was not considered human participant research as it used anonymous, publicly available web-based search data and, as such, was exempt from human participant approval.

Figure 1 displays a comparative timeline including each of the 39 states that implemented a statewide mask mandate. Fewer Republican states had a mask mandate than Democratic states. Moreover, Republican states that did have a mandate were generally delayed in their implementation, and the mandates were implemented for a shorter period than in Democratic states. Indeed, mask mandates in Republican states were in effect for an average of approximately 206 (SD 89) days in comparison with 430 (SD 162) days in Democratic states. New Jersey was the first state to implement a statewide mask mandate on April 8, 2020. Note that these mask mandates reflect policies that were effective at the state level—some states (eg, Georgia) had mandates at the city or county level, but these mandates were not considered under this investigation. In addition, some states may have ended their initial mask mandate and reinstituted it; however, only the first statewide mask mandate was considered and, therefore, reflected in Figure 1. The exact dates of when each mask mandate was in effect, as well as other state-level information included in these analyses (ie, political party elected, COVID-19 cases and deaths, urbanization, and per capita income), can be found in Table S1 in Multimedia Appendix 2.

The results of the missingness analyses can be found in Table S1 in Multimedia Appendix 1. After preprocessing the raw Google Trends data for the 19 aforementioned mental health search terms, 12 terms (“anxiety,” “depression,” “ocd,” “angry,” “afraid,” “restless,” “irritable,” “tense,” “tired,” “insomnia,” “suicidal,” and “suicide”) had <5% of values missing for aggregated daily hit counts. In other words, throughout all days from March 24, 2020, to March 29, 2021, for each of the 50 states, <5% (19/371) of days did not have a hit count associated with them for these terms. Consequently, data for these terms were included in the modeling portion of the analysis after performing zero imputation, as outlined in the Outcome Calculation and Imputation section. All other terms were dropped because of more extensive missingness that may have influenced the model results and comparisons. In addition, at the hourly level, all the 12 included terms had approximately 1% of values missing.

Figure 2 shows the distribution of daily hit counts collapsed across all states for each of the 12 mental health search terms included in the modeling analysis. Distributions of the outcomes varied in central tendency and skewness, but in general, the data were nonnormal and overdispersed, suggesting appropriate use of a negative binomial model fit in GAMM analyses. Figures S1 and S2 in Multimedia Appendix 3 plot the distributions of the daily hit counts collapsed across all states for the physical health terms both related and unrelated to COVID-19 included in modeling.

After preprocessing the raw Google Trends data for the non–COVID-19 physical health terms, 17 terms (“allergic,” “angina,” “apnea,” “bleeding,” “blister,” “bruising,” “constipation,” “discharge,” “fracture,” “hemorrhage,” “incontinence,” “inflammation,” “itching,” “lesions,” “rash,” “spasms,” and “swelling”) had <5% of values missing for daily aggregated hit counts. In addition, after preprocessing the raw Google Trends data for the COVID-19–related physical health terms, 9 terms (“bloating,” “congestion,” “coughing,” “diarrhea,” “dizzy,” “fainting,” “fever,” “strep,” and “stuffy”) had <5% of values missing for daily aggregated hit counts. Accordingly, the 17 non–COVID-19 physical health terms and the 9 COVID-19–related physical health terms were zero-imputed and incorporated into subsequent modeling. Table S2 in Multimedia Appendix 1 displays the full results of the missingness analyses for physical health terms both related and unrelated to COVID-19.

A GAMM model was run, in which the search volume of each of the 12 mental health search terms served as the outcome. The full summary of the model output can be obtained from the authors upon request; however, this analysis focused on the smooth terms that correspond to mask mandate and the interaction of mask mandate with political party elected, as outlined in the GAMM Approach section. The results for the smooth term corresponding to mask mandate (s₃) can be found in Table 1. The search volume for 17% (2/12) of search terms, “anxiety” and “restless,” had significant nonlinear associations with the timing of mask mandates. Analysis of the 95% CI for the estimated change in search volume for “anxiety” revealed that the trend in search volume did not differ significantly from those when the mask mandate was first implemented as the CI contained 0 at all values. However, the trend in search volume for “restless” was found to be significant. As shown in Figure 3, the trend is characterized by a steady increase directly after the mask mandate was implemented, peaking at approximately 400 searches above the initial search volume at approximately 6 months after the mandate and leveling off shortly thereafter.

The results for the smooth term corresponding to the interaction of mask mandate and political party affiliation (s₄) can be found in Table 1. The search volume for 58% (7/12; “afraid,” “anxiety,” “insomnia,” “irritable,” “restless,” “suicide,” and “tense”) search terms had significant nonlinear associations with the timing of mask mandates for Republican states in comparison with Democratic states. Figure 4 plots the associations among these search terms. Of these 7 search terms, 5 (71%; “afraid,” “insomnia,” “restless,” “suicide,” and “tense”) did not differ significantly from baseline (0 contained in the 95% CIs) and, thus, are not plotted in Figure 4. Of the remaining 2 terms, the search volume of “anxiety” decreased significantly after the mask mandate for Republican states relative to Democratic states, dropping to roughly 65 searches below baseline after 6 months. In addition, the search volume of “irritable” significantly increased after the mask mandate for Republican states relative to Democratic states, increasing to roughly 25 searches above baseline after 5 months before leveling off. In sum, although 58% (7/12) of the terms exhibited significant nonlinear associations with the timing of mask mandates for Republican states relative to Democratic states, only “anxiety” and “irritable” exhibited changes in volume that differed significantly from their respective baseline levels.

Figure 5 displays the number of states with mask mandates relative to the date of mask mandate implementation, or when the mandates first became effective in each state. A total of 39 states had mask mandates initially, with 16 of these states being Republican. After 210 days, or roughly 7 months, 28 of these mandates were still in effect, with 9 of these mandates being in Republican states. In the context of this analysis, as the associations shown in Figures 3 and 4 represent the difference between states with and without mask mandates and between Republican states with mandates and Democratic states with mandates, respectively, there were more states “driving” these associations earlier in time relative to mandate implementation than later in time. It is important to acknowledge the attrition of mandates over time as any significant changes in search volume that happened earlier in time relative to mandate implementation reflect changes that were common across a greater number of states, whereas any significant changes in search volume occurring later in time reflect changes that were common across fewer states. For instance, 210 days following the mandate, only the 9 Republican states with mask mandates were factored into the model that yielded the associations in Figure 4 compared with 16 states at day 0.

Table 1 shows the results of assessing the model fit, in which the marginal R² and conditional R² were calculated for each search term model. As indicated by the marginal R² values, the fixed effects of the models accounted for roughly between 11% and 20% of the variance in search term volume across models. The conditional R² values show that the combined fixed and random effects of the models explain roughly between 50% and 76% of the variance in search term volume. These values suggest that (1) the models explain a considerable portion of the variance in mental health search term volume and, (2) although the random effects may explain most of the variance in the outcome for each model, the fixed effects still explain a sizable portion of the variance in the outcome. This latter point provides confidence that the associations between mask mandate and political party elected revealed in Figures 3 and 4 are unlikely to be owing to chance or random noise in the data.

Additional GAMMs were fit for physical health search terms to determine whether the associations between mask mandate and political party elected were unique to mental health–related search term activity. For the physical health terms not related to COVID-19, a total of 17 GAMMs were fit—one model with the search volume of each individual term as the outcome. In an identical manner to the mental health search terms, the analysis focused on the smooth terms that corresponded to mask mandate and the interaction of mask mandate with political party elected. The search volume of 24% (4/17) of the search terms—“allergic,” “bruising,” “constipation,” and “rash”—had significant nonlinear associations with the timing of mask mandates. Figure S1 in Multimedia Appendix 4 plots the associations among these search terms. Regarding these 4 terms, the search volume for “allergic” and “rash” did not deviate significantly from baseline at any point (0 in the 95% CI at all time points). However, the search volume for “constipation” increased slowly after the implementation of a mask mandate, peaking approximately 7 months following the mandate. The search volume for “bruising” initially decreased in the 4 months after implementation of a mask mandate, but the volume approached baseline levels in the 4th through 7th months following the mandate, still staying below baseline levels. In addition, the search volume for 12% (2/17) of terms, “constipation” and “fracture,” had significant nonlinear associations with the timing of mask mandates for Republican states in relation to Democratic states, plotted in Figure S1 in Multimedia Appendix 5. The search volume for “fracture” did not deviate significantly from baseline at any point, but the search volume for “constipation” steadily decreased after the implementation of the mask mandate, ultimately becoming significantly different from baseline after 6 months.

For the COVID-19–related physical health terms, 9 GAMMs were fit, with each model using the search volume of 1 of the 9 COVID-19–related physical health search terms as the outcome. The search volume of 22% (2/9) of search terms, “bloating” and “coughing,” had significant nonlinear associations with the timing of mask mandates, plotted in Figure S1 in Multimedia Appendix 6. The search volume for “bloating” decreased steadily after the implementation of the mask mandate, whereas the search volume for “coughing” did not significantly deviate from baseline at any point and, thus, was not plotted. In addition, the search volume of 78% (7/9) of terms—“bloating,” “congestion,” “coughing,” “dizzy,” “fainting,” “fever,” and “strep”—had significant nonlinear associations with the timing of mask mandates for Republican states in relation to Democratic states. However, the search volume for none of these terms was significantly different from baseline levels, and thus, no terms were plotted for the associations with the timing of mask mandates for Republican states in relation to Democratic states.