Standard errors increase slightly in some estimations and decrease slightly in others, where our estimates in Tables 2, 3 and 4 that are significant at p < 0.01 remain so in every instance and are robust to variation in the distance of spatial correlation and the decay structure. As a check on our controls, we also employ k-fold LASSO , a machine-learning based regularization technique to select controls based on a penalty function that, along with minimizing the sum of the squared errors in the regression, penalizes the sum of the absolute values of the coefficients on variables in the estimation. All our controls are retained under a LASSO estimation with k=10. Additionally, we run Oster bounds tests to estimate the likelihood that endogeneity problems could affect our estimated relationships between COVID outcomes and political identity. Oster’s δ reflects the ratio of influence from unobserved variables relative to observed controls that would have to hold true to render the independent variable of concern statistically insignificant. We consistently find high and negative deltas in our estimates, with δ >1 holding for all results even at the stringent Rmax = 1 standard, indicating statistical significance through endogeneity to be extremely unlikely. We highlight three important conclusions from the empirical estimations in our research: First, our results suggest that the health costs of political identity in the United States during the COVID-19 pandemic have been high.
Using U.S. county-level data,cannabis growing system we estimate that a 10 percentage point increase in the county Trump vote to be associated with a 3.9 percentage point decrease in public mask-wearing, a 5.1 percentage point decrease in COVID vaccination rates, and a 0.23σ reduction in a COVID-safety index. Second, our results suggest that these differences in safety behavior across political identities have had large impacts on COVID cases and deaths. For every additional 10 percentage points in the county Trump vote, we estimate an increase of 1,394 COVID cases per 100,000 county residents and 27.6 COVID deaths . Thus, our results show a 10 percentage point increase in county-level Trump voter support in the 2020 Presidential election linked to a roughly 11% increase in COVID cases and 12.6% increase in deaths during the first twenty months of the pandemic. Third, we find indices of more traditional strains of American political conservatism– specifically social conservatism and libertarian conservatism–to exhibit low levels of explanatory power over COVID-safety behaviors, cases, and deaths after controlling for Trump voter support in the 2020 general election. Because our U.S. county data are observational rather than experimental, it is important to consider whether other factors that characterize high-Trump-support counties–apart from differences in COVID-safety behaviors–could be responsible for elevated COVID cases. The possibility of such unobservable confounders is included in our directed acyclic graph in Figure 1. To bias the present results, these unobservable confounders would have to 1) vary at the county level with Trump support , 2) fall outside of demographic, ethnic, economic, and co-morbidity controls in the estimations, and 3) result in higher levels of COVID infection. By the nature of COVID transmission, this unobserved confounder would almost certainly have to be related to higher levels of human contact.
A leading possibility would be that those living in Trump-voting counties have more frequent social contact with others in their communities, for example through higher frequency of church attendance or social gatherings, than those in non-Trump counties in their state. We view the possibility of our results being driven by this kind of unobservable confounder as unlikely for a number of reasons. One is that our estimates control for differences in county population density, which naturally affects the frequency of human interaction. The second is that church attendance is strongly correlated with the social conservatism index we include in Table 4. If higher levels of church attendance were responsible for higher levels of COVID infection, then this estimate would very likely be positive and significant, yet the estimate is actually negative . Our estimates of Oster bounds also appear to preclude results being driven by an unobserved confounder, where results on Oster’s delta make endogeneity-driven results extremely unlikely. Finally, the decline in mask-wearing and corresponding increase in COVID cases that we observe in the U.S., based on differences in political identity, is strikingly similar to the estimated causal effect of mask-wearing on symptomatic COVID infection found in the most substantive randomized trial on the effects of the mask wearing intervention. Put simply, there is little variation in U.S. COVID infections left to explain once one accounts for differences in COVID-safety behavior. The nature of how COVID-safety behavior became intertwined with political identity in the United States is outside our present scope, but has been studied closely in other work. Likely channels linking messaging from the Trump administration to a lack of adherence to COVID-safe behaviors include studies of both social media and conservative television outlets . This has occurred during the present social and political era in which issues, that in the past might have been viewed as non-partisan, quickly manifest polarizing stances.
Indeed the climate of extreme U.S. political polarization has created a tendency for views about how to address new global challenges to become politically polarized, both legislatively and socially, in ways they otherwise might not be . That COVID-safety behaviors have become subject to this U.S. political polarization has inflicted a tremendous aggregate cost to American health and lives. The estimates in Table 3 suggest that a difference in political identity consistent with a 10 percentage point higher county Trump vote projects over the U.S. population to an increase in approximately 4.60 million COVID cases 95% CI=[3.18m, 6.02m] and 91,000 COVID-related deaths 95% CI=[46,300, 135,800] in the first twenty months of the pandemic. These added infections and deaths would appear particularly tragic as non-adherence to health-safety behaviors does not seem to necessarily contravene deeply rooted values nested within traditional American conservatism. Indeed, there are numerous issues in which conservative mores advocate a sacrifice of individual liberty for the benefit of the larger community or nation, including restrictions on behaviors that cross a moral line, military service, and an adherence to law and order more generally.Normative prescriptions for an individual behavior serving the public interest may be equally vilified as an affront to liberty or hailed as an act of patriotism. In this research we do not find differences in COVID-19 safety behaviors as originating from political identity rooted in libertarian or social conservatism, but rather in the manner of their politicization by political leadership. Studies were included if they examined the association between depression or antidepressant use and motor vehicle crash risk; used an epidemiologic design , and were published in English between January, 1995, and October, 2015. Cross-sectional studies, qualitative studies, reviews, commentaries, opinion pieces, and magazine articles were excluded. In this review,hydroponics rack system depression was ascertained through questionnaires such as the Geriatric Depression Scale , General Health Questionnaire or their modified versions, as well as interviews and claims-based databases. Antidepressants included classes of medications prescribed for the treatment of depression such as tricyclics , monoamine oxidase inhibitors , selective serotonin reuptake inhibitors , serotonin-norepinephrine reuptake inhibitors , and atypicals .All studies that met the inclusion criteria were assessed for quality using the Newcastle-Ottawa Scale as recommended by the Cochrane Collaboration on bias assessment . For the study designs included in this study, the best possible score is 9. Better quality studies will have higher scores. Heterogeneity was assessed using the Q and I2 tests, with P ≤ .05 and I2 > 0.5 considered homogenous. Funnel plots were used to assess publication bias. Data abstracted from each study were entered in the Comprehensive Meta-Analysis software to compute the individual odds ratios and a summary odds ratio for each of the two analyses. A fixed effects model was used unless heterogeneity was present, in which case, a random-effects model would be preferred. Two forest plots were created; one to show the distribution of the effect of depression on car crash risk and another one to show the distribution of the effect of antidepressant use on car crash risk .Seven studies investigating the effect of depression on driving met the inclusion criteria . Rainio et al conducted a retrospective review of crashes in 2001 and 2002, using interviews of police, surviving drivers, and family members, medical records, and autopsy reports.
A physician on the team made the determination of the presence of depression. This Finnish study used the reports of the road crash investigation teams which investigate every fatal crash in Finland. Of the 640 crash deaths analyzed, 390 were drivers considered to be at fault. Of these 390, 6.4% had depression, while one non-fault driver and none of the non-fault passengers had depression . No medication data were included in the analysis. Sagberg used self-report questionnaires from 4448 crash-involved drivers in Norway and found an OR of 2.43 in persons reporting depression. Though antidepressants were included, the relationship between antidepressants and depression was not included in the analysis. Mann et al did a cross sectional telephone survey of adults in Ontario, Canada aged 18 and older. Depression and anxiety were determined using two sub-scales of the 12-item General Health Questionnaire: Depression-anxiety and social functioning. With a sample of 4935 adults, Mann et al found the odds of crash involvement increased significantly with an increase in the anxiety/ depression score, with a five percent increase in the risk of crash involvement for every unit of anxiety/depression increase . Medications for depression were not included in the study. Wickens et al , also with an Ontario sample of 12,830, studied self-reported anxiety and mood disorders, queried via telephone survey, and found an increase in crashes . There was no report of antidepressant use. Sims et al , in a prospective analysis of 174 older adults, used baseline and one year follow-up inperson assessments. They found that a Geriatric Depression Scale score ≥16 was associated with an increased crash rate . While antidepressant use was assessed, the relationship between depression, antidepressants and crashes was not reported, and in this study the relationship between antidepressant use and crashes was not significant. LeRoy found an OR of 3.99 for crash risk with depression in a case control study of 81,408 cases and 244,224 controls using claims-based population data. Another study of older adults, using the Geriatric Depression Scale, included depression in their analysis and found no association; however, this study was small, with a low prevalence of depression in the study population .Neither of the studies by Orriols included data on depression. Rapaport did a case-only time to event analysis in 159,678 persons in Ontario, also using provincial databases, and found that SSRI and SNRI antidepressants were associated with at-fault crashes alone , in combination with a benzodiazepine , and with an anticholinergic medication . The period within the first 3-4 months of antidepressant initiation also was associated with increased risk of crash . Bramness , using national databases, found non-sedating antidepressants increased the risk for traffic crashes more than sedating antidepressants . Neither Rapaport nor Bramness include depression as a variable. Sagberg , in a Norwegian study of 4,448 crash involved drivers, found the self-reported use of antidepressants had an adjusted OR for crash of 1.70 . While this study, and the following two studies included both depression and anti-depressants, the relationship between the two variables was not assessed. LeRoy , in a US-based case-control study of 5,398 cases and 16, 194 controls, using national survey and also claims-based databases, found that serotonin-2 antagonist/reuptake inhibitors had an OR for crash of 1.90 . Barbone , in a case crossover study in the UK, used a local prescription database and found no association between crash and tricyclic or SSRI antidepressants in the 1998 study of 19,386 crashes. Ravera in the Netherlands study of 3,963 cases and 18,828 controls used three existing Dutch population-based databases and found an OR of crash with SSRI of 2.03 . Gibson , using a primary care database, found that SSRI had no effect in the short term , but had a small increase in the risk of crash with extended use in a case crossover study of 49,821 crashes. None of these three studies included depression as a variable.Three cohort ; Margolis et. al ; Sims et.al and three case control studies ; Sagberg ; Wickens et. al that examined the association between depression and car crash risk were included in the quantitative meta-analysis .