State-Level COVID-19 Symptom Searches and Case Data: a Quantitative Analysis of Political Affiliation as a Predictor for Lag Time Using Google Trends and CDC Data.

Abstract

Across each state. the emergence of the COVID-19 pandemic in the United States was marked by policies and rhetoric that often correspond to the political party in power. These diverging responses have sparked broad ongoing discussion about how the political leadership of a state may affect not only the COVID-19 case numbers in a given state, but also the subjective individual experience of the pandemic. This study leverages state-level data from Google Search Trends and CDC daily case data in order to investigate the temporal relationship between increases in relative search volume for COVID-19 symptoms and corresponding increases in case data. I aimed to identify whether there are state-level differences in patterns of lag time across each of the four spikes in the data (RQ1) and whether the political climate in a given state is associated with these differences (RQ2). Using publicly available data from Google Trends and the Centers for Disease Control and Prevention, linear mixed modeling was utilized to account for random state-level intercepts. Lag time was operationalized as number of days between a peak (a sustained increase before a sustained decline) in symptom search data and a corresponding spike in case data and was calculated manually for each of the four spikes in individual states. Google offers a dataset that tracks the relative search incidence of more than 400 potential COVID-19 symptoms, which is normalized on a 0-100 scale. I used the CDC's definition of the eleven most common COVID-19 symptoms and created a single construct variable that operationalizes symptom searches. To measure political climate, I considered the proportion of 2020 Trump popular votes in a state as well as a dummy variable for the political party that controls the governorship and a continuous variable measuring proportional party control of federal Congressional representatives. The strongest overall fit was for a linear mixed model that included proportion of 2020 Trump votes as the predictive variable of interest and included controls for mean daily cases and deaths as well as population. Additional political climate variables were discarded for lack of model fit. Findings indicated evidence that there are statistically-significant differences in lag time by state but that no individual variable measuring political climate was a statistically-significant predictor of these differences. Given that there will likely be future pandemics within this political climate, it is important to understand how political leadership affects perceptions of and corresponding responses to public health crises. Although this study did not fully model this relationship, I believe that future research can build on the state-level differences that I identified by approaching the analysis with a different theoretical model, method for calculating lag time, and/or level of geographic modeling.