Background. Only 65.4% of the eligible population is vaccinated against COVID-19. Vaccine hesitancy is a complex phenomenon that plays an important role in explaining the low vaccination rates. Having a good understanding of the factors that drive hesitancy can improve our capacity to address it. In this study I use regularized regression to identify the most important predictors of vaccine hesitancy.

Methods. The CDC estimated the proportion of people (18 and over) hesitant or unsure to take the COVID-19 vaccine at the county level; May 2021, data available at HHS-ASPE. The prediction set included demographic variables (age, race, and education), the CDC social vulnerability index which measures the relative stress people is under in a community, and the proportion of votes the Republican party received in the 2020 presidential election collected by MIT as a proxy of political affiliation at the community level. I fitted a LASSO regression model, implemented via leave-one-out cross-validation to find the penalization that minimizes the mean average error (MAE) and excludes variables without explanatory power. I used bootstrap with one-thousand iterations to estimate coefficients’ confidence intervals.

Results. The final dataset included 3,111 counties. At optimum, the MAE was 2.7%, which denotes average prediction error of vaccine hesitancy. Among the most important drivers of hesitancy were proportion of people with some years of college (Coef: 12%; 95%CI: 7%, 16%), less than high-school diploma (10%; 5%, 15%), proportion of Black/African American (8%; 7%, 10%), and political affiliation (7%; 6%, 9%). The variables that reduce hesitancy were proportion of Asian population(-26%; -33%, -21%), people aged 65 and more (-21%; -26%, -13%), college graduates (-19%; -22%, -16%), and males (-17%; -25%, -9%).

Conclusions. The model demonstrated good prediction properties. These results can help in deepen our understanding of the drivers of vaccine hesitancy, especially in acknowledging its multifactorial nature.