Table 1.
Four categories of input features for the machine learning models.
Fig 1.
The architecture of the DNN model for estimating alcohol sales.
Fig 2.
Total spirits sales in 13 states: (a) sales in ethanol per capita from January 2018 to June 2020 (the dashed lines indicate the month of March in 2018, 2019, and 2020); (b) sales changes in March, April, May, and June 2020 as compared with the average sales in the same months during 2018 and 2019.
Fig 3.
Spirits sales in each of the 13 states (the dashed lines indicate the month of March in 2018, 2019, and 2020).
Fig 4.
Geographic differences for how the per capita sales of spirits changed during the pandemic.
The percentage change for each state is compared to the average sales during the same months in 2018 and 2019. Red colors indicate an increase in spirits sales, while blue colors indicate a decrease in sales. The darker the colors, the larger the changes.
Fig 5.
Total wine sales in thirteen states: (a) sales in ethanol per capita from January 2018 to June 2020 (the dashed lines indicate the month of March in 2018, 2019, and 2020); (b) sales changes in March, April, May, and June 2020 as compared with the average sales in the same months during 2018 and 2019.
Fig 6.
Geographic differences for how the per capita sales of wine changed during the pandemic.
The percentage change for each state is compared to the average sales during the same months in 2018 and 2019.
Fig 7.
Total beer sales in eleven states: (a) sales from January 2018 to June 2020 (the dashed lines indicate the month of March in 2018, 2019, and 2020); (b) sales changes in 2020 compared with the average sales in the same month of 2018 and 2019.
Fig 8.
Geographic differences for how the per capita sales of beer changed during the pandemic.
The percentage change for each state is compared to the average sales during the same months in 2018 and 2019.
Fig 9.
Visitation behavior to the four types of alcohol outlets in the U.S. during the first months of COVID-19.
(a) Time series describing the per capita visits for each month from January 2018 to June 2020 (which is averaged across 16 states). The dashed lines indicate the month of March in 2018, 2019, and 2020. Note that visits to drinking places are divided by 2 so that all curves can be shown within a similar scale for clear visualization. (b) For the months of March, April, May, and June 2020, we depict the changes to per capita visits. For each month, the percentage change was computed relative to the average value for the same month of 2018 and 2019.
Fig 10.
Visits to the four types of alcohol outlets in each of the 16 states.
The vertical dashed lines indicate the month of March in 2018, 2019, and 2020.
Fig 11.
Geographic differences for how the per capita visits to liquor stores changed in response to COVID-19.
For each month, the percentage change to visits was computed relative to the average for the same month in 2018 and 2019. Red colors indicate an increase in sales, while blue colors indicate a decrease in sales. The darker the colors, the larger the changes.
Table 2.
VIF values of the variables in three MLR models.
Fig 12.
Performances of the three machine learning models based on ten-fold cross-validation and walk-forward validation: Multiple linear regression (MLR), random forest (RF), and a deep neural network (DNN).
Table 3.
RMSE values of including and not including grocery store visits for alcohol sales prediction based on the RF model.
Fig 13.
Model estimates and the recorded per capita sales of spirits in individual states.
The vertical dashed lines indicate the month of March in 2018, 2019, and 2020. The model was trained on data before March 2020 and was then used to predict the sales for the months following (and including) March 2020.
Fig 14.
Feature importance of the numeric input variables in the optimized RF models for the estimation of spirits, wine, and beer sales.
Fig 15.
Partial dependence plots: (a) maximum latitude and spirits sales; (b) minimum longitude and wine sales; (c) month and beer sales.