Our analytical strategy was based on the following steps: first, linear regression models (OLS) accounting for fixed-effects (FE) were fitted, with weekly alcohol consumption as the outcome variable and the lockdown period as a main independent variable. Additionally, employment status was accounted for as a potential time-varying confounder (whereas time-constant confounders dropped out of the models). FE models focus on the changes within individuals throughout the observation period, net of time-constant unobserved confounding. Thus, they allow us to infer the impact of the lockdown on alcohol consumption. Based on this FE model, predicted margins estimated alcohol consumption levels at every time point. Sensitivity analyses with FE Poisson models were carried out, due to the large number of zero values of the outcome. As shown in the appendix -see table A2 and figure A2 -, the patterns are practically identical, although the coefficients differ because Poisson models are expressed different units, which makes them harder to interpret. In contrast, an advantage of OLS models is that they allow to interpret the coefficients in the same units of the outcome. For this reason, we decided to stick to OLS models.
