Patterns of Alcohol Consumption Among Individuals With Alcohol Use Disorder During the COVID-19 Pandemic and Lockdowns in Germany

Key Points Question Are COVID-19 lockdown measures associated with alcohol consumption (AC) and temporal patterns of AC? Findings In this cohort study of 189 participants who met the criteria for alcohol use disorder (AUD), high-frequency AC tracking comprising 14 694 smartphone ratings revealed no immediate negative association of lockdown measures with overall AC. Independent of the lockdown, intention to control AC was associated with less AC; however, a difference between AC on weekends vs weekdays decreased during lockdown measures and in individuals with severe AUD. Meaning Both holidays and weekly patterns were associated with drinking intention and lockdown measures, reflecting losing and regaining control over AC; these patterns may serve as targets for prevention and intervention of AUD.


eAppendix 2. Participant Recruitment and Characteristics
The recruitment process of this study had been continuous. The decision to start the analysis with the data beginning 10/02/20 was based on two considerations. First, for statistical power reasons and following multilevel analysis guidelines 4 as well as recent simulation studies 5 , up to this point in time a sufficient number of participants (n = 105) had already been included in the study to be able to detect the hypothesized effects. Second, with the data beginning 10/02/20, the coverage of a sufficient time period across the second wave of the COVID19 pandemic in Germany was achieved, with roughly comparable length of time intervals across distinct lockdown stages: 4 weeks pre lockdown; 6 weeks light lockdown; 10 weeks hard lockdown.
We examined potential differences between included and excluded participants (see eFigure 2) according to age, gender, and AUD-criteria. In particular, we compared the sample of included participants with people eligible and with at least 2 AUD-criteria but wished for no further contact. To test differences in AUD criteria, we conducted a Mann-Whitney-U test and found no evidence that excluded participants fullfilled a higher number of AUD criteria compared to included participants (P =.419). Moreover, we found the excluded participants to be younger (median = 32 years (IQR, 24-43) than included participants (median = 37 years, IQR 27.5-52) (P=.003). Conducting a Chi 2 -test, no significant difference was observed for the variable sex (P = .728).
None of the included participants dropped out during the 5 month study period. Data delivery differed between lockdown phases. In particular, 105 participants delivered data during the pre-lockdown, 146 participants during the light lockdown, and 189 participants during the hard lockdown. Multi-level models have been shown to be well suited to deal with data structure characterized by different amounts of data points. 4 eAppendix 3. e-Diary Items The data frame used to compute the multilevel models was structured as follows. First, we restructured the main outcome variable of interest, i.e., alcohol intake rated every other day over the previous 2 days separately. This resulted in a continuous data set with a daily resolution, i.e., each row represented one day including the alcohol consumption at this particular day. Second, we restructured the intentions to drink item. In particular, this resulted in the time frame referenced (i.e., over 8 days) of the intention predictor preceding the time frame for the outcome (i.e., alcohol consumption). Third we restructured the social isolation item by extending it towards the preceding day.

eAppendix 4. Statistical Analyses and Results
We estimated intraclass correlation coefficients (ICC) applying unconditional models. We received ICCs of 0.25 (AC), 0.60 (perceived social isolation), and 0.59 (drinking intention), indicating that about 75% of variance in AC, and each about 40% of variance in perceived social isolation and drinking intention was attributable to withinsubject fluctuations (level 1 in our hierarchical statistical model).

eAppendix 5. Main Multilevel Results
Multilevel statistics showed that at weekends compared to weekdays, AC was significantly heightened (ßcoefficient = -11. 38; 95%CI, 10-12.77; P < .001;  Table 2 and Figure 1 main text). Perceived social isolation had no statistically significant effect on alcohol use (ß-coefficient = -1.31; 95%CI, -2.89-0.27; P = .104). During the lockdown phase with hard restrictions, AC was decreased by 5.45 gram compared to pre lockdown (ß-coefficient = -5.45; 95%CI, -8 --2.9, P = .001; Table 2 and Figure 1 main text). Between pre lockdown and the light lockdown, no significant difference on AC was found (ß-coefficient = -1.30; 95%CI, -3.94-1.33; P = .333; Table 2 and Figure 1 main text). To assess the difference between the light lockdown and the hard lockdown phase, the same model was performed with the light lockdown as the reference group resulting in a significantly lower AC during the hard lockdown compared to the light lockdown (ß-coefficient = -4.15; 95%CI, -5.95 --2.35, P < .001). Drinking intention was negatively associated with AC (ß-coefficient (no more AC than usual) = -3.97; 95%CI, -6.56 --1.38; P = 0.003; ß-coefficient (less AC than usual) = -11.10; 95%CI, -13.63 --8.58; P < .001; reference = no particular resolutions; Table 2 main text), indicating that when participants intended to consume less alcohol, this resulted in a reduction in the amount of alcohol consumed. Translated to practice, if participants plans to limit their "alcohol consumption for the next eight days?" were indicated as "Yes, I want to drink not more than usual" or "Yes, I want to drink less than usual", these ratings were associated with less alcohol consumption self-reported every other day compared to participants response "No, I don´t have any particular resolutions".

Moderation analyses
To test whether the effect of drinking intention was different across lockdown measures, we applied multilevel moderation analysis, see equation 3: = 00 + 10 * + 20 * + 30 * * + µ + These analyses of the interaction effect of intention to drink alcohol and lockdown measures on AC revealed no significant effect (F(3967) = 0.63; P = .638), neither for intention and light lockdown, nor for intention and the hard lockdown (eTable 8). The results indicate that the intention to drink alcohol had a significant effect on consumption regardless of the lockdown measures. Please note that, following established procedures to maximize statistical power in multilevel moderation analyses, 4,5 we computed interactions in models with a reduced set of predictors.
To explore whether weekend drinking cycles of AC were influenced by AUD severity, we computed a multilevel moderation analysis as detailed in the following equation 4: ( ℎ ) = 00 + 01 * + 10 * + 20 * * + µ + We found a significant interaction effect for AUD criteria * weekend (F(12e 3 ) = 5.02; P = .025), i.e., participants with more AUD criteria showed a smaller difference of AC between weekend days and weekdays (eTable 9).
For interpretation and visualization purposes (see Figure 2a in the main text), we recomputed the moderation analyses AUD criteria * weekend using AUD categories of the DSM-5 7 (2-3 criteria = mild, 4-5 criteria = moderate, ≥ 6 criteria = severe). Again, we found a significant interaction effect (F(12e 3 ) = 4.89; P = .008) revealing significant differences between AC at weekend days vs. weekdays in the mild (difference in AC: 12.86 grams alcohol per day; SE = 0.93; P < .001), moderate (difference in AC: 11.50 grams alcohol per day; SE = 1.03; P < .001) and severe AUD groups (difference in AC: 6.60 grams alcohol per day; SE = 1.77; P < .001), respectively. The interaction effect was driven by differences between the mild and severe AUD groups (eTable 10).
To enable a more fine-grained interpretation we plotted the simple effects, i.e., AC at weekdays and weekend days by AUD category (Figure 3a). To explore whether weekend drinking cycles of AC were influenced by lockdown measures, we computed a further multilevel moderation analysis as detailed in the following equation 5: To enable a more fine-grained interpretation we plotted the simple effects, i.e., AC at weekdays and weekend days by lockdown phase (Figure 3b in the main manuscript).
We conducted three supplementary interaction analyses to examine a potential moderating role of age on the associations between i) lockdown measures, ii) weekend drinking cycles, iii) intention to drink and AC. While there was no moderation effect of age on the associations intention to drink and AC (F(4161) = 0.03; P = .9732), age significantly moderated the associations between lockdown measures and AC (F(4337) = 3.68; P = .0252); i.e., the older the participants were, the smaller the reduction of AC in the lockdown phases was. Moreover, age significantly moderated the association between weekend drinking cycles and AC with reduced weekday vs weekend AC differences as a function of increasing age (F(12000) = 12.72; P < .001); i.e., the older the participants, the smaller the weekday vs weekend AC difference was.

Holiday season as an additional lockdown phase
We conducted a supplementary analysis treating the hard lockdown as consisting of two distinct time periods. In particular, we extended our predictor lockdown and introduced the time periods between the build-up to Christmas and New Years ((2020/16/12) to (2021/01/02)) as an additional category. Results showed no difference between pre lockdown and the "holiday phase", but again confirmed the reduced consumption during the hard lockdown (see eTable 5)

Multi-level model with log-transformed AC outcome
Given the daily life consumption behavior of our sample (i.e., the at risk yet not highly alcohol-dependent participants did not consume alcohol each and every day), AC as the main outcome variable of interest is skewed and does show outliers. Therefore, we conducted a robustness check. Following established procedures for transforming such skewed data, 4,8,9 we log-transformed all AC-values using the natural logarithm and adding "1": logn (AC-values + 1). We recomputed the multi-level model and received the contentual same results compared to the multilevel model with non-transformed AC-values (see eTable 6). To judge whether this multilevel model is suited for dealing with the given data structure, we followed established procedures 4 and thus examined level-1 (assessment-level) residuals measuring deviations from the conditional mean (conditional residuals) which we derived from our multilevel model. Visual inspection confirmed that there was no obvious deviation from a normal distribution of the residuals. (see eFigure 4).

Controlling for between person effects
Following Wang and Maxwell (2015) 6 , we computed a multilevel model including all between-person effects (i.e., each within-person predictor aggregated on a between-person mean value, respectively) as a supplementary analysis. None of the between-person predictors yielded significant effects (see eTable 4 -In principle, staying in public is allowed only with the members of the own and one other household, a maximum of 10 people may come together.
-Citizens are asked to generally refrain from non-essential private travel and visits, including those of relatives. Overnight accommodations within the country will be made available only for necessary travel and explicitly non-touristic purposes.
-Institutions and facilities that are classified as recreational are to be closed, including recreational and amateur sporting operations at and in all public and private sports facilities -with the exception of individual sports alone, in pairs or with members one's own household only. -Restaurants and bars, clubs, pubs and similar establishments are to be closed. This excludes the delivery and collection of takeaway food for consumption at home and the operating canteens in workplaces. 2 hard lockdown (12/16/2020-02/28/2021) -Private meetings are limited to a maximum of 5 people from a maximum of two households for private meetings. -Retail and service businesses, such as hair salons and beauty salons, are to be closed from December 16. Exceptions apply to grocery stores, drugstores, pharmacies, opticians, gas stations, auto repair shops, banks, post offices, dry cleaners and Christmas tree dealers. Medically necessary treatments such as physical therapy also remain available. -At schools, contacts are to be significantly restricted from Dec. 16 until Jan.
10, 2021. Children are to be cared for at home whenever possible during this period. Therefore, schools will generally be closed or attendance will be suspended during this period. Emergency care is provided and distance learning is offered. An analogous approach is taken in daycare centers. Additional opportunities are created for parents to take paid leave for childcare.
-Employers are urged to consider whether business sites can be closed either through company vacations or generous home office solutions from Dec. 16 to Jan. 10, 2021.
-In the food service industry, delivery and pickup of take-out meals for home consumption by food service establishments will remain available. On-site consumption is prohibited.