Impact of COVID-19 Confinement on Alcohol Purchases in Great Britain: Controlled Interrupted Time-Series Analysis During the First Half of 2020 Compared With 2015–2018

Study design

We undertook controlled, interrupted time series regression analyses of the impact of the introduction of confinement in Great Britain (26 March 2020) on British household purchases of alcohol (including no- and low-alcohol products), using purchases made during the same calendar time period for 2015–2018 (averaged) as our control. We inferred that confinement would lead to excess purchases during confinement above the average level purchased during the same time period for 2015–2018, both to compensate for closure of bars, restaurants, etc.,—‘on-licenced premises’, and to use alcohol as an anxiolytic. Thus, we examine abrupt and persistent level changes in alcohol purchases, due to an event, the introduction of COVID-19 confinement.

Data source

Our data source is Kantar Worldpanel’s (KWP) household shopping panel. KWP comprises ~30,000 British households at any one time, recruited via stratified sampling, with targets set for region, household size, age of main shopper and occupational group. Households record all purchases (from all store types, including Internet shopping) brought into the home using barcode scanners. To be included in KWP’s final datasets, households must meet quality control criteria (meeting thresholds for data recording and purchasing volume/spend (based on household size) every 4 weeks). Panellists also upload digital images of checkout receipts, which KWP use to verify the accuracy of scanner data.

Households report the age of the main shopper, the number of adults in the household, income, social class and life stage of the household (e.g. with young family, retired, etc.). We grouped households into the following:

• (a) five age groups of the age of the main shopper: 18–34; 35–44; 45–54; 55–64 and 65+ years (which correlated with life stage of the household);

• (b) five social class groupings, AB, C1, C2, D and E, based on a system of socio-demographic classification used in the UK, originally developed by the National Readership Survey to classify readers. AB includes higher and intermediate managerial, administrative or professional workers; C1 supervisory or clerical and junior managerial, administrative or professional workers; C2 skilled manual workers; D semi-skilled and unskilled manual workers and E states pensioners, casual and lowest grade workers, and unemployed with state benefits only (National Readership Survey, 2019);

• (c) five income groups per adult per household per year of: £0–7.5 k, >£7.5–12.5 k; >£12.5–17.5 k, >£17.5–25 k and > £25 k; and

• (d) five geographic areas representing eleven regions across Great Britain, based on UK Office for National Statistics classifications for truncated postcode data look-ups: South England (South West, London, South East and Eastern regions); Mid England (West Midlands and East Midlands regions); Wales; North England (North West, North East and Yorkshire & Humber regions) and Scotland.

The distributions of household socio-demographic characteristics for all households that made at least one alcohol purchase for the first 6 months of 2015 to 2018 (averaged) and 2020 are shown in Table 1. Compared with the average of the first 6 months of each of 2015–2018, households for the first 6 months of 2020 tended to have a higher age of main shopper and a higher household income, but similar distribution of social class group and geographic area. We thus included the age of main shopper and household income as covariates in our models. We also undertook analyses separately for each household grouping, age, class group, household income and geographic area.

We obtained KWP data on all off-trade purchases of alcohol products in Great Britain, including non-alcoholic and low-alcohol products, for the 4 years covering 2015–2018, on which we have previously reported (O’Donnell et al., 2019; Anderson, Jané Llopis, Rehm, 2020; Anderson et al., 2020), and for the six and a quarter months between January and early July 2020. Because only items purchased and brought into the home are recorded, only off-trade alcohol purchases are included, that is, alcohol bought in supermarkets, other shops or via Internet shopping for consumption off the premises. Alcohol purchases are recorded daily. A detailed product description identifies the type and volume of the purchase using 19 drink categories, the brand and the alcohol by volume (ABV). Volume was combined with ABV to calculate grams of alcohol purchased.

We prepared data for the interrupted time series analyses by summing up daily records of purchases of alcohol in grams across all households, adjusted for the number of adults in each household, for each day during the first six and a quarter months of each year (2015–2018 and 2020). To take into account differences in the numbers of households between 2015–2018 and 2020, we express purchases per 100 households for each of the time periods. To match days of the week across the 6 years (i.e. Monday for Monday, etc.), we re-set day 1 for each year as the first Sunday after 1st January (2015: 4th January; 2016: 3rd January; 2017: 8th January; 2018: 7th January; 2020: 5th January). Confinement was introduced on 26th March 2020, day 82 for our re-set day numbers. We restricted our analyses for the re-set days 8–182. Re-set day 8 was chosen to give a greater number of days after New Year, when alcohol purchases are minimal. Day 182 was chosen as this coincides with 4th July 2020, when pubs were re-opened in England, where over 86% of the included households were located. We defined confinement as the period from re-set day 82 to re-set day 182. For our potentially confounding covariates, we used the mean age of the main shopper per day and the mean household income (adjusted for numbers of adults in the household) per day.

Statistical analyses

We adhered to published guidance for undertaking interrupted time series analyses in the health (Bhaskaran et al., 2013; Bernal et al., 2018) and addictions (Beard et al., 2019) fields. To conduct the main interrupted time series analyses, we adopted similar methodology to that previously used in our assessment of the impact of the introduction of minimum unit price in Scotland (O’Donnell et al., 2019), based on guidance for the use of controls in interrupted time series studies of public health interventions (Bernal et al., 2018). As such, we generated a new series of dependent variables representing the differences between 2020 and 2015–2018 (averaged) for each corresponding re-set day. The dependent variables were as follows:

• (a) differences between 2020 and 2015–2018 (averaged) in means of sum of grams of alcohol purchased per 100 households (adjusted per adult in the household) per day for all alcohol and by beverage category;

• (b) differences between 2020 and 2015–2018 (averaged) in the mean ABV of purchased beer per day; and

• (c) differences between 2020 and 2015–2018 (averaged) in the proportion of the volume of all purchased beer that had an ABV of 3.5% or less.

For the two potentially confounding covariates (mean age of main shopper and mean household income), we also generated a new set of independent variables, the differences between 2020 and the average of 2015–2018.

By creating the new series of differences for the dependent variable data, we largely removed seasonal variation in the data. For each dependent variable, we examined the distribution visually and with Q-Q plots and found all variables to be normally distributed. As recommended by a number of authors (Jebb et al., 2015; Beard et al., 2019), we used a time series modeler function (IBM Corp, 2017) to estimate best fitting ARIMA models that: (a) specify degrees of differencing and transformations to ensure a stationary series; and (b) specify autoregressive and moving average orders. This eliminated the need to identify an appropriate model through trial and error (Makridakis et al., 1983; McLaughlin, 1984).

Figure 1 plots the residuals, residual autocorrelations and residual partial autocorrelations for both the original and modelled series (from the time series modeller function) for our main dependent variable, the differenced sum of grams of purchased alcohol per 100 households. The plots of the modelled series, ARIMA (0,0,2, 1,1,1)₇, demonstrated a stationary series, with no evidence of autoregression. Our series was daily, and we set the seasonal function as daily, with a periodicity of 7 (i.e., a week). We repeated the time series modeller function separately for all dependent variables, including within sub-group analyses.

We examined abrupt level changes due to the single event, the introduction of confinement in Great Britain. The event variable was entered as a dummy variable coded with 0 for each day before the event and with 1 for each day from the event forwards. Thus, in our regression models, the dependent variables were ARIMA modelled purchased grams of alcohol, mean ABV of beer and proportion of beer with an ABV of 3.5% or less. The independent variable was the dummy variable confinement, with two covariates, the differences between 2020 and 2015–2018 (averaged) of mean age of main shopper and of mean household income.

The model we are testing is illustrated in Fig. 2, where, with ARIMA terms (0,0,2) (1,1,1)₇, the equation is as follows:

(1 − Φ₁B⁷) (1 − B⁷)(Y_t) = β_intercept + (1−θ₁B−θ₂B)(1−Θ₁B⁷) α_t + β_confine (1−Φ₁B⁷)(1−B⁷) X_t + β_a (1−Φ₁B⁷)(1−B⁷) A_t + β_i (1−Φ₁B⁷)(1−B⁷)I_t,

where we have:

(Seasonal AR (1) term) (seasonal difference (1) term) (dependent variable) = intercept + (non-seasonal MA (2) term) (seasonal MA (1) term) (random error) + (impact of confinement) (Seasonal AR (1) term) (seasonal difference (1) term) (confinement) + (coefficient age) (Seasonal AR (1) term) (seasonal difference (1) term) (age) + (coefficient income) (Seasonal AR (1) term) (seasonal difference (1) term) (income).

Y_t is the dependent variable at day t; Φ₁ is seasonal autoregressive operator at lag 1; B is the backshift operator; β_intercept is the pre-event intercept, in this case, the average of the differences between 2020 and 2015–2018 in purchases for each day across all days prior to the event; θ₁ is the non-seasonal moving average operator at lag 1, and θ₂ the non-seasonal moving average operator at lag 2; Θ is the seasonal moving average operator at lag 1; α_t is the error term; β_confine is the impact of confinement, in this case, the change of mean daily purchase in grams per household for the time period since the confinement took effect versus before confinement; X_t is confinement; β_a is the coefficient of age and A_t is age; and, β_i is the coefficient of income and A_t is income.

As results would be sensitive to the exact dates chosen for the event (introduction of confinement) and the end of the period of confinement chosen, we undertook sensitivity analyses for a range of alternative dates, detailed in the footnote to Table 3.

We repeated the ARIMA models for each of the four separate groupings of households (by age of main shopper, household income, class group and area of Great Britain), with the results reported per 100 households.

All analyses were performed with SPSSv25 (IBM Corp, 2017).

Lockdown and overall purchases

Figure 3 plots the average daily purchases of alcohol in grams per 100 households by day (adjusted for the numbers of adults in each household) for the 6-month period early January to early July (days 8–182 of our re-set days), for 2015–2018 (average) minus 2020. Prior to the date of confinement, purchases of alcohol were similar between 2015–2018 and 2020 (difference = 15.6 g (95% confidence intervals (CIs): −14.0 to 45.2). After confinement, excess purchases increased over nearly 3 weeks and then remained stable, suggesting no stockpiling.

The increase in purchases for the period of confinement was 178 g per 100 households per day for all alcohol (95% CI: 163–193) (Table 2, Column C) a 41% increase (95% CI: 39–46) (Table 2, Column D) over the estimated expected level of 438 g (95% CI: 417–458), Table 2 Column B. Increases were 49% for beer, 31% for wine and 45% for spirits. However, when adjusting for the proportion of grams of alcohol normally purchased from on-licences (Giles and Robinson, 2019) (Table 2, Column E), the real change in excess purchases for all alcohol was 0.68% (95% CI: 0.65–0.71) (Table 2, column G), based on the off-licence purchases that would have had to have been made for not purchasing any on-licence alcohol, Table 2, column F. The real changes were −40% for beer, +15% for wine and +22% for spirits. In other words, although the overall grams of alcohol purchased did not change, there was a relative shift in purchases from lower to higher strength products, when accounting for no on-licence purchases.

Sensitivity analyses

Table 3 summarizes the results of the sensitivity analyses, based on a range of different start and end dates for an analysed confinement period. The excess off-licence purchases due to confinement range from 25.6% with a start date of the Prime Ministers’ advice on 16th March to 13th May, when some partial relaxation of confinement occurred to 40.6% for a start date of 26th March, when confinement was actually put in place, to 4th July. The excess purchases for the different time periods match the plots of Fig. 3.

Purchases by socio-demographic characteristics of households

Proportional excess off-licence purchases increased the younger the age of the main shopper, were considerably lower among the lowest income households, were lower among households in class groupings D and E (in class group E, the age of the main shopper was 65 years or more in over two-fifths of households) and tended to be higher in the north as opposed to the south of Great Britain, Table 4.

Shifts in purchases of beers

Figure 4 plots the mean ABV% of purchased beer and the proportion (%) of all beer purchased with an ABV ≤ 3.5% by study day. Prior to confinement, the mean ABV(%) of beer was lower during 2020 than during 2015–2018 (−0.178), although not significant (95% CI: −0.40 to 0.045), but increased following confinement by 0.127 (95% CI: 0.012–0.241, P = 0.030). Prior to confinement, the proportion of beer with an ABV of 3.5% or less was higher during 2020 than during 2015–2018 (1.06% higher, 95% CI: 0.71–1.40), but decreased following confinement (1.48% lower, 95% CI: 1.30–1.66).