Quantifying the human impact of Melbourne's 111-day hard lockdown experiment on the adult population

Abstract

Lockdown was used worldwide to mitigate the spread of severe acute respiratory syndrome coronavirus 2 and was the cornerstone non-pharmaceutical intervention of zero-COVID strategies. Many previous impact evaluations of lockdowns are unreliable because lockdowns co-occurred with severe coronavirus disease related health and financial insecurities. This was not the case in Melbourne's 111-day lockdown, which left other Australian jurisdictions unaffected. Interrogating nationally representative longitudinal survey data and quasi-experimental variation, and controlling for multiple hypothesis testing, we found that lockdown had some statistically significant, albeit small, impacts on several domains of human life. Women had lower mental health (-0.10 s.d., P = 0.043, 95% confidence interval (CI) = -0.21 to -0) and working hours (-0.13 s.d., P = 0.006, 95% CI = -0.22 to -0.04) but exercised more often (0.28 s.d., P < 0.001, 95% CI = 0.18 to 0.39) and received more government transfers (0.12 s.d., P = 0.048, 95% CI = 0.001 to 0.24). Men felt less part of their community (-0.20 s.d., P < 0.001, 95% CI = -0.30 to -0.10) and reduced working hours (-0.12 s.d., P = 0.004, 95% CI = -0.20 to -0.04). Heterogeneity analyses demonstrated that families with children were driving the negative results. Mothers had lower mental health (-0.27 s.d., P = 0.014, 95% CI = -0.48 to -0.06), despite feeling safer (0.26 s.d., P = 0.008, 95% CI = 0.07 to 0.46). Fathers increased their alcohol consumption (0.35 s.d., P = 0.002, 95% CI = 0.13 to 0.57). Some outcomes worsened with lockdown length for mothers. We discuss potential explanations for why parents were adversely affected by lockdown.

Fig. 1 |. Description of the natural experiment with a timeline of COVID-19 daily infection numbers, lockdown dates and HILDA Survey data collection windows.

The figure was prepared by the authors based on data available from Our-World-in Data. HILDA Survey waves are collected annually between August and February the following year. A small number of individuals are interviewed in March. For information on the lockdown rollout, and data collection, protocol and composition, see Methods.

Fig. 2 |. Lockdown treatment effects (expressed in s.d.), separately for men and women.

Estimated treatment effects and their 95% CIs. Each estimate was obtained from separate difference-in-differences regressions of a specific outcome, estimated separately for men (blue) and women (red). Treatment effects are interpreted as outcome changes in Melbourne between 2019 and 2020 relative to outcome changes in Sydney between 2019 and 2020, holding constant location fixed effects, location-specific time trends (wave year indicators interacting with Melbourne indicators), individual-specific effects and time-varying observable confounders. Each regression applies sample weights to make the sample representative of the population (Methods, equation (1)). Human impact is measured across four domains of human life: (1) health (mental health, n = 24,357 for male and 27,854 for female), general health (n = 24,192 for male and 27,622 for female) and bodily pain (n = 24,336 for male and 27,825 for female); (2) health behaviours (BMI, n = 23,622 for male and 26,566 for female), frequency of alcohol consumption (n = 21,731 for male and 23,322 for female) and frequency of physical activity (n = 24,355 for male and 27,848 for female); (3) social connectedness (feeling safe, n = 27,939 for male and 31,246 for female), feeling lonely (n = 24,237 for male and 27,699 for female) and feeling part of the local community (n = 27,907 for male and 31,187 for female); (4) labour supply and income (weekly working hours (n = 27,964 for male and 31,272 for female), weekly income from all sources (n = 27,964 for male and 31,272 for female) and weekly government transfers excluding family benefits (n = 27,964 for male and 31,272 for female) (see Supplementary Table 1 for definitions). Each outcome is standardized to mean = 0 and s.d. = 1. Standard errors are clustered at the household level to account for the fact that all household members (aged 15 or older) were interviewed and that full households were mobility-restricted in the same location. Supplementary Tables 4 and 5 present the full model results.

Fig. 3 |. Lockdown treatment effects (expressed in s.d.), separately for men and women within socially relevant subgroups.

Estimates (and 95% CIs) obtained from separate difference-in-differences regressions of a specific outcome on the treatment group indicator. a, Mental health. b, General health. c, Bodily pain. d, BMI. e, Frequency of alcohol consumption. f, Frequency of physical activity. g, Feeling safe. h, Feeling lonely. i, Feeling part of the local community. j, Working hours. k, Weekly income. l, Government income support. Treatment effect is interpreted as s.d. differences from the mean outcome (standardized to 0). See Fig. 2 for details of the model specification. See Supplementary Tables 1 and 3 for variable definitions in each domain (1–4) and the five subgroup definitions (with young child, lone individual, had low mental health in 2019, bottom income below the median and lives in an apartment). Each panel presents the results for a different outcome. Sample sizes for each estimate are reported in Supplementary Table 15. Supplementary Tables 17–21 present the full model results.

Fig. 4 |. Lockdown treatment effect (expressed in s.d.) for women for selected outcomes according to length of exposure and family status.

Estimated treatment effects of lockdown for different exposure lengths to lockdown for women. a, Mental health. b, General health. c, Bodily pain. d, BMI. e, Frequency of alcohol consumption. f, Frequency of physical activity. g, Feeling safe. h, Feeling lonely. i, Feeling part of the local community. j, Working hours. k, Weekly income. l, Government transfers. Spikes represent the 95% CIs. Estimates are expressed in s.d. (from 0 mean) obtained from separate difference-in-differences regressions of a specific outcome on a treatment group indicator. We used the same specification as the benchmark model (Fig. 2), but the treatment indicator (wave 20 × Melbourne) was interacted with four dummy variables, each of which presents a different exposure length to lockdown: fewer than 40 d; 40–70 d; 71–111 d; and the days after lockdown was lifted. Note that vertical axes vary across the panels. Each panel presents results for different outcomes. The sample sizes for each estimation are reported in Supplementary Table 16. Supplementary Table 22 presents the full model results.

Fig. 5 |. Estimated treatment effect of lockdown (expressed in s.d.) for men and for selected outcomes according to the length of exposure and family status.

Estimated treatment effects of lockdown for different exposure lengths to lockdown for men. a, Mental health. b, General health. c, Bodily pain. d, BMI. e, Frequency of alcohol consumption. f, Frequency of physical activity. g, Feeling safe. h, Feeling lonely. i, Feeling part of the local community. j, Working hours. k, Weekly income. l, Government transfers. Spikes represent the 95% CIs. Estimates are expressed in s.d. (from 0 mean), obtained from separate difference-in-differences regressions of a specific outcome on a treatment group indicator. We used the same specification as the benchmark model (Fig. 2), but the treatment indicator (wave 20 × Melbourne) was interacted with four dummy variables, each of which presents a different exposure length to lockdown: fewer than 40 d; 40–70 d; 71–111 d; and the days after lockdown was lifted. Note that the vertical axes vary across panels. Each panel presents results for different outcomes. The sample sizes for each estimation are reported in Supplementary Table 16. Supplementary Table 23 presents the full model results.