Statistical analysis

Analyses were performed in STATA v15.0 (STATA Corp, College Station, Texas, USA). Summary statistics are presented as means with standard deviations (SD) or proportions. Mean and SD were calculated for each survey period and instrument.

We used a balanced panel approach. That is, only participants with data from the first mental health survey and data at each subsequent survey were included in the analyses n = 313. Resilience questions were launched later, following the end of the circuit breaker, and were analysed in a separate cohort (also a balanced panel). As recruitment was rolling and new participants joined each month, the resilience cohort was larger (n = 583). Demographics of the analysis cohort, full cohort, non-responders, and dropouts at each survey timing are included in Supplement 1. A dropout was defined as a participant who didn’t respond at a survey wave and all subsequent survey waves.

A linear panel regression model with random effects was used to assess the temporal relationships of: i) COVID-19 and mental health, well-being and resilience and; ii) to identify if participant characteristics (age, sex: male/female, ethnicity: Chinese, Indian, Malay, Others, education level: primary level or below, O-level/N-level, Diploma/A-level, higher degree, monthly household income level: < S$1,000, S$1,000–4,999, S$5,000–8,999, S$9,000–19,999, ≥ S$20,000, employment status: employed, in school, self-employed, not employed or in school, living alone or the presence of one or more medical conditions) were associated with mental health, well-being and resilience. We hypothesised that younger age, having children, female gender, unemployment, and minority groups would be at risk of poorer mental health and well-being scores [16, 19–24]. The baseline survey measure was used as the reference group in each panel regression. We calculated the Cohen’s _d for paired samples to interpret the magnitude of any clinical effects if statistically significant differences from baseline scores were observed [41]. An effect size of 0.2 is considered small, 0.5 moderate, and 0.8 large [41]. Statistical significance was set at p < 0.05.