Health behavior change among UK adults during the pandemic: findings from the COVID-19 cancer attitudes and behaviors study | BMC Public Health

Study design

The study used prospect observational data derived from a broader study of cancer symptom help-seeking and screening behavior in the UK adult population during COVID-19. The study protocol and analysis plans were pre‐registered on the Open Science Framework [26]. Findings are reported in accordance with the STROBE guidelines for surveys and observational studies. [27, 28].

Participants and data

Participants were English-speaking adults aged 18 years and over, resident in the UK and recruited using Cancer Research UK’s online panel provider (Dynata), HealthWise Wales (a national register of ‘research ready’ participants) [29] and social media platforms (Facebook and Twitter), then followed up over time. Representation of people who smoke and those from ethnic minority and lower socioeconomic groups was increased by using targeted advertising and placing quotas on sample size relative to the UK population statistics. Two online surveys were conducted in parallel: CABs (participants recruited via HealthWise Wales and social media) and the COVID-19 Cancer Awareness Measure (COVID-CAM; participants recruited via Dynata). Survey data were pooled where appropriate [25, 26]. Prospective data were collected in two phases from the same participants during August–September 2020 (phase 1) and February–March 2021 (phase 2). Sampling and recruitment methods are reported in detail elsewhere [25].


Measures of behavior change attempts were derived from the Cancer Awareness Measure 2019 [30]. Participants were asked “Are you currently trying to do any of the following?” for each of the following behaviours: “Reduce the amount you smoke”; “Stop smoking completely”; “Reduce the amount of alcohol you drink”; “Increase the amount of fruit and vegetables you eat”; “Increase the amount of physical activity you do”, and “Lose weight”. Response options were “Yes”, “No”, “Maybe”, “Prefer not to say”, and “This is not applicable to me”.

Smoking status was ascertained by asking participants “Which of the following best describes you?” with response options “I have never smoked”, “I used to smoke but have given up”, “I smoke but not every day”, “I smoke every day”, “Other” and “Prefer not to say”. Demographic variables included age, sex, ethnicity, employment and educational qualification [25].


Outcome variables were binary measures of smoking status and attempts to change five behaviors: smoking (ie attempts to reduce smoking and attempts to stop smoking), alcohol consumption, fruit and vegetable intake, physical activity and weight loss. The predictor variable was a time variable with a binary measure to indicate the first (August–September 2020) and second (February–March 2021) survey phases.

Handling missing data

The rate of missing values ​​in the outcome variables ranges from 0 to 8% (see Supplementary Table 1). We examined whether missingness in the health behavior variables was patterned by participants’ characteristics (gender, age, ethnicity, country, occupation, and educational qualification). Although missing values ​​in the outcome variables were mostly not patterned by participants’ demography, we found differences in the odds of missing data in some health behavior variables according to educational qualification (those with other qualifications or no qualification were more likely to have missing value for fruit and vegetable intake, physical activity and alcohol use than those with a degree) and occupation (those unemployed and retired were more likely to have missing data for alcohol use and weight loss compared to those employed).

In order to minimise bias, missing values ​​were handled using multiple imputation, which ensures that all observed values ​​in a dataset with some systematic differences between the missing and observed values ​​are retained [31, 32]. Outcome and demographic variables informed the multiple imputations model. Results were averaged across ten imputed datasets, and the sample characteristics of the imputed sample did not differ from those of the fully observed person-years.

Statistical analysis

Analyses were conducted using Stata (version 17). Descriptive statistics were used to analyze the proportion of participants (overall and by demographic variables) engaging in each health behavior attempt at each time point. Univariable and multivariable logistic regression analyzes were carried out to assess differences in attempts to change each behavior (smoking, alcohol, fruit/vegetable intake, physical activity and weight loss) between the two phases. Results for all models are presented as odds ratios with 95% confidence intervals. Participants’ age, sex, ethnicity, UK country, employment status and educational qualification were included as hypothesised confounders. Complete case analyzes were conducted, with sensitivity analysis using the imputed dataset to account for potential attrition bias.

Subgroup analyzes were pre-specified and conducted by introducing interaction terms between the predictor (time) and demographic variables (sex, ethnicity, employment and education).

Leave a Comment