Design and setting
A cross-sectional design was used, incorporating an online survey, appropriate both in terms of lockdown restrictions during the COVID-19 pandemic as well as familiarity of university students with online surveys.
Protective behaviour outcome or dependent variables were the domains of hand hygiene, and social distancing behaviours. Predictor (independent) variables for the factors or determinants comprised domains of sociodemographic factors (e.g. gender, age, ethnicity), knowledge (of disease or effectiveness of behaviours), and socio-cognitive factors, both motivational/pre-motivational (e.g. risk perception, attitude, social support, self-efficacy), and predisposing (e.g. habit/automaticity, time factors, trust in authorities’ policies). The study was informed by the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines .
The survey was conducted online with paid participants recruited by Prolific Academic Ltd , a crowdsourcing platform. Prolific identified 4,250 potential participants who were currently residing in the United Kingdom at the end of the initial “lockdown” period of the COVID-19 pandemic, were currently undertaking a university course at undergraduate, graduate or postgraduate level and were at least 18 years old. This was similar to the study’s eligibility criteria of current U.K. university students, aged 18 and over. From the sample size calculations for multiple linear regression, a minimum sample size of 206 would detect a medium effect size of 0.1, with 80% power and α error probability of 5%.The targeted sample size was 300, to account for responders who were not eligible but passed the pre-screening/profiling criteria, or missed responses/questions or where there were incorrect assumptions for the sample size calculation. Potential pre-screened participants enrolled in the study on a first-come, first-served basis.
The questionnaire was adapted from existing surveys, primarily a WHO longitudinal survey on monitoring behavioural insights for use during the COVID-19 pandemic [8, 33]. A subset of the WHO questions was used to allow focus on support for the predictor and outcome variables relevant to this study and limit the questionnaire to an acceptable length. Adaption was also required to ensure consistency with the UK government communications of the time, and information likely to be known from the media or more formal sources such as the WHO.
Questions were selected to address the I-Change pre-motivational factors such as knowledge and risk perception, as well as motivational factors such as attitudes (advantages, disadvantages), social support and self-efficacy as appropriate for the different behaviours of hand hygiene and social distancing . Additional predisposing factors were identified from the literature. Thus, hand hygiene questions included time factors and habit/automaticity, but not social support , while social distancing questions covered all 4 motivational factors from I-Change (advantages, disadvantages, social support, self-efficacy), but also perception of trust in government interventions . Risk perception items and knowledge of the effectiveness of protective behaviours also reflect the work of Rubin  for designing a questionnaire on perceptions and behaviour during an influenza pandemic. The questionnaire consisted of 98 items from 28 questions, primarily multiple choice and Likert scales and can be found as Additional File 1. It took 10–15 minutes to complete. A data dictionary can be requested from the corresponding author.
Reliability of the composite scales in the questionnaire was determined using Cronbach’s alpha coefficient. Pallant  notes that although a Cronbach alpha coefficient should be above 0.7, with short scales of less than 10 items, low Cronbach’s values such as 0.5 are commonly found. Therefore, a threshold of 0.5 was used for scales. Due to the heightened time scales, formal content validity was not performed. However, as well as the questions being informed by previous surveys, a ‘Think aloud’ review by departmental colleagues was performed to increase confidence in validity. In addition, a pilot of 10 participants matching the inclusion criteria was carried out before the survey was launched.
Demographics included age, gender (male, female, other), student status (UK, International), health/life-science related course (Yes, No), and ethnicity (white, black, asian, chinese, mixed, other), all of which could influence protective behaviours, as reviewed by Bish and Michie  and assessed by Ergin  and Seale et al . In addition, known exposure to infection (self or immediate social environment) was assessed .
The outcome or ‘dependent’ variables comprised the hand hygiene behaviour scale (α = 0.75) and the social distancing behaviour scale (α = 0.74). Scores for these were derived using three-point Likert scales of 8 and 10 questions respectively, based on a combination of current government guidance at the time , the WHO longitudinal survey  and, for handwashing, published surveys [18, 37, 38]. Options were “Always” (3 points), “Mostly” (1 point), or “Rarely/never” (0 points). The mean scores were used, adjusted in the case of hand hygiene for participants not required to answer the question on handwashing after touching pets, where that was not relevant. Due to the criticality of this score, the mean was also calculated for the 3 cases with individual item missing data in the social distancing behaviour, and thus in these cases the missing data could be considered imputed. Additional understanding of hand hygiene behaviour such as whether soap and water was used, if hands were washed for the minimum recommended 20 seconds, reasons for not washing hands, and whether handwashing had increased since the start of the pandemic, were also gathered.
Knowledge was divided into disease knowledge or that specific to the effectiveness of carrying out specific hand hygiene or social distancing behaviours. Disease knowledge comprised 4 questions, assessed by appropriate responses to identification of symptoms (“Related, “Not related”, “Don’t know”), at risk groups (“Are at risk”, “Are not at risk”, “Don’t know”), treatment availability, and incubation period (both multiple choice). Knowledge of the effectiveness of the behaviours was assessed by 8 questions each as “Yes”, “No”, “Don’t know”. Correctness of answers was determined based on independently verifiable knowledge at the time of the survey [3, 4, 33], with correct answers awarded one point. Knowledge scales were derived by adding correct answers. Total scores for each scale were determined, but as these scales consisted of multiple dimensions and could be considered an index, Cronbach’s α calculation was not performed .
Risk perception was assessed based on the perceived probability, susceptibility and severity components used in the WHO survey , and as defined by Brewer , but using a 9-point scale. The mean risk perception score was derived (α = 0.66).
Other socio-cognitive constructs
The socio-cognitive constructs were derived from 3 or 4 statements for each construct, using 5-point Likert scales (1 – “Strongly disagree” to 5 – “strongly agree”) as in Additional File 2 Tables S4 and S5. The individual’s mean score was derived for each construct. Individual missing items resulted in very few (≤ 5) cases per construct with missing data, and therefore minimal impact on sample size and ability to detect an effect. Such cases were omitted from the analysis, rather than imputed, which could otherwise potentially introduce bias.
For hand hygiene behaviour the following socio-cognitive constructs were developed:
Attitude (Hand hygiene): In order to assess attitude towards carrying out the behaviour, perceived advantages such as whether the activities prevented infection in self or others, and disadvantages such as the effort involved or potential to hurt hands, were assessed. These were measured through 4 advantages items and 3 disadvantages statements, with the latter reversed coded to support an overall attitudes (advantages minus disadvantages ) scale with Cronbach α = 0.58, although mean inter-item correlation was 0.19. Greater reliability may have been possible through use of the advantages (α = 0.61, mean inter-item correlation 0.31) and disadvantages (α = 0.55, mean inter-item correlation 0.33) subscales.
Self-efficacy (Hand hygiene): Confidence in ability to carry out hand hygiene behaviour was assessed using 3 items (α = 0.62), while the predisposing factor constructs of habit/automaticity and time factors were assessed with 4 and 3 items respectively (α = 0.64, 0.55).
Habit: Habit or the related automaticity (α = 0.64) was assessed by statements such as feeling strange if they do not wash hands after using the toilet, washing hands before eating being performed automatically, washing hands or having a tissue ready without realising it.
Time: Assessment of time factors (α = 0.55) was by use of concepts such as washing hands after the toilet even when busy, believing that hand washing before food preparation takes too much time (reverse coded), and seeing washing hands with soap and water as quick and easy.
All social distancing socio-cognitive factors were assessed as 3 item constructs.
Attitude (social distancing): Attitude (α = 0.46) was represented by an advantages scale (α = 0.85) covering perception of protection for self, others and the NHS, minus disadvantages scale (α = 0.48), which included perceptions of missing family/friends, job concerns or being bored (reverse coded). As the attitude scale did not demonstrate acceptable reliability, the subscales were used in association analysis. Additionally, as the disadvantages subscale also was below the internal acceptability threshold, a sensitivity analysis was performed using an item which was thought to best summarise the construct , in this case whether the respondent thought they would lose their job or that of someone close to them.
Social norms/support: Social support for carrying out social distancing included statements about family and friends avoiding crowds or social contacts, and a reverse coded item on being encouraged to meet against guidelines (α = 0.67).
Trust: The ‘trust’ construct was about perceptions of policies such as agreeing with restricting liberty rights, whether the decisions were fair, and whether they should be relaxed even while many new COVID-19 cases were appearing (α = 0.39). Due to the low alpha, the scale was re-assessed and two key single items used in association analysis, 1) fairness of decisions and 2) relaxation of restrictions (reverse coded).
Self-efficacy (Social distancing): Similar to hand hygiene behaviour, self-efficacy for social distancing behaviour involved confidence in ability to carry out the behaviour (α = 0.75).
Self-efficacy (Infection avoidance): Applicable to both behaviours, self-efficacy at the level of general infection avoidance was assessed from a 9-point scale from “Extremely difficult” to “Extremely easy”
Behaviour: Perception of behaviour compliance was assessed by two ‘Yes/No’ questions to determine participants’ perception of whether they complied with guidance for the behaviours.
Data quality checks were carried out such as time to perform the survey, and data scrutinised for consistent single response on Likert scale across all categories, with the result that all 293 participants were included. Statistical analysis was performed using IBM Statistical Package for the Social Sciences (SPSS) v 26.
The first objective of the study, to investigate the socio-cognitive perceptions towards hand hygiene and social distancing, was addressed through descriptive statistics, primarily frequencies and percentages for behaviour, knowledge, risk perception, and other socio-cognitive perceptions. All scales were found to show other than a normal distribution by the Kolmogorov-Smirnov test, and therefore medians were used, although mean and standard deviation (SD) were also reported.
The second objective of the study, to investigate which determinants (such as knowledge and socio-cognitive perceptions) explain hand hygiene and social distancing, was addressed through inferential statistics, notably associations such as in bivariate correlations and multiple linear regression. Due to the distribution, non-parametric tests such as Spearman’s rank correlation co-efficient or Chi-squared analysis were used for testing associations between each behaviour (scale variable), and the potential determinants (scale, dichotomous or nominal variables). Where minimum cell count assumptions for Chi-squared were not met, Likelihood Ratio was used.
Hierarchical multiple linear regression was performed to develop an explanatory model to fit the data and assess the main predictors or determinants influencing the behaviour scales . The model was performed as a hierarchical regression in 3 stages; an initial block of socio-demographic variables, as co-variates, some of which have been identified as determinants in previous studies [6, 44], a second block of potential predisposing (habit, time factors, trust), or pre-motivational (knowledge, risk perception) predictors identified previously in literature [18, 24, 45], followed by other potential motivational determinants (attitude, social support, self-efficacy) in a third stage. All potential determinants variables, including those not significant in bivariate correlation analysis, were therefore assessed in the model . Where Cronbach alpha internal reliability for scales was < 0.5, individual subscales, or items were entered into the analysis [28, 46] as a sensitivity analysis, as well as assessed in the correlation analysis. Final parsimonious linear regression models of the determinants with the highest standardised β coefficients were also produced. Separate analysis by gender was also performed.
For the hand hygiene linear regression, assumptions for regression models including absence of collinearity were essentially met. However, for the social distancing regression model, a second dataset was used for linear regression modelling and correlations, due to the presence of 12 extreme outliers (1st quartile – 3*interquartile range (IQR), or 1st quartile – 1.5* IQR for > 4 constructs (or 3rd quartile + 1.5*IQR equivalent))  which otherwise would have led to contravention of regression assumptions. Sample size was thus reduced to a maximum of 281 for social distancing inferential statistical analyses. To ensure robust regression modelling in the presence of scales which may have diverged from normal distribution, bootstrapping of 1,000 samples for calculation of bias-corrected and accelerated (BCa) 95% confidence intervals (CI) and significance, was used.
Correlation matrices for all general and hand hygiene related variables, as well all general and social distancing related variables, were also created to determine if there were any strong relationships between individual predictor variables which may have then impacted the multiple linear regression findings. Post hoc correlation and linear regression analysis was performed using individual components of risk perception (probability, susceptibility, severity), to further investigate associations between these aspects of risk perception, and hand hygiene and social distancing behaviours. Where key determinants from the literature did not show significant correlation or appear significant in the linear regression models in this study, post hoc analysis was undertaken by use of dichotomised variables (such as for ethnicity – white/minorities), or item by item correlation (knowledge of effectiveness of specific behaviour versus the matching behaviour) to see whether relationships could then be identified.
Statistical significance was set at p < 0.05, 2-tailed, and 95% confidence intervals were reported.