Participants
This study used data from Understanding Society: The UK Household Longitudinal Study (UKHLS), which provides high-quality longitudinal panel data comprising a stratified and clustered General Population Sample of around 40,000 households. These analyses used data from wave 2 (2010-2012) and wave 5 (2013-2015) when questions on arts participation/cultural events were included. Of 54,554 respondents in wave 2, 37,389 were followed up in wave 5, and 25,051 of these (around 67%) responded to self-completion questionnaires on health and arts. After deleting 1,391 cases with missing information (around 5%), the final analytic sample includes 23,660 respondents and 47,320 person-wave observations.
Measures
Arts engagement was measured using 28 separate questions that were categorised into participation in active arts participation (“arts participation”) or attending cultural events (“cultural attendance”). Arts participation included dance (including ballet), singing to an audience or rehearsed for a performance (not karaoke), playing a musical instrument, writing music, rehearsing/performing in a play/drama, opera/operetta or musical theatre, taking part in a carnival/street arts event, learning or practising circus skills, painting, drawing, printmaking or sculpture, photography, film or video making as an artistic activity, using a computer to create original artworks or animation, taking part in textile crafts, wood crafts or any other crafts such as embroidery, knitting, reading for pleasure (not newspapers, magazines or comics), writing any stories, plays or poetry, or being a member of a book club where people meet up to discuss and share books.
Cultural attendance included attending a film at a cinema or other venue, an exhibition or collection of art, photography, sculpture or a craft exhibition, an event which included video or electronic art, an event connected with books or writing, street arts or a public art display or installation, a carnival or cultural specific festival, a circus (not animals), a play/drama, pantomime or musical, an opera/operetta, a classical music performance, a rock, pop or jazz performance, a ballet, a contemporary dance performance, or an African people’s dance or South Asian and Chinese dance. For each question, frequency of arts engagement was measured using five categories for participation in arts participation (never, once/twice per year, once per month, once per week, more than once per week) and four categories for attendance at cultural events (never, once/twice per year, once per month, once per week or more).
Given the well-known distinctions between mental health and multidimensional wellbeing [31], we explored three different outcome variables. Mental distress was measured with GHQ-12 (General Health Questionnaire); a well-validated scale derived from 12 items to measure the levels of respondents’ psychiatric illness. Items include depressive and anxiety symptoms, sleeping problems, and overall happiness [32]. UKHLS converts the answers to GHQ-12 questions to a single continuous scale ranging from 0 (the least distressed) to 12 (the most distressed), with a lower score indicating better mental health.
Mental functioning was measured using SF-12 (12-Item Short Form Health Survey), a widely used and reliable instrument that measures respondents’ general quality of life and focuses both on mental and physical health. It places a particular emphasis on the implications of any problems for ability to function as normal in everyday life [33]. The survey contains eight indicators formed of 12 items: physical functioning, role limitations due to physical health problems, bodily pain, general health , vitality, social functioning, role limitations due to emotional problems , and mental health [33]. UKHLS calculates the SF-12 Mental Component Summary (MCS) score by assigning higher weights to mental health related items (the latter six items). The MSC score ranges from 0 (the lowest mental functioning) to 100 (the highest mental functioning).
Subjective wellbeing comprises both affective aspects (such as happiness and pleasure in daily life and being free from negative affect) as well as cognitive-evaluative aspects (such as life satisfaction) [34]. We focused specifically on life satisfaction. This was measured using a single-item “overall, how satisfied are you with your life nowadays?” Responses ranged from 1 (completely unsatisfied) to 7 (completely satisfied) [35].
We used directed acyclic graphs to identify potential confounding factors that could influence both mental health and arts engagement [36]. As our statistical approach controlled automatically for any time-constant, even if unobserved (see ‘Statistics’), we restricted our identification of further confounders to those that vary over time. Identified demographic confounders included age, age squared, marital status (never married, married/cohabited, divorced/separated/widowed), presence of children in the household (no children, preschool children aged 0-4, primary school children aged 5-11, middle school children aged 12-15), employment status (inactive, unemployed, working class, intermediate class, service class), number of people in household, logged household income, and data collection wave. In order to ascertain whether individuals who engaged in the arts simply led healthier lifestyles, which contributed to their mental health (perhaps as an underlying function of socio-economic status), we additionally controlled for a wide range of health behaviours which are often associated with mental health [37,38]. These included self-reported sports activity ranking (from 0 ‘doing no sport at all’ to 10 ‘very active through sport’), smoking behaviour (current smoker, ever smoked, never smoked), drinking frequency in the last year (from 1 - never drink - to 8 - drink every day), and portions of fruits or vegetables eaten per day. We also adjusted for the extent to which health limits moderate activities to try and capture the health selectivity in participating arts activities or events. Finally, in order to identify whether individuals who engaged in the arts simply had stronger social ties and more frequent social contact (which could support their mental health (Umberson & Montez, 2010)), we additionally controlled for family support and friend support measured using a 4-point scale from 1 (not at all/no family or friends) to 4 (a lot) for each of the following 3 questions: family/friends understand the way I feel; I can rely on family/friends; I can talk about my worries with family/friends. Principal component factor analyses were conducted to extract one factor for family support (eigenvalue = 2.35, variance explained = 78%, alpha = 0.86) and one factor for friend support (eigenvalue = 2.45, variance explained = 82%, alpha = 0.89). For more details about the distribution of variables, see Table 1.
Statistics
Using Stata 14, we performed FE regression analyses, a sophisticated statistical technique commonly used in causal inference research. Compared with ordinary least square regression, which does not distinguish between within- and between-person variation, FE regression only focuses on within-person variation, examining how changes in frequencies of art engagement are linked to changes in mental health within each individual over time [39]. In doing this, FE regression eliminates the confounding effects of all time-constant variables (e.g. gender, ethnicity, social class, personality, previous arts engagement, previous mental health, education etc). As such, these factors cannot explain any association found. Further, FE regression considers time-varying confounders at both waves, not just at baseline, capturing their dynamic relationship with the exposure and outcome to better estimate the causal relationship.
We fitted nested models adding covariates stepwise. Model 1 automatically adjusted for time-constant variables. Model 2 controlled for time-varying demographic characteritics and wave. Model 3 additionally controlled for time-varying health behaviors and social support. However, as the factors in Model 3 could be seen to lie on the causal pathway (which would make adjusting for them inappropriate), Model 2 may present more appropriate estimates. We further assessed whether age and gender were moderators through including interaction terms.
Although the panel data only consist of two waves, the key variables in this study such as mental health, frequencies of art engagement have enough within variation (on average 35% of total variation is from within variation) allowing for accurate estimation of FE regression analysis (Allison, 2009).
Data were strongly balanced. A Hausman test confirmed the selection of a fixed effects over a random effects model. The modified Wald test for group-wise heteroscedasticity was significant so sandwich estimators were applied. Coefficients for all years were not jointly equal to zero, so time-fixed effects were included in the model. Longitudinal weights provided by the UKHLS were used to adjust for the complex survey design, non-response rate, unequal selection probabilities and non-random attrition across waves.