The effects of the London 2012 Olympics and related urban regeneration on physical and mental health: the ORiEL mixed-methods evaluation of a natural experiment

Adolescent survey

Participants were recruited through secondary schools in two ways: (1) school-based enrolment of adolescents aged 11–12 years in year 7, and (2) recruitment of parents/carers through the surveyed adolescents. Schools were incentivised by a single donation of £1000 paid after completion of the baseline survey.

Adolescent recruitment began by asking the borough-level administrators (local education authorities/learning trusts) to encourage schools to participate. This approach has been previously successful in recruiting primary schools, albeit outside London. 24 However, all four boroughs suggested that we contact schools directly. A letter of invitation was sent to school principals (and members of their senior leadership team), followed by a telephone call. This resulted in the recruitment of 10 schools. The remaining schools were recruited from an e-mail campaign targeting heads of year and subject leaders who might find the ORiEL project of academic interest to their students (e.g. physical education, geography, sociology). Overall, 42 of the 48 eligible schools were approached to recruit the final 25 schools needed. The most common reason for schools refusing was ‘research fatigue’. This suggests that personal preferences of organising staff were a cause of refusal rather than pupil characteristics. Further details can be found in Smith et al. 14

Respondents were recruited from six schools in each of the London boroughs of Newham, Hackney, and Barking and Dagenham and from seven schools in Tower Hamlets (Table 2). The cross-sectional baseline survey respondents comprised 3095 adolescents in year 7 of secondary school (aged 11–12 years) who completed a paper-based questionnaire during the 6 months (January to July 2012) prior to the start of the London 2012 Olympic and Paralympic Games. To maximise the sample, the whole school year was surveyed in seven schools that had relatively small year groups. The remaining 18 larger schools provided an allocation of mixed-ability adolescents selected on the basis of school timetabling. Adolescents were followed up at approximately 6 months (January to July 2013) and 18 months (January to July 2014) post intervention. Schools were surveyed as close to the same month of each year, when possible, in order to minimise seasonality effects. The cross-sectional sample was larger at follow-up than at baseline owing to a deliberate oversampling strategy. This was a logistical requirement given that the cohort members became dispersed over time to many different classes across the school timetables at first follow-up (year 8).

The final sample featured single- and mixed-sex schools and drew on the largest and smallest schools in the four boroughs, which were affiliated to a range of religious denominations.

Table 3 compares the sociodemographic profile of the ORiEL sample at baseline with census data. The sociodemographic characteristics of the ORiEL baseline sample were broadly similar to the equivalent population observed by the 2011 census, with some exceptions. The ORiEL sample was slightly under-represented in terms of female respondents and Bangladeshi and white UK respondents; this ethnic difference contrasted with an ORiEL over-sample of white other and mixed white ethnic groups. The high proportion of white other groups included recent migrants from European Union states and will have contributed significantly to the higher than expected numbers of participants born overseas. Overall, the baseline response rate was 87% and the study sample (n = 3095 in school year 7) can be estimated at approximately 25% of the entire age group attending state schools in the catchment areas (n = 12,136 in school year 6).

Parent/carer survey

The recruitment of parents/carers was carried out under contract with an external market research agency using face-to-face interviewer-administered questionnaires. Owing to data protection legislation, schools were unable to supply the home addresses of parents/carers of children enrolled in the study. We therefore asked adolescents to volunteer their home address during the completion of their questionnaires and this information was then passed on to the market research organisation for recruitment. Parents/carers received a letter of invitation to participate along with an incentive of an automatic entry into a prize draw in which five people would win £100. The overall response rate was 60%, a high response rate for east London, which has been historically difficult to enumerate. 25 However, the achieved sample size of the baseline cross-section was a considerably smaller than the adolescent sample because invitations to participate could be sent only to homes with a valid address provided by the adolescent during the survey session (Table 4). We were also obliged under the terms of the Market Research Society guidelines to ask all adult participants whether or not they would be willing to participate in the follow-up interview. Consequently, 37% of respondents opted out of the potential follow-up cohort at baseline. Because of the loss to follow-up, the study protocol was amended for follow-up data collection. The modified parent/carer survey adopted a repeat cross-sectional design, although a nested cohort was retained at each wave of data collection. This approach was agreed with the National Institute for Health Research (NIHR) as an appropriate compromise.

Adolescent questionnaire

A paper-based questionnaire, based on validated tools and instruments listed below, was administered to assess individual and household sociodemographic characteristics, mental health and well-being, and physical activity of participating adolescents. Core questionnaire items are outlined below.

Parent/carer questionnaire

The content of the parent/carer questionnaire was similar to that of the adolescent questionnaire. The three primary outcomes (employment, mental health, and physical activity and behaviour) were identical but used instruments adapted for face-to-face adult interviews.

Fieldwork

Study protocols were drafted that detailed standard procedures in the preparation of fieldwork materials and duties and regulations during the in-school data collection. These quality management systems were implemented to ensure that all fieldworkers were trained to the same level and shared the same knowledge of the questionnaire. This helped to minimise the potential response bias introduced if fieldworkers provided differing levels of assistance or information to adolescents completing their questionnaire.

A pilot study was conducted on a subsample within a participating school to determine the appropriate length of the adolescent questionnaire, identify language or comprehension difficulties with the use of standard scales, and refine elements of the survey protocol focused on school and parental/carer consent, and adolescent assent. The questionnaire was designed so that the primary research outcomes were completed early in the schedule to ensure higher response rates, with secondary outcomes completed towards the end of the questionnaire, when there was a greater risk of non-response. For the main study at baseline, school-based data collection commenced in January 2012 and finished 3 weeks before the opening ceremony of the London 2012 Olympic and Paralympic Games, which were held on 27 July 2012.

Consent

One week prior to survey, the school provided each adolescent with an age-appropriate study information sheet and a study information sheet to take home to their parent/carer. The letter presented the opportunity for parents/carers to actively opt the adolescent out of the study at any time. Parental consent was therefore passively obtained if the opt-out form was not returned by the adolescent.

During the survey visit the questionnaire was explained orally prior to completion; all adolescents additionally provided active written assent prior to completing the survey, and all adolescents were reminded that they were free to withdraw at any time without consequence. Immediately following survey completion all students were provided with a copy of their assent form and a duplicate of the age-appropriate information sheet. They were invited to contact the ORiEL study team if they had further questions.

Written consent from the school Principal or authorised member of the school’s senior leadership team was obtained before fieldwork began in their school.

Response rates

Survey completion was defined as answering the final battery of questions. Any adolescent respondent who had a record of ‘non-compliance’ – either because the student was a random box ticker or it was obvious that the questionnaire was not being taken seriously – was removed from the database. This amounted to 18 individuals at wave 1, 15 individuals at wave 2 and 48 individuals at wave 3. Figures 2–4 show the percentage of respondents completing each item on the questionnaire at each wave, in which the x-axis represents the proportion of the questionnaire designated to each topic.

FIGURE 2.

The ORiEL baseline adolescent questionnaire: item completion by topic.

FIGURE 3.

The first ORiEL follow-up (wave 2) adolescent questionnaire: item completion by topic.

FIGURE 4.

The second ORiEL follow-up (wave 3) adolescent questionnaire: item completion by topic.

Full questionnaire completion at baseline was 50%, rising to 60% at first follow-up, increasing further still to 80% at second follow-up. Reasons for non-completion include unexpectedly short or interrupted questionnaire sessions; random sampling of streamed lower ability groups; higher levels of special educational needs; and lower than anticipated levels of literacy or English-language skills.

In order to invite parents/carers to participate in the study, we asked each adolescent participant for their home address. Overall, 88.8% of participants provided a valid home address at baseline. Addresses at first and second follow-up were derived by reconciling previous addresses with new details contained in the more recent follow-up questionnaires. This yielded addresses for 88.1% of participants at the first follow-up and for 86.7% at the second follow-up.

Parents/carers were also sent a letter of invitation describing the study’s aims and requesting participation in a doorstep interview. Those who did not wish to participate were asked to e-mail or ring the telephone number provided. If parents/carers did not opt out within 2 weeks of receiving the invitation, they were included in the doorstep interview schedule. Up to seven attempts to gain an interview were made before the participant was classified as a non-responder. Parents/carers were given the opportunity to refuse participation when first contacted by the market research fieldworkers and at any point during or after completion of the interview. Interviews were anonymised and parents/carers were linked to the relevant child.

Parent/carer interviews started in April 2012 and were completed by 27 July 2012. A market research agency administered the 35-minute face-to-face computer-assisted personal interview to consenting parents/carers and interviews were carried out at the parent/carer’s home address. To overcome any potential language problems, the parent/carer survey was translated into two of the most common non-English languages within the local community, namely Urdu and Bengali. Interviewers with particular language skills were allocated to specific participants, when required. All interviews were completed and had no missing data. The response rates were 60% at baseline, 51% at wave 2 and 57% at wave 3.

Data processing

Survey data entry was performed by an external agency with extensive experience in generating data files for longitudinal cohort studies. Variable names and coding structures were devised by the ORiEL research team and were implemented by the data entry contractor. Questionnaire data were double-punched and cleaned using range, consistency and logic checks. In a limited number of cases, the data were manually cleaned by ORiEL research staff when it was unclear to the third party what the correct coding should be.

Participants were allocated a unique identifier to allow us to track cohort members across waves while continuing to anonymise questionnaires. Adolescent and parent/carer names and addresses were stored separately from each other on encrypted USB drives. These were accessible by a single data custodian and were linked only temporarily by a unique identification number to produce lists of participants who were eligible for follow-up.

Proximity-based measures

Proximity to the nearest environmental resource is estimated according to the shortest path distance in metres on the road network. The road network is given by the OS survey MasterMap Integrated Transport Network. The resolution of the distance, in all but a small number of cases, is address-point to address-point. For an aggregate of environmental resource types, the minimum distance for those types represents the shortest distance to a member of that aggregate class. Metric creation used the Esri ArcGIS version 10.3 network (Esri, Redlands, CA, USA) analyst extension.

Proximity measures were created for food establishments, premises selling alcohol, green spaces (access points), and sports and recreation facilities.

Density-based measures

An adaptive kernel density estimation (aKDE) approach is used to compute density surfaces covering the study area for which comprehensive point location data have been collected. The aKDE is carried out in the R package (The R Foundation for Statistical Computing, Vienna, Austria) sparr (http://cran.r-project.org/web/packages/sparr/sparr.pdf); map outputs have a cell size of 25 m, and intensity values indicate the estimated relative density of a given environmental resource type per km² for the area covered by a given cell.

We used aKDE, as opposed to the standard fixed-bandwidth kernel density estimation, owing to the highly clustered nature of observations within the various environmental resource data sets. A fixed bandwidth approach would probably produce unsatisfactory results, over-smoothing areas with high resource numbers and under-smoothing areas with low numbers, resulting in a density surface that inadequately captures variation in density. An adaptive approach better preserves the locally varying densities across the study area.

The aKDE method computes a density based upon a Gaussian (normal) kernel, and employs adaptive smoothing on the basis of a ‘pilot bandwidth’. Fitting a locally adaptive kernel is an optimisation problem; the use of a pilot bandwidth simply acts to help limit the size of the bandwidth to within realistic bounds. The pilot bandwidth is first calculated in sparr using leave-one-out least-squares cross-validation (LSCV). In practice, this method can be unreliable for computing fixed bandwidths and has a tendency towards conservative estimates that over-smooth spatial data. However, as pilot bandwidth inputs for aKDE, they are effective at narrowing the search space so that a density surface can be computed in a reasonable time.

Density-based measures better represent the local availability of environmental resources than do proximity-based measures, although the two are closely related. Density-based measures incorporate the combined effect of all nearby resources, whereas proximity-based measures generally restrict investigation to the nearest resource, or the nearest k resources.

A density value representing environmental resources of different types per km² was calculated for each respondent, and for each environmental resource type. Density values were summed within environmental themes (e.g. food environment, alcohol environment) in order to provide a density for a chosen aggregate of resource types within that environment theme.

Metric creation used the ESRI’s ArcGIS version 10.3 spatial analyst extension, and R version 3.0.2 with sparr 0.3–4, and spatstat 1.36–0. 67

Density-based measures were created for food establishments, premises selling alcohol, and sports and recreation facilities. Density-based measures could not be created for green-space access as it is not possible to compute a kernel density value for polygon data.

Egocentric measures

An egocentric residential neighbourhood is created for each study respondent based on an 800-m road network buffer, using the same reference data as the proximity-based measure. This distance is widely used in epidemiological literature as representing the 10-minute walking distance of an average person, and is therefore said to represent the likely extent of the immediate residential neighbourhood. 68

Using respondent home address, we computed a set of bespoke neighbourhood buffers for each ORiEL respondent. We counted the occurrences of each environmental resource that fell within each buffer to measure how many resources an individual can reach for a given egocentric neighbourhood. When green space is concerned, we count the number of access points for distinct green spaces that are contained within each egocentric neighbourhood (e.g. parks), avoiding the repeated count of the same green space if an individual can access more than one entrance point.

Egocentric neighbourhoods also underlie the creation of the walkability index. All the components of walkability are captured subject to the 800-m network buffer. We used the egocentric neighbourhood as a ‘cookie cutter’ in order to compute the denominators for each walkability component as follows:

• Residential density is computed as the count of domestic addresses as a ratio of the area of residential building footprints within each egocentric 800-m neighbourhood.

• Intersection density is computed as the count of junctions connected to three or more edges (loosely, streets) as a ratio of the kilometres of road within each egocentric 800-m neighbourhood.

• Land use mix is computed using the normalised Shannon entropy formula69 for each egocentric neighbourhood, which effectively measures the evenness with which the three categories are distributed within the 800-m buffer. Residential, office and commercial land uses are computed as proportions of combined land use based on the building footprint area of buildings belonging to each category. Assigning building footprints to office, commercial or residential usage is done using the National Land Use Database classification implemented in the OS MasterMap AddressLayer 2 data.

Metric creation used the ESRI’s ArcGIS version 10.3 network, including the network analyst extension.

Egocentric measures were created for food establishments, premises selling alcohol, green spaces (access points to green spaces), sports and recreation facilities, and walkability.

Further details of analyses can be found in the relevant results chapters.

Intervention and comparison groups

Participants residing within Newham at baseline and subsequent waves of the study were considered to have received a greater ‘dose’ of urban regeneration. Participants residing in Barking and Dagenham, Hackney and Tower Hamlets at baseline and subsequent waves of the study were considered not to have received urban regeneration (or not such a strong dose of regeneration) and, therefore, formed the comparison group.

At baseline, participants were asked to give consent to be followed up even if they moved during the lifetime of the study. Pupils who moved to attend a different school already participating in the ORiEL study were followed up at their new school when possible. When pupils moved outside the study schools, these were no longer followed up as the administration cost involved was considered too high given the relatively small numbers involved. Very few participants moved schools across the waves, so an intent-to-treat (ITT) approach was taken for the analyses, based on the following approach. All participants are analysed as if they had remained living in the same borough that they lived in at baseline for subsequent waves.

This therefore involved analysis of the change in well-being, mental health and physical activity before the Olympics, 6 months after the Olympics and 18 months after the Olympics for Newham, compared with Tower Hamlets, Barking and Dagenham, and Hackney, using all those who took part at baseline. Some individuals may have moved boroughs during the project, swapping from control to intervention school, from intervention to control school, from one control school to another, or from one intervention school to another. More simply, they may have dropped out or moved out of the area completely. The ITT approach does not take into account any movement between schools or drop out after baseline. ITT is the typical approach taken in randomised controlled trials, based on the argument that it preserves randomisation. Alongside ITT, other sensitivity analyses were conducted and compared.

Attrition and missing data

Participant non-response was present in individual questionnaire items and by wave, introducing attrition. Attrition was explored through missing data patterns and logistic regression, investigating predictors of missingness by wave. Missingness by wave was not found to be associated with any predictors and supported the assumption that attrition followed a missing completely at random (MCAR) mechanism. With a MCAR mechanism we were able to create a non-biased cohort of adolescents present at all three waves.

For each analysis, the missing data mechanism was explored for individual questionnaire items. We found no violations against the assumption that the missing data mechanism is missing at random (MAR). Data were imputed using multiple imputation to gain statistical power, and sensitivity analysis was explored to check if inferences were robust to the MAR assumption. The imputation methodology used in each analysis is described in more detail in each of the results chapters.

Weighting

Design weights have been derived to address over- or under-sampling of specific cases or for disproportionate stratification and sample clustering. These weights are used when we want the sample under investigation to be representative of the population. Design weights are used in analyses in which clustering is considered a nuisance to be controlled for. When clustering or area effects are a point of interest, then a hierarchical multilevel model has been used to estimate these effects. Non-response weights for the data sets are not necessary; rather, we have adjusted our models for a range of covariates known to predict non-response such as age, gender and ethnicity.

Family sample

Participants were recruited via the parent/carer quantitative survey. We wrote to all survey participants who had indicated that they were willing to be contacted again with a view to participating in further research. These letters were followed up with a telephone call inviting them to participate in the qualitative study. In total, 130 people were contacted in this way, of whom 20 made themselves available for interview at wave 1. We asked that these core participants invite other members of their household to participate in interviews as they saw fit. In all, an additional 19 young people and one spouse also took part. At wave 2 this fell to 15 core participants and an additional 13 of their family members.

Adolescent sample

In total, 26 adolescent core participants (12 boys and 14 girls) were recruited from three participating schools at wave 1 (two in Newham and a pilot in a neighbouring borough). School contacts, who served as gatekeepers for qualitative recruitment, were each asked to select up to nine students from their year 8 and 9 cohorts to participate in a half-day video focus group workshop. At wave 2 the pilot school dropped out and a total of 12 adolescents participated in the focus group workshops.

Family narrative interviews

A family narrative interview was conducted with each of the families. The typical format for this was an interview with a parent/carer (core participant recruited from the adult survey) and their children. However, adult participants often invited other family members to participate, and sometimes the other family members decided themselves to join in. Two of the adult participants wanted to be interviewed alone without any other family members present. At wave 1, participants were asked to provide a narrative account of their experience of the Games and whether or not they felt that it affected them personally. The interview then moved on to focus on what had changed in their local area and daily life, and how they perceived their neighbourhood. At wave 2, we asked the families to revisit their original narratives and update us on events between waves. These interviews provided insight into how the Games were experienced as a spectacle in their own right and how participants positioned the Games and regeneration in relation to their own lives, trajectories and local areas.

Go-along interviews

A go-along interview is a mixture of observation and interview concentrated around a particular site, journey or activity. 70 The researcher can accompany the participant(s) on a routine journey or activity to a specific place or request that the participant give them a ‘tour’ of part of their familial environment. In this way the interviews can provide direct experience of the natural habitats of informants, and access to their practices and perceptions as they unfold in real time and space. 71 The accounts and narratives made by all participants in the preceding family interviews (described above) were extended in the go-along interviews with the aim of understanding how participants experienced these sites and how they related them to their own lives and practices.

School video focus group workshops

Half-day workshops were organised with three participating schools. Focus groups were used because these interviews generate rich information within a social context in which interaction between participants can reveal cultural values and group norms that may not arise in individual interviews. 72 The first half of the sessions was a focus group interview on participants’ perceptions and experiences of the Games and of their local neighbourhoods, followed by small group work in which students were split into groups and given the task of interviewing each other about aspects of neighbourhood experience arising from the focus group discussion. Safety and crime was a popular choice of topic for participants. The sessions were video recorded so that they could be viewed by these participants at the next wave of data collection and serve as a prompt for reflection and discussion. At wave 2 we showed the participants clips from wave 1 and asked them to update us and reflect on their contributions to the previous wave.

Outcome measures

As described in Chapter 2, validated instruments were deployed to assess a range of health outcomes. The main items of interest for this analysis were mental well-being, depression, physical activity and sedentary behaviour, self-rated health and long-term illness or disability. These were operationalised as described in the following sections.

Individual demographic, social and environmental factors

The distribution of health outcomes described above was explored across a range of individual demographic and household socioeconomic indicators as well as by individual perceptions of the local environment. These were are described in the following sections.

Mental health and well-being

The complete case-analysis observed that females self-reported significantly lower mental well-being than their male counterparts, but there were no differences according to ethnic group or generation (Table 6). Well-being was lowest in adolescents in Newham and was significantly higher in those attending schools in Tower Hamlets and Barking and Dagenham. Overall, there was a mixed relationship between well-being and socioeconomic disadvantage; there was a gradient effect, with well-being increasing significantly with increasing family affluence, but no differences were apparent according to free school meal status. After full adjustment, adolescents with a lone parent not in paid employment had significantly higher levels of well-being than adolescents with two parents/carers not in paid employment. However, for all environmental factors there was a statistically significant gradient effect, with those perceiving the neighbourhood more positively being more likely to report higher mental well-being scores. These differences remained after full adjustment. Such patterns were broadly similar for the prevalence of depressive symptoms (Table 7). In unadjusted and adjusted models, females were more likely to be at risk of depressive symptoms, with no variation by socioeconomic background. However, after full adjustment only Indian adolescents were significantly less likely to report depressive symptoms than the white UK group. As observed for mental well-being, there was a significant association between negative perceptions of neighbourhood safety and aesthetics and a greater risk of depressive symptoms. This observation was also significant after full adjustment.

Physical and sedentary activity

In unadjusted complete-case models, girls spent significantly fewer hours (12.6 hours) than boys (14.0 hours) participating in physical activity (Table 8). This was consistent with girls spending a significantly greater number of hours per week in sedentary activity (35.6 hours) than boys (33.2 hours) (Table 9). Indian adolescents reported significantly higher participation in physical activity than the white UK comparison group, but there were no other ethnic differences in either physical or sedentary behaviour. There was mixed evidence of a socioeconomic influence on activity. There was a significant increase in the hours spent physically active with increasing family affluence, but no differences in physical and sedentary activity were observed for other socioeconomic factors. In terms of neighbourhood effects on activity, adolescents who described their neighbourhoods as more amenable to walking and cycling were significantly more likely to be physically active and less likely to be sedentary than those who described the neighbourhood as more difficult to walk or cycle in. All differences remained significant in fully adjusted models. Variation in the number of hours spent sedentary was significant at school level.

Self-rated general health

The proportion of respondents reporting fair/very poor health by a range of factors is shown in Table 10. Over one-quarter of girls (28.6%) reported fair/very poor health, but the prevalence was significantly lower for boys (20.6%). Bangladeshi and black African adolescents were significantly more likely to report poor health than white UK adolescents after full adjustment. There was weak evidence of a socioeconomic gradient in self-rated health. Although adolescents with both parents/carers in employment were significantly less likely to report poor health, as were those who did not have free school meals, these differences were no longer significant in fully adjusted models. However, there was a strong and consistent association between positive perceptions of the neighbourhood and better reported general health across all three neighbourhood domains. These associations were also observed in fully adjusted models.

Long-term illness

Table 11 shows the prevalence estimates and factors associated with having a long-term illness. There were no gender differences in long-term illness. Prevalence varied widely among ethnic groups. Participants from the black Caribbean and white mixed groups were significantly more likely to report a long-term illness than their white UK counterparts in unadjusted models. After adjustment, participants from the ‘other’ ethnic group were also significantly more likely to report a long-term illness. Adolescents born overseas were significantly less likely to report having a long-term illness, but this finding was not significant following adjustment. There were no associations between any of the socioeconomic indicators and long-term illness. However, there was a graded increase in the odds of having a long-term illness as perceptions of neighbourhood safety and aesthetics worsened, although this finding was no longer significant for the case of aesthetics after adjusting for covariates. No association was observed for walk/cycleability or proximity to services in the local area.

The co-occurrence of selected health outcomes after full adjustment for all covariates is described in Table 12. Poor general health, having a long-term illness, lower levels of mental well-being and having depressive symptoms were all strongly associated with one another. However, there were no significant associations between the mean hours spent physically active or sedentary and all other health outcomes.

Study design and participants

This longitudinal quasi-experimental study followed a cohort of adolescents recruited from schools in the London boroughs of Barking and Dagenham, Hackney, Newham and Tower Hamlets over a 3-year period (see Chapter 2). The participants, in year 7 at baseline (age 11–12 years, January–June 2012), were first followed up in year 8 (wave 2, age 12–13 years, January–June 2013), 6–11 months after the London 2012 Games and again in year 9 (wave 3, age 13–14 years, January–June 2014), 18–23 months after the London 2012 Games. Timing of follow-up for each school was matched by month to reduce seasonality effects. The baseline survey sample comprised 3106 respondents. After excluding respondents who moved schools between waves (n = 8), the final longitudinal cohort for this analysis comprised 2254 adolescents who participated in all three waves, representing an overall retention rate of 73%.

Intervention and control areas

The London Borough of Newham, where the London 2012 Olympic Games were mainly hosted and where the majority of focused regeneration occurred, acted as the intervention area. In the context of this analysis, Olympic-led regeneration in Newham consisted primarily of the construction of infrastructure and facilities supporting the Olympic Park prior to the Games, and subsequent legacy redevelopment of the Queen Elizabeth Olympic Park, which reopened to the public in phases from July 2013 www.queenelizabetholympicpark.co.uk (accessed May 2018). Regeneration components plausibly affecting the outcomes reported here, and outlined in the Legacy Masterplan,13 include the development of ‘sustainable’ transport networks (rail and walking/cycling corridors) to promote active travel; the provision of new and refurbished recreation and leisure facilities to promote formal physical activity; and new and refurbished civic space, parks and green areas to promote leisure-time and routine physical activity.

Data and variables

Physical activity and screen time was assessed using the Y-PAQ. 39 The Y-PAQ is a validated self-reported tool that captures the frequency and duration of a range of physical and sedentary activities over the past 7 days. In addition, it records frequency and duration of screen-based activities, such as TV viewing, internet use and gaming. Respondents who reported > 75 hours of total activity per week were excluded from analysis. Total hours spent physically active or undertaking screen-based activities were calculated for each respondent. Time spent physically active per week was highly positively skewed, with the majority of adolescents reporting few hours of activity, whereas screen time per day was negatively skewed. Time spent physically active was therefore categorised to either ‘active’ or ‘inactive’ based on whether or not the adolescents met the current UK Department of Health and Social Sciences recommendation of 60 minutes of physical activity per day. 81 Screen time was used as a marker of sedentariness and was dichotomised into ‘sedentary’ or ‘not sedentary’ at the internationally recognised cut-off of no more than 2 hours per day. 82

Engagement with the London 2012 Olympic and Paralympic Games

Engagement with the London 2012 Games was assessed using a question from the national evaluation of the London 2012 education programme (Get Set) (ref 5). Respondents were asked ‘How excited do you feel about the 2012 Olympic Games being held in London?’, with responses ranked on a 1 to 10 scale, where 10 was the most excited. Data were categorised into low (1–4), moderate (5–7) and high (8–10) levels of engagement.

Use of the Olympic Park

Use of the Olympic Park was assessed using the question ‘How often do you visit the Olympic Park?’. Responses were every day, once a week, 2–3 times per month, once a month, less than once a month, and never visited. Data were categorised into never visited, less than once a month, and more than once a month in order to adequately power statistical analyses.

Covariates

A range of factors hypothesised to act as confounders between exposure to the games and its legacy were identified: age (in months), gender, ethnicity (white UK, white mixed, Asian Indian, Asian Pakistani, Asian Bangladeshi, black Caribbean, black African, other), means-tested free school meals (yes/no), home language (English/other), household composition (both parents/carers, single parent/carer or neither parent/carer at home), self-rated health (five-point categorical scale collapsed to very good/good versus fair/poor/very poor), long-term limiting illness (yes/no; defined as a health problem that has troubled them over a period of time, or is likely to affect them over a period of time), BMI (z-scored), days between surveys and days from opening ceremony.

Statistical analysis

Survey weights were applied to all analyses by the use of the .svy commands in Stata 14 SE (StataCorp LP, College Station, TX, USA). Weighting was deployed to account for differing probabilities of being selected in each school. Descriptive analysis of baseline data explores the extent to which physical and sedentary activity and screen time, as well as pre engagement with the Olympic Games, covaries across the range of selected social, economic and demographic characteristics. Data were imputed under a MAR assumption83 with a joint multivariate normal modelling approach through the Bayesian estimation method Markov chain Monte Carlo (MCMC). A burn-in of 20,000 iterations was used to stabilise the Markov chains and a further 25,000 iterations were run, creating 50 imputed data sets at every 500th iteration. We tested whether or not there were longitudinal changes in the proportion of individuals reporting > 60 minutes of physical activity per day and < 2 hours of screen time per day before and after the Games by (1) intervention versus control areas (at 6 and 18 months), (2) levels of pre-engagement (at 6 months) and (3) use of the Olympic Park (at 18 months). When appropriate, we present gender-stratified results.

Logistic regression models derived the fully adjusted odds of reporting > 60 minutes per day of physical activity post Games in the lesser engaged groups compared with the highly engaged group (reference category), after accounting for activity pre Games. This was repeated for screen time.

Legacy effects at 18 months

Table 20 shows changes in meeting physical activity recommendations between baseline (wave 1) and the 18-month follow-up (wave 3). In the fully adjusted models there were no statistically significant differences between intervention and comparison boroughs for all physical activity and screen time transitions.

In order to assess the effect of the infrastructural legacy of the Olympic Park we also explored whether or not visiting the Olympic Park at 18 months was associated with transitions in physical activity and screen time. Table 23 shows that there was a graded effect of visiting the Olympic Park (less than once a month to more than once a month) for those who remained inactive and those who became active. Those who visited the Olympic Park more than once a month were the least likely to remain inactive (RR 0.11, 95% CI 0.02 to 0.48) and the least likely to become inactive (RR 0.38, 95% CI 0.24 to 0.60) of those who were active at both baseline and at the 18-month follow-up. However, study respondents were also less likely to become active if they visited the Olympic Park more than once a month (RR 0.52, 95% CI 0.30 to 0.90) than the always active group. In gender-stratified models, males who visited the park more than once a month were less likely to remain inactive (RR 0.11, 95% CI 0.02 to 0.48) and less likely to become inactive (RR 0.42, 95% CI 0.23 to 0.76). For females, associations were observed for all three physical activity transitions, with respondents less likely to remain inactive (RR 0.31, 95% CI 0.13 to 0.74), less likely to become inactive (RR 0.36, 95% CI 0.18 to 0.73) and less likely to become active (RR 0.24, 95% CI 0.08 to 0.74) than the always active group.

For sedentary behaviour, Table 24 shows that females who visited the Olympic Park less than once per month were less likely to become less sedentary (RR 0.39, 95% CI 0.21 to 0.73). No other significant associations for sedentary behaviour were observed.

Measuring change in physical activity environments

As detailed in Chapter 2, we computed a range of environmental metrics measured at each wave from which to create measures of individual environmental change. Here we focus on access to sporting facilities and green space.

Access to sporting facilities

Redevelopment of the Olympic Park introduced several new opportunities to access sporting facilities and, therefore, we wanted to capture this change in provision. Density measures are particularly well suited to do this; however, the actual density value is difficult to interpret as differences are small, and the meaning of any change is hard to grasp. As a result, rather than using the absolute change in density we assessed relative change as a percentage difference from the measure at wave 1 (baseline) to wave 3. For each individual, for each sports facility category, we calculated the change in density at the point at which new Olympic sports facilities became available as an increase or decrease from the density of provision available to that individual at baseline. In this analysis we focus on all public sports facilities, which represent the wider sporting legacy of the games. Although we cannot fully distinguish Olympic-related provision from other changes across the study site, it is evident that the new venues within the Olympic Park were responsible for the majority of change in the density of public sports facilities. Figure 6 classifies the ORiEL study area based on the percentage change in density between wave 1 (baseline) and wave 3, demonstrating that the largest positive changes in access are centred around the Olympic Park area, extending westward. Much of Barking and Dagenham experienced some amount of negative change.

FIGURE 6.

Change in density of sports facilities in the ORiEL study boroughs.

Access to green space

Characterising individual change in access to green space was challenging, given that most change in green space is concentrated within the Olympic Park itself (see Figure 5) and not distributed across east London. In addition, it became clear that green space is relatively abundant within east London. This means that, for respondents, the new green-space provision catalysed by the development of the Olympic Park is unlikely to be the nearest green space for any participant. This suggests that simple ‘nearest neighbour’ accessibility metrics that are the standard approach in the field would not adequately capture environmental change. This suggests that we need to treat the Olympic Park as an addition to existing green infrastructure rather than exclusively as a new, closer destination for local people. As with the sports facilities metric, a density measure would be well suited here; however, we cannot calculate one for polygonal data, which are what the available green-space data are based upon. Therefore, for each individual we computed a measure of the average (mean) distance to their nearest three ‘metropolitan parks’ at wave 1 (baseline) and wave 3. We used changes in that average to quantify exposure to the new Olympic Park green space. This change is then expressed as the relative percentage change (as we used for the sports facilities metric). The ‘metropolitan parks’ categorisation used as the basis for this metric is an official park category (used in the London Plan)84 that denotes large size, high quality and high amenity.

Figure 7 demonstrates the percentage change in access to green space interpolated over the ORiEL study area and represented using four change categories. A similar kind of spatial pattern as in Figure 6 is evident, with few negative changes. The negative changes that do exist appear to be the result of changes to the road network structure, leading to some individuals travelling slightly further at wave 3 to get to the same metropolitan parks and hence forming a small negative change category.

FIGURE 7.

Change in access to green space in the ORiEL study boroughs.

Environmental change: use of a four-category change variable

The raw distributions of percentage environmental change for sports facilities and green space did not fit any known distribution, and the data could not be transformed for normality. Distributions are strongly multimodal, with the largest peaks clustered around zero and smaller peaks for a range of positive change values. This distribution is likely to be as a result of the spatially clustered nature of the sample around schools. In order to meaningfully use these data, we took a pragmatic approach. First, we acknowledged that change values were meaningful: values clustered around zero indicated no or little change between the two time points; negative change values for densities (of sport facilities) and positive change values for distances (to green space) indicated a worsening environment for a given individual; and positive change values for densities (or sport facilities) or negative change values for distances (to green space) indicated an improving environment. On this basis we first created four categories to accurately represent any meaningful change that occurred. These are presented in Tables 27 and 28. Negative change categories were generally small, and tended towards small changes; however, the positive change categories tended to include larger numbers of people and a larger spread of data values. The positive change group was multimodal so we split those who experienced positive change in two and created similarly sized groups.

Primary outcomes

Sociodemographic and neighbourhood covariates

The following covariates, available at each wave, were identified a priori from existing literature: age in cohort (months); gender; ethnicity (assessed using an 11-category variable based on the 2011 UK census); years in UK; number of parents/carers the participant lived with; parental income; bullied at school in the past 12 months; receiving free school meals; long-term illness; number of life events; whether or not moved neighbourhood since last wave; and the number of days since the opening ceremony of the Olympics that the participant completed the questionnaire. Social support from family and friends were assessed using the MSPSS,36 with scores divided into tertiles representing low-, medium- and high-level support. See Table 34 for further details of these measures.

Dealing with missing data and multiple imputation strategy

Weights were derived to take account of unequal probabilities of school and pupil selection. Prevalence and missing data rates for the outcomes and covariates were examined: missing values ranged from 0.0% to 45.2%. We found no evidence against the assumption that data were MAR. 83 We imputed the data using multilevel multiple imputation in the REALCOM software (Realcom Applications, Coleorton, Leicestershire, UK), which uses a joint multivariate normal modelling approach through the MCMC method. We imputed with two levels (first = wave and second = adolescent) with all of the outcomes and covariates as fixed effects. Interaction terms between gender and the intervention, and between free school meals and the intervention, were also included. The imputation model was chosen to be congenial86 with the most saturated model of interest; auxiliary variables were included to strengthen the MAR assumption. We used a ‘burn-in’ period of 25,000 iterations, followed by 50,000 iterations producing a data set every 1000th iteration, resulting in 50 imputed data sets. The MCMC chains were examined to check for convergence.

Statistical analyses

Analyses were carried out using Stata version 14 (StataCorp LP, College Station, TX, USA). Standard descriptive statistics are reported. For categorical outcomes, the prevalence and 95% CI were calculated; for continuous outcomes, the mean/median, standard deviation and range were calculated. Crude and adjusted multilevel linear and logistic regression models were run to assess the impact of urban regeneration (intervention vs. comparison) on short-term (wave 1 to wave 2) and longer term (wave 1 to wave 3) change in adolescent well-being and depressive symptoms. The models were adjusted for baseline demographic factors; household factors; family and friend social support; and psychological factors (see Table 34 for variable details). An additional adjustment for baseline WEMWBS was made to the adjusted model for well-being to test the sensitivity of the findings for baseline well-being. Interactions between the intervention and gender, and between the intervention and free schools meals, were tested: models were stratified when the interactions had a p-value of ≤ 0.05. Robustness of inference to the MAR assumption applied within the imputation of missing observations was explored through sensitivity analysis. Two methods were used: the first replaced missing observations with unrealistic extreme settings; if inferences were not robust, the second method was used, which applied more realistic settings through tipping point sensitivity analysis. The sensitivity analysis suggested that inferences were robust to realistic departures from the MAR assumption.

Descriptive analysis of adolescent mental health and well-being (waves 1, 2 and 3)

The following sections present data from 2254 adolescents who participated at all three waves (73% of the wave 1 sample). Table 34 shows the prevalence for the outcomes and covariates at each wave. In total, 27.9% of the participants were from schools in the intervention borough of Newham and 72.1% were from schools in the comparison boroughs (Barking and Dagenham, Hackney, and Tower Hamlets). Participants were ethnically diverse: the largest groups described themselves as white UK (16.9%), Asian Bangladeshi (15.0%) and white other (14.5%).

Over one-fifth of the sample reported depressive symptoms at each wave (wave 1, 21.7%; wave 2, 20.8%; wave 3, 24.2%). The mean well-being score was 51 at each wave. Baseline depressive symptoms were higher in the intervention group than in the comparison group (27% vs. 20%: unadjusted OR 1.49; p = 0.024) and baseline well-being scores were lower in the intervention group than in the comparison group [50.7 standard error (SE) (0.6) vs. 53.0 SE (0.2); p = 0.001]. These findings indicate that adolescents attending schools in the London Borough of Newham had higher rates of depressive symptoms and lower levels of well-being at the start of the study than adolescents attending schools in the boroughs of Barking and Dagenham, Hackney and Tower Hamlets.

Table 35 shows the change in SMFQ scores between wave 1 and wave 2. In total, 66.69% of the sample remained not depressed, with 11.36% becoming depressed, 10.24% remaining depressed and 11.71% being no longer depressed. Table 35 also shows the change in SMFQ scores between wave 1 and wave 3. In total, 64.50% of the sample remained not depressed, with 13.56% becoming depressed, 11.12% remaining depressed and 10.82% being no longer depressed.

Table 36 shows the change in well-being scores between wave 1 and wave 2, and between wave 1 and wave 3. In the total sample, changes in well-being scores were small in magnitude, with a positive change of 0.43 (95% CI –0.31 to 1.16) between wave 1 and wave 2, and a positive change of 0.09 (95% CI –0.74 to 0.92) between wave 1 and wave 3. Change scores for well-being were negative for females and positive for males at both time-points. Slightly larger positive change scores were observed for the intervention site (Newham) than the comparison group at wave 1 to wave 2. Positive change scores were observed for the intervention site (Newham) from wave 1 to wave 3, while a small negative change was observed for the comparison group. There was little difference in change scores in well-being by free school meal status at wave 1 and wave 2; participants receiving free school meals showed a small negative change in levels of well-being between wave 1 and wave 3, whereas participants not receiving free school meals showed a small positive change in levels of well-being between waves 1 and 3.

Change in depressive symptoms from wave 1 to wave 2

Table 37 shows the changes in SMFQ outcome between wave 1 and wave 2. In the fully adjusted models, adolescents who were ‘no longer depressed’ (RRR 1.53, 95% CI 1.07 to 2.20) or who ‘remained depressed’ (RRR 1.78, 95% CI 1.12 to 2.83) were more likely to be from the intervention borough than from the comparison boroughs, compared with those who were not depressed. ‘Remaining depressed’ was also associated with female gender (RRR 2.94, 95% CI 1.92 to 4.51), low family support (RRR 2.36, 95% CI 1.58 to 3.52), being bullied (RRR 2.53, 95% CI 1.69 to 3.78), long-term illness (RRR 1.98, 95% CI 1.26 to 3.09) and experiencing one, two, or three or more life events (RRR 1.88, 95% CI 1.22 to 2.91; RRR 2.24, 95% CI 1.56 to 3.21; RRR 2.91, 95% CI 1.64 to 3.09) and was more likely for Asian Bangladeshi adolescents (RRR 1.76, 95% CI 1.04 to 2.95). Becoming ‘no longer depressed’ was associated with low family support (RRR 1.70, 95% CI 1.21 to 2.37), being bullied (RRR 1.84, 95% CI 1.34 to 2.53), and experiencing two or three or more life events (RRR 2.03, 95% CI 1.37 to 2.99; RRR 2.12, 95% CI 1.26 to 3.58), and was less likely for Asian Indian adolescents (RRR 0.17, 95% CI 0.03 to 0.99). ‘Becoming depressed’ was more likely for females (RRR 2.34, 95% CI 1.71 to 3.21) and those experiencing three or more life events (RRR 1.69, 95% CI 1.05 to 2.71). No interactions between the intervention and gender or entitlement to free school meals were observed for depressive symptoms (p > 0.05).

Change in depressive symptoms between wave 1 and wave 3

Table 38 shows the changes in SMFQ outcome between wave 1 and wave 3. In the fully adjusted models, compared with adolescents who were not depressed, adolescents who ‘remained depressed’ (RRR 1.93, 95% CI 1.01 to 3.70) were more likely to be from the intervention borough than from the comparison boroughs. ‘Remaining depressed’ was also associated with female gender (RRR 3.98, 95% CI 2.70 to 5.86), low family support (RRR 2.11, 95% CI 1.34 to 3.32), being bullied (RRR 2.55, 95% CI 1.79 to 3.64), experiencing one, two, or three or more life events (RRR 1.78, 95% CI 1.17 to 2.70; RRR 2.42, 95% CI 1.53 to 3.83; RRR 2.82, 95% CI 1.61 to 4.94) and long-term illness (RRR 1.62, 95% CI 1.07 to 2.44), but was less likely for Asian Indian adolescents (RRR 0.39, 95% CI 0.18 to 0.82). The intervention was not associated with ‘becoming depressed’ or becoming ‘no longer depressed’. Becoming ‘no longer depressed’ was associated with living in the UK for less than 6 years (RRR 2.05, 95% CI 1.08 to 3.89), low family support (RRR 1.73, 95% CI 1.23 to 2.42), being bullied (RRR 1.70, 95% CI 1.22 to 2.37), long-term illness (RRR 1.46, 95% CI 1.01 to 2.11) and experiencing two or three or more life events (RRR 1.89, 95% CI 1.26 to 2.83; RRR 2.30, 95% CI 1.28 to 4.14). ’Becoming depressed’ was more likely for females (RRR 2.89, 95% CI 2.06 to 4.06), living in the UK for less than 6 years (RRR 1.75, 95% CI 1.17 to 2.62) and having one or three or more life events (RRR 1.50, 95% CI 1.05 to 2.14; RRR 1.79, 95% CI 1.16 to 2.77), but was less likely for Asian Indian adolescents (RRR 0.42, 95% CI 0.18 to 0.96). No interactions between the intervention and gender or between the intervention and entitlement to free school meals were observed for depressive symptoms (p > 0.05).

Change in psychological well-being between wave 1 and wave 2

There was only a small change in well-being scores between the waves (wave 1 to wave 2, µ = –0.02, SE 0.03). Table 39 shows that there intervention had no effect on change in well-being between wave 1 and wave 2. A negative change in well-being between wave 1 and wave 2 was predicted by baseline well-being (β = –0.72, 95% CI –0.78 to –0.66), being female (β = –2.04, 95% CI –3.12 to –0.97), having low or medium family support (β = –1.68, 95% CI –2.79 to –0.58; β = –1.42, 95% CI –2.64 to –0.21) and being bullied (β = –1.15, 95% CI –1.99 to –0.31). A positive change in well-being between wave 1 and wave 2 was predicted by being of Asian Indian (β = 2.61, 95% CI 0.14 to 5.07) or ‘other’ ethnicity (β = 1.31, 95% CI 0.11 to 2.50). In terms of well-being, no interactions were observed between the intervention and gender or between the intervention and entitlement to free school meals (p > 0.05).

Change in levels of well-being between wave 1 and wave 3

There was only a small change in well-being scores between the waves (wave 1 and wave 3, µ = –0.04, SE 0.04). Table 39 shows that there the intervention had no effect on change in well-being between wave 1 and wave 3. A negative change in well-being between wave 1 and wave 3 was predicted by baseline well-being (β = –0.71, 95% CI –0.77 to –0.64), being female (β = –3.65, 95% CI –4.72 to –2.58), having low family support (β = –1.33, 95% CI –2.22 to –0.45) and being bullied (β = –1.15, 95% CI –2.03 to –0.27). A positive change in well-being between wave 1 and wave 3 was predicted by being of Asian Indian (β = 4.23, 95% CI 2.60 to 5.87), black African (β = 2.29, 95% CI 0.71 to 3.88) or black other ethnicity (β = 1.98, 95% CI 0.41 to 3.54). In terms of well-being, no interactions were observed between the intervention and gender or between the intervention and free school meals (p > 0.05).

Primary outcomes

Sociodemographic, neighbourhood and discrimination covariates

A range of sociodemographic, neighbourhood and discrimination factors that might influence mental health were identified a priori from the literature. We investigated whether or not these factors were associated with parental/carer well-being, anxiety symptoms and depressive symptoms.

The following demographic factors were included: age (years); gender; ethnicity (assessed using an 11-category variable based on the 2011 UK Census); highest educational qualification; current employment status; highest NS-SEC in the household; marital status; home ownership; receipt of any benefit(s); born in the UK; and long-term illness.

The following accommodation and neighbourhood factors were included: access to a garden; satisfaction with accommodation; frequency of visits to green spaces; self-reports of damp, vibrations, dust and mould in the house; and the following scales from the MESA measure – neighbourhood aesthetics, neighbourhood safety, neighbourhood social cohesion and neighbourhood violence. 62

The following discrimination measures were included:64 insulted because of ethnicity in the past 12 months; ever refused a job as a result of discrimination; and ever treated unfairly at work because of ethnicity. See Table 40 for further details of these measures.

Missingness and multiple imputation models

Prevalence and missing data rates for the outcomes and covariates were examined: missing values ranged from 0.0% to 36.5%. We found no evidence against the assumption that data were MAR. 83 We imputed the data separately for each wave using multiple imputation by chained equations using the Stata ice package. We used 50 cycles of the imputation algorithm producing 25 data sets for wave 1, 35 for wave 2 and 40 for wave 3. The number of data sets produced for each wave was chosen to be greater than the variable with the largest percentage of missing data. The imputation model was chosen to be congenial86 with the most saturated model of interest.

Analysis

Cross-sectional univariate and multivariate logistic regression models were run to assess the impact of regeneration on rates of parental/carer mental health. The models were run in several stages to assess the effect of the intervention on parental mental health:

1. Model 1 – univariate regression analyses examined the initial associations between the intervention, sociodemographic, neighbourhood and discrimination factors with each outcome.

2. Model 2 – multivariate regression analyses simultaneously adjusted for all factors that were significantly associated (p ≤ 0.05) with the outcome in model 1, with the exception of the multiple poverty score.

3. Model 3 – model 2, with additional adjustment for the multiple poverty score.

Interactions between intervention and gender were examined in model 2. If the interaction was significant (p ≤ 0.05), models 1 and 2 were re-run stratified by gender.

Three sets of cross-sectional models were run examining each outcome separately at wave 1, wave 2 and wave 3.

Descriptive analysis of parent mental health and well-being (waves 1, 2 and 3)

In total, 1233 parents/carers completed the survey at wave 1, 1023 at wave 2 and 995 at wave 3. Table 40 shows the prevalence of the outcomes and covariates at each wave. At each wave, approximately one-third of the sample was from the intervention borough of Newham. In total, 74% of the sample at each wave was female. The sample was drawn from a range of ethnic groups at each wave. At waves 1 and 2, 42% of the participants were in employment, compared with 49% at wave 3. Over half the sample at each wave reported receiving benefits.