Can we empirically derive a geographic definition of ‘coastal’ for use in cancer data reporting? An ecological modelling study using England’s national cancer registry

Introduction

Inequalities in health are unfair, avoidable, systematic differences in population-level health that can lead to large disparities in outcomes like life expectancy.1 Much work in recent times has sought to identify, quantify and explain these differences found across many dimensions of geography and patient demography, with the aim of identifying levers to reduce their magnitude. NHS England identified several populations at risk of such inequality using their ‘Core20PLUS5’2 approach, to help target the most at-risk groups. One such group was ‘coastal communities with pockets of deprivation hidden among relative affluence’. This population was identified as one facing unacceptable health inequalities (compared with ‘non-coastal’ areas) by reports such as the Chief Medical Officer’s Annual Report 2021—Health in Coastal Communities.3 In this report, it was noted that there was no nationally agreed definition nor consensus on what constitutes a coastal community, and definitions that include areas as large as local authorities, or even larger areas, lead to the deprived communities of interest being masked by more affluent inland areas.

Having a coastal definition that assigns small areas as populations at risk of health inequalities would allow the NHS, Office for National Statistics (ONS) and other health bodies to publish data comparing those areas to others inland, allowing deeper understanding of the distributions of health outcomes, quantification of the ‘coastal excess’ in adverse outcomes that is not accounted for by deprivation and population age, and fulfilment of the NHS’s Public Sector Equality Duty.4

This study aims to develop and evaluate a range of empirically derived coastal definitions at the small area level for cancer data, to identify a candidate variable that captures the maximum possible variance in cancer-related health between coastal and non-coastal areas that is not explained by known demographic and geographic characteristics. The candidate variable chosen will be embedded within the National Disease Registration Service’s (NDRS) databases to allow future cancer data reporting to include breakdowns by coastal/non-coastal groupings. Our goal is to develop a resource that is easy to interpret and could be deployed by a range of organisations that may not have the analytical capacity to use more complicated methods for defining coastal areas.

As long ago as 2006, it was noted that coastal populations of the UK were experiencing dramatic demographic ageing.5 Alongside an ageing population, coastal communities tend to suffer from a range of socioeconomic disadvantages that have been linked to worse psychological and physical health,6 including reliance on the seasonally variable, service-led employment sector in the wake of reduced maritime industry, a comparatively low-skilled workforce with worse job security and a limited range of employment opportunities, and a higher density of houses of multiple occupation. The stress afforded by these conditions can increase the risk of non-communicable diseases like cancer.7 These issues are not confined to UK shores, as international work on coastal inequalities has shown.8 9 This work represents a cancer-focused first step towards reporting of coastal inequality, in the hopes that future collaboration can lead to effective reduction in health inequality across a wider set of health domains.

Materials and methods

Statistical analysis

Descriptive analyses were undertaken to explore the variation of our outcomes and candidate variables. We visualised candidate variables using geographic mapping. Spearman correlation coefficient for the association of each candidate with each covariate was computed to identify potential multicollinearity issues, with an a priori cut-off of r = ±0.4 (figure 1).

Request permission

Figure 1

Heatmap of Spearman correlation coefficients for the associations between each candidate coastal variable and each covariate. Non-statistically significant (p value >0.05) correlations are not shown. Darker blue boxes indicate larger absolute coefficient values. IMD, Index of Multiple Deprivation

A set of linear regression models was created. Each cancer outcome (age-standardised incidence rate, crude prevalence and age-standardised mortality rate) was regressed against each candidate variable, additionally adjusted for mean population age, proportion of female residents, proportion of white ethnicity residents, IMD quintile and rural classification. Where a candidate variable incorporated either rural classification or IMD quintile (‘GD’, ‘GR’ and ‘GDR’ candidates), those covariates were excluded from the adjustment set to avoid over fitting.

Residual plots for each model were visually inspected for outliers and heteroskedasticity, leading to the omission of two outlier LSOAs for which data quality issues may have caused misleading figures. After excluding the Isles of Scilly and City of London (n=7 LSOAs), the final dataset used in all models thus included complete data for 32 835 LSOAs.

For each metric, all candidate variables whose Wald p value was <0.05 within their fully adjusted model were ranked according to Bayesian Information Criterion (BIC), which was scaled within each cancer metric to have a mean of 0 and a SD of 1 (BIC within each metric was scaled by subtracting the mean and dividing by the SD of the metric). The candidate variable whose model produced the lowest scaled BIC was selected as the ‘best’ model for that outcome/statistic combination. By adjusting for key socioeconomic and demographic characteristics, we aimed to move beyond simply mapping prevalence, thus strengthening our confidence that the chosen variable would be capturing more of the as-yet ill-defined health inequality dimension that underpins the residual variation in outcomes.

Results

After data cleaning, there remained 32 835 LSOAs with complete data available for analysis, covering a population of 55.3 million inhabitants. Mean population size per LSOA was 1683 persons (range 422–11 514 persons).

Spearman correlation coefficients between each candidate variable (n=40) and each outcome metric (n=3) did not exceed r = ±0.4, suggesting that multicollinearity was limited (figure 1).

All 40 candidate coastal variables were significantly associated with at least one cancer metric (incidence, prevalence or mortality) in fully adjusted models (figure 2). All geography-only, deprivation and rurality candidates (‘G’, ‘GD’ and ‘GR’ candidates) significantly improved models of two of the three cancer statistics, whereas all but one of the ‘GDR candidates’ were significantly associated with incidence, prevalence and mortality.

Request permission

The ‘best performing’ (ie, lowest scaled BIC) candidate variable for capturing coastal-related variation in cancer incidence and prevalence was the geography-only model ‘G_25_5’. Table 1 describes the distributions of model covariates across those LSOAs deemed ‘coastal’ by the ‘G_25_5’ model. This candidate variable assigns LSOAs to the ‘coastal’ identifier if 25% or more of their postcodes have centroids within 5 km of the coastline. When considering cancer mortality however, no candidate variables were a significant addition to the model. In this case, the ‘best performing’ candidate variable for capturing coastal-related variation in cancer mortality was ‘GR_10_5’, which assigns LSOAs with 10 or more postcodes within 5 km of the coast, provided they are also rural (figure 2). This candidate variable was not found to significantly improve the model of mortality rate based on likelihood ratio tests, despite the increased mortality rate in coastal areas as defined by this variable (table 2). This suggests that the confounder variables included in our models were responsible for most of the variability in cancer mortality, leaving little explanatory power for the ‘GR_10_5’ variable to account for. We therefore conclude that, when considering LSOA-level cancer mortality, additional adjustment for coastal location is not advisable.

Table 2 presents summary statistics for the regression models presenting the associations between coastal regions (defined here using ‘G_25_5’) and cancer outcome statistics. We show that coastal areas have poorer cancer outcomes compared with non-coastal areas (after adjusting for known covariates). In coastal areas, age-standardised incidence was 9.1 per 100 000 persons higher (95% CIs=5.6 to 12.6), crude prevalence was 6.4 per 100 000 people higher (95% CIs=4.4 to 8.4) and age-standardised mortality was 1.6 per 100 000 people higher (95% CIs=−0.8 to 4.0) compared with non-coastal LSOAs.

Our first sensitivity analysis sought to augment our adjustment set to account for a wider range of potential confounders that may be associated with coastal health inequality. As such we added variables describing LSOA-level average distance to hospital, and to GPs, alongside a measure of air quality and one of mental health, to our existing adjustment. These additions did not alter the choice of ‘best’ coastal variable per outcome and made no significant difference to the adjusted R² values from the original models, suggesting overfitting (data not shown). We opted, therefore, to keep our adjustment set as originally described.

As a secondary sensitivity analysis to ensure that the chosen coastal variable was not only suitable for reporting of coastal health inequalities in cancer, we used LSOA-level CHD prevalence as an additional health outcome. The ‘G_25_5’ model was also a significant addition to the fully adjusted model of CHD (table 2).

To allow for swift uptake of our coastal variable to health data analyses, we have included in the supplement the list of 2011 coastal LSOAs (defined by ‘G_25_5’, see online supplemental file 3.1). Although data were not available at the time of writing to revalidate our candidate using more recent data, we have provided the list of coastal LSOAs from the 2021 census by applying our ‘G_25_5’ candidate method to these more recent boundaries (see online supplemental file 3.2).

Discussion

The National Disease Registration Service (NDRS) is committed to fulfilling the NHS’s Public Sector Equality Duty by publishing National and Official statistics broken down by health inequality dimensions, as standard, wherever this is feasible and compliant with data privacy rules. Currently, most regular reports of cancer epidemiology in England are provided broken down by age group, gender, ethnic background, IMD quintile and geography. Despite its inclusion in NHS England’s ‘Core20PLUS5’ framework, publishing by coastal status has not been implemented in these reports yet, because there is no agreed definition of ‘coastal communities’. This study has sought to rectify this deficit by nominating a simple, parsimonious definition that can be replicated throughout cancer data reporting. Our coastal definition is at LSOA-level because this is a widely used census geography in the UK, and because geographies of lower granularity mask important intercommunity variation in health outcomes. We have chosen a candidate variable that best explained differences in cancer outcomes in the presence of socioeconomic adjustment, but as the definition of ‘coastal’ chosen excluded these factors, our candidate variable still enables adjustment or stratification by accepted measures of deprivation and rurality, thus allowing multidimensional exploration of geographical variation in cancer outcomes.

Despite a range of evidence for health inequalities for coastal communities,3 5 6 8 25 data and insight into this area have been hampered by the lack of a coherent definition of ‘coastal’ that allows for the inclusion of both rural and urban coastal communities and is sufficiently granular to avoid the ‘masking’ of pockets of deprivation by more wealthy nearby areas. A report by Atterton et al (2006)5 employed the Vickers classification of Coastal Britain, which uses a local authority-level classification derived from a principal component analysis and k-means clustering algorithm, using initially 129 different geographical and demographical variables.26 A report by Public Health England in 2020 stated that those older people living in coastal areas may be at higher risk of social isolation and loneliness.25 This report comprised a literature review of studies that referenced coastal health/inequalities. Their working definition of ‘coastal’ was ‘any coastal settlement within a local authority area whose boundaries include the UK foreshore, including local authorities whose boundaries only include estuarine foreshore. Coastal settlements include seaside towns, ports and other areas which have a clear connection to the coastal economy’. Unfortunately, ‘clear’ in this context, was not defined. This report also noted that there was a paucity of data on coastal health in England, in comparison to studies of rural health inequalities.25

Asthana and Gibson (2022) mapped several health outcomes, including cancer, by Lower and Middle Super Output areas in England. Cardiovascular disease showed a distinct core/periphery distinction, as did several other health conditions. The working definition of ‘coastal’ in this study was LSOAs which include or overlap a built-up area of any size which lies within 500 m of the ‘Mean High Water Mark’ coastline. Again, it was noted that most health data reporting is made available at local authority or higher geographies, thus masking this core/periphery distinction. At the time of writing, several research groups are working to identify other definitions of ‘coastal’ for varied uses. This includes an ESRC funded project at the University of Plymouth and the ONS’s Built-Up Areas classification. These have been produced or are aimed at seeking definitions to support policy research, whereas our focus was purely associated with cancer data reporting. We anticipate that in future, a range of novel disease-specific empirically derived variables may be produced, which may share similarities with those reported here. At that point, it may be that a more harmonised, cross-disciplinary definition can be ascertained collaboratively, but this work must start with the development of more subject-specific solutions, as we provide here. It would also be interesting to compare how our ‘simpler’ methodology compares to these more complex approaches at capturing coastal inequalities.

Using an holistic definition of cancer, we sought to identify a coastal variable that would be broadly applicable to all cancers, but we recognise that, just as IMD quintile may not be the most appropriate measure of deprivation as it relates to all diseases, having a single definition of a concept enables acceleration in research and understanding of that concept. We anticipate that our selected candidate variable will not be the statistically superior version to capture coastal health variation across all cancer types and metrics, but statistical precision must be balanced against pragmatic utility. A single, easily computed variable can be widely implemented quickly across all analyses, providing comparability, consistency, and clarity, which is preferable to a suite of variables whose use varies per disease and which cannot then be directly compared.

Although our focus with this study was to produce a method to report cancer data by coastal status, it is unclear whether this description would be applicable in other disease settings. We report a simple sensitivity analysis showing that our coastal variable captures a significant amount of residual variation in CHD diagnoses above our adjustment set as an illustration, but further exploration of other diseases was beyond the scope of this work. Understanding coastal variation in other diseases represents an interesting future direction for research. Additionally, future research should apply this work to understand the causal processes through which coastal inequalities materialise to ensure that any measure of coastal status is capturing the correct information.