Study area
The study area conducted in the Pays de Brest region: an area located in western France, is subdivide into 79 municipalities and 174 census blocks. Census blocks are small, county statistical subdivisions that contain around 2000 inhabitants. They are designed to be relatively homogeneous with respect to population characteristics, socioeconomic status and living conditions (19).
Data sources
Brest stroke registry
Brest Stroke Registry is an ongoing prospective community-based stroke register covering a total population of 366 000 inhabitants in western France. Since 2008, it operates using multiple information sources for case identification: from public and private hospitals, radiology clinics performing brain imaging, neurologists and general practitioners. A validation comity ascertain cases and the Brest Stroke Registry performance was evaluated based on capture recapture analysis suggested level of completeness in excess of 90% (20). The characteristics of the study population, investigations and methods of assessment have been described in detail elsewhere (20). It was approved by the local ethics committee. Patients or their legal representatives gave their written informed consent for participation. Specific authorizations were obtained from the national “Comité consultatif sur le traitement de l’information en matière de recherche” and from the “Commission nationale informatique et liberté” for this study.
Stroke cases
All incident ischemic and hemorrhagic stroke cases aged 60 years or more living in Pays de Brest between 1 January 2008 and 31 December 2013 were extracted from the Brest stroke registry. Relevant ICD codes used were I63-I64 for stroke: 1) new focal neurological deficit with symptoms and signs coherent according to the criteria of the World Health Organizations, lasting for more than 24 hours or who died in the first 24 hours; 2) all neurological focal deficits lasting at least 1 hour or resolving within 1 hour but with abnormal brain imaging associated with a clinically relevant picture. Cases were geocoded to their respective census block based on the residential street address of patients.
Patient-level data were extracted from the Brest Stroke registry: the sociodemographic data (age and sex); clinical data including stroke type (ischemic or hemorrhagic), stroke severity (National Institute of Health Stroke Score (NIHSS) <6, 6-13, >13); the presence of cardiovascular risk factors before stroke (the presence of high blood pressure, cardiac arrhythmia, diabetes and dyslipidemia).
Socioeconomic and Urban-rural contextual effects
Firstly, the level of deprivation for each census block was estimated using the FDep index (French deprivation index) (21). This index have been defined and analyzed in previous studies in order to analyze environmental and health inequalities (14). It was generated using Principal component analysis (PCA); based on four variables from the national census of 2013 collected by INSEE (National Institute for Statistics and Economic Studies): average household income, percentage of high school graduates in the population aged ≤ 15 years, percentage of blue-collar workers in the active population, and unemployment rate.
Secondly, an index SES & urban-rural was generated. In fact, the study area consists in mostly rural areas and this index may not always give a comprehensive representation of the socioeconomic level according to the degree of rurality. The construction of the SES & urban-rural index was carried out using the same method: a Principal component analysis, including the same variables more a new variable urban-rural. This urban-rural variable is categorized as urban, suburban, isolated towns and rural depending on their number of towns and respective populations and come from INSEE. The SES & urban-rural index is composed of 4 categories presented in Figure1.
Statistical analysis
Incidence risk and standardization
To assess the frequency and the spatial distribution of this disease among census blocks, we computed age standardized stroke incidence risk with 95% confidence intervals using the indirect standardization. This procedure ensured that differences in the geographic distribution of stroke risk were not affected by geographic difference of the distribution of age in the population. The stroke incidence risk was computed as the observed number of stroke cases >60 years old during each age-period divided by the expected cases, reported per census blocks respectively. The population data by sex and age period come from INSEE. The age standardized stroke incidence risk were presented per 1 000 inhabitants-year.
Ordinary Poisson regression and Local Poisson geographically weighted regression
Firstly, we investigate potential associations between stroke incidence risk, and contextual socioeconomic deprivation and urban-rural level separately for men and women using Poisson regression models. Poisson regression is a suitable regression model fitted the data using the generalized linear model function, dealing with count data applied to a small area. Significance overdispersion was detected for women models making Poisson regression inappropriate, the analysis beyond this point have preceded with negative binomial models. Secondly, the application of Koenker (BP) Statistic aims to determine whether the explanatory variable in the ordinary model have a stationary relationship with the dependent variable throughout the study area. When the p value was<0.10, a local Geographic Weighted Regression (GWR) was carried out, meaning that clusters of census blocks with higher risk of stroke incidence or/and lower risk of stroke incidence exist. When the p value was high, ordinary Poisson or negative binomial model fits better the dataset.
GWR was used to identify interesting locations (areas of variation according to the level of socioeconomic deprivation) for reducing stroke incidence investigation. Local GWR model allow to estimate as many local regression coefficients as the number of locations in the study area. In case of local Poisson GWR, this model is parameterized as follows:
See Formula 1 in supplementary information.
Where Χ(k,i) and βk are the kth explanatory variable and its local regression coefficient that is unique to location U. Thus, the regression coefficients vary based on the spatial location Ui = (ui, vi). The expected value of response of the ith observation, E[yi], is related to the linear predictor via a Poisson link function.
To calibrate this formula, a bi-square adaptive weighting kernel function is used to account for spatial structure (density, shape and size of the census blocks) and the appropriate bandwidth was selected using the golden section method (22). The locations near to i have a stronger influence in the estimation of βj(ui,vi) than locations farther from i. In the GWR model localized parameter estimates can be obtained for any location i which in turn allows for the creation of a map showing a continuous surface of parameter values and an examination of the spatial variability (nonstationarity) of these parameters. Additionally, we used corrected Akaike Information Criterion (AICc) and pseudo adjusted coefficient of determination (Adjusted R²) as goodness of fit for comparing models. GWR models have been estimated with R using GWmodel (R Development Core Team, 2011). All maps layouts were performed using ArcMap v. 10.5 from ESRI.