Role of mammography accessibility, deprivation and spatial effect in breast cancer screening participation in France: an observational ecological study

Study setting

A descriptive, cross-sectional ecological study design was employed to assess the geographic variation in mammography screening participation in 2015 and 2016 in Lyon MA. Lyon MA is the third largest MA in the region of Auvergne-Rhône-Alpes in the Eastern region of France. It is composed of 59 municipalities and 510 CBs, with a total population of 1,381,349 inhabitants in an area of 534 km2 in 2016. This area was of interest because participation inequalities in mammography screening persist despite previous intervention studies implemented in highly deprived CBs [26].

Data and methodsData

For this study, we analyzed factors that influence the mammography participation rate. Due to the ecological design of the study, no patients or the public were involved in the study's planning, design, conduct, or reporting. Participation rates were calculated as the percentage of eligible women (50–74 years and invited by NMSP for a free mammography screening) in a CB in 2015–2016. Data were extracted from the NMSP 2015–2016 dataset of the Auvergne Rhone Alpes region. We focused our study on the spatial effects of mammography participation and the spillover sources, including spatial accessibility, deprivation level, modes of travel, and variables for social cohesion.

Geographic accessibility

Geographic accessibility factors include the travel distance to the closest accredited mammography service, the degree of urbanization and the density of GPs. To be considered accredited, the radiologists must be registered members of the NMSP. Travel distance was calculated with the Network Analyst function in ArcGIS (version 10.7) using the French road network. The geographic position of each CB is required as input for the spatial analyses, so we derived the position of the centroid (a measure for the geographical center point of a polygon) for each CB using a zonal geometry function in ArcGIS. The distance (kilometers) from each centroid of the CB to the closest accredited mammography service was calculated based on the shortest driving road. Thus, the travel distance is the measure of distance to be traveled by driving to reach the closest mammography service, considering the realistic road network, type of roads, and corresponding driving speeds along with possibilities of driving preferences such as avoiding highways and crossing the bridge.

The density of GPs was used to represent the professional’s advice and guidance for women to participate in the NMSP. The density of GPs was calculated as the number of general practitioners per 100 women invited to participate in mammography screening.

We used an urban–rural index to represent the degree of urbanization. This index was built using the Principal Components Analysis (PCA) method, a statistical technique for reducing the dimensionality of a dataset. The first component of a final PCA with selected variables corresponds to the score of an urban–rural index from residential and rural to urban CBs. Variables included in the PCA were obtained from the CORINE-Land-Cover [27] and French National Institute for Statistics and Economic Studies (INSEE) in the followings domains: housing characteristics, population density, professional mobility (population who work outside of their municipality), residential mobility, and green space. CBs were classified into three classes: urban, peri-urban, and rural (Fig. 2). CBs located in urban areas are characterized by small homes, unstable housing, working within municipalities, and high population density. CBs in rural areas are characterized by large houses with multiple cars, stable housing, working outside municipalities of their residence, green space, and low density.

Adjusted confounders

To account for the socioeconomic disparity, we classified CBs by the socioeconomic deprivation level, measured by a deprivation index. This index was defined and evaluated in previous studies that investigated environmental and health inequalities [6, 28], and it was built by performing multiple PCAs to study redundant variables. The first component of a final PCA with selected variables corresponds to the deprivation score. Data on the following domains were obtained from the French national census of 2014 collected from INSEE [28]: employment, single parent family, education, occupation, immigration status, and proportion of social housing. CBs were classified using tertiles as low (− 2.076, − 0.548), medium (− 0.549, 0.133), and high (0.134, 2.994) levels of deprivation according to the distribution of the score.

To account for the mobility of the population inside and across the CBs, we used two variables: the proportion of the population who used a car for their daily mobility and the proportion of people having no access to any transportation (public transport, car or bike). Data on modes of travel for 2014 were also obtained from INSEE [29]. To test the influence of social cohesion on mammography screening participation, we used two variables as a proxy for social cohesion: the proportion of the married population and the proportion of people living alone in a household. The data were derived from INSEE 2016.

Statistical analysis

To determine whether a significant difference in the outcome and covariates existed between the three groups of the degree of urbanization, non-parametric Kruskal–Wallis tests with a critical p-value of 0.05 were further used. We mapped spatial patterns in the distribution of mammography screening participation rates by CB and then measured the degree of spatial autocorrelation of the dependent and independent variables by calculating the Global Moran’s I [30]. Briefly, the Global Moran’s I falls between − 1 and 1; a positive Moran’s I value indicates a positive spatial autocorrelation—that the nearby areas have similar values (i.e., clustered)—and a negative value indicates a negative spatial autocorrelation—that the nearby areas have different values (i.e., dispersed). To calculate Moran’s I statistic, we defined a contiguity matrix W using queen continuity weights to define the neighbourhood structure to indicate whether or not CBs share a common boundary.

In this study, we were primarily interested in mammography participation rate patterns that can be explained by the geographic accessibility (i.e., travel distance, density of GPs, degree of urbanization with the rural CBs as references) or/and the deprivation level. To assess the contribution of these factors, we first applied ordinary least squares models adjusted with these factors alone (OLS1), adjusted for mobility and social cohesion (OLS2), and with the interaction between deprivation and degree of urbanization (OLS3). Because observations associated with spatial units may reflect measurement error, spatial autocorrelation was tested on the residuals of each OLS1, OLS2 and OL3 models using the Global Moran’s I [30]. The global Moran’s I was statistically significant (p < 0.05) suggesting the necessity to account for the spatial configuration of the CBs in the study (Additional file 1: Table S1).

If spatial autocorrelation was detected, a spatial autoregressive model is required to avoid violating the OLS assumption of independence between features and to ensure that our estimates are unbiased [13, 31]. The spatial autoregressive model incorporates a diffusion process across the geographical location in which the participation rates in a CB are affected by explanatory variables in the same CB and the adjacent ones, and at the same time is also influenced by the participation rates in the adjacent CBs [32].

To decide which spatial autoregressive model is more appropriate, we used the decision rule suggested by Florax and Rey [33] based on the Lagrange Multiplier tests and their robust counterparts by Anselin [34]. Briefly, the choice of model depends on the significance of LAG or error models and their robust forms: robust LM-lag and robust LM-error. If LM-lag is statistically significant and LM-error is not, then Spatial Lag Model (SLM) is appropriate and not the Spatial Error Model (SEM) model. Conversely, if LM-error is statistically significant and LM-lag is not, then the appropriate specification is a SEM and not a SLM model. We conducted diagnostics and the LM-lag of the mammography participation model was statistically significant (p < 0.001) and the LM-error was not (p > 0.05). Therefore, we selected a SLM for subsequent analyses (Additional file 2: Table S2).

An SLM assumes that there is a spatial dependence in the dependent variable, whereas SEM assumes that there is a spatial dependence in the error term. An SLM can be expressed as Eq. (1):

$$y = \rho Wy + \beta + \varepsilon$$

(1)

$$\varepsilon \sim N\left( } In} \right)$$

(2)

with y as the dependent variable (participation rate), W denotes the spatial weight matrix and ρWy is the spatially y value. β is a vector of coefficients of the explanatory variables X. The error term, ε, follows a normal distribution with a mean 0 and a variance σ2In, where In is a n x n identity matrix.

We first perform adjusted analyses with geographic accessibility factors (i.e., travel distance, density of GPs, degree of urbanization with the rural CBs as references) and the deprivation level alone (SLM1), adjusted for mobility and social cohesion (SLM2), and with the interaction between deprivation and degree of urbanization (SLM3). After each spatial model, the Akaike Information Criterion (AIC) and the pseudo-adjusted coefficient of determination (adjusted R2) were used as a goodness of fit estimation for model comparison. We tested the residuals of each spatial lag model for spatial autocorrelation using Moran’s I statistics. We chose a significance level of 0.05 to determine statistical significance. Statistical analyses were performed using R version 4.2.1 (2022-06-23). All GIS processes and map layouts were performed using ArcMap v.10.5 (ESRI, Redlands, CA, USA).

留言 (0)

沒有登入
gif