Geospatial and regression analyses
To examine whether and where clustering of high or low MIRs occurred by health regions across Canada, we calculated the Moran's I statistic , one of the most commonly used measurements of spatial autocorrelation in ecological, health, environmental and geological studies (\cite{pap1950},\cite{m2011}). Neighbouring regions with similar values (clustering) suggest spatial dependency, resulting in positive spatial autocorrelation; conversely, neighbouring regions with dissimilar values (dispersion) are inversely spatially dependent, resulting in negative spatial autocorrelation. Neighbouring regions with no spatial pattern (randomness) result in an autocorrelation value of zero. Applying the Local Indicators of Spatial Association (LISA) test then shows regions exhibiting clustering with high or low MIRs \cite{al-zahrani2013}. For this test, we chose a more stringent significance level of p ≤ .01 (rather than .05), to reduce the risk of Type I errors due to multiple comparisons \cite{1}.
To explore the variables that predicted MIR per health region, we first conducted univariate analyses with distance to radiotherapy center and each sociodemographic variable as the independent variables, and MIR as the dependent variable. All significant (p ≤ 0.05) independent variables in the univariate analyses were then included in a step-wise backwards elimination method of multiple regression analysis (ordinary least squares, OLS) to determine the best predictors of MIR per health region. Variables were dropped in order of least significance and if found to have high collinearity within the model, based on a Variance Inflation Score (VIF) of >10, until all remaining variables were significantly associated at p ≤ .05 .
Significant variables in the OLS model were then included in a geographically-weighted regression (GWR) analysis. GWR is a regression modeling method that accounts for spatial structure, and can therefore produce a more robust model than OLS when spatial dependency is suspected (\cite{Ford_2016},\cite{Nakaya_2005}). The global R2 of the OLS and GWR models were compared to assess the model of fit, with a higher R2 indicating that a greater proportion of the variance in MIR was explained by the model. The Akaike Information Criterion (AIC) was also used to compare the two models, with a lower value indicating higher accuracy (\cite{me1996},\cite{a2009}).
All statistical tests were conducted in JMP version 12 (XX), except for the calculation of Moran's I statistics and the GWR analyses, which were conducted using GeoDa software version 1.12.1.59 (XX). Choropleth maps were generated using Tableau version 10.4 (XX).
Saskatchewan – rates same in each region that were initially combined
Nova Scotia manipulated data for CCHS health regions (2013 vs 2015)
Role of the funding source
The funder had no role in the study design, collection,
analysis or interpretation of the data, or writing of the
report. JC had full access to all data used in the
study, and the final responsibility to submit for
publication.
RESULTS
All-cancer MIRs were calculated for the 112 health regions in Canada, for the 2010-2012 period, which revealed the lowest MIR in Ontario's York Regional Health Unit at 0.35, and the highest MIR in the territory of Nunavut at 0.88 (Figure 1). Overlaying the coordinates of the radiotherapy centers in Canada with the MIRs per health region showed that the vast majority of radiotherapy centers in 2012 were located along the country's southern border.