The research project has a double purpose: firstly to investigate statistical methods suited to the analysis of associations between the spatial variations of health indicators and risk factors, then to apply these methods to study the link between some indicators of low dose radiation exposure, industrial pollution and the mortality for cancer of specific sites by using aggregated data collected at a regional level.
The suitability of statistical methods was investigated for the analysis of models of association between spatially defined variables. The methods were then applied to the geographical association in France between the mortality rates for some radiosensitive sites (lung, breast, thyroid, leukemia) and background radiations, taking into account appropriate geographical confounders.
The development of parametric tests of association between spatially distributed variables was complemented by a study of nonparametric methods. As a preliminary step, the robustness of the modified tests of the correlation coefficient to some patterns of departure from normality was investigated. The performance of the modified tM-2 test for non-Gaussian variables, Spearman's rank correlation test for large samples and Monte Carlo tests were investigated.
The tests were then applied to the analysis of the correlation between population radiation exposure and cancer mortality. There was weak evidence of a link between lung cancer mortality and gamma radiation exposure for the period 1968 to 1969 with similar correlations coefficients for men and women, indoor or outdoor exposure. For the other 2 time periods, there was no evidence of any pattern and the correlations are nonsignificant contrary to what would be expected if the 1968 correlations reflected a direct link. No geographical association was found between lung cancer rates and radon exposure.
No positive association was found between gamma radiation exposure and either breast cancer or leukemia for any of the 3 periods analysed.
There was some evidence of geographical association between female thyroid cancer rates and gamma radiation. Simple correlation coefficients were broadly similar for the 3 periods even though the modified tM-2 statistic was only statistically significant for the first and third period. Some evidence of positive association was maintained after adjustment on a gradient of distance to sea, a covariate which is significantly linked to female thyroid cancer for the periods 1968 and in 1984. No geographical association was found for male thyroid cancer.
The only positive result to have emerged from the geographical analysis was that of a possible link of gamma radiation with thyroid cancer in females. For this site, mortality represented only a small fraction of the incident cases and hence the analysis would have been considerably more powerful if geographical incidence data were available. Thyroid cancer mortality finding for men. Further analyses at a more detailed geographical level would be interesting. For rare cancer sites such as thyroid cancer or leukemia, an analysis which takes into account the fluctuations of the observed mortality counts around their mean is also warranted.
Correlation and regression are the statistical methods that are commonly used for testing association between a set of variables. The standard framework for these methods assumes that the sample consists of variables measured at different geographical locations since typically these variables will exhibit some spatial dependence (autocorrelation). Hence modifications of the classical tests of the correlation or regression coefficients are necessary.
In the statistical part of this project, several ways of modifying or adapting the standard tests of association to take into account spatial autocorrelation are considered and compared. Firstly modified tests of simple or partial correlation coefficients are developed and their statistical performance is assessed. These tests are based on an estimation of an effective sample size which is typically less than the number of geographical locations in the case of positive autocorrelation. Secondly multiple regressions with different types of spatial structure for the errors are studied. This allows a sensitivity analysis of regression results to the choice of errors structure to be performed. Finally the different methods will be compared on an example.
The epidemiological part of this project consists of the analysis of a data file which contains mortality rates for some cancer sites (lung, breast, thyroid, leukaemia) for three time periods (per department); estimation of radiodiagnostic examinations exposure (per region); measures of radon and background irradiation exposures (some departments); cigarette consumption per inhabitant (per department); and percent of employed male population in specific industrial branches (per department). This analysis will be carried out using the different methods developed.