Extremes in health-meteorology relationship, applied to cardiovascular diseases in Québec, Canada

Cardiovascular diseases (CVDs) are serious events which can lead to hospitalisation or death. More and more scientific publications examine how meteorological conditions can affect them. They usually focus on the mean statistics of the health issues (e.g. mean numbers or rates of hospitalisations and deaths). However, extreme values are also problematic and cause management issues in the public health network (e.g. hospital emergencies). It is therefore essential to consider studying the extremes with the appropriate statistical tools. As most parts of the world, the province of Québec has experienced an increase in mean temperatures over the last 30 years which is expected to continue for the next 60 years. This would bring fewer cold spells, particularly in southern Québec, while heat waves would become more frequent and would last longer. Consequences could be important on public health since temperature has a strong effect on cardiovascular diseases. It is therefore necessary to better understand their relationship to meteorological conditions (temperature, humidity, atmospheric pressure, precipitations, snow), in order to prevent or at least to reduce climatic changes consequences and to improve health care management. The aim of this study is to focus on establishing relationships with extreme health issues using appropriate statistical tools and then consider the case of CVDs in Québec as an application.

Extremes techniques are required to specifically study extremes series of the sanitary variables. Two approaches are employed to this purpose: extreme value theory (EVT) and quantile regression (QR). EVT has been seldom used in a public health context, while it is well exploited in finance and hydrology. First, two methods contribute to select extreme series to be modeled: threshold and block methods. Then, the selected extreme observations are fitted with an extreme distribution, including in particular Generalized Pareto, Generalized Extreme Value, Lognormal, Gamma or Exponential. This univariate modeling is a required step, since EVT has never been used in the health context, not to mention CVDs. Then, in order to establish relationships between CVDs peaks and meteorology, we use Generalized Additive Models (GAM). GAM is largely employed in usual health-weather studies. The flexibility and nonlinearity of GAM is very useful in the case of temperature for instance, as more people die due to CVDs in both very cold and very hot situations, but there is a comfort zone between the two, implying a “U” or “J” relationship.

The second approach is QR. Even though QR is well known in medicine, it has not been used to study extremes of sanitary variables in connection with weather. Whereas classical regression models study the conditional mean response, QR focuses on a specified conditional quantile. Extremes can be classified via high quantiles, e.g. observations greater than the 95th or 99th quantile can be considered extremes. Thus, there is no need to extract extreme observations, as with EVT. This is interesting since we are now able to detect different trends (if any). A meteorological variable can be non-significant in the 90th quantile although it could become so in the 99th quantile, or it can have a different effect. Furthermore, QR is combined to non-linear techniques for more flexibility.

The obtained results show that temperature, relative humidity, precipitation and snow have an impact on extreme events of hospitalisations and deaths, whereas atmospheric pressure does not. In particular, temperature is found to be significant in most extreme quantiles for both hospitalisations and deaths (QR case). It contributes a lot more to the explained deviance of the established models than usual analytical models (EVT case). Moreover, in the case of sanitary extremes, the shapes of the relationships are different from what they are in the mean case (non extreme). Note that the proposed procedure is valid for other chronic diseases and other regions.