Pathogen seasonality and links with weather in England and Wales: a big data time series analysis

Cherrie, Mark P. C.; Nichols, Gordon; Iacono, Gianni Lo; Sarran, Christophe; Hajat, Shakoor; Fleming, Lora E.

The primary strength of the analysis is the large infectious disease dataset, which is nationally representative and has information on a wide range of pathogens. We have shown how a well-known clustering algorithm (k-means) can be applied to these data to classify pathogens by their relationship with weather variables. We have utilised a number of weather parameters from the MEDMI database, which allowed for subtle differences in correlation to be illustrated. The use of two methods to detail seasonal patterns was also a strength of the analysis. The advantages of using a TBATS model is that it automatically selects Fourier terms and other aspects of the model, whilst allowing for seasonality to change over time. Wavelet analysis could be used to test for the robustness of the findings in future analysis. By sub-setting the data on the basis of seasonality detected using the difference in model fit statistics between a 'seasonal' and 'non-seasonal' model, it was less likely that the correlations with climate in the following analysis were spurious. This is akin to defining an exclusion criterion in the design of an epidemiological study to reduce the effect of bias. Having detailed the strengths and limitations of the current analysis, in the following sections we aim to explain the results in relation to previously published work under headings based on the explanations for seasonality outlined by Grassly and Fraser [3] . The data linkage was at the England and Wales level which has certain advantages (reducing noise in the data), however public health applications often require predictions at a variety of smaller scales [29] . Analysis at a local level would complement the results presented here by showing the context in which national level predictors hold.