Using reproducible research pipelines to help disentangle health effects of environmental changes from social factors

Abstract

The scientific questions motivating this thesis relate to the health effects of environmental changes including droughts, bushfires, woodsmoke, duststorms and heatwaves. Such questions require us to attempt to disentangle health effects of environmental changes from social factors as all diseases have multiple causal factors. Environmental exposures should be explored in the context of many other variables that comprise the biological and socioeconomic milieu. Investigators often narrow the focus to a single environmental cause and health effect. A simple example is bushfire smoke and its direct effects on cardiorespiratory disease. A more complex example is the indirect relationship between drought and suicide. Even simple questions require consideration of numerous putative causes and confounders. Adequately controlling for all these factors in statistical models is difficult. Furthermore results might be sensitive to choice of analysis procedure, or otherwise error-prone due to the many steps. Such difficulties have led to what some researchers assert is a ‘reproducibility crisis’ where many scientific publications are difficult or impossible to reproduce. This, with fallacious findings, harms scientific credibility. Reproducibility of data analysis is defined as the ability to recompute results, given a dataset and knowledge of the method’s steps. A key problem impairing reproducibility is inadequate documentation of the numerous steps and decisions required for the computations. Reproducible research pipelines allow data and software (such as analysis code) to be disseminated with publications, enhancing reproducibility. However, this approach often places a considerable burden on the analyst. This thesis identifies effective methods for implementing reproducible research pipelines in environmental epidemiology, aiming to reduce this burden. In addition to the contribution to methodology which this constitutes, the thesis also includes a range of peer reviewed papers (along with accompanying datasets and software packages of code) published by the author, which also add to knowledge. Key findings include health effects of environmental changes relevant to debates about climate change. Reproducibility of these findings enhances their credibility in response to the heightened scepticism of those debates. Important insights included the finding that the risk of suicide in New South Wales increases in rural men during drought but decreases during droughts for women. Another striking finding was that while bushfire smoke and duststorms each increased cardiorespiratory mortality risk in Sydney, they appear to do so in different ways, with dust having a much higher risk estimate than biomass smoke. In cases such as this where findings are novel, unexpected or contradict accepted opinion, the scientific method stresses the need for scepticism and critical review. Reproducible research pipelines strengthen our ability to conduct such review beyond what was available in the traditional research model. Not only does the use of pipelines make methodological choices and assumptions more transparent; doing so also safeguards against data misuse by making errors easier to find. Encoding analysis steps in a computer ‘scripting’ language and distributing the data and code with publications aids readers to assess (and challenge) each choice of data or methods. This will help minimise mistakes in the execution or interpretation of research.