Moving beyond correlation in environmental epidemiology

Abstract

Discovering causal effects of specific environmental factors on population health remains a pressing task but also a major challenge, owing to the inherent complexity of environment-health relationships and the current evidence base premised largely on correlation and regression studies. The big data era has catalyzed causal science advancement, providing great opportunities to enrich the toolkit for causal studies using observational data. Besides, the present paradigm of environmental health research focuses more on avoiding the “bads” than maintaining the “goods”, which may compromise the efforts of promoting health in the whole population. Therefore, the motivation of this thesis is to move beyond correlation in environmental epidemiology and enhance causal inference from dynamic data by integrating multiple divergent but complementary methods to rigorously analyze the adverse or beneficial health effects of environmental factors. In practice, the traditional time series regression via generalized additive model (GAM) was combined with two novel causal discovery methods, namely graphical modelling algorithm Peter-Clark-momentary-conditional-independence plus (PCMCI+) and state space reconstruction–based approach empirical dynamic modelling (EDM) from chaos theory, of which the underlying assumptions and potential biases are completely different. I assessed the complementary nature of these three methods in enhancing causal inference, by the application in four environment-health case studies. Firstly, I found that ambient ozone could be another environmental driver (besides absolute humidity and temperature) of influenza in the USA which has a largely temperate climate. Ambient ozone potentially alleviates the population-level influenza epidemics through priming host immunity against infection. Secondly, the negative driving role of ozone in influenza activity was replicated in subtropical Hong Kong. In contrast to the ozone findings consistently supported by multiple methods, findings on absolute humidity and temperature varied with different methods of causal analysis. Thirdly, when examining the hypothesis of weather-triggered stroke hospitalizations in Hong Kong, GAM and PCMCI+ generated mixed effect estimates for temperature-related admission risks of hemorrhagic and ischemic stroke; EDM found that lower daily temperature would result in higher risk of stroke hospitalizations, especially hemorrhagic stroke. Fourthly, the three disparate methods have all negated the hypothesis of ambient temperature driving birth sex ratio (BSR) in Hong Kong; EDM detected a decreased (viz. female-biased) BSR following maternal exposures to temperature variability during the early trimester. This supports the classical “fragile boy” hypothesis that girls are favored during environmental hardship for higher chances to pass on the genes. This thesis is expected to raise the awareness of enriching the environmental epidemiology research toolkit and harnessing nonlinear dynamic approach as a complement in an integrative methodological framework for rigorous scientific inquiry, from which the findings can be referred to by greater research communities and policymakers, as applicable.