Refuting Causal Relations in Epidemiological Time Series

Abstract

Causal detection is an important problem in epidemiology. Specifically in infectious disease epidemiology, knowledge of causal relations facilitates identification of the underlying factors driving outbreak dynamics, re-emergence, and influencing immunity patterns. Moreover, knowledge of causal relations can help to direct and target interventions, aimed at mitigating outbreaks. Infectious diseases are commonly presented as time series arising from nonlinear dynamical systems. However, tools aiming to detect the direction of causality from such systems often suffer from high false-detection rates. To address this challenge, we propose BCAD (Bootstrap Comparison of Attractor Dimensions), a novel method that focuses on refuting false causal relations using a dimensionality-based criterion, with accompanying bootstrap-based uncertainty quantification. We test the performance of BCAD, demonstrating its efficacy in correctly refuting false causal relations on two datasets: a model system that consists of two strains of a pathogen driven by a common environmental factor, and a real-world pneumonia and influenza incidence time series from the United States. We compare BCAD to Convergent Cross Mapping (CCM), a prominent method of causal detection in nonlinear systems. In both datasets, BCAD correctly refutes the vast majority of spurious causal relations which CCM falsely detects as causal. The utility of BCAD is emphasized by the fact that our models and data displayed synchrony, a situation known to challenge other causal detection methods. In conclusion, we demonstrate that BCAD is a useful tool for refuting false causal relations in nonlinear dynamical systems of infectious diseases. By leveraging the theory of dynamical systems, BCAD offers a transparent and flexible approach for discerning true causal relations from false ones in epidemiology and may also find applicability beyond infectious disease epidemiology.