Finding leading indicators for disease outbreaks: filtering, cross-correlation, and caveats

Area of Science:

• Epidemiology
• Biostatistics
• Medical Informatics

Background:

• Emerging infectious diseases and bioterrorism necessitate rapid outbreak detection systems.
• Identifying leading indicators from diverse data sources is crucial for early disease outbreak warnings.
• The sample cross-correlation function (CCF) is commonly used to assess lead-lag relationships in time-series data.

Purpose of the Study:

• To evaluate the influence of methodological choices on the CCF's ability to detect lead-lag relationships in time-series data.
• To demonstrate the bias inherent in the sample CCF when analyzing surveillance data for outbreak detection.

Main Methods:

• Utilized simulation studies to assess the performance of the sample CCF.
• Drew upon methodologies from econometrics and environmental health research.
• Investigated the impact of different data scales on CCF results.

Main Results:

• The sample CCF is highly susceptible to bias, particularly from long-scale phenomena.
• Long-term trends can obscure shorter-term patterns, leading to misleading interpretations.
• Failure to account for scale can mask important features in surveillance data.

Conclusions:

• Researchers must specify the scale of interest (e.g., short-term spikes vs. long-term seasonality) when analyzing time-series data.
• Appropriate data filtering is essential to mitigate bias and reveal true lead-lag relationships.
• Unfiltered CCF analysis of bi-variate time-series data can yield ambiguous or incorrect conclusions for outbreak detection.