Identifying correlations driven by influential observations in large datasets.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• High-throughput data analysis enables simultaneous interrogation of numerous variables.
• However, large datasets are prone to spurious correlations and false positives.
• Existing methods often fail to identify outliers causing these inaccuracies.

Purpose of the Study:

• To introduce an open-source method for detecting false positives in correlation analysis.
• To address the challenge of spurious correlations in large-scale omics datasets.
• To provide a robust approach for identifying unreliable correlations and their sources.

Main Methods:

• Development of Correlations Under The InfluencE (CUTIE), a jackknifing-based statistical method.
• Application of CUTIE with both parametric and non-parametric correlation measures.
• Utilizing CUTIE to identify outlier variables or samples driving false correlations.

Main Results:

• CUTIE effectively detects false positives in correlation analyses of large datasets.
• The method can rescue previously missed or incorrectly signed correlations.
• A meta-analysis confirmed the pervasiveness of this issue across omics data, with microbiome data being particularly affected.

Conclusions:

• Outliers significantly impact correlation analysis in high-throughput data, leading to widespread false positives.
• CUTIE offers an efficient, automated solution for identifying unreliable correlations in large datasets.
• This method enhances the reliability of correlation findings in various scientific domains, especially omics research.