Skill: Pairwise Correlation Analysis

Use When

• User wants to scan for linear or monotonic relationships among multiple numeric variables
• User needs a correlation matrix with p-values for gene expression, neural activity, or metabolomics data
• User wants to identify highly correlated features before modeling (collinearity check)
• User is building a co-expression or co-activity network

Inputs

• Required:
  • Data matrix (CSV or TSV): rows = observations, columns = numeric variables
• Optional:
  • Correlation method: pearson (default), spearman, or kendall
  • Multiple-testing correction method: fdr (Benjamini-Hochberg, default) or bonferroni
  • Significance threshold for highlighting (default: 0.05)
  • Output directory (default: ./corr_output)

Workflow

1. Read the data matrix; keep only numeric columns.
2. Compute the full pairwise correlation matrix using the specified method.
3. Compute p-values for each pair using the appropriate test (Pearson: t-distribution; Spearman/Kendall: permutation or asymptotic).
4. Apply multiple-testing correction across all variable pairs.
5. Generate a correlation heatmap: color-coded by correlation coefficient, asterisks for significant pairs (adjusted p < threshold).
6. Generate a scatter plot matrix (pairs plot) for datasets with ≤ 10 variables.
7. Report the top 20 strongest positive and top 20 strongest negative correlations.
8. Write correlation matrix (TSV), adjusted p-value matrix (TSV), and plots (PDF) to output directory.
9. Save the executable analysis script (analysis_code.py or analysis_code.R) and software version manifest (session_info.txt) used to produce results.

Output Contract

• Correlation matrix (TSV): variable × variable, values = correlation coefficients
• Adjusted p-value matrix (TSV): variable × variable, values = corrected p-values
• Correlation heatmap (PDF)
• Pairs plot (PDF, only if ≤ 10 variables)
• Top correlations summary table (TSV): var1, var2, correlation, p_value, adj_p_value
• Reproducibility artifacts: analysis_code.py or analysis_code.R, plus session_info.txt

Limits

• Pearson correlation assumes linearity; use Spearman or Kendall for non-linear monotonic relationships or ordinal data.
• Correlation does not imply causation.
• Kendall's tau is slow for n > 10000; Spearman is preferred for large samples.
• Common failure: constant or near-constant columns produce undefined correlations — remove before running.