Read one cell completely
If the row is X and column is Y and the cell is r = −.42, the variables have a negative linear association under Pearson correlation. The mirrored Y–X cell repeats the same value. That does not establish causation, agreement, or a practically important effect.
Use correlation strength interpretation for attributed magnitude language and Pearson versus Spearman to identify the coefficient family.
A three-variable example
Suppose X–Y is .68, X–Z is .65, and Y–Z is .12. X is associated with both variables, but Y and Z are only weakly associated with each other. That pattern might reflect two distinct relationships with X, a grouping variable, nonlinearity, or different missing-data samples. The matrix alone cannot choose among those explanations.
Inspect the corresponding scatter plots and pairwise sample sizes. If the matrix displays p-value stars, remember that three variables create three unique tests, while 20 variables create 190. A star-filled large matrix can be driven more by multiplicity and sample size than by practical magnitude.
Matrix-level traps
- Multiple comparisons: dozens of pairs create many opportunities for extreme sample correlations.
- Pairwise missingness: different cells can use different observations and may not form a coherent positive-semidefinite matrix.
- Redundancy: very high predictor correlations can indicate overlapping measurement, not a reason to delete automatically.
- Clusters: blocks can reflect domains, methods, groups, or shared confounding.
- Color scales: asymmetric or truncated palettes can exaggerate weak differences.
Record n per pair where missingness varies. Confidence intervals or multiplicity-aware inference are more informative than stars alone.
Correlation is not a loading matrix
A PCA or factor-loading table maps variables to components or factors and is not read as pairwise association. PCA output and factor loadings covers that artifact.
A defensible reporting sentence
“Pearson correlations ranged from −.42 to .68. The strongest block involved the named variables, but pairwise sample sizes differed because of missingness, and intervals were examined before interpreting rank order.”