Correlation heatmap with non-significant cells masked

A correlation heatmap with 15 variables shows 105 correlations. Coloured in full, most of those colours are noise, and the eye finds patterns in noise very readily.

Why mask

Masking hides cells whose correlation is not statistically significant, leaving only those your data actually support. It turns a decorative grid into a figure a reader can rely on.

Common approaches: blank the cell, grey it out, or overlay a cross.

The multiple comparisons problem

With 105 correlations tested at p < .05, roughly five will appear significant by chance alone. Masking at an uncorrected threshold still leaves you presenting noise.

Apply a correction across the matrix. Benjamini-Hochberg (false discovery rate) is usually the right choice for exploratory correlation work: less severe than Bonferroni, and it controls the proportion of your reported findings that are false. State in the legend which correction you used.

Use a diverging colour scale

Correlations run from -1 to +1 with a meaningful midpoint at zero. That requires a diverging scale such as coolwarm or RdBu, centred on zero. A sequential scale like viridis implies that -1 and +1 are opposite ends of one continuum, when they are equally strong relationships in opposite directions.

Consider clustering

Reordering rows and columns by hierarchical clustering groups related variables together and makes structure visible that alphabetical ordering hides.

In Plotara

Heatmaps support correlation, direct and pivot modes, with significance masking, row and column clustering with a choice of linkage, and diverging colour scales by default for correlation data.