What is a correlation matrix?

You have a survey with a dozen rating questions and you want to know which ones travel together. Checking every pair by hand would take forever, and a pile of separate charts would hide the pattern. A correlation matrix puts all of those pairs on one page. It is the standard first step in correlation analysis in survey research, and the one table most analysts reach for before they build a model.

Key takeaways

• A correlation matrix is a table of correlation coefficients for every pair of variables, so you can see at a glance which survey measures move together.
• To read one, remember that the diagonal is all 1s, the table mirrors itself across the diagonal, and every value runs from -1 to +1.
• As a rule of thumb for survey data, an absolute r of 0.2 to 0.39 is weak, 0.4 to 0.59 is moderate, and 0.6 to 0.79 is strong.
• Match the coefficient to your data, using Pearson for numeric linear relationships and Spearman for ordered ratings such as 5-point scales.
• A matrix of 20 variables has 190 pairs, so some will look significant by chance, and surprising pairs should be treated as leads, not findings.

How do you read a correlation matrix?

The rows and columns list the same variables in the same order. Each cell shows the correlation coefficient (usually Pearson’s r) for the pair that meets there. Three features make the table quick to read.

• The diagonal is all 1s. Every variable correlates perfectly with itself, so the cells running from top left to bottom right always read 1.
• The table is symmetric. The correlation of A with B equals the correlation of B with A, so the numbers above the diagonal mirror those below it. Many reports show only one triangle.
• Every value sits between -1 and +1. The sign gives the direction and the size gives the strength.

A positive value means the two variables rise together. A negative value means one tends to fall as the other rises. A value near 0 means there is no linear relationship, though a curved relationship can still exist.

What counts as a strong correlation?

Cutoffs vary by field, so treat these as rules of thumb for survey and attitudinal data, based on the absolute value of r.

Absolute value of r	Rough label
0.00 to 0.19	Negligible or very weak
0.20 to 0.39	Weak
0.40 to 0.59	Moderate
0.60 to 0.79	Strong
0.80 to 1.00	Very strong

A correlation of -0.70 is as strong as +0.70. Only the direction differs.

How is each number calculated?

Each coefficient is the covariance of two variables divided by the product of their standard deviations. Covariance measures how much the two variables vary together. Dividing by the standard deviations strips out the units, which is why a rating scale and a dollar amount can sit in the same table.

Here is the arithmetic for five respondents who rated satisfaction (x) and likelihood to recommend (y) on a 0–10 scale.

Respondent	x	y	x minus 6.2	y minus 6	Product
1	3	2	-3.2	-4	12.8
2	5	5	-1.2	-1	1.2
3	6	5	-0.2	-1	0.2
4	8	8	1.8	2	3.6
5	9	10	2.8	4	11.2

The products sum to 29. The squared deviations sum to 22.8 for x and 38 for y. So r = 29 / √(22.8 × 38) = 29 / 29.44 = 0.985. A matrix repeats this for every pair.

What does a correlation matrix look like in practice?

Imagine a survey of 1,000 customers. They rate satisfaction, likelihood to recommend and perceived value on 0–10 scales, and we also record age. The numbers below are illustrative, not from a real study.

	Satisfaction	Recommend	Value	Age
Satisfaction	1.00	0.82	0.64	0.05
Recommend	0.82	1.00	0.58	0.09
Value	0.64	0.58	1.00	-0.31
Age	0.05	0.09	-0.31	1.00

Reading it:

• Satisfaction and Recommend (0.82): very strong and positive. Customers who rate their experience highly say they would recommend the brand.
• Satisfaction and Value (0.64) and Recommend and Value (0.58): moderate to strong. Perceived value tracks both outcomes, though less tightly.
• Value and Age (-0.31): weak and negative. Older respondents tend to rate value a little lower.
• Age and Satisfaction (0.05): negligible. Age tells you almost nothing about satisfaction here.

With 1,000 respondents, any r above roughly 0.06 in absolute value is statistically significant at the 5% level. That makes 0.09 (Age and Recommend) technically significant, yet far too small to act on. Significance says the relationship is probably not zero. It does not say the relationship matters.

What types of correlation matrix are there?

The matrix is a layout, and the coefficient inside it is a choice. Match it to your data.

• Pearson measures linear relationships between numeric variables. It is the default.
• Spearman and Kendall’s tau use ranks, so they suit ordered ratings and relationships that rise or fall steadily without being straight lines. They are less affected by outliers.
• Point-biserial compares a two-category variable, such as owner versus non-owner, with a numeric one.
• Biserial applies when a continuous variable has been split into two groups and you want the correlation with the underlying scale.

For a 5-point scale, Spearman is often the safer choice.

How do you make a correlation matrix?

1. Pick the variables, all measured in a form that suits your coefficient. Put them in columns, with one row per respondent.
2. Decide how to handle missing answers. Pairwise deletion uses every respondent who answered both questions in a pair. Listwise deletion drops anyone with any missing answer.
3. Calculate the coefficient for every pair, as in the arithmetic above. Software does this for you: CORREL in a spreadsheet, df.corr() in Python’s pandas, cor() in R, or the correlation tools in SPSS.
4. Lay the results out as a square table and format it for reading. Round to two decimals and consider a color scale.
5. Add the sample size, and flag which coefficients are statistically significant.

A heatmap, where color shows strength and direction, is the usual way to scan a large matrix. The example below shows a Pearson correlation matrix as a heatmap: red cells are positive, purple cells are negative, and one half of the grid is left blank because it would only repeat the other. For more on choosing the right chart, see our guide to visualizing correlations in survey analysis.

Does correlation mean causation?

No. A strong coefficient tells you two measures move together. It does not tell you why. Satisfaction might drive recommendation, recommendation might shape how people report satisfaction, or a third factor such as a recent good experience might lift both.

Use the matrix to find relationships worth investigating, then test them with a designed experiment, a regression model or a follow-up question. Once you fit a model, residuals show where its predictions miss.

What are the common mistakes?

• Reading 0 as “no relationship.” Zero means no linear relationship. A U-shaped pattern can have a coefficient near 0, so plot the pair before you rule it out.
• Ignoring outliers. A few extreme respondents can create or erase a Pearson correlation. Check scatterplots or use Spearman.
• Treating significance as importance. With large samples, tiny correlations are significant. Look at the size of r.
• Scanning too many pairs. A matrix of 20 variables has 190 unique pairs. Some will cross a significance threshold by chance, so treat surprising ones as leads, not findings.
• Correlating variables that overlap by design. Two questions that ask about the same idea will correlate highly. That is a measurement result, not a discovery.
• Using the wrong coefficient. Pearson on unordered categories gives meaningless numbers.

How do you build one in practice?

In Halo Reports you can cross survey questions, apply weighting and run significance testing, then publish the result as charts and an interactive report page. See Halo Reports for what it covers. If the matrix points to satisfaction or loyalty, you can see what drives them by segment.

For your next project, start small. Pick four or five variables you have a reason to compare, build the matrix, and plot any pair that surprises you before you report it. Insights teams with wide, complex studies can keep that work in one governed platform.