Correlation Matrix

What a correlation matrix shows

It summarises the direction and strength of their pairwise dependence on a standardised scale, with the same variables listed in the rows and columns.

The matrix helps analysts inspect relationships, diagnose modelling problems, and represent dependence in multivariable analysis or simulation. It does not by itself establish causation, describe every form of dependence, or show how relationships change after adjustment for other variables.

The mathematical structure

For variables $X_1,\ldots,X_p$, the correlation matrix $R$ has elements $\rho_{ij}=\operatorname{Corr}(X_i,X_j)$. A valid correlation matrix is symmetric, has ones on its diagonal, and is positive semidefinite.

$$R= \begin{bmatrix} 1 & \rho_{12} & \cdots & \rho_{1p}\\ \rho_{21} & 1 & \cdots & \rho_{2p}\\ \vdots & \vdots & \ddots & \vdots\\ \rho_{p1} & \rho_{p2} & \cdots & 1 \end{bmatrix}$$

Symmetry follows because $\rho_{ij}=\rho_{ji}$, while each diagonal entry equals one because a non-constant variable is perfectly correlated with itself. Positive semidefiniteness requires:

$$z^{\mathsf T}Rz\geq 0$$

for every real vector $z$. Equivalently, all eigenvalues of $R$ must be non-negative, allowing for a small numerical tolerance when the matrix has been estimated or stored in finite precision.

Calculating a Pearson correlation

Pearson's correlation measures linear association between two variables with finite, non-zero variances. It standardises their covariance by the product of their standard deviations, making the result unit-free.

For a sample of size $n$:

$$r_{XY}= \frac{\sum_{i=1}^{n}(x_i-\bar{x})(y_i-\bar{y})} {\sqrt{\sum_{i=1}^{n}(x_i-\bar{x})^2} \sqrt{\sum_{i=1}^{n}(y_i-\bar{y})^2}}$$

The coefficient lies between $-1$ and $1$. A positive value means that larger values of one variable tend to accompany larger values of the other, a negative value means that larger values tend to accompany smaller values, and a value near zero indicates little linear association rather than general independence.

A worked three-variable matrix

Suppose an analysis examines annual cost, QALYs, and adverse-event count. An illustrative correlation matrix is:

$$R= \begin{bmatrix} 1 & 0.40 & 0.30\\ 0.40 & 1 & -0.20\\ 0.30 & -0.20 & 1 \end{bmatrix}$$

The matrix states that cost and QALYs have correlation $0.40$, cost and adverse-event count have correlation $0.30$, and QALYs and adverse-event count have correlation $-0.20$. The determinant is:

$$\det(R)=1+2(0.40)(0.30)(-0.20) -0.40^2-0.30^2-(-0.20)^2=0.662$$

The positive determinant is compatible with positive definiteness in this three-variable example, but determinant alone is not a sufficient general validation test. A full check should confirm symmetry, diagonal values, coefficient bounds, and non-negative eigenvalues.

Correlation and covariance contain different information

Covariance retains the measurement units and scale of the variables, whereas correlation removes scale. A covariance matrix is needed when the magnitude of joint variation matters, while a correlation matrix is convenient for comparing dependence across differently scaled variables.

The conversion between correlation and covariance is:

$$\operatorname{Cov}(X_i,X_j)=\rho_{ij}\sigma_i\sigma_j$$

If the standard deviation of annual cost is £1,000, the standard deviation of QALYs is $0.20$, and their correlation is $0.40$, then:

$$\operatorname{Cov}(C,E)=0.40(1{,}000)(0.20)=80$$

The covariance has joint units of currency-times-QALYs. The complete covariance matrix can be recovered from a correlation matrix $R$ and diagonal standard-deviation matrix $D=\operatorname{diag}(\sigma_1,\ldots,\sigma_p)$ using $\Sigma=DRD$.

Choosing the type of correlation

Different coefficients capture different forms of association and rely on different assumptions. A matrix should state which coefficient was used rather than presenting every value generically as “correlation.”

Mixing coefficients within one matrix can make it unsuitable for decomposition or simulation unless the construction is theoretically and mathematically justified. The variable types, purpose, distribution, and downstream method should determine the coefficient and estimator.

Correlation does not imply causation

A matrix records pairwise statistical association, not the direction or mechanism of an effect. Confounding, reverse causation, selection, measurement error, common time trends, or mathematical coupling can create correlations without a direct causal relationship.

Conditioning on other variables can also change a correlation or even reverse its sign. Causal questions require a defined estimand, study design, temporal structure, and identification assumptions; a correlation matrix can support exploration but cannot replace causal inference.

Pairwise correlation can miss important structure

A single coefficient compresses a relationship into one number. A near-zero Pearson correlation can coexist with a strong curved relationship, distinct subgroups, heteroscedasticity, or dependence concentrated in the tails.

Scatterplots, marginal distributions, and subject-matter checks should accompany important matrix entries. Analysts should also distinguish marginal correlation from conditional association, within-person correlation, between-person correlation, serial correlation, and correlations induced by shared denominators or repeated observations.

Missing data can produce an incoherent matrix

With listwise deletion, every coefficient is estimated from the same complete-case sample, but the sample may be smaller and selected. With pairwise deletion, each coefficient may use a different subset of observations, increasing apparent coverage while potentially producing a matrix that is not positive semidefinite.

The reporting should state the missing-data rule and, where pairwise estimation is used, the sample size underlying each coefficient. Multiple imputation or model-based estimation can preserve more information under stated missingness assumptions, but correlations must be pooled or reconstructed using a method appropriate to the analysis rather than averaged mechanically.

Sampling uncertainty and multiple comparisons

Sample correlations are estimates and can vary substantially, especially with small samples or weak overlap. A matrix containing many variables also contains many pairwise coefficients, so isolated large values may arise by chance.

For Pearson correlation under suitable assumptions, Fisher's transformation is:

$$z=\operatorname{arctanh}(r) =\frac{1}{2}\log\left(\frac{1+r}{1-r}\right)$$

Its approximate standard error is:

$$SE(z)=\frac{1}{\sqrt{n-3}}$$

Intervals can be formed on the $z$ scale and transformed back with the hyperbolic tangent. These formulas rely on assumptions that may be poor for clustered data, heavy tails, complex survey designs, or data-dependent variable selection, where resampling or model-based methods may be more appropriate.

Checking whether a matrix is valid

A proposed matrix may contain individually plausible coefficients that cannot coexist jointly. This commonly occurs when correlations are assembled from different studies, rounded aggressively, edited manually, or elicited one pair at a time.

• Symmetry check: The matrix must satisfy $R=R^{\mathsf T}$.
• Diagonal check: Every diagonal entry must equal one for a correlation matrix.
• Bounds check: Every entry must lie in the interval $[-1,1]$.
• Eigenvalue check: The smallest eigenvalue must be non-negative within a declared numerical tolerance.
• Decomposition check: Cholesky decomposition should succeed when positive definiteness, rather than semidefiniteness, is required.
• Provenance check: Every coefficient should have a consistent population, definition, time point, estimator, and evidence source.

Replacing an invalid matrix with a nearest valid correlation matrix can be a practical repair, but it changes the supplied relationships. The repair method, distance criterion, altered entries, and decision impact should be documented and subjected to sensitivity analysis rather than concealed.

Correlated simulation using a matrix decomposition

Probabilistic sensitivity analysis often requires uncertain parameters to move together. For a positive-definite correlation matrix $R$, Cholesky decomposition provides a lower-triangular matrix $L$ such that:

$$R=LL^{\mathsf T}$$

If $Z$ is a vector of independent standard normal draws, then:

$$X=LZ$$

has correlation matrix $R$. To simulate non-normal marginal distributions, analysts may transform correlated uniform or normal variables through a copula or use an appropriate joint distribution; matching a rank or linear correlation after non-linear transformation is not automatic and should be verified empirically.

How correlation affects health-economic models

Costs, utilities, event risks, treatment effects, and model parameters often share data or causal drivers. Ignoring material dependence can distort the uncertainty of incremental cost, incremental effectiveness, net benefit, and the probability that an intervention is cost-effective.

For incremental cost $\Delta C$ and incremental effect $\Delta E$, the variance of incremental net monetary benefit at threshold $\lambda$ is:

$$\operatorname{Var}(INMB)= \lambda^2\operatorname{Var}(\Delta E) +\operatorname{Var}(\Delta C) -2\lambda\operatorname{Cov}(\Delta E,\Delta C)$$

The covariance term can either increase or decrease decision uncertainty depending on its sign. Correlations should therefore be estimated or elicited for the correct quantities—such as joint treatment-effect estimates or patient-level costs and QALYs—rather than borrowed because two parameters merely appear related.

Correlation matrices in regression and evidence synthesis

Highly correlated predictors can make regression coefficients unstable because several variables contain similar information. This multicollinearity affects precision and interpretation even when prediction remains adequate.

Evidence synthesis can also require within-study correlations among multiple outcomes, treatment contrasts, or time points. Assuming independence when estimates share participants or a common comparator can give incorrect standard errors, while imposing a correlation without evidence can create false precision; uncertainty about the correlation may itself require sensitivity analysis.

Spreadsheet implementation and checks

A spreadsheet can calculate a sample matrix and make its symmetry visible, but formula choice and data alignment require care. Blank cells, text, filtered observations, and inconsistent ranges can cause different pairs to use different records without an obvious warning.

• Use aligned ranges so every intended pair refers to the correct observations and time points.
• Report pairwise sample sizes when missingness means that coefficients use different records.
• Keep full precision in stored coefficients and apply display rounding separately.
• Check symmetry and diagonals with explicit audit formulas rather than visual inspection alone.
• Validate eigenvalues externally or with a tested numerical routine before using a manually assembled matrix for simulation.
• Compare simulated correlations with targets using a large diagnostic sample before accepting a correlation-generating procedure.

Excel's CORREL function calculates Pearson correlation for two aligned ranges, but it does not validate the complete matrix or its suitability for Cholesky decomposition. A polished matrix display is therefore not evidence that the joint dependence structure is mathematically valid.

A practical workflow

A defensible matrix begins with a clear purpose and consistent variable definitions. The following workflow connects estimation, validation, modelling, and reporting.

1. Define the variables, units, population, time points, and intended use of the matrix.
2. Inspect distributions, scatterplots, outliers, missingness, repeated observations, and subgroup structure.
3. Choose a correlation measure that matches the variable types and dependence being represented.
4. Align records and apply a prespecified missing-data method consistently.
5. Estimate coefficients with pairwise sample sizes and uncertainty where decision-relevant.
6. Assemble the matrix with identical row and column order, full precision, and traceable labels.
7. Validate symmetry, unit diagonal, coefficient bounds, eigenvalues, and decomposition requirements.
8. Diagnose influential observations, non-linearity, poor overlap, multicollinearity, and sensitivity to modelling choices.
9. Test downstream calculations or simulations against the intended covariance and decision outputs.
10. Document the estimator, data source, missingness, uncertainty, repairs, software, and limitations.

Interpreting and reporting the matrix

A correlation matrix is a compact description of pairwise dependence for specified variables in a specified dataset or model. Its entries should be interpreted with their sampling uncertainty, variable definitions, population, timing, missing-data handling, and coefficient type.

Common errors include treating correlation as causation, interpreting zero Pearson correlation as independence, mixing incompatible coefficients, using different samples without disclosure, rounding an almost singular matrix, and forcing an invalid matrix into a simulation. Transparent validation and provenance are essential whenever the matrix affects uncertainty or a decision result.