Condition Index

Condition Index: Concept Architecture

Orientation and learning roadmap

The page begins with the regression design matrix and why near-dependence among its columns matters. It then derives an index from the matrix's eigenvalues, works through a small numerical example, and shows how variance-decomposition proportions identify the coefficients involved. The final sections explain sensible responses and why a large index alone does not invalidate a health economic analysis.

Inspect the predictors as a system

In a regression model, two or more predictor columns can nearly reproduce one another. Such near-dependence makes it difficult to estimate their separate coefficients precisely, even if the combined prediction remains useful. Health economic regressions can encounter this problem when closely related disease-severity measures, utilization counts, or interaction terms are entered together.

The diagnostic uses the design matrix $X$ containing the predictor columns used in the fitted model, with a specified convention for the intercept and scaling. Near-dependence can involve several columns without any pair having an extreme correlation. A rank-deficient matrix has an exact linear dependency, for which at least one coefficient is not separately identifiable under ordinary least squares.

Calculate indices from a scaled matrix

A common collinearity diagnostic first scales the columns of $X$ to comparable lengths, then examines eigenvalues of the scaled cross-product matrix. Small eigenvalues correspond to combinations of columns containing little independent information. If $\lambda_{\max}$ is the largest eigenvalue and $\lambda_j$ is another eigenvalue, the associated condition index is

$$\kappa_j=\sqrt{\frac{\lambda_{\max}}{\lambda_j}},\qquad\lambda_j>0.$$

The largest index is the condition number of the scaled design matrix under this convention. If an eigenvalue is exactly zero, the corresponding index is unbounded or undefined because the columns are exactly dependent. Different software may report a raw-matrix condition number, a centred-predictor calculation, or an intercept-inclusive scaled calculation; compare results only after checking which matrix and convention were used.

The index is dimensionless after scaling, but it is sensitive to model specification and preprocessing. An unscaled matrix can appear ill-conditioned merely because predictors use different units, such as pounds versus thousands of pounds. Centring can change dependence involving an intercept and polynomial or interaction terms without necessarily eliminating the substantive dependence among predictors.

A two-predictor calculation

Suppose two centred, unit-length predictor columns have correlation $0.998$, so their scaled cross-product matrix is $\begin{pmatrix}1&0.998\\0.998&1\end{pmatrix}$. Its eigenvalues are $1.998$ and $0.002$. The indices are $\sqrt{1.998/1.998}=1$ and $\sqrt{1.998/0.002}=\sqrt{999}\approx31.6$.

The near-zero eigenvalue corresponds to a contrast between the two columns, which the data identify poorly. This numerical example is a deliberately extreme dependence, not a health-data estimate. It demonstrates the calculation; whether the resulting coefficient uncertainty matters depends on the coefficients, study purpose, and other diagnostics.

Locate the affected coefficients

A large index signals a near-dependent component of the design, but it does not identify by itself which coefficient estimates are affected. Variance-decomposition proportions associate each coefficient's estimated variance with the design's eigen-components. High proportions for two or more coefficients on the same high-index component support a diagnosis that those terms share a problematic dependency.

The diagnostic should be read as a table of eigenvalue, index, and proportions rather than reduced to one maximum number. A large component involving the intercept may reflect a different issue from a dependency between two scientific predictors, especially when centring choices differ. Heuristics such as an index around $30$ merit investigation, but they are not universal pass-or-fail cutoffs and should be interpreted with the variance proportions and substantive design.

Variance inflation factors (VIFs) offer a complementary predictor-specific measure. A condition-index table can reveal a dependency involving several predictors, while VIFs summarize inflation relative to a predictor's association with the others. Neither measure detects confounding, causal bias, model misspecification, or poor outcome measurement by itself.

Respond in line with the analysis goal

If separate effects are central, inspect the definitions and data lineage of the implicated predictors. Two columns may encode the same underlying construct, a redundant set of categorical indicators with an intercept, or a total and its components. Correct coding errors and exact redundancies before considering statistical fixes.

For genuine near-dependence, possible choices include collecting information that better separates the predictors, prespecifying one scientifically justified measure, combining related measures into a defensible construct, or using a method such as regularization with clear interpretation. Removing a confounder solely to lower an index can bias the target effect. Penalization can stabilize prediction but changes coefficient estimation and does not manufacture the missing independent information.

A cost prediction model and a causal treatment-effect regression have different priorities. Correlated predictors may still yield acceptable predictions in the observed domain while producing unstable individual coefficients or unreliable extrapolations. State which output matters, inspect its uncertainty and sensitivity, and avoid claiming that a good global fit validates every parameter.

Check implementation and interpretation

Record which observations and columns enter $X$, whether the intercept is included, the scaling and centring convention, and how categorical variables and interactions are encoded. Recalculate diagnostics after any change to the analysis sample or specification. Compare coefficients, standard errors, and predictions across scientifically reasonable formulations rather than treating the index as an isolated quality score.

For a simple reproducibility check, compute the singular values of the same scaled design matrix; the ratio of the largest to each singular value should match the corresponding indices up to numerical precision. Use stable numerical methods such as singular value decomposition instead of directly inverting a nearly singular cross-product matrix. When a near-zero singular value appears, trace its associated linear combination back to the original predictors.

A condition index describes numerical and inferential sensitivity in a chosen regression design. It cannot tell whether a treatment works, whether a price coefficient is causal, or whether a cost model will transport to another population. Report the diagnostic as part of a broader assessment of study design, model specification, and uncertainty.

Evidence used for this explanation

SAS and NIST provide primary technical descriptions of condition indices and their use in regression diagnostics. The Belsley, Kuh, and Welsch work develops the component and variance-decomposition approach. The two-predictor eigenvalue calculation is an original illustration that can be checked directly from the stated matrix.

• SAS/STAT User's Guide, “Collinearity Diagnostics”.
• NIST, “Condition Indices”.
• Belsley DA, Kuh E, and Welsch RE, Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. Wiley (1980).