Statistical Methods

Use a Null-Space Check to Explain Unidentified Coefficients

Define the task

When two coefficient vectors give the same fitted values, a model may be unable to identify their individual components. A null-space basis makes that ambiguity explicit: directions z satisfying Az=0 can be added to a coefficient vector without changing A times that vector. Begin with the actual design matrix, including intercepts, category coding and transformations. A numerical rank diagnosis does not by itself decide which scientific question should be changed.

Carry out the check

A singular-value decomposition can construct an orthonormal basis for the numerical null space. Record the relative threshold used to treat singular values as zero, along with dimensions and column scales. Check both Az near zero and orthonormality of the reported basis. Signs and rotations of basis vectors may vary across implementations while spanning the same space. Examine redundant columns, unsupported interactions or empty categories before imposing a constraint merely to obtain coefficients.

Worked hypothetical example

For hypothetical A with rows (1,1) and (2,2), z=(1,−1)/sqrt(2) lies in its null space. Coefficients (3,0) and (2,1) both yield fitted values (3,6). Only their sum is identified by this matrix. This algebraic example does not establish that any particular real study is unidentified. Keep a record of any recoding or constraint and explain how it changes coefficient interpretation. Near dependence requires a sensitivity assessment rather than a simple exact-zero rule.

Method checked 19 November 2026 against Primary method documentation: null_space.

Related reading

Reading path: Analyze data. Guide 179 of 182. Choose any guide to suit your project.

Statistical Methods

Center Predictors to Make Interaction Coefficients Interpretable

Change the reference point without claiming that centering removes an interaction or alters fitted predictions.

Statistical Methods

Explain How Case Mix Changes External Model Performance

Compare the validation population with the development setting before interpreting a performance difference.

Browse this reading pathTools for this stageGuides and downloads for this stage