Does it require likelihood derivatives?

The general construction does, although specialized implementations can simplify the calculation.

Can an incorrect model pass?

Yes; the two-point example satisfies the tested equality despite not being normal.

AI Glossary / Statistics and probability / Advanced

Information matrix test

Definition

The information matrix test checks a likelihood model by comparing two quantities that agree in expectation under correct specification and suitable regularity conditions: the outer product of the score and the negative Hessian of the log likelihood. A statistically significant discrepancy provides evidence that the specified model does not adequately describe the data-generating process.

What it means

Likelihood assumptions include variability and distributional shape. The information matrix equality checks their implications through likelihood derivatives.
The equality concerns expectations, not exact sample agreement. Sampling variation can produce a discrepancy even when the model is correct. Conversely, some incorrectly specified distributions can satisfy the tested moment restrictions, so agreement does not certify the entire model.

How it works

For each observation, calculate the score vector, the first derivative of its log-likelihood contribution, and its Hessian, the matrix of second derivatives. Form the score's outer product and add the Hessian. Under the relevant null, this matrix has expectation zero.
Average the nonredundant elements and assess their joint discrepancy using an appropriate covariance estimate. Parameter estimation must be accounted for: simply treating fitted parameters as known can give the wrong uncertainty. Under suitable assumptions, a quadratic-form statistic has an asymptotic chi-square reference distribution, with degrees of freedom determined by the independent tested restrictions.
Finite-sample calibration can be poor, and a suitable bootstrap may help, but neither a large-sample formula nor resampling should be applied without checking its conditions.

Three practical examples

An analyst checks whether a parametric demand model's likelihood assumptions are compatible with its fitted data. A result can motivate examining the distribution and variance specification, not only average prediction errors.
A choice-model researcher uses a model-appropriate information matrix diagnostic after maximum-likelihood estimation. Derivative calculations and numerical conditioning are checked before interpreting rejection.
A simulation study compares nominal and actual rejection rates for a proposed implementation. This helps distinguish genuine sensitivity to misspecification from an unreliable approximation in small samples.

Think of it this way

Imagine checking a machine using two gauges that should agree on average when the operating model is right. A persistent disagreement raises a question, but a single small difference may be ordinary measurement variation. Agreement also does not prove that every other part of the machine works correctly.

Worked example

Consider an invented model in which independent observations are assumed normal with unknown mean mu and variance fixed at 1. Only the mean is estimated. Ignoring constants, one observation contributes minus one-half times (x minus mu) squared to the log likelihood.
Differentiating with respect to mu gives the score x minus mu. Differentiating again gives the Hessian minus 1. Therefore, the negative Hessian is always 1, whereas the score's outer product, here just its square, is (x minus mu) squared.
Use eight observations: minus 2, minus 1, minus 1, 0, 0, 1, 1, and 2. Their fitted mean is zero. The squared scores are 4, 1, 1, 0, 0, 1, 1, and 4, which sum to 12. The average squared score is 12/8 = 1.5.
The average negative Hessian is 1. The observed difference is therefore 1.5 minus 1 = 0.5. Equivalently, the individual score-square-plus-Hessian values are 3, 0, 0, minus 1, minus 1, 0, 0, and 3; their average is 0.5.
This calculation shows the diagnostic moment, not a completed significance test. A formal test must standardize its sampling variation correctly, including the effect of estimating mu. This small dataset illustrates the calculation, not reliable asymptotic inference.
Now change the modeling assumption: allow the normal variance to be estimated as well. The maximum-likelihood variance estimate is 12/8 = 1.5, using divisor eight. For the mean-mean information component, the average negative Hessian becomes 1/1.5 = 2/3.
The squared mean-score average becomes 1.5 divided by (1.5 squared), also 2/3. Thus that particular component agrees exactly after estimating the variance. This does not establish that the entire normal model is correct. The expanded parameter vector introduces other information components, and the full diagnostic must reflect them.
For an additional population-level check, suppose observations take values minus 1 and 1 with equal probability. A working normal model with mean zero and fixed variance 1 has negative expected Hessian 1 and expected squared score 1. The equality holds even though this two-point population is not normal. This counterexample demonstrates why the test is a check of particular restrictions rather than a proof of complete distributional correctness.

Illustrative case study

Fictional company Maple Demand fits a likelihood model to operating data. Its analyst compares derivative-based information estimates and investigates whether any discrepancy reflects model inadequacy or numerical problems.
The team checks the variance assumptions, unusual observations, and whether measurements are genuinely independent. It evaluates revised models on additional data and keeps the diagnostic separate from business forecasting performance. A failed specification check can identify a reason to investigate without specifying the correct replacement. This scenario is fictional.

Common misconceptions and mistakes

The test is not the same as White's heteroskedasticity test, despite their association with the same researcher and possible connections in particular models.
The outer product must use individual score contributions. The outer product of the total fitted score can be nearly zero at a likelihood optimum and is not a substitute.
A significant result does not identify a unique cause. Several distributional or structural problems can violate the restrictions.
Failure to reject is not model validation. Low power, limited samples, and alternatives satisfying the same restrictions remain possible.
Singular or nearly singular covariance estimates require attention; blindly inverting them can invalidate the calculation.

Business use

Use the diagnostic alongside residual analysis, predictive checks, and substantive knowledge. Record the likelihood, parameterization, restrictions, covariance calculation, and reference distribution. Assess whether the implementation behaves appropriately at the available sample size. A statistical diagnostic should guide investigation rather than automatically approve or reject a business decision.

Check your understanding

What is the first example's discrepancy?

Answer: The average squared score is 1.5 and the average negative Hessian is 1, giving 0.5.

Why not square the total score?

Answer: Individual contributions contain the information being compared.

Does exact agreement in one component establish correct specification?

Answer: No.

Related terms

Hessian
Maximum likelihood
score vector
Fisher information
model misspecification
sandwich covariance
specification test

Learn next

Study information matrix equality, robust covariance estimation, and bootstrap calibration of specification tests. Compare moment diagnostics with predictive model checks.

Sources

Golden and colleagues, Generalized Information Matrix Tests for Detecting Model Misspecification, 2016, indexed article text: https://www.mdpi.com/2225-1146/4/4/46 . Horowitz, The Bootstrap, discussion of information matrix tests and finite-sample behavior: https://users.ssc.wisc.edu/~behansen/718/Horowitz%20Handbook.pdf . Numerical examples and company scenario are original illustrations.

Previous: Information integration Next: Information model