Why a Variance Estimate Can Be Zero
A longitudinal model may allow each subject’s biomarker trajectory to vary around a population mean trajectory. A random intercept represents subject-to-subject variation in baseline biomarker level; a random slope represents variation in rate of change. The corresponding variance components describe the amount of that variation. Because variances cannot be negative, their parameter space has an edge at zero. When the data do not support detectable variation in a particular random effect, the likelihood can be maximized at that edge.
A zero estimate is therefore not automatically a computational error. It can be a valid constrained estimate. But it does affect inference: a usual likelihood-ratio test calibrated to a chi-square distribution may be wrong when the null value lies on the edge of the parameter space.
For example, consider a Gaussian longitudinal mixed model
Here \(i\) indexes subjects, \(\boldsymbol{\beta}\) describes the population mean trajectory, and \(\mathbf{b}_i\) contains subject-specific random effects. The usual assumptions include independent subjects, mean-zero random effects and residual errors, and independence between \(\mathbf{b}_i\) and \(\boldsymbol{\varepsilon}_i\). Their covariance structures must be specified appropriately for the analysis. In the single-component case considered first, suppose a random intercept \(b_{0i}\) has variance \(\tau^2\geq0\). Testing whether this component is needed means testing \(H_0:\tau^2=0\) against \(H_a:\tau^2>0\), while retaining the same fixed-effects structure in both models.
Why the Likelihood-Ratio Reference Distribution Changes
The likelihood-ratio statistic compares the best fit under the alternative with the best fit under the null. Let \(\ell_1\) and \(\ell_0\) denote the maximized log likelihoods under the alternative and null, respectively. The statistic is \(T=2(\ell_1-\ell_0)\). Under regular conditions for an interior parameter value, a likelihood-ratio statistic often has an asymptotic chi-square distribution. Here, however, \(\tau^2=0\) is at the boundary: estimates can move only in the positive direction from the null. The ordinary interior-parameter argument does not apply.
For one variance component, the asymptotic null distribution is a mixture: it places probability one-half at zero and probability one-half on a chi-square distribution with one degree of freedom. The result requires a single tested variance component, regular identifiable nuisance parameters under the null, a correctly specified model, and enough independent subjects for the large-sample approximation. In particular, it is not a universal reference distribution for every test involving random effects.
A short local derivation shows why the mixture appears. Near the null, the log likelihood is approximately quadratic in the tested parameter. If the parameter were allowed to take either sign, its local unconstrained estimate, when standardized, would behave like \(Z\sim N(0,1)\) under \(H_0\). But the variance parameter must be nonnegative, so the constrained estimate is zero whenever the unconstrained estimate points below zero. The likelihood-ratio statistic is consequently approximated by
A standard normal variable is below zero with probability one-half, giving \(T=0\) with probability one-half. When \(Z>0\), squaring produces the positive part of the chi-square-one distribution. Equivalently, for \(t>0\),
The factor of one-half is essential. If \(T=0\), the upper-tail probability including the point mass is one; if \(T>0\), use the mixture tail formula above. This calculation concerns the likelihood-ratio test, not a general confidence interval or a test of a fixed-effect coefficient.
Worked Example: A Fit Reaches the Zero-Variance Boundary
Suppose 48 hypothetical subjects have biomarker measurements at several scheduled visits. A random-intercept model is fitted with the same population mean trajectory in the null and alternative models. The alternative-model optimizer returns \(\hat{\tau}^2=0\), and both models have maximized log likelihood \(-84.20\). Assume the model fit and optimization are otherwise valid.
Likelihood-ratio statistic. Substituting the two log likelihoods gives
The likelihood has not improved by allowing a nonnegative random-intercept variance. Since the observed statistic is zero, its upper-tail probability under the mixture distribution is \(P(T\geq0)=1\). This is not evidence against \(H_0\). The fitted model assigns zero estimated variance to subject-specific baseline deviations, but limited information or the assumed model structure may also contribute to the boundary estimate.
Calculating and Interpreting the Mixture Test
For a positive statistic, first calculate \(T\) from the two maximized log likelihoods and then use the one-half chi-square-one upper tail. Do not halve a rounded tail probability if the final answer needs additional precision: retain the calculator’s unrounded tail until the final rounding step. The following example also illustrates how the test is reported in the units and context of a biomarker analysis.
Worked Example: Testing a Random-Intercept Variance
In a hypothetical study of 60 subjects, the null model sets the random-intercept variance to zero. The alternative allows it to be nonnegative. The models have the same fixed-effects design, and their maximized log likelihoods are \(\ell_0=-122.50\) and \(\ell_1=-120.80\). Assume independent subjects, an appropriate Gaussian mixed model, and a single tested variance component.
Step 1: Calculate the statistic.
Step 2: Calculate the boundary-adjusted p-value. Using the upper tail of a chi-square distribution with one degree of freedom,
At significance level \(\alpha=0.05\), this p-value is below \(\alpha\), so the likelihood-ratio test provides evidence against zero random-intercept variance under the stated model. It supports subject-to-subject baseline variation in the biomarker beyond what the model’s residual variation accounts for. It does not establish that every subject has a distinctly different baseline, nor does it establish a causal effect.
The most direct interpretation of \(\hat{\tau}^2=0\) is specific to the random effect tested. For a random intercept, it means the fitted model estimates no between-subject variation in baseline levels after accounting for the fixed effects and the rest of the covariance model. It does not mean that biomarker measurements are identical: residual error can remain, and fixed effects can still describe a changing population mean trajectory. For a random slope, a zero slope variance would concern estimated variation in individual rates of change, not variation in baseline levels. A boundary estimate should therefore be described by naming the random effect and the model in which it was estimated.
Worked Example: A Modest Statistic and the Ordinary Chi-Square Test
Suppose a hypothetical longitudinal analysis compares two otherwise identical models and obtains \(\ell_0=-211.50\) and \(\ell_1=-210.60\). The alternative estimates a positive random-intercept variance of \(0.12\) biomarker-units-squared. Assume the single-component boundary-test conditions hold.
Statistic:
Mixture p-value:
At \(\alpha=0.05\), this test does not provide sufficient evidence to reject zero random-intercept variance. That is not the same as establishing that the population variance is zero. If one incorrectly used the ordinary \(\chi_1^2\) upper tail, the p-value would be about \(0.1797\), twice the boundary-adjusted value. The ordinary calibration is not the appropriate test here. The estimated variance is measured in squared biomarker units; its square root, \(\sqrt{0.12}\approx0.346\), is an estimated between-subject standard deviation in biomarker units, though the test result indicates uncertainty about whether this component is supported.
Assumptions, Diagnostics, and Alternatives
The mixture test is useful only when its setup matches the problem. The null and alternative must differ by the one variance component being tested, with the fixed-effects structure held constant. For ML fits, compare maximized likelihoods from those models. REML comparisons are appropriate for covariance structures when the fixed-effects design is the same; do not use REML likelihood comparisons to choose between models with different fixed effects. Subject independence and an adequately specified longitudinal distribution also matter. In small samples, the asymptotic mixture can be imprecise even when the model is otherwise well specified.
Before interpreting a zero estimate, check that the optimizer converged and that the result is not an artifact of poor numerical optimization. Inspect the reported variance estimate and likelihood, rerun with reasonable starting values if needed, and consider whether the assumed mean and residual covariance structures are plausible. A flat likelihood near zero can indicate weak information about the component. Residual patterns, unusually influential subjects, or few independent subjects can also make variance estimation unstable. These checks can reveal model or numerical problems, but they cannot turn a boundary estimate into proof that the population variance is exactly zero.
The one-half mixture should not be applied automatically to tests of several variance components at once. With multiple constraints, or with nuisance parameters that are not identified under the null, the limiting distribution can be more complicated. For example, testing a random-slope structure may involve additional covariance parameters and requires care beyond the single-component result here. In such cases, consult a method appropriate to that particular test. A parametric bootstrap—simulating data under the fitted null model and refitting both models—can be a useful alternative, especially when the asymptotic approximation is doubtful, provided the simulation model and refitting procedure are correctly specified. Ordinary chi-square calibration is simpler, but it ignores the boundary and is not a substitute for the correct reference distribution.
| Approach | When it applies | Main caution |
|---|---|---|
| One-half chi-square mixture | One variance component tested at zero under the stated regularity conditions | Large-sample approximation; not universal for multiple or more complex constraints |
| Ordinary chi-square-one test | Often used for regular interior-parameter tests | Does not account for the variance parameter’s boundary at zero |
| Parametric bootstrap | Potentially useful when sample size is small or the asymptotic reference is uncertain | Depends on a credible null simulation model and successful refitting |
In a joint analysis, a boundary estimate for a biomarker random effect describes the fitted longitudinal variation for that component. It should not be translated into a broad claim that subjects have no biological differences, or that the biomarker is unrelated to event outcomes. Those are different claims requiring their own model-based evidence. Report the component estimated at zero, the model and fitting method, and the boundary-aware test or other inference used.
Check Your Understanding
Answer each question using the single-component boundary-test setting unless stated otherwise.
- A test gives \(T=2.00\). What is the approximate mixture p-value, and what is the conclusion at \(\alpha=0.05\)?
- Why can the estimated random-intercept variance equal zero even though the residual variance is positive?
- Two models being compared have different fixed-effects designs and were fitted by REML. Is their REML likelihood-ratio comparison appropriate for this variance-component test?
- What does a zero random-slope variance estimate say about biomarker trajectories, and what does it not prove?
Solutions. (1) The mixture p-value is \(\frac{1}{2}P(\chi_1^2\geq2.00)\approx\frac{1}{2}(0.1573)=0.07865\approx0.0786\). Since \(0.0786>0.05\), the test does not reject \(H_0\). (2) The random-intercept variance is constrained to be nonnegative, and the likelihood may be maximized at zero when the data do not support additional estimated baseline variation. Residual measurement variation can remain positive because it is a different model component. (3) No. REML likelihood comparisons for covariance structures require the same fixed-effects design; use ML when comparing models with different fixed effects. (4) It says the fitted model estimates no between-subject variation in rates of biomarker change for that random-slope component. It does not prove that every subject’s true rate is identical, nor does it rule out other kinds of heterogeneity.