Why Zero Frailty Variance Is a Special Null Hypothesis
A frailty model represents unobserved heterogeneity by adding a latent random effect to the event process. For a shared frailty model, clusters have a common latent multiplier \(W\), conventionally normalized to have mean 1. Its variance, denoted here by \(\theta\), measures the degree of heterogeneity: \(\theta=0\) means \(W=1\) with probability 1, so there is no residual frailty variation. Testing \(H_0:\theta=0\) therefore asks whether the data support a positive frailty variance beyond the fitted no-frailty model.
The key difficulty is not the construction of a likelihood-ratio statistic. It is the parameter space. A variance cannot be negative, so the admissible values are \(\theta\geq 0\). Under \(H_0\), the parameter is at the edge of this space, not in its interior. The usual likelihood-ratio result for an interior, unrestricted parameter does not apply directly.
The notation \(\widehat{\eta}\) and \(\widehat{\eta}_0\) allows nuisance parameters, such as regression coefficients and baseline-hazard parameters, to be re-estimated in each model. Comparing the alternative with a null model in which those parameters are held fixed at their alternative estimates would generally not produce the likelihood-ratio test defined above.
Assumptions Behind the Boundary Reference Distribution
The half-and-half mixture is an asymptotic result, not a universal correction for every frailty comparison. Its usual form applies when the null and alternative are nested; \(\theta\) is the single constrained parameter of interest; the null value is exactly the boundary \(\theta=0\); and the nuisance parameters remain identifiable and are estimated regularly under the null. The independent cluster contributions should be described by the fitted likelihood, with sufficient information as the number of independent clusters grows. A finite, nonzero efficient information for \(\theta\) is also required.
In a standard shared frailty model, the no-frailty model is recovered at \(\theta=0\), with the remaining event-process parameters still estimable. The frailty distribution must be parameterized consistently so that its mean is fixed at 1 and its variance is the parameter being tested. Otherwise a change in frailty scale may be confounded with a change in the baseline hazard. The earlier material on identifiability of frailty variance and baseline hazard explains why this separation matters.
These conditions need scrutiny. If an additional parameter exists only when frailty is present, so that it is unidentified under \(H_0\), the simple half-\(\chi^2_1\) mixture may not hold. Nor does the result automatically apply to a test of several variance components, a misspecified likelihood, or data with too few independent clusters for an asymptotic approximation to be reliable. In such cases, the null distribution may require a more specialized derivation or simulation-based calibration.
Deriving the Mixture Distribution
The core argument is a local likelihood approximation after nuisance parameters have been profiled out. Let \(U\) be the efficient score for \(\theta\) at zero and \(I\) its efficient information. Under the null and regularity conditions, \(U/\sqrt I\) is approximately standard normal. For small nonnegative \(\theta\), the profile log likelihood has the quadratic approximation
If there were no constraint, the maximizer of this quadratic would be \(U/I\). With the required constraint, its maximizer is \(\widehat{\theta}=\max(0,U/I)\). Substituting that maximizer into twice the log-likelihood difference gives
The derivation shows exactly where the boundary enters. When \(Z\leq 0\), the likelihood would locally improve by moving to a negative variance, which is forbidden; the constrained maximum remains at zero and \(T=0\). When \(Z>0\), the squared positive score contributes. A standard normal variable is nonpositive with probability one-half and positive with probability one-half. Conditional on being positive, its square has the \(\chi^2_1\) distribution. Therefore \(T\) has probability one-half at zero and probability one-half distributed as \(\chi^2_1\).
For an observed statistic \(t>0\), the asymptotic boundary-adjusted p-value is
At \(t=0\), the inclusive upper-tail probability is 1 because the mixture places an atom at zero. For positive \(t\), the usual \(\chi^2_1\) p-value is twice the boundary-adjusted value. Thus a conventional \(\chi^2_1\) reference test is conservative for this particular one-sided boundary problem, not a substitute that gives the same p-value.
Worked Example: A Positive Variance Estimate Near the 5% Threshold
Suppose a hypothetical recurrent-event analysis compares a no-frailty model with a shared frailty model. The null-model maximized log likelihood is \(-154.85\), and the constrained alternative-model maximum is \(-153.40\), attained at a positive frailty variance. Assume the models are nested and the boundary conditions above are reasonable.
The likelihood-ratio statistic is $$ T=2\{-153.40-(-154.85)\}=2(1.45)=2.90. $$ For one degree of freedom, \(P(\chi^2_1\geq 2.90)\approx 0.08858\). Applying the boundary adjustment gives $$ p=\frac12 P(\chi^2_1\geq 2.90)\approx \frac12(0.08858)=0.04429. $$
At \(\alpha=0.05\), this asymptotic boundary test rejects \(H_0\). The evidence is just sufficient at this significance level to indicate positive frailty variance under the specified model. It does not establish a particular variance magnitude or prove that frailty caused the observed clustering. With the ordinary \(\chi^2_1\) reference, the p-value would be approximately \(0.0886\), leading to a different decision.
Equivalent Critical-Value Rule
For a positive statistic, rejection at level \(\alpha\) requires \(\tfrac12P(\chi^2_1\geq T)\leq\alpha\), or equivalently \(P(\chi^2_1\geq T)\leq 2\alpha\). At \(\alpha=0.05\), the \(\chi^2_1\) upper-tail probability must be at most \(0.10\), whose critical value is approximately 2.706. Thus the mixture test rejects for \(T>2.706\), subject to the convention at the boundary. This is a lower cutoff than the ordinary \(\chi^2_1\) 5% cutoff of approximately 3.841.
Worked Example: A Statistic Below the Boundary-Test Cutoff
For another hypothetical fit, suppose the maximized log likelihood increases by \(0.605\) when the frailty variance is estimated rather than fixed at zero. Both fits re-estimate their nuisance parameters, and the alternative estimate is nonnegative.
The statistic is $$ T=2(0.605)=1.21. $$ Since \(\sqrt{1.21}=1.10\), the \(\chi^2_1\) upper tail is \(P(|Z|\geq 1.10)\approx 0.2713\). The boundary-adjusted p-value is $$ p=\frac12(0.2713)\approx 0.1357. $$
Because \(0.1357>0.05\), the test does not reject \(H_0\) at the 5% level. This means the fitted data do not provide sufficient evidence against zero variance by this test; it does not prove that the true variance is exactly zero. The same decision follows from comparing \(1.21\) with the mixture critical value \(2.706\).
When the Constrained Fit Stays at Zero
A likelihood-ratio statistic should be nonnegative when the models are genuinely nested and both are maximized correctly, since the alternative includes the null fit at \(\theta=0\). The alternative maximum may therefore occur at zero. This is a valid boundary estimate, not by itself evidence of a software failure.
Worked Example: Constrained Maximum at the Boundary
Consider a hypothetical profiled log likelihood, for \(\theta\geq 0\), of $$ \ell_p(\theta)=-80.00-3\theta-2\theta^2. $$ Its derivative is \(\ell_p'(\theta)=-3-4\theta\), which is negative throughout the allowed parameter space. The maximum is therefore at \(\widehat{\theta}=0\), and both the null and alternative maximized log likelihoods equal \(-80.00\).
Thus \(T=2\{-80.00-(-80.00)\}=0\). Using the inclusive upper tail of the mixture distribution, the p-value is 1. There is no likelihood improvement from allowing positive variance. If an unconstrained numerical routine instead reports a negative variance, that value is outside the model and must not be treated as an estimate or used to form the test.
Diagnostics, Alternatives, and Limitations
The boundary test relies on correct fitting as well as correct asymptotic theory. Before interpreting it, check that the null and alternative likelihoods use the same observations, event definitions, censoring treatment, and nuisance-parameter structure. Verify that both optimizations converged and that the alternative has actually been constrained to \(\theta\geq 0\). A reported alternative log likelihood lower than the null maximum is a warning to investigate optimization, parameter constraints, or whether the models are truly nested. Refit from multiple starting values when results are sensitive to initialization.
The main alternatives answer different questions or rely on different approximations:
| Approach | What it does | Main caution |
|---|---|---|
| Ordinary \(\chi^2_1\) likelihood-ratio test | Uses the interior-parameter reference distribution | Does not account for the one-sided variance boundary; it gives a larger p-value here |
| Boundary-mixture likelihood-ratio test | Uses \(\tfrac12\chi^2_0+\tfrac12\chi^2_1\) for a single regular boundary parameter | Requires its regularity conditions and an adequate large-sample approximation |
| Wald test for variance | Compares an estimate with its estimated standard error using a local normal approximation | Can behave poorly near zero, where symmetry and an interior normal approximation are inappropriate |
| Null parametric bootstrap | Simulates data under the fitted no-frailty model and recalculates the statistic | Can help assess finite-sample calibration, but depends on a credible null model and adequate simulation |
The mixture result addresses a specified test of one frailty variance. If multiple variance components are tested jointly, the limiting distribution can involve a different mixture, with weights determined by the constraint geometry and information structure. If a nuisance parameter is unidentified under the null, the distribution can also differ. A bootstrap under the null may be useful for small samples, but it does not repair a misspecified event model or a poorly identified frailty variance.
A common mistake is to say, “The variance estimate is positive, so I can use the ordinary \(\chi^2_1\) p-value,” or to use a symmetric Wald interval/test as if zero were an interior value. The correction is to formulate the parameter space first: \(\theta\geq0\). Then maximize under that constraint and, when the single-boundary regularity conditions hold, use the half-\(\chi^2_1\) upper tail for positive test statistics. A positive estimate alone is not a test result; the likelihood improvement and its appropriate reference distribution determine the evidence.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why is \(H_0:\theta=0\) a boundary test rather than an ordinary interior-parameter test?
- A fitted model gives \(T=3.00\). If \(P(\chi^2_1\geq3.00)\approx0.0833\), calculate the boundary-adjusted p-value.
- At a 5% level, why is the boundary critical value about 2.706 rather than the ordinary \(\chi^2_1\) critical value 3.841?
- Name one circumstance in which the half-\(\chi^2_1\) reference distribution may not be justified.
Solutions. (1) The admissible variance values satisfy \(\theta\geq0\), so zero is an endpoint and negative values are not allowed. (2) \(p=\tfrac12(0.0833)=0.04165\), approximately \(0.0417\). (3) The mixture p-value is half the \(\chi^2_1\) upper tail, so a 5% test requires a \(\chi^2_1\) tail of 10%; its corresponding cutoff is approximately 2.706. (4) For example, the simple mixture may fail if a nuisance parameter is unidentified under the zero-variance null, or if multiple variance components are tested jointly.