Why the Reference Distribution Changes the P-Value
A likelihood-ratio statistic measures how much better the fitted alternative model is than the fitted null model. In a regular problem, its large-sample reference distribution is often an ordinary chi-square distribution. But when a null hypothesis places a variance component at zero, that value lies on the boundary of the parameter space. The reference law can then be a mixture: a point mass at zero combined with one or more chi-square distributions.
The likelihood-ratio statistic itself does not tell us which reference law to use. The null model, the boundary being tested, and the applicable asymptotic result determine that law. The earlier material on likelihood-ratio tests for random-effect variances and chi-bar-square laws establishes the relevant forms and weights. Here the task is to turn a specified mixture law into a p-value and interpret the result accurately.
The notation \(\chi^2_d\) here denotes a chi-square random variable with \(d\) degrees of freedom. A point mass at zero is sometimes written as a \(\chi^2_0\) component, but it is not an ordinary continuous chi-square distribution: all of its probability is concentrated at zero. For a positive observed statistic, its upper-tail probability is zero; at the observed value zero, its probability of being at least as large is one.
Calculating a Mixture-Tail Probability
A mixture p-value is a weighted average of component upper-tail probabilities. It is not generally the tail probability of a chi-square distribution whose degrees of freedom equal an average of the component degrees of freedom. The weights must be applied to the tail probabilities themselves.
Write down every component and its weight, including any atom at zero. Confirm that the weights sum to 1.
For each positive-degree chi-square component, find the probability that it is at least the observed statistic.
Multiply each component tail by its mixture weight and sum. If the statistic is positive, the point-mass component contributes zero.
Compare the p-value with the chosen significance level, and describe it as a probability under the null reference distribution—not as a probability that the null hypothesis is true.
The most familiar boundary reference law has weight one-half at zero and weight one-half on a chi-square distribution with one degree of freedom. It is often called the half-at-zero, half-chi-square-one law. For any positive observed statistic, its p-value is one-half the ordinary \(\chi^2_1\) upper-tail probability. This does not mean that every boundary test has this law: more complicated settings can have several positive-degree components and different weights.
Worked Example: The Half-at-Zero, Half-Chi-Square-One Law
Suppose a hypothetical random-intercept variance test has observed likelihood-ratio statistic \(t=2.71\), and the applicable null reference law is one-half point mass at zero and one-half \(\chi^2_1\). Calculate the p-value.
Because \(t>0\), the point mass at zero contributes zero. The \(\chi^2_1\) upper-tail probability at 2.71 is approximately 0.0997. Therefore,
At significance level \(\alpha=0.05\), this p-value is just below the threshold, so the test rejects \(H_0\) at the 5% level, subject to the assumptions supporting this reference law. In context, under the null model and the specified asymptotic mixture, the probability of a statistic at least as large as 2.71 is about 4.99%. The result is not a 4.99% probability that the null hypothesis is true.
For comparison, using the ordinary \(\chi^2_1\) tail would give approximately 0.0997. That is twice the mixture p-value in this example, because it omits the half weight assigned to the point mass at zero.
Mixtures with Several Chi-Square Components
Some boundary tests have a chi-bar-square reference distribution: a weighted combination of chi-square distributions, potentially including a point mass at zero. The weights reflect the boundary geometry and testing setup; they are not chosen to make a result significant. Once the relevant weights and degrees of freedom are established, calculating the p-value follows the same weighted-tail rule.
Worked Example: A Three-Component Mixture
For a hypothetical test, suppose the specified reference distribution places weight 0.25 at zero, weight 0.50 on \(\chi^2_1\), and weight 0.25 on \(\chi^2_2\). The observed statistic is \(t=4.00\). The weights sum to \(0.25+0.50+0.25=1.00\), as required.
The point mass contributes zero because 4.00 is positive. The component tails are approximately \(P(\chi^2_1\ge4.00)=0.0455\) and \(P(\chi^2_2\ge4.00)=0.1353\). Thus,
At \(\alpha=0.05\), the test does not reject \(H_0\), because \(0.0566>0.05\). This conclusion uses the specified mixture weights. The ordinary \(\chi^2_2\) tail, about 0.1353, is not the p-value for this test: it uses only one component and ignores both the weights and the other component.
A useful justification for the weighted-tail calculation is the law of total probability. Imagine first selecting a mixture component \(K\) with probabilities \(P(K=0)=w_0\) and \(P(K=j)=w_j\). Conditional on \(K=j\), the statistic follows the \(j\)th component distribution. For \(t>0\),
The zero term is zero because a point mass at zero cannot produce a value at least as large as a positive \(t\). At \(t=0\), every component—including the point mass—produces a value at least zero, so \(P(T\ge0)=\sum_{j=0}^m w_j=1\). This distinction matters because a mixture with an atom is not continuous everywhere.
Worked Example: A Statistic at the Point Mass
Suppose the reference law is half a point mass at zero and half \(\chi^2_1\), and a hypothetical fitted comparison yields \(t=0\). What is the upper-tail p-value?
By definition, the p-value is \(P(T\ge0\mid H_0)\). Both components place all their probability on values at least zero: the point-mass component equals zero, and the chi-square component is nonnegative. Therefore,
The p-value is 1, not 0.5. The weight 0.5 is the probability of drawing from the point-mass component; it is not the probability that the statistic is at least zero. Since the statistic cannot be negative under this reference law, it is always at least zero.
Assumptions, Diagnostics, and Alternatives
The arithmetic is only as appropriate as the reference distribution. A mixture p-value is justified when the test statistic and null hypothesis match a setting for which that mixture law applies, and when the asymptotic conditions behind the result are adequate. For example, the half-at-zero, half-\(\chi^2_1\) law is not a universal correction for any test involving a random effect. For other boundary configurations, the correct number of components and their weights may differ.
Before reporting a mixture p-value, check that the statistic is the intended likelihood-ratio statistic, calculated from the null and alternative models being compared on the same data and with a consistent likelihood convention. The statistic should be nonnegative apart from possible numerical error. Confirm that the reference weights are nonnegative and sum to one, and that each component's degrees of freedom and weight are the ones appropriate for the test. If a software output reports an ordinary chi-square p-value by default, verify that this is actually the correct reference law rather than assuming it is.
An appreciably negative likelihood-ratio statistic is a warning, not a value to insert uncritically into the mixture-tail formula. It can indicate inconsistent model fits, different data or likelihood definitions, or numerical optimization problems. A tiny negative value due to rounding may sometimes be reset to zero after checking the fits, but a material discrepancy should be investigated. Likewise, a mixture p-value can be poorly calibrated when the asymptotic approximation is unreliable. The earlier material on finite-sample calibration and parametric bootstrap describes simulation-based approaches that can be considered when the asymptotic reference law is in doubt.
| Approach | How the p-value is obtained | Main caution |
|---|---|---|
| Specified mixture reference law | Weight the component upper tails and add them. | Requires the correct null mixture and adequate asymptotic approximation. |
| Ordinary chi-square reference law | Use one chi-square upper tail. | Can be wrong for a boundary test if the null law is a mixture. |
| Parametric bootstrap calibration | Simulate statistics under the fitted null and estimate the upper-tail frequency. | Depends on the null model and simulation procedure; finite simulation introduces Monte Carlo uncertainty. |
These approaches answer the same broad calibration question in different ways, but they are not interchangeable by default. If the mixture law is justified and the asymptotic approximation is adequate, its p-value is a direct calculation. A parametric bootstrap instead approximates a null distribution by simulation and may be useful when finite-sample behavior is a concern. An ordinary chi-square calculation is suitable only when that ordinary law is the appropriate reference.
Common Mistakes and Interpretation
- Using the ordinary chi-square tail automatically. A likelihood-ratio statistic does not guarantee an ordinary chi-square null law. Use the reference distribution justified for the boundary test.
- Dropping the point-mass weight without explanation. For a positive statistic, the atom contributes zero to the upper tail, but the remaining components retain their original weights. Do not renormalize them to sum to one.
- Adding component tails without weighting. The mixture tail is \(\sum_j w_jP(\chi^2_{d_j}\ge t)\), not the unweighted sum or an average of degrees of freedom.
- Calling the p-value the chance that \(H_0\) is true. It is a probability of results at least as extreme as the observed statistic, calculated under \(H_0\) and the assumed reference law.
- Ignoring the atom when \(t=0\). At zero, the upper-tail probability includes the point mass and equals 1 for a nonnegative mixture law.
Check Your Understanding
Answer each question, then compare your work with the solutions below.
- A test has the half-at-zero, half-\(\chi^2_1\) reference law and observed statistic \(t=1.50\). If \(P(\chi^2_1\ge1.50)\approx0.2207\), calculate the mixture p-value and state the decision at \(\alpha=0.05\).
- A mixture has weights 0.20 at zero, 0.50 on \(\chi^2_1\), and 0.30 on \(\chi^2_2\). The component tails at the observed statistic are 0.08 and 0.12, respectively. Calculate the p-value.
- For a nonnegative mixture reference law, what is the upper-tail p-value when the observed statistic is exactly zero? Explain why.
- A colleague uses the \(\chi^2_2\) tail alone for a test whose specified reference law has weights 0.50 on \(\chi^2_1\) and 0.50 on \(\chi^2_2\), plus weight 0.25 at zero. Identify the problem with this reference law before calculating any p-value.
Solutions. (1) The point mass contributes zero because \(t>0\), so \(p=0.5(0.2207)=0.11035\approx0.1104\). Since this exceeds 0.05, do not reject \(H_0\) at the 5% level. (2) The weights sum to 1, and \(p=0.20(0)+0.50(0.08)+0.30(0.12)=0.040+0.036=0.076\). (3) The p-value is \(P(T\ge0)=1\), because every value in a nonnegative reference distribution is at least zero. (4) The proposed weights sum to \(0.50+0.50+0.25=1.25\), not 1. The stated mixture is invalid as written; the weights must be checked and corrected from the applicable test result before a p-value can be calculated.