Why a Boundary Changes the Test
A likelihood-ratio test compares how well two nested models fit the same observations. In a QSP model, the more extensive mechanism might add a nonnegative pathway rate \(k\). Setting \(k=0\) removes that pathway, so the null model is contained in the alternative. But \(k=0\) is not an interior point of the allowed parameter space \(k\geq 0\): estimates cannot move below zero. This constraint can change the null distribution of the test statistic, even when the models are nested.
That matters because the usual chi-square reference distribution for a likelihood-ratio statistic depends on regularity conditions, including that the null parameter value is in the interior of the parameter space. Using that reference distribution at a boundary can yield an incorrect p-value. The correct response is not to abandon likelihood comparison, but to check the conditions and choose a reference distribution that is justified for the specific mechanistic hypothesis.
In regular problems, \(T\) is often compared with a chi-square distribution whose degrees of freedom equal the number of restrictions imposed by the null. For this result, the models must be correctly specified for the likelihood calculation; the null must be nested in the alternative; the relevant parameters must be identifiable; and the usual smoothness and information conditions must hold. A parameter on a boundary violates the interior-point condition, so the usual result cannot simply be assumed.
A Regular One-Parameter Boundary
There is a standard limiting result for a particular, relatively simple boundary case. Suppose the alternative adds one scalar parameter \(k\geq 0\), the null fixes \(k=0\), and all nuisance parameters remain identifiable under the null. Assume the likelihood is smooth near the null, the information for the added parameter is positive, and the other regularity conditions for likelihood asymptotics hold. Then, under the null and for large samples, the likelihood-ratio statistic has a mixture distribution: probability one-half at zero and probability one-half from a chi-square distribution with one degree of freedom.
The mixture arises because random sampling variation can push the estimate toward either side of zero, but the constraint prevents a negative estimate. When the unconstrained estimate would be negative, the constrained estimate is exactly zero and the likelihood-ratio statistic is zero. When it would be positive, the statistic has the usual one-degree-of-freedom quadratic form.
A short local derivation makes this projection effect explicit. After accounting for identifiable nuisance parameters, write the local log-likelihood as a quadratic in the added parameter’s displacement \(h\geq 0\). Let \(S\) be the efficient score and \(I>0\) its information. Under the null, the score is approximately normal with mean zero and variance \(I\).
Maximizing this quadratic without the constraint gives \(h=S/I\); with the constraint, it gives \(\hat{h}=\max(S,0)/I\). Twice the maximized log-likelihood increase is therefore
Since a standard normal variable is nonpositive with probability one-half, \(T=0\) with probability one-half. Conditional on \(W>0\), squaring \(W\) has the same distribution as squaring a standard normal variable, which is \(\chi^2_1\). This yields the stated mixture. The argument depends on the regularity conditions: it is not a universal rule for every model in which something is absent.
Worked Example: Testing a Zero Pathway Rate
Suppose a hypothetical QSP analysis compares a null model with \(k=0\) against a nested alternative with \(k\geq 0\). Assume the pathway rate is the only added parameter, nuisance parameters are identifiable under the null, and the other regular boundary conditions are reasonable. The fitted log-likelihoods are \(\ell_0=-120.00\) and \(\ell_1=-118.645\).
The observed statistic is
For a one-degree-of-freedom chi-square variable, the upper-tail probability at \(2.710\) is approximately \(0.100\). The boundary-mixture p-value is half of that:
At significance level \(\alpha=0.05\), this result is at the decision threshold, subject to rounding and the asymptotic approximation. It is evidence against the zero-pathway null at about the 5% level under the stated model and assumptions. The p-value is not the probability that the pathway is absent, nor the probability that the null is true.
When Simulation Is Needed
The half-\(\chi^2_1\) mixture is justified only for its specific regular case. It may fail when the null makes another parameter unidentifiable, when several parameters are on boundaries, or when the likelihood has a more complicated local shape. An absent cell population is a useful example: if its abundance is zero, a growth or clearance rate that only governs that population may have no effect on the observations. That rate cannot be estimated under the null. The standard boundary derivation assumes the nuisance parameters remain identifiable, so neither the ordinary chi-square reference nor the simple half-\(\chi^2_1\) mixture is automatically valid.
A parametric bootstrap provides a practical calibration for such cases. It approximates the null distribution of \(T\) by repeatedly generating data under the fitted null model and refitting both the null and alternative to each simulated data set. The simulated data must preserve the study design, observation schedule, error model, and any other features included in the likelihood. The statistic in each replicate must be calculated using the same fitting and constraints as for the observed data.
Estimate its parameters from the observed data and retain the fitted null as the generating model for the bootstrap replicates.
Generate \(B\) data sets using the fitted null parameters and the original design and observation model.
For each simulated data set, maximize the null and alternative likelihoods using the same parameter bounds and fitting safeguards as for the observed data.
Calculate \(T_b=2(\ell_{1,b}-\ell_{0,b})\) for each replicate and count how many simulated statistics are at least as large as the observed \(T\).
Use \((r+1)/(B+1)\), where \(r\) is the number of simulated statistics satisfying \(T_b\geq T_{\mathrm{obs}}\). Report the number of replicates and recognize the simulation error in this estimate.
The added one in the numerator and denominator avoids reporting a zero p-value merely because no finite set of simulations exceeded the observed statistic. The resulting value is a Monte Carlo estimate, not an exact p-value. It also uses a plug-in null: the null parameters are estimated from the observed data, so the calibration inherits uncertainty and possible misspecification in that fitted model.
Worked Example: Bootstrap Calibration for a Boundary Test
In a hypothetical study, a pathway-rate test gives \(T_{\mathrm{obs}}=4.600\). Because the fitted null makes a nuisance rate weakly identified, the analyst does not rely on the regular one-parameter mixture. The analyst fits the null, simulates \(B=1{,}999\) data sets under that fitted null while retaining the original observation schedule and error model, and refits both models to each replicate. In \(r=79\) replicates, \(T_b\geq 4.600\).
The bootstrap estimate is
At \(\alpha=0.05\), this estimated p-value is below the threshold, so the observed statistic is relatively unusual under the fitted null. The Monte Carlo standard error is approximately
This standard error describes simulation variability, not uncertainty about the biological mechanism. Since the estimate is near the decision threshold, more replicates could clarify the Monte Carlo estimate. The conclusion also depends on the adequacy of the fitted null and on successful optimization in the observed and simulated data sets.
Worked Example: An Absent Population with an Unidentified Rate
Consider a hypothetical model in which a cell population has abundance \(A\geq 0\). When \(A>0\), its dynamics depend on a rate \(\lambda\). Under the null \(A=0\), changing \(\lambda\) does not change the predicted observations. Thus \(\lambda\) is not identifiable under the null, and the regular one-parameter boundary result does not apply.
Suppose fitting the null and alternative gives \(T_{\mathrm{obs}}=5.200\). A naive \(\chi^2_1\) reference gives an upper-tail probability of approximately \(0.0226\). Applying the half-\(\chi^2_1\) formula would give approximately \(0.0113\), but that formula is also unjustified here because \(\lambda\) is unidentified under the null.
Instead, suppose 999 null simulations are generated using the fitted null and the same observation and fitting procedures. If 91 simulated statistics are at least \(5.200\), then
The simulation-calibrated result is not significant at \(\alpha=0.05\). The substantial difference from the naive chi-square value illustrates why a reference distribution should not be selected just from the number of added parameters when the null changes identifiability.
Diagnostics, Alternatives, and Limitations
Simulation calibrates the statistic only if the simulation and refitting steps represent the test actually being performed. Check whether the null and alternative are genuinely nested, whether parameter constraints are implemented consistently, and whether each optimizer reaches a comparable maximum. In mechanistic models, multiple starting values can be important: a failed alternative fit can make \(T\) too small, while a failed null fit can make it too large. Inspect convergence and boundary estimates across the observed fit and replicates, and investigate failures rather than silently discarding difficult replicates.
The main choices can be summarized as follows. The standard chi-square reference is appropriate for regular interior-point tests. The half-\(\chi^2_1\) mixture is a sharper choice for the regular case of one added nonnegative parameter at zero with identifiable nuisance parameters. Parametric bootstrap calibration is useful when the reference distribution is not known or the regularity conditions fail, but it depends on the fitted null, the data-generating assumptions, and adequate simulation and optimization.
A finite simulation also limits p-value resolution: with \(B\) replicates, the smallest value from the plus-one formula is \(1/(B+1)\). When the estimated p-value is close to \(\alpha\), increase \(B\) and report the simulation count so that the precision is transparent. More simulations cannot repair a misspecified null model or an inappropriate observation model. Simulation-based calibration is therefore a way to approximate the null distribution under stated assumptions, not a guarantee that those assumptions are correct.
Common mistake. “One parameter was added, so I should compare the statistic with \(\chi^2_1\).” The correction is to inspect both the parameter space and identifiability under the null. If the added parameter is a regular nonnegative scalar fixed at zero and other conditions hold, use the half-\(\chi^2_1\) mixture. If the null also makes parameters unidentified or otherwise violates those conditions, calibrate the test under the fitted null by simulation, and qualify the result by the simulation and model assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A nested test adds one nonnegative rate fixed to zero under the null. What conditions are needed before using the half-\(\chi^2_1\) reference?
- For \(T=3.84\), suppose \(P(\chi^2_1\geq 3.84)\approx 0.050\). What is the regular boundary-mixture p-value?
- In a null simulation with \(B=999\), 24 replicates have \(T_b\geq T_{\mathrm{obs}}\). Calculate the plus-one p-value.
- Why can an absent-population hypothesis invalidate the simple half-\(\chi^2_1\) result?
Solutions. (1) The null must be nested; the added parameter must be a single scalar constrained to be nonnegative and fixed at zero; nuisance parameters must remain identifiable under the null; and the likelihood must satisfy the relevant smoothness and information conditions. (2) The mixture p-value is \(\tfrac12(0.050)=0.025\). (3) The estimate is \((24+1)/(999+1)=25/1{,}000=0.0250\). (4) A rate governing the absent population may have no effect on observations when its abundance is zero, so that rate is unidentified under the null; the regular mixture derivation assumes identifiability.