Why Boundary Estimates Need Special Attention
A selection-model likelihood can favor a parameter value at the edge of the values permitted by the model. This can happen when a selection parameter is deliberately restricted to a sensitivity range, when a probability reaches zero or one, or when a variance parameter is estimated as zero. It can also appear when an optimizer has not actually found the maximum. These situations can produce the same printed estimate—an endpoint—but they call for different conclusions.
A boundary estimate is therefore a diagnosis to investigate, not a conclusion about the underlying science. The relevant questions are whether the endpoint belongs to the specified parameter space, whether the likelihood truly maximizes there, whether the result is stable to numerical checks, and how the boundary affects uncertainty calculations. In a selection model, these checks matter especially because the observed-data likelihood may be weakly informative about how the outcome distribution and response mechanism combine.
The definition concerns the parameter space actually used in the model and analysis. For example, if a response probability \(q\) is allowed to range over \([0,1]\), then zero and one are boundary values. A logit coefficient \(\gamma\in\mathbb{R}\), by contrast, has no finite boundary point: probabilities approaching zero or one correspond to \(\gamma\) tending to negative or positive infinity. Such a limiting estimate should be reported as an infinite or diverging estimate, not silently described as a finite boundary estimate.
Selection-model parameters may also be bounded by an explicit sensitivity restriction, such as \(\gamma\in[L,U]\). An estimate at \(L\) or \(U\) is then a boundary estimate for that restricted analysis. It does not establish that the unrestricted likelihood has a finite maximum there; the endpoint may simply be the limit of the range the analyst chose to examine.
How to Check Whether an Endpoint Is a Maximum
For a scalar parameter \(\theta\), hold nuisance parameters fixed or profile them out. The profile log likelihood is \(\ell_p(\theta)=\max_{\lambda}\ell(\theta,\lambda)\), where \(\lambda\) denotes nuisance parameters and the maximization uses their permitted values. A boundary estimate is plausible when the profile likelihood is largest at an endpoint and moves downward as one moves from that endpoint into the parameter space.
The directional-derivative condition makes this precise. Suppose \(\theta\in[a,b]\), the log likelihood is differentiable near \(b\), and \(b\) is a local maximum over the permitted interval. A feasible movement from \(b\) has the form \(b+td\), with \(t>0\) small and \(d<0\). At a local maximum, the directional derivative cannot be positive:
At the lower endpoint \(a\), feasible directions have \(d>0\), so the corresponding condition is \(\ell'(a)\leq 0\). These are necessary local conditions, not proof of a global maximum. A likelihood can be irregular or have more than one local maximum; compare the endpoint with interior values and examine the profile across the permitted range.
For a multivariate parameter, the same logic applies to every feasible direction: at a constrained local maximum, the directional derivative of the log likelihood must be nonpositive for each direction that remains inside the parameter space. With active constraints, the ordinary unconstrained score need not equal zero. Therefore, treating “score equals zero” as a universal test can wrongly label a valid endpoint maximum as a failed fit—or can overlook a failure to optimize over feasible directions.
Worked Example: A Profile Likelihood Maximized at a Sensitivity Limit
Consider a hypothetical analysis of 180 patients, 54 of whom have missing outcomes. The analyst examines a bounded selection sensitivity parameter \(\theta\in[0,1]\). After maximizing over the other parameters, suppose the observed-data profile log likelihood is \(\ell_p(\theta)=-4(\theta-1.2)^2\), up to an additive constant. The interval is an explicit sensitivity restriction, not an inherent bound on every selection coefficient.
Differentiate the profile: \(\ell_p'(\theta)=-8(\theta-1.2)\). For \(0\leq\theta\leq1\), this derivative is positive because \(\theta-1.2<0\). Thus the profile increases throughout the permitted interval, and its constrained maximum is at \(\hat{\theta}=1\). The endpoint condition is also satisfied: \(\ell_p'(1)=1.6>0\), so moving left from 1 decreases the log likelihood to first order.
For a numerical comparison, \(\ell_p(1)=-4(1-1.2)^2=-0.16\), while \(\ell_p(0.9)=-4(0.9-1.2)^2=-0.36\). The log-likelihood difference is \(-0.16-(-0.36)=0.20\), so twice that difference is \(0.40\). This supports a maximum at the upper limit within the stated range. It does not show that the unrestricted selection parameter has a finite maximum at 1, nor that the imposed sensitivity range is scientifically adequate.
Distinguishing an Endpoint Maximum from an Optimization Failure
A numerical optimizer may stop near an endpoint because of a poor step, a convergence criterion, or a parameter transformation—even when the likelihood improves at a feasible point farther inside. Conversely, an optimizer may report a small gradient after transforming a constrained parameter, although the original-scale likelihood still has a feasible direction of improvement. Review the likelihood on the original parameter scale as well as the optimizer’s status.
Useful checks include evaluating the log likelihood at the reported estimate and at nearby feasible values, fitting from more than one starting value, checking whether nuisance parameters were also optimized, and comparing gradients or one-sided profile slopes on the original scale. These checks do not replace careful numerical analysis, but they can reveal a false endpoint diagnosis. Starting-value selection and the behavior of multiple local maxima require separate attention; an endpoint check alone cannot rule them out.
Worked Example: An Optimizer Stops Near an Endpoint, but the Maximum Is Interior
Suppose a hypothetical selection-model fit uses a parameter \(\theta\in[0,1]\), and its profiled log likelihood is \(\ell_p(\theta)=-3(\theta-0.35)^2\). An optimizer reports \(\theta=0.99\). Check whether this is a boundary maximum by comparing it with feasible values.
The derivative is \(\ell_p'(\theta)=-6(\theta-0.35)\), which is negative at 0.99. Moving left therefore increases the profile. Direct calculation gives \(\ell_p(0.99)=-3(0.64)^2=-1.2288\) and \(\ell_p(0.35)=-3(0)^2=0\). The interior value is higher by \(1.2288\) log-likelihood units, so the reported estimate is not a maximum.
The endpoint comparison leads to the same conclusion: \(\ell_p(1)=-3(0.65)^2=-1.2675\), which is even lower than the value at 0.99. The fact that the optimizer returned a value near 1 does not validate it. The profile has an interior maximum at \(\theta=0.35\); the fit should be investigated and re-optimized rather than reported as a genuine boundary estimate.
A profile can also look nearly flat near its largest value. In that case, a tiny numerical difference may determine which parameter is printed, without providing meaningful evidence that one endpoint is preferred. Inspect the scale of the likelihood differences, the numerical precision, and the stability of the profile. A plateau, a discontinuous profile, or implausibly different results from equivalent evaluations can point to numerical or identification problems rather than a decisive boundary maximum.
Consequences for Likelihood-Based Inference
At a boundary, the usual interior approximation based on a locally quadratic likelihood can be unreliable. For example, a variance parameter is restricted to be nonnegative, so a fitted value of zero has no feasible values below it. A symmetric standard error or Wald interval centered at zero can extend into impossible negative values. More generally, the likelihood may be asymmetric, sharply curved on one side and nearly flat on the other, or poorly approximated by a quadratic near the endpoint.
Profile likelihood can make the one-sided shape visible. But an interval still requires an appropriate likelihood-ratio cutoff and the assumptions that justify it. When a null value or a fitted value lies on a boundary, the regular interior likelihood-ratio reference distribution need not apply. The appropriate calibration depends on the model, the constraints, and the hypotheses. The earlier material on Nonstandard Likelihood-Ratio Calibration addresses this calibration issue; finding a boundary estimate is a reason to check that issue, not a justification for choosing a familiar cutoff automatically.
Worked Example: A Variance Estimate at Zero and an Asymmetric Profile
Suppose a hypothetical longitudinal selection model includes a nonnegative variance parameter \(\omega\geq0\). For a fictional data set of 240 patients, 72 with missing outcomes, assume the profiled log likelihood is \(\ell_p(\omega)=-10(\omega+0.10)^2\), up to an additive constant. This example illustrates the consequences of a boundary estimate; it does not claim that the profile form is typical of every selection model.
The unconstrained quadratic has its maximum at \(\omega=-0.10\), which is not permitted. On the permitted range, \(\ell_p'(\omega)=-20(\omega+0.10)<0\), so the profile decreases as \(\omega\) increases from zero. Its constrained maximum is therefore \(\hat{\omega}=0\). For instance, \(\ell_p(0)=-10(0.10)^2=-0.10\), while \(\ell_p(0.10)=-10(0.20)^2=-0.40\).
The result says that the maximum under the specified model and constraints occurs at zero; it does not prove that the true variance is exactly zero. The profile is asymmetric at the boundary, and negative values are infeasible. A symmetric Wald interval based on an ordinary interior approximation would not reflect this geometry. Inference should use a method and calibration appropriate to the constrained parameter and the model rather than interpreting the endpoint estimate as a conventional interior estimate.
Common Pitfalls, Comparisons, and Limits
A boundary estimate is not automatically evidence of strong selection, no selection, or a scientifically meaningful extreme. Its meaning depends on the parameterization and constraints. In particular, an endpoint imposed for sensitivity analysis says what the likelihood prefers within that range; it does not establish what would happen outside it. And if the likelihood continues to improve as an unconstrained coefficient grows without bound, the issue is a limiting or non-existent finite maximum, not a finite estimate on the boundary of \(\mathbb{R}\).
The main alternatives answer related but distinct questions. An unconstrained fit asks where the likelihood is maximized over the full permitted parameter space; a constrained fit may instead answer where it is maximized under a scientific or sensitivity restriction. A Wald analysis summarizes local curvature at an estimate and can be particularly fragile at a boundary. Profile-likelihood analysis displays how fit changes across values of the target parameter, but still needs justified inferential calibration. Numerical optimizer diagnostics assess whether the stated maximum was found; they do not determine whether the model or constraint is scientifically appropriate.
A practical diagnosis should report the parameterization, the imposed bounds, whether the endpoint is finite or a limit at infinity, and evidence that the likelihood is maximized there. Include relevant profile behavior and numerical checks. If the profile is flat, results may be sensitive to small computational changes; if it is sharply sloped at the allowed endpoint, conclusions may depend heavily on the chosen restriction. Neither pattern alone resolves weak identification or makes the selection assumptions empirically verifiable.
Identify the parameterization and whether the endpoint is a model constraint, a sensitivity limit, or merely an apparent limiting value.
Compare the reported endpoint with nearby feasible values and, where available, check the one-sided profile slope or directional score.
Verify likelihood evaluations and nuisance-parameter optimization, and investigate convergence warnings or near-flat profiles.
Do not assume interior Wald or likelihood-ratio approximations apply; use a justified method and state the relevant constraints.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For \(\theta\in[0,2]\), a differentiable log likelihood has \(\ell'(2)=0.7\). What does this derivative imply about small feasible movements from the upper endpoint?
- An unconstrained logit coefficient \(\gamma\in\mathbb{R}\) grows larger on every optimization iteration and the fitted response probability approaches one. Is this a finite boundary estimate?
- A constrained optimizer reports \(\theta=0.98\) for \(\theta\in[0,1]\), but the profile log likelihood at 0.80 is higher. What should be concluded?
- Why may an ordinary symmetric Wald interval be misleading for a nonnegative variance parameter estimated as zero?
Solutions. For question 1, feasible movements from 2 have \(d<0\). The directional derivative is \(0.7d<0\), so sufficiently small movements left decrease the log likelihood to first order, consistent with a local maximum at the endpoint; a global check is still needed. For question 2, no. The coefficient space is the real line and has no finite boundary point. The sequence may indicate a limiting estimate at infinity, separation, or other instability, which must be diagnosed rather than reported as a finite boundary estimate. For question 3, the reported value is not a maximum over the permitted space because a feasible value has higher likelihood. Recheck the optimization and do not report 0.98 as a genuine boundary maximum. For question 4, zero is an endpoint and negative variance values are infeasible; the likelihood can be asymmetric there, so an interior quadratic approximation and symmetric interval may misrepresent both the parameter space and likelihood shape.