Parameters Near Bounds and One-Sided Evidence

Why Boundaries Change the Interpretation

QSP parameters often have restricted ranges: a rate constant may be nonnegative, a fraction may lie between zero and one, and a concentration cannot be negative. When a fitted parameter approaches one of these limits, the profile likelihood may provide one-sided rather than two-sided evidence. That shape matters: an endpoint of an interval can be set by the data, by the allowed parameter space, or by an arbitrary numerical constraint.

The key question is not merely whether an estimate is “at the boundary.” It is whether the likelihood profile has enough information to identify a threshold crossing before reaching that boundary. If the profile keeps improving toward an imposed limit, the data may not locate the parameter’s optimum within the searched range. Reporting that limit as a data-supported estimate would overstate what the fit establishes.

Definition: Let \(\theta\) be a parameter of interest, with feasible range \(\Theta\), and let \(\eta\) denote nuisance parameters. The profile log likelihood at \(\theta\) is \(\ell_p(\theta)=\max_{\eta}\ell(\theta,\eta)\), where the maximization is over the allowed nuisance-parameter values. Let \(\hat{\theta}\) be the parameter value at the overall constrained maximum likelihood fit, and define the profile likelihood-ratio statistic \(q(\theta)=2\{\ell(\hat{\theta},\hat{\eta})-\ell_p(\theta)\}\). The profile threshold set at a chosen cutoff \(c\) is the set of feasible values satisfying \(q(\theta)\le c\).

These definitions assume that the likelihood is appropriate for the observations, the bounds are stated explicitly, and the nuisance-parameter optimization is sufficiently reliable at each fixed \(\theta\). The cutoff \(c\) is an input to this discussion, not an automatic guarantee of a particular coverage level. In particular, standard likelihood-ratio calibration can require special care near a boundary.

What One-Sided Evidence Means

Suppose a rate constant \(k\) is constrained by \(k\ge 0\), and the constrained maximum occurs at \(\hat{k}=0\). If the profile likelihood decreases as \(k\) moves above zero, the data favor the boundary value over nearby feasible positive values under the fitted model. If the profile then crosses the chosen threshold at a positive value, the threshold set may have the form \([0,U]\): it gives an upper limit, not a conventional two-sided interval centered on an interior estimate.

This does not, by itself, prove that the true rate is exactly zero. The maximum may occur at zero because the likelihood is relatively flat near the boundary, because the observations do not resolve small positive rates, or because of model misspecification. The profile describes relative likelihood support under the specified model; it does not turn a boundary estimate into a mechanistic claim.

A different pattern occurs when the profile is still improving as the parameter approaches a bound and does not turn back within the feasible range. If it reaches the bound without crossing the selected threshold, then the profile has not supplied a finite threshold endpoint on that side. The bound may truncate the support set. If that bound was imposed only for computation, the apparent endpoint is an imposed-bound artifact, not evidence that the data locate the parameter there.

Interpretation: A threshold crossing inside the feasible parameter range is data-dependent, conditional on the model and chosen cutoff. A threshold set that runs into a computational boundary without a crossing is censored by the search range on that side. The imposed boundary is not a likelihood-based confidence limit.

Worked Examples

Worked Example: An Upper Limit From a Boundary Maximum

In a hypothetical QSP model, \(k\) is a degradation rate in day\(^{-1}\), constrained by \(k\ge 0\). The fitted profile log likelihood is \(-100.0\) at \(k=0\), \(-100.8\) at \(k=0.05\), and \(-102.1\) at \(k=0.10\). Suppose, for illustration, that the analysis uses cutoff \(c=2.0\). Find an approximate upper endpoint by linear interpolation between the latter two profile values.

The maximum is at the feasible boundary \(k=0\), so \(\hat{k}=0\) and \(q(k)=2\{\ell_p(0)-\ell_p(k)\}\). At \(k=0.05\), \(q=2[-100.0-(-100.8)]=1.6\). At \(k=0.10\), \(q=2[-100.0-(-102.1)]=4.2\). The cutoff lies between those values.

Linear interpolation gives \(k_U=0.05+\frac{2.0-1.6}{4.2-1.6}(0.10-0.05) =0.05+\frac{0.4}{2.6}(0.05) \approx 0.05769\ \text{day}^{-1}\). Thus the illustrative threshold set extends from the boundary to an approximate upper endpoint of \(0.05769\ \text{day}^{-1}\). The calculation supports a one-sided limit under this chosen cutoff; it does not establish that the biological rate is exactly zero. The interpolation is approximate, so the endpoint should be checked by evaluating the profile more closely near the crossing.

Worked Example: A Lower Endpoint Before the Bound

In another hypothetical fit, a nonnegative pathway rate \(a\), in day\(^{-1}\), has an interior profile maximum at \(\hat{a}=0.08\). Its profile statistic is \(q(0)=4.8\), \(q(0.02)=3.6\), and \(q(0.05)=1.8\). Use the illustrative cutoff \(c=2.0\) to estimate the lower crossing by linear interpolation.

The cutoff is between \(q(0.02)=3.6\) and \(q(0.05)=1.8\). Assuming a straight line over this short segment, the fraction of the distance from \(0.02\) to \(0.05\) needed to decrease from \(3.6\) to \(2.0\) is \((3.6-2.0)/(3.6-1.8)=1.6/1.8\). Therefore, \(a_L=0.02+(1.6/1.8)(0.05-0.02)=0.04667\ \text{day}^{-1}\), approximately.

Since \(q(0)=4.8>2.0\), the zero boundary is outside this illustrative threshold set, while the estimated crossing lies inside the feasible range. This is different from reporting zero as the lower limit simply because the parameter is constrained to be nonnegative. The interpolated value is only an approximation; the fitted profile should be evaluated near \(0.04667\ \text{day}^{-1}\) to confirm the crossing.

Worked Example: Checking Candidate Roots Against a Bound

A numerical profile calculation produces the candidate threshold-crossing equation \(\theta^2+0.40\theta-0.0304=0\), where \(\theta\) is a dimensionless parameter constrained by \(\theta\ge 0\). Solve the equation and identify which root is feasible.

The discriminant is \(0.40^2-4(1)(-0.0304)=0.1600+0.1216=0.2816\), and \(\sqrt{0.2816}\approx0.5307\). Hence \(\theta=\frac{-0.40\pm0.5307}{2}\), giving \(\theta\approx0.06533\) and \(\theta\approx-0.4653\). Only \(0.06533\) satisfies \(\theta\ge0\), so it is the feasible candidate root.

A substitution check confirms the arithmetic: \(0.06533^2+0.40(0.06533)-0.0304\approx0.004268+0.02613-0.0304\approx0\). The negative root also solves the equation algebraically, but it is outside the stated parameter space. A feasible root is only a candidate crossing: it must still be checked against the original profile statistic and the optimization results that produced the equation.

Why Bounds Can Produce an Artifact

An imposed bound can affect the numerical result even when it has no biological justification. For example, an optimizer may be instructed to search only \(k\ge 0.01\), although the scientific parameter space is \(k\ge0\). If the profile likelihood improves continuously as \(k\) decreases to \(0.01\), then \(0.01\) is where the search stopped—not necessarily where the likelihood has a maximum or where the data provide a lower limit. The missing range \(0\le k<0.01\) could contain a better fit or a profile threshold crossing.

A physical constraint and a computational bound should therefore be distinguished. A constraint such as a fraction lying in \([0,1]\) may be part of the model definition. A tighter limit chosen for numerical convenience is not automatically a scientific constraint. When the profile runs into the latter, expand the range if scientifically permissible and repeat the fit. If a genuine physical boundary is reached, report that the likelihood support is truncated there if no crossing occurs; do not describe the physical boundary as a data-derived limit.

For a parameter with a lower bound, the profile often has a characteristic one-sided shape near the boundary. A short justification comes from a local expansion. If the profile is differentiable from the right at zero, then for small feasible \(k\), \(\ell_p(k)\approx\ell_p(0)+\ell_p'(0)k+\frac{1}{2}\ell_p''(0)k^2\). If \(\ell_p'(0)<0\), the profile initially decreases as \(k\) increases from zero, consistent with a local maximum at the boundary. If instead the profile increases as \(k\) moves away from zero, the boundary cannot be a local maximum under this smooth approximation. The expansion is local, and does not establish global behavior; the full profile must be inspected.

Diagnostics and Alternative Summaries

Before interpreting a near-bound estimate, inspect the profile over the full relevant feasible range, not only at the reported estimate. In particular, check whether the profile turns around, crosses the chosen threshold, or continues toward the boundary. At fixed parameter values, confirm that nuisance parameters were re-optimized rather than held at their overall estimates. Compare fits from multiple starting values and verify that apparent profile changes are not caused by failed optimization. If a profile endpoint coincides with a computational limit, repeat the analysis with a wider range where appropriate.

Wald intervals summarize uncertainty using local curvature near the estimate and are often symmetric on the parameter scale. Near a boundary, that symmetry can yield an impossible negative rate or conceal the one-sided profile shape. A log transformation can enforce positivity and make a symmetric interval on the log scale correspond to an asymmetric interval on the original scale. It does not, however, create information about values the likelihood does not distinguish, nor does it resolve an arbitrary bound placed on the transformed parameter. Profile likelihood is useful here because it follows the fitted likelihood across feasible parameter values, while still requiring trustworthy optimization and an appropriate cutoff.

Boundary cases can also affect the reference distribution used to calibrate likelihood-ratio evidence. Thus a familiar cutoff should not be assumed to have its usual interpretation simply because a profile statistic has been computed. The distinction between the shape of the profile and the calibration of a test is essential: the profile reveals where a chosen threshold is crossed, while the suitability of that threshold depends on the inferential setting.

Common Mistakes and Limitations

Calling a bound a confidence endpoint. A profile that reaches an imposed limit without crossing the chosen threshold has not located a data-supported endpoint on that side.
Claiming the parameter equals zero. A boundary maximum says that zero gives the largest likelihood among feasible values under the fitted model; it does not prove the true mechanism is absent.
Keeping an infeasible root. Solve the crossing equation, then check every candidate against the model’s parameter domain and the original profile.
Using a standard cutoff without checking boundary conditions. A profile shape and a calibrated inferential statement are distinct; boundary behavior can affect calibration.

Profile likelihood remains conditional on the model, observation assumptions, and optimization. A well-behaved curve cannot compensate for a misspecified observation model, and a numerical threshold crossing cannot establish that the parameter is practically estimable at a useful precision. These results should be reported with the parameter units, the relevant bound, the chosen cutoff and its justification, and a clear statement about any side where the profile was truncated.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A rate \(r\ge0\) has a profile maximum at zero and crosses the chosen threshold at \(0.12\ \text{day}^{-1}\). What kind of limit does this give, and what does it not prove?
A profile is still improving as a parameter decreases to a computational lower bound of 0.05, and no threshold crossing occurs. What should be concluded about that endpoint?
For a dimensionless parameter \(\phi\ge0\), suppose \(q(0)=5.1\) and \(q(0.03)=1.7\), with illustrative cutoff \(c=2.0\). Is the boundary included in the threshold set?
Why can a symmetric Wald interval be misleading for a nonnegative rate near zero?

Solutions: (1) It gives a one-sided upper limit of \(0.12\ \text{day}^{-1}\) at the selected cutoff, conditional on the model and calibration. It does not prove that the true rate is exactly zero. (2) The endpoint is imposed by the search range; the profile supplies no threshold crossing there. Expand the range if appropriate and refit, or report that the profile is truncated at the genuine physical boundary. (3) No. Since \(q(0)=5.1>2.0\), zero is outside the threshold set; the profile’s crossing must be located between the supplied values. (4) A symmetric interval may extend below zero, outside the feasible parameter space, and may fail to represent the asymmetric likelihood shape near the boundary.