Why a Boundary Estimate Is Not a Conclusion
A fitted effect of zero can look decisive: the optimizer has reached the smallest allowed value, so the model appears to say that the effect is absent. But the same estimate can arise for quite different reasons. The data may strongly favor an effect close to zero, or they may contain too little information to distinguish zero from a range of nonzero effects. In the second case, zero is a boundary solution, not a well-supported scientific conclusion.
The distinction matters in pharmacometric models because effect parameters often govern predictions across dose, concentration, time, or covariate values. A parameter fixed at its permitted limit can conceal substantial uncertainty in those predictions. The diagnostic question is not only whether the estimated parameter equals zero, but also whether the likelihood and the predictions rule out scientifically important alternatives.
The distinction is evidential rather than a claim that the data can prove an exact zero. With finite data, a model that estimates zero does not by itself demonstrate that the population parameter is identically zero. A useful conclusion may instead be that the data constrain the effect to a range that is negligible for a specified purpose, or, alternatively, that they leave important effects plausible.
Two Checks: Likelihood Support and Prediction Change
Let \(\eta\) represent the model's other parameters. For each fixed value of \(\beta\), the profile log-likelihood is the largest log-likelihood obtainable by adjusting those other parameters:
This comparison allows nuisance parameters to adapt at each effect value, rather than unfairly holding them at their estimates from the boundary fit. Let \(\hat{\ell}\) be the largest log-likelihood over all permitted values of \(\beta\) and \(\eta\). The profile likelihood drop at \(\beta\) is
A larger drop means the data fit less well at that fixed value, after the other parameters have been re-optimized. The relative likelihood of that value compared with the best-fitting value is \(\exp\{-\Delta(\beta)/2\}\). This is a comparison of likelihoods, not the probability that a parameter value is true, and it is not a p-value. In particular, do not apply a generic likelihood-ratio reference distribution at a boundary without addressing the boundary conditions; the drop itself can still describe relative fit.
Prediction comparisons must use the same target observations or covariate settings. For each fixed \(\beta\), predictions can be calculated using the corresponding profiled estimates of \(\eta\). Specify in advance what change would matter: for example, a clinically relevant difference in an exposure, response, or dose recommendation. Without such a criterion, “small” prediction change is subjective. Also distinguish the change in the model's predicted value from the full predictive uncertainty, which includes sources of variability beyond uncertainty in \(\beta\).
Worked Interpretation
Worked Example: A Boundary Estimate With Evidence Against Important Effects
Hypothetical data. A concentration-effect model constrains a positive effect parameter \(\beta\) to be nonnegative. Its maximum log-likelihood is \(\hat{\ell}=-412.0\), attained at \(\hat{\beta}=0\). The profile log-likelihood is \(-414.5\) at \(\beta=0.10\) and \(-420.0\) at \(\beta=0.20\). In the target exposure range, the model predicts a response of 8.0 mg/L at \(\beta=0\), 8.3 mg/L at \(0.10\), and 8.7 mg/L at \(0.20\). Suppose, for this hypothetical decision, that a 1.0 mg/L change is the prespecified threshold for a material response change.
At \(\beta=0.10\), the profile drop is
Its relative likelihood is \(\exp(-5.0/2)=\exp(-2.5)\approx 0.0821\). At \(\beta=0.20\),
Thus these larger values fit substantially less well than the boundary estimate. Their predicted response changes are \(8.3-8.0=0.3\) mg/L and \(8.7-8.0=0.7\) mg/L, respectively; both are below the stated 1.0 mg/L materiality threshold. Taken together, the profile and predictions support an effect that is negligible for this target range and decision. They do not prove that the true \(\beta\) is exactly zero, and the conclusion need not apply to another exposure range or decision threshold.
Worked Example: A Boundary Estimate With Inadequate Information
Hypothetical data. In a second model, \(\hat{\beta}=0\) and \(\hat{\ell}=-220.0\). The profile log-likelihoods at \(\beta=0.10\) and \(0.20\) are \(-220.1\) and \(-220.4\). Predictions at the target setting are 10.0 mg/L at zero, 8.8 mg/L at \(0.10\), and 7.5 mg/L at \(0.20\). A change of 1.0 mg/L or more is material for this hypothetical application.
At \(\beta=0.10\), the profile drop and relative likelihood are
At \(\beta=0.20\),
The likelihood has changed little at either nonzero value. Yet the predicted changes relative to zero are \(8.8-10.0=-1.2\) mg/L and \(7.5-10.0=-2.5\) mg/L, both material in magnitude. Therefore, the data have not distinguished the boundary from effects with important prediction consequences. The appropriate interpretation is weak information about \(\beta\), not evidence that the effect is absent. The likelihood ratios here are relative fit measures; neither 0.905 nor 0.670 is a probability that the corresponding value is true.
Worked Example: Little Prediction Impact in the Observed Range
Hypothetical data. Consider a model with predicted response \(m(x,\beta)=10\exp(\beta x)\) response units, where \(x\) is a dimensionless exposure index and \(\beta\geq 0\). The estimate is \(\hat{\beta}=0\), but the profile remains relatively flat through \(\beta=0.50\), where \(\Delta(0.50)=0.40\). The observed exposure indices lie between 0 and 0.10. Suppose a 10% prediction change is material for use within that observed range.
The relative likelihood at \(\beta=0.50\) is \(\exp(-0.40/2)=\exp(-0.20)\approx 0.819\), so the profile alone does not strongly distinguish that value from zero. At the largest observed \(x=0.10\), the prediction at \(\beta=0.50\) is
Relative to the prediction \(m(0.10,0)=10\), the change is \(0.51\) response units, or approximately \(5.1\%\). It is below the stated 10% threshold throughout the observed range because \(m(x,0.50)\) increases with \(x\) for \(\beta\geq0\), and \(x\leq0.10\). For a much larger, unobserved index \(x=2\), however,
a large change from 10 response units. The parameter remains weakly identified by the profile. Predictions are practically insensitive to it over the observed range under the stated criterion, but that does not support extrapolating the zero-effect conclusion to \(x=2\) or claiming the parameter is exactly zero.
A Short Justification for the Two-Part Assessment
The profile and prediction checks answer different questions. The profile measures how much fit is lost when the effect is fixed at an alternative value and all other parameters are allowed to adjust. Since a likelihood is proportional to relative support for model parameter values under the assumed model, a large value of \(\Delta(\beta)\) means that alternative fits worse. Specifically,
But fit is not the same as consequence. Two parameter values can have similar likelihoods while producing meaningfully different predictions, as in the second example. Conversely, predictions at the observed covariates can be very similar even when the parameter itself is weakly identified, as in the third example. Therefore, neither the profile nor the predictions alone answer both the information and practical-impact questions. The joint interpretation is conditional on the model, data, target setting, and stated materiality criterion.
Assumptions, Diagnostics, and Alternatives
These comparisons are useful only if the likelihood evaluations and predictions are trustworthy. At minimum, the same observations, likelihood definition, and fixed model structure must be used throughout the profile. For each fixed \(\beta\), the nuisance-parameter optimization should be adequate; otherwise an artificially low profile value can make an alternative effect look more strongly excluded than it is. The boundary fit itself should also be a credible optimum, rather than an optimizer failure or a result sensitive to starting values. These checks complement the estimation and curvature diagnostics used for boundary and weak-identification problems.
A practical diagnostic sequence is:
Identify effect values and prediction changes that would matter in the intended population, exposure range, or decision. Do not define “negligible” only by whether an estimate rounds to zero.
At selected fixed effect values, re-optimize the other parameters. Check optimizer status, repeat from suitable starts, and verify that the resulting likelihood values are stable.
Ask whether meaningful nonzero values fit poorly, and whether they materially change predictions at the settings that matter. Report the setting and units for each prediction comparison.
State whether the data support a negligible effect, leave material effects plausible, or show negligible prediction impact only in a restricted range. Do not generalize beyond the profile range or prediction domain assessed.
A Wald standard error and symmetric interval can be misleading at a boundary: the local quadratic approximation may not describe the one-sided likelihood shape, and a symmetric interval may extend into values the model does not permit. A likelihood profile directly examines fit over permitted values, though its computation and boundary interpretation still require care. Comparing a model with \(\beta=0\) against one with \(\beta\) free can summarize a model comparison, but a single comparison does not show how much nonzero effect remains plausible or how predictions vary across it. Boundary likelihood-ratio calibration is not automatically the usual interior-parameter calibration; choosing a valid inferential reference distribution is a separate issue from describing the relative profile fit.
Warning signs include a jagged or discontinuous profile, materially different results from repeated optimizations, abrupt nuisance-parameter changes, or predictions that jump without a scientific explanation. Such patterns can indicate numerical or model-fitting problems rather than evidence for or against a zero effect. A profile that is nearly flat over the scientifically important range is a different warning: the model may fit successfully, but the data do not resolve the effect. More data, informative design variation, or a simpler scientific question may be needed; forcing a zero estimate does not create information.
Common Mistake
Wrong reasoning: “The fitted effect is exactly zero, so the drug has no effect.” The estimate reports the best fit under the specified constraints and model. It does not tell whether plausible nonzero values fit nearly as well, whether those values change predictions, or whether the model has adequate information to estimate the effect.
Correction: Check the likelihood deterioration for scientifically meaningful alternatives and calculate their prediction changes in the relevant setting. A sharply deteriorating profile, coupled with small prediction changes, can support a negligible effect for that use. A flat profile that permits material prediction changes indicates inadequate information. A flat profile with small prediction changes supports practical insensitivity only over the settings assessed—not proof of an exact zero parameter.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A boundary estimate is zero. The profile is nearly flat across nonzero values that cause important changes in predictions. Is this evidence for an absent effect, or for weak information?
- At a fixed effect value, \(\hat{\ell}=-80.0\) and \(\ell_p(\beta)=-82.0\). Calculate \(\Delta(\beta)\) and the relative likelihood.
- A flat profile permits a range of effects, but predictions change by less than a prespecified materiality threshold over the observed covariate range. What conclusion is justified, and what conclusion is not?
- Why should nuisance parameters be re-optimized at each fixed effect value when constructing a profile?
Solutions. (1) This indicates weak information: important prediction-changing effects remain compatible with nearly the same likelihood, so the boundary estimate does not establish absence. (2) \(\Delta(\beta)=2\{-80.0-(-82.0)\}=4.0\). The relative likelihood is \(\exp(-4.0/2)=\exp(-2)\approx0.135\). This is a relative fit measure, not a probability or p-value. (3) It is justified to say predictions are practically insensitive to the effect over the assessed covariate range and under the stated threshold. It is not justified to claim the parameter is exactly zero or to extend that conclusion to unassessed settings. (4) Re-optimizing allows the remaining parameters to take their best-fitting values for each fixed effect. Otherwise, the likelihood comparison can reflect an artificial restriction on nuisance parameters rather than the fit achievable at that effect value.