Selecting Plausible Parameter Ranges

Why Parameter Bounds Matter

A robust design is only as meaningful as the parameter values against which it is judged. A range that is too narrow can hide important loss of performance when estimates are wrong. A range that is too broad can force the design to protect against values that are technically possible but biologically or empirically implausible. The result may be a design that performs poorly across the parameter values that are actually credible.

Parameter ranges are inputs to a design analysis, not facts revealed by an optimization algorithm. They should therefore be supported by evidence and documented alongside the model, design criteria, and assumptions. In this tutorial, a parameter range means the specified set of parameter values treated as plausible for a design decision. It is not, by itself, a probability distribution.

Definition: A plausible parameter set, written \(\mathcal{R}\), is a specified set of parameter values supported by estimation uncertainty, biological or model constraints, and relevant external evidence. A parameter range gives the lower and upper bounds for one parameter. For a parameter vector, \(\mathcal{R}\) may also impose constraints on combinations of parameters.

For a scalar parameter \(\theta\), a simple range is \(\mathcal{R}=[L,U]\). For several parameters, using separate intervals creates a rectangular set, such as \(L_j\leq\theta_j\leq U_j\) for every component \(j\). That rectangle may contain combinations that are inconsistent with the data, biology, or model. A more defensible set can retain restrictions such as correlations, ratios, or bounds on derived quantities.

This distinction matters because scenario-based, maximin, and minimax-regret criteria assess performance over specified parameter values. Changing the range can change which values count as worst cases and can therefore change the selected design. Range selection is part of the scientific design problem, not a neutral formatting choice.

Evidence for Bounds

Estimation uncertainty is one important source of range information. If a parameter estimate and its uncertainty are available from a relevant model, an interval can be constructed on the scale on which the parameter was estimated, then transformed back if needed. The interval level should be stated. A confidence interval describes the behavior of an estimation procedure under repeated sampling; it does not mean that there is a specified probability that the fixed parameter lies inside this particular interval.

Biological and model constraints provide a different kind of evidence. Examples include positive clearance and volume, nonnegative rates, or bounds implied by a plausible half-life. Such constraints can rule out impossible or incompatible values, but they do not automatically identify the most likely values within the allowable region. External evidence, including estimates from relevant compounds or populations, can help distinguish a biologically possible range from one that is credible for the intended application. Its relevance should be assessed rather than assumed.

A defensible range often combines these sources. For example, an estimate-based interval may be narrowed by a genuine biological constraint, or a broad biological range may be made more relevant by evidence from a similar population. Record which source supports each bound and why it applies. If evidence sources conflict, report the conflict and examine its effect rather than silently choosing the most convenient bound.

Key distinction: A bound states which values are included in the design analysis. It does not state how probable those values are. Assigning probabilities to parameter values is a separate modeling choice; do not interpret an unweighted range as a prior distribution.

Worked Examples of Constructing and Using Ranges

Worked Example: Transforming Estimation Uncertainty for Clearance

Suppose a hypothetical analysis of a relevant population estimates \(\log(CL/(1\ \mathrm{L/h}))=1.61\), with standard error \(0.12\). Assume the estimate is approximately normally distributed on the log scale for the purpose of constructing a 95% confidence interval. Find the interval on the clearance scale.

On the log scale, the interval is estimate \(\pm 1.96\) standard errors:

1.61\pm 1.96(0.12)=1.61\pm0.2352=[1.3748,\ 1.8452].

Exponentiating both endpoints gives:

CL\in[\exp(1.3748),\exp(1.8452)]\ \mathrm{L/h} =[3.95,\ 6.33]\ \mathrm{L/h}.

The estimated 95% confidence interval is therefore approximately 3.95 to 6.33 L/h. For a design analysis, an analyst might round this to a documented working range of 4.0 to 6.3 L/h, provided that rounding does not exclude relevant values. The interpretation is about the confidence-interval procedure, not a 95% probability that this fixed clearance lies in this particular interval. Nor does the interval alone establish that the source population and model are appropriate for the intended design.

Worked Example: Combining Biological and External Constraints

Consider a hypothetical drug with clearance \(CL\) between 2 and 8 L/h and volume \(V\) between 20 and 50 L, based on preliminary information. Similar compounds provide external evidence supporting half-lives from 2 to 10 hours. Use \(t_{1/2}=0.693V/CL\) to express the resulting plausible set and check whether the rectangular bounds alone are adequate.

The biologically and externally constrained set is:

\mathcal{R}=\{(CL,V):2\leq CL\leq8,\ 20\leq V\leq50,\ 2\leq0.693V/CL\leq10\}.

The outer rectangle includes combinations that violate the half-life evidence. At \(CL=2\) L/h and \(V=50\) L, the implied half-life is \(0.693(50)/2=17.3\) hours, exceeding 10 hours. That corner should not be treated as plausible under the stated evidence. By contrast, \(CL=8\) L/h and \(V=20\) L imply \(0.693(20)/8=1.73\) hours, below 2 hours, and are also excluded. A combination such as \(CL=4\) L/h and \(V=30\) L gives \(0.693(30)/4=5.20\) hours and satisfies all constraints.

The resulting set is not simply the four independent interval limits: the half-life condition links \(CL\) and \(V\). Keeping that dependence avoids evaluating designs at combinations that conflict with the evidence. The numbers are hypothetical, and the external half-life range would need a scientific justification for the actual application.

Worked Example: How an Unsupported Extreme Can Change a Design

Consider a simplified hypothetical one-parameter model with mean concentration \(\mu(t;k)=\exp(-kt)\), where \(k\) is an elimination rate in \(\mathrm{h}^{-1}\), \(t\) is in hours, and the observation error is independent normal with known standard deviation 1 concentration unit. Compare one observation at \(t=1\) hour with one at \(t=4\) hours. Assume the model and error variance are correct and fixed. For this scalar parameter, the Fisher information from one observation is the squared derivative of the mean with respect to \(k\), divided by the error variance:

I(t;k)=\frac{1}{1^2}\left(\frac{\partial\mu(t;k)}{\partial k}\right)^2 =t^2\exp(-2kt).

First use a broad range \(0.1\leq k\leq1.0\ \mathrm{h}^{-1}\). For each sampling time, the information decreases with \(k\), so the worst value is at \(k=1.0\). At \(t=1\), the worst-case information is \(1^2\exp(-2)=0.1353\). At \(t=4\), it is \(4^2\exp(-8)=0.005367\). Under a maximin-information criterion over this range, the 1-hour observation is selected.

Now suppose relevant evidence supports the narrower range \(0.1\leq k\leq0.4\ \mathrm{h}^{-1}\). Again the worst case for each time is at the upper endpoint. The 1-hour information is \(\exp(-0.8)=0.4493\) at \(k=0.4\) only if the formula were \(t^2\exp(-kt)\); using the stated information formula gives \(\exp(-0.8)=0.4493\) as well, because \(t=1\) and the exponent is \(-2kt=-0.8\). At \(t=4\), the worst-case information is \(16\exp(-3.2)=0.6522\). Thus the 4-hour observation has higher worst-case information in the narrower range.

The design choice changes because the assumed range changes, not because the information calculation changed. The broad range makes the late observation vulnerable to a high elimination rate, but if that extreme is not supported for the target application, protecting against it can lead to an unhelpful choice. Conversely, narrowing a range without evidence can create false reassurance. The calculation is illustrative: real designs may have multiple observations, nuisance parameters, residual-error models, and operational constraints.

Justifying and Checking a Range

A practical range should be traceable to evidence and compatible with the model. The following workflow makes the choices explicit.

Specify the target application.
Identify the population, dosing context, model, and design decision. Evidence from a substantially different setting may not support the same bounds.

Assemble uncertainty and constraints.
Collect estimates and their uncertainty, plausible biological limits, model-implied restrictions, and relevant external evidence. State the scale and units for every parameter.

Construct the set, not just its endpoints.
Check whether parameters are linked by derived quantities or joint evidence. Retain scientifically supported dependence rather than including every combination in a rectangular product of intervals.

Evaluate boundary behavior and sensitivity.
Inspect design performance near the bounds and at relevant interior values. If a selected design is determined by one boundary, verify that the boundary is supported and examine a justified alternative range.

For smooth information functions, evaluating endpoints may be enough only when monotonicity has been established. Otherwise, an interior value can be the worst case. In multidimensional problems, checking a few corners does not guarantee that the worst case has been found; the shape of the criterion and the constraints matter. Numerical optimization over a range should therefore be checked with additional starts, grids, or other suitable diagnostics.

Warning signs include a maximin design that changes sharply when a single unsupported endpoint is added, parameter combinations that violate known relationships, and boundary values far outside the estimation evidence with no biological rationale. Also check that transformed bounds remain in the parameter space permitted by the model, and that units and transformations have been applied consistently. These diagnostics do not prove that the range is wrong; they indicate where its justification or the optimization needs closer review.

Alternatives, Limitations, and Common Mistakes

A nominal design optimizes performance at one chosen parameter estimate. It is simpler and can be efficient near that value, but it does not directly protect against uncertainty away from the nominal point. Range-based robust criteria assess performance across specified values: maximin focuses on the worst value, while other criteria may aggregate or compare losses differently. Their conclusions depend on the chosen set and criterion.

A probability-weighted approach instead requires a distribution over parameter values and evaluates performance across that distribution. That can distinguish common from rare values, unlike an unweighted range, but it adds assumptions about probability weights. Selecting plausible bounds is not a substitute for specifying or assessing those weights. Likewise, an interval from estimation uncertainty is not automatically a defensible prior distribution.

A range cannot repair a misspecified structural or residual-error model, and evidence may be weak for a new drug, population, or dose. A deliberately conservative range can be useful, but the resulting design may sacrifice performance over the central region to protect its edges. Report that trade-off rather than presenting the selected range as uniquely correct.

Mistake: treating all mathematically possible values as plausible. The correction is to distinguish the model's parameter space from the evidence-supported set used for design.
Mistake: taking independent lower and upper bounds as proof that every combination is credible. The correction is to retain relevant joint constraints or explain why independence is a reasonable approximation.
Mistake: assuming a wider range is always more robust. The correction is to examine whether the added values are supported and how they alter the design objective and choice.
Mistake: interpreting a range as a probability distribution. The correction is to state that it defines the values considered; probability weights require a separate specification.

Key takeaway: Defensible robust design starts with a defensible parameter set. Use estimation uncertainty, biological limits, and relevant external evidence; preserve important dependence; and verify that worst-case values are both computationally identified and scientifically credible.

Check Your Understanding

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

A parameter estimate has a standard error. Why is a confidence interval based on that estimate not automatically a probability distribution for robust design?
Suppose a design parameter is clearance \(CL\), and the model also uses volume \(V\). Give one reason why independent bounds on \(CL\) and \(V\) may define an unsuitable plausible set.
For the simplified model in the third worked example, calculate the information at \(t=1\) hour and \(k=0.4\ \mathrm{h}^{-1}\). Include units if information is reported with units.
A broad range makes the design change because of one extreme value. What should be checked before narrowing the range?
In one sentence, distinguish a range-based robust analysis from a probability-weighted analysis.

Solutions. (1) A confidence interval has a repeated-sampling coverage interpretation; it does not assign probabilities to parameter values. A probability distribution requires an additional modeling choice. (2) The implied half-life \(0.693V/CL\) may violate biological or external constraints for some combinations, so the rectangle can include unsupported pairs. (3) \(I(1;0.4)=1^2\exp[-2(0.4)(1)]=\exp(-0.8)=0.4493\) in inverse squared units of \(k\), or \(\mathrm{h}^2\), under the stated unit-variance concentration error. (4) Check the extreme's evidence and relevance, parameter dependence, model validity, and whether the optimization has correctly identified the worst case; then assess a narrower range only if evidence supports it. (5) A range-based analysis evaluates specified values without probability weights, whereas a probability-weighted analysis averages performance according to an explicitly specified distribution.