Screening Implausible Parameter Draws

Why Screen Parameter Draws?

A parameter draw can be mathematically sampleable yet unsuitable for the intended simulation. A draw might assign a negative value to clearance, produce a prediction outside a scientifically defensible range, or cause the model to return a non-finite result. Screening is the explicit process of detecting and handling such cases. Its purpose is not to make an inconvenient simulation look reasonable; it is to identify what the draw implies and make any exclusion rule visible.

A crucial distinction is whether a draw is invalid under the model or implementation, or merely implausible for the scientific question. For example, a negative clearance is incompatible with the usual interpretation of clearance in a standard pharmacokinetic model. A very high but positive clearance might instead be unusual under a particular fitted model or target population. The latter cannot be rejected just because its prediction is surprising: the rule needs a scientific justification tied to the intended use.

The earlier material on posterior draws and parameter uncertainty provides the sampled parameter vectors to be evaluated. Screening acts on those vectors and their implications. It does not by itself provide a new source of uncertainty or establish that the retained draws represent the original uncertainty without change.

Specify What Counts as a Failure

A useful screening plan separates checks into layers. First, check whether parameter values satisfy the model's mathematical domain: for example, positive clearance and volume in a standard one-compartment model. Second, check whether model execution succeeded: outputs should be finite, and any required state variables should obey the model's structural constraints. Third, evaluate scientific plausibility using criteria appropriate to the intended scenario, such as a justified prediction range at a specified time. A draw that fails one layer should not be described as failing another.

The rules should be chosen before reviewing which draws pass. Record the parameterization, units, model version, simulation settings, numerical tolerances, each criterion, and the disposition of every draw. If a prediction-based bound is used, state its endpoint, time, scenario, and scientific rationale. A bound is not universal merely because it is convenient. Also retain the original draws and a record of rejected ones; otherwise, the screening decision cannot be audited or its effect on uncertainty assessed.

Checks should usually apply to the whole parameter vector. A combination of individually plausible values may imply implausible behavior, particularly when the model is nonlinear. Conversely, screening each component separately can destroy the dependence structure represented by joint draws. When one criterion fails, retain the original vector in the audit record rather than replacing its failing component with a boundary value.

A Prediction-Based Check in a Simple Model

For a transparent illustration, suppose a hypothetical model describes an instantaneous intravenous bolus in a one-compartment system with first-order elimination. Let Dose be the administered amount, \(V\) the volume of distribution, and \(CL\) clearance. Under these assumptions, concentration at time \(t\) is:

Here Dose is in mg, \(V\) in L, \(CL\) in L/h, \(t\) in h, and \(C(t)\) in mg/L. The formula assumes instantaneous distribution, a constant volume, and first-order elimination; it is an illustration of screening, not a general concentration model. For this hypothetical screening exercise, suppose the model owner has specified component bounds \(1 \leq CL \leq 10\) L/h and \(10 \leq V \leq 80\) L, and a scenario-specific check that \(C(24)\) must be between 0.05 and 2.00 mg/L. Those limits are stipulated for the example and should not be taken as clinical standards.

Why Rejection Changes the Distribution

The effect of rejection can be stated precisely. Let \(\theta\) denote a random parameter vector with density \(f(\theta)\), and let \(A\) be the set that passes the screening rule. If draws are repeatedly generated from \(f\) and only those in \(A\) are kept, the retained draws have the conditional density:

where \(\mathbf{1}(\theta\in A)\) equals 1 inside \(A\) and 0 outside. This follows from the definition of a conditional density: the joint density of \(\theta\) and the event \(\theta\in A\) is \(f(\theta)\mathbf{1}(\theta\in A)\); dividing by the event probability normalizes the density over the retained region. Thus, unless the rule excludes only impossible or erroneous draws that were never part of the intended distribution, rejection represents a conditional distribution, not the original one. In a Bayesian setting, rejecting posterior draws can be interpreted as conditioning on the additional restriction, but that is a change in the inferential target and needs justification.

Diagnostics and Alternatives

A screening procedure needs diagnostics of its own. Report the fraction retained and tabulate rejection reasons; a single total can conceal whether most failures arose from a parameter-domain check, a numerical issue, or a scientific prediction rule. Inspect the distribution of rejected and retained draws, the locations of failures in parameter space, and whether failures cluster in combinations of parameters. For prediction-based screening, examine which times, scenarios, and outputs trigger the rule. A sharp change in conclusions under small, defensible changes to a cutoff is evidence that results depend strongly on the screening choice.

Unexpectedly frequent numerical failures may indicate an implementation problem rather than implausible parameter uncertainty. Check input units, parameter mapping, initial conditions, solver settings, and model execution. A non-finite result should be diagnosed and recorded; it should not automatically be called scientifically impossible. The earlier material on model execution and cross-software equivalence is relevant when the same vectors behave differently across implementations.

Screening is one option among several, and the appropriate choice depends on why draws fail:

A transformation may be preferable when a parameter must be positive by definition; the earlier material on transforming parameter draws covers that construction. It does not justify transforming a draw after seeing an undesirable result without assessing the induced distribution. Clipping is especially hard to interpret: replacing a clearance of 11 L/h by 8 L/h does not preserve the original draw or implement ordinary rejection. If rejection is needed to obtain a fixed number of usable draws, generate additional draws under the same documented rule and report both the number attempted and the number retained.

Common Pitfall: Treating a Cutoff as Harmless Cleanup

The wrong reasoning is: “These few draws look extreme, so I will drop them; the remaining draws still represent the same parameter uncertainty.” That conclusion does not follow. If the draws came from an intended distribution and the cutoff excludes a region with positive probability, the retained distribution is conditional on passing the rule. Its mean, spread, correlations, and downstream predictions can all differ. The correction is to define and justify the rule, quantify how many and which draws it affects, state the target distribution after screening, and compare conclusions with and without the restriction when both are scientifically meaningful.

A second mistake is to reject a whole vector because one simulated output is surprising, without checking whether the output is genuinely impossible, simply rare, or caused by a coding or solver issue. Diagnose the failure first. If the draw is rare but valid under the intended parameter distribution, excluding it changes uncertainty rather than repairing an error.