Plug-In Simulation at the Population Parameter Estimates

What a Plug-In Simulation Produces

A fitted NLME model can be evaluated under a proposed dose, covariate pattern, and observation schedule to produce a model-based profile. In a plug-in simulation at the population parameter estimates, the estimated fixed effects are inserted into the model and held fixed. Evaluating the model at successive times or scenarios creates a profile, but repeating the same calculation with the same inputs gives the same values.

That reproducibility is useful: it makes the model's fitted structural predictions easy to inspect and compare across regimens or covariate values. It is also important to name the result carefully. A profile calculated at fixed effects alone is not automatically a distribution of outcomes in a population. In particular, it does not by itself include uncertainty in the estimated parameters, between-subject random effects, or residual error.

The distinction between these quantities matters because NLME models describe several sources of variation. The plug-in calculation substitutes estimates for unknown fixed effects; it does not draw alternative plausible parameter values. Setting \(b=0\), when that is the chosen prediction, gives the model's structural output at the reference random-effect value. It does not average predictions over the random-effects distribution. Nor does a structural prediction include observation-level residual error unless that error is deliberately added in a separate step.

Model, Inputs, and Assumptions

Consider a simple hypothetical one-compartment concentration model after an intravenous bolus. Let \(C(t)\) be concentration in mg/L, \(D\) the dose in mg, \(V\) volume in L, and \(CL\) clearance in L/h. With first-order elimination, the structural prediction is

Suppose clearance depends on body weight \(w\) through \(CL(w)=CL_{70}(w/70)^{0.75}\), where \(CL_{70}\) is clearance at 70 kg. In the examples, all values are hypothetical. The plug-in calculation uses \(D=420\) mg, \(V=42\) L, and fitted estimates \(\widehat{CL}_{70}=5.2\) L/h and \(\hat{V}=42\) L. These estimates are treated as fixed inputs for the calculations.

The profile is meaningful only relative to its model and inputs. Here, the equation assumes an instantaneous intravenous dose, a one-compartment disposition structure, and first-order elimination over the time range considered. The units are compatible: \(D/V\) is mg/L, while \(CL\,t/V\) is dimensionless because \((\mathrm{L/h})(\mathrm{h})/\mathrm{L}=1\). A different route, dosing history, or structural model would require the corresponding model equation and inputs rather than reuse of this one.

Calculating Profiles Across Scenarios

A plug-in profile is obtained by evaluating the same fitted model for each desired time or scenario. The estimates remain fixed; the inputs such as time, weight, or dose change. This makes it straightforward to compare model predictions under controlled changes, provided each scenario is within a range where the model and its covariate relationships are credible.

Why Fixed Effects Alone Do Not Give a Population Distribution

It is tempting to interpret a zero-random-effect profile as the average concentration in the population. That interpretation is generally unjustified in a nonlinear model. Even if a random effect \(b\) has mean zero, averaging the model output over \(b\) need not give the output at \(b=0\). A short Taylor expansion shows why.

For a scalar random effect with mean zero and variance \(\omega^2\), expand \(f\) around zero, holding other inputs fixed. To second order,

Taking expectations and using \(E(b)=0\) and \(E(b^2)=\omega^2\) gives

Unless the curvature term is zero (or negligible), the marginal mean differs from the zero-random-effect prediction. The approximation is a local justification, not an exact formula for every NLME model. It establishes the key point: in a nonlinear model, substituting a reference random-effect value is not generally equivalent to averaging predictions across the random-effects distribution.

There is a separate omission: plug-in estimates are uncertain because they were estimated from finite data. Holding \(\hat{\theta}\) fixed does not account for that estimation uncertainty. A profile may therefore look smooth and precise while still being uncertain as a prediction. Parameter uncertainty, random-effect variability, and residual error answer different questions; they should not be treated as interchangeable sources of spread.

Checks, Alternatives, and Limitations

Before interpreting a plug-in profile, verify the calculation and its scope. Check that units agree, the dosing history and time origin are correct, the covariates are in the units and coding expected by the model, and the output is in the intended quantity. Compare the profile with the fitted data and model diagnostics available for the relevant range. Predictions far outside observed covariate or dose support rely more heavily on the model's functional form and should be described as extrapolations.

A useful diagnostic is to recompute a few points independently, including a simple reference case such as \(t=0\), for which this model gives \(C(0)=D/V\). If a purported plug-in profile changes between runs despite identical inputs and estimates, some quantity is being varied or randomly generated; the calculation is no longer solely the deterministic evaluation described here. Conversely, identical repeated outputs do not demonstrate that the model is correct or that individual outcomes are identical.

These alternatives are not substitutes for one another: each targets a different quantity. A fixed-effects plug-in profile is often a useful first calculation because its inputs and structural consequences are transparent. Its limitations are equally important. It does not, by itself, provide prediction intervals, subject-level distributions, or simulated observed data. Those require additional model components and simulation choices, which should be stated rather than silently implied.

Common Mistakes

• Calling the plug-in profile “the population mean.” A prediction at \(b=0\) is generally not the average of predictions over random effects in a nonlinear model. Name it as a zero-random-effect or fixed-effects prediction unless an appropriate population averaging calculation was made.
• Reading a smooth curve as evidence of low uncertainty. The curve is smooth because the same fitted estimates are used at every point. Smoothness does not account for estimation uncertainty or subject variability.
• Assuming a simulation must be random. Evaluating a structural model on a grid is a deterministic simulation. Random draws are needed only when the target quantity calls for simulated variability or uncertainty.
• Rounding intermediate values too early. Retain full precision for intermediate calculations, then round the final output. Small rounding differences can otherwise alter the reported concentration.
• Extrapolating without qualification. The model will return a numerical value for many inputs, but a computed value outside data support is not automatically reliable.