Standardizing Predictions To A Reference Population

Why Standardize Predictions?

A pharmacometric model can predict an outcome for a subject with specified covariates, but a population-level summary also depends on which covariates are represented and how common they are. A prediction averaged over a population with mostly normal renal function may differ from one averaged over a population with more renal impairment, even when the same model, dose, and prediction time are used. Standardization makes that dependence explicit: first predict for subjects represented in a defined reference population, then average those predictions according to that population’s covariate distribution.

The reference distribution is part of the quantity being reported, not a cosmetic choice made after modeling. Two standardized predictions can be compared meaningfully only when their model prediction, regimen, endpoint, and prediction time are aligned, or when differences in those ingredients are deliberately part of the comparison.

Definition and Calculation

Definition: Let $X$ denote the vector of covariates and $F^*$ a specified reference distribution for those covariates. For a fixed model parameter vector $\theta$, regimen $d$, and prediction time $t$, let $m(d,t,x;\theta)$ be the model-predicted conditional mean outcome for covariates $x$, under a stated convention for random effects and residual error. The standardized population prediction is $$ \psi(d,t;F^*,\theta)=\int m(d,t,x;\theta)\,dF^*(x). $$ For reference records $x_1,\ldots,x_N$ with nonnegative weights $w_i$ summing to one, its empirical estimate is $$ \widehat{\psi}(d,t;F^*,\theta)=\sum_{i=1}^{N}w_i\,m(d,t,x_i;\theta). $$

With equally weighted records, $w_i=1/N$. With unequal weights, each record contributes in proportion to its representation in the target population. The prediction $m$ must be defined consistently: for example, a typical-subject prediction that fixes random effects at zero is not the same target as a prediction averaged over a random-effects distribution. If the target is an expected outcome, residual variability is normally integrated out rather than adding a fresh residual error realization to each prediction.

This is a model-based population average, not necessarily the average of observed outcomes in a data set. It also is not automatically an average over parameter uncertainty. The definition above holds $\theta$ fixed. When parameter uncertainty is to be included, that is a separate layer of calculation: predictions can be standardized for each parameter draw and the resulting distribution summarized. Keeping these quantities distinct makes clear whether a reported number is conditional on one fitted parameter vector or incorporates uncertainty in that vector.

Why the Average Must Follow Individual Prediction

The standardization formula follows from conditioning on covariates. If $Y$ is the outcome under the specified regimen and time, and the model gives $E(Y\mid X=x)=m(d,t,x;\theta)$, then the law of total expectation gives $$ E(Y)=E\{E(Y\mid X)\}=\int m(d,t,x;\theta)\,dF_X(x). $$ Setting $F_X=F^*$ defines the corresponding model-based mean for the reference population. This argument requires that the conditional prediction be defined for covariate values represented by $F^*$, and that the reference distribution describe the population to which the result is intended to apply.

For a finite reference sample with equal weights, the average of the predictions is a direct approximation to that expectation. In particular, if the records are sampled independently from $F^*$, then $$ E\left\{\frac{1}{N}\sum_{i=1}^{N}m(d,t,X_i;\theta)\right\} =\frac{1}{N}\sum_{i=1}^{N}E\{m(d,t,X_i;\theta)\} =\psi(d,t;F^*,\theta). $$ Thus, averaging subject-level predictions estimates the target-population average under that sampling scheme. If the records were selected with a design that over- or under-represents parts of the reference population, their unweighted mean need not estimate the desired average; appropriate weights are required.

Worked Example: Standardizing Across Two Renal-Function Groups

Suppose a hypothetical model predicts steady-state trough concentration at a fixed dose and time. For subjects in a normal-renal-function group, the subject-level predictions are represented by 12 mg/L; for subjects in an impaired-renal-function group, they are represented by 8 mg/L. Assume these values are the relevant group means under the model and that the reference population is 70% normal and 30% impaired. Calculate the standardized prediction.

The reference proportions are the weights: $w_{\text{normal}}=0.70$ and $w_{\text{impaired}}=0.30$. They are nonnegative and sum to one. The weighted average is $$ \widehat{\psi} =0.70(12\ \text{mg/L})+0.30(8\ \text{mg/L}) =8.4\ \text{mg/L}+2.4\ \text{mg/L} =10.8\ \text{mg/L}. $$

The standardized prediction is 10.8 mg/L for a population with the stated renal-function composition. It is not the prediction for every member of that population, nor does it imply that an individual has a 10.8 mg/L concentration. It summarizes the specified subject-level predictions using the reference population’s group proportions.

If the reference population instead had 40% normal and 60% impaired renal function, the calculation would be $0.40(12)+0.60(8)=9.60$ mg/L. The model predictions did not change; the standardized result changed because the reference covariate distribution changed.

Nonlinearity and the Prediction at Mean Covariates

A frequent alternative is to calculate a prediction for a hypothetical subject with the reference population’s mean covariates. That is a different calculation. In general, $$ E\{m(X)\}\ne m\{E(X)\}, $$ because model predictions can be nonlinear functions of covariates. Equality holds for an affine prediction function, but should not be assumed for nonlinear functions. The standardized prediction uses the distribution of covariates and averages predictions across it; a prediction at mean covariates substitutes a single covariate vector into the model.

Worked Example: A Nonlinear Exposure Prediction

Consider a hypothetical model in which clearance is $CL=2+0.04\,WT$ L/h, where $WT$ is weight in kg. At a dose of 100 mg, suppose the exposure metric is $AUC=100/CL$, in mg·h/L. A reference population consists of three equally weighted subjects weighing 50, 70, and 90 kg. Calculate the standardized mean exposure and compare it with the prediction at the mean weight.

First calculate each clearance and exposure. At 50 kg, $CL=2+0.04(50)=4.00$ L/h, so $AUC=100/4.00=25.00$ mg·h/L. At 70 kg, $CL=2+0.04(70)=4.80$ L/h, so $AUC=100/4.80=20.83$ mg·h/L. At 90 kg, $CL=2+0.04(90)=5.60$ L/h, so $AUC=100/5.60=17.86$ mg·h/L.

The reference weights are all $1/3$, so the standardized mean is $$ \widehat{\psi} =\frac{1}{3}(25.00+20.83+17.86) =21.23\ \text{mg·h/L}. $$ The mean weight is $(50+70+90)/3=70$ kg. Predicting at that mean gives $100/\{2+0.04(70)\}=100/4.80=20.83$ mg·h/L. Therefore, the prediction at mean weight is 0.397 mg·h/L lower than the standardized mean in this example.

Both calculations are arithmetically valid, but they answer different questions. The standardized result averages predictions for the three covariate profiles. The other result describes a single hypothetical subject at the mean weight. Here the reciprocal relationship between clearance and exposure makes the distinction consequential.

Stratified Samples and Reference-Population Comparisons

Stratified covariate sampling can improve representation of groups that are uncommon in a raw sample, but a deliberately balanced sample does not automatically represent the target population. If the sampling fractions differ from the target proportions, standardize using target weights. More generally, when predictions are summarized within strata $h=1,\ldots,H$, the target mean is $$ \widehat{\psi}=\sum_{h=1}^{H}\pi_h^*\,\bar{m}_h, $$ where $\pi_h^*$ is the proportion of the reference population in stratum $h$, the proportions sum to one, and $\bar{m}_h$ is the mean subject-level prediction in that stratum. This is a weighted average of stratum predictions, not necessarily the unweighted average across strata.

Worked Example: Correcting an Equal-Stratum Sample to Two Targets

A hypothetical simulation draws the same number of covariate profiles from each of three disease-severity strata. The average predicted trough concentrations within the strata are 18 mg/L for mild, 14 mg/L for moderate, and 9 mg/L for severe disease. The equal allocation was chosen for simulation convenience. Calculate the standardized predictions for two target populations: Target A has proportions 0.20, 0.50, and 0.30; Target B has proportions 0.40, 0.40, and 0.20, respectively.

For Target A, the target proportions sum to $0.20+0.50+0.30=1.00$. Reweight the stratum means by those proportions: $$ \widehat{\psi}_{A} =0.20(18)+0.50(14)+0.30(9) =3.60+7.00+2.70 =13.30\ \text{mg/L}. $$ For Target B, the proportions also sum to one: $$ \widehat{\psi}_{B} =0.40(18)+0.40(14)+0.20(9) =7.20+5.60+1.80 =14.60\ \text{mg/L}. $$

The standardized prediction for Target B is 1.30 mg/L higher than for Target A. This difference reflects the specified composition and the different stratum predictions; it is not a change in the model’s within-stratum predictions. Simply averaging the three strata equally would give $(18+14+9)/3=13.67$ mg/L. That number corresponds to a reference population with equal proportions in the three strata, not to either target described here.

Assumptions, Diagnostics, and Alternatives

A standardized result is interpretable only if the population definition and prediction calculation agree. Before reporting one, check the following:

Reference population: State which covariates define it, the intended population, and how its distribution or weights were obtained.
Common prediction target: Use the same endpoint, units, dose or regimen, prediction time, and random-effects convention for every profile being averaged.
Valid covariate profiles: Check that profiles are internally coherent and that the model can provide meaningful predictions for them. Predictions for unsupported combinations should not be treated as reliable merely because the averaging formula can be evaluated.
Weights: Confirm that weights are nonnegative and sum to one. For stratified sampling, verify that target proportions—not simulation allocation proportions—are used when the goal is a target-population average.
Calculation checks: Inspect subject-level and stratum-level predictions for unexpected values, verify units, and confirm that the weighted result lies between the minimum and maximum predictions when all weights are nonnegative.

That last check follows directly from the weighted-average form. If $m_{\min}\le m_i\le m_{\max}$ and $w_i\ge0$ with $\sum_i w_i=1$, then $m_{\min}\le\sum_iw_im_i\le m_{\max}$. A result outside this range often signals a weight, unit, or implementation error.

Standardization is not the only way to summarize predictions. A prediction at mean covariates is simpler but can miss nonlinear averaging. An unweighted mean of observed or simulated records is appropriate when those records represent the intended reference distribution; otherwise, it can describe the sample’s composition instead. A stratum-weighted average is a convenient implementation when the reference distribution is specified by strata, while record-level weighting can retain finer variation within strata. These approaches target the same quantity when they use the same model predictions and correctly represent the same reference distribution.

Common Pitfall

Common mistake: “The simulation sampled equal numbers from three strata, so the average of the three stratum means is the population prediction.” Equal simulation allocation describes the sample design, not necessarily the target population. If the target proportions are unequal, use those proportions as weights. Also, do not replace the weighted average of subject-level predictions with a prediction at mean covariates unless that is explicitly the intended quantity and the distinction is acceptable for the model.

A useful reporting statement names both the prediction and its reference. For example: “The model-predicted mean trough concentration at the specified regimen and time, standardized to a reference population with 20% mild, 50% moderate, and 30% severe disease, was 13.30 mg/L.” This makes the result reproducible and prevents a population-specific average from being mistaken for a universal model prediction.

Check Your Understanding

Try each problem, then compare your reasoning with the solutions.

A model predicts 6 mg/L for one covariate group and 10 mg/L for another. The reference proportions are 25% and 75%. What is the standardized prediction?
For a reference population with equally weighted covariate values 2 and 6, a model predicts $m(x)=x^2$. Compare the standardized prediction with the prediction at the mean covariate value.
A simulation samples equal numbers from two strata, but the target proportions are 80% and 20%. The stratum mean predictions are 5 mg/L and 15 mg/L. What target-standardized prediction should be reported?
Name two conditions that should be checked before interpreting a standardized prediction as applying to a target population.

Solutions. (1) $0.25(6)+0.75(10)=1.50+7.50=9.00$ mg/L. (2) The subject-level predictions are $2^2=4$ and $6^2=36$, so their mean is $(4+36)/2=20$. The mean covariate is 4, and the prediction there is $4^2=16$; the results differ because the model is nonlinear. (3) Use target rather than sampling proportions: $0.80(5)+0.20(15)=4+3=7.00$ mg/L. (4) Answers include confirming that the reference covariate distribution represents the intended target, that all profiles receive predictions for the same regimen and time, that weights are nonnegative and sum to one, and that predictions are meaningful for the profiles used.