End-To-End Virtual Population Construction Workshop

From Targets to a Qualified Virtual Population

A virtual population is useful only when its individuals jointly represent the population relevant to the QSP question. Matching each trait separately is not enough: a generated population can reproduce the marginal distributions of body weight and clearance yet miss their association, or match overall phenotype prevalence while misrepresenting a clinically important subgroup. Construction is therefore an end-to-end task: define the target population, generate bounded correlated traits, assign categorical phenotypes, and evaluate fit and diversity against criteria set in advance.

This workshop uses hypothetical targets and generated summaries. The calculations illustrate a construction and qualification process; they are not results from a clinical study. The goal is not to force every empirical summary to equal its target, but to determine whether the generated population represents the specified targets adequately for its intended use.

Specify the Target Before Generating Individuals

Begin by listing the traits and phenotypes required for the QSP application. For each continuous trait, specify units, plausible bounds, and the target marginal distribution or its available summaries. For each categorical phenotype, specify its target prevalence. If prevalence depends on a subgroup or on other traits, state those conditional targets as well. Finally, identify which associations must be retained, such as a correlation between two continuous traits or different phenotype prevalences across strata.

The target definitions must refer to a coherent population. For example, a body-weight distribution from one eligibility-defined population should not be combined uncritically with phenotype prevalences from a different population. If targets are reported as ranges rather than point values, use the interval-valued target as stated rather than silently replacing it with a midpoint.

Construction choices depend on the target information. A fully specified marginal distribution permits direct transformation; means and quantiles alone may not uniquely determine a distribution. Likewise, a target correlation is a feature of the observed traits, not automatically the correlation to use in a latent-variable generator.

Generate Bounded, Correlated Continuous Traits

One flexible approach is a Gaussian copula. First generate correlated standard normal latent variables, then transform each latent variable through the standard normal cumulative distribution function and the inverse target marginal distribution. If \(Z_j\) is a latent variable and \(F_j\) is the target cumulative distribution function for trait \(X_j\), define:

where \(\Phi\) is the standard normal cumulative distribution function. When the target marginal has bounded support, the transformed trait has that bounded support. The correlation matrix of the latent normal variables determines the copula dependence, but nonlinear marginal transformations generally change Pearson correlations. Therefore, a latent correlation should not be reported as though it were the resulting observed-scale correlation. Compare the generated observed-scale association with its target, and calibrate the latent dependence if needed.

A short justification clarifies the distinction between attained values and support. If \(Z\sim N(0,1)\), then \(\Phi(Z)\) is uniform on \((0,1)\). For \(X=L+(U-L)\Phi(Z)\), any interval \((L,L+\epsilon)\), with \(0<\epsilon<U-L\), has probability \(\epsilon/(U-L)>0\). The same is true near \(U\). Thus values arbitrarily close to either endpoint occur with positive probability, even though the exact endpoints have probability zero; the support is \([L,U]\), not merely the set of attained values \((L,U)\).

Assign Categorical Phenotypes Without Losing Subgroup Targets

Phenotypes can be assigned after generating continuous traits, provided the assignment mechanism respects the intended relationships. If a phenotype has one overall target prevalence and is meant to be independent of the traits, an assignment with that prevalence may be suitable. If its prevalence differs by stratum, use stratum-specific probabilities or allocations. If it depends on continuous traits, specify and implement a conditional assignment rule. In every case, check both the overall prevalence and the conditional prevalences that matter to the application.

Evaluate Fit, Diversity, and Subgroups

Qualification should examine the dimensions specified by the target, not a single convenient summary. Marginal checks can compare means, standard deviations, quantiles, or distributional distances. Joint checks can compare correlations or other specified dependence summaries. Categorical checks should include overall prevalence and relevant conditional prevalences. Diversity checks ask whether the population retains adequate variation and coverage rather than concentrating on a few individuals or leaving portions of the target support unrepresented.

A practical report distinguishes target values, generated values, differences, and acceptance criteria. For a continuous summary \(T\), a simple signed discrepancy is \(T_{\mathrm{generated}}-T_{\mathrm{target}}\); a scaled discrepancy can divide this difference by a prespecified scale, such as the target standard deviation. Such discrepancies describe fit but do not, by themselves, define an acceptable population. Set tolerances before examining the generated results, and explain why each diagnostic is relevant to the intended QSP use.

An End-to-End Construction Sequence

The following sequence keeps the construction choices and the qualification decision connected. The steps apply whether the final population is formed by direct generation, filtering, or resampling; the mechanics of those algorithms differ, but their outputs still require the same target-based checks.

Alternatives, Diagnostics, and Limits

Independent marginal sampling is simple and can be appropriate when independence is a defensible target. It is not a neutral default: it removes dependence, so it can produce unrealistic combinations even when every marginal fits. A copula construction is useful when target marginals and a dependence structure must be combined, but its dependence parameters may need calibration on the observed scale. Constraint filtering can enforce eligibility or other hard requirements, but may alter marginal distributions and subgroup representation; these must be rechecked after filtering. Weighted resampling can emphasize target-compatible individuals, but concentrated weights can reduce diversity. These approaches are alternatives for construction, not substitutes for qualification diagnostics.

Warning signs include trait values outside stated bounds, a marginal that fits while its relevant subgroup does not, an observed correlation inconsistent with its target, phenotype counts that look adequate overall but leave a subgroup poorly represented, and a narrow set of individuals carrying most of the population weight. The interpretation depends on the target and design: a small count in a rare but important phenotype may be a substantive coverage concern even when overall prevalence is close to target.

No finite population can match every feature of a continuous target distribution exactly. A diagnostic is informative only in relation to its sampling variability, target precision, and intended use. If the target is uncertain or only partially specified, apparent precision in a generated summary does not remove that uncertainty. Qualification also cannot rescue incompatible targets or establish validity for outcomes that were not part of the qualification criteria.

Common Pitfall

A common mistake is to set the latent-normal correlation equal to the desired correlation between bounded traits, then declare the relationship matched. That reasoning is generally wrong because nonlinear transformations change Pearson correlation. The correction is to calculate the correlation in the generated observed traits and compare it with the target; if necessary, adjust the latent dependence and regenerate. Similarly, matching an overall phenotype prevalence does not prove that conditional prevalences are correct. Report each targeted subgroup diagnostic separately.