Why Define the Target Covariate Population?
A simulated prediction is only as relevant as the population it represents. A model may produce a precise exposure prediction for each virtual individual, yet an average over the wrong covariate mix can answer a different question from the one intended. Defining the target covariate population makes explicit whose covariates the simulation is meant to represent and why.
This is a specification task, not a choice of a sampling algorithm. The definition should precede decisions about how covariate records will be generated. In particular, the target should not be identified automatically with the dataset used to fit the model: that dataset may be a selected sample, may have been collected in a different setting, or may not cover the population for which predictions are needed.
The target may be defined by a calendar period, care setting, geographic region, treatment context, or other features that determine who the prediction is about. The covariate vector should include the characteristics required by the prediction model and the question. For example, if a model’s prediction depends on age, body weight, and renal function, the target description must clarify the definitions and units of those variables, not just say “adult patients.”
Four Parts of a Useful Specification
A reproducible target definition answers four related questions. The answers should be concrete enough that another analyst can decide whether an individual or record belongs in the target and can determine what covariate distribution the simulation is intended to represent.
- Prediction purpose and population. State the quantity to be predicted and the people or settings to which it applies. Distinguish, for example, a prediction for eligible patients beginning a regimen from a prediction for all patients currently receiving it.
- Data source and time frame. Identify the source used to describe target covariates, its setting, and the period represented. State whether it is intended as a census, a sample, a registry, or another source. A source is evidence about the target; it is not, by itself, the target definition.
- Inclusion and exclusion criteria. Specify eligibility in operational terms. Include definitions of key variables, measurement timing, and any exclusions. Rules should be applied consistently and should match the intended population rather than merely making a dataset easier to analyze.
- Subgroups and their target proportions. Name any subgroups that matter for the prediction and state the proportion each contributes to the target. Explain the basis for those proportions. Observed source proportions are not automatically target proportions.
These choices can change a population-level prediction even when the individual-level model remains unchanged. Let \(m(\mathbf{c};\theta)\) denote the model prediction for covariates \(\mathbf{c}\) at parameter values \(\theta\). A target-average prediction is
where the expectation is taken over the specified target covariate distribution. This quantity answers a population-level question; an individual prediction at one covariate value does not. If the prediction also averages over individual random effects or another source of variation, that averaging must be stated separately. The covariate target alone does not define those other simulation choices.
Worked Example: Writing an Operational Target Definition
Hypothetical scenario. A team needs exposure predictions for adults starting a specified oral regimen in outpatient care. A regional registry contains 1,200 records from the intended region and calendar period. The prediction model uses age, body weight, and baseline renal function. The intended target excludes patients younger than 18 years and patients receiving dialysis.
A defensible specification begins with the prediction purpose: represent adults who start this regimen in the region during the stated period, not every person in the registry and not every patient treated in a hospital setting. The source is the regional registry, with its collection dates and outpatient setting recorded. Eligibility is age at treatment start of at least 18 years and no dialysis at treatment start. The team also defines how baseline renal function is measured and its units, and how body weight and age are recorded.
Suppose 80 of the 1,200 source records fail those eligibility criteria. The eligible source count is \(1{,}200-80=1{,}120\). This count describes the filtered source records; it does not establish that the eligible records are representative of every target patient. The specification should state whether any further exclusions, such as missing baseline measurements, are applied and how that choice affects the population being represented.
The specification is not complete until the team states how the intended prediction use determines subgroup representation. If predictions are meant to describe the eligible regional population, subgroup proportions should be justified as estimates of that population’s composition. If the aim instead concerns a particular subgroup, that subgroup must be named as the target rather than silently retaining the full-population mix.
Subgroup Mixtures and Their Proportions
Suppose the target population is partitioned into mutually exclusive, collectively exhaustive subgroups \(G=1,\ldots,K\), with target proportions \(\pi_g=P_T(G=g)\). For a subgroup with positive target proportion, its subgroup-specific covariate distribution is \(P_T(\mathbf{C}\mid G=g)\). The overall target distribution is a mixture of these subgroup distributions.
Here \(A\) is any set of covariate values. The proportions satisfy \(\pi_g\geq 0\) and \(\sum_{g=1}^{K}\pi_g=1\). The sum is explicitly restricted to positive-proportion groups because conditional probability given an event of probability zero is not defined by the usual ratio definition. Equivalently, without conditional probabilities, the identity can be written for all listed groups as
This follows because the events \(\{G=g\}\) partition the population: the joint events \(\{\mathbf{C}\in A,G=g\}\) are disjoint, and their union is \(\{\mathbf{C}\in A\}\). For each \(g\) with \(\pi_g>0\), the multiplication rule gives \(P_T(\mathbf{C}\in A,G=g)=P_T(\mathbf{C}\in A\mid G=g)\pi_g\). A group with \(\pi_g=0\) contributes zero to the joint-probability sum and is omitted from the conditional-probability sum.
For a target-average prediction, the same partition gives
Thus subgroup-specific prediction averages and subgroup proportions both matter. The subgroup proportions must describe the intended target, not merely the available source records. If a target subgroup has zero proportion, it should not be treated as a subgroup to be averaged into the target. If the question is specifically about that group, the target definition must instead be changed to make that group the population of interest.
Worked Example: How Target Proportions Change an Average
Hypothetical values. Suppose a model gives mean predicted exposure of \(8.0\) mg·h/L for a younger-adult subgroup and \(11.0\) mg·h/L for an older-adult subgroup. These are subgroup-average predictions under the same specified parameter values. The target population is defined to be 70% younger adults and 30% older adults.
The target-average prediction is the weighted sum:
The two contributions are \(5.60\) mg·h/L and \(3.30\) mg·h/L, respectively, so
If the source dataset were 50% younger and 50% older, averaging with those source proportions instead would give \((0.50)(8.0)+(0.50)(11.0)=9.50\) mg·h/L. That is a different weighted prediction, not an arithmetic error. It answers the target question only if the 50:50 mix is justified for the intended population. The calculation does not show that age causes the difference between subgroup predictions; it shows how the specified mixture affects a model-based population average.
Source Data, Selection, and Evidence for the Target
A source dataset can inform a target covariate distribution only to the extent that its records align with the target definition. Before using source summaries as target descriptions, examine how individuals entered the source, which settings and dates it covers, and whether eligibility and measurements match the prediction purpose. A hospital registry, for example, may overrepresent patients with more severe illness relative to a target defined as all outpatient treatment initiators.
When source and target compositions differ, the specification should say which proportions define the target and why. Those proportions might come from a population summary, an external source, or an explicit scenario assumption. Their provenance and uncertainty should be recorded. A scenario assumption is legitimate when clearly labeled, but it should not be presented as an observed population fact.
Missing covariates also affect what population has actually been described. If individuals with missing renal function are excluded, the resulting records represent those with observed renal function unless a defensible approach establishes otherwise. The target description should therefore report the number and characteristics of excluded records and whether the intended prediction population includes them. Silent complete-case restriction can redefine the target without acknowledgment.
The intended prediction use also determines whether a broad target is appropriate. A population-average exposure prediction for a region and a prediction for a narrowly defined subgroup are different targets. Predictions should not be generalized beyond the population and covariate support represented by the model without a separate justification. Defining a target does not ensure that the model is valid for it.
Worked Example: A Zero-Proportion Subgroup
Hypothetical scenario. A simulation specification lists two categories of renal-function status: “not severely impaired” and “severely impaired.” The target is explicitly a population of patients without severe impairment. Therefore, the target proportions are \(\pi_1=1.00\) for the first group and \(\pi_2=0.00\) for the second.
For any covariate event \(A\), the joint-probability form gives
Because \(P_T(G=2)=0\), the joint event \(\{\mathbf{C}\in A,G=2\}\) has probability zero. The target distribution is therefore determined entirely by the first group. The conditional form is correctly written as
There is no term involving \(P_T(\mathbf{C}\in A\mid G=2)\), because conditioning on a zero-probability target subgroup is undefined under the usual definition. If the scientific question instead asks for predictions in the severely impaired group, this target is unsuitable: the target definition must be revised, and the model’s ability to support that prediction must be assessed separately.
Diagnostics, Alternatives, and Limitations
A target definition is going wrong when its population cannot be reconstructed from the stated rules, its data source does not match its setting or time frame, or its proportions have no stated interpretation. A short review can expose many such failures:
- Ambiguous eligibility: phrases such as “typical patients” do not establish operational inclusion rules.
- Unexplained proportions: source frequencies are used as target proportions without evidence that the source mix represents the intended population.
- Incomplete subgroup partition: groups overlap, omit eligible individuals, or have proportions that do not sum to one.
- Inconsistent variable definitions: a covariate is measured at different times or in different units across the stated source and prediction purpose.
- Unacknowledged exclusions: filtering for complete covariate records silently changes who is represented.
- Unsupported prediction use: the target includes covariate patterns or settings for which the model provides no adequate basis for prediction.
A useful comparison is between defining a target by an eligible population and defining it by a set of explicit subgroup proportions. The first is often easier to explain when the prediction question is tied to a clearly described population and credible population data are available. The second makes the intended mixture transparent, especially when the prediction is deliberately standardized to specified subgroup shares or when a scenario is being explored. These are complementary ways of expressing the target, not competing statistical estimators.
Either approach has limitations. Eligibility criteria can be too broad to identify a meaningful population or too narrow to match actual use. Proportions can be uncertain, time-dependent, or based on sources that differ from the target. A single set of proportions can conceal meaningful variation within subgroups. In addition, a well-defined covariate target cannot repair model misspecification, unreliable parameter estimates, or an unsupported extrapolation. The target defines the intended population average; it does not validate the prediction model.
A common mistake is to say, “The dataset contains 40% older adults, so the target population is 40% older adults.” That reasoning confuses the observed source composition with the target composition. The correction is to state why the source represents the target, or to specify and justify a different target proportion. The same source can support different targets for different prediction purposes.
Check Your Understanding
Answer each question, then compare with the solutions.
- A source dataset contains 45% patients aged 65 years or older. What additional justification is needed before using 0.45 as the target proportion?
- A target is divided into mutually exclusive subgroups with proportions 0.25, 0.50, and 0.25. The subgroup-average predictions are 6, 8, and 10 mg·h/L. Find the target-average prediction.
- Why should a subgroup with target proportion zero be omitted from the conditional-probability mixture formula?
- A covariate is missing for 12% of eligible source records, and those records are excluded. What should the target specification report or reconsider?
- Explain why a target covariate definition does not, by itself, establish that model predictions are valid for that population.
Solutions. (1) Establish that the source’s age composition represents the intended target population and prediction setting, or justify the proportion using another source or an explicitly labeled scenario assumption. (2) The weighted average is \((0.25)(6)+(0.50)(8)+(0.25)(10)=1.5+4.0+2.5=8.0\) mg·h/L. (3) The usual conditional distribution given a subgroup is defined only when that subgroup has positive probability; use the joint-probability identity for all groups or sum conditional terms only over positive-proportion groups. (4) Report the exclusion and its effect on the represented population, examine whether missingness changes who is represented, and clarify whether the intended target includes those individuals. (5) Defining the target says who the prediction is meant to describe; model validity, parameter adequacy, and support for the target covariates require separate assessment.