From Observed Summaries To Population Targets

From a Reported Number to a Population Target

A published summary is not yet a complete specification for a virtual population. “Mean clearance: 42” does not say whether 42 is a sample mean or a known population value, what units apply, how many observations contributed, or how uncertain the estimate is. Those distinctions matter when the summary is used to guide a population of mechanistic parameter vectors: an unsupported narrow band can exclude plausible patients, while a very wide band may provide little constraint.

The task here is to translate reported summaries into explicit, traceable targets. It is useful to keep three items separate: the population quantity of interest, its estimate from observed data, and the tolerance or uncertainty range used in a decision. A confidence interval describes statistical uncertainty under stated assumptions. An operational tolerance band is a practical criterion chosen for a particular purpose; it may be informed by a confidence interval, but it is not automatically the same thing.

A reported \(n\) is the number of observations contributing to a summary, not necessarily the number of virtual patients to generate. It helps determine the uncertainty of the estimate, but does not by itself define the target. The target must also match the intended population: age range, disease status, treatment, measurement time, and other inclusion conditions can change the relevant summary.

Means and Their Uncertainty

Suppose \(x_1,\ldots,x_n\) are independent observations from a population with mean \(\mu\). Their sample mean is \(\bar{x}\), and their sample standard deviation is \(s\). If the population is normal, or if the sample is sufficiently large for the sampling distribution of \(\bar{x}\) to be adequately approximated by a normal distribution, a two-sided \(100(1-\alpha)\%\) confidence interval for \(\mu\) is

Here \(t_{1-\alpha/2,\ n-1}\) is the corresponding quantile of the \(t\) distribution with \(n-1\) degrees of freedom. The standard error \(s/\sqrt{n}\) quantifies the estimated sampling variability of the sample mean. This interval does not describe the spread of individual patients; that spread is represented by quantities such as \(s\), not by the standard error.

For a virtual population, the confidence interval can inform how much precision the observed mean supports. A separate operational rule might require the generated population mean to fall within a prespecified band. That band should be justified by the application, measurement precision, and intended population—not silently substituted for the confidence interval.

Variances: A Distributional Assumption Matters

A variance target describes population dispersion, in squared units. For a random sample from a normal population, \((n-1)s^2/\sigma^2\) has a chi-square distribution with \(n-1\) degrees of freedom. Inverting this result gives the following \(100(1-\alpha)\%\) confidence interval for \(\sigma^2\):

The lower endpoint uses the larger chi-square quantile. This interval depends strongly on the normality assumption, particularly for small samples. If the data are markedly skewed or contain influential outliers, the interval may not have its stated coverage.

The calculation illustrates why a target should include its scale and units. A variance in \((\mathrm{mg/L})^2\) cannot be compared directly with a standard deviation in mg/L. If a report gives only a standard deviation, square it to obtain a variance target; if uncertainty for that target is needed, the sampling assumptions and sample size are also required.

Proportions and Sample Size

For a binary characteristic, let \(p\) be the population proportion and \(\hat{p}=x/n\) the observed proportion, where \(x\) of \(n\) observations have the characteristic. When the sample is independent and both expected counts \(n\hat{p}\) and \(n(1-\hat{p})\) are sufficiently large, a normal-approximation interval is

This approximation is most defensible when neither category is rare. It can produce endpoints outside zero to one in small or highly imbalanced samples; in those settings, use an interval designed for proportions rather than treating impossible endpoints as meaningful. The sample size \(n\) is essential: the same observed percentage can imply quite different precision in a small and a large sample.

Quantiles Require Rank Information

A quantile target identifies a location in a distribution, such as the median or the 90th percentile. A reported quantile alone does not usually provide enough information to quantify its sampling uncertainty. That uncertainty depends on sample size and on the data near the quantile. With raw observations, a distribution-free interval for a population median can be obtained from order statistics, provided the observations are independent and drawn from a continuous population with a well-defined median.

For a sample of size \(n\), let \(X_{(1)}\leq\cdots\leq X_{(n)}\) be the sorted observations. If the population median is \(m\), the number \(B\) of observations below \(m\) follows a binomial distribution with \(n\) trials and success probability \(1/2\). Consequently, the interval \([X_{(k)},X_{(n-k+1)}]\) has coverage \(P(k\leq B\leq n-k)\). This gives a direct justification for rank-based uncertainty: the interval covers the median when the count on either side places it between the selected order statistics.

For other quantiles, the same principle uses the probability mass below the target quantile rather than one half. In practice, a report should provide a confidence interval or enough underlying data to construct one. A quantile point estimate should not be given an arbitrary symmetric tolerance and presented as statistically derived.

Building a Defensible Specification

A concise target record can include: the population or subgroup; the quantity and units; the estimate and sample size; the assumptions or method used for uncertainty; the interval, if available; and any separate operational acceptance band. This makes clear whether the range reflects sampling uncertainty, a scientific constraint, or a decision rule.

A confidence interval and an operational tolerance answer different questions. The interval asks how much uncertainty the observed sample leaves about a population quantity under a model. A tolerance band states what discrepancy is acceptable for a particular use. For example, a project may choose an allowed deviation from a mean because of measurement precision or a downstream decision threshold. The choice should be documented and justified; it is not licensed solely by the confidence level.

A common mistake is to treat a reported sample mean as the exact population target and then choose a narrow band around it without considering \(n\) or variability. The correction is to separate the point estimate from uncertainty: for a mean, for example, the estimated standard error is \(s/\sqrt{n}\), so precision depends on both spread and sample size. Another mistake is to read a 95% confidence interval as containing 95% of patients. It concerns uncertainty in a population summary; patient-level variation is a different feature of the distribution.

Summaries also have limits. Marginal means, variances, proportions, and quantiles do not determine the joint distribution of patient traits, nor do they guarantee physiological feasibility. A reported mean may conceal skewness or subgroups, and a sample may not represent the intended target population. When essential information is absent—such as sample size, units, or the definition of a subgroup—the honest specification records that limitation rather than manufacturing precision.