From Population Priors to Patient-Specific PK Distributions

From Population Variation to an Individual Distribution

A population pharmacokinetic model describes how parameters vary across people, but a dosing decision often concerns one person whose parameter values are not known exactly. Concentration observations provide information about that person; the population distribution provides information before those observations are considered. Combining the two gives a patient-specific distribution, not merely a single estimated parameter value.

The distinction matters because the population distribution and the patient-specific distribution answer different questions. The former describes plausible parameter values across a defined population. The latter describes uncertainty about one patient's parameter after incorporating that patient's data, conditional on the model and the information used to construct the prior.

Definition: A population prior is a probability distribution for an individual's PK parameter before using that individual's concentration observations, conditional on the chosen population model and any covariates already included. The likelihood describes how probable the observed data are at each candidate parameter value. The posterior distribution is proportional to the prior density multiplied by the likelihood, and is normalized to have total probability one.

The parameter is treated as unknown for the patient, rather than as a random value that changes during the observation period. A population model can represent variation between patients, while a posterior distribution represents uncertainty about the parameter for the particular patient under analysis. The posterior depends on the population distribution, the observation model, and the observed data; it is not an intrinsic property of the patient alone.

A One-Compartment Model and a Normal Update

Consider a one-compartment IV bolus model with known dose \(D\), known volume \(V\), and elimination rate \(k=CL/V\), measured in inverse time. Its concentration at time \(t\) is \(C(t)=D/V\exp(-kt)\). For a simple updating calculation, suppose measurement error is normal on the log-concentration scale:

\log C_{\mathrm{obs}}=\log(D/V)-kt+\epsilon,\qquad \epsilon\sim N(0,\sigma_e^2).

Here \(C_{\mathrm{obs}}\) is the measured concentration and \(\sigma_e\) is the standard deviation of log-scale measurement error. Rearranging gives a concentration-based estimate \(y\) of \(k\), with a corresponding standard error:

y=\frac{\log(D/V)-\log C_{\mathrm{obs}}}{t},\qquad y\mid k\sim N\left(k,\left(\frac{\sigma_e}{t}\right)^2\right).

This representation isolates the update algebra: the likelihood for \(k\) is normal, centered at \(y\). The transformation is used here for a known dose, known volume, and a single post-bolus observation; it is not a general recipe for constructing a concentration likelihood in other designs.

Suppose the population prior for \(k\) is normal with mean \(\mu_0\) and variance \(\tau_0^2\). For the moment, ignore the positivity boundary. Multiplying the prior density by the normal likelihood and completing the square gives a normal posterior:

k\mid y\sim N(\mu_1,\tau_1^2),\qquad \frac{1}{\tau_1^2}=\frac{1}{\tau_0^2}+\frac{1}{\tau_y^2},\qquad \mu_1=\tau_1^2\left(\frac{\mu_0}{\tau_0^2}+\frac{y}{\tau_y^2}\right),

where \(\tau_y=\sigma_e/t\). The posterior precision is the sum of prior precision and data precision. The posterior mean is a precision-weighted average: the more precise source of information contributes more. This result assumes the stated normal prior and normal likelihood, with known variances and conditionally independent measurement error as specified by the model.

Worked Example: One Concentration With a Positive Elimination Rate

Suppose a hypothetical patient receives \(D=500\) mg IV, has known \(V=50\) L, and has a concentration of approximately \(7.286\) mg/L measured at \(t=10\) h. Assume log-concentration error has standard deviation \(\sigma_e=0.200\). A population model supplies \(k\sim N(0.040,0.050^2)\), restricted to \(k>0\). Find the patient-specific posterior distribution.

First transform the observation to the normal likelihood scale. Since \(D/V=10\) mg/L,

y=\frac{\log(10/7.286)}{10}\approx 0.0316630\ \mathrm{h}^{-1}, \qquad \tau_y=\frac{0.200}{10}=0.0200\ \mathrm{h}^{-1}.

The normal update before enforcing positivity has precision \(1/0.050^2+1/0.0200^2=400+2500=2900\ \mathrm{h}^2\). Its variance is \(1/2900=0.00034483\ (\mathrm{h}^{-1})^2\), so its standard deviation is \(\tau_1=0.01857\ \mathrm{h}^{-1}\). The mean is

\mu_1=\frac{0.040(400)+0.0316630(2500)}{2900} =\frac{95.1575}{2900} =0.032813\ \mathrm{h}^{-1}.

Because \(k\) cannot be negative, the posterior is this normal distribution conditioned on \(k>0\). Its standardized lower boundary is \(z=(0-\mu_1)/\tau_1\approx-1.767\). For a normal variable with lower truncation at standardized value \(z\), define \(\lambda=\phi(z)/[1-\Phi(z)]\), where \(\phi\) and \(\Phi\) are the standard normal density and cumulative distribution function. Here \(\lambda\approx0.08711\). Thus the normalized posterior is a lower-truncated normal, and its mean and standard deviation are approximately

E(k\mid y,k>0)=\mu_1+\tau_1\lambda =0.032813+0.01857(0.08711) =0.03443\ \mathrm{h}^{-1},

SD(k\mid y,k>0)=\tau_1\sqrt{1+z\lambda-\lambda^2} \approx 0.01700\ \mathrm{h}^{-1}.

The truncation raises the mean slightly because negative values are excluded. It does not turn the posterior into a point estimate: it remains a distribution over positive elimination rates. The posterior probability assigned to an interval should be interpreted conditional on the stated model, prior, and observation.

Why the Truncation Adjustment Has This Form

The positivity adjustment follows directly from conditioning a normal distribution on being above a boundary. Let \(Z\sim N(0,1)\), and retain only values \(Z>z\). Since the standard normal density satisfies \(\phi'(u)=-u\phi(u)\),

E(Z\mid Z>z) =\frac{\int_z^\infty u\phi(u)\,du}{1-\Phi(z)} =\frac{\phi(z)}{1-\Phi(z)} =\lambda.

For an untruncated variable \(K=\mu_1+\tau_1 Z\), the condition \(K>0\) is equivalent to \(Z>-\mu_1/\tau_1\). Therefore its conditional mean is \(\mu_1+\tau_1\lambda\). The conditional variance is \(\tau_1^2[1+z\lambda-\lambda^2]\), obtained by integrating \(Z^2\phi(Z)\) over the same tail and subtracting the squared conditional mean. This establishes why the boundary changes both the mean and spread.

A normal prior truncated at zero, multiplied by a normal likelihood, produces the corresponding updated normal density truncated at zero: the likelihood-prior product is a normal kernel, and the parameter support remains \(k>0\). This convenience relies on the normal kernel and a fixed truncation boundary. Other parameterizations or prior distributions need not yield a truncated normal posterior.

Worked Example: Combining Two Concentration Observations

For a second hypothetical patient, assume known \(D\) and \(V\), and suppose two post-bolus measurements have already been converted using the stated model to \(y_1=0.100\ \mathrm{h}^{-1}\) with standard error \(0.030\ \mathrm{h}^{-1}\), and \(y_2=0.080\ \mathrm{h}^{-1}\) with standard error \(0.040\ \mathrm{h}^{-1}\). Take a population prior \(k\sim N(0.120,0.040^2)\), restricted to positive values. Assume the two errors are independent conditional on \(k\).

The combined likelihood precision is \(1/0.030^2+1/0.040^2=1111.111+625=1736.111\ \mathrm{h}^2\). Adding the prior precision \(1/0.040^2=625\ \mathrm{h}^2\) gives posterior precision \(2361.111\ \mathrm{h}^2\). Therefore

\mu_1=\frac{0.120(625)+0.100(1111.111)+0.080(625)}{2361.111} =\frac{236.111}{2361.111} =0.1000\ \mathrm{h}^{-1},

\tau_1=\sqrt{\frac{1}{2361.111}}=0.02058\ \mathrm{h}^{-1}.

The positive-boundary standardized value is \((0-0.1000)/0.02058=-4.86\), so truncation has a negligible numerical effect in this example. The posterior is approximately \(N(0.1000,0.02058^2)\) on the positive support. An approximate central 95% interval is \(0.1000\pm1.96(0.02058)\), or \(0.05966\) to \(0.1403\ \mathrm{h}^{-1}\). Under this model, the interval contains 95% of the posterior probability; it is not a claim that the fixed parameter has a 95% chance of changing into the interval.

Prior, Likelihood, and Posterior Need Not Be Normal

Normal conjugacy is useful because it gives a compact formula, but the basic updating principle is broader. For a continuous parameter \(k\), write the prior density as \(p(k)\) and the likelihood of observed data \(y\) as \(L(k;y)\). Then

p(k\mid y)=\frac{L(k;y)p(k)}{\int L(u;y)p(u)\,du}.

The denominator is the normalizing constant: it integrates the unnormalized product over the allowed parameter values. This expression clarifies the roles. The prior represents population information before the patient's observations; the likelihood compares candidate parameter values using those observations; the posterior is the updated, normalized distribution. The likelihood is not itself a probability distribution over \(k\), because its integral over parameter values need not equal one.

Worked Example: Updating a Two-Component Population Prior

Suppose, for illustration, a population prior places probability \(0.60\) at \(k=0.050\ \mathrm{h}^{-1}\) and probability \(0.40\) at \(k=0.150\ \mathrm{h}^{-1}\). A hypothetical concentration observation yields \(y=0.080\ \mathrm{h}^{-1}\) with normal likelihood standard deviation \(0.030\ \mathrm{h}^{-1}\). Find the posterior probabilities for the two values.

The likelihood densities share the same normalizing constant, so their relative likelihoods are \(\exp[-(y-k)^2/(2(0.030)^2)]\). At \(k=0.050\), the standardized difference is \(1\), giving relative likelihood \(\exp(-0.5)=0.6065\). At \(k=0.150\), it is \(-2.333\), giving relative likelihood \(\exp[-(2.333)^2/2]\approx0.06573\). Multiplying by the prior probabilities gives unnormalized posterior masses

0.60(0.6065)=0.3639,\qquad 0.40(0.06573)=0.02629.

Their sum is \(0.3902\). Dividing each mass by that sum gives posterior probabilities of approximately \(0.9326\) at \(0.050\ \mathrm{h}^{-1}\) and \(0.0674\) at \(0.150\ \mathrm{h}^{-1}\). The lower-rate value becomes more probable because it is more compatible with the observed \(y\), while the higher-rate value retains some probability because it had prior support and its likelihood is not zero. This discrete example illustrates Bayes' rule; it is not an assertion that real PK parameters occur at only two values.

Diagnostics, Alternatives, and Limitations

Before interpreting an individual posterior, check the assumptions that produce it. In the one-compartment calculation, verify that dose and sampling time are known, volume is treated as known, the post-bolus concentration model is suitable, and the stated log-scale error model is plausible. For multiple observations, conditional independence must be justified or replaced by an error model that accounts for dependence. Check units at each step: \(y\), \(k\), and their standard errors have units of inverse time, while log concentrations are dimensionless.

A useful diagnostic is whether observed concentrations are broadly compatible with the concentration range implied by values receiving substantial prior probability. A posterior that shifts sharply toward a support boundary, or stays concentrated near the prior despite observations that appear inconsistent, calls for checking the model, data, and prior rather than automatic acceptance of the result. A normal approximation can also place appreciable probability on impossible negative rates; truncation addresses support, but does not guarantee that the normal shape is appropriate.

A plug-in alternative might use the population mean for every patient or report only the likelihood-based estimate from that patient's data. The population mean ignores the patient's observations; an individual-only estimate can be unstable with sparse or noisy data. A posterior distribution combines the two sources and retains uncertainty, but depends on both being adequately specified. A point summary, such as the posterior mean, cannot show asymmetry, multiple modes, or tail probability by itself.

The normal-normal formula is limited to a normal prior kernel and a normal observation model with known variance. Nonlinear concentration models, uncertain volume, multiple correlated PK parameters, or non-normal residual errors may require numerical integration or other computational methods. A posterior is conditional on the chosen model: it does not correct an inaccurate dose history, an omitted covariate effect, or a misspecified structural model. Population information must also represent a relevant target population; a poorly matched prior can pull an individual distribution in an inappropriate direction.

Common mistake: “The population mean is the patient's parameter, so I can use it as the patient-specific distribution.” The population mean is a single summary, not the distribution and not the posterior. A patient-specific posterior requires a population prior distribution and the patient's likelihood; using only the mean discards both prior spread and the information in that patient's observations.

Check Your Understanding

Answer each question, then compare with the solutions below.

In a Bayesian update, which quantity represents population information before the patient's concentration data are used: prior, likelihood, or posterior?
A one-compartment observation gives \(y=0.060\ \mathrm{h}^{-1}\) with standard error \(0.020\ \mathrm{h}^{-1}\). The prior is \(N(0.080,0.040^2)\), before any positivity truncation. Calculate the normal posterior mean and standard deviation.
Why does conditioning on \(k>0\) raise the posterior mean in Worked Example 1?
True or false: the likelihood is always a probability distribution over parameter values that integrates to one. Explain.

Solutions. 1. The prior. 2. The prior precision is \(625\), the data precision is \(2500\), and total precision is \(3125\ \mathrm{h}^2\). Thus the posterior variance is \(1/3125=0.0003200\ (\mathrm{h}^{-1})^2\), the standard deviation is \(0.01789\ \mathrm{h}^{-1}\), and the mean is \([0.080(625)+0.060(2500)]/3125=0.06400\ \mathrm{h}^{-1}\). 3. The untruncated normal assigns some probability below zero; conditioning removes those values, shifting the remaining distribution's mean upward. 4. False. The likelihood is a function of the parameter for fixed data and need not integrate to one over parameter values; multiplying by a prior and normalizing produces the posterior distribution.