From an Individual Posterior to a Future Concentration
An individual’s posterior PK parameter distribution summarizes what is known about that individual after combining prior information with observed concentration data. A forecast asks a different question: what future concentration could be observed, given the data available now and the dosing history specified for the forecast? A single curve evaluated at a single parameter estimate does not answer this fully. It gives a conditional model prediction, but does not represent the uncertainty in the individual’s parameters or the variability of a future measurement.
A posterior predictive concentration forecast combines these sources of uncertainty. For a future observation, it averages the observation model over the individual’s posterior distribution. The forecast therefore describes a distribution, not just one predicted concentration. Its center may be summarized by a mean or median, and its spread by prediction intervals or probabilities of exceeding a specified concentration.
The dosing history and forecast time matter because they determine the model prediction. They must be specified before generating a forecast. The posterior is conditional on the observations used for updating; the forecast conditions describe what is assumed to happen after those observations. If the future dose history is uncertain, treating it as fixed will not account for that uncertainty.
Latent Concentrations and Future Measurements
Let \(C^*\) denote the latent concentration predicted by the PK model at the future time, before residual measurement or observation variability is added. Conditional on \(\theta\) and the dosing history, the PK model may give \(C^*=f(\theta,a^*)\). A future measured concentration \(Y^*\) is then generated from a residual error model, for example \(Y^*=C^*+\epsilon^*\) under an additive model. The distinction matters: uncertainty about \(C^*\) comes from uncertainty in the PK parameters, while the observation distribution can add residual variability.
In a log-additive model, a positive measurement might instead be represented by \(\log Y^*=\log C^*+\epsilon^*\). The appropriate model depends on the residual error specification used for the data. A forecast should use that specified model rather than choosing an error distribution solely because it is convenient for calculation.
The standard construction assumes the posterior represents uncertainty about the parameters conditional on the observed data, the PK model and the assumed dosing record; the future residual error is generated from the stated error model; and any dependence structure required by the model is preserved. For example, if residual errors across future samples are correlated, they should be generated jointly rather than independently. A forecast also assumes the model is applicable to the specified future conditions. Posterior averaging does not repair a misspecified PK model or an incorrect dose history.
Worked Example: Averaging Future Concentration Exceedance Probabilities
Suppose two equally weighted draws from a hypothetical individual posterior produce latent concentrations of 4.8675 and 4.5735 mg/L at a specified future time. For transparency, these are the unrounded outputs of a one-compartment IV bolus calculation, \(C^*=(D/V)\exp(-CLt/V)\), under the parameter values in the corresponding posterior draws. Assume a future measured concentration follows an additive normal error model, \(Y^*\mid C^*\sim N(C^*,0.5^2)\), with residual standard deviation 0.5 mg/L. Calculate the posterior predictive probability that the future measurement exceeds 4.5 mg/L.
For each draw, calculate the conditional probability using the normal model. The first draw gives $$ P(Y^*>4.5\mid C^*=4.8675)=1-\Phi\left(\frac{4.5-4.8675}{0.5}\right) =1-\Phi(-0.7350)\approx 0.7688. $$ The second draw gives $$ P(Y^*>4.5\mid C^*=4.5735)=1-\Phi\left(\frac{4.5-4.5735}{0.5}\right) =1-\Phi(-0.1470)\approx 0.5584. $$
Average the conditional probabilities using the posterior draw weights: $$ P(Y^*>4.5\mid D,a^*)\approx \frac{0.7688+0.5584}{2}=0.6636. $$ Thus, under these hypothetical posterior draws and the stated residual model, the posterior predictive probability of a future measured concentration above 4.5 mg/L is approximately 0.664. This is a probability about the future measurement under the model and posterior, not the probability that a fixed unknown parameter exceeds a threshold.
Why Posterior Averaging Is Required
For any particular parameter value, the observation model supplies a conditional predictive distribution. But the parameter is not known exactly: its posterior distribution assigns weight to multiple plausible values. The probability of a future event must therefore average its conditional probability over that posterior. If \(A\) is an event such as \(Y^*>c\), then
This follows by integrating the joint conditional distribution over \(\theta\): \(p(y^*,\theta\mid D,a^*)=p(y^*\mid\theta,D,a^*)p(\theta\mid D)\). Integrating both sides with respect to \(\theta\) yields the posterior predictive density. The same identity applies to event probabilities by integrating over the values \(y^*\in A\).
A useful implication is that the predictive variance has two components. If the relevant moments exist, the law of total variance gives
The first term averages the conditional future observation variance, including residual variability. The second describes variation in the conditional mean across posterior parameter values. Omitting either term generally changes the forecast spread. This decomposition is a useful justification for generating both posterior parameter draws and future residual errors.
Worked Example: Forecasting With a Normal Posterior Predictive Distribution
Suppose, for a hypothetical individual and fixed future conditions, the model’s conditional mean concentration is \(m(\theta)\), and the posterior distribution of that mean is \(m(\theta)\mid D\sim N(5.0,0.6^2)\) mg/L. Suppose future additive residual error is independent of \(\theta\) and follows \(N(0,0.4^2)\) in units of mg/L. Thus \(Y^*=m(\theta)+\epsilon^*\). Find the predictive distribution and the probability that \(Y^*>5.5\) mg/L.
The sum of independent normal variables is normal. Its mean is the sum of the means, and its variance is the sum of the variances: $$ Y^*\mid D,a^*\sim N(5.0,\,0.6^2+0.4^2)=N(5.0,0.52). $$ The second argument is the variance, in \((\text{mg/L})^2\). Therefore the predictive standard deviation is \(\sqrt{0.52}=0.7211\) mg/L. The exceedance probability is $$ P(Y^*>5.5\mid D,a^*)=1-\Phi\left(\frac{5.5-5.0}{0.7211}\right) =1-\Phi(0.6934)\approx 0.244. $$
Under the stated assumptions, the forecast probability of a future measured concentration above 5.5 mg/L is about 0.244. The posterior spread in the conditional mean and the future residual spread both contribute to the predictive variance. This normal result depends on the assumed normal distributions, independence, and additive error structure; it is not a general formula for every PK model.
Generating Forecasts in Practice
For nonlinear PK models, the integral defining the posterior predictive distribution usually has no simple closed form. Posterior simulation provides a direct approximation. For each replicate, draw an individual parameter vector from the posterior, evaluate the PK model at the specified forecast time and dosing history, then draw a future observation from the residual model. The resulting simulated values approximate the posterior predictive distribution.
Set the future time, dose history and any other conditions. State whether the target is a latent model concentration or a future measured concentration.
For each replicate, sample the relevant individual parameters jointly from their posterior. Preserve posterior dependence rather than sampling each parameter independently.
Use each draw and the specified conditions to compute the corresponding latent concentration. Do not replace the draws with a single posterior point estimate if a predictive distribution is required.
For measured-concentration forecasts, draw residual error from the specified observation model, including any uncertain residual parameters or required dependence among samples.
Use the simulated distribution to estimate its mean or median, prediction interval, or probability of an event. Report the summary and its concentration units or event definition.
The empirical proportion of simulated values above a threshold estimates the posterior predictive probability of exceeding that threshold. With \(M\) independent predictive replicates, this estimate is the number above the threshold divided by \(M\). Monte Carlo error remains: increasing \(M\) can reduce simulation noise, but cannot correct a poor posterior, a wrong residual model, or misspecified forecast conditions.
Worked Example: A Lognormal Forecast and an Exceedance Probability
Consider a hypothetical positive concentration forecast. Suppose posterior uncertainty in the log of the model-predicted concentration is \(N(\log 7,0.12^2)\), and the future log-scale residual error is independent and \(N(0,0.16^2)\). Concentrations are measured in mg/L. Find the posterior predictive median and the probability that the future measurement exceeds 6 mg/L.
The future log concentration is the sum of the uncertain log model prediction and the residual error. Independence gives $$ \log Y^*\mid D,a^*\sim N\left(\log 7,\,0.12^2+0.16^2\right) =N(\log 7,0.20^2). $$ The median of a lognormal distribution is the exponential of its log-scale mean, so the predictive median is \(\exp(\log 7)=7.00\) mg/L.
For the exceedance probability, standardize the log threshold: $$ P(Y^*>6\mid D,a^*)=1-\Phi\left(\frac{\log 6-\log 7}{0.20}\right) =1-\Phi(-0.7708)\approx 0.780. $$ Thus, under the assumed lognormal forecast model, the probability of a future measured concentration exceeding 6 mg/L is approximately 0.780. This probability includes both the posterior spread in the log model prediction and the future log-scale residual variability.
Diagnostics, Alternatives, and Limitations
Before interpreting a forecast, check that its inputs and simulated distribution behave as intended. Confirm that the future time and dose record match the question, that units are consistent, and that posterior draws are taken jointly from the intended individual posterior. Compare latent concentration draws with measured-concentration draws: the latter should reflect the residual model. Inspect simulated values for implausible negative concentrations under an additive model, extreme tails, or unexpected discontinuities. These can indicate an unsuitable error model, a coding error, or a model whose assumptions do not fit the application.
A common mistake is to evaluate the PK model at the posterior mean parameter and call that value the posterior predictive concentration. This is generally only a plug-in conditional prediction: for a nonlinear model, \(f(E[\theta\mid D])\) need not equal \(E[f(\theta)\mid D]\), and it has no future residual variability. The correction is to evaluate the model over posterior draws and, for a future measured concentration, generate the corresponding residual variability as well. Even a forecast mean can differ from the plug-in value, and a plug-in value alone cannot provide a predictive interval.
A latent-concentration forecast is a reasonable alternative when the target is the model-implied concentration trajectory rather than a future assay result. It should not be presented as a prediction interval for a future measured concentration. Conversely, adding residual error is necessary when predicting future observations under the observation model. A point estimate, such as the posterior mean prediction, can be useful as a compact summary, but it is not a replacement for the predictive distribution when uncertainty or threshold probabilities matter.
Posterior predictive forecasts remain conditional on the model, prior, observed data and specified future conditions. They do not automatically include uncertainty about model structure, an unrecorded future dose, or a change in physiology not represented by the PK model. Extreme or multimodal posterior predictions may make a single mean misleading; report quantiles or the distribution’s shape when needed. Monte Carlo summaries should also be checked for adequate simulation precision, particularly for probabilities in the tails.
Check Your Understanding
Answer each question using the distinction between posterior parameter uncertainty and future observation variability.
- In one sentence, define the posterior predictive distribution for a future measured concentration.
- A model gives a latent concentration prediction of 8 mg/L for every posterior draw, while the residual standard deviation is 1 mg/L under an additive normal model. Is the future measured concentration fixed at 8 mg/L? Explain.
- If the posterior standard deviation of a conditional mean is 0.3 mg/L and independent additive residual error has standard deviation 0.4 mg/L, what is the predictive standard deviation?
- Why can evaluating a nonlinear PK model at the posterior mean parameters fail to give the posterior predictive mean?
Solutions. (1) It is the distribution of a future measured concentration obtained by integrating its conditional observation distribution over the individual’s posterior parameter distribution. (2) No. The latent prediction is 8 mg/L, but a future measurement is \(Y^*=8+\epsilon^*\), with \(\epsilon^*\sim N(0,1^2)\) under the stated model. (3) The predictive variance is \(0.3^2+0.4^2=0.25\) \((\text{mg/L})^2\), so the predictive standard deviation is \(0.5\) mg/L. (4) In general, a nonlinear function does not commute with averaging: \(f(E[\theta\mid D])\) need not equal \(E[f(\theta)\mid D]\). Posterior draws must be passed through the PK model to represent the distribution of predictions.