Why the Choice of Point Estimate Matters
After combining a patient’s concentration data with a population-based prior, the result is a patient-specific posterior distribution for individual PK parameters. A single number can be convenient for reporting or calculation, but there is no universally best way to reduce that distribution to one value. The maximum a posteriori (MAP) estimate, posterior mean, and posterior median summarize different features and answer different decision questions.
The distinction matters especially when a posterior is skewed or a PK prediction is a nonlinear function of its parameters. For example, a forecast calculated at mean clearance need not equal the posterior-average forecast. The appropriate summary depends on the target of the decision, the loss attached to errors, and—in the MAP case—the parameterization used to describe the posterior.
These definitions do not make the three summaries interchangeable. The posterior mean balances squared deviations, the median balances absolute deviations, and the MAP identifies the highest point of a density or mass function. A mode describes where the posterior is most concentrated locally; it does not generally describe the center or an average.
Which Summary Fits Which Objective?
A point estimate is meaningful only in relation to what an error costs. Under squared-error loss, \(L(a,\theta_i)=(a-\theta_i)^2\), the posterior expected loss for choosing \(a\) is minimized by the posterior mean, provided the posterior second moment is finite. Under absolute-error loss, \(L(a,\theta_i)=|a-\theta_i|\), a posterior median minimizes expected loss. The MAP is often reported as a maximum-density summary, but it is not generally the estimate that minimizes either of these losses.
The finite-second-moment condition matters: without it, this expansion does not establish a finite squared-error risk or its minimizer. Under the condition, write \(\mu_i=E(\theta_i\mid y_i)\) and complete the square:
The final two terms do not depend on \(a\), while \((a-\mu_i)^2\) is minimized at \(a=\mu_i\). Thus the posterior mean minimizes posterior expected squared error when the second moment is finite. When \(E(|\theta_i|\mid y_i)\) is finite, a posterior median minimizes expected absolute loss; moving a proposed value toward a median reduces risk when the proposed value is not a median. Without this condition, expected absolute loss may be infinite for every finite choice. If there is a range of medians, each minimizes that loss.
Worked Example: Three Summaries From a Discrete Clearance Posterior
Suppose a hypothetical patient has the following posterior distribution for clearance, \(CL_i\), after the concentration data have been combined with the prior:
| Clearance (L/h) | Posterior probability |
|---|---|
| 2.00 | 0.20 |
| 4.00 | 0.50 |
| 10.0 | 0.30 |
The posterior mean is the probability-weighted average:
The MAP is \(4.00\ \text{L/h}\), because that value has the largest posterior probability, 0.50. The posterior median is also \(4.00\ \text{L/h}\): the cumulative probability through 2.00 L/h is 0.20, and through 4.00 L/h it is 0.70, so 4.00 is a value with at least half the probability on either side. Under squared-error loss, the mean is the appropriate summary; under absolute-error loss, the median is appropriate. The MAP answers the different question of which listed value has the largest posterior mass.
Parameter Summaries and PK Forecasts
Forecasting often applies a function \(g\) to an individual parameter. Two distinct quantities are then available: the forecast at a chosen parameter summary, \(g(E(\theta_i\mid y_i))\), and the posterior-average forecast, \(E\{g(\theta_i)\mid y_i\}\). In general, these are not equal. They coincide for an affine function, but not necessarily for a nonlinear one.
For example, under a one-compartment IV bolus model with fixed volume \(V\), the concentration at time \(t\) is \(C(t)=C_0\exp(-CL_i t/V)\). Clearance enters the exponential, so averaging clearance first and then calculating concentration does not generally give the average concentration across the posterior. If the target is expected concentration under squared-error loss, the relevant quantity is the posterior-average model concentration, not automatically the concentration at the posterior mean clearance.
Worked Example: Average Forecast Versus Forecast at Mean Clearance
Use the hypothetical posterior in the previous example. Suppose a one-compartment IV bolus forecast has \(C_0=8.00\ \text{mg/L}\) and \(t/V=0.125\ \text{h/L}\). Thus, at clearance \(CL_i\), the model concentration is \(8.00\exp(-0.125CL_i)\ \text{mg/L}\). The same formula and fixed \(C_0,t,V\) apply at each posterior clearance value.
First calculate the forecast at each value, keeping unrounded values for the weighted average:
The posterior-average model concentration is:
By contrast, the forecast at the posterior mean clearance of \(5.40\ \text{L/h}\) is:
The posterior-average model concentration exceeds the forecast at mean clearance by about \(4.359815-4.073251=0.286564\ \text{mg/L}\), or \(0.287\ \text{mg/L}\) to three significant figures. This difference arises from the curvature of the exponential relationship and the asymmetry of the clearance distribution. The forecast at the MAP or median clearance, \(4.00\ \text{L/h}\), is about \(4.852\ \text{mg/L}\); it is another distinct summary, not the posterior-average forecast.
Why Parameterization Changes a MAP
The posterior mean and median describe a parameter on its stated scale, with the mean requiring an appropriate moment. The MAP, however, is the maximum of a density, and a density changes with the coordinate used to express the parameter. A nonlinear transformation introduces a Jacobian factor. Therefore, maximizing a posterior density for clearance is not generally equivalent to maximizing the density for log-clearance and then transforming that mode back.
Worked Example: Log-Clearance and the Clearance-Scale MAP
Suppose, hypothetically, the posterior for \(\eta_i=\log(CL_i)\) is normal with mean \(\mu=\log(4.00)\) and variance \(\sigma^2=0.250\). Since \(CL_i=e^{\eta_i}\), the posterior median clearance is \(e^\mu=4.00\ \text{L/h}\). The posterior mean is:
The density of \(\eta_i\) is maximized at \(\eta_i=\mu\), so transforming that log-scale mode gives \(4.00\ \text{L/h}\). To find the mode of the density for clearance itself, include the transformation Jacobian:
Ignoring constants, its log density is \(-\log c-(\log c-\mu)^2/(2\sigma^2)\). Differentiating with respect to \(c\) and setting the derivative to zero gives:
Thus the clearance-scale MAP is about \(3.115\ \text{L/h}\), while the transformed log-scale mode is \(4.00\ \text{L/h}\). For the same hypothetical forecast function as above, the forecast at the clearance-scale MAP is:
The forecast at the posterior median clearance is \(8.00e^{-0.500}\approx4.852\ \text{mg/L}\), and at the posterior mean clearance it is \(8.00e^{-0.125(4.533)}\approx4.540\ \text{mg/L}\). These are forecasts at three different clearance summaries; none should be called the posterior-average forecast without averaging the forecast function over the posterior distribution.
Diagnostics, Alternatives, and Limitations
Before reporting a point estimate or using it in a forecast, check that the summary is defined and that it corresponds to the stated target. A posterior mean requires a finite first moment; a squared-error justification requires a finite second moment. A median can remain well-defined even when a mean does not exist. A MAP requires a proper density or mass function with an attained maximum; continuous densities can have boundary or multiple modes, and the reported mode can be sensitive to parameterization or numerical approximation.
For individual PK applications, also check whether the posterior summary is being reported on the original parameter scale or a transformed scale. State the scale explicitly, especially for MAP estimates. For a nonlinear concentration equation, calculate the forecast at each relevant posterior draw or integrate over the posterior when the desired target is a posterior-average forecast. A plug-in forecast is not automatically wrong: it may be the intended decision rule. The problem is treating it as equivalent to a posterior average without justification.
| Summary | What it describes | Typical decision connection | Main caution |
|---|---|---|---|
| Posterior mean | Posterior average on the stated scale | Minimizes expected squared error when the second moment is finite | Can be pulled toward a long tail; nonlinear forecasts need not equal the average forecast |
| Posterior median | A central value with at least half the posterior probability on each side | Minimizes expected absolute error | Does not generally equal the mean or mode |
| MAP | Location of maximum posterior density or mass | Useful as a mode-based point summary | For continuous parameters, depends on parameterization and may be nonunique |
A frequent mistake is to claim that the MAP is the “most probable exact parameter value” for a continuous parameter. Any single exact value has probability zero under a continuous distribution; the MAP is instead the point of greatest density, relative to the specified parameterization. Another mistake is to calculate a forecast at the posterior mean parameter and label it the posterior mean forecast. The correction is to keep the two operations distinct: \(g(E(\theta_i\mid y_i))\) is a plug-in forecast, whereas \(E\{g(\theta_i)\mid y_i\}\) averages the forecast function over the posterior.
These point summaries do not convey how spread out the posterior is or the range of plausible future concentrations. Choosing or comparing them is different from quantifying parameter uncertainty or predictive uncertainty. The summary should be reported with its scale, units, and intended loss or forecasting target; it should not be treated as a complete description of the patient-specific posterior.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- For a posterior with probabilities 0.25 at 2 L/h, 0.50 at 5 L/h, and 0.25 at 9 L/h, calculate the posterior mean and give a posterior median.
- Which posterior summary minimizes expected squared error, and what moment condition is needed for that result?
- For \(g(CL)=e^{-0.2CL}\), explain why \(g(E(CL\mid y))\) need not equal \(E\{g(CL)\mid y\}\).
- Why can a MAP estimate change when clearance is replaced by log-clearance?
Solutions: (1) The mean is \(0.25(2)+0.50(5)+0.25(9)=5.25\ \text{L/h}\). A median is \(5\ \text{L/h}\), since cumulative probability reaches 0.75 there and each side has at most half. (2) The posterior mean minimizes expected squared error when the posterior second moment is finite; expanding the risk yields a quadratic minimized at the mean. (3) The exponential is nonlinear, so applying it after averaging is not generally the same as averaging its values; calculate or integrate the transformed values for the posterior-average forecast. (4) The transformed density includes a Jacobian factor, so its maximum need not map to the maximum of the original-scale density. State which parameterization defines the MAP.