Why examine prior sensitivity?
An individual posterior combines the information in the observed concentrations with information supplied by the prior. When an individual has sparse data, a clinically important forecast can therefore depend on how the prior is specified, including its variance and distributional shape. A single posterior calculation does not reveal that dependence. Prior sensitivity analysis asks whether reasonable alternative priors lead to materially different posteriors or clinical forecasts, with the observed data and likelihood held fixed.
The aim is not to select whichever prior gives a preferred answer. Rather, the analysis makes explicit which conclusions are supported across plausible prior specifications and which rely strongly on assumptions about the individual parameter distribution. This is especially useful when a forecast lies near a clinical threshold: modest posterior changes can alter the probability of crossing that threshold even when posterior means appear similar.
The worked calculations use a deliberately simplified individual model. Let \(\theta\) be the log of an individual's clearance relative to a reference clearance, so \(\theta\) is dimensionless and may take any real value. A transformed concentration summary \(z\) is modeled as \(z=\theta+\epsilon\), where \(\epsilon\sim N(0,s^2)\). In the examples, \(z=0.20\) and the known standard deviation \(s=0.15\). This scalar likelihood represents a reduced illustration of individual information; it is not a substitute for constructing the likelihood from a full PK model.
For a clinical illustration, suppose the concentration at a specified future time under a fixed regimen and fixed values of all other PK quantities is \(C=10\exp(-\theta)\) mg/L. This deterministic mapping captures an inverse relation between clearance and concentration. It isolates how uncertainty about \(\theta\) changes the parameter-based forecast. It does not add residual measurement error or other sources of future variability.
Normal-prior updating and its variance
Begin with a normal prior \(\theta\sim N(\mu_0,\tau^2)\), independent of the sampling error in the likelihood. The likelihood is proportional to \(\exp\{-(z-\theta)^2/(2s^2)\}\). Multiplying it by the prior density gives a posterior density proportional to \(\exp\{-\frac{1}{2}[(\theta-\mu_0)^2/\tau^2+(z-\theta)^2/s^2]\}\). Collecting the terms involving \(\theta\) and completing the square yields the posterior variance and mean.
Thus \(\theta\mid z\sim N(m_{\text{post}},v_{\text{post}})\). The posterior precision, \(1/v_{\text{post}}\), is the sum of prior precision and likelihood precision. If the prior mean and observed summary differ, the posterior mean is a precision-weighted average of them. A smaller prior variance means greater prior precision, so the prior generally has more influence on the posterior. This result depends on the stated normal likelihood, known \(s^2\), and normal prior; it should not be applied automatically to a nonlinear, nonnormal individual likelihood.
Worked Example: Updating Under a Concentrated Prior
Use the hypothetical observation \(z=0.20\), likelihood standard deviation \(s=0.15\), and prior \(\theta\sim N(0,0.10^2)\). The scalar normal model is assumed to hold, the likelihood variance is treated as known, and the prior is independent of the observation error.
The prior variance is \(0.0100\), and the likelihood variance is \(0.0225\). Applying the precision formulas, \[ v_{\text{post}}=\left(\frac{1}{0.0100}+\frac{1}{0.0225}\right)^{-1} =(100+44.4444)^{-1}=0.00692308. \] The posterior standard deviation is \(\sqrt{0.00692308}=0.08321\), in log-clearance units.
The posterior mean is \[ m_{\text{post}}=0.00692308 \left(\frac{0}{0.0100}+\frac{0.20}{0.0225}\right) =0.00692308(8.88889)=0.06154. \] The observation is above the prior mean, but the posterior mean remains closer to zero because this prior is more precise than the likelihood. The result is a parameter posterior, not by itself a complete future-concentration predictive distribution.
Translate posterior changes into a clinical forecast
A variance comparison is more informative when it is tied to the clinical quantity that will be used. Under the illustrative concentration mapping, \(C>8\) mg/L is equivalent to \(10\exp(-\theta)>8\), or \(\theta<\log(10/8)=0.223144\). Because a normal posterior for \(\theta\) maps monotonically to a distribution for \(C\), its threshold probability can be obtained from the normal cumulative distribution function \(\Phi\): \(P(C>8\mid z)=\Phi((0.223144-m_{\text{post}})/\sqrt{v_{\text{post}}})\).
Worked Example: Comparing Two Prior Variances
Keep the same hypothetical likelihood, \(z=0.20\) with \(s=0.15\), and compare two plausible priors with the same mean \(\mu_0=0\): \(N(0,0.10^2)\) and \(N(0,0.50^2)\). Assume the scalar likelihood and fixed-regimen concentration mapping stated above. The first posterior was calculated in the preceding example. For the broader prior, \(\tau^2=0.25\), so \[ v_{\text{post}}=\left(\frac{1}{0.25}+\frac{1}{0.0225}\right)^{-1} =(4+44.4444)^{-1}=0.0206422, \] and \[ m_{\text{post}}=0.0206422 \left(\frac{0}{0.25}+\frac{0.20}{0.0225}\right) =0.18349. \] Its posterior standard deviation is \(0.14367\).
For the concentrated prior, the standardized threshold is \[ \frac{0.223144-0.0615385}{0.083205}=1.942, \] so \(P(C>8\text{ mg/L}\mid z)=\Phi(1.942)\approx0.974\). For the broader prior it is \[ \frac{0.223144-0.18349}{0.14367}=0.2760, \] so \(P(C>8\text{ mg/L}\mid z)=\Phi(0.2760)\approx0.609\). The two priors produce substantially different probabilities of exceeding the illustrative threshold, even though both use the same mean prior and the same data.
The posterior median concentration is \(10\exp(-m_{\text{post}})\) mg/L. It is about \(9.403\) mg/L under the concentrated prior and \(8.324\) mg/L under the broader prior. These are parameter-based forecasts under the specified mapping. The threshold probabilities use the full posterior uncertainty in \(\theta\), not merely these medians.
A useful comparison therefore reports more than whether posterior means differ. It can include posterior standard deviations or intervals, the shape of the posterior, a concentration summary, and a decision-relevant probability. In this example, both posterior medians exceed 8 mg/L, yet the probability of exceeding 8 mg/L differs markedly. A decision based only on the median would hide that sensitivity.
Distributional shape can matter as well
Changing prior variance is only one sensitivity analysis. Two priors can have similar centers and marginal variances but differ in shape, such as a single normal distribution versus a mixture representing two plausible subgroups. A mixture prior can yield a posterior with distinct components, particularly when the individual data do not clearly distinguish between them. The component probabilities must be updated using the marginal likelihood under each component, not by simply retaining the prior weights.
Worked Example: A Two-Component Prior
Use the same likelihood, \(z=0.20\) and \(s=0.15\), but take a hypothetical prior \(\theta\sim0.5N(-0.20,0.10^2)+0.5N(0.20,0.10^2)\). The component means represent two possible clearance patterns. The likelihood variance is known and the same in both components. Conditional on either component, normal updating gives posterior variance \(0.00692308\) and standard deviation \(0.083205\).
For the lower-mean component, the posterior mean is \[ 0.00692308\left(\frac{-0.20}{0.0100}+\frac{0.20}{0.0225}\right) =0.00692308(-11.1111)=-0.07692. \] For the higher-mean component, it is \[ 0.00692308\left(\frac{0.20}{0.0100}+\frac{0.20}{0.0225}\right) =0.2000. \] To update component weights, note that the observation under a component with mean \(\mu_j\) has marginal distribution \(z\mid j\sim N(\mu_j,0.0100+0.0225)\). At \(z=0.20\), the likelihood ratio for the higher-mean component versus the lower-mean component is \[ \frac{p(z\mid\text{high})}{p(z\mid\text{low})} =\exp\left(\frac{(0.40)^2}{2(0.0325)}\right) =\exp(2.46154)\approx11.72. \] With prior weights of 0.5 each, the posterior weight for the higher-mean component is \(11.72/(1+11.72)\approx0.9214\); the lower-mean weight is approximately \(0.0786\).
The event \(C>8\) mg/L again corresponds to \(\theta<0.223144\). The conditional event probabilities are \[ \Phi\left(\frac{0.223144-0.2000}{0.083205}\right)\approx\Phi(0.2781)\approx0.6095 \] for the higher-mean component and \[ \Phi\left(\frac{0.223144-(-0.07692)}{0.083205}\right) \approx\Phi(3.606)\approx0.9998 \] for the lower-mean component. Averaging with the posterior weights gives \(0.9214(0.6095)+0.0786(0.9998)\approx0.640\). The mixture posterior forecast probability is therefore about 0.640, distinct from the single-normal results above. The calculation also shows why inspecting only a posterior mean can conceal structure: the mixture retains contributions from both components.
How to assess the comparison
A defensible sensitivity analysis begins by identifying which prior features are uncertain and specifying alternatives before interpreting the resulting forecasts. Alternatives should be scientifically plausible, for example a range of variances supported by relevant population information or a mixture motivated by real heterogeneity. An intentionally extreme prior can be a stress test, but it should be labeled as such rather than presented as equally plausible evidence.
For each prior, use the same observed data, likelihood, parameterization, dosing scenario, and forecast definition. Then compare posterior and clinical quantities side by side. A compact report might include:
- the prior center, variance or distributional form, and rationale for each alternative;
- posterior location, uncertainty, and any important shape features;
- the forecast summary or probability of crossing a clinically specified threshold;
- whether the clinical interpretation or proposed action changes across plausible priors.
Prior-data conflict diagnostics address whether the observed data are surprising under the prior predictive distribution. Sensitivity analysis asks a related but different question: how much do posterior conclusions change across plausible prior choices? A conflict signal can motivate careful investigation, but a lack of obvious conflict does not guarantee that every clinically relevant forecast is insensitive to prior assumptions.
Vary the prior feature under examination, such as variance or distributional shape, and record why each alternative is scientifically defensible.
Use the same individual likelihood, observations, parameter definitions, and forecast scenario for every prior so differences are attributable to prior specification.
Calculate posterior summaries and clinically relevant forecast probabilities, then determine whether plausible prior choices change the interpretation or decision.
Alternatives, limitations, and common pitfalls
The main alternative to a prior sensitivity analysis is to report results under one selected prior. That can be appropriate when the prior is strongly justified and the result is robust, but it does not show robustness by itself. A broader prior may be used as a reference comparison, while a distributional alternative can reveal sensitivity to tail behavior or multimodality. Neither alternative automatically represents the truth: the comparison is conditional on the candidate priors and the likelihood being considered.
Sensitivity analyses also have limits. If all candidate priors are implausible, stability across them is not reassuring. If the likelihood itself is misspecified, changing priors cannot repair the model. In a full PK analysis, uncertainty in other parameters, dosing history, residual error, and future variability may also affect forecasts; those elements were fixed or omitted in the scalar examples. A deterministic mapping from parameter to concentration must not be described as a complete predictive distribution when additional sources of variability are relevant. Finally, if alternative priors produce materially different clinical conclusions, that is a finding to report and investigate, not a reason to choose the most convenient result.
Common mistake: “The posterior means are close, so the forecast is insensitive to the prior.” This reasoning is incomplete: a threshold probability depends on the full posterior distribution, including its variance and shape. In the variance example, the posterior means are both below the threshold value for \(\theta\), but the probability of \(C>8\) mg/L changes from about 0.974 to 0.609. The correction is to compare the clinical forecast itself, including uncertainty and decision-relevant probabilities, under each plausible prior.
Another error is to change the prior and the likelihood assumptions at the same time, then attribute all differences to prior sensitivity. Keep the likelihood fixed for a prior sensitivity comparison. If the likelihood is also uncertain, examine that separately or clearly label a broader model sensitivity analysis. The reported probability is conditional on the model, prior, data, and fixed forecast scenario; it is not the probability that a prior is correct.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In the scalar normal model, what happens to posterior precision as the prior variance decreases, with the likelihood held fixed?
- For the concentrated-prior posterior in the first worked example, calculate the probability that \(C>8\) mg/L using the stated mapping.
- Why must a two-component prior's weights be updated when data are observed?
- Give one reason that a prior sensitivity analysis can be reassuring but still fail to establish that a forecast is reliable.
Solutions. (1) Prior precision is \(1/\tau^2\), so decreasing \(\tau^2\) increases prior precision and increases total posterior precision \(1/\tau^2+1/s^2\). (2) The event is \(\theta<0.223144\). With \(m_{\text{post}}=0.0615385\) and posterior standard deviation \(0.083205\), the standardized value is \((0.223144-0.0615385)/0.083205=1.942\), giving \(\Phi(1.942)\approx0.974\). (3) Bayes' rule weights each component by its prior weight multiplied by the probability of the observed data under that component; here the observation is much more likely under the higher-mean component. (4) All candidate priors might be implausible, or the likelihood or forecast model might be misspecified. Sensitivity to the chosen priors is only one part of assessing model reliability.