Diagnosing Prior-Data Conflict in Sparse PK Estimation

Why Diagnose Prior-Data Conflict?

An individual PK estimate combines population information with that individual’s concentration data. With sparse observations, the population-derived prior can have substantial influence. That influence is often appropriate: one noisy concentration should not automatically override reliable population information. But if the data are difficult to reconcile with the prior and the assumed PK model, the resulting estimate may be misleading. The practical question is therefore not simply whether the estimate moved, but whether the observed data are unexpectedly far from what the prior and model jointly predict, and which parameter directions or observations account for that tension.

A prior-data conflict diagnostic is a signal for investigation, not a decision rule that proves the prior is wrong. Sparse PK data may be surprising by chance, a sampling time or dose record may be incorrect, the likelihood may be misspecified, or the individual may genuinely differ from the population. The diagnostics below help distinguish these possibilities without treating any one numerical threshold as decisive.

Definition: Let $\theta$ denote an individual’s PK parameter vector, $y$ the individual’s concentration data, and $x$ the available patient covariates. A population-derived individual prior is $p(\theta\mid x)$, and the individual likelihood is $p(y\mid\theta)$. The posterior is proportional to their product: $$ p(\theta\mid y,x) \propto p(y\mid\theta)\,p(\theta\mid x). $$ Prior-data conflict is a diagnostic description for data that are unusually discrepant with the distribution implied jointly by this prior and the specified observation model. It is not a statement that either the prior or the data must be incorrect.

Three Complementary Diagnostics

The diagnostics address different questions. A posterior shift asks how far the estimate moved from the prior. A prior-predictive discrepancy asks whether the observed data themselves are surprising before updating. Parameter-specific likelihood evidence asks which observations and parameter directions make the likelihood favor values away from the prior’s center. Agreement across these views is more informative than any one measure alone.

Posterior shift: For a scalar parameter represented on a chosen scale, a descriptive standardized shift is the posterior mean minus the prior mean, divided by the prior standard deviation. It measures movement in units of prior spread. It is not a universal conflict test, and there is no generally valid cutoff that declares conflict.

The scale matters. For a positive parameter such as clearance, a shift on the original scale answers a different question from a shift on the log scale. A log-scale shift describes multiplicative change and often suits a lognormal prior. In a correlated parameter vector, marginal shifts can also conceal movement along a combination of parameters. A large shift can be reasonable when the likelihood is informative; a small shift can occur simply because the likelihood is weak.

Prior-predictive discrepancy: The prior-predictive distribution of the data is $$ p(y\mid x)=\int p(y\mid\theta)\,p(\theta\mid x)\,d\theta. $$ It includes uncertainty in both the individual parameters under the prior and the observation process. A discrepancy is assessed by comparing the observed data, or a prespecified summary of them, with this distribution.

For a scalar observation with a normal prior-predictive distribution, a standardized discrepancy can be calculated as the observation minus its predictive mean, divided by its predictive standard deviation. For multiple observations, a discrepancy statistic could summarize a residual pattern or a distance from the joint predictive distribution. The statistic should be selected with the data structure in mind; examining many summaries and reporting only the most extreme one exaggerates apparent surprise.

A prior-predictive tail area is a model-based measure of how often data at least as discrepant would arise under the assumed prior and observation model. A small tail area is a warning, not the probability that the prior is wrong. In particular, it does not identify whether the cause is a prior mismatch, an unusual individual, a model limitation, or a data problem.

Worked Examples

Worked Example: Standardizing a Shift in Log Clearance

Suppose a hypothetical patient has a population-derived prior for $\eta=\log(CL)$, where $CL$ is clearance in L/h. Let $\eta\sim N(\log 5,0.25^2)$. A sparse-data likelihood is summarized by a noisy estimate $z=\log 8$, with $z\mid\eta\sim N(\eta,0.15^2)$. These assumptions give a normal-normal update on the log scale.

The posterior variance is the reciprocal of the sum of prior and likelihood precisions, and the posterior mean is the precision-weighted average:

v_1=\left(\frac{1}{0.25^2}+\frac{1}{0.15^2}\right)^{-1} =\left(16+44.4444\right)^{-1}=0.0165441, $$ $$ m_1=v_1\left(\frac{\log 5}{0.25^2}+\frac{\log 8}{0.15^2}\right) =\log 5+\frac{0.0625}{0.085}(\log 8-\log 5) =1.95503.

The posterior standard deviation is $\sqrt{0.0165441}=0.1286$. The standardized shift from the prior mean is $(1.95503-\log 5)/0.25=0.34559/0.25=1.382$ prior standard deviations. This is appreciable movement, but it does not by itself establish conflict: the likelihood summary is fairly precise relative to the prior.

Exponentiating the posterior mean on the log scale gives $\exp(1.95503)=7.06$ L/h, the posterior median clearance. Because the posterior for $CL$ is lognormal, its posterior mean is instead $\exp(m_1+v_1/2)=\exp(1.96330)$, approximately $7.12$ L/h. These summaries are not interchangeable. Nor does the shift alone tell us whether the original concentration data are surprising under the prior predictive distribution.

Worked Example: Checking a Prior-Predictive Concentration

For a hypothetical sampling time, suppose the prior and PK model imply that a patient’s true concentration has a normal distribution with mean $3.0$ mg/L and standard deviation $1.0$ mg/L. Assume the observation error is independent, normal, and has standard deviation $0.25$ mg/L. The measured concentration is $5.4$ mg/L. Under these assumptions the prior-predictive concentration includes both sources of uncertainty.

The predictive mean is $3.0$ mg/L. Independence gives the predictive variance as the sum of the parameter-prediction and observation variances:

SD_{\mathrm{pred}}=\sqrt{1.0^2+0.25^2} =\sqrt{1.0625}=1.031\ \text{mg/L}.

The standardized discrepancy is $(5.4-3.0)/1.031=2.328$. Under the stated normal predictive distribution, the two-sided tail area is approximately $0.0199$. Thus this concentration is relatively unusual under the specified prior and model, and merits checking. The tail area is not a 1.99% probability that the prior is correct or incorrect; it is the probability, under this predictive model, of a discrepancy at least this large in either direction.

The calculation relies on the stated predictive mean and variance, normality, and independence. If the PK predictive distribution is skewed or multimodal, a normal standardized score may not describe its tails well. A suitable predictive distribution or discrepancy measure must then be used rather than forcing this calculation.

Locating the Likelihood Evidence

A global discrepancy can flag tension without showing what drives it. Inspection of observation-level likelihood contributions can identify which samples disagree most with predictions at a prior-typical parameter value. Parameter-direction scores add information about which way the likelihood locally favors movement. These are diagnostics of the likelihood at a specified point, not a decomposition of the posterior into independent “votes” for each parameter.

For conditionally independent observations, the log likelihood is a sum of observation-level terms. Its derivative with respect to a parameter is likewise a sum of observation-specific score contributions. A score is local: its magnitude depends on the parameter scale and the point at which it is calculated. In a nonlinear model, the direction that improves fit may change as parameters move. Correlation in the prior can also mean that changing one parameter in isolation is not a plausible direction of individual variation.

Worked Example: Finding Which Samples Drive a Likelihood Shift

Consider a hypothetical one-compartment intravenous bolus model with dose $D=100$ mg, volume $V=20$ L, and clearance $CL=4$ L/h at the center of the individual prior. Suppose the two observed concentrations are $4.5$ mg/L at 1 h and $0.75$ mg/L at 8 h. Assume independent normal errors with standard deviation $0.20$ on the log-concentration scale. For this diagnostic, evaluate the likelihood and its local scores at the prior center.

Writing $v=\log V$, $c=\log CL$, and $k=CL/V=0.2$ h$^{-1}$, the predicted log concentration is $\log C(t)=\log D-v-kt$. The prior-center predictions are $1.40944$ at 1 h and $0.00944$ at 8 h. The observed log concentrations are $\log 4.5=1.50408$ and $\log 0.75=-0.28768$, giving residuals $e_1=0.09464$ and $e_2=-0.29712$.

For normal errors, each observation’s contribution to minus twice the log likelihood, apart from a constant common to all parameter values, is $e_i^2/\sigma^2$. Here the contributions are $0.09464^2/0.20^2=0.224$ at 1 h and $(-0.29712)^2/0.20^2=2.207$ at 8 h. The later sample contributes much more to the discrepancy at this prior-center prediction.

To inspect parameter directions, differentiate the predicted log concentration: $\partial\log C/\partial c=-kt$ and $\partial\log C/\partial v=-1+kt$. Since the log-likelihood score contribution is $e_i(\partial\log C/\partial\theta)/\sigma^2$, the total scores at the prior center are

\frac{\partial\ell}{\partial c} =\frac{0.09464(-0.2)+(-0.29712)(-1.6)}{0.20^2} =11.41, $$ $$ \frac{\partial\ell}{\partial v} =\frac{0.09464(-0.8)+(-0.29712)(0.6)}{0.20^2} =-6.350.

Locally, the likelihood increases as $c$ increases and as $v$ decreases from this point. The late sample supplies most of the positive score for $c$, while both samples contribute negative score for $v$. This does not prove that clearance alone is too low or volume alone is too high: the model is nonlinear, and the two parameters jointly affect the predictions. A posterior fit and the parameter correlation structure are needed to interpret the resulting movement.

Justification and Practical Workflow

The normal-normal calculation illustrates why a posterior shift and a predictive discrepancy are related but not identical. If $z\mid\eta\sim N(\eta,s^2)$ and $\eta\sim N(m_0,\tau^2)$, integrating over $\eta$ gives $z\sim N(m_0,\tau^2+s^2)$. The predictive variance includes uncertainty in the prior parameter and in the observation. By contrast, the posterior shift is the likelihood-weighted fraction of the difference $z-m_0$:

m_1-m_0=\frac{\tau^2}{\tau^2+s^2}(z-m_0).

A precise observation can therefore produce a substantial posterior shift even when its predictive discrepancy is not extreme; a noisy observation can be quite far from the prior center but have limited influence on the posterior. This is why neither diagnostic substitutes for the other.

A practical sequence is to establish that the data and model inputs are credible, evaluate prior-to-posterior movement on interpretable parameter scales, compare observations with their prior-predictive distributions, and inspect observation-specific likelihood and parameter-direction evidence. For correlated parameters, examine relevant joint directions rather than interpreting marginal movement as independent evidence. If computation uses an approximation or sampling method, also check that the reported posterior summaries are numerically reliable before diagnosing conflict.

Verify the setup.
Check concentration units, dose history, sample times, covariates, and the assumed observation model before attributing a discrepancy to the prior.

Quantify movement.
Compare prior and posterior summaries using a stated parameter scale and prior uncertainty; do not interpret movement alone as evidence of conflict.

Locate the tension.
Check prior-predictive discrepancies and determine which observations and parameter directions contribute to the likelihood’s preference away from the prior center.

Interpret cautiously.
Consider data problems, model inadequacy, an unusual individual, and ordinary chance variation; a diagnostic does not distinguish these causes by itself.

Comparisons, Failure Modes, and Limitations

Diagnostic	Question answered	Main limitation
Standardized posterior shift	How far did a parameter summary move relative to prior spread?	Depends on scale and can be large under informative data without model conflict.
Prior-predictive discrepancy	How surprising are the observations under the prior and observation model?	Can flag tension but does not identify its cause.
Likelihood contributions and scores	Which observations and local parameter directions favor movement?	Local and parameterization-dependent; does not provide a unique allocation of evidence among correlated parameters.

A common mistake is to see a posterior estimate several prior standard deviations from the prior mean and conclude that the prior is invalid. That reasoning ignores how informative the data are: a precise, correctly modeled likelihood can rationally move the posterior far from the prior. The correction is to treat the shift as a descriptive measure, then ask whether the observations are surprising under the prior predictive distribution and whether the data, model, and parameter directions have been checked.

The reverse mistake is to call the prior-data relationship acceptable because the posterior barely moved. With very sparse or noisy data, the posterior may remain close to the prior even when an observation is discrepant. Also, a sum of likelihood contributions or a local score is not a definitive ranking of parameter importance: changing the parameter scale, moving away from the evaluation point, or accounting for prior correlation can change the interpretation.

All these diagnostics are conditional on the stated PK model, prior, and observation assumptions. Sparse sampling may leave multiple explanations compatible with the same concentrations. A conflict flag should prompt review and clinical judgment, not automatic rejection of population information or an immediate dose change. It is also important not to mistake a model-based predictive tail area for a posterior probability about the cause of the discrepancy.

Check Your Understanding

Answer each question, then compare with the solutions.

A prior for a log-scale parameter has mean 0 and standard deviation 0.4; the posterior mean is 0.6. What is the standardized shift, and what does it establish?
A measurement is far from the prior-predictive mean but has a very large predictive standard deviation. Why might the posterior move only slightly?
In a nonlinear PK model, one sample has the largest likelihood residual contribution at the prior center. Does this prove that the sample is erroneous?
Why should scores for two correlated PK parameters not automatically be interpreted as independent evidence about each parameter?

Solutions. (1) The shift is $(0.6-0)/0.4=1.5$ prior standard deviations. It describes movement only; it does not establish conflict. (2) The observation may still be fairly ordinary relative to the wide predictive distribution, and a noisy likelihood generally has limited precision to move the posterior. (3) No. It identifies a large contribution to the discrepancy at that parameter value; chance variation, prior mismatch, or model inadequacy are also possible explanations. (4) The likelihood may favor a joint direction involving both parameters, while the prior constrains their combinations. Marginal scores evaluated one parameter at a time do not isolate independent causes.