Updating Integrated Evidence as Data Accumulate

Why Updating Integrated Evidence Requires Care

An integrated model is rarely built from a single, final data set. Pharmacokinetic observations, exposure-response results, and clinical outcomes may arrive at different times. The central task is to incorporate each new piece of evidence while preserving what earlier data have already established. Updating too little leaves relevant evidence unused; updating too much, by reusing the same observations as if they were new, makes the evidence appear stronger than it is.

A further challenge is that model components are linked. New clinical outcomes may change predictions from an exposure-response component and, when the linked components share uncertain parameters, may also change beliefs about quantities in an upstream component. A reliable update therefore carries forward distributions and their dependence, not only point estimates.

Definition: An incremental evidence update revises the current posterior distribution using the likelihood contribution of data that have not already been incorporated. In a linked model, the update is made to the joint distribution of the relevant parameters or model quantities, so that uncertainty and dependence can be propagated to downstream predictions.

Bayesian Updating Across Stages

Let \(\theta\) denote parameters or quantities in an integrated model, and let \(D_{\text{old}}\) be the data already used. The current posterior is \(p(\theta\mid D_{\text{old}})\). If \(D_{\text{new}}\) is genuinely new data, Bayes' rule gives

p(\theta\mid D_{\text{old}},D_{\text{new}}) \propto p(D_{\text{new}}\mid \theta,D_{\text{old}}) p(\theta\mid D_{\text{old}}).

When the old and new data are conditionally independent given \(\theta\), the new-data likelihood simplifies to \(p(D_{\text{new}}\mid\theta)\). This is the usual sequential update: the current posterior serves as the prior for the next stage, and only the new likelihood contribution is applied. Conditional independence is a modeling assumption, not a consequence of the data arriving at different times. For example, repeated measurements on the same participant may remain dependent after conditioning on population parameters unless the model accounts for participant-level effects.

If components have parameters \(\theta_1\) and \(\theta_2\), the relevant object may be their joint posterior \(p(\theta_1,\theta_2\mid D)\), rather than two separate marginal distributions. Replacing that joint distribution with independent marginal summaries discards dependence. That loss can misstate uncertainty in a prediction that combines both components.

A Normal Update for a New PK Summary

Worked Example: Updating a Mean Exposure Estimate

Suppose \(\theta\) is a population mean exposure metric in mg·h/L. Before a new clinical PK analysis, the current posterior is approximated by \(\theta\mid D_{\text{old}}\sim N(10,4)\). A new, non-overlapping data set gives an estimated mean \(z=11\) mg·h/L with known standard error \(1.5\) mg·h/L. Assume the sampling model \(z\mid\theta\sim N(\theta,2.25)\), and assume this estimate is conditionally independent of the earlier data given \(\theta\).

For a normal prior and normal observation model, the posterior variance is the reciprocal of the sum of prior and data precisions; the posterior mean is the precision-weighted average:

V_{\text{new}}= \left(\frac{1}{V_{\text{old}}}+\frac{1}{V_z}\right)^{-1}, \qquad m_{\text{new}}= V_{\text{new}} \left(\frac{m_{\text{old}}}{V_{\text{old}}}+\frac{z}{V_z}\right).

Substituting the stated values,

V_{\text{new}}= \left(\frac{1}{4}+\frac{1}{2.25}\right)^{-1} =(0.25+0.4444)^{-1} =1.44\;(\text{mg·h/L})^2,

m_{\text{new}}= 1.44\left(\frac{10}{4}+\frac{11}{2.25}\right) =1.44(2.500+4.889) =10.64\;\text{mg·h/L}.

The updated posterior is approximately \(N(10.64,1.44)\), with posterior standard deviation \(1.20\) mg·h/L. Under this normal model, a 95% posterior credible interval is \(10.64\pm1.96(1.20)=[8.288,12.992]\) mg·h/L. Conditional on the model and data, the posterior probability that \(\theta\) lies in this interval is approximately 95%. The calculation relies on the stated normal approximations, known sampling variance, and non-overlap of the new data with the old data.

Updating Linked Components and Predictions

A linked update is not always confined to the component that generated the new observations. If an upstream quantity and a downstream outcome are dependent in the current posterior, data about the outcome can revise the posterior for both. This is not an extra pathway for reusing old data; it is the consequence of conditioning their joint distribution on a new likelihood.

For a multivariate normal current posterior \(x\sim N(m,\Sigma)\), suppose a new observation follows \(y=Hx+\epsilon\), with independent \(\epsilon\sim N(0,R)\). The conditional update is

m_{\text{new}}=m+\Sigma H^\mathsf{T}(H\Sigma H^\mathsf{T}+R)^{-1}(y-Hm),

\Sigma_{\text{new}}=\Sigma-\Sigma H^\mathsf{T}(H\Sigma H^\mathsf{T}+R)^{-1}H\Sigma.

The term \(y-Hm\) is the difference between the observation and its current model-based mean. The update to a component depends on its covariance with the observed quantity: if that covariance is zero, the observation does not shift that component's mean in this Gaussian update. The covariance matrix also changes, so a downstream prediction should be recalculated from the updated joint distribution.

Worked Example: New Clinical Outcomes Update a Linked Prediction

Suppose a current integrated analysis has a joint normal posterior for \(\theta\), a mean exposure quantity in mg·h/L, and \(q\), a model-derived mean clinical response in response units. The current distribution has means 10 mg·h/L and 0.60 response units, variances 4 \((\text{mg·h/L})^2\) and 0.25 response-units squared, and covariance 0.50 mg·h/L times response units. Thus

\begin{pmatrix}\theta\\q\end{pmatrix} \sim N\left[ \begin{pmatrix}10\\0.60\end{pmatrix}, \begin{pmatrix}4&0.50\\0.50&0.25\end{pmatrix} \right].

A new, non-overlapping clinical data set estimates the response as \(y=0.90\) response units. Assume \(y=q+\epsilon\), with \(\epsilon\sim N(0,0.09)\), independent of the current parameter uncertainty. The observation variance in the update is \(0.25+0.09=0.34\) response-units squared, and the covariances of \(\theta\) and \(q\) with \(y\) are 0.50 and 0.25, respectively. The observed-minus-predicted difference is \(0.90-0.60=0.30\) response units.

The updated means are

E(\theta\mid y)=10+\frac{0.50}{0.34}(0.30) =10.441\;\text{mg·h/L},

E(q\mid y)=0.60+\frac{0.25}{0.34}(0.30) =0.8206\;\text{response units}.

The updated variances and covariance are obtained by subtracting the relevant covariance products divided by 0.34:

\operatorname{Var}(\theta\mid y)=4-\frac{0.50^2}{0.34}=3.2647, \qquad \operatorname{Var}(q\mid y)=0.25-\frac{0.25^2}{0.34}=0.06618,

\operatorname{Cov}(\theta,q\mid y) =0.50-\frac{(0.50)(0.25)}{0.34} =0.1324.

The clinical observation shifts the exposure quantity because the current integrated model gives \(\theta\) and \(q\) positive posterior covariance. For a future comparable response \(Y_{\text{new}}=q+\epsilon_{\text{new}}\), with independent residual variance 0.09, the predictive mean is 0.8206 response units and the predictive variance is \(0.06618+0.09=0.15618\) response-units squared. The predictive standard deviation is 0.3952 response units, giving a normal-approximation 95% posterior predictive interval

0.8206\pm1.96(0.3952) =[0.0460,1.595]\;\text{response units}.

This interval describes uncertainty for a future response, including both uncertainty in \(q\) and residual variability. It is wider than an interval for the posterior mean \(q\) alone because the future observation has additional residual variation. The calculations assume the stated joint normal posterior, correct covariance, and independent future residual.

New Evidence, Reused Summaries, and Sequential Workflows

A report can be newly received without containing new evidence. For updating, what matters is whether its underlying observations, or their likelihood contribution, are already represented in the current posterior. A posterior estimate from an earlier analysis is not an independent observation that can simply be added again. Likewise, a clinical report that reprints an earlier PK estimate does not justify applying that estimate's likelihood a second time.

Worked Example: Adding a Distinct Data Set Without Recounting the First

Use the posterior from the PK example: \(\theta\mid D_1\sim N(10.64,1.44)\). A separate, non-overlapping clinical PK data set now provides \(z_2=12\) mg·h/L with known standard error \(2\) mg·h/L. Assume \(z_2\mid\theta\sim N(\theta,4)\) and conditional independence from \(D_1\) given \(\theta\).

The correct sequential update is

V_2=\left(\frac{1}{1.44}+\frac{1}{4}\right)^{-1} =1.0588\;(\text{mg·h/L})^2,

m_2=1.0588\left(\frac{10.64}{1.44}+\frac{12}{4}\right) =1.0588(7.3889+3.000) =11.00\;\text{mg·h/L}.

Thus the updated posterior is approximately \(N(11.00,1.0588)\). In contrast, suppose a later report merely repeats the first estimate \(z_1=11\) mg·h/L and its standard error of 1.5 mg·h/L. That is not another observation. Applying its likelihood again would give a false variance \((1/1.44+1/2.25)^{-1}=0.8780\) and a false mean of approximately 10.7805 mg·h/L. The apparent extra precision comes entirely from counting the same data twice.

A practical update should keep an auditable record of the data contributing to each component, the likelihood used for each update, the model version, and the posterior passed forward. When a source supplies only a summary, establish whether it is a raw-data estimate, a likelihood contribution, or a posterior that already incorporates earlier evidence. A summary may support a valid update if its sampling distribution and dependence on existing evidence are adequately represented; a point estimate without its uncertainty and provenance is generally insufficient.

Justification, Checks, and Alternatives

The sequential rule follows directly from Bayes' rule. If the full data are conditionally independent given \(\theta\), then \(p(D_{\text{old}},D_{\text{new}}\mid\theta)=p(D_{\text{new}}\mid\theta)p(D_{\text{old}}\mid\theta)\). Substituting this factorization into the posterior and grouping the prior with the old-data likelihood gives the current posterior:

p(\theta\mid D_{\text{old}},D_{\text{new}}) \propto p(D_{\text{new}}\mid\theta) p(D_{\text{old}}\mid\theta)p(\theta) \propto p(D_{\text{new}}\mid\theta)p(\theta\mid D_{\text{old}}).

This derivation shows both why sequential updating works and where it can fail: the factorization is invalid if dependence between old and new data is ignored. If data overlap, share participants, or are linked through unmodeled study effects, the joint likelihood must represent that structure rather than treating the records as independent.

Useful checks include confirming that the new likelihood corresponds to new observations, checking units and sampling variances, and comparing the new observation with the current predictive distribution. An unexpectedly large discrepancy can indicate a data error, an inadequate likelihood, a change in population or measurement conditions, or a misspecified link between components. A small posterior shift does not by itself show that the new data are uninformative: the prior may be precise, the data noisy, or the new observation close to the current prediction. Conversely, a large shift warrants investigation rather than automatic rejection of either source.

Approach	Strength	Key caution
Sequential posterior update	Efficient when the current posterior and new-data likelihood are available and correctly specified.	Requires sufficient posterior information and a valid account of dependence and data overlap.
Joint refit to all data	Can represent the full likelihood and dependence directly.	May require more computation and careful harmonization of data and model components.
Plug-in handoff of point estimates	Simple to communicate and implement.	Discards uncertainty and often covariance, potentially understating predictive uncertainty.

Sequential updating is mathematically equivalent to a joint analysis when the same model, prior, and likelihood factorization are used and the full posterior is carried forward. In practice, approximations to a posterior, changes in model assumptions, or loss of covariance can break that equivalence. A plug-in handoff may be a deliberate operational simplification, but it should not be mistaken for a full uncertainty-propagating update.

Common Pitfalls and Limitations

Counting a summary twice. The wrong reasoning is, “This estimate has arrived in a new report, so it is new evidence.” The correction is to trace the estimate to its underlying data and include each likelihood contribution once.
Passing only a point estimate. This treats an uncertain upstream result as known exactly. Carry a posterior distribution, or a justified approximation that retains relevant uncertainty and covariance.
Assuming every new study is independent. Different study dates do not guarantee conditional independence. Shared participants, sites, assays, or study-level effects may induce dependence that belongs in the model.
Interpreting an update as proof of causation or model validity. A posterior shift is conditional on the likelihood, prior, and link assumptions; it does not establish that those assumptions are correct.

Updates are only as reliable as the model and the information retained between stages. Approximate summaries may omit skewness, multimodality, or dependence; a normal approximation can then produce misleading predictions. New data from a materially different population may also require a model that represents the difference, rather than a direct update under an unchanged sampling model. These limitations should be reported alongside the updated predictions.

Key takeaway: Update an integrated analysis with the likelihood contribution of genuinely new data, conditional on what is already known. Preserve joint uncertainty across linked components, and verify that a newly received summary does not reuse evidence already included.

Check Your Understanding

Answer each question, then compare with the solutions below.

A current posterior is \(N(5,1)\). A genuinely new independent estimate is 7 with known sampling variance 3. What are the updated mean and variance?
A downstream response quantity has zero posterior covariance with an upstream parameter. Under the multivariate normal update described above, will new data measuring only the downstream quantity shift the upstream posterior mean? Why?
A report repeats a study estimate already included in the current posterior. What likelihood contribution should be applied for this report?
Why can a future-outcome predictive interval be wider than a credible interval for the posterior mean response?

Solutions. 1. The updated variance is \((1/1+1/3)^{-1}=0.75\). The mean is \(0.75(5/1+7/3)=0.75(7.3333)=5.50\). 2. No. The Gaussian mean update for the upstream parameter is its covariance with the observation divided by the observation variance, multiplied by the observed-minus-predicted difference; with zero covariance, that shift is zero. 3. No new likelihood contribution should be applied for the repeated estimate. Its underlying data are already represented; applying the likelihood again would double count them. 4. A future outcome includes residual variability as well as posterior uncertainty about the mean response. For independent residual error, the predictive variance adds the residual variance to the posterior variance of the mean response.