Using Cut Models to Limit Feedback

Why Block Feedback?

Integrated models often have a directional scientific structure. For example, a pharmacokinetic (PK) model may estimate exposure, which is then passed to an exposure-response model. If the PK component is well supported but the response component is weakly validated, fitting both jointly can let uncertain response evidence alter the PK estimates. That feedback may be mathematically coherent under the joint model, yet scientifically undesirable if the response model is not trusted to inform PK.

A Bayesian cut distribution formalizes a one-way information rule: the upstream component informs the downstream component, but downstream data do not update the upstream component. The cut does not make the components independent. It retains the forward dependence and propagates uncertainty in the upstream output.

Definition: Let $\theta$ denote parameters in an upstream module, $\phi$ parameters in a downstream module, $y_A$ data used by the upstream module, and $y_B$ data used by the downstream module. A two-module cut distribution is $$ p_{\mathrm{cut}}(\theta,\phi\mid y_A,y_B) = p(\theta\mid y_A)\,p(\phi\mid\theta,y_B). $$ The upstream marginal remains $p(\theta\mid y_A)$, while downstream data update $\phi$ conditional on each value of $\theta$. Thus information flows from $\theta$ to $\phi$, but not from $y_B$ back to $\theta$.

How a Cut Differs From Ordinary Joint Updating

Suppose the joint model has prior $p(\theta)p(\phi\mid\theta)$ and likelihood $p(y_A\mid\theta)p(y_B\mid\theta,\phi)$. Ordinary Bayesian updating gives a joint posterior proportional to

p(\theta,\phi\mid y_A,y_B) \propto p(y_A\mid\theta)\,p(y_B\mid\theta,\phi)\,p(\phi\mid\theta)\,p(\theta).

To see the feedback, integrate out $\phi$. Define the downstream marginal likelihood $m_B(y_B\mid\theta)=\int p(y_B\mid\theta,\phi)p(\phi\mid\theta)\,d\phi$. Then the joint posterior marginal for the upstream parameter is

p(\theta\mid y_A,y_B) \propto p(\theta\mid y_A)\,m_B(y_B\mid\theta).

Unless $m_B(y_B\mid\theta)$ is constant in $\theta$, downstream evidence changes the upstream distribution. A cut instead holds the upstream marginal at $p(\theta\mid y_A)$. The conditional downstream distribution is still updated by $y_B$:

p(\phi\mid\theta,y_B) = \frac{p(y_B\mid\theta,\phi)p(\phi\mid\theta)} {\int p(y_B\mid\theta,\phi')p(\phi'\mid\theta)\,d\phi'}.

This establishes the key trade-off: a cut deliberately departs from the ordinary joint posterior to enforce a chosen direction of learning. It is not an approximation to joint Bayes that is always closer to the “truth”; it is a modeling decision reflecting which information pathways are considered credible.

Worked Example: Discrete Modules Show the Feedback

Suppose $\theta$ and $\phi$ are binary, each taking values 0 or 1. Before observing $y_A$, let $P(\theta=1)=P(\theta=0)=0.5$. The trusted upstream data have likelihoods $P(y_A\mid\theta=1)=0.8$ and $P(y_A\mid\theta=0)=0.2$. In the downstream module, let $P(\phi=1\mid\theta)=P(\phi=0\mid\theta)=0.5$. For the observed downstream data $y_B$, the likelihoods are:

$\theta$	$P(y_B\mid\theta,\phi=1)$	$P(y_B\mid\theta,\phi=0)$
1	0.18	0.02
0	0.98	0.82

First update the trusted module. Its posterior probability of $\theta=1$ is

P(\theta=1\mid y_A) = \frac{0.8(0.5)}{0.8(0.5)+0.2(0.5)} = 0.8.

The downstream marginal likelihood is $P(y_B\mid\theta=1)=0.5(0.18)+0.5(0.02)=0.10$, and $P(y_B\mid\theta=0)=0.5(0.98)+0.5(0.82)=0.90$. Ordinary joint Bayes therefore changes the upstream probability to

P(\theta=1\mid y_A,y_B) = \frac{0.8(0.10)}{0.8(0.10)+0.2(0.90)} = \frac{0.08}{0.26} = 0.3077.

The downstream evidence strongly favors $\theta=0$, so joint updating moves the upstream probability from 0.8 to 0.3077. Under the cut, $P_{\mathrm{cut}}(\theta=1\mid y_A,y_B)=0.8$ by construction. The downstream conditional probabilities are $P(\phi=1\mid\theta=1,y_B)=0.18/(0.18+0.02)=0.9$ and $P(\phi=1\mid\theta=0,y_B)=0.98/(0.98+0.82)=0.5444$. The cut marginal for $\phi$ is consequently

P_{\mathrm{cut}}(\phi=1\mid y_A,y_B) = 0.8(0.9)+0.2(0.5444) = 0.8289.

For comparison, the joint posterior gives $P(\phi=1\mid y_A,y_B)=0.3077(0.9)+0.6923(0.5444)=0.6538$. This difference is not an arithmetic discrepancy: it is the consequence of allowing or blocking downstream feedback.

Propagating Upstream Uncertainty

A cut should generally pass a distribution of upstream values, not only a point estimate. The downstream posterior is a mixture over that upstream distribution:

p_{\mathrm{cut}}(\phi\mid y_A,y_B) = \int p(\phi\mid\theta,y_B)\,p(\theta\mid y_A)\,d\theta.

This retains uncertainty about the upstream component while preventing the downstream likelihood from changing that component’s posterior. Replacing $\theta$ with its posterior mean instead is a plug-in analysis; it can understate downstream uncertainty and can miss nonlinear changes in $p(\phi\mid\theta,y_B)$.

Worked Example: A Cut Mixture for a Downstream Parameter

Suppose $\theta$, an exposure quantity measured in mg/L, has trusted posterior values 8, 10, and 12 mg/L with probabilities 0.25, 0.50, and 0.25. A downstream response parameter $\phi$, in response units per mg/L, has prior $N(0,1)$. For one hypothetical response observation $y_B=6$ response units, assume $y_B\mid\theta,\phi\sim N(\theta\phi,16)$, where 16 is the observation variance.

For fixed $\theta$, normal conjugate updating gives posterior variance $v(\theta)=1/(1+\theta^2/16)$ and mean $m(\theta)=6\theta/(16+\theta^2)$. Substituting each exposure value:

$\theta$ (mg/L)	$m(\theta)$	$v(\theta)$
8	$48/80=0.6000$	$16/80=0.2000$
10	$60/116=0.5172$	$16/116=0.1379$
12	$72/160=0.4500$	$16/160=0.1000$

The cut posterior for $\phi$ is the corresponding mixture, not a single normal distribution. Its mean is

E_{\mathrm{cut}}(\phi\mid y_A,y_B) = 0.25(0.6000)+0.50(0.5172)+0.25(0.4500) = 0.5211.

By the law of total variance, its variance is the weighted average of the conditional variances plus the variance of the conditional means:

\begin{aligned} \operatorname{Var}_{\mathrm{cut}}(\phi\mid y_A,y_B) &= \sum_j w_j v(\theta_j) +\sum_j w_j\{m(\theta_j)-0.5211\}^2\\ &\approx 0.14397+0.00283 =0.1468. \end{aligned}

Thus the cut posterior standard deviation is $\sqrt{0.1468}=0.3831$ response units per mg/L. Fixing exposure at 10 mg/L would give the conditional posterior mean 0.5172 and standard deviation $\sqrt{0.1379}=0.3714$. That plug-in calculation omits variation across the three plausible upstream values. The cut mixture preserves that variation and its effect on the response parameter.

Derivation and Practical Computation

The cut construction can be justified directly by combining the two required conditional distributions. Draw $\theta$ from the upstream posterior, then draw $\phi$ from its downstream posterior conditional on that draw. The resulting joint density is the product $p(\theta\mid y_A)p(\phi\mid\theta,y_B)$, which integrates to one because each factor is normalized. Integrating over $\phi$ returns the unchanged upstream marginal:

\int p_{\mathrm{cut}}(\theta,\phi\mid y_A,y_B)\,d\phi = p(\theta\mid y_A)\int p(\phi\mid\theta,y_B)\,d\phi = p(\theta\mid y_A).

Integrating over $\theta$, in contrast, gives the mixture posterior for $\phi$. A practical Monte Carlo procedure follows the definition:

Fit the upstream module.
Obtain draws $\theta^{(1)},\ldots,\theta^{(M)}$ from $p(\theta\mid y_A)$, retaining relevant dependence among upstream parameters.

Update downstream conditionally.
For each upstream draw, obtain one or more draws from $p(\phi\mid\theta^{(m)},y_B)$, using the downstream likelihood and its specified conditional prior.

Summarize the cut draws.
Use the resulting paired draws to estimate downstream summaries and predictions. Do not reweight the $\theta$ draws using the downstream marginal likelihood if feedback is meant to be blocked.

The last step is crucial: weighting upstream draws by $m_B(y_B\mid\theta)$ returns to joint Bayesian updating. If the downstream conditional fit is computationally difficult, numerical integration or nested simulation can approximate the same mixture, but approximation error should be distinguished from the deliberate cut.

When a Cut Is Defensible—and How to Check It

A cut is most defensible when the direction of information flow follows the scientific design and there is a clear reason not to trust the reverse update. For example, the upstream model may be supported by measurements specific to that component, while downstream observations depend on a weakly validated structural assumption. The decision is not a general remedy for model misspecification: it states which evidence is permitted to update which parameters.

Before using a cut, specify the modules, their data, their shared or interface parameters, and the direction of permitted information. The conditional downstream model must still be sensible across the range of upstream values that the cut passes forward. If the downstream likelihood or prior is wrong, cutting feedback does not make downstream inference reliable. Likewise, if data are reused across modules without accounting for their dependence, a cut does not by itself prevent double counting.

Useful diagnostics include checking convergence and effective sample sizes within each module, examining conditional downstream fits at low, central, and high upstream draws, and confirming that the empirical upstream draws are unchanged after the downstream step. If a downstream fit fails for some plausible $\theta$, that is a substantive warning: the interface or conditional model may not support the upstream uncertainty range. Sensitivity analysis can compare the cut results with a joint fit and with plausible alternative module specifications. The comparison reveals how consequential the feedback decision is; it does not determine automatically which analysis is scientifically preferable.

Alternatives and Limitations

Approach	Information flow	Main implication
Joint Bayesian model	Both modules can update shared or linked parameters.	Coherent under the specified joint model, but vulnerable to feedback through a poorly validated component.
Cut distribution	Upstream information flows downstream; reverse updating is blocked.	Protects the selected upstream posterior while propagating its uncertainty; generally not the posterior from one ordinary joint model.
Plug-in sequential analysis	A point estimate or fixed summary is passed downstream.	Simpler, but generally discards upstream uncertainty and may distort downstream summaries.

A cut can be conservative about feedback, but it is not automatically conservative about every quantity. It can reduce precision in the downstream module by retaining upstream uncertainty, while also preventing downstream data from improving upstream precision. Its results depend on the module boundary, the conditional prior, and the chosen data allocation. There may also be no single conventional joint posterior whose conditionals equal the cut distribution; that is an intentional consequence of enforcing modular learning, not a numerical defect.

Common pitfall: “The downstream model is uncertain, so I will keep only its best-fitting upstream value and pass that value forward.” This both blocks feedback and discards upstream uncertainty. A cut instead retains $p(\theta\mid y_A)$ and averages downstream inference over that distribution. Another error is to use downstream likelihood weights on upstream draws while calling the result a cut; those weights restore feedback.

Check Your Understanding

Work through each question, then compare with the solution provided.

An upstream posterior assigns probabilities 0.6 and 0.4 to $\theta=1$ and $\theta=0$. The downstream marginal likelihoods are 0.2 and 0.8, respectively. What is the cut probability of $\theta=1$, and what is the joint posterior probability?
Solution: The cut probability remains 0.6. The joint probability is $0.6(0.2)/[0.6(0.2)+0.4(0.8)]=0.12/0.44=0.2727$. The difference arises from downstream feedback.
In one sentence, what does a cut do to the conditional distribution of the downstream parameter?
Solution: It updates the downstream parameter using downstream data conditional on each upstream value, then averages over the upstream posterior.
Why is fixing an uncertain upstream parameter at its posterior mean not generally equivalent to a cut?
Solution: A cut integrates $p(\phi\mid\theta,y_B)$ over all $\theta$ values in $p(\theta\mid y_A)$; a plug-in analysis uses only the conditional distribution at one value. In general, the mixture is not equal to that single conditional distribution.
A downstream model fails to fit for some upstream draws that have appreciable posterior probability. Is this evidence that the cut has successfully protected the analysis?
Solution: No. Blocking feedback does not establish downstream adequacy. The failures indicate that the conditional model or interface may be incompatible with plausible upstream values and require investigation.

\(\theta\) (mg/L)	\(m(\theta)\)	\(v(\theta)\)
8	\(48/80=0.6000\)	\(16/80=0.2000\)
10	\(60/116=0.5172\)	\(16/116=0.1379\)
12	\(72/160=0.4500\)	\(16/160=0.1000\)