Contamination Priors and Limited Borrowing

Why Add a Contamination Component?

Historical data can improve precision when the historical and current trial settings are sufficiently compatible. But when the historical evidence conflicts with the current trial, a prior that concentrates heavily near the historical estimate can pull the current-trial inference in an undesirable direction. A contamination prior represents uncertainty about whether that concentrated historical-informed component is adequate: most prior mass is placed on it, while a smaller amount is assigned to a protective component that is less restrictive.

The protective component does not automatically identify the right explanation for a discrepancy. It provides a route for the current data to receive substantial posterior weight even when those data would have been surprising under the historical-informed component. The key calculation is therefore not just the prior mixture weight: it is the posterior mixture weight, which responds to how well each component predicts the current data.

Definition: Let $\theta$ be the parameter of interest, $D_H$ the historical data, and $D$ the current-trial data. A contamination prior is a mixture $$ \pi(\theta\mid D_H)=(1-\epsilon)\pi_H(\theta\mid D_H)+\epsilon\pi_C(\theta), $$ where $\pi_H$ is the historical-data-informed component, $\pi_C$ is a protective component, and $0<\epsilon<1$ is the prior contamination probability. The component $\pi_C$ is usually broader or otherwise more protective of the current-trial inference than $\pi_H$.

This construction is meaningful when both components are proper distributions and the current-data likelihood is evaluated under the same sampling model for either component. The model must describe the current observations conditional on $\theta$; the prior components encode different degrees of reliance on historical information. In a clinical trial, $\theta$ might be a treatment effect or a control-event probability. The component weights do not, by themselves, account for every source of bias or differences in trial conduct.

How Current Data Update the Mixture

Write $L(\theta;D)$ for the likelihood of the current data. For component $j\in\{H,C\}$, its prior predictive density (or probability, for discrete data) is $m_j(D)=\int L(\theta;D)\pi_j(\theta)\,d\theta$. It measures how well that component predicts the observed current data before those data are used to update the component-specific distribution.

\pi(\theta\mid D,D_H) =w_H(D)\pi_H(\theta\mid D,D_H) +w_C(D)\pi_C(\theta\mid D), \qquad w_H(D)+w_C(D)=1,

w_H(D)=\frac{(1-\epsilon)m_H(D)} {(1-\epsilon)m_H(D)+\epsilon m_C(D)}, \qquad w_C(D)=\frac{\epsilon m_C(D)} {(1-\epsilon)m_H(D)+\epsilon m_C(D)}.

Here $\pi_H(\theta\mid D,D_H)$ is the historical component updated with the current data, and $\pi_C(\theta\mid D)$ is the protective component updated with those data. These are ordinary component-specific posterior distributions. The overall posterior is their mixture, with weights adjusted by the components’ predictive performance.

Theorem: The posterior odds of the historical component against the protective component equal the prior component odds multiplied by the ratio of their current-data predictive densities: $$ \frac{w_H(D)}{w_C(D)} =\frac{1-\epsilon}{\epsilon}\frac{m_H(D)}{m_C(D)}. $$ Thus, if $m_H(D)/m_C(D)\le r$, then $$ w_H(D)\le \frac{(1-\epsilon)r}{(1-\epsilon)r+\epsilon}. $$

To justify the result, multiply the prior mixture by the common likelihood. The contribution from component $H$ to the unnormalized posterior is $(1-\epsilon)L(\theta;D)\pi_H(\theta)$; integrating over $\theta$ gives $(1-\epsilon)m_H(D)$. The corresponding integral for component $C$ is $\epsilon m_C(D)$. Dividing each contribution by their sum gives the two posterior weights. Taking their ratio gives the odds formula. For the bound, substitute $m_H(D)/m_C(D)\le r$ into the odds expression and convert odds $a$ to weight $a/(1+a)$, an increasing function of $a$.

This is the precise sense in which the contamination component can limit borrowing under conflict: poor predictive performance of the historical component reduces its posterior weight. It does not impose a universal upper bound of $1-\epsilon$ on that weight. Compatible data can increase the historical component’s weight; sufficiently conflicting data can reduce it. If the protective component assigns zero predictive probability to the observed data, it cannot protect against that outcome, so positive support over plausible current-trial outcomes is essential.

Worked Normal Example: Strong Conflict

Worked Example: A Current Estimate Far From the Historical Center

Suppose $\theta$ is a treatment effect measured in log-hazard-ratio units. A hypothetical current trial gives $Y\mid\theta\sim N(\theta,0.25)$, with observed estimate $y=3$ and known sampling variance $0.25$. Assume the historical-informed prior is $N(0,0.25)$, the protective prior is $N(0,4)$, and $\epsilon=0.10$. These are modeling assumptions for this example, not data from a real study.

The component predictive distributions for $Y$ are normal with mean zero. Their variances are the sum of the prior variance and sampling variance: $0.25+0.25=0.50$ for component $H$, and $4+0.25=4.25$ for component $C$. At $y=3$,

m_H(3)=\frac{1}{\sqrt{2\pi(0.5)}}\exp\left(-\frac{3^2}{2(0.5)}\right) \approx 6.963\times10^{-5},

m_C(3)=\frac{1}{\sqrt{2\pi(4.25)}}\exp\left(-\frac{3^2}{2(4.25)}\right) \approx 0.06712.

The posterior historical-component weight is therefore

w_H(3)=\frac{0.9(0.00006963)} {0.9(0.00006963)+0.1(0.06712)} \approx 0.00925, \qquad w_C(3)\approx0.99075.

The component posterior means are obtained by normal updating. For component $H$, the posterior variance is $(1/0.25+1/0.25)^{-1}=0.125$, and the posterior mean is $1.5$. For component $C$, the posterior variance is $(1/4+1/0.25)^{-1}=1/4.25$, and the posterior mean is $3(4)/(4.25)\approx2.824$. The mixture posterior mean is

E(\theta\mid y)=0.00925(1.5)+0.99075(2.824)\approx2.811.

Although the historical component had 90% prior weight, it has less than 1% posterior weight here. Under the assumed model, the observed estimate is much more predictive under the broad protective component. The posterior mean remains slightly below the current estimate because the historical component has nonzero posterior weight; it is not forced to equal the current estimate.

Worked Normal Example: Compatible Data

Worked Example: The Same Prior Under Compatibility

Keep the two prior components, the sampling variance $0.25$, and $\epsilon=0.10$ from the previous hypothetical setup, but suppose the observed current estimate is $y=0$. The model and component variances remain unchanged. The predictive densities at zero are

m_H(0)=\frac{1}{\sqrt{2\pi(0.5)}}\approx0.5642, \qquad m_C(0)=\frac{1}{\sqrt{2\pi(4.25)}}\approx0.1935.

Consequently,

w_H(0)=\frac{0.9(0.5642)} {0.9(0.5642)+0.1(0.1935)} \approx0.9633.

The current result is well aligned with the concentrated historical component, so its posterior weight rises from 0.90 to about 0.963. The contrast with the $y=3$ calculation shows why $\epsilon$ alone is not the amount of borrowing: compatibility, as measured by the component predictive densities, also matters.

Worked Beta-Binomial Example: A Binary Endpoint

Worked Example: Updating After One Current-Trial Non-Event

For a hypothetical binary endpoint, let $p$ be the current-trial event probability. Suppose the historical-informed component is $p\sim\operatorname{Beta}(8,2)$, the protective component is $p\sim\operatorname{Beta}(1,1)$, and $\epsilon=0.10$. One current-trial participant is observed and has no event, so $n=1$ and $x=0$. Assume the Bernoulli model and conditional independence of outcomes given $p$.

The predictive probability of no event is the prior expectation of $1-p$. Under the historical component it is $2/(8+2)=0.20$; under the uniform protective component it is $1/2=0.50$. Thus

w_H(0)=\frac{0.9(0.20)}{0.9(0.20)+0.1(0.50)} =\frac{0.18}{0.23}\approx0.7826.

The updated component distributions are $\operatorname{Beta}(8,3)$ and $\operatorname{Beta}(1,2)$, with means $8/11\approx0.7273$ and $1/3\approx0.3333$, respectively. The posterior mixture mean is

E(p\mid x=0)=0.7826\left(\frac{8}{11}\right) +0.2174\left(\frac{1}{3}\right) \approx0.6416.

The single non-event favors the protective component predictively, but not enough to overcome the prior odds strongly. The posterior mean event probability is about 0.642. This example also emphasizes that a protective component need not dominate after every observation that moves away from the historical mean; the size of the update depends on the predictive evidence and the prior mixture weight.

Interpretation, Diagnostics, and Limitations

For any scalar parameter with finite component posterior means, the overall posterior mean has the form $w_H E_H(\theta\mid D)+w_C E_C(\theta\mid D)$. The historical component’s contribution to the difference from the protective-component mean is therefore $w_H\{E_H(\theta\mid D)-E_C(\theta\mid D)\}$. A low $w_H$ limits that weighted contribution. This is a statement about the mixture representation, not a guarantee that the numerical estimate is close to a particular target when component means are extremely far apart.

Useful diagnostics include examining the two predictive densities or their ratio, the posterior weights, and the component-specific posterior summaries. For continuous data, predictive densities have units and should not be interpreted as probabilities of a single exact observation; their ratio is the relevant dimensionless comparison when the same data are evaluated under both components. For discrete data, the corresponding quantities are predictive probabilities. Unexpectedly tiny predictive density under both components can indicate that neither component adequately represents the current data, rather than evidence that one component is satisfactory.

Check predictive support. If plausible trial outcomes have negligible predictive density under both components, broadening or revising the model may be needed; reweighting two poor models does not fix their misspecification.
Inspect sensitivity to $\epsilon$ and component scales. Posterior weights can be sensitive to these choices, especially when the predictive-density ratio is near the threshold where component weights change substantially.
Check model comparability. The likelihood and outcome definition should be appropriate for the current trial. Differences in eligibility, endpoint ascertainment, or follow-up may make apparent conflict reflect non-comparability rather than random variation alone.

A common mistake is to say, “The protective component has prior probability $\epsilon$, so borrowing can never exceed $1-\epsilon$.” That is incorrect: $1-\epsilon$ is the prior historical-component weight, while the posterior weight is multiplied in odds form by $m_H(D)/m_C(D)$. If the data strongly favor the historical component, $w_H$ can exceed $1-\epsilon$; if they favor the protective component, it can become very small.

A contamination prior differs from simply discounting historical data by a fixed factor. Discounting changes the effective strength of a single prior contribution; a mixture instead preserves distinct component distributions and allows current data to reweight them. A single broad prior avoids a concentrated historical component altogether, but it also forgoes its potential precision gain when compatibility holds. The mixture’s flexibility comes with limitations: conclusions depend on the protective distribution, on $\epsilon$, and on the adequacy of the likelihood. The contamination probability is a model weight, not a measured probability that the historical study is “wrong,” and the mixture does not itself diagnose why the data differ.

Key takeaway: Limited borrowing arises through predictive reweighting. A protective component can retain posterior influence when historical-informed predictions fit poorly, but no fixed posterior borrowing cap follows from the prior contamination probability alone.

Check Your Understanding

Answer each question, then compare with the solutions below.

If $\epsilon=0.05$ and $m_H(D)/m_C(D)=0.10$, calculate $w_H(D)$.
In the normal example, why is the predictive variance for component $C$ equal to $4.25$?
Can a posterior historical-component weight exceed $1-\epsilon$? Explain using the odds formula.
What does it imply if both components assign extremely low predictive density to the observed current data?

Solutions. (1) $w_H=0.95(0.10)/\{0.95(0.10)+0.05\}=0.095/0.145\approx0.6552$. (2) The protective prior variance is $4$, and the sampling variance is $0.25$; independent normal prior and sampling variation add in the prior predictive distribution, giving $4+0.25=4.25$. (3) Yes. The posterior odds equal prior odds times $m_H/m_C$; if the observed data are more predictive under $H$, the ratio exceeds one and the historical-component weight can rise above its prior value $1-\epsilon$. (4) Neither component predicts the data well, so the result may indicate model misspecification or omitted features; relative reweighting alone does not establish that either component is adequate.