Normal-Normal Robust Mixture Example

Why work through a normal-normal mixture?

When a clinical trial uses historical information, a single prior centered on the historical estimate can pull the current estimate toward that history even when the trial data disagree. A robust mixture makes that borrowing conditional: before seeing current data, the prior assigns some probability to a historical-data-centered component and some to a protective component. The likelihood updates both components and also changes their probabilities. This tutorial works through those calculations for a normal outcome with known sampling variance.

The examples use a hypothetical continuous endpoint measured in points. Let \(\theta\) be the current trial control-arm mean change in the endpoint. A positive value means a larger mean change in the chosen scale; its clinical interpretation depends on how the endpoint is defined. The calculations describe uncertainty about this mean and do not, by themselves, establish a treatment effect.

Model and component-specific update

Definition: Suppose the observed current-trial sample mean is \(\bar{x}\), based on \(n\) independent observations from a normal distribution with mean \(\theta\) and known standard deviation \(\sigma\). Thus \(\bar{x}\mid\theta\sim N(\theta,\sigma^2/n)\). Let \(s^2=\sigma^2/n\) denote the known sampling variance of \(\bar{x}\). A two-component robust mixture prior is \(\theta\sim w_H N(\mu_H,\tau_H^2)+(1-w_H)N(\mu_P,\tau_P^2)\), where \(H\) denotes the historical borrowing component, \(P\) the protective component, and \(w_H\) is the prior probability of the historical component. The mixture notation means that one component is selected with its stated probability, then \(\theta\) is drawn from that component.

The historical component is centered at \(\mu_H\), typically informed by historical data. The protective component has its own specified center \(\mu_P\) and scale \(\tau_P\); it is chosen to limit reliance on historical information when current data are discordant. It is not automatically flat or noninformative. The component choices, scales, and prior weight must be clinically and statistically defensible.

Assume the observations are independent conditional on \(\theta\), the normal likelihood and known \(\sigma\) are adequate, and the prior parameters and \(w_H\) are fixed before observing the current data. For component \(j\in\{H,P\}\), write its prior mean and variance as \(\mu_j\) and \(\tau_j^2\). Conjugacy gives a normal posterior within each component:

v_j=\left(\frac{1}{\tau_j^2}+\frac{1}{s^2}\right)^{-1}, \qquad m_j=v_j\left(\frac{\mu_j}{\tau_j^2}+\frac{\bar{x}}{s^2}\right).

These are conditional posteriors: \(m_H\) is the posterior mean if the historical component was selected, and \(m_P\) is the posterior mean if the protective component was selected. The posterior probability of each component is a separate calculation. Its update uses the component's predictive density for the observed sample mean:

q_j(\bar{x})= \frac{1}{\sqrt{2\pi(\tau_j^2+s^2)}} \exp\left\{-\frac{(\bar{x}-\mu_j)^2}{2(\tau_j^2+s^2)}\right\}, \qquad w_H^*= \frac{w_H q_H(\bar{x})} {w_H q_H(\bar{x})+(1-w_H)q_P(\bar{x})}.

The predictive density follows because, under component \(j\), the sample mean can be represented as \(\bar{x}=\theta+\epsilon\), with independent \(\theta\sim N(\mu_j,\tau_j^2)\) and \(\epsilon\sim N(0,s^2)\). Their sum has distribution \(N(\mu_j,\tau_j^2+s^2)\). Bayes' rule then multiplies each component's predictive density by its prior probability and normalizes the two products. The posterior probability of the protective component is \(1-w_H^*\).

The full posterior is still a mixture, now with updated weights \(w_H^*\) and \(1-w_H^*\). Its mean is \(w_H^*m_H+(1-w_H^*)m_P\). If uncertainty is summarized by a variance, it must include both the within-component variances and the separation between component means:

\operatorname{Var}(\theta\mid\bar{x}) =w_H^*\{v_H+(m_H-\bar{m})^2\} +(1-w_H^*)\{v_P+(m_P-\bar{m})^2\}, \qquad \bar{m}=w_H^*m_H+(1-w_H^*)m_P.

Worked Example: Updating a Historical and Protective Component

Suppose a hypothetical trial has \(n=25\) independent control-arm observations, known outcome standard deviation \(\sigma=10\) points, and observed sample mean \(\bar{x}=14\) points. The resulting standard error is \(SE=\sigma/\sqrt{n}=10/5=2\) points, so \(s^2=4\) points squared. Use the prior mixture with historical component \(N(10,4)\), protective component \(N(0,100)\), and prior historical weight \(w_H=0.80\). The historical center is 10 points; the broader protective component is centered at 0 points. The calculation assumes the specified normal sampling model and known \(\sigma\).

1. Update within each component. For the historical component, \(v_H=(1/4+1/4)^{-1}=2\) points squared and \(m_H=2(10/4+14/4)=12\) points. For the protective component, \(v_P=(1/100+1/4)^{-1}=3.846\) points squared and \(m_P=3.846(0/100+14/4)=13.46\) points.

2. Calculate predictive densities. The historical predictive variance is \(4+4=8\) points squared, so \(q_H(14)=\{2\pi(8)\}^{-1/2}\exp\{-(14-10)^2/(2(8))\}=0.05187\) per point. The protective predictive variance is \(100+4=104\) points squared, so \(q_P(14)=\{2\pi(104)\}^{-1/2}\exp\{-(14-0)^2/(2(104))\}=0.01524\) per point.

3. Reweight the components. The unnormalized historical and protective weights are \(0.80(0.05187)=0.04150\) and \(0.20(0.01524)=0.003048\). Therefore \(w_H^*=0.04150/(0.04150+0.003048)=0.9316\). Given these assumptions and this sample mean, the posterior probability of the historical component is about 93.16%, and that of the protective component is about 6.84%.

4. Summarize the mixture. The overall posterior mean is \(0.9316(12)+0.0684(13.46)=12.10\) points. This is not either component mean: it averages the component posterior means using their updated probabilities. The posterior variance is approximately \(0.9316\{2+(12-12.10)^2\}+0.0684\{3.846+(13.46-12.10)^2\}=2.262\) points squared, giving posterior standard deviation \(\sqrt{2.262}=1.504\) points.

How conflict changes the borrowing probability

The mixture does not make a fixed decision to borrow or not borrow. The current data affect component probabilities through their predictive densities: a component that makes the observed sample mean relatively plausible gains posterior probability. This reweighting is distinct from the within-component update, which moves each component's posterior mean toward the observed sample mean.

Worked Example: Strong Disagreement With the Historical Center

Keep the hypothetical model, prior components, prior weight, and standard error from the previous example, but suppose the current trial instead observes \(\bar{x}=24\) points. These data are 14 points from the historical center and 24 points from the protective center. The sample size, known standard deviation, and normal model are unchanged, so the likelihood assumptions remain the same.

The predictive densities are \(q_H(24)=\{2\pi(8)\}^{-1/2}\exp\{-(24-10)^2/16\}=6.75\times10^{-7}\) per point and \(q_P(24)=\{2\pi(104)\}^{-1/2}\exp\{-(24-0)^2/208\}=0.002452\) per point. Weighting by the prior probabilities gives \(0.80q_H(24)=5.40\times10^{-7}\) and \(0.20q_P(24)=0.0004904\). Thus \(w_H^*\) is approximately \(0.00110\): the posterior probability of the historical component is about 0.11%, and the protective component has probability about 99.89%.

The component posterior means are \(m_H=2(10/4+24/4)=17\) points and \(m_P=3.846(0/100+24/4)=23.08\) points. The overall posterior mean is approximately \(0.00110(17)+0.99890(23.08)=23.07\) points. The protective component dominates because its predictive density for these data is much larger. It does not discard all prior influence: within that component, the posterior mean is still slightly below the observed 24 points.

Posterior probabilities and interpretation

Because a normal mixture is generally not a single normal distribution, probability statements should be calculated component by component and then combined with posterior weights. For an event such as \(\theta>c\), the calculation is \(P(\theta>c\mid\bar{x})=w_H^*P_H(\theta>c\mid\bar{x})+(1-w_H^*)P_P(\theta>c\mid\bar{x})\). Each conditional probability is evaluated under its component's normal posterior.

Worked Example: Probability the Mean Exceeds a Threshold

Return to the hypothetical \(\bar{x}=14\) analysis, where the component posterior means and variances are \(m_H=12\), \(v_H=2\), \(m_P=13.46\), \(v_P=3.846\), and \(w_H^*=0.9316\). Calculate the posterior probability that the mean change exceeds 12 points.

Under the historical component, the standardized threshold is \((12-12)/\sqrt{2}=0\), so \(P_H(\theta>12\mid\bar{x})=0.5000\). Under the protective component, it is \((12-13.46)/\sqrt{3.846}=-0.745\), giving \(P_P(\theta>12\mid\bar{x})\approx0.772\). Therefore the mixture probability is \(0.9316(0.5000)+0.0684(0.772)=0.5186\). Under this specified model and prior, the posterior probability that the mean exceeds 12 points is about 51.86%.

This probability is about the parameter conditional on the stated model, prior, and data. It is not the posterior probability that a particular prior component is correct in an absolute sense: the component indicator is a device in the specified mixture model, and its posterior probability is conditional on that specification.

Checks, alternatives, and limitations

Check the likelihood scale.
Confirm \(SE=\sigma/\sqrt{n}\) for the assumed independent observations and that the variance used in predictive densities is \(SE^2+\tau_j^2\), not \(SE+\tau_j\).

Check the component weights.
Each unnormalized weight is prior probability times predictive density. After normalization, the two posterior weights must be between 0 and 1 and sum to 1.

Check the conditional updates.
Each component posterior mean is a precision-weighted average of its prior mean and \(\bar{x}\), so it must lie between those two values when both variances are positive.

Check sensitivity and fit.
Inspect whether the observed mean is surprising under both component predictive distributions. If both densities are extremely small, reweighting alone does not make the assumed model adequate; investigate the endpoint, sampling model, and prior specification.

A single historical-centered normal prior is simpler, but it cannot shift posterior probability away from historical borrowing when data conflict with that prior. A broad protective prior alone avoids a strong historical pull but forgoes the mixture's possibility of substantial borrowing when data are compatible. The robust mixture combines these possibilities, at the cost of having to choose and justify two components and an initial mixture weight. Its behavior depends on all those choices, not just on the current data.

The calculation also relies on a known sampling variance and a normal model for the sample mean. If the outcome distribution, independence assumption, or variance specification is poor, the component probabilities can be misleading. A small predictive density identifies tension between the data and a component under the model; it does not diagnose why the tension occurred. The mixture is not a substitute for checking data quality, clinical comparability, or the endpoint definition.

Common mistake

Common pitfall: A reader may see a large posterior historical weight and conclude that the current estimate is “93% historical data,” or may calculate the overall posterior mean using the original prior weight of 0.80. Neither conclusion is correct. The posterior weight is the conditional probability of a component under the specified mixture, updated by the current data; the overall mean uses that posterior weight, not the prior weight. It is not a literal percentage of observations contributed by the historical data.

Check Your Understanding

Use the model and formulas in this tutorial. The solutions are included so that each calculation can be checked.

For a component prior \(N(5,9)\) and a sample mean with known sampling variance \(s^2=4\), calculate the component posterior variance and mean when \(\bar{x}=9\).
In one sentence, explain why the predictive variance for a component is \(\tau_j^2+s^2\).
If the historical and protective components have equal prior weights, but their predictive densities for the observed data are 0.03 and 0.01, respectively, what is the posterior historical weight?
For a two-component posterior with weights 0.7 and 0.3, component means 2 and 5, and component variances 1 and 4, calculate the overall posterior mean.
Why is it incorrect to describe the posterior historical weight as the percentage of current patients whose outcomes came from the historical component?

Solutions. (1) The posterior variance is \((1/9+1/4)^{-1}=36/13=2.769\). The posterior mean is \((36/13)(5/9+9/4)=7.615\). (2) Under the component, the sample mean is the sum of a draw of \(\theta\), with variance \(\tau_j^2\), and independent sampling error, with variance \(s^2\); variances add. (3) The posterior historical weight is \(0.5(0.03)/\{0.5(0.03)+0.5(0.01)\}=0.75\). (4) The mean is \(0.7(2)+0.3(5)=2.9\). (5) The weight is a posterior probability for a component in the specified prior model, not a count or fraction of patients or observations.