Mixture Priors for Treatment Effects

Why Represent Treatment Effects With a Mixture?

A single prior distribution can summarize uncertainty about a treatment effect, but a mixture prior makes distinct scientific possibilities explicit. For example, the evidence available before an interim analysis might support both a skeptical view, in which the effect is small, and a more optimistic view, in which it is meaningfully beneficial. A predictive probability calculation can show not only the overall chance of meeting the success criterion, but also how much each possibility contributes to that chance.

This tutorial considers success at one planned final analysis, given the data observed at an interim analysis. It uses a normal, independent-increments model with known information. The introductory material on predictive probability establishes the general idea of averaging future success over uncertainty; the focus here is how that averaging works when treatment-effect uncertainty is represented by distinct mixture components.

Definition: A mixture prior for a treatment effect \(\theta\) has the form \(p(\theta)=\sum_{k=1}^{K}\pi_k p_k(\theta)\), where \(p_k\) is the prior distribution under component \(k\), \(\pi_k\geq 0\), and \(\sum_{k=1}^{K}\pi_k=1\). The component label \(k\) identifies a plausible treatment-effect model. After interim data are observed, the component weights generally change; the updated weights are the posterior probabilities of the components.

Model and Assumptions

Let \(I_t\) be the information at the interim analysis and \(I_T\) the planned total information at the final analysis, with \(0<I_t<I_T\). The effect estimate at the interim analysis is \(\hat{\theta}_t\). Assume that, conditional on the true effect \(\theta\), the interim estimate follows \(\hat{\theta}_t\mid\theta\sim N(\theta,1/I_t)\). Assume also that the future information increment is independent of the interim data conditional on \(\theta\), and that the final statistic uses the information accumulated from both parts.

For a one-sided efficacy test in which larger positive values favor treatment, define the final statistic as \(Z_T=\sqrt{I_T}\hat{\theta}_T\). Success occurs if \(Z_T\geq c\), where \(c\) is the prespecified final efficacy boundary. This setup requires a normal approximation with known information, a fixed planned final information level, and a valid independent-increments representation. If the design instead has a different final success rule or uses a different sampling model, the component probabilities must be calculated for that rule and model.

The mixture prior is specified before seeing the interim data. Once \(\hat{\theta}_t\) is observed, both the plausibility of each component and, for continuous components, uncertainty about \(\theta\) within that component must be accounted for. Weighting component success probabilities by the original \(\pi_k\) alone would generally ignore the information in the interim estimate.

From Interim Data to Predictive Success

Under the independent-increments model, the future estimate based only on the remaining information has conditional distribution \(N(\theta,1/(I_T-I_t))\). Combining that future estimate with the interim estimate gives the final statistic’s conditional distribution:

Z_T\mid \hat{\theta}_t,\theta \sim N\left( \frac{I_t}{\sqrt{I_T}}\hat{\theta}_t+ \frac{I_T-I_t}{\sqrt{I_T}}\theta,\, \frac{I_T-I_t}{I_T} \right).

The first term in the mean reflects the information already observed; the second reflects the future information under effect \(\theta\). The variance is due to future sampling, not uncertainty about \(\theta\). Therefore, with a fixed value of \(\theta\), the conditional probability of final success is:

g(\theta)=P(Z_T\geq c\mid\hat{\theta}_t,\theta) = \Phi\left[ \frac{ \frac{I_t}{\sqrt{I_T}}\hat{\theta}_t+ \frac{I_T-I_t}{\sqrt{I_T}}\theta-c }{ \sqrt{(I_T-I_t)/I_T} } \right],

where \(\Phi\) is the standard normal cumulative distribution function. This is a conditional success probability at a particular effect, not yet a predictive probability under the mixture prior.

Updating and Combining the Mixture Components

Write \(q_k=P(k\mid\hat{\theta}_t)\) for the posterior probability of component \(k\). Bayes’ rule gives:

q_k= \frac{\pi_k\,m_k(\hat{\theta}_t)} {\sum_{j=1}^{K}\pi_j\,m_j(\hat{\theta}_t)}, \qquad m_k(\hat{\theta}_t)= \int f(\hat{\theta}_t\mid\theta)p_k(\theta)\,d\theta.

Here \(m_k(\hat{\theta}_t)\) is the interim-data density under component \(k\). Within each component, average \(g(\theta)\) over that component’s posterior distribution for \(\theta\). Call the resulting probability \(p_k=P(Z_T\geq c\mid\hat{\theta}_t,k)\). The overall predictive probability is then:

PP=\sum_{k=1}^{K}q_k p_k.

This decomposition separates two sources of information: \(p_k\) describes the chance of success if component \(k\) applies, while \(q_k\) describes how plausible that component is after the interim data. The product \(q_kp_k\) is component \(k\)’s contribution to the overall predictive probability.

Derivation for normal components: Suppose component \(k\) has prior \(\theta\mid k\sim N(\mu_k,\tau_k^2)\), with \(\tau_k^2>0\). Combining it with \(\hat{\theta}_t\mid\theta\sim N(\theta,1/I_t)\) gives posterior variance \(v_k=(1/\tau_k^2+I_t)^{-1}\) and posterior mean \(m_k=v_k(\mu_k/\tau_k^2+I_t\hat{\theta}_t)\). Conditional on the interim data and component \(k\), the mean of \(Z_T\) is \(I_t\hat{\theta}_t/\sqrt{I_T}+(I_T-I_t)m_k/\sqrt{I_T}\). Integrating over the posterior uncertainty in \(\theta\) adds \(((I_T-I_t)/\sqrt{I_T})^2v_k\) to the future-sampling variance. Thus \(Z_T\mid\hat{\theta}_t,k\) is normal with that mean and variance \((I_T-I_t)/I_T+((I_T-I_t)^2/I_T)v_k\). Its upper-tail probability above \(c\) is \(p_k\). This follows because the final statistic is a linear function of the normal posterior effect and an independent normal future increment.

If component \(k\) instead assigns a fixed effect \(\theta_k\), its posterior variance is zero and \(p_k=g(\theta_k)\). Its posterior component weight is proportional to its prior weight times the interim likelihood evaluated at \(\theta_k\). Thus point-mass mixtures are a direct special case of the same weighted calculation.

Worked Example: Updating a Two-Point Mixture

Suppose a hypothetical trial has \(I_t=25\), \(I_T=100\), and interim estimate \(\hat{\theta}_t=0.12\), in treatment-effect units. Use a one-sided final boundary \(c=1.645\). The prior assigns probability \(0.5\) to each fixed effect, \(\theta_1=0.05\) and \(\theta_2=0.25\). Calculate the posterior component weights and the predictive probability of success.

The interim standard error is \(1/\sqrt{25}=0.20\). The likelihood ratio for component 2 versus component 1 is:

\frac{f(0.12\mid\theta_2)}{f(0.12\mid\theta_1)} = \exp\left\{-\frac{25}{2}\left[(0.12-0.25)^2-(0.12-0.05)^2\right]\right\} = \exp(-0.15) = 0.8607.

With equal prior weights, the posterior weights are \(q_2=0.8607/(1+0.8607)=0.4625\) and \(q_1=1-q_2=0.5375\). The final-statistic standard deviation conditional on a fixed effect is \(\sqrt{(100-25)/100}=\sqrt{0.75}=0.8660\). Its mean is \(25(0.12)/10+75\theta/10=0.30+7.5\theta\). For \(\theta_1=0.05\), this mean is \(0.675\), so \(p_1=\Phi((0.675-1.645)/0.8660)=\Phi(-1.120)=0.1313\). For \(\theta_2=0.25\), it is \(2.175\), so \(p_2=\Phi((2.175-1.645)/0.8660)=\Phi(0.612)=0.7300\).

The component contributions are \(q_1p_1=0.5375(0.1313)=0.0706\) and \(q_2p_2=0.4625(0.7300)=0.3376\). Therefore \(PP=0.0706+0.3376=0.4082\), or about 40.8%. The more favorable effect has a somewhat smaller posterior weight, but supplies most of the predictive probability because success is much more likely under that effect.

Worked Example: Normal Components Include Within-Component Uncertainty

Consider another hypothetical calculation with \(I_t=25\), \(I_T=100\), \(\hat{\theta}_t=0.12\), and \(c=1.645\). Let the prior be a two-component mixture: component 1 is \(N(0,0.10^2)\) with weight \(0.60\), and component 2 is \(N(0.30,0.10^2)\) with weight \(0.40\). Find the updated weights, component predictive probabilities, and overall probability.

The marginal interim estimate under either component has variance \(0.10^2+1/25=0.05\). The likelihood ratio of component 2 to component 1 is:

\frac{m_2(0.12)}{m_1(0.12)} = \exp\left\{-\frac{(0.12-0.30)^2-(0.12-0)^2}{2(0.05)}\right\} = \exp(-0.18) = 0.8353.

Thus \(q_2=0.40(0.8353)/[0.60+0.40(0.8353)]=0.3577\), and \(q_1=0.6423\). In each component, the posterior variance is \(v=(100+25)^{-1}=0.008\). The posterior means are \(m_1=0.008(25(0.12))=0.024\) and \(m_2=0.008(100(0.30)+25(0.12))=0.264\).

The predictive variance of \(Z_T\) within either component is \(0.75+(75^2/100)(0.008)=1.20\), with standard deviation \(1.095\). The predictive means are \(0.30+7.5(0.024)=0.480\) and \(0.30+7.5(0.264)=2.280\). Therefore \(p_1=\Phi((0.480-1.645)/1.095)=\Phi(-1.063)\approx0.144\), and \(p_2=\Phi((2.280-1.645)/1.095)=\Phi(0.580)\approx0.719\). The weighted contributions are approximately \(0.6423(0.144)=0.0925\) and \(0.3577(0.719)=0.2572\). Their sum is \(PP\approx0.350\), or about 35.0%.

Although the interim estimate shifts weight toward component 1 relative to the original 0.60/0.40 weights, the optimistic component still contributes more to the success probability. Its contribution is a weighted probability, not a posterior probability that the component is true.

Worked Example: Reading Contributions in a Three-Component Mixture

Suppose a hypothetical trial has \(I_t=25\), \(I_T=100\), \(\hat{\theta}_t=0.20\), and \(c=1.645\). A prior mixture assigns weights \(0.20\), \(0.50\), and \(0.30\) to fixed effects \(-0.05\), \(0.10\), and \(0.30\), respectively. Calculate the component contributions and identify which component contributes most to predicted success.

The interim standard error is \(0.20\), so the likelihood for an effect \(\theta\), apart from a common factor, is \(\exp[-12.5(0.20-\theta)^2]\). The three weighted likelihoods are \(0.20e^{-0.78125}=0.0916\), \(0.50e^{-0.125}=0.4412\), and \(0.30e^{-0.125}=0.2648\). Their sum is \(0.7976\), giving posterior weights \(q_1=0.1148\), \(q_2=0.5534\), and \(q_3=0.3319\).

For fixed effects, the future standard deviation is \(0.8660\) and the final-statistic mean is \(25(0.20)/10+7.5\theta=0.50+7.5\theta\). The means for the three components are \(0.125\), \(1.250\), and \(2.750\). Hence \(p_1=\Phi((0.125-1.645)/0.8660)\approx0.0396\), \(p_2=\Phi((1.250-1.645)/0.8660)\approx0.324\), and \(p_3=\Phi((2.750-1.645)/0.8660)\approx0.899\).

The contributions are \(0.1148(0.0396)=0.0045\), \(0.5534(0.324)=0.1793\), and \(0.3319(0.899)=0.2983\). Thus \(PP\approx0.482\), or 48.2%. The third component contributes most, despite having a smaller posterior weight than the second, because its probability of success conditional on that effect is much larger.

Interpretation, Checks, and Alternatives

A mixture calculation is useful when materially different effects remain plausible and a single plug-in effect would hide that uncertainty. Reporting \(PP\), the posterior weights, the component probabilities, and the weighted contributions makes the calculation interpretable: readers can see whether predicted success is broad-based or depends mainly on a low-weight optimistic component.

Several checks help reveal errors or fragile conclusions:

Confirm that the posterior component weights are nonnegative and sum to one, and that each component contribution is \(q_kp_k\).
Check that each \(p_k\) is between zero and one and that \(PP\), as a weighted average, lies between the smallest and largest \(p_k\).
Verify the direction of the boundary and the sign of the treatment effect. If larger effects are beneficial, increasing a fixed \(\theta\) should increase the one-sided success probability.
Make sure the calculation includes uncertainty within continuous components. Using only each component’s posterior mean in \(g(\theta)\) omits posterior uncertainty and is not generally the same as integrating over the component posterior.
Check that the chosen information levels and independent-increments approximation are appropriate for the design. A changing analysis plan, a materially inaccurate normal approximation, or an incorrectly specified effect scale can invalidate the calculation.

A common mistake is to use the original mixture weights \(\pi_k\) after seeing the interim data. For example, that would retain the prior weights 0.60 and 0.40 in the normal-component example, even though the observed estimate updates them to about 0.642 and 0.358. The correction is to calculate \(q_k\) from each component’s marginal interim likelihood before averaging component success probabilities.

Another tempting alternative is plug-in conditional power: select one assumed treatment effect and calculate future success at that value. It is simpler, but does not average over distinct effect possibilities or update their relative plausibility. A mixture predictive probability does both, at the cost of depending on the selected components and prior weights. It is not an objective probability that the trial will succeed under every reasonable prior, nor does a high value establish that the treatment works. Results can be sensitive to component choices, weights, the success boundary, and the assumed model. The mixture representation should therefore be scientifically defensible and its components should be reported clearly.

The formulas here concern one planned final look. With multiple future looks, success can occur at more than one analysis and the future statistics are dependent; a single-final-look component probability cannot simply be treated as the probability of crossing any boundary in that design.

Check Your Understanding

Answer each question using the definitions and formulas above. A short solution follows the questions.

Why are prior mixture weights generally not the weights used to combine component success probabilities after an interim analysis?
A two-component mixture has posterior weights \(q_1=0.7\) and \(q_2=0.3\), with component predictive probabilities \(p_1=0.2\) and \(p_2=0.8\). What is \(PP\), and which component contributes more to it?
For a normal prior component, what two sources of variability contribute to the predictive variance of the final statistic?
If every component probability is between 0.25 and 0.60, can the mixture predictive probability equal 0.75? Explain.
Give one limitation of interpreting a mixture predictive probability as a general statement about the treatment.

Solutions. (1) Interim data change the relative plausibility of the components, so Bayes’ rule updates \(\pi_k\) to \(q_k\). (2) \(PP=0.7(0.2)+0.3(0.8)=0.38\); component 1 contributes \(0.14\), while component 2 contributes \(0.24\), so component 2 contributes more. (3) Future sampling variability and posterior uncertainty about \(\theta\) within the component both contribute. (4) No. A weighted average with nonnegative weights summing to one must lie between the component probabilities, so it cannot exceed 0.60. (5) The result depends on the specified components, weights, and model; it is a predictive probability under those assumptions, not evidence that the treatment works under every plausible model.