Bayesian Copula Models for Combination Toxicity

Why Model Dependence Between Toxicities?

At a combination dose, a patient may experience toxicity attributed to drug A, toxicity attributed to drug B, both, or neither. Modeling each drug's toxicity probability separately does not determine the probability of experiencing both toxicities. That missing joint probability matters whenever the dose-limiting event is defined as toxicity from either drug: the overall risk depends on how often the two events overlap.

A copula provides a way to connect marginal probabilities to a joint probability. In this tutorial, the two binary outcomes are \(Y_A\) and \(Y_B\), where \(Y_A=1\) denotes the specified toxicity outcome associated with drug A and \(Y_B=1\) denotes the specified outcome associated with drug B. The outcomes are assessed over the same defined observation window. The combination DLT is the event \(Y_A=1\) or \(Y_B=1\). If the clinical endpoint is defined differently, the model's event definitions and resulting risk calculation must be changed accordingly.

Specify the Margins and the Copula

The marginal models describe dose effects; the copula describes dependence between the two toxicity outcomes after conditioning on the dose pair and model parameters. A simple starting specification uses logistic dose-response margins. Define \(x_A=\log(d_A/d_{A,\mathrm{ref}})\) and \(x_B=\log(d_B/d_{B,\mathrm{ref}})\), where the reference doses are positive and are chosen for convenient interpretation. One parsimonious model is

Here \(a_A\) and \(a_B\) are reference-dose log-odds, and \(b_A\) and \(b_B\) are dose slopes. Positive slopes encode increasing marginal toxicity risk with the corresponding drug dose. This basic specification assumes that the A-toxicity margin depends on drug A's dose and the B-toxicity margin on drug B's dose. If scientific knowledge or data support cross-drug effects on an individual toxicity outcome, terms for the other drug's dose can be added to that margin. Dependence in the copula is not a substitute for such cross-dose effects on the margins.

One commonly used positive-dependence choice is the Clayton copula:

The independence model is the limiting case \(\theta=0\), for which \(C(u,v;0)=uv\). Larger positive \(\theta\) gives stronger positive dependence for this copula. The Clayton family is not a universal choice: it represents positive association and has a particular dependence structure. If negative association is scientifically plausible or the observed joint pattern is poorly represented, another copula family or a different joint model may be needed.

From Joint Risk to Combination Risk

For paired outcomes, the copula determines the probability of both toxicities. The other three cell probabilities follow from the margins, and their non-negativity is necessary for a valid model:

Adding the first three cells gives the risk of at least one toxicity. This also follows directly from inclusion–exclusion: \(P(Y_A=1\text{ or }Y_B=1)=P(Y_A=1)+P(Y_B=1)-P(Y_A=1,Y_B=1)\). The overlap is subtracted once because it is counted in both marginal probabilities.

For any valid joint distribution, the probability of both events must satisfy \(\max(0,p_A+p_B-1)\leq P(Y_A=1,Y_B=1)\leq\min(p_A,p_B)\). These are the Fréchet bounds. They imply that the overall risk cannot be calculated from the margins alone: different permissible overlaps yield different probabilities of at least one event.

Bayesian Likelihood and Dependence Learning

If both binary outcomes are recorded for every patient, each patient contributes one of the four cells above to the likelihood. At patient \(i\)'s dose pair, let the model give cell probabilities \(q_{11,i},q_{10,i},q_{01,i},q_{00,i}\). Conditional independence across patients gives a likelihood equal to the product of the relevant cell probabilities. Equivalently, for counts \(n_{jk}\) at a common dose pair, the likelihood is proportional to \(q_{11}^{n_{11}}q_{10}^{n_{10}}q_{01}^{n_{01}}q_{00}^{n_{00}}\); the multinomial coefficient does not depend on the parameters.

A Bayesian specification assigns priors to the marginal-model parameters and to the dependence parameter, then updates them using this likelihood. For example, priors for \(a_A,a_B\) and positive slopes \(b_A,b_B\) can encode plausible reference risks and dose directions; a prior supported on \(\theta>0\) can express uncertainty about positive dependence. An independence point mass can be included if independence is a scientifically relevant possibility. Prior choices should be checked against implied risks and outcome patterns at the dose combinations under study, rather than assessed only by inspecting the parameters individually.

When only the composite endpoint “any toxicity” is recorded, a patient's likelihood contribution is \(p_{\mathrm{any}}\) if a DLT occurs and \(1-p_{\mathrm{any}}\) otherwise. This can inform \(\theta\) through its effect on \(p_{\mathrm{any}}\), but it supplies less direct information about which outcome occurred and how often toxicities co-occurred. Dependence can then be weakly identified, particularly if marginal parameters are also uncertain. The copula formulation does not create information absent from the observations.

What to Check, and What the Model Cannot Do

Before using the model to describe combination risks, check that the dose-response margins yield plausible probabilities throughout the intended dose region, that the assumed dose directions are supported, and that every calculated four-cell distribution is valid. Prior predictive simulation is useful: draw parameters from the priors, calculate margins and joint risks at relevant dose pairs, and inspect whether the implied outcome patterns are credible. Posterior predictive checks can compare replicated and observed counts of each paired outcome, especially co-occurrence. If only a composite endpoint is recorded, check its dose-specific frequency while recognizing that it cannot diagnose the observed overlap directly.

A further diagnostic is sensitivity to the dependence specification and prior. If a risk estimate changes substantially under reasonable priors or alternative copulas, the data may not pin down dependence well enough for that estimate to be robust. Sparse observations of the both-toxicities cell, limited variation in dose pairs, or highly uncertain marginal risks can all contribute. A model that returns precise-looking posterior risks is not necessarily strongly informed by the data.

The main alternatives serve different purposes. Independence fixes the joint probability at \(p_Ap_B\), simplifying calculations but imposing zero association. A different copula can represent a different dependence pattern, including negative association if scientifically defensible. A saturated four-cell model at each dose pair can describe cell frequencies flexibly, but it may be difficult to estimate with sparse dose-combination data and does not automatically provide a coherent dose-response structure. The copula approach separates marginal dose-response assumptions from a dependence assumption, which aids interpretation but makes adequacy of both components important.

The model is not a guarantee that toxicity events have distinct biological causes, nor does association imply that one event causes the other. Its usefulness depends on meaningful outcome definitions and a suitable observation window. Also, an overall risk for “either toxicity” can hide differences in severity, attribution, and clinical consequences between the events; those distinctions require endpoint definitions and analyses that represent them.