Copula Models For Joint Dose Outcomes

Why a Copula Is Needed

At a dose \(d\), an efficacy model can estimate the probability of response and a toxicity model can estimate the probability of an adverse event. Those two marginal probabilities do not, by themselves, determine how often both outcomes occur in the same patient. A joint model supplies that missing dependence information, which matters when a dose decision distinguishes patients who benefit without toxicity from those who experience both.

The earlier material on dependence between efficacy and toxicity introduces why the association matters. Here, the focus is a particular way to build a joint distribution from the two marginal dose-response models. The copula leaves the marginal probabilities intact and specifies how they fit together.

Gaussian Copula Construction

A Gaussian copula is indexed by a dependence parameter \(\rho\), with \(-1\leq\rho\leq1\). For \(0<u,v<1\), define

where \(\Phi\) is the standard normal cumulative distribution function and \(\Phi_2(a,b;\rho)\) is the bivariate standard normal probability \(P(Z_E\leq a,Z_T\leq b)\) when the two standard normal variables have correlation \(\rho\). For interior values \(-1<\rho<1\), use this copula with the marginal models to set

Equivalently, the event probabilities can be represented by thresholding a pair of standard normal variables with correlation \(\rho\). This representation defines the Gaussian copula construction; it does not assert that the trial outcomes are continuous or that an unobserved biological mechanism has been established.

Once \(q_{11}(d)\) is calculated, the other cells are fixed by subtraction:

The first subscript denotes efficacy and the second toxicity. Thus \(q_{10}\) is efficacy without toxicity, while \(q_{01}\) is toxicity without efficacy. These cells sum to one and reproduce both marginal probabilities. A basic coherence check is that each cell is nonnegative.

What the Dependence Parameter Means

The parameter \(\rho\) is the correlation in the Gaussian copula, not generally the correlation between the observed binary outcomes. With binary margins, the observed correlation depends on \(\rho\) and on both marginal probabilities. A positive \(\rho\) tends to increase the probability that efficacy and toxicity occur together relative to independence; it does not change either marginal probability.

For example, when both margins equal \(0.50\), the Gaussian copula has a convenient exact expression. If \(\rho=0.50\), then

The remaining cells are \(q_{10}=0.50-1/3=1/6\), \(q_{01}=1/6\), and \(q_{00}=1-0.50-0.50+1/3=1/3\). The observed binary correlation is

Thus a copula parameter of \(0.50\) produces an observed binary correlation of \(1/3\) in this example. A different pair of marginal probabilities would generally give a different binary correlation for the same \(\rho\).

Derivation: Independence and Dependence Limits

At \(\rho=0\), the bivariate standard normal distribution factors into independent standard normal margins. Therefore \(\Phi_2(a,b;0)=\Phi(a)\Phi(b)\), and substituting \(a=\Phi^{-1}(p_E)\) and \(b=\Phi^{-1}(p_T)\) gives \(q_{11}=p_Ep_T\). This establishes the independence case directly.

The joint probability must also obey bounds imposed by the binary margins. From the requirement that all four cells are nonnegative, \(q_{11}\geq0\), \(p_E-q_{11}\geq0\), \(p_T-q_{11}\geq0\), and \(1-p_E-p_T+q_{11}\geq0\). Rearranging gives the Fréchet bounds:

For the Gaussian copula, as \(\rho\) approaches \(1\), the two copula coordinates move together and \(q_{11}\) approaches the upper bound \(\min(p_E,p_T)\). As \(\rho\) approaches \(-1\), they move oppositely and \(q_{11}\) approaches the lower bound \(\max\{0,p_E+p_T-1\}\). These are limiting cases; they do not mean that every application should assume perfect dependence.

Assumptions, Diagnostics, and Alternatives

The construction depends on more than arithmetic. First, efficacy and toxicity must be defined for the same patient and the same dose, with clear outcome windows and coding. Second, the marginal dose-response models must be appropriate: their probabilities should be calibrated to the intended population and estimand. Third, the selected copula must be a reasonable representation of dependence. If one common \(\rho\) is used across doses, that also assumes the copula dependence structure is constant across dose, even though the observed binary association can change as the margins change.

For discrete outcomes, the margins do not uniquely determine an unrestricted copula at all points. Selecting a Gaussian copula is therefore a modeling choice, not a consequence of fitting the two marginal models. At the observed binary thresholds, the chosen copula supplies a joint probability; other copula families can supply different joint probabilities while preserving the same margins.

Useful checks include verifying that all four computed cells are in \([0,1]\), that they sum to one, and that their row and column sums reproduce the fitted margins. Compare predicted joint cell frequencies with observed counts by dose where data permit. Large systematic discrepancies, sparse cells, or estimates of \(\rho\) near its limits can indicate weak information or model strain. If fitted values violate the Fréchet bounds or produce negative cells, the implementation or parameterization is invalid.

The main alternatives answer different needs. Independence fixes the joint probability at \(p_Ep_T\), which is simple but imposes no residual association. A saturated four-cell model at each dose can represent the observed joint distribution flexibly, but requires more data and may not borrow strength across doses. Other copula families change the dependence pattern, including how strongly the model links rare or extreme outcomes. The Gaussian copula is a compact choice, not a universally correct one.

Common Pitfall and Practical Interpretation

A common mistake is to describe \(\rho\) as the efficacy-toxicity correlation observed among patients. That is wrong because \(\rho\) indexes the Gaussian copula, while the correlation of the two binary indicators depends on their margins as well. The correction is to report \(\rho\) as the copula dependence parameter and, when relevant, calculate the implied binary association or joint cell probabilities at the dose of interest.

Another limitation is that the joint probabilities are conditional on the chosen marginal and dependence models. A plausible set of cells is not evidence that the copula captures the clinical mechanism. In dose-finding work, communicate the marginal probabilities, the dependence assumption, and the resulting joint probabilities together. This lets readers see both what the marginal models imply and how the dependence choice changes the balance of benefit and harm.