Therapeutic Window Objectives from Efficacy and Toxicity Thresholds

From Clinical Thresholds to a Dose Objective

A dose can produce substantial efficacy and still be unattractive if many patients experience clinically important toxicity. Conversely, a well-tolerated dose may provide too little benefit. When clinicians can specify a minimum acceptable efficacy and a maximum acceptable toxicity, these thresholds define a clinically interpretable region of outcomes: the therapeutic window. A dose-finding objective can then be the fraction of patients whose outcomes fall inside that region.

The relevant quantity is a joint probability. A patient counts as inside the window only if the efficacy condition and the toxicity condition are both met. This is not generally the same as multiplying the efficacy rate by the nontoxicity rate: the two outcomes may be dependent.

The inequalities reflect the example convention that larger efficacy values are better and larger toxicity values are worse. If efficacy is measured by a time or burden for which smaller values are better, the efficacy inequality must be reversed. For a binary toxicity event, \(Y_T(d)\) can be coded as 1 if the event occurs and 0 otherwise; a threshold \(c_T=0\) then means that patients with the event do not meet the toxicity condition. Thresholds, endpoint definitions, assessment times, and boundary conventions should be stated in clinical terms before comparing doses.

The target population matters. If patients differ in covariates such as disease severity or organ function, the overall fraction averages over the covariate distribution in the population to which the decision will apply. A fraction calculated in a trial with a different mix of patients need not represent the target population without standardization.

Why the Joint Probability Matters

Write \(A(d)=\{Y_E(d)\geq c_E\}\) for meeting the efficacy threshold and \(B(d)=\{Y_T(d)\leq c_T\}\) for meeting the toxicity threshold. Then \(p_W(d)=P\{A(d)\cap B(d)\}\). It is generally not equal to \(P\{A(d)\}P\{B(d)\}\); that product is correct only if the two threshold events are independent at that dose in the population being considered.

A useful check is the pair of probability bounds \( \max\{0,P(A)+P(B)-1\}\leq P(A\cap B)\leq\min\{P(A),P(B)\} \). The upper bound follows because an intersection cannot be larger than either event. For the lower bound, \(P(A\cup B)\leq 1\) and \(P(A\cup B)=P(A)+P(B)-P(A\cap B)\), so \(P(A\cap B)\geq P(A)+P(B)-1\); probabilities are also at least zero. If reported marginal rates and a proposed joint rate violate these bounds, there is an arithmetic or modeling inconsistency.

Constructing and Comparing Dose Objectives

Once the clinical event is defined, the objective can be stated directly: choose a dose with a large \(p_W(d)\). For a finite set of candidate doses, this means estimating the window fraction at each dose and comparing the estimates. In a dose-response model, the same quantity can be calculated from the joint predictive distribution at each dose. The prediction must represent the efficacy and toxicity outcomes together, not just two separate curves without a dependence assumption.

For a target population with covariates \(Z\), the population fraction can be written as an average over those covariates. Conditional on \(Z=z\), let \(p_W(d\mid z)\) be the probability of meeting both thresholds. Then \( p_W(d)=\int p_W(d\mid z)\,dF_Z(z) \), where \(F_Z\) is the target population's covariate distribution. For discrete strata, the integral becomes a weighted sum. This is standardization: each stratum's window probability is weighted by its proportion in the target population.

A window objective differs from several nearby decision summaries. An efficacy target-attainment objective asks how often an efficacy target is met, without necessarily requiring that toxicity also be acceptable. A toxicity rate alone ignores whether patients benefit. A mean efficacy outcome also does not reveal how many patients meet both clinical thresholds. The window fraction answers a distinct question: how frequently are both requirements satisfied in the same patient?

Another possible decision rule is to impose a maximum acceptable toxicity rate and then maximize efficacy among doses meeting that constraint. This may be suitable when toxicity is a hard constraint, but it is not identical to maximizing the joint window fraction. Likewise, choosing a dose that maximizes the window fraction is one explicit objective, not a universal clinical rule. A decision maker may instead require the fraction to exceed a prespecified minimum, or compare it with other clinical considerations. The thresholds and the way the objective will guide a decision should therefore be agreed before examining dose rankings.

Estimation, Assumptions, and Diagnostics

In an observed sample of \(n\) patients treated at dose \(d\), a direct estimator is the number meeting both thresholds divided by \(n\). With unequal sampling probabilities, trial weights, or a target population different from the enrolled sample, an appropriate weighted or standardized estimator is needed. In model-based dose finding, the corresponding estimate is obtained from the model's joint outcome distribution and then averaged over the target population. In either case, uncertainty in the estimated fraction should be reported when it is relevant to the dose decision; a point estimate alone can make close dose rankings appear more decisive than they are.

A sound interpretation relies on several conditions. The efficacy and toxicity thresholds must represent clinically meaningful definitions, including assessment times and handling of boundary values. The model or data must support prediction of both outcomes at the candidate doses. The target population and its covariate distribution must be appropriate to the decision. Finally, the joint outcome representation must preserve, or explicitly model, dependence between efficacy and toxicity. Marginal predictions cannot recover an unknown joint distribution by themselves.

Diagnostics should examine more than the final ranking. Check that predicted marginal rates are plausible, that joint probabilities are between zero and one and obey the probability bounds, and that the four categories—both conditions met, efficacy only, toxicity condition only, and neither—sum to one. Compare observed and predicted joint category frequencies where data are available, including across important patient strata and doses. If a model predicts sensible marginal response and toxicity rates but systematically misses co-occurrence, its window predictions can still be wrong.

A further distinction is between patient variability and uncertainty about the model parameters. The population fraction \(p_W(d)\) describes the proportion of patients expected to meet both conditions under the specified model and population. Uncertainty intervals for its estimate describe uncertainty in that quantity due to limited data and model estimation; they are not themselves the distribution of individual patient outcomes. Both matter, but they answer different questions.

Common Pitfalls and Limitations

• Multiplying marginal rates without justification. The product is the joint fraction only under independence of the two threshold events. Use a joint model or observed joint data when dependence may be present.
• Using thresholds that are not clinically anchored. A mathematically convenient cut point does not define a meaningful therapeutic window unless its clinical interpretation is clear.
• Changing the population implicitly. Comparing unstandardized fractions from populations with different covariate mixes can confound the dose comparison with differences in who was represented.
• Treating the largest estimate as a certain winner. Close estimates may be uncertain, and the ranking may depend on model assumptions or target-population weights.
• Assuming the window objective captures every relevant tradeoff. It treats outcomes on each side of the thresholds as meeting or failing the specified condition. Differences in severity within those categories are not represented by the fraction alone.

The window fraction is valuable because its definition is transparent and its interpretation is direct: it is the proportion of the target population predicted to satisfy both clinical conditions. Its limitations follow from the same threshold structure. Patients just below a threshold and patients far below it are both counted as outside, while patients meeting both conditions are counted alike. Such distinctions may call for a different objective, but they should not be smuggled into the window fraction by changing its interpretation.