Why Interpretation Errors Matter
A joint dose model can produce several quantities that sound similar but answer different questions. The probability of efficacy at a dose is not the probability of efficacy without unacceptable toxicity. An observed association between toxicity and efficacy does not establish that toxicity causes efficacy. And a high posterior probability that a dose meets prespecified criteria is evidence under a model, not certainty that the dose is truly acceptable.
These distinctions matter because a dose-finding decision combines multiple outcomes and uncertainty. The interpretation must preserve the target population, outcome definitions, time window, and decision thresholds used by the model. The introductory material on joint dose models establishes the setting; here the emphasis is on diagnosing errors in the quantities and claims drawn from it.
Mistake 1: Treating Marginal Probabilities as Joint Probabilities
A marginal probability describes one outcome while averaging over the other. In contrast, a joint probability describes a combination of outcomes. Knowing that efficacy occurs in 60% of patients and toxicity in 30% does not, by itself, tell us how often efficacy occurs without toxicity. That answer depends on how often efficacy and toxicity occur in the same patients.
For binary outcomes, the joint probability of efficacy without toxicity can be written using the overlap between efficacy and toxicity:
The expression subtracts patients who have both efficacy and toxicity from all patients with efficacy. It is also possible to write the quantity as \(P(E=1\mid d)P(T=0\mid E=1,d)\), if that conditional probability is available. Multiplying the two marginal probabilities, \(P(E=1\mid d)P(T=0\mid d)\), is justified only under independence of efficacy and toxicity at that dose. Independence is a substantive assumption, not a default arithmetic rule.
Worked Example: Recovering a Joint Probability
Suppose a hypothetical dose group has 100 patients. Efficacy is observed in 60 patients, toxicity in 30, and both outcomes in 20. Calculate the proportion with efficacy and no toxicity, and compare it with an independence calculation.
The observed marginal probabilities are \(P(E=1\mid d)=60/100=0.60\), \(P(T=1\mid d)=30/100=0.30\), and the joint overlap is \(P(E=1,T=1\mid d)=20/100=0.20\). Therefore,
The same result follows from the count: 60 patients had efficacy, and 20 of them also had toxicity, leaving 40 with efficacy and no toxicity. If one incorrectly assumes independence, the calculation is \(0.60(1-0.30)=0.60(0.70)=0.42\). That is 2 percentage points higher than the observed joint proportion. The independence result is not supported by these data: the observed probability of both outcomes is 0.20, whereas independence would imply \(0.60(0.30)=0.18\).
There is a useful check even when the overlap is not directly known. For any two events with marginal probabilities \(p_E\) and \(p_T\), their intersection must satisfy
The upper bound holds because an intersection cannot be larger than either event. For the lower bound, the probability of the union cannot exceed 1, while \(P(E=1\text{ or }T=1)=p_E+p_T-P(E=1,T=1)\). Rearranging gives the lower bound \(p_E+p_T-1\), along with the fact that probabilities cannot be negative. These bounds imply that marginals alone may leave a wide range of possible joint benefit-risk probabilities.
Worked Example: Checking What Marginals Can Establish
At a hypothetical dose, suppose \(P(E=1\mid d)=0.55\) and \(P(T=1\mid d)=0.25\). Without joint information, find the possible range for efficacy and no toxicity.
First bound the overlap:
Since \(P(E=1,T=0\mid d)=0.55-P(E=1,T=1\mid d)\), the smallest possible value is \(0.55-0.25=0.30\), and the largest is \(0.55-0=0.55\). Thus the marginals are compatible with any efficacy-without-toxicity probability from 0.30 through 0.55. Multiplying the marginals would give \(0.55(0.75)=0.4125\), but that is only one value under an independence assumption; it is not identified by the two marginal estimates.
Mistake 2: Reading Association as Causation
A joint model can describe whether efficacy and toxicity tend to occur together, separately at each dose or after adjustment for covariates. Such a description is an association. It does not establish that toxicity causes efficacy, that efficacy causes toxicity, or that changing one outcome would change the other. A shared biological pathway, patient characteristics, measurement processes, or selection into dose groups can produce an association.
Causal claims about dose require a different justification. In a randomized dose comparison, random allocation can support causal contrasts between assigned doses, subject to consistency, adequate allocation to the doses being compared, and appropriate handling of missing outcomes and intercurrent events. In an adaptive trial, allocation can depend on past trial information and may also depend on patient covariates. The relevant causal comparison must then account for the allocation process and the target population; an unadjusted contrast between patients who happened to receive different doses need not be causal.
Worked Example: A Dose Association Created by Patient Mix
Consider a hypothetical adaptive trial in which clinicians tend to assign a lower dose to patients at higher baseline toxicity risk. Each dose group has 80 patients. Among low-dose patients, 60 are high-risk and 20 are low-risk; 30 high-risk and 2 low-risk patients experience toxicity. Among high-dose patients, 20 are high-risk and 60 are low-risk; 10 high-risk and 6 low-risk patients experience toxicity. Compare the crude and risk-stratified toxicity proportions.
The crude low-dose toxicity proportion is \((30+2)/80=32/80=0.40\). The crude high-dose proportion is \((10+6)/80=16/80=0.20\). The crude comparison suggests 20 percentage points less toxicity at the high dose.
Within the high-risk stratum, the low-dose rate is \(30/60=0.50\) and the high-dose rate is \(10/20=0.50\). Within the low-risk stratum, the corresponding rates are \(2/20=0.10\) and \(6/60=0.10\). In each risk stratum, the observed proportions are equal across doses. The large crude difference arises because the low-dose group contains many more high-risk patients, not because these data show a lower toxicity rate at high dose within either stratum.
This example does not prove that dose has no causal effect; observational stratum-specific comparisons can still be affected by other differences or sampling variation. It shows why the crude dose association alone cannot establish a dose effect. A joint efficacy-toxicity association within a dose group likewise does not establish a causal effect of toxicity on efficacy.
Mistake 3: Treating Posterior Acceptability as Certainty
A Bayesian dose-finding analysis may define acceptability using prespecified criteria, for example an efficacy probability above a target and a toxicity probability below a limit. Let \(\mathcal{A}_d\) denote the event that dose \(d\) meets all such criteria. The posterior probability \(P(\mathcal{A}_d\mid \text{data})\) quantifies uncertainty about whether the modeled dose-specific quantities satisfy those criteria, conditional on the observed data, the model, and its prior distribution. It is not a guarantee about future patients or a probability-free certification that the dose is acceptable.
The posterior probability also differs from the posterior mean efficacy or toxicity probability. A mean is an average over posterior uncertainty; acceptability is the probability of a specified event. A mean alone may conceal substantial uncertainty around a decision threshold. Conversely, a high probability of meeting thresholds still does not show that the thresholds are clinically appropriate or that the model represents the target population well.
Worked Example: Interpreting a Posterior Acceptability Probability
Suppose a hypothetical analysis defines dose \(d\) as acceptable when its efficacy probability is at least 0.50 and its toxicity probability is at most 0.25. Across 1,000 posterior draws, 900 satisfy the efficacy criterion, 850 satisfy the toxicity criterion, and 760 satisfy both. Estimate the posterior probability of acceptability.
The relevant event is the intersection of the two criteria, so the estimated probability is the fraction of draws satisfying both:
The separate criterion fractions, 0.900 and 0.850, cannot simply be multiplied to obtain the joint probability unless the two criterion events are independent under the posterior draws. Here their observed overlap is 0.760. The appropriate interpretation is: under the specified posterior model and decision criteria, the estimated posterior probability that this dose meets both criteria is 76.0%. It is not correct to say the dose is certainly acceptable or that 76% of future patients will benefit without toxicity. The latter is a patient-level predictive statement and is a different quantity.
The estimate is also based on a finite number of posterior draws. Its approximate Monte Carlo standard error is \(\sqrt{0.76(1-0.76)/1000}=0.0135\), or about 1.35 percentage points. This numerical approximation error is distinct from posterior uncertainty about the dose parameters.
Diagnostics, Alternatives, and Limits
When an interpretation seems too simple, return to the event being estimated and the assumptions connecting it to the data. The following checks help expose common errors:
- Write the event explicitly. Distinguish \(P(E=1\mid d)\), \(P(T=1\mid d)\), and \(P(E=1,T=0\mid d)\). Check that the reported quantity matches the clinical decision question.
- Check overlap and dependence. If a joint probability is inferred from marginals, identify the independence or dependence model that justifies the calculation. Compare implied joint probabilities with observed outcome patterns where possible.
- Inspect patient mix and allocation. Compare covariates and allocation probabilities across dose groups. In an adaptive design, changing patient composition can make crude rates misleading. Model fit and posterior predictive checks can reveal discrepancies between the fitted model and observed joint outcomes.
- Inspect the decision event. Confirm that the posterior probability refers to all criteria jointly, the intended dose, and the intended population. Examine sensitivity to plausible priors, model assumptions, and thresholds when these materially affect the conclusion.
Separate marginal models can be useful when the scientific questions concern efficacy and toxicity rates individually, but they do not automatically provide a valid joint probability. A joint model can represent outcome dependence and estimate combinations directly, yet its conclusions depend on adequate specification and data. An informal correlation or cross-tabulation can describe observed co-occurrence, but it is not a substitute for a causal analysis or a posterior decision probability. No one summary answers every question: use the quantity aligned with the estimand and decision, and state its assumptions.
A common pitfall is to report a single dose as “safe and effective” because its marginal efficacy is high, its marginal toxicity is low, or its posterior acceptability exceeds a chosen threshold. The correction is to name the joint event, give the posterior probability or estimate with uncertainty, state the criteria and target population, and avoid implying certainty. A posterior probability can guide a decision under prespecified rules, but the decision rule and the probability are not interchangeable.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- At a dose, efficacy is 0.70, toxicity is 0.20, and both occur with probability 0.15. What is the probability of efficacy without toxicity?
- If only the two marginal probabilities in Question 1 are reported, can the joint probability be determined? State why.
- A dose group has a lower crude toxicity rate, but patients assigned that dose had lower baseline toxicity risk. What causal conclusion can be drawn from the crude comparison alone?
- A posterior analysis gives a 0.92 probability that efficacy exceeds its threshold and a 0.88 probability that toxicity is below its limit. Is the posterior probability of meeting both criteria necessarily \(0.92(0.88)\)?
Solutions. (1) \(0.70-0.15=0.55\). (2) No. The overlap of efficacy and toxicity, or an assumption specifying their dependence, is needed. (3) None about the causal dose effect from the crude comparison alone; patient mix may explain the association, so the allocation and covariates must be considered. (4) No. The product is justified only if the posterior events are independent; the joint fraction or joint posterior distribution is needed.