Why Dependence Changes Cumulative Success
A development program may need to meet efficacy, safety, and operational requirements for its overall success. Multiplying three probabilities is appropriate only under a particular assumption: the corresponding events are independent. That assumption can be consequential. For example, an underlying biological feature may make both efficacy and an adverse effect more likely, while a site or enrollment problem may affect several operational milestones at once. A cumulative probability calculated as if risks were unrelated can then misstate the chance that all requirements are met.
The earlier material on conditional versus cumulative probability distinguishes the probability of success at one stage from the probability of success across a sequence. Here, the focus is how dependence among different success requirements changes that cumulative calculation. Each event must be defined clearly, including what counts as success, before its probability can be combined with others.
“Safety success” here means meeting a prespecified safety criterion, not the absence of all adverse events. Likewise, an operational success event needs a defined criterion, such as completing enrollment within a stated time. The exact criteria affect the probabilities and their dependence; they are not interchangeable labels.
Calculating Joint Success with Conditional Probabilities
The probability chain rule gives a general way to express joint success. It does not require independence. Provided the conditioning events have positive probability, the joint probability can be factored into a sequence of conditional probabilities.
This is an identity, not an approximation. To see why, the definition of conditional probability gives \(P(E\cap S)=P(E)P(S\mid E)\). Applying the same definition again gives \(P(E\cap S\cap O)=P(E\cap S)P(O\mid E\cap S)\). Substituting the first equality into the second yields the chain rule above. Other orderings are also valid, such as \(P(S)P(O\mid S)P(E\mid S\cap O)\), if the conditional probabilities in that ordering are available.
By contrast, the product of marginal probabilities follows only if the relevant independence conditions hold. For the ordering above, it is enough that \(P(S\mid E)=P(S)\) and \(P(O\mid E\cap S)=P(O)\), with the conditional probabilities defined. Then the chain rule reduces to the product of marginals. Without such conditions, using the product is an assumption rather than a consequence of knowing the marginal probabilities.
Worked Example: Sequential Conditional Probabilities
Suppose a hypothetical development team assigns \(P(E)=0.60\), \(P(S\mid E)=0.70\), and \(P(O\mid E\cap S)=0.60\). These are elicited conditional probabilities, not marginal estimates for safety and operations. Calculate the joint probability of all three success events.
Using the chain rule,
Thus, under these hypothetical inputs, the probability of meeting all three requirements is 25.2%. This result does not require independence because the inputs are conditional in the stated order. Multiplying these three numbers as if they were marginal probabilities would happen to give the same arithmetic, but would mislabel the quantities and invite an invalid interpretation.
For comparison, suppose the team separately specifies the marginal probabilities \(P(S)=0.80\) and \(P(O)=0.75\). The independence calculation would be \(0.60\times0.80\times0.75=0.360\), or 36.0%. That is not the joint probability implied by the conditional inputs. The two calculations use different assumptions and different probability inputs; they should not be treated as competing estimates of one fully specified model.
Joint Distributions and Shared Conditions
A joint distribution describes probabilities for combinations of event outcomes. For three binary success events, there are eight combinations: all three succeed, exactly two succeed in one of three ways, exactly one succeeds in one of three ways, or none succeeds. The probabilities assigned to these eight mutually exclusive and exhaustive outcomes must each be between zero and one and must sum to one. Marginal probabilities are obtained by adding the probabilities of all outcomes in which the event of interest occurs.
A joint distribution can represent dependence directly, while conditional probabilities summarize selected parts of that distribution. Either representation is useful; the choice depends on what evidence or judgments are available. A conditional-probability tree is often convenient when the program is described sequentially. A joint table can make the assumed combinations easier to inspect, especially when risk factors influence more than one event.
Worked Example: Dependence Through a Shared Operating Condition
Consider a hypothetical program with a favorable operating environment \(C\), such as adequate site capacity and timely supply, with \(P(C)=0.70\). Suppose that conditional on \(C\), the three success events are independent, with \(P(E\mid C)=0.70\), \(P(S\mid C)=0.90\), and \(P(O\mid C)=0.90\). In an unfavorable environment, suppose they are conditionally independent with probabilities \(P(E\mid C^c)=0.30\), \(P(S\mid C^c)=0.60\), and \(P(O\mid C^c)=0.50\). Calculate joint success and compare it with the product of marginal probabilities.
First calculate success in each environment. Conditional independence within each environment gives
and
The two environments partition all possibilities, so the law of total probability gives
The marginal probabilities, averaging over the two environments, are
and
Multiplying those marginal probabilities gives \(0.580\times0.810\times0.780=0.366\) to three significant figures. It is below the joint probability, 0.424. The difference arises because programs in the favorable environment tend to do well on all three events, while those in the unfavorable environment tend to do worse on all three. The events are dependent overall even though they were assumed conditionally independent within each environment.
The shared condition in this example is a representation of one possible source of dependence, not proof that a particular operational factor causes clinical efficacy or safety. A common factor may create statistical association, and a proposed causal explanation requires evidence beyond the observed association. Dependence can also arise through other mechanisms, including shared patient characteristics, shared sites, or correlated judgments in probability elicitation.
Checking Inputs and Diagnosing Dependence Assumptions
Before using conditional probabilities or a joint distribution, check that the inputs refer to the same program, time horizon, and success definitions. A conditional probability such as \(P(S\mid E)\) concerns safety success among cases where efficacy success occurs. It is not the probability of safety success in the entire program. Mixing probabilities conditioned on different populations or decision points can produce a number that is mathematically calculable but substantively incoherent.
Useful checks include the following:
- Check bounds and totals. Every probability must lie between zero and one. Joint-outcome probabilities must sum to one, and each marginal must agree with the sum of joint probabilities that include its event.
- Check conditional denominators. \(P(A\mid B)\) is defined only when \(P(B)>0\). In practice, also ask whether the conditioning event is sufficiently well represented for the probability to be meaningful.
- Compare assumptions with mechanisms. Ask whether the events share patients, sites, operational resources, biological drivers, or evidence sources. A plausible shared driver is a reason to examine dependence, not by itself a quantified dependence estimate.
- Vary uncertain dependence assumptions. Recalculate joint success under plausible conditional probabilities or alternative joint distributions. If the decision-relevant conclusion changes substantially, the dependence assumption deserves attention.
A discrepancy between the product of marginals and a conditional or joint calculation is not automatically evidence that one calculation is wrong. It may reveal that the models encode different assumptions. Conversely, matching values do not demonstrate independence: different dependence patterns can sometimes yield the same probability of the all-success outcome. A single cumulative probability does not describe the full joint distribution.
Comparing Approaches and Their Limits
| Approach | What it represents | Main caution |
|---|---|---|
| Product of marginal probabilities | Joint success under mutual independence | Can misstate cumulative success if events are dependent |
| Conditional probability chain | Joint success using probabilities conditioned on preceding events | Requires conditional inputs that match the stated order and context |
| Joint distribution | Probabilities for every combination of success and failure | Many probabilities may be difficult to estimate or elicit |
| Shared-condition model | Dependence induced by a modeled common environment or factor | Results depend on the chosen factor and conditional assumptions |
The independence product is simple and transparent when independence is defensible, but it is not a default truth. A conditional chain avoids that assumption and is often easier to elicit when experts can reason about a sequence of conditions. A joint distribution is more complete, but can demand more information than is available. A shared-condition model can explain how dependence might arise, yet its additional structure should be justified and tested rather than treated as known.
All these approaches have limitations. Probabilities may be uncertain, based on sparse evidence, or influenced by how experts interpret success criteria. A conditional probability may change if the conditioning event is redefined. A model with a shared factor can omit other dependencies or imply conditional independence that is not credible. When the evidence cannot identify the dependence structure, report the assumption plainly and show how conclusions vary across plausible alternatives. Do not present a precise cumulative probability as if it were independent of those modeling choices.
A Practical Calculation Sequence
State the efficacy, safety, and operational criteria and the time horizon for each probability.
Use conditional probabilities, a joint distribution, or an explicit shared-condition model according to the available evidence.
Apply the chain rule or sum the joint probabilities for combinations in which all required events succeed.
Compare with the independence product when useful, then state which assumptions account for any difference and how uncertain they are.
Check Your Understanding
Answer each question, then compare with the solutions below.
- If \(P(E)=0.50\), \(P(S\mid E)=0.80\), and \(P(O\mid E\cap S)=0.75\), what is \(P(E\cap S\cap O)\)?
- Suppose \(P(E)=0.60\), \(P(S)=0.70\), and \(P(O)=0.80\). What additional assumption is needed to calculate joint success as the product of these three numbers?
- In the shared-condition example, why did conditional independence within each environment not imply unconditional independence?
- A proposed joint distribution assigns probabilities \(0.20, 0.25, 0.30,\) and \(0.20\) to four mutually exclusive and exhaustive outcomes. What must be checked before using it?
Solutions. (1) By the chain rule, \(0.50\times0.80\times0.75=0.300\), or 30.0%. (2) Mutual independence of the three events is sufficient; without it, the product is not generally the joint probability. (3) The operating environment changes the probabilities of success for multiple events. Averaging over that common condition links the events, even when they are independent conditional on its value. (4) The probabilities must be nonnegative, no greater than one, and sum to one. Here they sum to \(0.95\), so the distribution is incomplete or invalid as stated; it cannot describe exhaustive outcomes until reconciled.