Stage-Gate Policies as Sequential Decisions

From a Gate Decision to a Policy

A development program rarely involves one decision made once. Evidence arrives at successive gates, and the appropriate action may depend on what that evidence shows. A team might continue after strong evidence, modify the program after an ambiguous result, and stop after evidence that makes further investment unattractive. The object being evaluated is then not a single action, but a policy: a rule that specifies what to do at each possible evidence state.

A policy is useful only if its evidence states and actions are stated clearly. “Continue if the results look good” is not operational: different decision makers may classify the same results differently. A usable policy defines the evidence categories or quantities that trigger each action, identifies what each action entails, and assigns a value to the consequences that follow.

Definition: A stage-gate policy is a mapping from each decision-relevant evidence state at each gate to an available action, such as continue, modify, or stop. Its expected value is the probability-weighted value of the consequences generated by following that mapping, including the costs and outcomes at later gates.

Write $e_t$ for the evidence state observed at gate $t$, and $a_t$ for the action selected there. A policy $\pi$ specifies $a_t=\pi_t(e_t)$. If the relevant history is not summarized by $e_t$, the policy must instead map the full decision-relevant history to an action. For example, a “high biomarker response” state at a later gate may need to be interpreted alongside the dose or regimen that produced it.

Expected-value evaluation requires a defined perspective and a consistent accounting horizon. The calculations below assume that decision makers are risk-neutral with respect to the monetary values shown, that probabilities represent the relevant conditional probabilities, and that all stated costs and benefits are included on a common value basis. Costs already incurred before a gate are sunk for that gate’s continue-or-stop choice, though they may be included when reporting the program’s total value from an earlier point. The actions considered must also be feasible, and their consequences must be compared on the same basis.

Evaluating a Policy at One Gate

At a gate, each evidence state can lead to a different conditional expected value for each action. The probability of reaching a state determines how much that state contributes to the policy’s overall expected value. If $E$ denotes the set of possible states at the gate, the value of a policy at that gate is

EV(\pi)=\sum_{e\in E}P(e)\,V\bigl(\pi(e)\mid e\bigr),

where $P(e)$ is the probability of state $e$, and $V(a\mid e)$ is the expected value of taking action $a$ conditional on that state. A stop action may have value zero at the gate if no further costs or benefits are counted from that point. If stopping has a salvage value or other consequence, that value should be included rather than assumed to be zero.

Worked Example: Continue, Modify, or Stop After Phase II

Hypothetical data. Before a Phase II evidence review, the probability of a favorable state is 0.40 and of an unfavorable state is 0.60. All values are in millions of dollars, measured from this gate onward. Continuing costs $30 million and has a 0.70 probability of technical success after favorable evidence and 0.25 after unfavorable evidence. Success produces $120 million of value; failure produces no additional value. Modifying costs $40 million and has corresponding success probabilities of 0.55 and 0.45. Stopping has value $0 million.

The assumptions for this calculation are that the stated success probabilities are conditional on the evidence state and chosen action, the stated costs apply whether or not the program succeeds, and the success payoff is the same under both actions. For any action, its conditional expected value is success probability times $120 million, less that action’s cost.

After favorable evidence, continuing is worth $0.70(120)-30=84-30=\$54$ million. Modifying is worth $0.55(120)-40=66-40=\$26$ million. Stopping is worth $0 million. The policy therefore continues after favorable evidence.

After unfavorable evidence, continuing is worth $0.25(120)-30=30-30=\$0$ million. Modifying is worth $0.45(120)-40=54-40=\$14$ million, and stopping is worth $0 million. The policy modifies after unfavorable evidence.

The resulting policy is $\pi(\text{favorable})=\text{continue}$ and $\pi(\text{unfavorable})=\text{modify}$. Its expected value is

EV(\pi)=0.40(54)+0.60(14)=21.6+8.4=\$30.0\text{ million}.

For comparison, always continuing has value $0.40(54)+0.60(0)=\$21.6$ million, and always modifying has value $0.40(26)+0.60(14)=\$18.8$ million. Stopping in both states is worth $0 million. These comparisons are valid because the same state probabilities and valuation basis are used for each policy.

Sequential Gates and Conditional Values

In a multi-stage program, an action at one gate may lead to another evidence state, where a further action is chosen. The value at the earlier gate must account for that later decision, rather than treating the next-stage outcome as fixed. In a finite sequence of gates, this can be evaluated backward: first determine the best available action at the last gate, then calculate the conditional value of the actions that lead to it at the preceding gate.

Let $V_t(e)$ be the best expected value available at gate $t$ after observing state $e$. For an action $a$, let $u_t(e,a)$ be its immediate expected payoff, including any cost or payoff incurred at that gate. If the action can lead to a later state $e'$, its value is

Q_t(e,a)=u_t(e,a)+\sum_{e'}P(e'\mid e,a)V_{t+1}(e').

The value at the current gate is the greatest value over the actions available in that state: $V_t(e)=\max_a Q_t(e,a)$. For a stop action, the value can be represented directly in $u_t(e,a)$, with no later-state term. The conditional probabilities in this calculation must match the action and evidence history; they need not be independent across stages.

Worked Example: A Two-Gate Policy With a Final Launch Decision

Hypothetical data. A program first incurs an $8 million evidence-study cost. The study has a 0.60 probability of favorable evidence and a 0.40 probability of weak evidence. At this gate, the team can continue to a trial for $12 million, modify the program for $18 million, or stop. Conditional on favorable evidence, the trial evidence is positive with probability 0.80 under continue and 0.70 under modify. Conditional on weak evidence, those probabilities are 0.35 and 0.60, respectively.

At the second gate, positive trial evidence gives a launch action an expected net value of $90 million; negative evidence gives launch an expected net value of $-\$20$ million. Stopping at this final gate is worth $0 million in either state. These launch values already include the consequences assigned to the launch decision, so no additional launch cost or payoff is added here. The trial-stage costs are paid before the second-gate result and are included in the values below.

At the second gate, the best action after positive evidence is launch, worth $90 million rather than $0 million from stopping. After negative evidence, stopping is better than launch, because $0> -20$. Thus the final rule is launch after positive evidence and stop after negative evidence.

Now calculate the values at the first gate. Under continue, the expected value after favorable evidence is $-12+0.80(90)+0.20(0)=\$60$ million; after weak evidence it is $-12+0.35(90)+0.65(0)=\$19.5$ million. Under modify, the corresponding values are $-18+0.70(90)=\$45$ million after favorable evidence and $-18+0.60(90)=\$36$ million after weak evidence.

After favorable evidence, continue is best because $60>45>0$. After weak evidence, modify is best because $36>19.5>0$. The policy therefore continues after favorable study evidence and modifies after weak study evidence; at the later gate, it launches only after positive trial evidence.

Before the study result is observed, the policy’s expected value, including the study cost, is

EV(\pi)=-8+0.60(60)+0.40(36)=-8+36+14.4=\$42.4\text{ million}.

The conditional trial probabilities were used only on the branches that reach the trial, and the final action was selected separately for each trial-evidence state. This is why multiplying one unconditional probability of success by a single payoff would not represent this policy.

Deriving Action Regions From a Success Probability

A policy can sometimes be written as a threshold rule in a probability or other evidence summary. The threshold is not universal: it depends on the action-specific payoffs and costs. The following derivation illustrates how a modify region can occur between a stop region and a continue region.

Worked Example: When Is Modification the Preferred Action?

Hypothetical data. Let $p$ be the posterior probability of technical success after the current evidence. Continuing yields $120 million if successful and $0 if not, and costs $35 million. Modifying yields $90 million if successful and $0 if not, and costs $15 million. Stopping is worth $0 million. These are incremental values from the current gate, in millions of dollars.

The three conditional expected values are $V_C=120p-35$, $V_M=90p-15$, and $V_S=0$. Modification is at least as good as stopping when

90p-15\geq 0 \quad\Longleftrightarrow\quad p\geq \frac{1}{6}\approx 0.1667.

Modification is at least as good as continuing when

90p-15\geq 120p-35 \quad\Longleftrightarrow\quad 20\geq 30p \quad\Longleftrightarrow\quad p\leq \frac{2}{3}.

Thus, for $1/6\leq p\leq 2/3$, modification is at least as good as both alternatives. Below $1/6$, stopping is better than modifying, and continuing also has negative value there. Above $2/3$, continuing is better than modifying and has positive value. At a boundary, the relevant actions tie, so a decision rule needs a tie-breaking convention if exactly one action must be selected.

For example, at $p=0.40$, continuing is worth $120(0.40)-35=13$, modifying is worth $90(0.40)-15=21$, and stopping is worth $0$, all in millions of dollars. Modification is preferred. The threshold calculation depends on the stated payoff structure; it is not a generic rule that a particular probability always warrants modification.

Why Statewise Optimization Works—and When It Can Fail

For a fixed set of evidence states and a fixed probability model, choosing the best action separately in each state maximizes the policy’s expected value. To see why, consider any alternative policy $\tilde{\pi}$. Its value is a sum over states. In each state, the selected action under $\pi$ has value at least as large as the action selected under $\tilde{\pi}$, so

V\bigl(\pi(e)\mid e\bigr)\geq V\bigl(\tilde{\pi}(e)\mid e\bigr) \quad\text{for every }e.

Multiplying each inequality by $P(e)\geq 0$ and summing gives $EV(\pi)\geq EV(\tilde{\pi})$. This argument assumes the choice at one state does not constrain the action available in another state. If, for example, a budget limit requires allocating a shared resource across branches or programs, the decisions may be linked and state-by-state maximization need not produce a feasible overall policy.

Several checks help identify a policy evaluation that is going wrong:

Check the state definition. Evidence categories should be mutually exclusive and collectively cover the possible outcomes used in the calculation. Their probabilities should sum to 1.
Check conditional probabilities. A transition probability should be conditioned on the evidence and action that define that branch. Reusing an unconditional rate on every branch can misstate the expected value.
Check cost timing. Include costs when they will be incurred under the action being evaluated. Do not make a future investment appear avoidable if it is already committed, or charge a sunk cost to a decision made after it was incurred.
Check consequence consistency. Use the same payoff horizon, valuation basis, and treatment of success, failure, and stopping across the actions being compared.
Check feasibility and omitted constraints. A policy can look valuable in separate branch calculations yet be infeasible if decisions share a budget, personnel, capacity, or ethical constraint.

A policy also depends on the evidence classification being operationally reproducible. If the measured result is close to a boundary, different classifications can change the selected action and therefore the expected value. Another limitation is that a single expected value conceals the distribution of possible outcomes: two policies can have the same expected value but different probabilities of severe loss. If risk tolerance, uncertainty in probabilities, or non-monetary consequences matter, the decision criterion must represent them explicitly rather than treating the expected monetary value as a complete description.

Common Mistake: Evaluating Only the Most Likely Branch

A common error is to choose the action that looks best in the most likely evidence state and then report that state’s conditional value as the value of the whole policy. That omits the other branches. The correction is to evaluate the specified action in every state and weight each state’s conditional value by its probability. In the first worked example, the favorable state alone contributed $0.40(54)=\$21.6$ million; the unfavorable branch contributed another $0.60(14)=\$8.4$ million. The policy’s value was their sum, not $54 million and not $21.6 million.

A related error is to describe a recommendation as “continue” without stating what happens after later evidence arrives. A sequential policy is incomplete until its rule is specified at every gate in the modeled horizon, including the final stop or launch decision. Conversely, a detailed policy should not imply precision beyond the evidence: state definitions, transition probabilities, and action payoffs are assumptions to be checked and updated as development evidence accumulates.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A gate has two evidence states, $E_1$ and $E_2$, with probabilities 0.30 and 0.70. A policy yields conditional values of $10 million and $4 million, respectively. What is its expected value?
In a state, continue is worth $8 million, modify is worth $11 million, and stop is worth $0 million. Which action does an expected-value-maximizing policy select in that state?
Why must the probability of a positive trial result under a modified regimen be conditioned on the modification action, rather than automatically copied from the continuation branch?
For the threshold example, what action is preferred at $p=0.10$, and what are the three action values?

Solutions. (1) $0.30(10)+0.70(4)=3+2.8=\$5.8$ million. (2) Modify, because $11 million exceeds both $8 million and $0 million. (3) The action may change the regimen and therefore the chance of each later evidence state; the relevant transition probability is conditional on the action as well as the preceding evidence. (4) Stop is preferred: continue is $120(0.10)-35=-\$23$ million, modify is $90(0.10)-15=-\$6$ million, and stop is $0 million.