Valuing Biomarker and Dose-Finding Studies

From Study Results to Development Value

A biomarker study can sharpen predictions about who may benefit from a treatment. A dose-finding study can help identify which dose offers the best balance of benefit and risk. But learning more is not, by itself, a reason to conduct either study. The decision question is whether the possible results could change a later development choice—and whether those changed choices are worth enough to offset the study’s cost.

The value calculation requires an explicit link between uncertainty and action. For a biomarker study, the uncertain state might describe whether a target population is responsive. For dose finding, it might describe which dose range has the most favorable benefit–risk profile. The study produces an outcome; the decision-maker uses that outcome to select a later action, such as advancing, stopping, enriching a trial population, or choosing a dose.

The probabilities \(P(\theta)\) represent uncertainty before the study, and \(P(Y=y\mid\theta)\) describes how likely each result is under each state. Together they determine the marginal probability of each result and the state probabilities conditional on that result. The payoffs should be on a common, decision-relevant scale—for example, expected net program value in millions of dollars—and should represent the later actions being compared.

The calculation assumes that the states and actions cover the decision being valued; that the prior probabilities and study-result likelihoods are appropriate for the program; and that the payoff model is consistent across actions and states. It also assumes the study result informs the choice through the stated likelihoods, and that the decision-maker can implement the action selected after seeing the result. If a biomarker does not affect a feasible treatment or development choice, its scientific interest alone does not create decision value in this model.

Why Information Has Value Only Through Decisions

For each possible result, calculate the expected payoff of every available action conditional on that result, then select the action with the greatest conditional expected payoff. Average those selected payoffs across possible results. The result is the expected payoff when the action can depend on the study. Subtract the best expected payoff available before the study to obtain EVSI.

A useful justification is that the decision-maker can always ignore the result and repeat the action that was best before the study. Choosing the best action separately after each result cannot produce a lower expected payoff than that fallback choice. Therefore EVSI is nonnegative under the stated assumptions. Information is valuable when it permits better choices, not merely when it reduces uncertainty.

The expected value of perfect information (EVPI) is the increase in expected payoff if the state were known before choosing an action. It provides an upper bound: a real study cannot be more informative than perfect knowledge of the state. Thus, under the same states, actions, and payoff model, \(0\leq\operatorname{EVSI}\leq\operatorname{EVPI}\). This bound is a useful check on a calculation.

Valuing Dose-Finding Information

Dose-finding studies are valued in the same decision framework, but their states and actions should represent the dose choice of interest. A state might identify which dose range has the most favorable benefit–risk profile; an action might be to take a candidate dose into a later trial. The payoff for each state–dose combination should reflect the program consequences of choosing that dose, rather than treating the numerically highest dose or response as automatically best.

Comparing Designs and Checking the Decision Model

A biomarker study and a dose-finding study can be compared by using the same decision framework, but their states, outcomes, actions, and payoffs should reflect their different purposes. A biomarker result may support a choice about population enrichment or whether to continue. A dose-finding result may change the dose selected for later development. In either case, a design’s value depends on how much it can change the best action and on the payoff difference between actions—not solely on how precise its measurements are.

Expected value of perfect information is an upper-bound comparison, not an estimate of what a feasible study will deliver. A decision threshold or go/no-go rule can be a simpler alternative when the decision is already well summarized by one probability and one payoff comparison. It does not, by itself, value a study with multiple possible results; that requires evaluating the action after each result. EVSI makes that comparison explicit, but its numerical result is only as credible as the probabilities, likelihoods, and payoffs supplied to it.

Practical checks include confirming that result probabilities sum to one, that conditional probabilities follow from the joint probabilities, and that each action is compared on the same payoff scale. Check that the selected action really maximizes conditional expected payoff for every result, and verify that EVSI is nonnegative and no larger than EVPI. If no result changes the selected action, EVSI will be zero under this model, even if the study would substantially reduce uncertainty. If actions change but their payoffs are similar, value may still be small.

Limitations and Common Pitfalls

The states and likelihoods simplify a complex development program. A biomarker may be measured with error, may not predict treatment benefit in the intended setting, or may be unavailable for a later decision. A dose’s benefit–risk profile may vary across patient groups, and the study’s results may not generalize to the later trial. These issues should be reflected in the states, likelihoods, available actions, or payoffs when they matter to the decision. Payoffs also need to include the relevant program consequences; omitting an important consequence can change which action is best.

This calculation treats the study’s cost as a direct subtraction from gross EVSI and assumes all values are comparable at the same decision point. If the study changes the timing of development or has other consequences, those effects must be represented in the valuation rather than silently ignored. The result is not a claim that the study will improve outcomes in every realized case; it is an expected comparison across possible states and study results under the stated model.

• Mistake: valuing information by how much it reduces uncertainty. A more precise biomarker estimate may not change any action. Correction: calculate whether results change the best action and the expected payoff from that action.
• Mistake: subtracting study cost before calculating gross EVSI and then subtracting it again. Correction: calculate gross EVSI first, then subtract each included study cost once to obtain net value.
• Mistake: choosing a dose from response alone. Correction: define state-specific payoffs that capture the benefit–risk consequences relevant to the later development choice.
• Mistake: treating a positive biomarker result as proof of treatment benefit. Correction: use the modeled conditional probabilities and decision payoffs; a study result updates uncertainty but does not establish causation by itself.