Why Timing Changes Information Value
A study can improve a development decision and still be a poor investment. Information has value because it may change which action is chosen; however, collecting it takes time and resources. While a study is underway, a decision may be delayed, potential benefits may arrive later, and development opportunities may change. The relevant question is therefore not only how much better decisions could be with information, but whether the net value today of the study policy exceeds the value of the best action available now.
Earlier material on expected value of sample information (EVSI) defines the expected gain from sample information relative to deciding without that information. Here, the key extension is to place the study cost and the consequences of delay on a common time and monetary scale. This requires making the timing of every payoff explicit.
In the examples, monetary values are in millions of dollars, time is measured in years, and the annual effective discount rate \(r\) is constant. Thus \(D(T)=(1+r)^{-T}\). This convention discounts a value received at the end of year \(T\) to time zero. If costs or benefits occur on different dates, discount each at its own date rather than treating them as if they all occurred at the same time.
Assumptions and the Timing Comparison
A simple timing-adjusted EVSI calculation relies on several assumptions. First, the actions being compared are genuinely available at the stated decision times. Second, the predictive distribution for the study result and the payoffs conditional on that result are specified consistently. Third, the discount rate and timing of cash flows are appropriate for the decision. Fourth, the study does not change the underlying state except through the information it provides. If disease progression, competitor entry, patent expiry, or operational changes can alter future payoffs, those effects must be represented in the future payoff model rather than ignored.
To see the logic, suppose the expected optimized value after the study is \(W_T\), and the same decision opportunity, evaluated at time \(T\), would be worth \(V_T\) without the study. The gross EVSI at time \(T\) is \(W_T-V_T\). If the no-study opportunity is still available at time \(T\), comparing the discounted study policy with the current policy gives:
This expression includes both the information effect and the time shift. It is not generally equal to \(D(T)(W_T-V_0)-C_0\) unless the baseline value \(V_0\) is also being treated as a payoff at time \(T\). A clean comparison keeps each option’s cash flows on their actual dates. In many practical cases, the no-study action can be taken now and the study option pays later; then the formula above is the direct comparison.
The short derivation is simply the expected-value accounting identity: the study policy produces the future expected decision value \(W_T\), whose present value is \(D(T)W_T\), and requires the present cost \(C_0\). Subtracting the value \(V_0\) of the best action now gives the incremental value. This is a decision comparison, not a claim that discounting itself measures the scientific quality of the evidence.
Worked Example: A Short Study That Adds Net Value
Suppose a hypothetical program can be advanced now for a payoff of \(40\) if the underlying state is good and \(-20\) if it is poor; stopping has payoff \(0\). The probability of the good state is \(0.40\). A two-year study produces a positive or negative signal. Assume \(P(+\mid\text{good})=0.80\) and \(P(+\mid\text{poor})=0.25\). The study costs \(1.00\) million at the start. Use a 10% annual effective discount rate.
Value of acting now. The expected payoff from advancing is \(0.40(40)+0.60(-20)=4.00\) million. Stopping is worth zero, so \(V_0=4.00\) million. The decision to advance now is therefore preferred to stopping.
Expected value after the signal. The probability of a positive result is \(P(+)=0.40(0.80)+0.60(0.25)=0.47\). By Bayes’ rule, \(P(\text{good}\mid +)=0.32/0.47\approx0.6809\). Given a positive result, the expected payoff from advancing is \(40(0.6809)-20(1-0.6809)\approx20.85\) million, so advancing is preferred to stopping.
For a negative result, \(P(-)=0.53\) and \(P(\text{good}\mid -)=0.40(0.20)/0.53=0.08/0.53\approx0.1509\). The expected payoff from advancing is \(40(0.1509)-20(1-0.1509)\approx-10.94\) million, so stopping is preferred. The expected optimized value at the end of the study is therefore \(W_T=0.47(20.85)+0.53(0)\approx9.80\) million. Equivalently, the positive-result contribution is \(0.40(0.80)(40)-0.60(0.25)(20)=9.80\) million.
Discount and compare. The discount factor is \(D(2)=1/(1.10)^2=1/1.21\approx0.8264\). Thus \(V_{\text{study}}=9.80/1.21-1.00\approx7.099\) million, and \(\Delta=7.099-4.00\approx3.099\) million. Under these assumptions, the study is preferred. Its expected optimized value at completion exceeds the value of acting now, and the remaining advantage is still positive after discounting and paying the study cost.
How Delay Can Reverse the Decision
An informative study does not guarantee a positive net value. The information may raise expected decision value at the time it arrives, while discounting and study costs reduce its value today. Moreover, the delay may cause real losses, such as a shorter period of market exclusivity or the loss of a timely treatment opportunity. Those losses should be represented in the payoff at the appropriate date. If they are already included in \(W_T\), they must not be subtracted a second time as a separate delay penalty.
Worked Example: A Slow Study Reduces Net Value
Consider a hypothetical program with a best expected payoff of \(12.00\) million if the team acts now. Suppose a proposed study would lead to an expected optimized decision value of \(29.40\) million when its results arrive eight years from now. The study costs \(3.00\) million at the present, and the annual effective discount rate is 10%. Assume the future value already reflects the development choices available after the study, and that there are no other costs or benefits.
The discount factor is \(D(8)=1/(1.10)^8=1/2.14358881\). The present value of the informed decision is \(29.40/(1.10)^8\approx13.72\) million. After subtracting the upfront study cost, \(V_{\text{study}}\approx13.72-3.00=10.72\) million. Using unrounded values, \(V_{\text{study}}\approx10.7153\) million, so \(\Delta=10.7153-12.00\approx-1.2847\) million, or about \(-1.28\) million. The study is informative, but under these assumptions it is inferior to acting now.
The example also shows why rounding should be done at the end. The unrounded present value is approximately \(13.7153\) million; rounding it first to \(13.72\) gives the stated net value and incremental comparison to two decimal places. Reporting \(13.71\) would be an incorrect rounding of that value.
Study-Cost Timing and Break-Even Comparisons
Costs also need dates. A cost paid at the start is subtracted without discounting. A cost paid only when results arrive is a future cash flow and should be discounted along with other end-of-study values. If a study has payments at several times, its present cost is the sum of those discounted payments. This distinction can matter when a study is long or expensive.
Worked Example: Moving a Cost to the End of the Study
Suppose a hypothetical decision is worth \(8.00\) million if made now. A three-year study is expected to produce an optimized decision value of \(15.00\) million at completion. The discount rate is 8%, and the study cost is \(2.00\) million. First assume the cost is paid at the start.
The discount factor is \(D(3)=1/(1.08)^3=1/1.259712\approx0.7938\). The present value of the future decision is \(15.00/1.259712\approx11.9075\) million. With the \(2.00\) million upfront cost, \(V_{\text{study}}\approx11.9075-2.00=9.9075\) million and \(\Delta=9.9075-8.00=1.9075\) million. The study is preferred on expected value.
Now suppose the same \(2.00\) million cost is paid at completion, rather than at the start. Its present value is \(2.00/1.259712\approx1.5877\) million. The study policy is worth \((15.00-2.00)/1.259712\approx10.3198\) million today, giving an incremental value of \(10.3198-8.00\approx2.3198\) million. The difference is not a change in study effectiveness; it is the effect of paying the same cost later. In an actual budget, payment timing should reflect the contracted cash flows, not a convenient calculation.
A useful screening calculation is the break-even condition. If the study cost is paid at the start and the expected optimized value \(W_T\) is fixed, the study is preferred when \(D(T)W_T-C_0>V_0\). Equivalently, its future value must exceed the threshold \((V_0+C_0)/D(T)\). This makes the impact of duration visible: with positive discounting, a longer wait raises the future value needed to justify the same present alternative and upfront cost.
For example, if \(V_0=8\), \(C_0=2\), and \(D(3)=1/1.259712\), the future value must exceed \(10/0.7938\approx12.60\) million to beat acting now. A value of \(15.00\) million clears that threshold. This is a threshold for the specified assumptions, not a universal minimum EVSI: changing the comparator, the payment schedule, or the effects of delay changes the calculation.
Diagnostics, Alternatives, and Limitations
Before interpreting a timing-adjusted result, check the model in concrete terms:
- Check dates and units. Identify when each study cost, decision, and payoff occurs. Do not discount a value twice, and do not combine amounts in different currencies or units without conversion.
- Check the comparator. Compare the study with the best feasible action now, not with a deliberately unattractive alternative. If a no-study decision can also be made later, evaluate that option on the same future date.
- Check what the future value includes. Confirm whether it already accounts for launch delay, exclusivity, changing disease burden, or other time-sensitive consequences. Add a delay penalty only if that loss is not already reflected in the payoff.
- Check assumptions about information and execution. A study may fail operationally, produce unusable data, or change the set of available actions. If these possibilities are material, represent them in the probability model and payoff calculation.
- Check sensitivity. Recalculate over plausible study durations, discount rates, costs, and future payoffs. A small positive \(\Delta\) may not be robust to uncertainty in these inputs.
The main alternative to a study is often to act now, as in the examples. Other relevant alternatives can include stopping, using existing evidence to choose a different dose or population, or collecting a smaller or faster sample. Each option should be valued as a policy with its own costs and payoff dates. The value-of-information comparison does not itself decide which study design is best; it evaluates the specified information-gathering option against specified alternatives.
A common mistake is to say, “The expected value after the study is larger, so the study must be worthwhile.” That reasoning compares \(W_T\) with a present value \(V_0\) without accounting for timing or cost. The correction is to calculate each policy’s value on a common present-value basis, as in \(\Delta=D(T)W_T-C_0-V_0\) when the study payoff arrives at \(T\) and its cost is paid now. Another mistake is to discount \(W_T\) and then subtract a separate delay penalty even though that penalty is already embedded in \(W_T\); this double-counts the loss. Finally, discounting is a financial convention, not a substitute for modeling real changes in clinical or commercial opportunity.
Check Your Understanding
For each question, first identify the timing convention, then compare the options on the same present-value basis. Solutions follow each problem.
- A study costs \(1.5\) million now and produces an expected optimized value of \(10\) million in two years. The discount rate is 5%, and the best action now is worth \(7\) million. What is the study’s incremental net value, and which option is preferred?
- Why is a positive EVSI measured at study completion not enough, by itself, to conclude that the study is worthwhile?
- A study’s future value already includes a loss from delayed market entry. Should the analyst also subtract that same delay loss as a separate cost? Explain.
- A study has value \(18\) million at completion in three years, and its entire \(2\) million cost is paid at completion. Write the present-value expression for the study policy using discount factor \(D(3)\).
Solutions. (1) \(D(2)=1/(1.05)^2=1/1.1025\approx0.9070\). The study value is \(10/1.1025-1.5\approx9.070-1.5=7.570\) million. Thus \(\Delta=7.570-7.000\approx0.570\) million, so the study is preferred under the stated assumptions. (2) EVSI at completion does not include, by itself, the effect of waiting and study costs. Discount the future value, place costs on their actual dates, and compare the resulting present value with the best available alternative. (3) No. If the loss is already included in the future value, subtracting it again double-counts the delay effect. (4) Both value and cost occur at completion, so the present value is \(D(3)(18-2)\) million. With an upfront cost instead, it would be \(D(3)(18)-2\) million.