Why the Timing of Futility Spending Matters
A futility rule can allocate its allowed stopping probability differently over the course of a trial. A front-loaded schedule assigns a relatively large share early; a back-loaded schedule reserves more for later looks. Both can have the same total spending by the end of the planned information period, yet imply different lower boundaries and different chances of stopping at each look.
The comparison is made under a specified design alternative \(H_A\). The spending target describes the probability, under that alternative, of having crossed a futility boundary by a given look. It is not automatically the overall probability of a type II error: that also depends on the efficacy rule and on what happens to trials that do not stop for futility.
A useful illustrative pair is the power family \(B_\gamma(t)=\beta_F t^\gamma\). For \(0<\gamma<1\), spending is front-loaded relative to the linear schedule; for \(\gamma>1\), it is back-loaded. These curves specify cumulative probability targets, not the \(Z\)-boundaries themselves. Boundary calibration must account for the correlation among interim statistics and the probability of crossing at earlier looks.
Model and Boundary Calibration
For the examples, use the canonical group-sequential normal model. Information \(I_k\) increases across looks, and under the design alternative the joint distribution of the standardized statistics is \(Z_k\sim N(\theta_A\sqrt{I_k},1)\), with \(\operatorname{Corr}(Z_i,Z_j)=\sqrt{I_i/I_j}\) for \(i<j\). Here \(\theta_A\) is the effect on the canonical standardized scale. This model assumes the stated information values are appropriate, the increments of information are independent, and the design alternative is specified. It is a planning model; departures from it can change the attained crossing probabilities.
Define \(U_k=Z_k-\theta_A\sqrt{I_k}\). Under \(H_A\), each \(U_k\) is standard normal, and the \(U_k\)'s have the same correlations as the \(Z_k\)'s. A boundary \(l_k\) is therefore equivalent to the centered threshold \(a_k=l_k-\theta_A\sqrt{I_k}\). With two looks, and temporarily considering only futility crossings, the cumulative crossing probability is \(P_A(Z_1\leq l_1\ \text{or}\ Z_2\leq l_2)\). For centered thresholds \(a_1,a_2\) and correlation \(\rho\), this is \(\Phi(a_1)+\Phi(a_2)-\Phi_2(a_1,a_2;\rho)\), where \(\Phi\) is the standard normal CDF and \(\Phi_2\) is the bivariate standard normal CDF.
The subtraction removes paths counted at both looks. Thus setting \(\Phi(a_2)\) equal to the cumulative target at look 2 would generally be wrong: it ignores first-look crossings, including their overlap with second-look crossings. In an actual design with an efficacy boundary, calibration must also account for paths that have already stopped for efficacy. The numerical illustration below deliberately omits efficacy boundaries so that the futility-spending comparison is transparent.
Worked Boundary Comparison
Worked Example: Calibrating Front-Loaded and Back-Loaded Boundaries
Suppose the planned total futility-crossing probability under \(H_A\) is \(\beta_F=0.20\). Consider two futility looks at \(t_1=0.25\) and \(t_2=0.50\), with schedules \(B_F(t)=0.20\sqrt{t}\) (front-loaded) and \(B_L(t)=0.20t^2\) (back-loaded). For this hypothetical design, let the mean \(Z\)-statistics under \(H_A\) be 1.40 and 1.98 at these looks. The information ratio is \(I_1/I_2=0.50\), so the centered statistics have correlation \(\rho=\sqrt{0.50}=0.7071\). Assume no efficacy stopping before these looks, as in the illustrative calibration above.
At the first look the crossing event is marginal, so \(a_1=\Phi^{-1}\{B(t_1)\}\) and \(l_1=1.40+a_1\). For the front-loaded schedule, \(B_F(0.25)=0.20\sqrt{0.25}=0.1000\); hence \(a_1\approx-1.282\) and \(l_1\approx1.40-1.282=0.118\). For the back-loaded schedule, \(B_L(0.25)=0.20(0.25)^2=0.01250\); hence \(a_1\approx-2.241\) and \(l_1\approx1.40-2.241=-0.841\).
At look 2, the targets are \(B_F(0.50)=0.1414\) and \(B_L(0.50)=0.05000\). For each schedule, choose \(a_2\) to satisfy \(\Phi(a_1)+\Phi(a_2)-\Phi_2(a_1,a_2;0.7071)=B(0.50)\), then set \(l_2=1.98+a_2\). Numerical evaluation of the bivariate normal probability gives approximately \(a_2=-1.375\) for the front-loaded schedule and \(a_2=-1.70\) for the back-loaded schedule. Thus the second-look boundaries are approximately \(l_2=1.98-1.375=0.605\) and \(l_2=1.98-1.70=0.280\), respectively. Rounding explains small discrepancies in the displayed crossing probabilities.
The front-loaded rule has the higher lower boundary at both looks, so it stops for futility more readily under the same observed \(Z\)-statistic. Its cumulative crossing target is also larger at both information fractions. The comparison is specific to the stated model and \(H_A\); these boundary numbers are not portable to a different effect assumption or information pattern.
From Cumulative Targets to Look-Specific Stopping
If the cumulative targets are attained exactly and the only stopping event considered is futility, the probability of first crossing at look \(k\) is the increase in cumulative crossing probability from the preceding look. This follows because the event “first cross at look \(k\)” is the difference between “cross by look \(k\)” and “cross by look \(k-1\).” Conditional on reaching look \(k\), the stopping probability is that increment divided by the probability of having avoided all earlier futility crossings.
Worked Example: Comparing the Probability of Stopping at the Second Look
Continue the hypothetical design. By the second look, the front-loaded schedule has cumulative target \(0.1414\), of which \(0.1000\) has already been allocated by the first look. Its target increase is \(0.1414-0.1000=0.0414\). The back-loaded schedule has target \(0.05000\), after \(0.01250\) at the first look, so its increase is \(0.05000-0.01250=0.03750\).
Under the stated futility-only calculation, these are the unconditional probabilities of first crossing at look 2 under \(H_A\). Among trials that reach look 2 without a prior futility crossing, the corresponding probabilities are \(0.0414/(1-0.1000)=0.0460\) for the front-loaded schedule and \(0.03750/(1-0.01250)=0.0380\) for the back-loaded schedule. The conditional probabilities are not the spending increments themselves: their denominators differ because the schedules leave different proportions of trials at risk of crossing.
This distinction matters when describing a design. Saying that the front-loaded schedule “spends 14.14% at look 2” is inaccurate: 14.14% is its cumulative target by that look. Its increase since the prior look is about 4.14 percentage points.
Worked Example: How Later Increments Reverse the Timing Contrast
At \(t=0.75\), the front-loaded curve has \(B_F(0.75)=0.20\sqrt{0.75}=0.1732\); the back-loaded curve has \(B_L(0.75)=0.20(0.75)^2=0.1125\). From \(t=0.50\) to \(t=0.75\), the front-loaded target increases by \(0.1732-0.1414=0.0318\), while the back-loaded target increases by \(0.1125-0.0500=0.0625\). The back-loaded schedule therefore assigns more of its cumulative crossing probability to this later interval.
If the targets are calibrated as cumulative first-crossing probabilities, the conditional stopping probabilities among trials that have avoided futility through \(t=0.50\) are \(0.0318/(1-0.1414)=0.0370\) for the front-loaded schedule and \(0.0625/(1-0.0500)=0.0658\) for the back-loaded schedule. These calculations show the intended timing contrast directly: the front-loaded schedule permits more early crossing, whereas the back-loaded schedule permits more crossing later. They do not imply that the later boundary must be lower or higher by a fixed amount; that depends on the joint distribution and effect assumption.
Derivation, Checks, and Design Choices
The two-look calibration equation follows from the addition rule for events. Let \(A_1=\{U_1\leq a_1\}\) and \(A_2=\{U_2\leq a_2\}\). A futility crossing by look 2 is \(A_1\cup A_2\), so \(P_A(A_1\cup A_2)=P_A(A_1)+P_A(A_2)-P_A(A_1\cap A_2)\). Under the canonical model, these three terms are \(\Phi(a_1)\), \(\Phi(a_2)\), and \(\Phi_2(a_1,a_2;\rho)\), respectively. This proves the calibration formula and explains why correlation is essential: it determines how much overlap to subtract.
In practice, numerical calibration can use bivariate or higher-dimensional normal probabilities, or an equivalent sequential computation. Check that the schedule begins at zero, is nondecreasing, stays within its planned total, and that the attained cumulative crossing probabilities reproduce the targets to numerical tolerance. Also check that the correct design alternative, information fractions, crossing direction, and prior stopping rules were used. A boundary that looks unexpectedly permissive or severe can signal a wrong tail, an incorrect correlation, or a confusion between marginal and cumulative probabilities.
A spending function is one way to control the timing of futility crossings. A fixed lower \(Z\)-cutoff is simpler to state, but its cumulative crossing probability is determined by the joint model and may not match a desired timing profile. Front-loaded and back-loaded spending instead specify the cumulative targets and require calibrated boundaries. A nonbinding rule is another operational choice: it may be monitored without being enforced, but its actual stopping probabilities then differ from those for an enforced rule. When efficacy boundaries are present, binding status and joint crossing calculations must be handled consistently.
The central limitation is that beta spending is alternative-dependent. A schedule calibrated under \(H_A\) does not guarantee the same crossing probabilities under another effect size. The information model also matters: misspecified information, dependence, or analysis timing can alter the joint distribution and hence the attained probabilities. Finally, \(\beta_F\), the planned probability of futility crossing, should not be reported as the overall type II error unless the design's full efficacy and futility operating characteristics justify that interpretation.
- Common mistake: Treating \(B(t_k)\) as \(P_A(Z_k\leq l_k)\). Correction: \(B(t_k)\) is a cumulative crossing target; calibration accounts for earlier crossings and their overlap with the current look.
- Common mistake: Assuming two schedules with the same terminal spending have the same boundaries. Correction: Their intermediate targets differ, and boundaries also depend on the alternative and joint distribution.
- Common mistake: Calling a target increment the conditional probability of stopping among those still in the trial. Correction: Divide the unconditional increment by the probability of reaching that look without an earlier futility crossing.
Check Your Understanding
Answer each question, then compare with the solution immediately below it.
- For \(\beta_F=0.16\), calculate \(B(0.25)\) under a front-loaded power schedule with \(\gamma=0.5\) and a back-loaded schedule with \(\gamma=2\).
Solution: The front-loaded value is \(0.16\sqrt{0.25}=0.0800\). The back-loaded value is \(0.16(0.25)^2=0.0100\). Both reach \(0.16\) at \(t=1\), but their early targets differ. - A schedule has cumulative targets \(0.06\) at look 1 and \(0.15\) at look 2. In a futility-only calculation, what is the unconditional probability of first crossing at look 2, and what is the conditional probability among trials without an earlier crossing?
Solution: The increment is \(0.15-0.06=0.09\). The conditional probability is \(0.09/(1-0.06)=0.0957\), or about 9.57%. - Why is \(\Phi(a_2)=B(t_2)\) generally not a valid two-look calibration equation?
Solution: It uses only the marginal probability of crossing at look 2. The cumulative probability is \(\Phi(a_1)+\Phi(a_2)-\Phi_2(a_1,a_2;\rho)\), which includes look-1 crossings and subtracts paths crossing at both looks. - Two schedules have the same \(\beta_F\) at \(t=1\). Does that ensure equal lower boundaries at every interim look?
Solution: No. Their intermediate cumulative targets can differ, and each target is translated into a boundary using the joint distribution under \(H_A\). Equal terminal spending alone does not equalize either targets or boundaries earlier.