Choice of futility boundaries for group sequential designs with two endpoints
 2.5k Downloads
 2 Citations
Abstract
Background
In clinical trials, the opportunity for an early stop during an interim analysis (either for efficacy or for futility) may relevantly save time and financial resources. This is especially important, if the planning assumptions required for power calculation are based on a low level of evidence. For example, when including two primary endpoints in the confirmatory analysis, the power of the trial depends on the effects of both endpoints and on their correlation. Assessing the feasibility of such a trial is therefore difficult, as the number of parameter assumptions to be correctly specified is large. For this reason, socalled ‘group sequential designs’ are of particular importance in this setting. Whereas the choice of adequate boundaries to stop a trial early for efficacy has been broadly discussed in the literature, the choice of optimal futility boundaries has not been investigated so far, although this may have serious consequences with respect to performance characteristics.
Methods
In this work, we propose a general method to construct ‘optimal’ futility boundaries according to predefined criteria. Further, we present three different group sequential designs for two endpoints applying these futility boundaries. Our methods are illustrated by a real clinical trial example and by MonteCarlo simulations.
Results
By construction, the provided method of choosing futility boundaries maximizes the probability to correctly stop in case of small or opposite effects while limiting the power loss and the probability of stopping the study ‘wrongly’. Our results clearly demonstrate the benefit of using such ‘optimal’ futility boundaries, especially compared to futility boundaries commonly applied in practice.
Conclusions
As the properties of futility boundaries are often not considered in practice and unfavorably chosen futility boundaries may imply bad properties of the study design, we recommend assessing the performance of these boundaries according to the criteria proposed in here.
Keywords
Group sequential design Stopping for futility Two endpoints Intersectionunion testBackground
In recent years, the flexibility and efficiency of clinical trials became increasingly important, in particular for trials from the pharmaceutical industry. Group sequential designs give the opportunity to stop the study early during an interim analysis, thereby saving time and financial resources. Generally, the study can either be stopped for efficacy if the study goal is prematurely achieved or for futility, if reaching the aim of the trial seems desperate. Decision rules to stop a trial early for efficacy have been broadly investigated. Since the pioneering works of Pocock [1] and O’Brien and Fleming [2], these designs have been implemented since long in clinical trial routine. A comprehensive overview and general concepts of the statistical methodology in group sequential designs are provided by Jennison and Turnbull [3]. With respect to futility stopping, there are mainly two fundamental approaches in the statistical literature [3, 4]. The first approach is a conditional one, where the study is stopped for futility, if the conditional power falls under a prespecified threshold. This conditional approach can further be divided into stochastic curtailment [5, 6], a frequentist approach, and methods based on the predictive power or the predictive probability [4], which are partially or fully Bayesian methods [7, 8]. The second approach is to define futility stopping boundaries either in isolation or in conjunction with group sequential efficacy boundaries [3]. The choice of adequate boundaries to stop the study early for futility has been discussed only briefly in these works. This is astonishing as unfavorable futility boundaries may have serious consequences with respect to the performance of the study design.
In our work, we propose a general method for the construction of futility boundaries. In addition, we introduce several intuitive criteria which should be considered when defining ‘optimal’ futility boundaries described as follows. First, unless offset by increased sample size, early stopping for futility reduces the overall power. The aim is thus to avoid a too high loss in power. Second, the study should be stopped early for futility if the unknown true effect is far away from the anticipated effect under the alternative hypothesis (‘correctly’ stopping for futility). Finally, if the unknown true effect corresponds to the anticipated effect under the alternative hypothesis, the study should not be stopped for futility (‘wrongly’ stopping for futility).
In principle, these optimality criteria might be applied in any twostage design. In the literature, group sequential designs with futility stopping are usually formulated for a single primary endpoint [9, 10, 11, 12]. Beside several works on interim efficacy evaluation in group sequential designs with multiple endpoints [13, 14, 15, 16, 17, 18] there are only few works suggesting group sequential methods for multiple endpoints including futility stopping [19, 20]. However, we believe that such designs are of particular relevance for clinical trials with multiple primary endpoints for the reason specified in the following. When considering two endpoints, the power of the trial depends on the expected effects of both endpoints and on the correlation between them. Assessing the feasibility of the trial is therefore a particular challenge, as the number of required parameter values to be correctly specified in the planning stage is large. Therefore, the option for an early futility stop is attractive to account for these uncertainties. Moreover, the required sample size for a clinical trial with several primary endpoints is usually higher than for a single primary endpoint and therefore the need to stop the trial early in case of unfavorable outcomes is of particular importance.
In this work, we apply the general construction method of futility boundaries to three different twostage group sequential designs for two endpoints. For these designs, we propose an algorithm to choose ‘optimal’ futility boundaries with respect to the above mentioned criteria. Thereby as Jennison and Turnbull [3], we define group sequential futility stopping boundaries in conjunction with efficacy boundaries.
This paper is organized as follows. The “Methods” section introduces first the study design, the test problem and general principles of group sequential designs and then describes the new approach of choosing ‘optimal’ futility stopping boundaries as motivated above. Moreover, different group sequential designs for two endpoints profiting from these optimal futility boundaries are proposed. Furthermore, the derivation of ‘optimal’ futility boundaries is described. The “Results” section investigates the performance characteristics of the new methods by means of MonteCarlo simulations cased on a real clinical trial example. Finally, we conclude with a discussion.
Methods
Standard group sequential methods allow stopping a study early at an interim analysis where the outcome is observed and analyzed for a part of the maximum number of patients. The study can thereby either be stopped for efficacy or for futility. For simplicity, we restrict our considerations to a controlled clinical trial with one interim analysis and one final analysis resulting in a group sequential design with two stages. In this section, we first consider the situation of a single endpoint in order to describe the new approach of choosing futility boundaries generally. After that, group sequential designs for two endpoints are proposed.
Test problem and general principles
where δ=0 implies a onesided superiority test problem and δ<0 corresponds to a test for noninferiority. Considering binary or continuous outcome measures, θ is given by the absolute risk difference or mean difference, respectively, whereas for timetoevent data θ indicates the logarithm of the hazard ratio. The standard approaches to test for superiority are given by the chisquare test for binary data, the ttest for continuous data and the logrank test for timetoevent data [21]. Note that all these test statistics are approximately standard normally distributed under the corresponding null hypothesis.
In a twostage group sequential design the test statistics T _{1} and T _{2} used at the interim analysis and at the final analysis, respectively, corresponds to the standard test statistics of a fixed design using all data collected so far with some modifications for the case of timetoevent data [22]. The correlation between the normally distributed test statistics T _{1} and T _{2} then exclusively depends on the (assumed) information fraction at interim which can be specified in the planning stage. After the correlation between the test statistics has been determined, the adjusted (onesided) local significance levels α _{1} and α _{2} can be defined by taking this correlation into account. The local significance levels in a group sequential design can be chosen in various ways, e.g. as constant or increasing in time [1, 2, 23, 24, 25]. Then the null hypothesis given in (1) is rejected at interim whenever the onesided pvalue p _{1} referring to T _{1} fulfills p _{1}≤α _{1}. The null hypothesis is rejected at the final analysis whenever the onesided pvalue p _{2} corresponding to T _{2} fulfills p _{2}≤α _{2}.
In general, stopping for futility without compromising the type I error is possible at any time and independent of any predefined rules as an early acceptance of H _{0} decreases the actual type I error rate. In the context of group sequential designs, it can generally be differentiated between binding and nonbinding stopping for futility rules, compare also Bretz et al [26]. ‘Binding’ means that stopping for futility at the interim analysis is obligatory whenever the futility criteria are met. When the data suggest stopping for futility, it is thus not allowed to continue the trial for other external reasons. If a binding futility rule is applied, the local significance levels can be increased in order to fully exhaust the global significance level which is otherwise no longer guaranteed as futility stopping implies a lower probability of rejecting the null hypothesis. In contrast, the nonbinding version does not commit early futility stopping. Therefore, there may be situations were the data advise stopping for futility, but the study is continued nonetheless for other reasons, e.g. as new external information suggests that the futility criteria might be too strong. As a consequence, the local significance levels cannot be adjusted and the global significance level is not fully exhausted. In clinical trial applications nonbinding futility boundaries are usually applied because they allow reacting flexibly to interim results such as adverse events or new external information. However, quantifying the performance properties (in terms of power loss or ‘correctly’ and ‘wrongly’ stopping for futility) of nonbinding rules is impossible as the study progress is not predictable from the observed effect at interim.
For the reasons specified above, we will focus in this work on binding stopping criteria at interim but without increasing the local significance levels. In general, a futility rule can equivalently be expressed either in terms of a boundary for the test statistic or as an upper bound for the pvalue. We will use the latter approach without loss of generality, that is the study is stopped for futility at interim whenever p _{1}>α _{ f }, where α _{ f } is the futility boundary.
Defining optimal binding futility boundaries
Choosing adequate futility boundaries is an important challenge as unfavorable futility boundaries may have serious consequences with respect to the performance of the study design. In case of ‘strong’ futility boundaries, for example, if the study is stopped for futility whenever the onesided pvalue is larger than 0.2, the overall power loss can be large and the study might be stopped for futility in too many situations caused by only small but nonrelevant deviations from the planning assumptions. In such cases, the probability of ‘wrongly’ stopping for futility is high. In the case of ‘liberal’ futility boundaries, given, for example, as a lower bound for the pvalue of 0.8, the overall power loss is quite small but at the same time small or opposite effects often do not result in an early stop for futility. Hence, the probability of ‘correctly’ stopping for futility is low in this case [9]. The idea of ‘optimal’ futility boundaries proposed here is to provide a high rate of ‘correctly’ stopping for futility and to simultaneously restrict the loss in power and the rate of ‘wrongly’ stopping for futility. To provide ‘optimal’ futility boundaries in this sense, an ‘admissibility condition’ is defined as follows.
Definition 1

1. \(P_{H_{1,\theta _{1}}^{\delta }}(H_{0}^{\delta }\,\) is rejected in stage 1) + \(P_{H_{1,\theta _{1}}^{\delta }}(H_{0}^{\delta }\) is rejected in stage 2 and \(H_{0}^{\delta }\) is neither rejected nor accepted in stage 1) ≥1−β−β _{ l },

2. \(P_{H_{1,\theta _{1}}^{\delta }}\)(The study is stopped for futility based on α _{ f } in stage 1) ≤γ,
where \(P_{H_{1,\theta _{1}}^{\delta }}(\cdot)\)denotes the probability under the assumption that \(H_{1,\theta _{1}}^{\delta }\) holds true.
For predefined values of β _{ l } and γ there generally exist several β γadmissible futility boundaries. These boundaries differ in the probability of ‘correctly’ stopping for futility as there is no condition on exhausting the admissible power loss β _{ l } or the probability of ‘wrongly’ stopping for futility γ.
In order to determine ‘optimal’ futility boundaries, the probability of early stopping for futility should be preferably high in case of a small or opposite effect which deviates considerably from the anticipated treatment effect θ _{1}. This motivates the following definition.
Definition 2
The β γoptimal futility boundary defines a lower bound for the β γadmissible boundaries, as all boundaries that are larger than the β γoptimal boundary automatically meet the β γadmissible conditions. Note that for group sequential designs for a single endpoint (as well as for designs with two endpoints we consider in this work) the optimal futility boundary could also be determined by maximizing the probability of correctly stopping under the null hypothesis effect θ instead of under θ ^{∗}. However, in more complex multiple endpoint group sequential designs this monotonicity property might no longer hold true. The derivation of β γoptimal futility boundaries will be described below in “Derivation of the local significance levels and β γoptimal futility boundaries” section.
In general, the optimality of group sequential designs is usually assessed by means of the average sample size, where a low average sample size is preferable. Although the average sample size is a common criteria to judge the performance of a group sequential design, it has also major shortcomings when applied as the unique measure of performance. For example, the power loss of a group sequential design compared to a correctly specified singlestage design is a further performance criterion. Liu et al. proposed a performance score combining both criteria (average sample size and power loss) [27]. The application of the performance score for the situation of an intersectionunion test is, e.g., provided by Kieser et al. [28]. Despite these important new aspects discussed in literature, there further remain some open topics with respect to a performance assessment of a group sequential design: The average sample size is a summary measure which does not necessarily show the true sample sizes because the variability of the sample size is completely ignored. Therefore, a low average sample size is only a good optimality criteria if the variability of the sample size is also low. Instead of looking at the average sample size and its variability, an alternative approach could be to judge the correctness of early stopping (implying a low sample size) or continuing (implying a high sample size). The latter is what we have investigated in our work.
Investigated group sequential designs with futility stop based on two endpoints
Motivated by the fact that the application of adequate binding futility boundaries is of particular interest when analyzing several endpoints, we consider three different group sequential designs that incorporate two endpoints and different futility stopping rules. The aim is to define a test procedure which offers a maximal gain in information from two endpoints of interest but simultaneously requires a minimal number of patients to save resources, especially when the effects are lower than originally anticipated. These specific twostage designs will subsequently be used to illustrate the impact of β γoptimal futility boundaries. In the following, the indexes EP _{1} and EP _{2} will denote the affiliation to the two endpoints under investigation.
where the local test hypotheses are given as stated in (1). The most rigorous requirement for a clinical trial with two primary endpoints is to base the efficacy proof on demonstrating superiority for both equivalently relevant endpoints (\(H_{0}^{IUT, 0, 0}\) versus \(H_{1}^{IUT, 0, 0}\)). In the case that the endpoints are of different relevance, e.g. an efficacy and a safety endpoint, a less rigorous test procedure may also be appropriate. Therefore, Approach 2 combines a superiority test for the efficacy endpoint with a noninferiority test for the safety endpoint or the endpoint of less clinical relevance (\(H_{0}^{IUT, 0, \delta _{2}}\) versus \(H_{1}^{IUT, 0, \delta _{2}},\ \delta _{2}<0\)). Note that for the ease of representation, it is assumed without loss of generality that the efficacy endpoint corresponds to EP _{1}. A group sequential test procedure including binding futility stopping rules for \(H_{0}^{IUT,0,0}\) or \(H_{0}^{IUT,0,\delta _{2}}\), respectively, is defined as follows.

Stage 1: ∙ The study is stopped early with rejection of \(H_{0}^{IUT,\delta _{1},\delta _{2}}\) if \(p_{1}^{EP_{1}} \leq \alpha _{1}^{EP_{1}}\) and \(p_{1}^{EP_{2}} \leq \alpha _{1}^{EP_{2}}.\)

∙ The study is stopped early for futility (with acceptance of \(H_{0}^{IUT,\delta _{1},\delta _{2}}\)) if \(p_{1}^{EP_{1}} \ge \alpha _{f}^{EP_{1}}\) or \(p_{1}^{EP_{2}} \ge \alpha _{f}^{EP_{2}}.\)

Stage 2: At the final analysis, \(H_{0}^{IUT,\delta _{1},\delta _{2}}\) is rejected if \(p_{2}^{EP_{1}} \leq \alpha _{2}^{EP_{1}}\) and \(p_{2}^{EP_{2}} \leq \alpha _{2}^{EP_{2}}.\) Otherwise \(H_{0}^{IUT,\delta _{1},\delta _{2}}\) is accepted.
which corresponds to the definition given in (1) with δ=0 and the related group sequential procedure is defined as follows.

Stage 1: ∙ The study is stopped early with rejection of \(H_{0}^{EP_{1},0}\) if \(p_{1}^{EP_{1}} \leq \alpha _{1}^{EP_{1}}.\)

∙ The study is stopped early for futility if \(p_{1}^{EP_{1}} \ge \alpha _{f}^{EP_{1}}\) or \(p_{1}^{EP_{2}} \ge \alpha _{f}^{EP_{2}}.\)

Stage 2: At the final analysis, \(H_{0}^{EP_{1},0}\) is rejected if \(p_{2}^{EP_{1}} \leq \alpha _{2}^{EP_{1}}.\) Otherwise \(H_{0}^{EP_{1},0}\) is accepted.
Unlike for the Approaches 1 and 2, the efficacy proof is now only based on EP _{1}. However, the other endpoint EP _{2} still influences the study result, as early stopping for futility due to EP _{2} is possible.
Derivation of the local significance levels and β γoptimal futility boundaries
As motivated above, we do not increase the local significance levels to fully exhaust the type I error. Therefore, the local levels can be chosen as usual in group sequential designs without taking futility stopping into account, e.g. constant or increasing [1,2]. The derivation of increased local significance levels in conjunction with futility stopping (in order to exhaust the overal significance level) is described in the Supplementary Material (see Additional file 1).
To calculate the β γoptimal futility boundaries for a specific study situation, we implemented a search algorithm in R which at first determines all constellations of futility boundaries for both endpoints which simultaneously maximally exhaust the prespecified thresholds for the probability of ‘wrongly’ stopping for futility γ and for the power loss β _{ l }. Note that these are the β γadmissible boundaries. Subsequently the constellations of boundaries which maximize the probability of ‘correctly’ stopping for futility are determined, which are the β γoptimal boundaries.
Results
To illustrate application of the proposed methods, the three group sequential designs incorporating two endpoints are applied (by simulations) to a real clinical trial example. For these three designs, β γadmissible and β γoptimal futility boundaries are calculated and furthermore investigated in terms of overall power and probability of ‘correctly’ and ‘wrongly’ stopping for futility.
Clinical trial example and simulation design
The RENAAL study was a randomized, doubleblind, placebocontrolled trial conducted to investigate the effect of losartan on renal and cardiovascular outcomes in patients with type 2 diabetes and nephropathy [29]. The primary outcome was a timetofirst event composite endpoint, where the events correspond to doubling the baseline serum creatinin concentration, endstage renal disease, and death. The recruitment time was fixed to 2 years and the minimal followup duration to 3.5 years. Patients were allocated to the placebo and the intervention group in a 1:1 ratio. The original planning assumptions used for sample size calculation were given by estimated 5 year event rates of 0.58 and 0.464 in the placebo and the intervention group, respectively [30]. Assuming constant hazard functions over time (exponentially distributed survival times), the underlying hazards can be directly calculated from the given event rates [31]. The resulting hazard ratio for the composite endpoint is given by \(\lambda _{CE}^{C}/ \lambda _{CE}^{I}=0.0145/0.0104=1.394\), wherethe parameter λ denotes the hazard function, which is assumed to be constant here, the index CE stands for ‘composite endpoint’ and the group affiliation is expressed by the indexes I and C for the intervention and the control group, respectively.
The clinically most relevant component of the composite endpoint defined above is clearly given by death. Within the context of a composite endpoint, current guidelines on clinical trials systematically recommend to investigate the components of a composite endpoint separately, in particular the most relevant components [32, 33, 34]. Therefore, it might provide a relevant gain in information to include this endpoint in the analysis strategy instead of exclusively considering the composite endpoint. The two endpoints under consideration are thus given by a composite endpoint combining doubling of baseline serum creatinin concentration, endstage renal disease, and death and by the endpoint death alone. Thus, whenever a death occurs, this corresponds to an event in both endpoints. However, the composite endpoint consist of more events of other types. This illustrates that the two endpoints are correlated by construction. The assumed event rates for the endpoint death have not been published. For the sake of illustration, we assume a hazard ratio for death given by \(\lambda _{MC}^{C}/\lambda _{MC}^{I}=0.01/0.0074=1.351\), where the index MC stands for ‘main component’.
The presented example is appealing in the sense that the two endpoints correspond to a composite endpoint and a main component. In the specific case of a composite endpoint, any additional confirmatory information on the components provides an important gain in information. Therefore, application of all three group sequential approaches presented in this manuscript can be illustrated by means of a unique example. In most other clinical trial applications only one of the proposed test problems fits the specific confirmatory requirements.
‘Correctly’ and ‘wrongly’ stopping for futility
In order to differentiate between ‘correctly’ and ‘wrongly’ stopping for futility, we modify the underlying hazard assumptions. Consequently, we consider different hazard ratios for the main component that deviate from the original planning assumption given by \(\lambda _{MC}^{C}/ \lambda _{MC}^{I}=0.01/0.0074=1.351\). Of course, deviations from the planning assumptions could also occur in the composite endpoint or in both endpoints. As every event in the main component also corresponds to an event in the composite endpoint and for sake of an easier illustration, we restrict our considerations to deviations in the main component. The following 7 hazard ratio scenarios for the main component are considered {1.351;1.3;1.25;1.2;1.15;1.1;1.05}, where without loss of generality the hazard in the intervention group was fixed to \(\lambda _{MC}^{I}=0.0074\) and deviations in the hazard ratio are due to variations from the assumptions for \(\lambda _{MC}^{C}\). Scenario 1 corresponds to the original planning assumptions for the main component while the remaining Scenarios 2 to 7 correspond to decreasing treatment effects in the main component. To determine β γoptimal futility boundaries according to Definition 2, the largest effect under the alternative hypothesis has to be determined for which stopping the study for futility would still be considered as ‘correct’. In the RENAAL study, a hazard ratio of \(\lambda _{MC}^{C}/\lambda _{MC}^{I}=1.2\) might be a reasonable choice for the main component. Consequently, misspecifications of the hazard ratio given in Scenarios 2 and 3 might be acceptable, whereas for Scenario 4 to 7 it would be justified to stop the study early for futility. This threshold should be based on aspects of clinical relevance and should be discussed with clinical experts. Note that we base for sake of simplicity the threshold for ‘correctly’ stopping for futility in this example only on deviations from the planning assumptions in the main component. This threshold can also depend on the composite endpoint or even on both endpoints.
Sample size considerations
On the way to determine β γoptimal futility boundaries and especially to assess the overall power loss β _{ l }, at first the reference group sequential design without stopping for futility has to be fixed. For this illustrating example, we apply a twostage group sequential design where the interim analysis is performed after an anticipated information fraction of 0.5 and where the local significance levels are adjusted according to Pocock [1]. Moreover, we assume that patient recruitment is stopped during the interim analysis. All these settings could be chosen differently and are only of illustrative purpose here. We assume a target power of β=0.90 and an overall onesided significance level of α=0.025 for the reference design without stopping for futility. The required sample sizes for these reference designs are then given by 1260 (630 per group) for Approach 1 and by 820 (420 per group) for the less stringent Approach 2 where the main component is tested for noninferiority with a noninferiority margin of 0.9 in terms of the hazard ratio. The correlation of the test statistics are then given by r _{1}=0.83 and r _{2}=0.82 for Approach 1 and Approach 2, respectively. For Approach 3, testing exclusively the composite endpoint yields a sample size of 730 (365 per group). An analytical derivation of the required sample sizes and the correlations between the test statistics is difficult and the results are thus based on simulations. More details on the sample size derivations are described in the Supplementary Material (see Additional file 2).
Simulation results
For each scenario described above, we simulated 10.000 data sets and applied the corresponding groupsequential designs. From these, rates of overall power and probability of wrongly or correctly stopping for futility, respectively, are estimated for different constellations of stopping for futility boundaries. The power loss is given as the anticipated power minus the observed overall power. The rates of early stopping in Scenario 1 (according to the planning assumptions) or Scenario 4 lead to the probability of wrongly or correctly stopping for futility, respectively. For varying constellations of futility boundaries the corresponding values of overall power and probability of stopping for futility differ consequentially. A search algorithm (compare “Defining optimal binding futility boundaries” section) finally chooses the constellation of futility boundaries yielding an acceptable power loss while minimizing wrongly stopping for futility. This algorithm can be obtained from the authors on request.
Optimal to nonoptimal futility boundaries
In order to calculate the β γadmissible and β γoptimal futility boundaries the maximal acceptable power loss with stopping for futility is set to β _{ l }=0.05 and the admissible ‘wrongly’ stopping for futility rate is chosen as γ=0.025.
For Approach 1, futility bounds given by \(\alpha _{f}^{CE}=0.43\) and \(\alpha _{f}^{MC}=0.44\) maximize the ‘correctly’ stopping for futility rate at 0.117. In Approach 2, the β γoptimal futility boundaries are given by \(\alpha _{f}^{CE}=0.65\) and \(\alpha _{f}^{MC}=0.37\) with a probability of ‘correctly’ stopping for futility of 0.087. Approach 3 yields β γoptimal boundaries of \(\alpha _{f}^{CE}=0.69\) and \(\alpha _{f}^{MC}=0.65\) with a probability of ‘correctly’ stopping for futility of 0.085.
Comparison of design performance between optimal and nonoptimal futility boundaries
As a reference, we compare the β γoptimal futility bounds with a common futility boundary of \(\alpha _{f}^{CE}=\alpha _{f}^{MC}=0.5\) which is often applied in practice. For Approach 1 and Approach 2 this choice is more liberal than the β γoptimal boundaries. Therefore, the probability of ‘correctly’ stopping for futility is smaller with 0.09<0.117 for Approach 1 and 0.047<0.087 for Approach 2. Neither the probability of ‘wrongly’ stopping for futility (given as 0.016 for Approach 1 and 0.018 for Approach 2) nor the overall power loss (given as 0.017 for Approach 1 and 0.006 for Approach 2) exhaust the maximal admissible values γ=0.025 and β _{ l }=0.05. However, a common futility bound of 0.5 is contained in the set of admissible boundaries for the first two approaches. For Approach 3 the reference futility boundaries correspond to stricter values than the β γoptimal ones. Therefore, the probability of ‘correctly’ stopping for futility is higher with 0.158>0.085 but to the prize of a higher probability of ‘wrongly’ stopping for futility with 0.06 which exceeds the admissible value γ=0.025 considerably. At the other hand the overall power loss of 0.04 does not exhaust the maximal admissible value β _{ l }=0.05.
Impact of the planning assumptions on power loss and probability of stopping for futility
The right plot shows the overall power where for the β γoptimal futility boundaries the line of the original assumption (black line) should not fall below the admissible power 1−β−β _{ l }=0.90−0.05=0.85 (horizontal dotted line) for a sample size of 630. Decreasing the sample size or decreasing the hazard ratio (which corresponds to a larger deviation from the original planning assumptions) results in a monotone loss in overall power. Note that the global power depends on the correlation of the two test statistics. If the correlation is unknown in the planning stage, we recommend investigating the power by simulation. In general, the power increases with increasing absolute value of the correlation.
The left plot of Fig. 1 displays the probability of stopping for futility. In case of the original planning assumption (black line), the probability of ‘wrongly’ stopping for futility is given by 0.025 for n _{ group }=630. As a hazard ratio of 1.2 was considered as the largest effect under the alternative hypothesis for which stopping the study for futility would still be considered as correct, the probability of ‘correctly’ stopping for futility is given by 0.117. Decreasing the sample size or decreasing the hazard ratio (which corresponds to a larger deviation from the original planning assumptions) results in a monotone increase of the probability of stopping for futility. Note that for decreasing probability of ‘wrongly’ stopping for futility the probability of ‘correctly’ stopping for futility also decreases.
For Approach 2, the corresponding plots look similar, but additionally the noninferiority margin of the main component mainly influences the required sample size. For Approach 3, the displayed curves are much closer to each other, which means that on the one hand the loss in power is less prominent but on the other hand the probabilities of ‘correctly’ stopping for futility are also smaller. Figures for Approach 2 and Approach 3 are provided as Supplementary Material (see Additional files 3 and 4).
Influence of the choice of futility boundaries on power and probability of stopping  Admissible and optimal futility boundaries
Approach 3 shows the largest futility boundaries as the efficacy proof is based only on the composite endpoint and the overall power loss is thus most vulnerable to decreasing futility boundaries. Comparing the β γoptimal futility boundaries of the first two approaches, Approach 2 shows a larger value for the composite endpoint but a smaller for the main component. This is due to the fact that basing the efficacy proof on a noninferiority test of the main component but on superiority for the composite assigns a higher impact to the composite endpoint. Consequently, the overall power loss is more vulnerable to a decreasing futility boundary in the composite endpoint. For the specific clinical trial situation at hand, all approaches show medium and similar ‘correctly’ stopping for futility rates but Approach 1 by far requires the largest sample size. Approach 2 clearly provides more gain in information compared to Approach 3 as both endpoints are included in the efficacy claim. For this reason, Approach 2 should be recommended here.
Discussion
By construction, the provided method of choosing futility boundaries maximizes the probability of detecting small or opposite effects while limiting the power loss and the probability of stopping the study ‘wrongly’. The simulation results provided in “Simulation results” section clearly demonstrate the benefit of using such ‘optimal’ futility boundaries in the considered scenarios, especially compared to a futility bound of α _{ f }=0.5 which is commonly applied in practice. However, these criteria are generally applicable to any type of group sequential designs. In order to determine the most suitable twostage design out of the three approaches presented in “Investigated group sequential designs with futility stop based on two endpoints” section, we recommend to compare the performance properties of the designs under β γoptimal futility boundaries, which usually differentiate between the designs. The design is then chosen by considering the specific study situation, the gain in information provided by the design, and the probability of ‘correctly’ stopping for futility under the β γoptimal futility boundaries.
In this work we investigated binding futility boundaries in order to quantify the performance properties of the corresponding group sequential designs. However, the resulting performance properties of a binding rule can also be used to approximate the performance properties of more liberal nonbinding rules as some deviations from the binding rule will not importantly influence the performance.
The performance investigations and calculations of the β γoptimal futility boundaries in “Clinical trial example and simulation design” section were done within the context of a composite endpoint and one relevant main component which refers to a situation of two correlated endpoints. We considered only deviations from the original planning assumptions in the main component, which is motivated by the fact that every event in the main component corresponds to an event in the composite endpoint. Similar investigations could be made for deviations in the other endpoint or simultaneous deviations in both endpoints. Generally, our proposed designs could equivalently be applied to endpoints with other, potentially differing, scale levels. In particular, it would also be of interest to consider the performance properties of our designs for uncorrelated endpoints. A systematic investigation of all possible sequential designs and endpoint settings within the current work was not feasible. Therefore, we encourage to perform further simulations in order to determine a suitable twostage design for the specific situation at hand. Further, in this work we chose the Pocock approach for the local significance levels, which facilitates stopping the study early at the interim analysis. Allowing decision criteria that are more conservative in stopping at the interim analysis as, for example, proposed by O’Brien and Fleming [2] will improve the overall power.
The general principles of optimal futility boundaries might be transferred similarly to groupsequential designs with more than two stages or patientwise interim looks as long as the number of interim looks is fixed in advance. The implementation of our ideas in these situations will be the task of future work.
Conclusion
In this paper, we presented general optimality criteria for the choice of suitable futility boundaries which maximize the probability of detecting small or opposite treatment effect while limiting the power loss and the probability of stopping the study ‘wrongly’. We illustrated the criteria on three different group sequential designs including two endpoints, which are motivated by the fact that in many clinical trial applications it is not sufficient to consider only one primary endpoint in order to adequately describe the efficacy of a new treatment. As the properties of futility boundaries are often not considered in practice and unfavorably chosen futility boundaries may have serious consequences with respect to the performance of the study design, we recommend assessing the impact of these boundaries according to the proposed admissibility and optimality criteria.
Notes
Acknowledgements
We acknowledge financial support by Deutsche Forschungsgemeinschaft and RuprechtKarlsUniversität Heidelberg within the funding programme Open Access Publishing.
Funding
Not applicable.
Availability of data and materials
Not applicable.
Authors’ contributions
SS developed the methods, conducted the underlying simulations and exemplary applications and was the major contributor in writing the manuscript. MK and GR critically reviewed, discussed and adapted the methodology. All authors read and approved the final manuscript.
Ethics approval and consent to participate
For this type of study formal consent is not required.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary material
References
 1.Pocock S. Group sequential methods in the design and analysis of clinical trials. Biometrika. 1977; 64(2):191–9.CrossRefGoogle Scholar
 2.O’Brien P, Fleming T. A multiple testing procedure for clinical trials. Biometrics. 1979; 35(3):549–56.CrossRefPubMedGoogle Scholar
 3.Jennison C, Turnbull B. Group Sequential Methods with Applications to Clinical Trials. Boca Raton: CRC Press; 1999.CrossRefGoogle Scholar
 4.Snapinn S, Chen M, Jiang Q, Koutsoukos T. Assessment of futility in clinical trials. Pharm Stat. 2006; 5(4):273–81.CrossRefPubMedGoogle Scholar
 5.Lan K, Simon R, Halperin M. Stochastically curtailed tests in long–term clinical trials. Seq Anal. 1982; 1(3):207–19.Google Scholar
 6.Lachin J. A review of methods for futility stopping based on conditional power. Stat Med. 2005; 24(18):2747–64.CrossRefPubMedGoogle Scholar
 7.Spiegelhalter D, Freedman L, Parmar M. Bayesian approaches to randomized trials. J R Stat Soc Ser A Stat Soc. 1994; 157:357–416.CrossRefGoogle Scholar
 8.Heitjan D. Bayesian interim analysis of phase ii cancer clinical trials. Stat Med. 1997; 16(16):1791–802.CrossRefPubMedGoogle Scholar
 9.Emerson S, Kittelson J, Gillen D. On the use of stochastic curtailment in group sequential clinical trials. 2005. UW Biostatistics Working Paper Series. Working Paper 243. (See http://biostats.bepress.com/uwbiostat/paper243/).
 10.He P, Lai T, Liao O. Futility stopping in clinical trials. Stat. Interface. 2012; 5:415–23.CrossRefGoogle Scholar
 11.Pampallona S, Tsiatis A. Group sequential designs for onesided and twosided hypothesis testing with provision for early stopping in favor of the null hypothesis. J Stat Plan Infer. 1994; 42(1):19–35.CrossRefGoogle Scholar
 12.Pampallona S, Tsiatis A, Kim K. Spending functions for the type i and type ii error probabilities of group sequential tests. Technical report, Technical Report, Dept. of Biostatistics, Harvard School of Public Health, Boston. 1995.Google Scholar
 13.Asakura K, Hamasaki T, Sugimoto T, Hayashi K, Evans S, Sozu T. Sample size determination in groupsequential clinical trials with two coprimary endpoints. Stat Med. 2014; 33(17):2897–913.CrossRefPubMedPubMedCentralGoogle Scholar
 14.Cheng Y, Ray S, Chang M, Menon S. Statistical monitoring of clinical trials with multiple coprimary endpoints using multivariate bvalue. Stat Biopharm Res. 2014; 6(3):241–50.CrossRefGoogle Scholar
 15.Hamasaki T, Asakura K, Evans S, Sugimoto T, Sozu T. Groupsequential strategies in clinical trials with multiple coprimary outcomes. Stat Biopharm Res. 2015; 7(1):36–54.CrossRefPubMedPubMedCentralGoogle Scholar
 16.Kosorok M, Yuanjun S, DeMets D. Design and analysis of group sequential clinical trials with multiple primary endpoints. Biometrics. 2004; 60(1):134–45.CrossRefPubMedGoogle Scholar
 17.Tamhane A, Mehta C, Liu L. Testing a primary and a secondary endpoint in a group sequential design. Biometrics. 2010; 66(4):1174–84.CrossRefPubMedPubMedCentralGoogle Scholar
 18.Glimm E, Maurer W, Bretz F. Hierarchical testing of multiple endpoints in groupsequential trials. Stat Med. 2010; 29(2):219–28.PubMedGoogle Scholar
 19.Cook R, Farewell V. Guidelines for monitoring efficacy and toxicity responses in clinical trials. Biometrics. 1994; 50(4):1146–52.CrossRefPubMedGoogle Scholar
 20.Jennison C, Turnbull B. Group sequential tests for bivariate response: interim analyses of clinical trials with both efficacy and safety endpoints. Biometrics. 1993; 49:741–52.CrossRefPubMedGoogle Scholar
 21.Chow S, Wang H, Shao J. Sample Size Calculations in Clinical Research. Boca Raton: CRC Press; 2007.Google Scholar
 22.Wassmer G. Statistische Testverfahren Für Gruppensequentielle und Adaptive Pläne in Klinischen Studien: Theoretische Konzepte und Deren Praktische Umsetzung mit SAS. Köln: Mönch; 1999.Google Scholar
 23.Lan K, DeMets D. Discrete sequential boundaries for clinical trials. Biometrika. 1983; 70(3):659–63.CrossRefGoogle Scholar
 24.Kim K, DeMets D. Design and analysis of group sequential tests based on the type i error spending rate function. Biometrika. 1987; 74(1):149–54.CrossRefGoogle Scholar
 25.Hwang I, Shih W, De Cani J. Group sequential designs using a family of type i error probability spending functions. Stat Med. 1990; 9(12):1439–45.CrossRefPubMedGoogle Scholar
 26.Bretz F, Koenig F, Brannath W, Glimm E, Posch M. Adaptive designs for confirmatory clinical trials. Stat Med. 2009; 28(8):1181–217.CrossRefPubMedGoogle Scholar
 27.Liu G, Zhu G, Cui L. Evaluating the adaptive performance of flexible sample size designs with treatment difference in an interval. Stat Med. 2008; 27(4):584–96.CrossRefPubMedGoogle Scholar
 28.Kieser M, Rauch G. Twostage designs for crossover bioequivalence trials. Stat Med. 2015; 34(16):2403–16.CrossRefPubMedGoogle Scholar
 29.Brenner B, the RENAAL study Investigators. Effects of losartan on renal and cardiovascular outcomes in patients with type 2 diabetes an nephropathy. N Engl J Med. 2001; 345:861–9.CrossRefPubMedGoogle Scholar
 30.Brenner B, Cooper M, de Zeeuw D, Grunfeld J, Keane W, Kurokawa K, McGill J, Mitch W, Parving H, Remuzzi G, et al. The losartan renal protection study–rationale, study design and baseline characteristics of renaal (reduction of endpoints in niddm with the angiotensin ii antagonist losartan). J ReninAngiotensinAldosterone Syst. 2000; 1(4):328–35.CrossRefPubMedGoogle Scholar
 31.Koller M, Raatz H, Steyerberg E, Wolbers M. Competing risks and the clinical community: irrelevance or ignorance?Stat Med. 2012; 31(11–12):1089–97.CrossRefPubMedGoogle Scholar
 32.ICH. Statistical principles for clinical trials  e9. Technical report. 1998.Google Scholar
 33.CPMP. Points to consider on multiplicity issues in clinical trials. Technical report. 2002.Google Scholar
 34.IQWIG. General Methods, Version 4.2. Technical report. 2015.Google Scholar
Copyright information
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.