Adaptive Dose Response Studies Brenda Gaydos, Eli Lilly; Michael Krams, Wyeth; Inna Perevozskaya, Merck, Frank Bretz, Novartis; Qing Liu, Johnson&Johnson, Paul Gallo, Novartis; Don Berry, MD Anderson Cancer Center; Christy Chuang-Stein, Pfizer; José Pinheiro, Novartis; Alun Bedding, Eli Lilly Abstract. Insufficient exploration of the dose response is a shortcoming of clinical drug development, and failure to characterize dosing early is often cited as a key contributor to the high late-stage attrition rate currently faced by the industry. Adaptive methods, for example, make it feasible to design a proof-of-concept study as an adaptive dose response trial. Efficient learning about the dose response earlier in development will ultimately reduce overall costs and provide better information on dose in the filing package. This paper presents the PhRMA working group’s main recommendations regarding adaptive dose response studies. As background, traditional fixed and adaptive dose response designs are briefly reviewed. Information on developing a Bayesian adaptive dose design and some monitoring and processing issues are also discussed. 1. Introduction Insufficient exploration of the dose response is often a key shortcoming of clinical drug development. Initial proof-of-concept (PoC) studies often rely on testing just one dose level (e.g. the maximum tolerated dose) without much information on which to base the decision, assuming “more is better” and hoping the “right” dose was selected. Adaptive designs (defined in [1]) offer efficient ways to learn about the dose response and guide decision making on which dose(s) to select for further development, or whether to discontinue a program. It is both feasible and advantageous to design a PoC study as an adaptive dose response trial. The continuation of a dose response trial into a confirmatory stage in a seamless design is a further opportunity to increase information on the right dose(s) earlier in development and reduce the total clinical development timeline; see [2]. We envisage that efficient learning about the dose response earlier in development will reduce overall costs and provide better information on dose in the filing package. Our goals for the development of adaptive dose response designs are to facilitate: more feasible approaches for gaining information on the dose response profile earlier in development; increasing the probability of technical success by taking the correct choice of dose(s) forward into phase III, and thereby, reducing the late-stage attrition rate; reducing the clinical development timeline by stopping at the earliest possible timepoint either for futility or efficacy. This paper primarily focuses on phase Ib and II study designs, using endpoints that support filing or are predictive of the filing endpoint. For example, a biomarker might be used as the primary endpoint in a small PoC study, or to adapt a trial where the primary endpoint is not immediately observable, or be used more indirectly to improve a predictive model which is driving dose selection. We do not discuss here exposureresponse modelling; however, we want to acknowledge its importance and note that many of the adaptive design principles discussed directly apply. Exposure-response models are crucial to understanding activity at the patient level. Knowledge of exposureresponse increases our understanding of dose response, and these models are useful in developing adaptive dose response designs. Section 2 presents our key recommendations. In Section 3 we briefly review traditional fixed dose response designs. In Sections 4 and 5 we review adaptive dose response designs spanning phases I-III, and provide information on developing Bayesian adaptive dose designs in Section 6. We discuss monitoring issues and process in Section 7, and introduce a relatively recent class of dose response designs (Rolling Dose) in Section 8. 2. Recommendations In general, we recommend the following: Consider adaptive dose response designs in exploratory development. Both FDA and EMEA are supportive and encourage the deployment of appropriately planned adaptive designs in this segment of the critical path. Whether or not an adaptive design is ultimately deployed should be the result of a critical feasibility debate (see [3]), but if adaptive designs are not considered as an option, the debate would not even occur. Consider adaptive dose response designs to establish proof-of-concept. An estimate of the dose response profile is more informative than an estimate of the response at a single dose. The former is not frequently done, however. One reason for this is the perception that the dose response cannot be estimated sufficiently well with the limited budgets often available for very early go/no-go PoC studies. Our objective is to show that adaptive designs could be viewed as an enabler merging “PoC” and “dose response” into a single research question. Whenever possible use an approach that incorporates a model for the dose response. The intention is to make the best use of information that accrues in the trial. Potential applications include longitudinal models where an early readout of a biomarker predictive of final outcome is used to assist in decision making. This may lead to faster learning, and allow earlier futility decisions based on the 2 biomarker readout. However, it is critical to assess the model assumptions before deployment. Consider seamless approaches to improve the efficiency of learning. Logistic advantages include reduction of downtimes, and synergistic effects in setting up study sites and other infrastructure. Integrating traditional phases into a single experiment may enrich the information value derived from such studies. One example of benefit could be an ongoing longer-term exploration of the relationship between biomarkers and clinical endpoints. Define the dose assignment mechanism prospectively and fully evaluate its operational characteristics through simulation prior to initiating the study. We strongly recommend building up experiences with adaptive designs that fully characterize the decision rule in advance. This will help to ensure the validity of study conclusions as well as facilitate ethical review and implementation. Health authorities are rightly concerned that the concept of adaptive designs could be misused for ad-hoc changes to remedy poor planning. A detailed layout of possible decisions a priori and their potential impact on study conduct and conclusions will help alleviate this concern. Stop the trial at the earliest time point when there is enough information to make the decision. Typically, a maximum sample size will need to be identified. The design could include planning for the maximum sample and stopping early, or planning to extend the trial until sufficient certainty (precision) has been achieved up to some maximum sample. The later case allows for new information to be considered in setting the maximum. In either case, we recommend routinely incorporating futility analyses into early dose response studies due to the low probability of success for novel candidates. A committee must monitor the study on an ongoing basis to verify that the performance of the design is as expected. The objectives of monitoring include ensuring that errors in programming are not occurring, and that no adverse responses are occurring that may affect the dose assignment but were not accounted for in the adaptive algorithm. For early blinded studies, this could be a sponsor-internal committee, as long as its members are adequately firewalled from the project team and other parties involved in the conduct of the study [3,4]. Engage the committee early in scenario simulations prior to protocol approval. The responsibilities of the committee would include overseeing performance of the treatment allocation algorithm. This requires insight into both the clinical aspects and the statistical methodology and operational characteristics of the algorithm. 3 3. Leverage the information from disease state and exposure-response models to design studies. These models should be instrumental in developing the adaptive mechanism and in assessing the performance of the algorithm under a variety of scenarios of truth, especially for first-in-patient dose response studies. For Bayesian applications, this information can also be incorporated into the definition of the prior distributions for the dose-response model parameters of interest. Traditional methods to explore the dose response Fixed-dose parallel-group designs are typically driven by the desire to understand the average population response at a dose and the shape of the population dose response curve for both efficacy and safety. Population average dose response information is useful in determining the starting dose, and for adjusting dose for a population. In addition, the shape of the population average dose response curve can help identify the smallest dose with a discernible beneficial effect, or a maximum dose beyond which no further benefits are seen. These considerations have led to study designs that treat doses equally. Such designs attempt to estimate the mean response at different doses with nearly identical precision. There are two possible downsides to this approach. One is the potential to allocate a fair number of patients to several non-informative doses, resulting in sub-optimal efficiency in learning about the dose response. The other is that sample size considerations often limit the number of doses feasible to explore. Therefore, if little is known about the true dose response a priori, there can be a high likelihood that the study will not be sufficiently informative. Historically, four study design types have been used to explore dose response with some regularity. They are the parallel group design, cross-over design, forced titration design, and optional titration design. The latter three are primarily aimed at learning about individual dose-response. Despite the differences in the primary objectives among these four designs, they can be useful in describing the dose response relationship if properly planned, executed and analyzed. These designs are covered in ICH E4 [5]. Other recommended references include [6-8]. 4. Adaptive dose response methods for early exploratory studies This section provides a brief overview of traditional phase I designs with respect to estimation of the Maximum Tolerated Dose (MTD), and discusses novel adaptive design approaches aimed at improving the relatively poor performance of traditional designs (for more detail see [9,10]). While the majority of these methods originated from oncology 4 and were designed to target the MTD only, there is methodological overlap with dose response methods discussed later. The applicability of these methods can be generalized from MTD determination to learning about the dose response profile for any defined response (e.g. tolerability, safety or efficacy measure), and some of these approaches are also discussed in this section. That being said, more work generally needs to be done outside the area of cancer research since most of the literature on such methods is restricted to the oncology context. A monotonic relationship between dose and toxic responses is typically assumed in phase I trials. There are two major philosophies regarding MTD definition: 1. The dose which, if exceeded, would put patients at unacceptable risk of toxicity. Such a definition is vague from the statistician’s point of view since “unacceptable risk” may not be defined quantitatively. 2. Specifying unacceptable risk of toxicity as a probability; therefore MTD is an unknown parameter to be estimated corresponding to the specified probability. These two definitions lead to two different approaches for designing phase I clinical trials: 1. Conventional up-and-down designs (such as 3+3 design for cancer) 2. Model-based designs (MTD is a quantile to be estimated), e.g.: Random Walk Rule Bayesian methods (continual reassessment method, escalation with overdose control, decision-theoretic approaches, Bayesian D-optimal design) Conventional 3+3 designs employ an ad-hoc approach to efficiently screen dose levels and identify the MTD; no estimation in a traditional sense is involved. Patients are treated in groups of three starting with the initial dose. “Toxicity” is defined as a binary event. The algorithm iterates moving the dose up or down depending on the number of toxicities observed, and MTD is defined as the highest dose studied with less than, say, 1/3 toxicities. These designs rely on a response rate within a range of 20%-30% to be applicable and are not well suited for efficacy studies; these designs often provide poor estimates of MTD, i.e. the probability of stopping at an incorrect dose level is higher than generally perceived [11]. Random-Walk-Rule Designs (RWR) (“biased coin designs”) are non-parametric modelbased approaches to MTD estimation, i.e. the MTD is treated as a quantile of a dose 5 response distribution, but there is no underlying parametric family. RWR designs are a generalization of conventional up-and–down methods, but, unlike those methods, they provide a unified approach targeting any quantile of interest. As in conventional designs, patients are treated sequentially and dose escalation occurs when no toxicities are observed. However, the dose escalation rule is different: instead of applying a deterministic rule, a ‘biased coin’ is flipped after observing each response; the algorithm escalates to the next dose with probability p, where p depends on the targeted level of response. RWR designs have advantages of being non-parametric, having a workable finite distribution theory, and being simple and intuitive to implement. They generate a cluster of doses around the quantile of interest and also posses some optimality properties [12]. The Continual-Reassessment Method (CRM) originated as a Bayesian method for phase I cancer trials of cytotoxic agents (Section 6 provides more details of Bayesian methods). For a predefined set of doses and a binary response, CRM estimates MTD as the dose level that yields a particular target proportion of responses (e.g., TD20). It assumes a particular model (such as the logistic function) and the assignment of doses converges to the MTD. The method operates by updating the posterior distribution of the model parameters after observing each response. The dose for the next patient is chosen as the one with (Bayesian current) probability of toxicity closest to the target response level. A potential downside to the original CRM method [13] is that it can escalate doses too quickly. Several modified CRM approaches which mitigate risk of toxicity exposure above acceptable limits have been developed and implemented [13-17]. An example of non-oncology CRM application can be found in [17]. Other Bayesian Designs (1) Escalation With Overdose Control (see [18]). This is similar to CRM, but utilizes a more flexible family of underlying dose response models and controls the predicted probability of next assignment exceeding MTD. The latter is particularly important in oncology trials due to the severity of side effects. (2) Designs based on Bayesian Decision Theory. This is a wide class of methods applying Bayesian Decision Theory tools to achieve various goals, such as shortening trials, reducing sample size, maximizing information, and reducing cost. These designs use gain functions (based on the desired goal) that are sequentially updated after each patient’s (cohort’s) response. The next dose assignment is determined by maximising the gain function. Whitehead and Brunier [19] describe a design minimizing posterior variance of the MTD estimator. This methodology is extended in [9,20] to looking at pharmacokinetic data with two gain functions: the first chooses the highest dose possible subject to a safety constraint, and the second uses D-optimality in optimizing the parameters 6 of the model. The methodology is extended further to evaluate efficacy and toxicity dose response relationships in [21,22]. Section 6 provides an example of application in a phase II trial. (3) Bayesian D-optimal Designs (see [23]). Similar to decision-theoretic approaches [19], these designs are concerned with the efficiency of estimation. Originally developed for a two parameter logistic family for the underlying dose response model, they can be used for a variety of models (i.e. wide class of parametric families). These designs introduce a formal optimality criterion (D-optimality) minimizing the determinant of the variance-covariance matrix of the model parameter estimates. After a small internal “pilot study” designed to obtain an initial idea about the distribution of model parameters, a sequential procedure begins to allocate incoming patients to doses yielding maximum information about the dose-response experiment (measured as the determinant of the inverse of the variance-covariance matrix). The optimal allocation changes with each subsequent posterior distribution update (after each patient or cohort). Additional constraints can prevent assigning patients to highly toxic doses. These designs target the overall dose response curve rather than the MTD only; therefore, any level of response can be estimated. Simultaneous assessment of efficacy and toxicity: Penalized D-optimal Designs (see [24]). These non-Bayesian designs accomplish “learning” from the information accumulated during the trial by updating the model parameters using the likelihood function after each patient’s response (instead of updating a posterior distribution). Similar to the Bayesian D-optimal design, the D-optimality criterion is applied at each step of the sequential trial to maximize the expected increment of information about efficacy and toxicity dose response. The optimization is done subject to a constraint which can reflect ethical concerns, cost, sample size, etc. This flexibility in the constraint and the underlying bivariate model make this approach particularly useful in early phase trials, as the scope extends beyond MTD estimation, and allows a number of questions involving efficacy and safety dose response relationships to be addressed simultaneously. 5. Adaptive frequentist approaches for late stage exploratory development This section focuses on designs more typically applicable to phase II studies. Several methods are available for adaptive dose selection in multistage clinical trial designs which strongly control the overall type I error rate of incorrectly rejecting any null hypothesis of no dose level effect. In the context of adaptive dose response trials, multiplicity concerns typically arise due to (i) comparison of several doses with a control and (ii) multiple interim looks at the data for decision making. Performing each hypothesis test at the nominal level α intended for the whole trial would inflate the 7 overall type I error; the significance levels of the individual tests therefore have to be adjusted appropriately. For more details on multiple testing see [25,26]. Classical Group Sequential Designs are a type of adaptive design where the planned sample size or amount of information may be reduced based on outcome data if the trial or an arm of the trial is stopped early. In classical group sequential trials, test statistics produced at interim analyses are compared to upper and/or lower stopping limits that have been determined to jointly ensure the desired operating characteristics, and in particular to control the overall type I error rate. Different methods have been proposed to determine stopping boundaries [27-30]. Further details on group sequential methods are given in [31]. Stallard and Todd [32] extended classical group sequential procedures to multi-armed clinical trials in which patients are initially randomized between several dose groups and a control. During the course of the trial, the best experimental treatment is identified based on the maximum standardized test statistic at interim. Adaptive Designs have been developed that offer far more flexibility for adaptation within the multistage framework. These methods offer a high level of flexibility for decision making during the trial, such as increasing sample size based on the observed effect, modifying the patient population, or adapting doses; see [33,34] for reviews. They require little pre-specification of decision rules before the beginning of a trial, and therefore, the total information available at each interim time point can be used in designing or adjusting the next stage. However, some of these approaches may lose efficiency relative to classical group sequential designs [35,36]. The various approaches differ in how evidence is combined across stages, and for all approaches this has to be pre-specified. When early stopping is allowed, the decision to continue the trial to the next stage or to stop for futility is based on unblinded data collected up to the interim time point. If the trial continues to the final stage, the final analysis combines the information across all stages [37]. Lehmacher and Wassmer [38] established a connection between adaptive and classical group sequential designs. They suggested combining evidence from different stages using weighted sums of the inverse normal distribution function calculated at the observed p-values. A general framework for adaptively changing hypotheses at interim was presented in [39]. This method is based on the closure principle [40] for controlling the overall type I error rate. The closure principle considers all intersection hypotheses constructed from the initial hypotheses set of all pairwise comparisons with a control; a dose is declared significant if all intersection hypotheses related to the dose are also rejected. 8 Further Methods. A different approach is to use a standard single-stage multiple testing procedures to account for the multiplicity arising from multiple comparisons versus the control. The conduct of the study remains the same as in the previous methods, with the understanding that doses may be dropped for statistical or non-statistical reasons during the course of the trial. Irrespective of how many treatment arms continue to the end, the final analysis uses standard multiplicity adjustments such as Dunnett [41] or Bonferroni. As shown in [42], such approaches are often equally as powerful as adaptive design test procedures, while being simpler to perform and allowing the derivation of simultaneous confidence intervals. In late-stage clinical trials for which the clinical endpoint of interest is only slowly available relative to the patient accrual rate, adaptation based on the clinical endpoint becomes impractical, if not impossible. To circumvent this problem, Liu and Pledger [43] proposed a two-stage design combining phase II and III trials into a single study in which an early efficacy endpoint during the phase II component is used in the interim analysis as a surrogate for the clinical endpoint of interest. The methodology allows dropping of low doses for lack of efficacy, or high doses for safety or tolerability concerns. Doses not dropped at interim continue into the phase III component of the trial. The final analysis uses combined pairwise test statistics of the phase II stage, and adaptive trend test statistics of the phase III stage, where scales for trend are derived from a sigmoid Emax modeling of the phase II data. Incorporation of the dose-response information in the adaptive trend statistic results in a procedure that is both powerful and robust. Another example of combining phase II and III using a surrogate is shown in [44]. 6. Developing a Bayesian adaptive dose design The prime advantage of the Bayesian approach [45-47] is its focus on updating information as it accrues [48]. This makes it ideal for sequential experimentation, including clinical trials. The Bayesian approach also enables calculation of predictive probabilities of future results for any particular design, which allows comparison of designs on the basis of probabilities of their consequences. It also allows for addressing the increment in information about a dose response curve depending on the dose assigned to the next patient. Although control of the type I error rate is not a relevant property of a Bayesian design, simulations can be used to design a Bayesian adaptive trial that also maintains frequentist properties. A potential downside to the Bayesian approach is the computational complexity coupled with the absence of commercial software packages to assist with study design and analysis. A critical step in the Bayesian approach is positing a model for the data to be observed in the trial. The model can be flexible and have many parameters. It need not be tractable mathematically, but must have parameters whose distributions can be updated from the accruing trial data. An additional ingredient in the Bayesian approach is a prior 9 distribution for unknown parameters. This distribution can be “non-informative” or can include historical information that may be available about the natural history of the disease or the experimental agent or related agents. Efficiency is gained by incorporation objective information in the prior, appropriately. Simulations should be performed to understand the potential impact of the prior on the operating characteristics of the design and the potential impact on the resulting posterior distribution. Bayesian approaches are standard in phase I cancer trials of cytotoxic agents (see Section 4) and are increasingly used in phase II dose response trials. Again the advantage lies in allowing updating of information about unknown aspects of the dose response curve. For example, Berry et al. [49] describe the design of a trial to investigate a neuroprotective agent for stroke and show how to proceed sequentially, analyzing the data in the trial as it accumulates. A flexible model was used (normal dynamic linear model, [50]) to estimate the dose response curve. This model does not restrict the shape of the curve, and allows for non-monotonicity if indicated by the data. In that example there are two trial stages, first dose-ranging and then confirmatory (if the latter is warranted). The dose-ranging stage continues until a decision is made that the drug is not sufficiently effective to pursue future development, or that the optimal dose for the confirmatory stage has been sufficiently well identified. Switches to the confirmatory stage can be made seamlessly [2]. Each entering patient is assigned the dose (one of 16 including placebo) that maximizes information about the dose response relationship, given the results observed so far. In particular, patients are not assigned doses in regions where evidence suggests that the dose response curve is flat. Note that the principle applies to whatever utility we choose to optimize our learning against: the dose response: the ED95 or ED50, a gain function, etc. In the dose-ranging stage, neither the number of patients assigned to any particular dose nor the total number of patients assigned in this stage are fixed in advance. The doseranging sample size can be large when the drug has marginal benefit, when the dose response curve is gently sloping, or when the standard deviation of the responses is moderately large. It tends to be small if the drug has substantial benefit or has no benefit, if the dose response curve rises over a narrow range of doses, or if the standard deviation of the responses is small. (In addition, and somewhat non-intuitively, the dose-ranging stage is small if the standard deviation of responses turns out to be very large. The reason is that a sufficiently large standard deviation implies that a very large sample size is required to show a beneficial effect. The required sample size may be so large that it makes it impossible to study the drug and so the trial stops in the dose-ranging phase before substantial resources go down the drain.) The stroke trial considered in [49,51] is important in showing how long-term endpoints can be handled in the Bayesian approach. Again, the answer is in modeling. The ultimate endpoint is the improvement in the stroke scale from baseline to 13 weeks. If 10 the accrual rate is sufficiently large then the benefit of adaptive assignment is limited by delays in obtaining endpoint information. However, each patient’s stroke scale was assessed weekly between baseline and week 13. Within-patient measurements are correlated, with correlations greater if they are closer together in time. A longitudinal model enabled Bayesian predictions of the ultimate patient-specific endpoint based on the current patient-specific information. The circumstances of this trial are typical of many types of trials. The adaptive nature of the stroke trial would be less efficient if early endpoints were not exploited. Many diseases and conditions are characterized by the availability of such early endpoints: information about how a patient is doing (local control of the disease, biomarkers, etc.) before reaching the primary endpoint. Bayesian modeling is an ideal way to incorporate such information. In comparison to a standard fixed design, a Bayesian adaptive dose response design is more effective in identifying “the right dose.” Additionally, it usually identifies this dose with a smaller sample size than when using fixed-dose assignments. Another advantage is that many more doses can be considered - even though some doses will be little used in the trial, and some might never be used. Also, the Bayesian adaptive approach is efficient at identifying drugs that have little or no benefit, as in the case of the stroke trial [51]. 7. Monitoring issues and processes in adaptive dose response trials Adaptive dose response trials can of course occur throughout most stages of clinical development, from phase I exploratory trials to confirmatory studies (and may in fact combine stages or minimize traditional interpretations of development stages). All adaptive trials which utilize unblinded data for within-trial decision making, or in which specific changes made can convey to observers some information about interim results, raise issues and potential concerns about how the adaptive aspects of the trials should best be implemented in the interest of enhancing the credibility of trial conclusions. As with most types of interim data monitoring, it is beneficial with regard to minimizing potential for bias in the trial and enhancing the credibility of its conclusions for trial participants and sponsor representatives not to have access to unblinded interim results, beyond the extent necessary to meet the needs of the trial and the clinical program [52]. Thus, it is preferable to have a separate body without other direct trial responsibilities review the interim results and make recommendations for the adaptations. Operational models for monitoring ongoing trials are discussed in [4]. The nature of recommended adaptations may convey some information to trial participants. For example, if the methodology being used involves adaptive randomization, as discussed earlier in this paper, it may be unavoidable for the change in 11 the randomization scheme to provide some limited comparative information regarding how some dose groups are performing relative to others. Discontinuation of dose groups in an adaptive dose response trial, or the addition of new doses, may convey some information about the shape of the dose response curve. However the knowledge conveyed about the magnitude of treatment effects and any resulting potential for bias would be expected to be quite minimal as long as the specific numerical information remains confidential, and should thus not be viewed as compromising the trial. If possible, some detail on the statistical methodology and thresholds for adaptations might be withheld from the trial protocol and documented elsewhere, to limit the extent of information which might be inferred by observers, as long as this information were not considered necessary for review boards, investigators, or patients to have knowledge of in order to be sufficiently informed for their participation in the trial. For trials which are planned to be confirmatory, and even for non-confirmatory trials which could potentially play a supportive role in a regulatory submission, it is important to try to adhere closely to current monitoring and confidentiality conventions and to limit knowledge of interim results, in order to enhance the credibility and interpretability of the trial results. 8. Rolling Dose Studies The term Rolling Dose Studies (RDS) refers to a broad class of designs and methods that allow flexible, dynamic allocation of patients to dose levels as the trial progresses. RDS are not a distinct set of methods, and many of the approaches discussed in this paper could be considered to be within this class. In RDS, dose level arms can be started or discontinued during the trial. Unlike classical group sequential methods and multi-stage adaptive designs aimed at preserving the overall type I error rate, RDS are intended to maximize the learning of the underlying dose response relationship and/or the precision of estimated doses to achieve certain target effects (efficacy and safety). To achieve these goals, RDS typically rely more on modelling techniques than on hypothesis testing approaches. The Bayesian adaptive dose response stroke study discussed in Section 6 is an example of a RDS; however, not all RDS are Bayesian. Accruing unblinded data are utilized in RDS, often in combination with dose response models, to determine the current best allocation of patients to dose levels according to some optimality criterion (e.g., Doptimality, or minimum variance of target dose estimator) related to dose response or target dose estimation. Alternative RDS approaches differ in the criteria used to define the optimal allocation of patients to dose, the frequency of interim looks for recalculating allocation fractions, and stopping rules (success in estimating dose response and/or target doses, or futility). 12 PhRMA has constituted a working group on RDS as part of its Pharmaceutical Innovation Steering Committee (PISC) initiative. This group is evaluating and developing different RDS methods, and comparing them to more traditional non-adaptive dose finding approaches (e.g., ANOVA with multiplicity correction) via a comprehensive simulation study. The conclusions from the simulation and recommendations from the group on the practical use of RDS will be included in a white paper to be submitted for publication. 9. Conclusion We recommend routinely assessing the appropriateness of adaptive designs, including adaptive dose response designs, when developing plans for clinical programs. The development of commercial software packages could greatly facilitate this. Adaptive designs present an opportunity to efficiently gain more information about the dose response at any point in development; however, we recommend gaining this information early for maximum benefit. A good example of a PoC study in pain employing a Bayesian adaptive dose design early is presented in [53]. It seems reasonable to expect that obtaining more information on the dose response earlier in development could enable confidence in a more streamlined phase III clinical plan, reducing timelines and costs, yet increasing information at time of filing. We point out, however, that adaptive designs are not necessarily always better than fixed designs, and that there are many adaptive design possibilities to select from. Thorough early planning is required to ensure that benefits are achieved. The added scientific and operational complexity should be justified. Simulations are often needed, under realistic scenarios of truth, to assess how the design will perform and to compare performance across designs. The adaptive mechanism should be detailed enough to enable simulations prior to finalizing the protocol. Dose selection should not be a black box. In addition to streamlining implementation, clarity on the criteria of dose selection will facilitate ethical review. We cannot emphasize strongly enough the importance of planning. Adaptive dose response designs have built-in flexibility to address areas of uncertainty about the dose response. As with any trial design, protocol amendments may be required, but may be less likely to be needed than in a fixed design where learning is not built in. Emphasis needs to be made upfront on the questions to be addressed regarding the dose response, and the adaptive trial needs to be designed accordingly. For example, it might be desirable to address one or all of the following to different degrees: minimally effective dose; ED95; a particular effect level beyond that of a comparator treatment. With a Bayesian dose-adaptive approach, for example, the utility function can be designed so that the adaptive algorithm will prospectively weight future treatment assignments to optimize information around the key questions. 13 We recognize that there are limited examples of adaptive dose response applications published in the literature, and more practical experience is needed. We thank Andy Grieve for some additional references [54-57]. We welcome further exploration and publications in this area. References 1. Dragalin V. Adaptive designs: terminology and classification. Drug Inf J. 2006 (submitted) 2. Maca J, Bhattacharya S, Dragalin V, Gallo P, Krams M. Adaptive seamless phase II / III designs – Background, operational aspects, and examples. Drug Inf J. 2006 (submitted). 3. Quinlan JA, Krams M. Implementing adaptive designs: logistical and operational considerations. Drug Inf J. 2006 (submitted). 4. Gallo P. Confidentiality and trial integrity issues for adaptive designs. Drug Inf J. 2006 (submitted). 5. International Conference on Harmonisation Expert Working Group. Guideline for industry: Dose response information to support drug registration. Federal Register 1994;59(216):55972-55976. 6. Ting N, ed. Dose Finding in Drug Development. Springer;2006. 7. Ruberg SL. Dose response studies. I. Some design considerations. J Biopharm Stat. 1995;5(1):1-14. 8. Ruberg SL. Dose response studies. II. Analysis and interpretation. J Biopharm Stat. 1995;5(1):15-42. 9. Whitehead J, Zhou Y, Patterson S, Webber D, Francis S. Easy-to-implement Bayesian methods for dose-escalation studies in healthy volunteers. Biostatistics 2001;2:47-61. 10. Rosenberger WF, Haines LM. Competing designs for phase I clinical trials: a review. Stat Med. 2002;21:2757-2770. 11. Reiner E, Paoletti X, O’Quigley J. Operating characteristics of the standard phase I clinical trial design. Computational Statistics and Data Analysis 1999;30:303-315. 12. Durham SD, Flournoy N. Random walks for quantile estimation. In Gupta SS, Berger JO, ed. Statistical Decision Theory and Related Topics. New York: Springer;1994:467–476. 14 13. O’Quigley J, Pepe M, Fisher L. Continual reassessment method: a practical design for phase 1 clinical trials in cancer. Biometrics 1990;46:33–48. 14. Faries D. Practical modifications of the continual reassessment method for phase I cancer clinical trials. J Biopharm Stat. 1994;4:147-164. 15. Korn EL, Midthune D, Chen TT, Rubinstein LV, Christian MC, Simon RM. A comparison of two phase I designs. Stat Med. 1994;13:1799-1806. 16. Goodman SN, Zahurak ML, Piantadosi S. Some practical improvements in the continual reassessment method for phase I studies. Stat Med. 1995;14:1149-1161. 17. Dougherty TB, Porche VH, Thall PF. Maximum tolerated dose of Nalmefene in patients receiving epidural fentanyl and dilute bupivacaine for postoperative analgesia. Anesthesiology 2000;92(4):1010-1016. 18. Babb J, Rogatko A, Zacks S. Cancer phase I clinical trials: efficient dose escalation with overdose control. Stat Med. 1998;17:1103-1120. 19. Whitehead J, Brunier H. Bayesian decision procedures for dose determining experiments. Stat Med. 1995;14:885-893. 20. Patterson S, Jones B. Bioequivalence and Statistics in Clinical Pharmacology London: Chapman & Hall;2005. 21. Whitehead J, Zhou Y, Stevens J, Blakey G. An evaluation of a Bayesian method of dose escalation based on bivariate binary responses. J Biopharm Stat. 2004;14(4):969-983. 22. Braun T. The bivariate continual reassessment method: extending the CRM to phase I trials of two competing outcomes. Contr Clin Trials 2002;23:240-256. 23. Haines LM, Perevozskaya I, Rosenberger WF. Bayesian optimal designs for phase I clinical trials. Biometrics 2003;59:561-600. 24. Dragalin V, Fedorov V. Adaptive designs for dose-finding based on efficacy-toxicity response. Journal of Statistical Planning and Inference 2005;136:1800-1823. 25. Hochberg Y, Tamhane AC. Multiple Comparison Procedures. Wiley: New York;1987. 26. Hsu JC. Multiple Comparisons. London: Chapman and Hall;1996. 15 27. Lan KKG, DeMets DL. Discrete sequential boundaries for clinical trials. Biometrika 1983;70:659-663. 28. Pocock SJ. Group sequential methods in the design and analysis of clinical trials. Biometrika 1977;64:191-199. 29. O'Brien PC, Fleming TR. A multiple testing procedure for clinical trials. Biometrics 1979;35:549-556. 30. Whitehead J. The Design and Analysis of Sequential Clinical Trials. Revised 2nd Edition, Chichester: Wiley;1997. 31. Jennison C, Turnbull BW. Group Sequential Methods with Applications to Clinical Trials. London: Chapman and Hall;2000. 32. Stallard N, Todd S. Sequential designs for phase III clinical trials incorporating treatment selection. Stat Med. 2003;22:689-703. 33. Jennison C, Turnbull BW. Meta-analyses and adaptive group sequential designs in the clinical development process. J Biopharm Stat. 2005;15:537-558. 34. Bauer P, Brannath W. The advantages and disadvantages of adaptive designs for clinical trials. Drug Discovery Today 2004;9(8):351-357. 35. Tsiatis AA, Mehta C. On the inefficiency of the adaptive design for monitoring clinical trials. Biometrika 2003;90:367-378. 36. Brannath W, Bauer P, Posch M. On the efficiency of adaptive designs for flexible interim decisions in clinical trials. Journal of Statistical Planning and Inference 2006;136:1956-1961. 37. Bauer P, Köhne K. Evaluation of experiments with adaptive interim analyses. Biometrics 1994;50:1029-1041. 38. Lehmacher W, Wassmer G. Adaptive sample size calculations in group sequential trials. Biometrics 1999;55:1286-1290. 39. Hommel G. Adaptive modifications of hypotheses after an interim analysis. Biometrical J. 2001;43:581-589. 40. Marcus R, Peritz E, Gabriel KB. On closed testing procedures with special reference to ordered analysis of variance. Biometrika 1976;63:655-660. 16 41. Dunnett CW. A multiple comparison procedure for comparing several treatments with a control. J Am Stat Assoc. 1955;50:1096-1121. 42. Bretz F, Schmidli H, König F, Racine A, Maurer W. Confirmatory seamless phase II/III clinical trials with hypothesis selection at interim: General concepts. Biometrical J. 2006 (in press). 43. Liu Q, Pledger WG. Phase 2 and 3 combination designs to accelerate drug development. J Am Stat Assoc. 2005;100:493-502. 44. Todd S, Stallard N. A new clinical trial design combining phases II and III: sequential designs with treatment selection and a change of endpoint. Drug Inf J. 2005;39:109118. 45. Spiegelhalter DJ, Abrams KR, Myles JP. Bayesian Approaches to Clinical Trials and Health-Care Evaluation. Chichester: Wiley;2004. 46. Berry DA. Statistical Innovations in Cancer Research. In Holland J, Frei T ed. Cancer Medicine e.7. Ch 29. London: BC Decker;2005. 47. Berry DA. Bayesian clinical trials. Nature Reviews Drug Discovery 2006;5:27-36. 48. Berry DA. Statistics: A Bayesian Perspective. Belmont, CA: Duxbury Press;1996. 49. Berry DA, Müller P, Grieve AP, Smith M, Parke T, Blazek R, Mitchard N, Krams M. Adaptive Bayesian Designs for Dose-Ranging Drug Trials. In Gatsonis C, Carlin B, Carriquiry A ed. Case Studies in Bayesian Statistics V 99-181. New York: SpringerVerlag;2001. 50. West M, Harrison J. Bayesian Forecasting and Dynamic Models, 2nd edition. Springer Verlag;1997. 51. Krams M, Lees KR, Hacke W, Grieve AP, Orgogozo J-M, Ford GA. Acute stroke therapy by inhibition of neutrophils (ASTIN): An adaptive dose response study of UK-279,276 in acute ischemic stroke. Stroke 2003;34:2543-2548. 52. US Food and Drug Administration. Guidance for Clinical Trial Sponsors on the Establishment and Operation of Clinical Trial Data Monitoring Committees. Rockville MD: FDA;2006. (Available at http://www.fda.gov/cber/gdlns/clintrialdmc.htm) 53. Smith MK, Jones I, Morris MF, Grieve AP, Tan K. Implementation of a Bayesian adaptive design in proof of concept study. Pharmaceutical Statistics 2006;5:39-50. 17 54. Roon KI, Olesen J, Diener HC, Ellis P, Hettiarachchi J, Poole PH, Christianssen I, Kleinermans D, Kok JG, Ferrari MD. No acute antimigraine efficacy of CP-122,288, a highly potent inhibitor of neurogenic inflammation: Results of two randomized, double-blind, placebo controlled clinical trials. Annals of Neurology 2000;47(2):238241. 55. Farge D, Marolleau JP, Zohar S, Marjanovic Z, Cabane J, Mounier N, Hachulla E, Philippe P, Sibilia J, Rabian C, Chevret S, Gluckman E. Autologous bone marrow transplantation in the treatment of refractory systemic sclerosis: early results from a French multicentre phase I-II study. Br J Haematology 2002;119(3):726-739. 56. Camorcia M, Capogna G, Lyons G, Columb M. Epidural test dose with levobupivacaine and ropivacaine: determination of ED50 motor block after spinal administration. Br J Anaesthesia 2004;92(6):850-853. 57. Desfrere L, Zohar S, Morville P, Brunhes A, Chevret S, Pons G, Moriette G, Rey E, Treluyer JM. Dose-finding study of ibuprofen in patent ductus arteriosus using the continual reassessment method. Journal of Clinical Pharmacy & Therapeutics 2005;30(2):121-132. 18