Home | About | Journals | Submit | Contact Us | Français |

**|**Front Psychiatry**|**v.3; 2012**|**PMC3504325

Formats

Article sections

- Abstract
- Introduction
- Recent Developments in Dose-Finding
- Illustration
- Locating the Most Successful Dose
- Modeling Overall Success
- Dose Allocation
- Studies with Topiramate
- Conclusion
- Conflict of Interest Statement
- References

Authors

Related links

Front Psychiatry. 2012; 3: 100.

PMCID: PMC3504325

Edited by: Wendy J. Lynch, University of Virginia, USA

Reviewed by: Zachary A. Rodd, Indiana University School of Medicine, USA; Ludovico Abenavoli, University Magna Graecia, Italy

*Correspondence: Nolan A. Wages, Department of Public Health Sciences, University of Virginia, P.O. Box 800717, Charlottesville, VA 22908, USA. e-mail: nwages/at/virginia.edu

This article was submitted to Frontiers in Addictive Disorders and Behavioral Dyscontrol, a specialty of Frontiers in Psychiatry.

Received 2012 July 6; Accepted 2012 November 1.

Copyright © 2012 Wages, Liu, O’Quigley and Johnson.

This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in other forums, provided the original authors and source are credited and subject to any copyright notices concerning any third-party graphics etc.

This article has been cited by other articles in PMC.

In alcohol dependence studies, the treatment effect at different dose levels remains to be ascertained. Establishing this effect would aid us in identifying the best dose that has satisfactory efficacy while minimizing the rate of adverse events. We advocate the use of dose-finding methodology that has been successfully implemented in the cancer and HIV settings to identify the optimal dose in a cost-effective way. Specifically, we describe the continual reassessment method (CRM), an adaptive design proposed for cancer trials to reconcile the needs of dose-finding experiments with the ethical demands of established medical practice. We are applying adaptive designs for identifying the optimal dose of medications for the first time in the context of pharmacotherapy research in alcoholism. We provide an example of a topiramate trial as an illustration of how adaptive designs can be used to locate the optimal dose in alcohol treatment trials. It is believed that the introduction of adaptive design methods will enable the development of medications for the treatment of alcohol dependence to be accelerated.

Alcoholism and alcohol abuse compose a large, worldwide public health problem that is responsible for significant morbidity and mortality (Mokdad et al., 2004). To this point, the most common form of treatment for alcohol dependence in the USA has been group counseling and referral to community support groups. Three medications – disulfiram, naltrexone, and acamprosate – have been approved for the treatment of alcohol dependence although the use of these medications is limited. Further, there is no single medication that is effective in every case or every person. Craving is an instrumental component of alcohol dependence and can involve a desire for the reward provided by alcohol, the need for relief from tension, or an obsessive loss of control over one’s thoughts about alcohol; hence, the most promising and efficacious medications are those that interfere with the neurotransmitters involved in craving mechanisms (Addolorato et al., 2005). The development of new and more effective medications to treat alcoholism remains a high priority (Willenbring, 2007).

Many clinical trials have been used to evaluate the efficacy and safety of new medications to treat alcoholism. Most of them involve two arms: a treatment arm and a control arm. It is often of particular interest to clinicians, however, to determine the optimal dose from a range of doses. In this case, two-arm studies are insufficient. For instance, in the single-site topiramate study (Johnson et al., 2003), topiramate’s (or matching placebo’s) dose started at 25mg for week 1, with a 25-mg increment in weeks 2–4 and a 50-mg increment in weeks 5–8 (up to a total dose of 300mg). The topiramate dose of 300mg was maintained between weeks 8 and 12. A similar dose-escalating scheme was employed in the multi-site topiramate study (Johnson et al., 2007). These proof-of-concept trials established the overall topiramate treatment effect at improving drinking outcomes. However, the topiramate effect at different dose levels remains to be established so that we can identify the best dose that has the satisfactory efficacy while minimizing the rate of adverse events.

A possible solution to this problem lies in the use of an adaptive design made up of two parts. The goal of the first part would be to determine the most promising dose of topiramate and to optimize the number of patients treated at that dose level while including enough patients at neighboring doses to examine accurately the relationship. In other words, we want to locate the dose that provides the best chance for success from among a set of doses. In the second part of the design, the optimal dose found in the first stage would be compared with a placebo arm in a randomized study to establish the statistical significance of the treatment. This stage is imperative because it guards against the unlikely situation in which the optimal dose, although more efficacious than any other dose, is not more successful than placebo.

The motivation behind adaptive designs is to make use of the statistical advantages of a sequential design in combination with the ethical considerations of treating as many patients as possible at a dose believed to be the best, given prior knowledge, and accumulated data. For traditional dose-finding designs in cancer, aimed at controlling adverse events, the optimal dose is defined in correspondence to a tolerable level of toxicity, i.e., maximum tolerated dose (MTD). For designs whose aim is to identify the most successful dose (MSD), the optimal dose is the one that maximizes the overall success rate, considering both treatment benefit (efficacy) and failure. Here, failure would be defined as either unacceptable toxicity or dropout as a result of not being able to tolerate the treatment or the absence of sufficient benefit. To address the questions raised above, we can make use of the dose-finding methodology that has been used successfully in the cancer and HIV settings over the last 30years. One such method is the continual reassessment method (CRM; O’Quigley et al., 1990), which makes use of working statistical models that have some optimal operating characteristics.

However, the implementation of adaptive designs is often challenging and is generally not readily available to practitioners. Consequently, these designs are not commonly applied in alcohol dependence trials. In this article, we will give a review of such methods and illustrate how we can apply them in the alcohol treatment field. The paper is organized as follows. In “Recent developments in dose-finding”, we give a basic introduction to the dose-finding background, mostly the original CRM to locate the MTD and MSD. In “Studies with topiramate”, we provide an example of identifying the optimal dose from a range of doses in an alcohol dependence trial. We conclude the paper with some discussion on future work.

Typically in a dose-finding (Phase 1) trial, we desire to administer as high a dose as possible without inducing too much toxicity. Usually, an increase in dose coincides with the number of patients who will experience a dose-limiting toxicity (DLT), which is typically defined by side effects that are sufficiently severe. However, an increase in dose is also accompanied by an increase in the number of patients who will benefit from the treatment. On the other hand, an absence of toxicity will be accompanied by a lack of treatment benefit. Consequently, the primary objective of a Phase 1 trial is to identify a dose with an “acceptable” toxicity rate. Furthermore, while the aim of the trial is to enhance the treatment of future patients, the best possible treatment must be administered to the patients enrolled in the study in order to adhere to certain ethical considerations. The highest dose that can be administered with an acceptable level of toxicity is the MTD.

Suppose there is a discrete set of *k* doses available, denoted *d*_{1}…*d*_{k}, that are ordered in terms of their probabilities of DLT, *R*(*d _{i}*), at each of the levels. That is,

The method requires estimates of the probability of toxicity at the available dose levels. Using the data accumulated for each patient in the trial thus far, we can calculate the maximum likelihood estimate (MLE) of *a*, which we denote *â*. Once *â* has been found, we can obtain an estimate of the probability of toxicity at each dose level by raising α* _{i}* to the power

We recall a brief example outlined in O’Quigley and Shen (1996). The simulated example investigates the operating characteristics of the CRM in a trial of 16 patients. The trial had six dose levels with true toxicity probabilities 0.03, 0.22, 0.45, 0.60, 0.80, and 0.95. The target toxicity rate was 0.20, indicating that dose level 2 is the correct MTD level with toxicity probability equal to 0.22. The initial guesses of the toxicity probability estimates were α_{1}=0.04, α_{2}=0.07, α_{3}=0.20, α_{4}=0.35, α_{5}=0.55, α_{6}=0.70. To be able to maximize the likelihood and generate the MLE, we need to have heterogeneity in patient observation in terms of toxicity. Thus, the trial is not considered fully underway until we have observed at least one DLT. O’Quigley and Shen (1996) advocated the use of two-stage designs with an initial escalation scheme to achieve the required heterogeneity in the responses. Some initial escalating scheme is required, and, here, we base this on a simple algorithm that includes patients in groups of three at a time. If all three remain on treatment and there is no toxicity, then the dose level is escalated. If a single subject experiences toxicity, then we remain at the same dose level. If two or more toxicities are observed, then the dose level is lowered. We continue in this way until the first DLT is encountered. As soon as we have both a toxicity and a non-toxicity, the first stage is closed; the second stage is subsequently opened, and we are in a position to fit the parameters of the under-parameterized model using the data accumulated thus far in the trial. Even though the first stage is closed, the response information accrued by the initial scheme is retained and used in the second stage.

For example, the first three patients were entered at dose level 1, and none of the three experienced a DLT. The trial was then escalated to dose level 2, and none of the three patients entered on level 2 experienced a DLT. Escalation then took place to dose level 3, where a DLT was observed in two of the three patients. Having achieved the required heterogeneity in the responses, the MLE of *a* exists. Based on the data accumulated from the first nine entered patients, the MLE is calculated to be *â* = 0.715. It follows, from raising our α* _{i}* values to this power, that we have estimated toxicity probabilities (

The original CRM design focuses only on how to locate the MTD. O’Quigley et al. (2001) proposed designs that simultaneously evaluate efficacy of treatment and toxicity in dose-finding methodology for HIV studies. The clinical setting for HIV studies allowed information on benefit to be available in a time frame similar to information on toxicity. Therefore, it was possible to evaluate overall success, which is unlike dose-finding designs for cytotoxic agents in cancer. In cancer studies, either we do not expect to observe much response or, if obtained, it is not clear how to utilize the information gathered on efficacy. Zohar and O’Quigley (2006) outlined dose-finding designs that take into account both toxicity and efficacy in the cancer setting.

In this setting, we are interested in describing both safety and efficacy by two one-parameter power models (O’Quigley et al., 2001). Safety is indicated by the presence or absence of DLT, and efficacy by the presence or absence of response to the treatment. The probability of DLT at a dose level is defined as the one-parameter power model in the previous section. Similarly, the probability of response given no toxicity at dose level *d _{i}* is given by . where β

Overall, the choice of the α* _{i}* and β

As discussed above, a requirement to be able to estimate the parameters is that we have heterogeneity in patient observation in terms of toxicity and response. Thus, the trial is not considered fully underway until we have observed at least one toxicity and one response. This can be done in two stages. As soon as we have both a positive response and a dropout or toxicity, the first stage is closed; the second stage is subsequently opened, and we are in a position to fit the models using the data accumulated thus far in the trial. We are able to reassess the working dose-toxicity and dose-response relationships through the estimated parameters, and calculate the probability of success for each dose level. The optimal dose for the next patient will be identified as that with the best probability of success. Repeating this procedure for all the remaining subjects will yield the final optimal dose: the MSD.

In this section, we will take the topiramate trial as an example to illustrate how CRM can be used to locate the MSD in alcohol treatment trials. Although topiramate has been shown to be effective at reducing drinking for alcoholics in two proof-of-concept trials (Johnson et al., 2003, 2007), the optimal dose has not be established. In the double-blind, randomized, placebo-controlled, 14-week clinical trial (Johnson et al., 2007), six doses were adopted in the dose-escalating scheme (in mg/day): 25, 50, 100, 150, 200, and 300. Suppose we wanted to choose the optimal dose in consideration of both efficacy and safety among the six doses.

The efficacy outcome of interest is the percentage of subjects with no heavy drinking days (PSNHDDs), an efficacy endpoint recommended by the Food and Drug Administration (Falk et al., 2010). It is a dichotomous endpoint measuring the frequency of heavy drinking days, whereby 0 heavy drinking days (e.g., during the last month of follow-up) is considered a good outcome and one or more heavy drinking days is considered risky drinking or a poor outcome. We also are interested in the safety measure, i.e., the adverse events. Johnson et al. (2007) found that 34 of 183 who received topiramate had adverse events. We expect that a higher dosage of topiramate is associated with a higher rate of adverse events. One of these levels will offer the greatest chance of overall success, and our purpose is to identify this level. Initially, we will treat the first cohort of patients at 25mg/day. The dose will be increased or decreased in accordance with observations on tolerance and efficacy. Once we are able to fit the stochastic models for the rates as a function of dosage, we will use CRM-type designs to concentrate treatment on the level corresponding to the optimal level.

On the basis of our fitted model, which will enable us to provide estimates of both the probabilities of success and those of failure (failure being unacceptable toxicity, dropout, or inadequate efficacy), we will take the statistical product of the conditional probability of success multiplied by one minus the probability of treatment failure. These products will vary with dose, and the one that provides the greatest value is the one that is estimated to have the greatest probability of overall success. This level is then designated the optimal one for topiramate. This optimal level accounts for dropout, loss to follow-up, unacceptable toxicity, and true efficacy.

If the optimal level is identified before all patients have been included, then this first part of the study can be brought to an early close and the second part of the study, involving comparison with a placebo, can begin. In the second part, we could perform a double-blind trial in which alcohol-dependent patients will be randomized to either the optimal dose of topiramate found in the first part or a placebo. Applying dose-finding designs in the context of pharmacotherapy research in alcoholism could be crucial to the success of a study since an incorrectly defined dose for the second part could lead us to conclude that the treatment is ineffective, either because the treatment is not well tolerated or because it fails to deliver enough of the drug to be effective.

Determining the optimal dose is of considerable clinical and scientific impact because it will enhance the safety, tolerability, and generalizability of its use. An important advantage of introducing this new methodology into the pharmacotherapy field for alcoholism is that it is more economical of time, resources, and human subjects compared with the traditional method of testing multiple fixed doses using a factorial design. Determining the optimal dose of topiramate will have an important effect on the clinical care of alcohol-dependent patients. It will widen considerably the population of those who can be provided with topiramate, enable those on the medication to derive the maximum benefit with a more tolerable adverse event profile, and improve medication compliance.

For simplicity, we make the assumption that we are dealing with a homogeneous group regarding the probability of both success and treatment failure. However, realistically, it is likely that there is some significant heterogeneity in the responses. The models described above have been extended to deal with heterogeneity, and we will be keeping this in mind as the study progresses. If we are able to identify significant sources of heterogeneity, it will be possible, within the context of a single study, to obtain more than a single successful dose, each one corresponding to a particular prognostic group (O’Quigley and Paoletti, 2003). The cost of this is a higher sample size, typically a sample of around 20% more than that encountered in the case of homogeneity.

Another question lies in clinical trials for a combination of different medications. For example, in the Combined Pharmacotherapies and Behavioral Interventions for Alcohol Dependence (COMBINE) study, a 2 (acamprosate/placebo)×2 (naltrexone/placebo)×2 [Cognitive Behavioral Intervention (CBI)/no CBI] factorial design was used, and a ninth group received CBI alone (Anton et al., 2006). In an ongoing trial, we are interested in the combination of ondansetron and naltrexone. Each medication has three dose levels: zero, low, and high. So, altogether, we have nine arms. Finding the best dose combination for these trials entails a large sample size, which could be prohibitive in terms of both time and money. Further, since there are two components, the ordering of the combinations, in terms of probabilities of toxicity and/or dropout and the probability of seeing a successful reduction in alcohol use, may not be fully known. For instance, it is not possible, *a priori*, to determine the toxicity order between the combination of low-dose ondansetron and high-dose naltrexone vs. that of high-dose ondansetron and low-dose naltrexone. Consequently, standard adaptive dose-finding methods may fail in this situation. To deal with this partial ordering, we could make use of a CRM design for partial orders proposed by Wages et al. (2011).

Another specific problem arising in the context of alcohol studies is that the recorded information can be subject to errors. We will, however, pay particular attention to robustness. That is, we want to know how changes in the recorded data influence the recommendations that are based on models. There are two main directions to this problem. The first is robustness of the choice of working model. It is important to underline our purpose here, which is not to see how well any given model identifies the true MSD. This problem has been studied in several simulation studies. Even had we, by good fortune, selected exactly the model that generates the observations, then, as a result of the inherent randomness in the experiment, the observed toxicity rate could have led us to recommend an incorrect level. What we wish to know is whether a different choice of model would have led us to this same recommendation, whether correct or not. It is thus important to separate out the task of identifying the correct level from that of verifying whether an identified level, under some model assumption, would remain the same under a different model assumption. Second, we need to deal with errors following model specification. Since toxicities themselves can be recorded with errors, it is important to investigate the influence of these errors on final recommendations. Sensitivity analysis could be used to find out how the recording error can impact the final decision of optimal dose. More work is needed to tackle both robustness issues simultaneously.

Prof. Johnson has served as a consultant to Johnson & Johnson (Ortho-McNeil Janssen Scientific Affairs, LLC), Transcept Pharmaceuticals, Inc., D&A Pharma, Organon, ADial Pharmaceuticals, LLC (with which he also serves as Chairman), Psychological Education Publishing Company (PEPCo), LLC, and Eli Lilly and Company. Drs. Wages, Liu, and O’Quigley report no biomedical financial interests or potential conflicts of interest.

This research was supported by NIAAA grant RC1 AA019274.

- Addolorato G., Abenavoli L., Leggio L., Gasbarrini G. (2005). How many cravings? Pharmacological aspects of craving treatment in alcohol addiction: a review. Neuropsychobiology 51, 59–66. doi: 10.1159/000084161. [PubMed] [Cross Ref]
- Anton R. F., O’Malley S. S., Ciraulo D. A., Cisler R. A., Couper D., Donovan D. M., et al. (2006). Combined pharmacotherapies and behavioral interventions for alcohol dependence: the COMBINE study: a randomized controlled trial. JAMA 295, 2003–2017. doi: 10.1001/jama.295.17.2003. [PubMed] [Cross Ref]
- Cheung Y. K., Chappell R. (2002). A simple technique to evaluate model sensitivity in the continual reassessment method. Biometrics 58, 671–674. doi: 10.1111/j.0006-341X.2002.00089.x. [PubMed] [Cross Ref]
- Falk D., Wang X. Q., Liu L., Fertig J., Mattson M., Ryan M., et al. (2010). Percentage of subjects with no heavy drinking days: evaluation as an efficacy endpoint for alcohol clinical trials. Alcohol. Clin. Exp. Res. 34, 2022–2034. doi: 10.1111/j.1530-0277.2010.01290.x. [PubMed] [Cross Ref]
- Johnson B. A., Ait-Daoud N., Bowden C. L., DiClemente C. C., Roache J. D., Lawson K., et al. (2003). Oral topiramate for treatment of alcohol dependence: a randomised controlled trial. Lancet 361, 1677–1685. doi: 10.1016/S0140-6736(03)13370-3. [PubMed] [Cross Ref]
- Johnson B. A., Rosenthal N., Capece J. A., Wiegand F., Mao L., Beyers K., et al. (2007). Topiramate for treating alcohol dependence: a randomized controlled trial. JAMA 298, 1641–1651. doi: 10.1001/jama.298.14.1641. [PubMed] [Cross Ref]
- Lee S. M., Cheung Y. K. (2009). Model calibration in the continual reassessment method. Clin. Trials 6, 227–238. doi: 10.1177/1740774509352199. [PMC free article] [PubMed] [Cross Ref]
- Mokdad A. H., Marks J. S., Stroup D. F., Gerberding J. L. (2004). Actual causes of death in the United States, 2000. JAMA 291, 1238–1245. doi: 10.1001/jama.291.10.1238. [PubMed] [Cross Ref]
- O’Quigley J., Hughes M. D., Fenton T. (2001). Dose-finding designs for HIV studies. Biometrics 57, 1018–1029. doi: 10.1111/j.0006-341X.2001.01018.x. [PubMed] [Cross Ref]
- O’Quigley J., Paoletti X. (2003). Continual reassessment method for ordered groups. Biometrics 59, 430–440. doi: 10.1111/1541-0420.00050. [PubMed] [Cross Ref]
- O’Quigley J., Pepe M., Fisher L. (1990). Continual reassessment method: a practical design for phase 1 clinical trials in cancer. Biometrics 46, 33–48. doi: 10.2307/2531628. [PubMed] [Cross Ref]
- O’Quigley J., Shen L. Z. (1996). Continual reassessment method: a likelihood approach. Biometrics 52, 673–684. doi: 10.2307/2532905. [PubMed] [Cross Ref]
- Shen L. Z., O’Quigley J. (1996). Consistency of continual reassessment method under model misspecification. Biometrika 83, 395–405. doi: 10.1093/biomet/83.2.395. [Cross Ref]
- Wages N. A., Conaway M. R., O’Quigley J. (2011). Continual reassessment method for partial ordering. Biometrics 67, 1555–1563. doi: 10.1111/j.1541-0420.2011.01560.x. [PMC free article] [PubMed] [Cross Ref]
- Willenbring M. L. (2007). Medications to treat alcohol dependence: adding to the continuum of care. JAMA 298, 1691–1692. doi: 10.1001/jama.298.14.1691. [PubMed] [Cross Ref]
- Zohar S., O’Quigley J. (2006). Identifying the most successful dose (MSD) in dose-finding studies in cancer. Pharm. Stat. 5, 187–199. doi: 10.1002/pst.209. [PubMed] [Cross Ref]

Articles from Frontiers in Psychiatry are provided here courtesy of **Frontiers Media SA**

PubMed Central Canada is a service of the Canadian Institutes of Health Research (CIHR) working in partnership with the National Research Council's Canada Institute for Scientific and Technical Information in cooperation with the National Center for Biotechnology Information at the U.S. National Library of Medicine(NCBI/NLM). It includes content provided to the PubMed Central International archive by participating publishers. |