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

**|**Biostatistics**|**PMC4834949

Formats

Article sections

- Abstract
- 1. Introduction and motivation
- 2. A decision-theoretic design
- 3. Trial design
- 4. Simulation study
- 5. Discussion
- Supplementary material
- Funding
- Supplementary Material
- References

Authors

Related links

Biostatistics. 2016 April; 17(2): 304–319.

Published online 2015 November 9. doi: 10.1093/biostatistics/kxv045

PMCID: PMC4834949

Juhee Lee^{*}

Department of Applied Mathematics and Statistics, Baskin School of Engineering, University of California, 1156 High Street, Mail Stop SOE2, Santa Cruz, CA 95064, USA

Peter F. Thall

Department of Biostatistics, University of Texas M.D. Anderson Cancer Center, Houston, TX, USA

Yuan Ji

Program of Computational Genomics & Medicine, NorthShore University Health System, Evanston, IL, USA and Department of Public Health Sciences, The University of Chicago, Chicago, IL, USA

Department of Mathematics, University of Texas, Austin, TX, USA

To whom correspondence should be addressed. Email: ude.cscu.eos@eeleehuj

Received 2015 March 3; Revised 2015 October 14; Accepted 2015 October 15.

Copyright © The Author 2015. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com.

This paper is motivated by a phase I–II clinical trial of a targeted agent for advanced solid tumors. We study a stylized version of this trial with the goal to determine optimal actions in each of two cycles of therapy. A design is presented that generalizes the decision-theoretic two-cycle design of Lee *and others* (2015. Bayesian dose-finding in two treatment cycles based on the joint utility of efficacy and toxicity. *Journal of the American Statistical Association*, to appear) to accommodate ordinal outcomes. Backward induction is used to jointly optimize the actions taken for each patient in each of the two cycles, with the second action accounting for the patient's cycle 1 dose and outcomes. A simulation study shows that simpler designs obtained by dichotomizing the ordinal outcomes either perform very similarly to the proposed design, or have much worse performance in some scenarios. We also compare the proposed design with the simpler approaches of optimizing the doses in each cycle separately, or ignoring the distinction between cycles 1 and 2.

This paper is motivated by the problem of designing a dose-finding trial of a new agent for cancer patients with advanced solid tumors. The agent aims to inhibit a kinase, which regulates cell metabolism and proliferation, in the cancer cells to reduce or eradicate the disease. The agent is given orally each day of a 28-day cycle at one of five doses, 2, 4, 6, 8, or 10mg, combined with a fixed dose of standard chemotherapy. Because both efficacy and toxicity are used for dose-finding, it is a phase I–II trial (Thall and Cook, 2004; Yin *and others*, 2006; Zhang *and others*, 2006; Thall and Nguyen, 2012). Both outcomes are 3-level ordinal variables, with toxicity defined as None/Mild (grade 0,1), Moderate (grade 2), or Severe (grade 3,4) and efficacy defined in terms of disease status compared with baseline, with possible values progressive disease (PD), stable disease (SD), or partial or complete response (PR/CR).

We study a stylized version of this trial with the more ambitious goal to determine optimal doses or actions for each patient in each of two cycles of therapy. This is a major departure from conventional dose-finding designs, which focus on choosing a dose for only the first cycle. While virtually all clinical protocols for dose-finding trials include rules for making within-patient dose adjustments in cycles after the first, this aspect usually is ignored in the trial design. In practice, each patient's doses in cycle 2, or later cycles, are chosen subjectively by the attending physician. To choose a patient's cycle 2 dose using a formal rule, it is desirable to use the patient's dose-outcome data from cycle 1, as well as data from other patients treated previously in the trial. Thus, ideally, a decision rule that is adaptive both within and between patients is needed.

Recent papers on designs accounting for multiple treatment cycles include Cheung *and others* (2014) and Lee *and others* (2015). In this paper, we build on the latter, who use a decision-theoretic approach for dose-finding in two cycles based on joint utilities of binary outcomes in each cycle. We extend the model to accommodate ordinal outcomes, and use a decision criterion that accounts for the many possible (efficacy,toxicity) outcomes in each of the two treatment cycles, including the risk-benefit trade-offs between the levels of efficacy and toxicity. In the stylized version of the trial described above, since there are 3-level ordinal toxicity and efficacy outcomes in each cycle, accounting for two cycles there are 81 possible elementary outcomes for each patient. Consequently, dose-finding is a much more complex problem than in a conventional phase I–II trial with two binary outcomes that chooses a dose for cycle 1 only.

Aside from the issue of accounting for two cycles, an important question is whether the additional complexity required to account for ordinal outcomes provides practical benefits compared with the common approach of dichotomizing efficacy and toxicity, which would allow the two-cycle design of Lee *and others* (2015) to be applied. Simulations, described in Section 4.4 of the main text, Figure Figure2,2, and Section 3 of Supplementary Material (available at *Biostatistics* online), show that reducing ordinal outcomes to binary variables produces a design that either performs very similarly to the proposed design, or has much worse performance in certain scenarios. Moreover, the behavior of the simplified design depends heavily on how one chooses to reduce the two ordinal outcomes to two binary variables.

Plots of for a comparison of DTD-O2 vs. a design with binary outcomes. Here, , , and represent empirical mean utilities of patients treated in the trial, true mean utilities of treatments given to patients in the trial, and true expected utilities chosen **...**

A naive design might aim to optimize the doses given in the two cycles separately. This may be not optimal. To see this, denote a patient's toxicity outcome by and efficacy outcome by for and denote the current data from patients by We include as a possible action in either cycle for cases where it has been determined that no dose is acceptable, so the action in each of cycles may be either to choose a dose or , that is, with 1 and denoting the minimum dose and the maximum dose levels, respectively. Suppose that some optimality criterion has been defined. If one derives optimal adaptive actions for cycle 1 and for cycle 2 separately, each based on the current data an inherent flaw is that in choosing one for all patients it ignores each patient's cycle 1 data. As in Lee *and others* (2015), we derive optimal decision rules = with the important property that = is a function of the first cycle decision and response . This is implemented by applying backward induction (Bellman, 1957, etc.). The method accounts for the patient's cycle 1 dose and outcomes, as well as other patient's data, in making an optimal decision for cycle 2.

Iasonos *and others* (2011) and Van Meter *and others* (2012) studied the use of ordinal toxicity outcomes for a generalized continual reassessment method and reported that gains in performance of their ordinal toxicity designs are not substantial in comparison to binary toxicity designs. However, the comparison looks quite different for the model-based two-cycle design for bivariate ordinal (efficacy, toxicity) outcomes that we propose in this paper. In simulations described in Section 4.4, we compare the proposed design with designs that do not properly model association between cycles. In simulations reported in Section 4.5, we show that the use of ordinal rather than binary outcomes can substantially improve design performance in our setting.

Section 2 describes the proposed decision-theoretic method for ordinal outcomes in two cycles (DTD-O2). Sections 3 and 4 include decision criteria using utilities and a simulation study. The last section concludes with a final discussion.

For notational convenience, we denote the possible levels of toxicity by and efficacy by . For the motivating trial, these are for None/Mild, 1 for Moderate, and 2 for Severe, and for PD, 1 for SD, and 2 for CR/PR, so . If the adaptively chosen cycle 1 action for any patient, then the trial is stopped and no more patients are enrolled. Otherwise, the patient receives a dose of the agent in cycle 1. A cycle 2 action is a function mapping the cycle 1 dose and outcomes, to an action in For example, if the cycle 1 action produced None/Mild toxicity (), one possible cycle 2 action is if , and if or 2. That is, if there was little or no toxicity but PD in cycle 1, then the action increases the dose in cycle 2, but if the patient had SD or better then it repeats the cycle 1 dose. The design thus involves an alternating sequence of decisions and observed outcomes, , , and .

We apply a Bayesian decision-theoretic paradigm to determine an optimal decision rule. First, focus on cycle 1, and temporarily ignore cycle 2. The general setup of a Bayesian decision problem involves actions , observable data , parameters that index a sampling model for the data, and a prior probability model for the parameters. We discuss specification of in more detail below. A utility function formalizes relative preferences for alternative actions under hypothetical outcomes and assumed truth . Starting from first principles, one can then argue ((Robert, 2007, Chapter 2)) that a rational decision-maker chooses the action that maximizes utility in expectation, that is

(2.1)

The integral is the expected utility with the expectation taken with respect to . To simplify notation, we will henceforth suppress conditioning on in the notation.

In the two-cycle dose-finding problem, the sequential nature of the within-patient decisions complicates the solution. In the second cycle, the utility is replaced by the expected utility under optimal continuation. Denote and . We get an alternating sequence of optimization and expectation

(2.2)

with the second cycle expected total utility as a function of , and the optimal second cycle decision . When we substitute and take the expectation with respect to we obtain

(2.3)

which is maximized to determine the optimal decision for cycle 1, . This alternating sequence of maximization and expectation, called dynamic programming, is characteristic of sequential decision problems. While it often leads to intractable computational problems ((Parmigiani and Inoue, 2009, Chapter 12)), in the present setting with ordinal outcomes the problem is solvable. Dynamic programming recently has been applied in other clinical trial design settings (Murphy, 2003; Zhao *and others*, 2011; Lee *and others*, 2015; Cheung *and others*, 2014).

We construct a utility function

(2.4)

as a sum over cycle-specific utilities , where is a scale parameter. If then the cycle 2 utility is ignored in selecting while corresponds to treating utilities in the two cycles equally. Optimal decisions may change under different values of . Even with , however, the importance of jointly modeling the two cycles remains in that inference on can be enhanced through borrowing information across cycles. For the simulations in Section 4, we used . A sensitivity analysis in is reported in the Supplementary Materials (available at *Biostatistics* online). The utility function (2.4) focuses on the clinical outcomes and is a function of only. That is, the inference on does not affect utility, and we do not initially consider preferences across doses . We thus drop and from the arguments of hereafter.

In practice, numerical utilities of the elementary must be elicited from the clinical collaborators, with specific numerical values reflecting physicians' relative preferences (cf. Thall and Nguyen, 2012). In our stylized illustrative trial, we fix the utilities of the best and worst possible outcomes to be and . In general, any convenient function with and that gives higher utilities to more desirable outcomes may be used. For future reference, we note that is the expected utility corresponding to i.e. do not treat the patient. Table Table11 shows the utilities that will be used for our simulation studies.

To reduce notation, we denote the utility as a function of hypothetical outcomes , and drop the arguments and . Upper case denotes expected utility, with data , removed by marginalization and decisions , substituted by maximization, as in (2.2). In addition to the cycle index , the arguments of clarify the level of marginalization and maximization. Maximizing in (2.1) and inside the integral in (2.2) yields the optimal action pair , where is either a dose or , is applicable only when is a dose, and is a function of and the patient's cycle 1 outcomes, . Assuming that the utility function takes the additive form (2.4), we define cycle-specific expected utilities, with the expected utility for cycle 2 given by

(2.5)

Figure Figure1(a)–(c)1(a)–(c) illustrates under the assumed simulation truth of Scenario 3 (discussed in Section 4.2), and shows how changes with , given . Figure Figure1(d)1(d) illustrates the assumed true over for the simulation scenarios discussed in Section 4.2.

(a)–(c) The true expected cycle 2 utilities of taking given , with for scenario 3. Each panel corresponds to one of the three possible outcomes of . is acceptable only when its expected utility is greater than that of , . is marked with a **...**

Some practical guidelines of using utility functions for a design with ordinal outcomes in the two-cycle setting are provided in Section 1 of the Supplementary Material (available at *Biostatistics* online).

Equation (2.2) includes two maximizations to determine and In the discussion thus far, we have not used the particular elements of and they might have been any actions. In actual dose-finding, ethical and practical constraints are motivated by the knowledge that, in general, higher doses carry a higher risk of more severe toxicity. We thus require a more restrictive action set, with additional conditions for the acceptability of a dose assignment.

The first additional criterion is that we *do not skip untried dose levels* when escalating. This rule is imposed almost invariably in actual trials with adaptive dose-finding methods. Let denote the highest dose level among the dose levels that have been tried in cycle 1 and the highest dose level among those that have been tried in either cycle 1 or cycle 2. The search for the optimal actions is constrained such that and . In addition, we do not escalate a patient's dose level in cycle 2 if severe toxicity was observed in cycle 1 (). Both restrictions are due to safety concerns.

A third safety restriction is defined implicitly in terms of the cycle-specific utility A patient is not treated () if there is no dose with expected utility . For the expected utility is compared with the expected utility of not receiving any treatment in both cycles, (horizontal dotted line in Figure Figure1(d)).1(d)). Any with below the line is not considered acceptable treatment. For , the expected utility is similarly compared with the expected utility of , (horizontal dotted line in Figure Figure1(a)–(c)),1(a)–(c)), and any with below the line is not acceptable.

At any interim point in the trial, let denote the current data, including dose assignments for previously enrolled patients. The three conditions together make the action sets for and dependent on , and . We let and denote the action sets for and , respectively, that are implied by these three restrictions.

Thus far, our discussion of optimal decisions has not included a particular probability model. We will assume a 4D ordinal probit model for with a regression on doses and standardized to the domain [0, 1], with and . Let denote a vector of latent probit scores for the th patient and let and denote fixed cutoffs that define if and if . While varying the mean of distributions of and across cycles, the same cutoffs are used for all cycles. The and are multivariate normal probit scores,

(2.6)

and . The covariance matrix implies associations across cycles and across outcomes through and . Given that the ordinality of the outcomes is accounted for by the latent probit scores and fixed cutoff parameters and a simple yet flexible model for regression on dose is obtained by assuming = with for toxicity and for efficacy. A discussion of nonlinear dose–response models is given by Bretz *and others* (2005). We assume that the toxicity and efficacy probabilities increase monotonic in dose by requiring and . Denote , and . We complete the model with a normal prior , .

Denote . Although, in terms of the utility-based objective function, yields the best clinical outcomes for the next patient, the performance of the design, in terms of frequentist operating characteristics, can be improved by including adaptive randomization (AR) among actions giving values of the objective function near the maximum at . Using AR decreases the probability of getting stuck at a suboptimal and also has the effect of treating more patients at doses having larger utilities, on average. The problem that a “greedy” search algorithm may get stuck at suboptimal actions, and the simple solution of introducing additional randomness into the search process, are well known in the optimization literature (cf. Tokic, 2010). This has been dealt with only very recently in dose-finding (Bartroff and Lai, 2010; Azriel *and others*, 2011; Braun *and others*, 2012; Thall and Nguyen, 2012).

To implement AR, we first define to be a function decreasing in patient index and denote . We define the set of -*optimal doses for cycle 1* to be

(3.1)

The set, contains doses in whose is within of the maximum posterior mean utility. Similarly, we define the set of -optimal doses for cycle 2 given to be

(3.2)

in (3.1) is based on (2.5). Our design randomizes patients uniformly among doses in for and for which we call AR(). Numerical values of depend on the range of , and are determined by preliminary trial simulations in which is varied.

Our illustrative trial studied in the simulations is a stylized version of the phase I–II chemotherapy trial with five dose levels described in Section 1, but here accounting for two cycles of therapy. The maximum sample size is 60 patients with a cohort size of 2. Based on preliminary simulations, we set for the first 10 patients, for the next 10 patients, and for the remaining 40 patients. An initial cohort of 2 patients is treated at the lowest dose level in cycle 1, their cycle 1 toxicity and efficacy outcomes are observed, the posterior of , and is computed, and actions are taken for cycle 2 of the initial cohort. If then patient does not receive a second cycle of treatment. If , then AR() is used to choose an action for cycle 2 from . When the toxicity and efficacy outcomes are observed from cycle 2, the posterior of is updated. The second cohort is not enrolled until the first cohort has been evaluated for cycle 1. For all cohorts after the first, after the outcomes of all previous cohorts are observed, the posterior is updated, the posterior expected utility, is computed using , and is determined. Using and , we find and search for . If for any interim then , and the trial is terminated. If , we then choose a cycle 1 dose from using AR(). Once the outcomes in cycle 1 are observed, the posterior is updated. Using and , is searched. If contains only, then and no cycle 2 dose is given to patient Otherwise, is selected from using AR(). The toxicity and efficacy outcomes are observed from cycle 2 and the posterior of is updated. The above steps are repeated until either the trial has been stopped early or has been reached. At the end of the trial, we record as recommended first cycle dose and as optimal policy . If the trial is early terminated, let and for all .

Let DTD-O2 denote the proposed decision-theoretic two-cycle design. We compare DTD-O2 with three other designs. The first is obtained by reducing each 3-level efficacy and toxicity outcome to a 2-category (binary) variable by combining categories, but using the same probability model to ensure a fair comparison. The next two comparators are single cycle designs. The first, called Single Cycle Comparator 1 (SCC1), assumes no association between cycles and optimizes and separately. The second, called Single Cycle Comparator 2 (SCC2), does not distinguish between cycles and treats the two cycles identically.

For SCC1, we assume patient-specific random probit scores, independent over cycles, , where and is the covariance matrix. We let be the upper-left partition of in (2.6). Owing to the independence of probit scores over cycles within a patient, SCC1 models the association between and within the same cycle only and does not assume any association between outcomes in different cycles, for example, and . The other model specification including the regression of on the dose in Section 2.4 stays the same. For SCC2, in addition to having patient- and cycle-specific random probit scores as in SCC1, we assume that the mean dose effects are identical in the two cycles by dropping the cycle index from in Section 2.4, i.e. setting , for all . For these two methods, we apply the acceptability rules in Section 2.3 and the AR rules in Section 3.1 for each cycle separately. For example, a trial is terminated if for all and is defined with only. Also, the no-escalation rule after , no-skipping rule and AR similar to those implemented in the proposed method are implemented to SCC1 and SCC2.

We simulated trials under each of 8 scenarios using each of the designs. A total of trials were simulated for each design under each scenario. The simulation scenarios were determined by fixing a set of marginal probabilities and regression coefficients on probit scores, given in Table Table22 and Supplementary Material Table S1 (available at *Biostatistics* online). Each simulation scenario is specified by the marginal distributions of and . Table Table22 gives the true and under each scenario. The corresponding probit scores are and , where is the cumulative distribution function of the standard normal distribution. To ensure a fair comparison, we intentionally define a simulation truth that is different from the assumed model used by the design methodology. The simulation model is best described as a generative model, first for , then given , and then given .

*Generating *: We first generate from the distribution specified by where . For later reference, we define a rescaled variable as which is evenly spaced in .

*Generating *: Conditional on we specify a distribution of by letting

with coefficient . Here, induces association between the cycle 1 outcomes, and . A negative value of leads to a negative association between and , that is, , . For later use, we define by rescaling to be evenly spaced in similarly to

*Generating*
: We generate using

Here, is a standardized dose in . We restrict and to induce a positive association of with and and negative association with . Here, determines how and jointly affect . A large negative value of implies that given that (severe toxicity) is observed at , the probability of observing , greatly increases for all . Similarly, observing (mild toxicity) at greatly increases the probability of observing for all , implying a large positive value of .

*Generating *: We use

where and . Similar to , determines a joint effect of and on . The detailed specification of the coefficients, and for each simulation scenario is described in the Supplementary Materials (available at *Biostatistics* online). Table Table33 shows the optimal actions, and , over two cycles under each of the 8 simulation scenarios under the simulation truth. For example, in Scenario 3, the optimal cycle 1 action is to give dose level 3, and the optimal cycle 2 action is to treat patients with at , and at if .

We calibrate the fixed hyperparameters, , for and and the cutoff points, , using effective sample size (ESS), described in the Supplementary Materials (available at *Biostatistics* online). We set and the cutoffs, and , and simulate 1000 pseudo-samples of , , and . We then compute probabilities of interest based on the pseudo-samples, such as and , . For all simulations, we determined to give each prior ESS between 0.5 and 2, using the approximation obtained by matching moments with a Dirichlet distribution. We used the same for SCC1 and SCC2.

We evaluate design performance for the patients treated in the trial using three different summary statistics, , , and . Recall that in a trial we record the clinical outcomes of the patients with their assigned doses and recommended doses for future patients, , , and , and respectively. We index the simulated replications of the trial by . We define average utility for the patients in the th simulated trial in two different ways; and Note that is a function only of occurred outcomes, , whereas depends on the true utilities of assigned doses . For and , is used as the utility for patients with . The empirical mean total payoffs taken over all simulated trials are

One may regard and as indexes of the ethical desirability of the method, given

The proposed method gives an optimal action for cycle 1, and policy for cycle 2. We let for all if so the trial is terminated early. We use and to evaluate performance in terms of future patient benefit. Under SCC1 and SCC2, is not a function of . For SCC2, and are identical. Assuming that the simulation truth is known, we define the expected payoff in cycle 1 of giving action to a future patient as for . That is the expected utility with respect to the assumed distribution of when is given. For , let . This expectation is computed under the distribution of given . If the rule is used, the expected cycle 2 payoff is

where becomes if . The total expected payoff to a future patient treated using the optimal regime = is defined to be .

We first compare DTD-O2 with designs obtained by collapsing each trinary toxicity and efficacy outcome to a binary variable. This mimics what often is done in practice in order to apply a phase I–II design based on binary efficacy and toxicity. We use an appropriately reduced version of our assumed underlying model to ensure a fair comparison. Since this reduction is not unique, we exhaustively define binary outcomes in four different ways, binary cases 1–4, given in Section 4 of the Supplementary Material (available at *Biostatistics* online). The utilities associated with the binary outcomes are defined accordingly based on the utilities in Table Table1.1. The results, in terms of
and are summarized graphically in Figure Figure2.2. Scenario 8 is not included in Figure Figure22 because the optimal action is in both cycles, and in this case all designs stop the trial early with high probability, The figure shows that reducing to binary outcomes can produce designs with much worse performance than DTD-O2, while for some cases the performance may be comparable. The binary outcome design's performance also varies substantially with the particular dichotomization used. Since different physicians may combine ordinal categories in different ways, the practical implication is that the additional complexity of the ordinal outcome design is worthwhile, in terms of benefit to both the patients treated in the trial and future patients.

The simulation results for DTD-O2, SCC1, and SCC2 are summarized in Figure Figure3.3. Scenarios 1–4 have the same marginal toxicity and efficacy probabilities, but different values of coefficients (), yielding different probit scores and different association structures of , , and . Scenario 1 has large and , so that the cycle 1 toxicity outcome greatly affects cycle 2 expected utilities in the simulation truth. As shown in Table Table3,3, the optimal action in cycle 2 after observing severe toxicity in cycle 1 is regardless of the cycle 1 efficacy outcome. Scenario 4 is similar to Scenario 1 but the cycle 1 efficacy outcome heavily affects the cycle 2 treatment in that all cycle 2 treatments are less desirable than when PD is observed in cycle 1. In Scenarios 2 and 3, the two cycle 1 outcomes jointly determine the cycle 2 treatment as shown in the tables. Scenario 3 has larger association between and within each cycle. In Scenarios 1–4, modeling dependence across cycles improves the performance, as shown in Figure Figure3,3, where DTD-O2 is superior to SCC1 and SCC2 in terms of all the three criteria, , , and . Since the only difference between DTD-O2 and SCC1 is whether the two cycles are modeled jointly or separately, the results show that the joint modeling significantly improves the performance. Differences in the performance are smaller for Scenarios 1 and 4. This may be because the true structure that one cycle 1 outcome dominates cycle 2 decisions in the scenarios is not easily accommodated under the assumed covariance structure in (2.6) and each trial gets only a small number of patients. In such a case, separate estimation for the two cycles may not be a very poor approach. In addition, the three methods are compared using and based on the last 20 patients in each trial for the three designs (not shown). This comparison shows that the improvement by DTD-O2 over the other two methods becomes greater, especially for Scenarios 1 and 4. It may imply that learning takes more patients for DTD-O2 when there is a discrepancy between the truth and the model assumption.

Plot of for a comparison with SCC1 and SCC2. Here, , , and represent empirical mean utilities of patients treated in the trial, true mean utilities of treatments given to patients in the trial, and true expected utilities chosen for future patients, **...**

Scenarios 5–7 have different shapes for as a function of . The cycle 1 utilities are U-shaped in Scenario 5, monotone increasing in Scenario 6, and monotone decreasing in Scenario 7. Very mild associations between outcomes and between cycles are assumed for these scenarios. For Scenarios 5 and 6, DTD-O2 achieves notably better performance (see Figure Figure3),3), with and similar to each other for DTD-O2. This implies that DTD-O2 identifies desirable actions early in the trial, treats many of the patients with the desirable actions, and has a high probability of selecting truly optimal actions at the end of a trial. In Scenario 7, DTD-O2 shows slightly worse performance (see the rightmost of Figure Figure3).3). In the simulation truth of Scenario 7, the cycle 1 expected utility does not change much with but the cycle 2 expected utility is very sensitive to , , and . This is a very challenging case for DTD-O2, and not modeling dependence between the cycles leads to better performance than incorrectly modeling in this particular scenario. Scenario 8 has no acceptable dose in either cycle. All the three methods terminate the trials with probability 1 in this case, with mean sample sizes 9.11, 8.33, and 8.29.

In all 8 scenarios, SCC2 yields better results than SCC1. This may be because and happen to be identical in many cases, so combining outcomes from the two cycles works well. However, the results for Scenarios 1–4 show that using each patient's cycle 1 dose and outcomes to select gives significantly superior performance in cases where there is significant dependence between the two cycles. More results are summarized using empirical toxicity and efficacy probabilities in Section 3 of Supplementary Material (available at *Biostatistics* online).

We carried out a sensitivity analysis in under Scenarios 2 and 5, including the four binary outcome designs, SCC1, SCC2, and DTD-O2, for , 0.4, 0.8, and 1.0. The results, given in Section 5 of Supplementary Material (available at *Biostatistics* online), show that changes in design performance with are very small, but corresponding to no use of cycle 2 utility in making a decision at cycle 1, yields higher early termination probabilities for binary outcome cases 1 and 3.

We have extended the decision-theoretic two-cycle phase I–II dose-finding method in Lee *and others* (2015) to accommodate ordinal outcomes. Our simulations show that incorporating cycle 1 information into the cycle 2 treatment decision yields good performance for both patients treated in a trial and future patients. The simulations in Figure Figure22 show that this extension may greatly improve design performance, quantified by
and compared with using binary toxicity and efficacy indicators. The proposed model and method also compared quite favorably with either assuming the two cycles are independent or ignoring the distinction between cycles 1 and 2.

In theory, DTD-O2 could be extended to more than two cycles. For this to be tractable, additional modeling assumptions may required to control the number of parameters, since decisions must be made based on small sample sizes. Two possible approaches are to model dependence among cycles as a function of distance between cycles, or to make a Markovian assumption.

Supplementary material is available at http://biostatistics.oxfordjournals.org.

Y.J. research is supported in part by NIH R01 CA132897. P.F.T. research was supported in part by NIH R01 CA 83932. P.M. research was supported in part by NIH 1-R01-CA157458-01A1. This research was supported in part by NIH through resources provided by the Computation Institute and the Biological Sciences Division of the University of Chicago and Argonne National Laboratory, under grant S10 RR029030-01.

We specifically acknowledge the assistance of Lorenzo Pesce (University of Chicago). *Conflict of Interest*: None declared.

- Azriel D., Mandel M., Rinott Y. (2011). The treatment versus experimentation dilemma in dose finding studies. Journal of Statistical Planning and Inference 1418, 2759–2768.
- Bartroff J., Lai T. L. (2010). Approximate dynamic programming and its applications to the design of phase I cancer trials. Statistical Science 255, 245–257.
- Bellman R. (1957)
*Dynamic Programming*, 1 edition Princeton, NJ, USA: Princeton University Press. - Braun T. M., Kang S., Taylor J. M. G. (2012). A phase I/II trial design when response is unobserved in subjects with dose-limiting toxicity. Statistical Methods in Medical Research. 0962280212464541. [PMC free article] [PubMed]
- Bretz F., Pinheiro J. C., Branson M. (2005). Combining multiple comparisons and modeling techniques in dose–response studies. Biometrics 613, 738–748. [PubMed]
- Cheung Y. K., Chakraborty B., Davidson K. W. (2014). Sequential multiple assignment randomized trial (smart) with adaptive randomization for quality improvement in depression treatment program. Biometrics 712, 450–459. [PMC free article] [PubMed]
- Iasonos A., Zohar S., O'Quigley J. (2011). Incorporating lower grade toxicity information into dose finding designs. Clinical Trials 84, 370–379. [PMC free article] [PubMed]
- Lee J., Thall P. F., Ji Y., Müller P. (2015). Bayesian dose-finding in two treatment cycles based on the joint utility of efficacy and toxicity. Journal of the American Statistical Association 110510, 711–722. [PMC free article] [PubMed]
- Murphy S. A. (2003). Optimal dynamic treatment regimes. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 652, 331–355.
- Parmigiani G., Inoue L. (2009) Decision Theory: Principles and Approaches. New York: Wiley.
- Robert C. P. (2007) The Bayesian Choice: From Decision-Theoretic Foundations to Computational Implementation, 2nd edition Berlin: Springer.
- Thall P. F., Cook J. D. (2004). Dose-finding based on efficacy–toxicity trade-offs. Biometrics 603, 684–693. [PubMed]
- Thall P. F, Nguyen H. Q. (2012). Adaptive randomization to improve utility-based dose-finding with bivariate ordinal outcomes. Journal of Biopharmaceutical Statistics 224, 785–801. [PMC free article] [PubMed]
- Tokic M. (2010). Adaptive -greedy exploration in reinforcement learning based on value differences. In:
*KI 2010: Advances in Artificial Intelligence*. Berlin: Springer, pp. 203–210. - Van Meter E. M., Garrett-Mayer E., Bandyopadhyay D. (2012). Dose-finding clinical trial design for ordinal toxicity grades using the continuation ratio model: an extension of the continual reassessment method. Clinical Trials 93, 303–313. [PubMed]
- Yin G., Li Y., Ji Y. (2006). Bayesian dose-finding in phase i/ii clinical trials using toxicity and efficacy odds ratios. Biometrics 623, 777–787. [PubMed]
- Zhang W., Sargent D. J., Mandrekar S. (2006). An adaptive dose-finding design incorporating both toxicity and efficacy. Statistics in Medicine 2514, 2365–2383. [PubMed]
- Zhao Y., Zeng D., Socinski M. A., Kosorok M. R. (2011). Reinforcement learning strategies for clinical trials in nonsmall cell lung cancer. Biometrics 674, 1422–1433. [PMC free article] [PubMed]

Articles from Biostatistics (Oxford, England) are provided here courtesy of **Oxford University Press**

PubMed Central Canada is a service of the Canadian Institutes of Health Research (CIHR) working in partnership with the National Research Council's national science library 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. |