## Summary

The clinical trial example used in this case study is based on the Phase III clinical trial presented in Keystone et al. (2004). The trial was conducted to evaluate the efficacy and safety of a novel treatment in the population of patients with rheumatoid arthritis who had an inadequate response to methotrexate. The primary analysis was based on the American College of Rheumatology (ACR) definition of improvement and patients were classified as responders if they experienced an improvement of at least 20% in the core criteria at Week 24 (the resulting primary endpoint is known as ACR20). Two doses of the experimental treatment were tested in the trial against a placebo.

## Define a Data Model

The data model defines parameters used for generating patient data in the clinical trial. In this case study, a set of outcome distribution parameters is specified for each of the three samples in the data model to represent different assumptions on the expected proportions of patients who meet the ACR20 response criterion at Week 24. The sets correspond to the three scenarios defined in the following table :

Outcome parameter set Response rate
Placebo Dose L Dose H
Scenario 1 30% 50% 50%
Scenario 2 30% 40% 50%
Scenario 3 30% 50% 45%

The outcome parameters are specified using the following R code.

The sample size is set to 100 patients per arm and a simplified assumption that all patients complete the trial is made in this case study. In the Mediana package, a dropout mechanism can be introduced to model more realistic settings with incomplete outcomes. This can be easily accomplished by specifying the dropout distribution and associated parameters (enrollment distribution, follow-up duration) in a Design object. For example, the dropout process can be modeled using an exponential distribution. As the primary endpoint is evaluated at Week 24, the follow-up duration is considered fixed for all patients, i.e., 24 weeks, and the enrollment period is estimated to be approximately 2 years, i.e., 104 weeks.

The data model is initialized using the DataModel object and each component of this object can be added one at a time. The outcome distribution is defined using the OutcomeDist object with the BinomDist distribution, the common sample size per arm is specified in the SampleSize object and the outcome distribution parameters for each trial arm are specified in Sample objects.

## Define an Analysis Model

The analysis model is composed of the two statistical tests that will be performed to compare Doses L and H to Placebo as well as the multiplicity adjustment procedures used to control the Type I error rate.

Each dose-placebo comparison will be carried out using a two-sample test for proportions, defined in the Test object with the PropTest method. Since the statistical tests are one-sided, the order of the two samples in the samples argument is important. If a higher numerical value of the endpoint indicates a beneficial effect, which is the case with the ACR20 response rate, the first sample must correspond to the sample expected to have the lower value of the endpoint. Thus "Placebo" is included as the first sample in the test specification.

Concerning the multiplicity adjustment procedures, three procedures are defined in this analysis model using the MultAdjProc object. First, in order to request a straightforward analysis without any adjustment, an empty object is set up, i.e., MultAdjProc(proc = NA). The fixed-sequence procedure (Procedure F) is defined using a FixedSeqAdj method while the Hochberg adjustment procedure (Procedure H) is defined using the HochbergAdj method.

It is worth noting that, by default, the multiplicity adjustment procedures defined in the analysis model will be applied to all tests included in the AnalysisModel object in the order specified. Specifically, the fixed-sequence procedure will begin with the test of Dose H versus Placebo and then examine the test of Dose L versus Placebo.

## Define an Evaluation Model

The evaluation model specifies the metrics for assessing the performance of the analysis model based on the assumptions defined in the data model. In this case study, several criteria are evaluated to assess the performance of the three analysis methods (no multiplicity adjustment, Procedure F and Procedure H). First, the disjunctive criterion, is defined as the probability to reject at least one null hypothesis, i.e., the probability of demonstrating a statistically significant effect with at least one dose. This criterion is defined using the DisjunctivePower method.

The second criterion is based on simple weighted power, which corresponds to the weighted sum of the marginal power functions of the two dose-placebo tests. This criterion can be defined in the evaluation model using the WeightedPower method, where the weights associated with each test are specified in the par argument.

Finally, the partition-based weighted criterion is a custom criterion, which is not currently implemented in the Mediana package. Nonetheless, a custom criterion function can be created and used within the evaluation model. This custom function includes three arguments. The first argument (test.result) is a matrix of p-values associated with the tests defined in the tests argument of the Criterion object. Similarly, the second argument (statistic.result) is a matrix of results corresponding to the statistics defined in the statistics argument of the Criterion object. In this case study, no Statistic objects were specified in the analysis model and this criterion will only use the p- values returned by the tests associated with the two dose-placebo comparisons. Finally, the last argument (parameter) contains the optional parameter(s) defined in the par argument of the Criterion object. For the partition-based weighted power, the par argument must contain the weights associated with each of the three outcomes:

• Reject H1 only

• Reject H2 only

• Simultaneously reject H1 and H2

The mult.cs1.PartitionBasedWeightedPower function evaluates the probability of each outcome and then computes the weighted sum.

It needs to be noted that the order of tests in the tests argument of this custom function is important. The first test should be the one corresponding to the comparison of Dose H versus Placebo and the second test is based on the comparison of Dose L versus Placebo. Finally, the evaluationmodel can be constructed by specifying each Criterion object. In addition, marginal power of each test can be computed using the MarginalPower method.

## Perform Clinical Scenario Evaluation

Using the data, analysis and evaluation models, simulation-based Clinical Scenario Evaluation is performed by calling the CSE function: