Case study 3
Summary
The clinical trial example used in this case study is based on the panitumumab trial in a population of patients with metastatic colorectal cancer (Amado et al., 2008). Using the same patient population, consider a Phase III trial for a novel treatment versus control (best supportive care) that utilizes a balanced design with a 1:1 randomization scheme. The treatment is a fully human antibody against the EGFR (epidermal growth factor receptor) and is expected to benefit mostly patients in a prespecified subset. This subset is defined based on each patient’s KRAS (Kirsten rat sarcoma viral oncogene homolog) status and includes patients with wildtype KRAS. These patients will be referred to as biomarkerpositive patients and patients with a mutated KRAS status will be referred to as biomarkernegative patients. For simplicity, it will be assumed that the KRAS status can be ascertained in all patients. The treatment effect in this twoarm trial will be evaluated based on progressionfree survival (PFS).
The decisionmaking rules will be set up using the influence and interaction conditions defined in Section 3.2.3 of the Clinical Trial Optimization Using R.
The influence condition has been defined in Case study 2. Additional restrictions based on the influence condition as well as the interaction condition are required in a more general setting with three potential efficacy claims (broad, restricted and enhanced claims) to prevent illogical conclusions.
Beginning with a basic decision rule for formulating efficacy claims based on statistical considerations, the trial’s sponsor can consider pursuing the enhanced claim of a beneficial effect in the overall population as well as the selected subpopulation (combination of the broad and restricted claims) if the two null hypotheses of interest (H0 and H+) are simultaneously rejected. It is immediately clear that this rule does not account for the influence condition and, as a consequence, may lead to erroneous conclusions discussed above. Another potential problem with this naive rule is that the enhanced claim can be recommended even if the treatment effect is homogeneous across the biomarkerpositive and biomarkernegative subsets. This may happen if the biomarker that was believed to have predictive properties turned out to be noninformative.
To address the limitations of the basic decision rule, the interaction condi tion needs to be applied in conjunction with the influence condition as shown in the Figure below:
This figure defines the recommended set of decision rules for trials with three potential claims. The influence and interaction conditions are applied sequentially. The latter assesses the degree of treatmentbybiomarker interaction or strength of a differential effect in the biomarkerpositive and biomarkernegative subgroups. The condition is satisfied if there is a clinically important interaction that supports the conclusion that the treatment provides a substantial additional benefit in patients with a biomarkerpositive status compared to patients in the overall population. The interaction condition can be assessed using any treatmentbybiomarker interaction test. This condition can be defined using a simple test based on the ratio of the observed effect sizes in the biomarkerpositive and biomarkernegative subgroups (i.e., the interaction condition is satisfied if this ratio is greater than a predefined threshold).
The figure shows that the enhanced claim can be made only if the interaction condition is satisfied. If no differential effect is observed and thus the interaction condition is not met, the biomarker is not useful in terms of predicting patients who experience enhanced benefit. In this case the best course of action is to focus on treatment effectiveness in the overall population of patients and pursue the broad claim.
Define a Data Model
A data model specifies a scheme for generating individual patients’ data in the set of predefined samples, i.e., nonoverlapping homogeneous groups of patients, in a clinical trial. In this case study, the overall population of patients is naturally split into four samples that are defined as follows:

Sample 1 (
Placebo BioNeg
) includes biomarkernegative patients in the placebo arm. 
Sample 2 (
Placebo BioPos
) includes biomarkerpositive patients in the placebo arm. 
Sample 3 (
Treatment BioNeg
) includes biomarkernegative patients in the treatment arm. 
Sample 4 (
Treatment BioPos
) includes biomarkerpositive patients in the treatment arm.
Using this definition of samples, the trial’s sponsor can model the fact that the treatment’s effect is most pronounced in patients with a biomarkerpositive status.
For each sample in the data model, the parameters of the outcome distribution (i.e., hazard rate) defined in the following Table are listed in a single set of outcome parameters.
Sample  Median time (months)  Hazard rate 

Placebo BioNeg  7.5  0.092 
Placebo BioPos  7.5  0.092 
Treatment BioNeg  8.5  0.082 
Treatment BioPos  12.5  0.055 
The outcome parameters are specified using the following R code.
It is important to note that, if no censoring mechanism is specified in a data
model with a timetoevent endpoint, all patients will reach the endpoint of
interest (i.e., progression) and thus the number of patients will be equal to the
number of events. Using this property, it is sufficient to define the number of
patients in each sample, according to the prevalence of biomarkernegative and
biomarkerpositive patients. The prevalence of biomarkerpositive patients in
the general population (prevalence.pos
) is set to 0.55.
Finally, the data model can be set up by initializing the DataModel
object
and adding each component to it. The outcome distribution is defined using
the OutcomeDist
object with the ExpoDist
distribution. The data model
is shown below.
By default, the outcome type is set to fixed, which means that a design
with a fixed patient followup is assumed even though the primary endpoint
in this clinical trial is a timetoevent endpoint. This is due to the fact that,
as was explained earlier, no censoring is assumed in this trial and all patients
are followed until the event of interest (disease progression) is observed. In the
presence of censoring, the outcome type will be set to event
and the design
parameters, e.g., length of the enrollment and followup periods, will need to
be specified as well.
Define an Analysis Model
The analysis model in this clinical trial is very similar to the one defined in Case study 2. The only components of the model that need to be modified
are the statistical method utilized in the primary analysis (method =
"LogrankTest"
) and method for computing the effect size in the biomarkernegative
subpopulation (method = "EffectSizeEventStat"
). In addition,
the ratio of effect sizes between the biomarkerpositive and biomarkernegative
subpopulations needs to be computed. This is accomplished by specifying
a Statistic
object with the RatioEffectSizeEventStat
method. This
method computes the ratio of effect sizes for exponentially distributed endpoints.
Define an Evaluation Model
As in the evaluation model used in Case study 2, custom functions need
to be written to support the evaluation of the probabilities of the individual
claims and weighted power. Note that the definitions of broad, restricted
and enhanced claims need to be updated to account for the interaction condition
(see Summary). As an illustration, a custom function for computing weighted power (subgroup.cs3.WeightedPower
) is defined below. The statistic.result
argument in the
subgroup.cs3.WeightedPower
function is a matrix containing the value of
the effect size in the biomarkernegative subpopulation and the ratio of effect
sizes between the two subpopulations.
A similar approach can be applied to create a custom function for computing the marginal probability of a broad, restricted and enhanced claims. These functions are defined below.
The evaluation model is presented below. The influence threshold (influence_threshold
)
and interaction threshold (interaction_threshold
) are set to their optimal
values, i.e., 0.15 and 1.5, respectively, in the computation of the probabilities
of the three claims of interest and weighted power.
Perform Clinical Scenario Evaluation
Using the data, analysis and evaluation models, simulationbased Clinical Scenario Evaluation is performed by calling the CSE
function:
Download
Click on the icons below to download the R code used in this case study and report that summarizes the results of Clinical Scenario Evaluation: