Chapter 5 - Clinical Trial Simulation

Chapter 5 - Clinical Trial Simulation#

General Steps	Applied to te coin toss example
Specify numerical values for:
- Simulation parameters	Probability of heads = 0.50
- Sample size	10 tosses of the coin
- Analysis/test statistics	Number of heads out of the 10 tosses
For each replication:
1. Generate data	Generate 10 random coin tosses
2. Run Analysis	Count the number of heads out of 10 tosses
3. Keep track of the performance	Keep track of 2 and repeat 1,000 times

In Clinical trial, we can’t repeat the trial like coin flip.
The Utility of simulation studies in a clinical design prcess has become very important.

Operating characteristics - description of how a design performs under various scenarios.
Scenario - List of values specifying the ‘true’ underlying parameters, such as true response rates, recuritment rate, etc.
Simulation - Repeated analyses of randomly generated datasets drawn under specific assumptions.
Simulation report - Written documents that provides the explanation, details, and justification of the simulation study.
Trial design - Details about how the trial is conducted.
Virtual trial - Clinical trial being simulated in computer.

The goal of cilnical trial sumulation is to make virtual trial design match as closely as possible how the trial will be conducted.
General steps of simulation for an adaptive clinical trial would involve: ** Step 1: Simulate the arrival time of the next patient. ** Step 2: Check the stopping rules to see if the trial should be stopped. ** Step 3: If the trial not stopped, enroll more patients into the nexy interim analysis. ** Step 4: Simulate the patient outcome as a fuction of the treatment. ** Step 5: Repeat steps 1-4

Simulation-guided trial design encourages evaluating multiple design options early to ensure the most efficient and ethical choice is made for each clinical question.
Adaptive designs can improve trial success by anticipating potential regrets and adjusting accordingly, serving as a form of ‘insurance’ against trial failure.

Simulation studies help explore the trade-offs between trial designs beyond average behavior.
Rather than recommending one design, a Bayesian adaptive approach is used to review different metrics.

No.	Trial designs	Scenarios
1	Equal randomisation	Equal treatment effect/response rate (Null)
	- 1:1 allocation to E and S	- pS = 0.3 and pE = 0.3
2	RAR with no minimum/maximum allocation	10% higher response rate in E
	- At interim analysis, adapt the	- pS = 0.3 and pE = 0.4
	allocation probability for E to
	probability that E has higher
	response rate than S (pE)
	- Allocation of S would be 1 − pE
3	RAR with minimum/maximum allocation	20% higher response rate in E
	- Same as Design 2 with 10% minimum	- pS = 0.3 and pE = 0.5
	and 90% maximal allocation
	probabilities, to ensure that the
	allocation does not get ‘stuck’
	on one study arm

Patient enrollment followed a Poisson process with interim analysis every 10 subjects after a 30-subject burn-in.
Response probabilities:
- pS ~ Beta(0.2, 0.8)
- pE ~ Beta(0.2, 0.8)
Stopping rules:
- Superiority threshold = 0.995
- Futility threshold = 0.005
- Ensures Type I error < 3.3% under null

Figures display randomization probabilities over time for virtual trials
Design 2 shows large imbalance in some trials, e.g., 172:26 (SOC:E) in null scenario
Design 3 mitigates imbalance via min allocation threshold
In Scenario 2 and 3, some trials assigned more to inferior treatment → RAR risk
P20 exposes ethical risk of RAR misallocating patients
Even fixed randomization (Design 1) can show imbalance without blocked randomization
Design evaluation must go beyond averages — include worst-case risks

Essential for adaptive and complex trials
DIA-ADSWG recommends including:
- Objectives, inputs, design rules
- Scenario definitions, operating characteristics
- Discussion and justification of chosen design (Table 5.4)

Commercial: EAST®, ADDPLAN®, COMPASS®, FACTS
- Advantages: ready-to-use, no coding required
- Limitations: cost, limited flexibility for custom designs
Open-source (R-based):
- Notable packages: rpact, MAMS, Mediana, multiarm, adaptr, OCTOPUS, gsDesign
- Many offer vignette tutorials
- Available on CRAN and GitHub

Simulation is critical to evaluate trial design operating characteristics
Adaptive designs may improve ethics but carry trade-offs
No universal best design — simulation helps identify the most efficient and ethical option per research question