Lecture 9: Randomized Controlled Trials (RCTs)

Counterfactual Ideal and RCTs

• The lecture discusses advanced topics in randomized controlled trials (RCTs).

• RCTs are essential for determining whether a treatment is beneficial or harmful.

• The counterfactual ideal involves comparing a person both exposed and non-exposed to a treatment simultaneously, which is impossible in reality.

• RCTs aim to approximate this ideal by selecting a non-exposed group that closely resembles the counterfactual comparison. Methods include randomized experiments, cohort studies with matching, propensity score matching, inverse probability of treatment weighting, and crossover studies.

• Main concern: ensuring comparability between treatment groups.

Phases of Drug Trials

• Drug trials are common for evaluating new drugs against standard treatments or placebos.

• Several phases of drug trials exist:

  • Phase 0: Very small number of participants, short duration; focuses on kinetics, dynamics, and micro-dosing.

  • Phase 1: Slightly larger sample size, longer duration; assesses tolerability and safety in healthy volunteers.

  • Phase 2: More participants, longer follow-up; aims to establish proof of concept and therapeutic effect in patients.

  • Phase 3: Large sample size; seeks significant proof of efficacy for therapy approval by regulatory bodies like the FDA or EMA.

  • Phase 4: Post-marketing studies; evaluate effectiveness in real-world settings, monitor rare side effects, and assess long-term harm.

• Key difference: Phase 3 focuses on efficacy under ideal conditions, while Phase 4 focuses on effectiveness under normal routine conditions.

• Efficiency: Minimizing resources for interventions with known efficacy and effectiveness (can also be a goal of Phase 4).

Superiority Trials

• Superiority trials, or überlegenheitsstudie, are classic RCTs that aim to demonstrate the superiority of a new treatment over a standard treatment or placebo.

• They start with the assumption of equipoise, meaning uncertainty about whether a treatment is beneficial.

• Null hypothesis: there is no treatment effect.

• Alternative hypothesis: there is a difference, treatment affects the disease, either beneficial or harmful (two-sided).

• The goal is to find evidence to reject the null hypothesis in favor of the alternative.

• Some trials are designed as one-sided, assuming the new treatment cannot be harmful, which requires fewer participants but is often criticized.

Equipoise

• Equipoise is necessary to initiate a trial and provides an ethical basis for allocating patients to different treatment arms.

• It requires true uncertainty about a treatment's benefits and absence of evidence suggesting any treatment is superior.

• Conducting a trial without equipoise would be unethical.

RCT Designs

• Parallel group design: classic two-arm (or three-arm) design.

• Crossover design: participants receive both active treatment and placebo in different periods with washout periods in between.

  • Advantage: Fewer patients needed because each participant serves as their own control.

  • Disadvantage: More complex to conduct due to washout periods and potential adherence problems and unintended crossing over.

• Factorial design: tests two hypotheses simultaneously (e.g., 2x2 factorial design).

  • Example: Physicians Health Study testing aspirin and beta-carotene.

    • Hypothesis 1: Aspirin reduces mortality.

    • Hypothesis 2: Beta-carotene reduces cancer incidence.

  • Efficient but requires a large number of participants.

• The choice of design depends on the research question, number of participants available, budget, and feasibility.

Randomization, Allocation Concealment, and Blinding

• Randomization: Every participant has an equal chance of being allocated to one treatment arm or another.

  • Ensures similar characteristics (measured and unmeasured) in both groups.

  • Prevents conscious and unconscious bias by physicians and patients.

  • Minimizes confounding and creates comparable groups.

  • Not the same as random sampling.

• Unmeasured confounding: Factors that are not routinely or easily assessed (e.g., genetics, social support) but could influence results.

Example: Political attitude

• Allocation concealment: Hiding the allocation sequence from those assigning participants to intervention groups.

  • Inadequate concealment is a leading cause of bias in clinical trials.

  • Methods: Computerized telephone systems, blocking, stratification.

  • Goal: Balance but not predictability.

  • Avoid using non-re-sealable envelopes in former times, subject to tampering to manipulate sequence in smaller experiments.

• Block randomization: Maintaining balance among groups using blocks of participants.

  • Permuted blocks of the same length can be predictable.

  • Attributed blocks with variable length are less predictable but can lead to unbalanced groups if the trial stops early.

Stopping Trials Early

• Reasons to stop a trial early include:

  • Significant treatment effect seen in an interim analysis (beneficial or harmful).

  • Dangerous or unbearable side effects.

• A Data and Safety Monitoring Board (DSMB) conducts interim analyses and makes decisions about continuing or stopping the trial.

Cluster Randomization

• Cluster randomization: Randomizing groups of subjects (e.g., primary care practices, schools, hospitals) rather than individual subjects.

  • Used for interventions that cannot be directed at individuals.

  • Avoids contamination.

  • More complex design and analysis.

Stepped Wedge Design

• Stepped wedge design: Clusters are randomly allocated to treatment arms, with a gradual introduction of the intervention over time.

  • Addresses the issue of preparing all clusters to start the intervention simultaneously.

  • Clusters are divided into groups, and the intervention is introduced to each group in a staggered schedule until all groups receive it.

Blinding

• Blinding: Concealing the treatment assignment from participants and/or investigators.

  • Single blinding: Only the patient is unaware.

  • Double blinding: Both patient and physician/investigator are unaware.

  • Triple blinding: Patient, physician, investigator, and statistician are all blinded.

• Blinded outcome assessment: Assessor of the outcome is unaware of the treatment assignment to avoid bias (observation bias).

  • For example: using a blinded assessor for ECGs to determine myocardial infarction.

• Blinded trials are more complex due to procedures for immediate unblinding in emergencies.

• Blinding may not always be possible (e.g., surgical vs. non-invasive procedures).

• The use of placebos is essential to implement blinding

Placebos

• Placebos are essential to blind participants in case intervention cannot be blinded with surgical, lifestyle or other external factors involved.

• Placebos in drug trials must look and taste like the real thing.

Sample Size Calculation

• Sample size to deliver definitive results.

• Reduces the role of chance.

  • A solid sample size calculation is the basis to judge how realistic it is to conduct the trial successfully.

• Key sentence: A clinical trial should be designed to deliver definitive results, whether positive or negative.

• Variables required for sample size calculation:

  • Research question and assumed benefit of the intervention.

  • Estimated effect size (clinical difference you want to demonstrate).

  • $Power needs to be at \ge 80\%$.

  • Alpha (type one error); typically \< 0.05.

  • Estimated number of events per year (deaths, MIs, strokes, etc.).

  • Realistic estimation needed factoring literature, experts etc.

  • Realistic assumptions about dropouts.

• Power: The probability of detecting a true effect if it is present.

• Alpha: The probability of erroneously detecting a treatment effect.

Example : Treatment X leads to survival of three years
$Power = 80\%$, $\,\alpha = 0.05$, 15 patients per arm.
Reduced to Treatment X --> $1.5$ years increases patient load to $200$ Patients

• The formula to choose depends on the design and on the analysis plan.

• If realizing it won't be doable lower event rate or modify the treatment or experiment goal.*

Analysis of Superiority RCTs

• Intention-to-treat analysis: Once randomized, always analyzed in the treatment arm to which they were assigned, regardless of adherence. This prevents confounding and reflects the intention to treat participants.

• Per-protocol analysis: Only analyzing those who adhered to the protocol, which can lead to biased results as it messes up randomization schemes.

  Adherence

• Adherence means the extent to which a person's behaviour, taking medication, following a diet, executive lifestyle changes, corresponds with agreed recommendations from a healthcare provider.

• Reasons why people do not adhere - Side effects, complex therapy, withdraw of consent, rapid worsening of conditions.

• Good communication can prevent this and help with adherence.

Non-Inferiority and Equivalence Trials

• The lecture introduces non-inferiority and equivalence trials, which are used when it is not ethical or feasible to conduct a superiority trial with a placebo arm.

• Non-inferiority Trial

  • Designed to show that a new treatment is not unacceptably worse than the current standard of treatment.

  • Only requires a smaller benefit compared to the reference product instead of proof of the best available product.

    • Example transcatheter aortic valve replacement is non-inferior to surgical aortic valve replacement.

• Equivalence Trial

  • Is designed to show if two treatments are equally effective.

    • a new anti-hypertensive drug can be equivalent to standard therapy. Results can differ.

Superiority Graphical Display

• Displaying the result in graph form.

  • If the treatment is better or the standard is better with the one in the confidence interval or the null in terms of effect.

  • If it is superior including the confidence limits. It is all on the right side.

Non-Inferiority Statistics

• Need to define a delta margin. The delta margin needs to be defined that is only acceptable.

• Is all one-sided

• H0 = Inferiority

• H1 = Non-inferiority.

Analysis on Confidence Interval.
Analysis only to be claimed if the lower bound confidence interval of the treatment is less than the predefined minus delta.

Equivalence Statistics Graphical Displays

• Have to define your margins on both sides. If your results, including the confidence limits, stay within theses two margins than your equivalent.

• Interpretation can be more complex.

• Needs a lot of Patients.
  The Delta is a ethical question. How much mortality are we ready to trade?

  Superiority, Statistics, Risks and problems

  *Risk of incorrectly claiming non-inferiority exposing all the patients to the risk of receiving an not efficient therapy.

Both results should be per-protocol tested.