Week 3: Questions ;)

Critical understanding deck

When is an experimental study more appropriate than an observational one?

When the researcher can ethically manipulate the exposure to test causality; experimental designs (RCTs) reduce confounding by randomization.

How does randomization strengthen causal inference?

It balances both known and unknown confounders across groups, making any observed outcome differences more likely due to the intervention.

Why might randomization fail to eliminate bias in practice?

Poor allocation concealment, small sample size, or differential loss to follow-up can unbalance confounders and reintroduce bias.

How does allocation concealment differ from blinding?

Allocation concealment prevents selection bias before assignment; blinding prevents measurement bias after assignment.

Why is blinding important for subjective outcomes?

Knowledge of treatment can influence how participants report symptoms or how investigators measure outcomes.

What is the trade-off between internal and external validity in an RCT?

Highly controlled studies maximize internal validity but may limit generalizability (external validity) to real-world settings.

Why is “intention-to-treat” analysis preferred?

It maintains comparability created by randomization and preserves real-world effectiveness, even with non-adherence or crossover.

When might a per-protocol analysis be justified?

When you want to estimate biological efficacy among those who fully followed the assigned treatment, knowing it may sacrifice randomization balance.

Why use a cluster RCT rather than an individual RCT?

When the intervention operates at a group level (e.g., school, clinic) or contamination between individuals would bias results.

What potential problem arises from cluster randomization?

Fewer independent units reduce statistical power and may create intracluster correlation that requires design adjustment.

Why conduct a crossover trial?

Each participant acts as their own control, reducing confounding—but only works if the disease and intervention effects are reversible.

What design feature of the Salk polio trial increased its validity?

Double-blind, placebo-controlled randomization of over a million children minimized bias and ensured reliable causal inference.

What threatens internal validity most in an RCT?

Differential loss to follow-up or non-comparable measurement across arms.

If an RCT finds no effect, how can you tell whether the study was underpowered?

Check expected effect size, α, β, and sample size; a small n or rare outcome may produce a false negative despite a real effect.

Why might an RCT with perfect internal validity still lack policy impact?

Limited external validity—participants or conditions differ from the populations or settings where policy would apply.

Why use survival analysis instead of a simple proportion of deaths?

It accounts for differing follow-up times and censored individuals, providing more accurate time-to-event estimates.

What does “censoring” mean in survival analysis?

The participant’s follow-up ended before experiencing the event (lost, withdrew, or study ended). They contribute information until censoring.

Why can’t censored individuals be counted as alive indefinitely?

Doing so would overestimate survival; they are included only up to their last known event-free time.

How is survival probability at time t estimated in a KM curve?

Multiply the conditional survival probabilities up to that point: S(t)=\prod(1-di/ni). Each event time lowers S(t).

What does a steep drop on a KM curve indicate?

A high event rate during that interval—rapid decline in survival probability.

Why are censoring marks shown as tick marks on a KM curve?

To visually separate loss of follow-up from true events; they indicate information stops, not that the person died.

How does the life-table method differ from KM estimation?

It groups time into intervals and assumes uniform event probability within each; KM uses exact event times.

What is the median survival time?

The time point where S(t)=0.5; half the population has experienced the event.

Interpret a hazard ratio (HR) of 0.6.

The treatment group has 40% lower instantaneous risk of the event compared to control at any given time.

What does HR > 1 imply?

Higher hazard (greater event risk) in the treatment/exposed group relative to control.

How does a hazard ratio differ from relative risk?

HR compares instantaneous risks over time; RR compares cumulative probabilities over a fixed period.

If survival curves cross midway, what does that suggest?

Treatment effects vary over time; hazards are not proportional—KM curves may overlap despite early differences.

Why is proportional hazards assumption important?

Many survival models (like Cox regression) rely on constant relative hazards over time for valid inference.

How is the log-rank test used?

To statistically compare two or more KM survival curves; it assesses whether observed event patterns differ beyond chance.

What does α = 0.05 represent?

A 5% chance of falsely detecting an effect when none exists (Type I error).

What does β = 0.20 represent?

A 20% chance of missing a true effect (Type II error).

If you lower α without changing n, what happens to power?

Power decreases—stricter significance thresholds require stronger evidence to reject the null.

If you increase power (e.g., 0.8 → 0.9), what happens to sample size?

It must increase; higher certainty needs more participants to detect the same effect.

How do event frequency and sample size interact?

Fewer expected events (low incidence) require a larger n to achieve the same power.

How does effect size influence power?

Larger true effects create larger group differences, increasing power; small effects are harder to detect.

Why is the observed 95% VE in the Pfizer trial associated with extremely high power?

The effect size (difference between vaccine and placebo infection rates) was so large that chance alone could not explain it—near-certain detection.

If an RCT has high power but finds no effect, what can you conclude?

The null result is likely true; with sufficient power, absence of evidence suggests true absence of effect rather than insufficient sample.