Clinical trials
Test knowledge
- Randomization: equal distribution of factors in both groups, don’t have to present p-
values of differences between baseline characteristics, because differences occur by
chance. It prevents confounding bias. If groups are randomized, you don’t have to
test if there are statistically significant differences between the groups (if there are,
they’re based on chance).
- Stratification: randomization based on specific factors. equally distribution of a
certain factor. high weight equally distributed in both groups
- Attrition bias: form of selection bias, chance that patients leave the study depends
on the group they are in??
- Intention to treat analysis: analysis according to randomization, confounders
remain equally distributed. It provides a pragmatic estimate of the benefit of
intervention. Underestimates efficacy
- Per protocol analysis: included only those participants who completed the protocol
for their allocated treatment. Overestimation of efficacy, because patients who leave
or change groups are excluded from analyses, while they could have left due to side-
effects or something like that.
- Power calculation: required difference is referred to the smallest effect of clinical
interest, effect that you can show with enough power. The higher the power, the
more patients you need. Clinical relevance is not the same as statistically relevance.
Trial is overpowered if there are more patients than needed. Unethical. Have to
adjust for drop-out rate.
- Primary and secondary outcomes should be declared before the start of the study.
- Surrogate outcome: measure something of the mechanism?
- Composite outcome: all cause mortality not due to primary outcome? e.g. all cause
mortality.
Lecture 1: Randomization, baseline differences & power
Issue in causal research: confounding. You should know the confounder, measure
the confounder, and adjust for the confounder, but confounders are often
unknown or difcult to measure → incomplete protocol, invalid inference.
No statistical method can achieve comparability on unknown/unmeasured factors in analysis
phase.
Comparability of prognosis (due to confounding by indication: differences in prognosis),
comparability of extraneous effects (change behavior), comparability of observation
(subjective). Adjust for baseline incomparability:
- Orthodox view: randomization assures comparability for known and unknown factors.
If you only adjust for known factors and forget unknown factors, it biases the result.
Don’t adjust for bias/unequal distributed factors.
- Pragmatic view: identify baseline variables of prognostic value on basis of all
available evidence and adjust for everything you know (covariance/regression).
Unknown variables might become unbalanced. If there is a difference present both
(adjusted as secondary analysis)??
1
,Randomization: Each group is a random sample of eligible patients, so both are
representative of that same population → they are prognostically comparable.
Inadequate randomization of chance variation (sampling error; smaller chance in
bigger trials) can cause imbalances in baseline variables between randomized
groups → diferent prognosis between the groups → confound the outcome.
Prevent imbalances using:
- Solid randomization
- Large sample size, stratified
- Stratified randomization (on variables/scores)
- Random permuted blocks: allocated sequences; imbalance can never be substantial.
No allocation concealment > also change length of blocks, combine with stratification
otherwise only treatment randomized
- Minimization: solution for smaller trials, people are evenly distributed.
Power
- Type 1 error (5%): there is no effect of treatment, but we say there is (false positive)
- Type 2 error (20%): there is an effect of treatment, but we say there is not (false
negative)
- Power: 1 - type 2 error = sensitivity (true positive: say there is effect when there truly
is)
- Power always calculates the number of participants per arm
- The smaller the difference you want to show, the larger the population
- Lower numbers of participants needed if the outcomes are continuous
- Calculate power on the outcome that is most relevant
Superiority trials: aim is to show that treatment A is better than B.
COMET database → core outcomes (agreed standardised set of outcomes that
should be measured and reported, as a minimum, in all clinical trials in specifc
areas of health(care). Core outcome sets increases consistency across trials,
maximise potential for trial to contribute to syst. review of these outcomes, more
likely to measure appropriate outcomes, major reduction in selective reporting.
Efficacy: does it work when you apply it
Effectiveness: does it work when you apply it in general practice
2
, Carry-over effect: patients are used for treatment A and re-used for treatment B, so
treatment A can influence the effect of treatment B.
2x2 factorial, cross-over and comparative effectiveness designs
2x2 factorial designs: randomize over 2 treatments at the same time (mostly used in
experiments or drug trials)
- Aim: to evaluate 2 interventions compared to a control in a single experiment
- Possibility of interaction
- Efficient use of money, time, personnel, patients
- Three test: efect factor A, efect factor B and interaction → difcult
statistical analysis
- A, B, A+B and placebo (e.g. aspirin (+control), b-carotene (+control), aspirin+b-
carotene, placebo/standard practice)
- Advantages: if no interaction, 2 experiments at the same time with less patients +
can examine interaction if this is of interest
- Disadvantages: difficult to randomize (within columns and within rows) = complex,
potential for adverse effects (side effects due to a combination of both drugs)
- Analyze it using HR, RR, OR etc. → create 2x2 tables frst
Cross-over design: Same person gets both treatments (only used in chronic conditions,
e.g. phase 2 trials)
- Subjects are randomized to sequences of treatments, so in a way that all
persons get both treatments: (run in period → same baseline) A→ wash out
→ B, or B→ wash out → A
- Advantage: SD is smaller, because each person is his own control, need
less patients. Treatment comparison is only subject to within-subject
variability (not between-subject variability → reduced sample sizes)
- Washout-period (time between treatment periods) to prevent carry-over effects:
treatment effect in first period affects outcomes in 2nd period.
- Period effect: effect in period differs (e.g. study infectious diseases in different
seasons). Study this effect by calculate both treatment sequences.
- Appropriate for conditions that are expected to return to baseline levels after one
treatment (e.g. in surgery this design isn’t possible)
- Assumptions: no carry-over effects (difficult to test)
- Disadvantages: strict assumptions about carry-over effects, inappropriate if you have
irreversible situations (e.g. surgical procedure)/acute diseases, drop-outs before
second period (requires compliance to study procedures for long-time period),
3
