Alec WalkerSeptember 2014
Core Characteristics of Randomized Clinical Trials
Preplanned Analysis
Goals Primary Secondary Strategy for unanticipated resultsStudy size Statistical power Stopping rules
2
Preplanned Data Collection
Baseline Characteristics of each participant’s history Concomitant illnesses Diagnostic tests and procedures Medical examinationTreatment Dose, route, frequency, duration, timingEndpoints Symptoms and tests required for diagnosisSafety Adverse outcomes Contemporaneous assessment of causality
Entry Criteria
Treated disease Definition Severity, PrognosisOther health conditions Concomitant diseasesOthers Demographics – age, sex, raceImplicit Criteria Populations served by participating clinical sites See also Informed Consent laterGoals 1o Clarify the comparison 2o Generalizabilty to target population
Comparison of Results in Groups
Groups Anecdotes, however persuasive, are set aside Frequency of outcome is the measure of effect We are examining net effects
Improved as a result of treatment No effect of treatment Deleterious effect of treatment
Comparisons Substitution of another person’s experience for the
impossible “What if?” question. (Counterfactual: What would have been the treated person’s experience if there had been no treatment?)
Informed Consent
Introduced for ethical reasons Patients should be aware that they are participating
in an experiment Actively agree to enter
A subtle selection criterion Language skills Education Trust in the medical care system Inclination to follow directions
Random Assignment of Treatments
“Coin flip” metaphor Mechanical process Assignment not systematically associated with any
patient characteristics At the discretion of the investigator:
Number of compared treatments Allocation ratio Blocking
Effects Expectation of similar outcomes between groups
under the Null Hypothesis Justification for the calculation of p-values
Randomization
Treatment allocation is determined by a process That generates
An expectation of zero correlation between
treatment and predictors of outcome. The Predictors may be
Known or unknown to the experimenter Measured or unmeasured Measured poorly or well
Balance
All characteristics other than treatment are balanced in expectation Measured and unmeasured Predictors and correlates of predictors The intermediate states that later arise from these
Balance
All characteristics other than treatment are balanced in expectation Measured and unmeasured Predictors and correlates of predictors The intermediate states that later arise from these
Unadjusted estimates are unbiased estimates of treatment effect Differences, ratios, more complex functions of Risk, rates, hazards, survival, … Costs, QoL, … Even of dependent happenings, like epidemics
(provided that exposure groups are not intermixed)
Treatment Adherence
Commitment from patients Encouragement from staff Monitoring
Pill counts Blood level
Dedicated Outcome Data Collection
Disease prespecified Expert consensus on diagnosis
Symptoms Signs Diagnostic testing
Recurrent monitoring
Limited Follow-up
Need to get drug to market For chronic conditions, no amount of follow-up will
reproduce ultimate conditions of use Surrogate outcomes
Examples Control of blood pressure or HbA1c Patient-reported outcomes
Desiderata Well established correlates of clinically important Generally not important clinically in themselves Manifest earlier
Real clinical outcomes can be addressed later
N Engl J Med. 2010 Apr 1;362(13):1192-202
Bala
nce
Delta = Treatment Effect
Delt
a =
Tre
atm
en
t Eff
ect
Differential outcome identification?
Protocol provided similar surveillance Scheduled visits Scheduled biopsies
Work-up of symptoms was inevitably differential because the symptomatic outcomes were differential Urinary retention, urinary tract infection, BPH surgery Presumably differences in symptoms that fell short of
study outcomes The protective association with cancer was small and
much smaller than corresponding protective association with noncancer outcomes – the algebra of detection bias works out with plausible assumptions
This same problem will often be part of the mix of uncertainty in observational studies.
18
Differential outcome identification?
Years 1 & 2 Years 3 & 4
Treatment | Grade 5-7 8-10 5-7 8-10
Finasteride 417 18 223 12
Placebo 558 17 274 1
19
The authors raise the possibility that early detection in the placebo group may have reduced the risk of more advanced prostate cancer in those patients in later years.
What should we consider to be the treatment effect?
Does dutasteride increase the risk of advanced prostate cancer?
Problems Solved, Problems Remaining
Randomization in RCTs provides the gold standard for inference No hypothesis of confounding Frequentist interpretation of measures is supported
by the structure of the trial RCT populations may be atypical
In baseline characteristics In adherence to therapy In care of follow-up
RCT follow-up may be short
20