STATISTICAL CONSIDERATIONS IN THE INTENTION-TO-TREAT PRINCIPLE
John M. Lachin
The Biostatistics Center
The George Washington University
jml@biostat.bsc.gwu.edu
www.bsc.gwu.edu
--------------------------------------------------------------------------------------------------------------------------------The Intention-To-Treat Principle
Criteria for the evaluation of the benefits and risks of a new therapy
FDA reviews and NIH trials
Contrasted with Efficacy or “Per Protocol” Analyses
Settings
Multi-component Intervention trials, or stepped therapies
Hypertension, CASS
Simple Active Drug Vs. Placebo
EFFICACY VERSUS EFFECTIVENESS
Efficacy: Pharmacologic Effect
The expected outcome among appropriate patients in whom
The drug is tolerated
The patient is compliant
The agent is effectively administered (bioavailable)
Efficacy (Per Protocol) Analysis
Effectiveness: Clinical Effect
The expected outcome among all eligible patients, allowing for side-effects and
incomplete administration.
Intention-to-treat design and analysis
--------------------------------------------------------------------------------------------------------------------------------Efficacy Subset Analysis
Subset selected post-randomization based on the desired efficacy criteria
Various exclusions of patients and selected patient data
No randomized basis for comparison of groups
Placebo group often not subject to the same post-randomization exclusion processes
(e.g. drug-induced hepatotoxicity)
Exclusions  various types of bias in the analyses.
Results do not apply to initially treated population
Nevertheless useful.
No difference implies lack of pharmacologic efficacy, or a negative bias
Difference implies efficacy, or a positive bias
The Intention-To-Treat Philosophy
The overall objective of a clinical trial is to provide a valid prospective assessment of
the difference between two (or more) initial treatment assignments to the general
population of patients with respect to a clinically relevant outcome assessment at a
later point in time.
Intention-to-Treat Analysis
Comparison of the ultimate outcome between two or more populations which are initially
assigned to receive different treatments, irrespective of tolerance or compliance.
Assesses the long term effects of an initial treatment decision.
--------------------------------------------------------------------------------------------------------------------------------Validity
Statistical validity = unbiased
^
E( )  
If p = 0.05, Type I error = 0.05.
Statistically unbiased when all patients randomized are evaluated and included in the
analysis
Randomization
Necessary but alone is not sufficient.
Two other requirements are
The collection of patients analyzed should be completely randomized.
 Missing Completely at Random (MCAR)
The outcome assessments should be obtained in a like and unbiased manner for all
patients
Masking
A Sufficient Strategy:
All patients randomized are evaluated as scheduled as objectively as possible and are
included in all analyses, regardless.
Requires continued follow-up of all patients.
Example:
Randomized
50
A
50
B
100
Evaluated at end
35
25
60
Missing data
15
25
40
60 evaluated: Still completely randomized??
Analysis unbiased only if Missing Completely at Random (MCAR)
MISSING COMPLETELY AT RANDOM
Observed versus missing occurs at random,
purely by chance
Unrelated to treatment assignment
Some examples:
Entered study later
MAR ?
Yes
Moved away
Probably
Lost to follow-up
Maybe
Terminated due to
transaminase elevation
Likely not
Terminated due to
“treatment failure”
NO
Terminated due to Recovery
NO
Testing the Assumption
Compare frequency of missing data between the groups (35 in A Vs 25 in B)
Compare reasons for missing data between the groups
Compare baseline characteristics of subset observed (A: 35 Vs B: 25)
Compare characteristics of those with missing versus those with complete data
Combined groups:
Within each group:
NO SINGLE DIRECT TEST
60 Vs 40
A: 35 Vs 15
B: 25 Vs 25
Statistical Strategies for Dealing With Incomplete Follow-up
MCAR, or Censoring Completely at Random
Untestable
Missing at random (MAR):
Missing is conditionally independent of the observable value, given other measurements
within the context of a given model.
Untestable
Last Observation Carried Forward (LOCF)
Constant value imputation  incorrect variances
Model Unrealistic
Informative Censoring Models:
To some extent, require MAR
Model dependent
--------------------------------------------------------------------------------------------------------------------------------Some Examples
National Cooperative Gallstone Study
transaminasemia, early withdrawals
Tacrine for Alzheimer’s Disease
Poor compliance but continued follow-up
Diabetes Control and Complications Trial
Continued Follow-up, little non-compliance
THE DIABETES CONTROL AND COMPLICATIONS TRIAL
1441 subjects with IDDM randomly assigned to Intensive vs. Conventional Therapy
Intensive: Goal of near normal blood glucose levels using any means possible while
avoiding hypoglycemia.
Conventional: Maintenance of clinical well being with a conventional therapy.
Principal Outcome: Progression of retinopathy on fundus photographs obtained 6-monthly
and centrally graded.
Mean 6.5 years follow-up (4-9 years)
8 of 1430 patients still alive did not complete the final end-of-study assessment
32 of 1441 patients were inactive for some period, the majority later returning to follow-up.
Patients remained on assigned therapy for 97% of the time in study
--------------------------------------------------------------------------------------------------------------------------------Conclusion
Continue follow-up of all patients randomized
Until death of patient
Until study end
Unless patient refuses
Both an intention-to-treat and efficacy analyses possible as well as other analyses which
“adjust” for full compliance.
Power
Intention-to-treat power decreases exponentially when proportion non-compliers decreases.
Extreme Case: Non-compliers in experimental group have the same expected outcome as
controls.
N = Sample size required for 100% compliers
R = proportion of non-compliers in treated group
N* = Sample size required to maintain same level of power with non-compliers
R
N*/N = 1/(1 - R)2
0.0
1.0
0.1
1.23
0.2
1.56
0.3
2.04
Does the Intention-to-treat analysis have greater power than an efficacy analysis in other
cases?
---------------------------------------------------------------------------------------------------------------------------------
Intention-to-treat analysis is more powerful than an efficacy analysis when
Lingering pharmacologic effectiveness.
Treatment arrests progression.
Exponential disease progression when untreated.
A landmark analysis is performed at the end of follow-up
Example: The DCCT
DESIGN CONSIDERATIONS
Early Toxicity:
Consider a non-randomized trial period where all subjects receive active therapy for a
period of 3 months. Any free of side effects then randomized to active Vs placebo.
Competing Risks:
Consider a combined outcome so that all “competing risk events” are counted as
outcomes. Otherwise have an informatively missing problem.
Patient and Physician Education:
Emphasize the scientific importance of adherence to the treatment regimens and
complete follow-up:
Before launching the study
Before randomizing each patient
Pay a premium for compliant patients with complete follow-up
Criteria For Withdrawal of Treatment:
Safety considerations ONLY.
Not lack of efficacy
Not failure to comply
Specific a priori specified criteria
Criteria For Withdrawal From Study
Death or end of study ONLY
Drop Dropouts
Temporarily Inactive:
At patient insistence or due to external influence (relocation, imprisonment)
ONLY.
Allows patient to be fully reinstated under active follow-up, and if indicated,
randomized treatment.
Lost to Follow-up: Only assigned to those still inactive at study end.
Data Management
Management of the trial, monitoring of patient and clinic adherence requires real time
data collection and data management.
Primary Outcome Analysis
Consider a landmark analysis at the targeted conclusion of the study.
Maximal power against DCCT-like treatment effects
Allows an outcome which includes 3 states
Worse
Unchanged
Improved
--------------------------------------------------------------------------------------------------------------------------------CONCLUSIONS
The Intention-to-treat analysis is far more defensible statistically than is an efficacy analysis.
Common study designs which allow withdrawal from study do not satisfy the intention-to-treat
principle, even in so-called intention-to-treat analyses.
An intention-to-treat analysis requires an intention-to-treat design which fosters maximal
adherence to the treatment regimens and to the follow-up schedule.
In a well designed study, the differences between the intention-to-treat and efficacy analyses
should be minimal.