Академический Документы
Профессиональный Документы
Культура Документы
VIEWPOINT
Challenges of Non–Intention-to-Treat Analyses
David L. DeMets, PhD Over the past 5 decades, the randomized clinical trial to the extent possible, all participants should be followed
Department of has become the gold standard for evaluation of the risks up for the full duration of the study regardless of their ad-
Biostatistics and and benefits of new interventions, including drugs, medi- herencetoassignedtreatment;“offtreatment”shouldnot
Medical Informatics,
cal devices, and surgical procedures.1 imply “off study.” In placebo-controlled trials in which it is
University of
Wisconsin–Madison, To justify the use of randomization, it is important likely that nonadherence to assigned treatment “dilutes”
Madison. to note that in a nonrandomized study comparing 2 in- the effect of the treatment, study power may be reduced
terventions, a small P value (of which P < .05 is gener- unless sample sizes are increased. However, in active-
Thomas Cook, PhD ally considered statistically significant) for a statistical controlled trials, nonadherence may not always result in
Department of
comparison between groups can be due to 1 of 3 sources: such dilution; for example, nonadherence in the control
Biostatistics and
Medical Informatics, chance, causation, or confounding. Because random- group may actually increase the observed difference be-
University of ized assignments cannot be associated with partici- tween groups.
Wisconsin–Madison, pant characteristics, effective randomization elimi- Stratifying the analysis by predefined baseline co-
Madison.
nates the third possibility, enabling direct assessment of variates is a standard analysis, but the stratification must
potential causal relationships provided that the study not depend on outcomes or experiences of a participant
is designed, conducted, and analyzed properly. that occur after randomization. A few examples illus-
Proper trial conduct and analysis include ensuring trate the problem. One of the earliest randomized clini-
complete follow-up and unbiased ascertainment of the cal trials in cardiology, the Coronary Drug Project, which
outcomes of interest among all randomized partici- compared 5 cholesterol-lowering drugs with a placebo
pants. The exclusion of participants or outcomes from the group, followed the ITT principle for its primary analyses
analysis, because outcomes either are missing or are delib- but illustrates how analysis based on adherence can yield
erately omitted from the analysis, reintroduces the third results that are clearly not interpretable.3 Participants in
possibility—thatobserveddifferencesareduetoconfound- the placebo group were divided into those who took at
ing.Theintention-to-treat(ITT)principleformalizesthisidea least 80% of their assigned medication and those who did
and that, up to the desired level of statistical significance not. Analysis of those 2 groups strongly indicated that the
(eg, P < .05), observed differences between groups are former had better mortality results than the latter, which
due to the causal effect of assignment to the intervention. is highly unlikely to be due to a causal effect of placebo.
The P value or corresponding confidence interval already This and other similar analyses clearly demonstrated that
accounts for any chance baseline covariate imbalances. adherence is associated with underlying risk and should
In trials for which all randomized participants are fully be considered an outcome rather than a predictor.
adherenttotheirassignedintervention,thereisagreement The Anturane Reinfarction Trial4 was a randomized
that use of ITT for the analysis accurately represents the clinicaltrialthatcomparedthedrugsulfinpyrazonewithpla-
effects. However, when not all participants are adherent cebo in a population of patients with recent myocardial in-
totheassignedintervention,investigatorsmaybetempted farction. Eligibility for analyses, predefined in the protocol,
to try to correct for this by use of “as-treated” analyses— requiredthatparticipantsadheretotheirassignedinterven-
analyses that reassign participants to the intervention that tion for at least 7 days; however, the application of this rule
theyactuallyreceived;“per-protocol”analyses—thosethat varied by how eligibility was assessed or who made the
exclude participants with adherence below a threshold; or assessment.2 In this case, classification of eligibility was
“on-treatment” analyses—analyses that terminate follow- done after the trial ended. when unblinding or partial un-
up at or shortly after discontinuation of the intervention. blinding was possible. Furthermore, exclusions based on
Because each of these approaches violates the ITT prin- postrandomizationmeasurementsreintroducedconfound-
ciple, each is subject to confounding, and there is no rea- ing and thereby subverted the randomization. In addition,
son to believe that such approaches can correct the “prob- notalleventsthatoccurredinthefollow-upperiodwerein-
lem.” The underlying problem is in the conduct of the trial, cluded.Includingornotincludingalleventsandallpatients
not in the analysis. It is impossible to reliably recover the in their randomized groups resulted in a large shift in the
causal effect of full adherence from a trial in which full ad- P values from significant (ie, P < .05) to nonsignificant.
Corresponding
herence is not achieved. Similarly, in the APPROVE trial,5 which was a trial of
Author: David L.
DeMets, PhD, Because the degree of potential confounding when rofecoxib vs placebo for cancer prevention, the idea that
Department of outcomes are missing cannot be determined (the re- “off treatment” implies “off study” was written into the
Biostatistics and quired data are by definition missing) the only statisti- protocol; follow-up ceased 14 days after participants
Medical Informatics,
University of
cally sound approach is sensitivity analyses, which in- withdrew from their assigned intervention, and cardio-
Wisconsin–Madison, volve assessing the robustness of the result to a range of vascular outcomes occurring 14 days after participants
610 Walnut St, WARF plausible mechanisms responsible for the missing data. stopped were unreported. Subsequently, a full year of
Building, Room 201,
May et al2 have described problems arising from fail- additional follow-up was conducted and vital status was
Madison, WI 53726
(demets@biostat ure to follow the ITT principle and noted implications for obtained for more than 95% of randomized partici-
.wisc.edu). study design and conduct. For example, ITT requires that, pants. The original “on-treatment” analysis suggested an
ment plus 30, 60, or 180 days for the placebo group. Assuming that
Figure. Cumulative Mortality Assessed by Intention-to-Treat Analysis
and by Variations of “On-Treatment” Analyses for the EVEREST Trial the placebo has no effect on mortality, the ITT estimate should rep-
resent the true mortality risk, whereas each variation of end of treat-
0.4 ment shows lower mortality. The most plausible explanation of this
result is that censoring is informative—that is, associated with
mortality—and the on-treatment estimates of cumulative mortality
0.3 are biased. More important, on-treatment comparisons of mortality
Cumulative Mortality
ARTICLE INFORMATION REFERENCES 5. Baron JA, Sandler RS, Bresalier RS, et al.
Published Online: December 17, 2018. 1. Friedman L, Furberg C, DeMets D, et al. Cardiovascular events associated with rofecoxib:
doi:10.1001/jama.2018.19192 Fundamentals of Clinical Trials. 5th ed. New York, NY: final analysis of the APPROVE trial. Lancet. 2008;
Springer Science + Business Media LLC; 2015. 372(9651):1756-1764.
Conflict of Interest Disclosures: Dr DeMets
reported receipt of grants from the National 2. May GS, DeMets DL, Friedman LM, et al. The 6. Yang F, Wittes J, Pitt B. Beware of on-treatment
Institutes of Health and personal fees randomized clinical trial: bias in analysis. Circulation. safety analyses [published online November 16,
from Frontier Science, D. L. DeMets 1981;64(4):669-673. 2018]. Clin Trials. doi:10.1177/1740774518812774
Consulting LLC, Actelion, the Duke Clinical 3. Coronary Drug Project Research Group. 7. Scirica BM, Bhatt DL, Braunwald E, et al.
Research Institute, the Population Health Influence of adherence to treatment and response Saxagliptin and cardiovascular outcomes in patients
Research Institute [Canada], AstraZeneca, of cholesterol on mortality in the Coronary Drug with type 2 diabetes mellitus. N Engl J Med. 2013;
GlaxoSmithKline, DalCor, LivaNova, Boston Project. N Engl J Med. 1980;303(18):1038-1041. 369(14):1317-1326.
Scientific, Medtronic, the Patient-Centered 8. Konstam MA, Gheorghiade M, Burnett JC Jr,
Outcomes Research Institute, Amgen, 4. Anturane Reinfarction Trial Research Group. The
Anturane Reinfarction Trial: reevaluation of et al. Effects of oral tolvaptan in patients
Merck, and Sanofi. No other disclosures hospitalized for worsening heart failure:
were reported. outcome. N Engl J Med. 1982;306(16):1005-1008.
the EVEREST outcome trial. JAMA. 2007;297(12):
1319-1331.