Translumina.in

The pros and cons of noninferiority trials Medical Statistics Unit, Department of Epidemiology and Population Health, London School of Hygiene andTropical Medicine, Keppel Street, London, WC1E 7HT UK Noninferiority trials comparing new treatment with an active standard control are becoming increasingly common. This article discusses relevant issues regarding their need, design, analysis and interpretation: the appropriate choice of control group, types of noninferiority trial, ethical considerations, sample size determination and Received 12 January 2001;revised 1 March 2001;accepted 13 January 2003 Correspondence and reprints:[email protected] (or not) of a noninferiority trial strategy will depend on the particular circumstances. While general guidance The term ‘noninferiority trial’ is commonly used to refer can be given, the relative merits of noninferiority active to a randomized clinical trial in which a new test control trials or placebo-controlled trials aimed at treatment is compared with a standard active treatment demonstrating superiority for evaluating any specific rather than a placebo or untreated control group. A prior new treatment rests on a complex of issues requiring judgement is made, that for the new treatment to be of wise judgements and continued open debate.
merit it only needs to be as good as the active controlregarding appropriate outcome measure(s) of response.
C H O I C E O F C O N T R O L G R O U P : While the superiority of the new treatment over active control would be an added (perhaps unrealistic) advant-age, the clear demonstration of noninferiority in one or One starting principle is that no patient is denied a more specific criteria of patient response is the desirable known effective treatment by entering a clinical trial. An goal which motivates such a trial. The term ‘equivalence equally important principle is that the degree of scientific trial’ is sometimes used in this context but does not rigour adopted in the evaluation of a new treatment is reflect so well the (usually) one-sided nature of this sufficient to prevent any ineffective, unsafe or inferior noninferiority question, and is implicitly dismissive of the treatments obtaining regulatory approval or gaining desirable option that the new treatment could actually widespread use. Both principles highlight the ethical be superior to the active control treatment. So the more responsibility of a society and its medical researchers to appropriate term ‘noninferiority’ is used hereon.
facilitate the best possible health care at present and also The aim of this article is to present a balanced view of the role of noninferiority trials in the development of safe The first principle is more easily grasped because it and effective new treatments. This involves a mix of relates immediately to the individual rights of the next ethical, scientific, statistical and practical considerations.
patient. Of importance here is the distinction between a This article elucidates some of the pros and cons of treatment known to be effective, and one thought to noninferiority trials and offers some pointers on how be effective, hoped to be more effective, believed to to enhance their public health value. The desirability be effective or in widespread use without evidence of Ó 2003 Blackwell Publishing Fundamental & Clinical effectiveness. Arguments against the use of placebo This ongoing dilemma for clinical trials research can controls are put forward because treatment practice be summarized by the wish to avoid two types of error: involves other active treatments. The question is: how a type I error would be the acceptance of a useless convincing is the evidence that such active treatments treatment into widespread use, and one needs to are better than placebo in aspects that genuinely benefit consider the increased risk of this error occurring by patient welfare? While one needs to consider the not using placebo controls and instead pursuing a understandable wish to do something positive for every noninferiority (equivalence) trial design with an active patient, one needs to draw a clear distinction between control group. The consequences of such an error will desire for benefit in a supposedly active potential control depend on the nature of the treatment and disease. If the treatment and hard evidence of benefit from previous ineffective treatment has substantial side-effects then great harm could ensue, if it is expensive then it detracts So, what is the extent of evidence? Is it ‘proof beyond from more fruitful use of health care costs, if it is a safe, reasonable doubt’ of patient benefit derived from several useless, cheap ‘placebo’ for a minor condition then large studies generalizable to the relevant patient pop- ulation or is it just one or two statistically significant Even if there is an effective active control treatment, results, perhaps on short-term studies of limited size there can still be problems in the design, conduct, studying surrogate end points rather than overall patient analysis and interpretation of a noninferiority trial that benefit? P < 0.05 for a treatment difference in a clinical could lead to such a type I error. Such problems are trial, does not equate with proof of effect.
outlined in the rest of this article.
Even the relevance of P < 0.0001 for a treatment A type II error is the failure to use an effective active difference can be questioned on several grounds: might control treatment by adopting a placebo control group the trial have been biased in some aspect of its design or instead. As expressed above, the degree of certainty with analysis; were the patients, the delivery of treatment, the which such an error occurs depends on the extent of outcome measure used and the length of follow-up prior knowledge that the active control is truly effective.
sufficiently relevant to normal clinical practice and In addition, the severity of this error will depend on patient benefit; was the absolute magnitude of benefit particular circumstances. At one extreme it would be sufficiently large taking account of any adverse side- absolutely intolerable to deny a known effective agent effects of a treatment; how many trials were performed that reduces mortality in a rapidly progressing life- and on how many patients. That is, in sizing up the threatening condition. However, in a more minor evidence that a pre-existing active treatment is superior ailment in which recovery often happens on placebo or to placebo one needs to exercise one’s constructive no treatment, the denial of a known active agent in a critical faculties when appraising its clinical trial short-term placebo-controlled trial, after which all evidence for internal validity, external validity, overall patients can go on to receive active treatment, has patient benefit and extent of research.
much less serious consequences for patient welfare.
Should the overall evidence for patient benefit on a In many trials, having a placebo control group does pre-existing active treatment be less than totally con- not that mean such patients receive no active interven- vincing, then the dangers of exclusively adopting that tion. Often, all randomized patients undergo normal treatment as an active control group (instead of a accepted care, including other active drugs as appropri- placebo control group) for the evaluation of other new ate, but the addition of a new treatment is compared treatments are substantial. In certain areas, this problem with addition of a placebo. It is often debated whether is tackled by having both an active control group and a such ancillary care and supplementary drugs should be placebo control group, so that noninferiority compared according to a fixed protocol or pragmatically left to with the former and superiority over the latter can be individual clinical judgement as in routine clinical evaluated within the same trial, or set of trials.
practice. It is also relevant to ask, might the patient It is important to recognize that approval of a drug by have got the active control treatment if they were not regulatory authorities as being safe and effective for a included in a clinical trial. In some circumstances the specific condition does not in itself imply that the use of answer is ‘no they would not’, in which case there that drug as an active control without a placebo group appears a perverse twist in the ethical argument would provide a reliable basis for a noninferiority trial of whereby trialists are required to adopt more stringent a new drug, (see Temple and Ellenberg 2000).
ethical standards than regular treating physicians.
Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490 The pros and cons of noninferiority trials Thus, every time one chooses between active controls (1) The new treatment has less side-effects. For instance, or placebo controls in planning a randomized trial, one low dose aspirin vs. anticoagulation following thrombo- has to consider the risks of type I and type II error, lysis after a myocardial infarction may be equivalent as taking account of the likelihood of them occurring and regards recurrence of infarct or cardiac death, but the the severity of the consequences. The consequent deci- former produces less bleeding complications. Aspirin is sions are not easily taken: neither passionate one-sided hardly a new drug! However, in the context where ethical arguments against placebo controls in general anticoagulation had become the norm it was a new nor scientific pleas for mandatory up front demonstra- tion of new treatment superiority over placebo should (2) The new treatment is less invasive. Carotid endarter- dominate this thinking and planning. Rather one aims ectomy is a surgical procedure for patients at high risk for an ethical balance of the genuine needs of the next of a stroke. If carotid stenting could be demonstrated patient in a trial to receive good care and the longer as equally efficacious, for many patients it might be term public health need to only allow marketing the treatment of choice, being less invasive. This is approval and widespread use of treatments that actually the motivation behind a proposed National Institute of Health-funded trial comparing these two interventionstrategies.
(3) The new treatment is cheaper. An analogous situation T Y P E S O F N O N I N F E R I O R I T Y T R I A L concerns the relative merits of bypass surgery and cor- There are many different circumstances that may lead to onary angioplasty for patients with angina. Trials have undertaking a noninferiority trial design. The simplest shown that the prognosis death and/or myocardial case is where one wishes to demonstrate (if true) that the infarction appears similar for both intervention strat- efficacy of a new drug is the same as an existing active egies, the former provides better symptomatic relief drug, and one is not anticipating any other differences.
initially but is more invasive. However, any health care This will be particularly plausible if the drugs are of the strategy must take costs and cost-effectiveness into same class, in which case such a ‘me too’ drug account, and much interest in these trials has focussed development could lead to an additional marketable on the reduced initial costs of an angioplasty and product but not a substantial improvement in thera- whether that gain is maintained over several years peutic care. The merits of such a narrow diversity of products within a specific drug class may appear rather The complexity of these situations arises from the fact small, except as regards company profits. However, even that one is looking at a trade-off between efficacy and if the average benefits and safety profiles of two drugs in other issues regarding side-effects, patient acceptability a class appear identical, it is possible that individual and costs. Although such trials are often presented as patients may benefit more from one drug than the other.
noninferiority trials, it is possible that some modest Either before or after marketing approval, there is reduction in efficacy may be acceptable alongside the sometimes a need for a large randomized controlled other benefits of a new treatment. One such example safety study to evaluate a concern that may have arisen of this may be the acellular pertussis vaccines, which from observational adverse event reporting. For instance, are used in preference to whole cell pertussis vaccines the European Post-Operative NSAID Study Group evalu- because of fewer adverse events, although their efficacy ated in 11 302 patients undergoing surgery the safety of one nonsteroidal anti-inflammatory drug (NSAID) pain Another issue is whether any particular trial compar- relief drug ketorolac compared with two others, diclof- ing a new treatment with an active standard treatment enac and ketoprofen. This noninferiority safety study should be formulated as a noninferiority trial or not.
was motivated by concerns in the European regulatory Given the relative state of ignorance with which one authority, the Committee on Proprietary Medicinal starts any new study, one is often unsure whether to optimistically pursue the prospect of demonstrating a A more complex and interesting scenario arises when new treatment’s superiority or whether to settle for the aim is to demonstrate equivalence (noninferiority) of demonstrating noninferiority on the basis that, that it a new treatment to an active control, while knowing or will be still good enough to make the new treatment of suspecting that the two treatments will differ in some some value. Although the statistical power calculations other important respects. Possibilities here are: differ somewhat for these two scenarios (i.e. the latter Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490 reverses the roles of null and alternative hypotheses) – deny any claim of a new treatment’s noninferiority. We the underlying statistical and scientific intent is unal- then choose a sample size sufficiently large such that if tered. What one wants is a sufficiently large and unbiased there is true equivalence of new and control treatments study that the true magnitude of treatment difference is there is a high probability that the confidence interval for estimated precisely. That is, when the trial is completed treatment difference will be wholly to one side (the good the point estimate and confidence interval for the appropriate measure(s) of treatment difference contain The simplest case to quantify is for a binary response all the relevant evidence on which to hang claims of (success or failure). Let p be the anticipated percentage of success on each treatment if true equivalence exists.
Rigorous adherence to a single prespecified criterion of Let d be the ‘minimum clinically relevant difference’.
noninferiority, except for the convenience of planning Suppose that results will be expressed as an estimated the size of a trial, may not necessarily be the most percentage of treatment difference with a 95% confid- sensible way of interpreting a trial’s results. Nevertheless, ence interval around it. Also, suppose one wants to be formulating one’s realistic goals, and hence the required 90% sure that if treatments are truly identical then the number of patients, is an important feature of any confidence interval will exclude d, in a more favourable noninferiority trial’s planning and the next section deals A simple commonly used formula in this instance is A P P R O P R I A T E G O A L S A N D S A M P L E S I Z E S F O R N O N I N F E R I O R I T Y T R I A L S First it is essential to realize that failure to demonstrate More complex refinements exist and alternative but a statistically significant difference between two treat- similar formulae exist for other types of outcome data, ments does not allow one to assert that the two treat- such as comparison of risk ratios or means of a ments are equivalent, or even similar, in their efficacy.
quantitative measure, but this formula will adequately Obviously, the fewer patients there are in a trial the less illustrate the problems of choosing the size of a non- power to detect any meaningful difference so that nonsignificance in a conventional test of a null hypo- The difficulty lies in choosing appropriate values for thesis is a hopeless criterion for inferring noninferiority.
p and d, especially the latter. For example, consider a It would actually encourage the pursuit of smaller trials! noninferiority trial comparing a new drug with omep- Instead, the most widely accepted approach to deter- razole for treatment of Helicobacter pylori infection. The mine the required size of a noninferiority trial is to first binary response is eradication of infection (yes or no).
define the smallest true magnitude of inferiority that From past experience with omeprazole, p ¼ 85% was would be regarded as unacceptable, assuming that one the anticipated eradication rate. For trial planning d was has already chosen a single primary outcome measure set at 15%. This means that the new drug would be of response for this purpose. Anything truly bad or regarded as noninferior provided that the possibility of its even worse needs to be detected reliably so that any claim eradication rate being 15% worse than omeprazole could of noninferiority for the new treatment can then be ruled be ruled out (in the sense that the 95% confidence out. However, one is prepared to accept more minor interval for the treatment difference in eradication differences from true equivalence as being ‘good enough’.
rates would not include a 15% inferiority relative to The logical basis here is that even if one carries out an extremely large clinical trial, one never fully proves that Hence the trial required 2n ¼ ð4 Â 10:5 Â 85 Â 15Þ= two treatments are truly identical in their efficacy. The confidence interval for the treatment difference gets Leaping ahead to the actual results of this trial, the smaller and smaller as the sample size increases, but observed eradication rates on new drug and omeprazole proof of equivalence would require a confidence interval were 109 of 126 (86.5%) and 110 of 129 (85.3%), centred on zero and with zero width. An impossible task! In a spirit of achievable compromise one sets out to The 95% confidence interval for the treatment differ- arbitrarily choose this minimum clinically relevant ence was )7.5% to +9.8%. A difference of )15% is difference, commonly called delta, which if true would clearly ruled out of consideration, and on that basis the Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490 The pros and cons of noninferiority trials trial data support the new drug’s noninferiority relative above formulae have reversed the usual concepts of null to omeprazole. Of course, one still cannot claim with and alternative hypothesis, and type I and II errors a, b.
certainty that the new drug is identical in efficacy to For a noninferiority trial, a/2 is the probability that the omeprazole. After all, the confidence interval for treat- 100(1 ) a)% confidence interval excludes d when the ment difference does go beyond 5% in both favourable null hypothesis, treatment difference ¼ d, is in fact true.
and unfavourable direction. But according to the pre- b is the probability that the confidence interval includes defined goal, adequate evidence of noninferiority is d (or worse) when the alternative hypothesis of no treatment difference is in fact true.
Now just suppose that the new drug had had an In practice, there seems to be a pragmatic acceptance eradication rate of 99 of 126 (78.6%) instead of the by trialists and regulatory authorities that fairly gener- above 109 of 126. In that case, the 95% confidence ous choices of d, a and b are allowed in order not to interval for the treatment difference would have been demand inordinately large numbers of patients in from )16.1% to +2.7%. This would have been a most noninferiority trials. Should we be concerned therefore unhelpful result as one could neither claim noninferior- that the adoption of noninferiority designs with gener- ity (as the confidence interval would include )15%) nor ously large choices of d be permitting treatments with could one rule out equivalence of the two treatments more modest but important extents of inferiority to be (because the confidence interval includes no difference).
falsely accepted as noninferior? This is an inherent Fortunately, this did not happen in reality, but it weakness of noninferiority trials as currently performed.
illustrates the inconclusiveness that can easily arise if We do take a sizeable risk that some truly inferior noninferiority trials are conducted with fairly modest treatments will slip through the net.
Perhaps one could draw a distinction between (1) trials It seems quite fashionable to choose d ¼ 15% in comparing two drugs in the same class where there noninferiority trials. Indeed the regulatory authority may exist a high prior belief that treatments truly should did approve such a choice for the above trial, and there be equally efficacious and (2) trials comparing quite are more general regulatory guidelines for choice of delta contrasting treatments, e.g. drugs of differing types, or in such anti-infective trials. But it is hard to come up radically differing intervention strategies where there is with an objective reasoning behind this apparently no firm grounds on which to anticipate noninferiority.
arbitrary often used choice. Why not d ¼ 10% instead? A generous d leading to a smaller required sample size That would require 2n ¼ 535 patients, more than twice seems more permissible in the first instance (as was as many. d ¼ 5% might seem a plausibly tight safety the case from the above H. pylori example). The more margin, on the basis that only the slightest possible contrasting the treatments the harder it often is to inferiority of the new drug should be allowable, but recruit patients, but that is just the instance when large that requires nine times as many patients, a staggering sample sizes are needed in order to be confident of true These calculations are all based on 95% confidence The appropriate choice of d is particularly important and being 90% sure or ruling out noninferiority.
when a noninferiority trial vs. active control is taking More generally if one requires 100(1 ) a)% confidence place because it is considered unethical to proceed with a and wants 100(1 ) b)% surity, then one requires placebo control group. The worst that could happen is that the noninferiority margin is set so wide that a new and Zb are standardized normal deviates associated with treatment not much (if at all) better than placebo gets one-tail probabilities a/2 and b, respectively.
accepted as ‘noninferior’ on such an unduly lose For instance, with a 90% confidence interval and only criterion. Hence, it is useful to infer, preferably from 80% surity, each of the above sample sizes is reduced by past placebo-controlled trials, the magnitude of superi- 40%. However, with the tougher demands set by a 99% ority of the active control over placebo. The choice of d confidence interval and being 95% sure of rejecting a both for planning trial size and interpreting results of the difference d and claiming noninferiority when equival- new trial needs to be substantially smaller than this ence truly exists, each of the above sample sizes increases (1) The magnitude of superiority of active control over For those more used to power calculations for trials placebo may well be an overestimate. The placebo- aimed at detecting differences, it is worth noting that the controlled evidence may be limited, past trials may have Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490 some biases present and to carry forward one (perhaps where strict adherence to protocol inevitably does not lucky) possibly exaggerated effect of active treatments always happen, means that the real-life benefit of a new into future planning, reflects a lack of scientific caution.
treatment is seen for what it really is. That is, a new (2) Without a direct comparison with a placebo-control treatment has to fight its way through the hiccups, group, one is using an indirect argument via a compar- failings, frailties and unpredictability of human beings, ison with active control to infer that a new treatment is (both trialists and patients), in order to demonstrate its worthwhile. For a whole variety of reasons discussed in the next section (e.g. patient selection, noncompliance), The great difficulty with noninferiority trials is that the circumstances of the noninferiority trial may be their very motivation is to demonstrate the similarity of sufficiently different from the placebo-controlled trials to new and standard treatments, so that all these same cast doubt on the appropriateness of the active treat- problems work towards achieving this goal even if it is ment’s apparent magnitude of superiority over placebo.
not true. The anti-conservatism of a poorly designed A safety margin of less ambitious efficacy may be in and poorly conducted noninferiority trial can greatly enhance the risk of a type I error, the adoption of a (3) Any new treatment needs to have a certain useless treatment whose inadequacies could not be minimum magnitude of efficacy compared with placebo in order to be of worth, especially if other considerations One could argue that the unscrupulous investigator (side-effects, costs, inconvenience) come into play. Thus, has every intention to undertake a sloppy noninferiority even supposing issues (1) and (2) above did not apply trial. For instance, with selection of inappropriate (they usually do though!), one would still want d to be patients, poor compliance with intended treatments, much smaller than the active control’s superiority use of nondiscriminatory outcome measures, inconsis- tencies between observers, too short a follow-up and a These issues are linked to the earlier arguments substantial amount of missing data, it would not be concerning the choice between placebo and active surprising if the results showed closely comparable controls. The ideal circumstance for a noninferiority results even if the real treatments properly given to the trial is when the superiority of active control compared right patients were substantially different in real patient with placebo is irrefutable and well-documented, the noninferiority trial can be conducted in very similar So in noninferiority trials it is especially important to conditions and the choice of delta is small enough to adhere to a well-defined relevant study protocol, and also convince one that any new treatment passing such a to document that such adherence is successfully noninferiority test is truly of therapeutic value.
achieved. Some of the principal difficulties to bear in Fundamentally, many noninferiority trials are not be large enough to satisfy these requirements, meaning that (1) Selection of patients. It is important to select the type the risk of a type I error (as discussed in section 2 above), of patient for whom the efficacy of the active control false acceptance of a useless treatment, is often greater treatment has been clearly established. For instance, were one to deviate, even in part, from the patientpopulation in whom superiority over placebo hadpreviously been demonstrated, then any claim regarding T H E P O T E N T I A L I N F E R I O R I T Y O F a new treatment’s merits could not well distinguish between genuine noninferiority or inappropriate selec- For conventional clinical trials aimed at exploring the tion of patients. Informative generalizability depends on potential superiority of a new treatment over standard a representative patient sample of the same kind as had treatment (whether placebo or active), many of the previously demonstrated efficacy for the active control.
pitfalls that can arise operate in the direction of making it (2) Treatment compliance. The first requirement is that harder to detect a genuine treatment difference. Such one chooses a genuinely efficacious active control conservatism leads to the observed treatment difference treatment, and that it be given in the same form, dose being a dilution of the true effects under ideal conditions.
and quality as was previously used to demonstrate that This is often seen as an appropriate pragmatism, efficacy. One then requires that for both new and active whereby the attempted unbiased comparison of new treatment groups a satisfactorily high level of patient and standard treatment policies in a practical setting compliance is achieved, and that appropriate measures Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490 The pros and cons of noninferiority trials of such compliance are recorded. Any reasons for noninferiority by diluting some real treatment differ- alteration or discontinuation of treatments need docu- ence. More weight may instead be attached to per menting. Also, use of concommitant nonrandomized protocol analyses (which focus on patient outcome treatments needs documenting (and possibly standard- amongst compliers only or up until compliance ceases izing) as any differential use of other efficacious treat- in each patient) in the hope that they may reveal ments could conceivably mask the inferiority of a new undesirable treatment differences. But this in turn has problems as compliers are a select group of patients (3) Outcome measures. One needs to choose outcome who may give a favourably biased view (e.g. if the measures that reflect genuine patient benefit (i.e. surro- treatment is not helping, you drop out). These difficul- gate markers may well not suffice), and which were ties become particularly problematic if compliance previously used to demonstrate the efficacy of the active control treatment. Each such measure (or end point) Thus, for noninferiority trials there is no single ideal needs consistent well-defined criteria, with appropriate analysis strategy in the face of substantial noncompli- steps to reduce observer variation or bias. In addition, ance or missing data, and both analysis by intention to rigorous, objective reporting of adverse events is an treat and well-defined per protocol analyses would seem important issue, as any noninferiority needs to concern warranted. Such noncompliance will inevitably lead to a degree of concern over any affirmative claims of non- (4) Duration of treatment and evaluations. In any non- inferiority, which feeds back to the need to minimize any inferiority trial, the randomized treatments need to be given for long enough and the patient response evalu- In general, statistical considerations in the design, ated over a long enough period so that any potential monitoring and analysis of noninferiority are less well treatment differences have a realistic opportunity to established than for superiority trials, making it a fruitful reveal themselves. Due attention needs to be given to the area for further methodological research.
durations of treatment and follow-up in previous trialsdemonstrating efficacy of the active control treatment, and also the intended duration of treatment in futureclinical practice.
The most important advantage of a noninferiority trial (5) Statistical analysis issues. Any noninferiority trial is that faced with clear evidence of efficacy for an requires a well-documented statistical analysis plan.
existing standard treatment, it would be ethically There will often be a single primary outcome measure unacceptable to proceed with a placebo or inactive with a predefined noninferiority criterion and method of control group in the evaluation of a new treatment for analysis, but this should not preclude appropriate the same condition. In any particular circumstance an secondary analyses of other outcome measures, which important reservation before jumping to that conclusion could become important if they exhibit any signs of the is that such evidence of efficacy really is strong enough new treatment’s inferiority or if interpretation of the to warrant exclusion of placebo controls. Too lax an primary outcome findings is not clear-cut.
acceptance of noninferiority trials, with a less than In major phase III trials aimed at detecting treatment convincing active control treatment, could potentially differences, analysis by intention to treat is routinely lead to the adoption of more and more ineffective highlighted. That is, one analyses the complete follow- treatments, and this would be a misguided over-reac- up results for all randomized patients regardless of tion to the ethical concerns in conducting randomized- their compliance with intended treatment in a spirit of comparing the treatment policies as actually given.
So when one has made the right judgement to Although this may dilute any idealized treatment differ- undertake a noninferiority trial with an active control ences under (unrealistic) circumstances of 100% com- treatment, all necessary steps need to be taken to ensure pliance, that is generally considered wise pragmatism that any failings in the trial design, conduct or analysis compared with any potential exaggerations of efficacy could not artificially dilute out any real treatment that could arise from focussing on treatment compliers differences. That is, false claims of noninferiority need only. The dilemma for noninferiority trials is that faced with non-negligible noncompliance analysis by inten- Inevitably one can never prove that two treatments tion to treat could artificially enhance the claim of are identical, and hence some degree of compromise Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490 is required so that realistically achievable but ade- non-zero risk difference or non-unity relative risk. Stat. Med.
quately large numbers of patients are randomized in a noninferiority trial. Thus, any clinically important 3 Garbe E., Ro¨hmel J, Gundert-Remy U. Clinical and statistical issues in therapeutic equivalence trials. Eur. J. Clin. Pharmacol.
treatment difference can be demonstrated not to exist with reasonable confidence. One suspects that in too 4 International Conference on Harmonisation. Choice of Control many instances sample size determination for noninfe- Group in Clinical Trials. Federal Register (1999) 64 51767– riority trials is based on too generous a criterion of what constitutes a minimum clinically important treat- 5 Jones B., Jarvis P., Lewis J.A., Ebbutt A.F. Trials to assess ment difference, and this increases the risk that some equivalence: the importance of rigorous methods. BMJ (1996) inferior treatments may gain regulatory approval and ¨ hn A. Comparison of tests and sample size formulae for proving therapeutic equivalence based on the So placebo controls may rightly need to be ruled out in difference of binomial probabilities. Stat. Med. (1995) 14 certain areas of clinical research, but it would be wrong to rush too enthusiastically into more widespread use of 7 Senn S. Inherent difficulties with active control equivalence noninferiority (equivalence) trials without full consid- studies. Stat. Med. (1993) 12 2367–2375.
8 Senn S. Statistical issues in drug development. Wiley, 9 Snapinn S.M. Noninferiority trials. Curr. Cont. Trials B I B L I O G R A P H Y F O R F U R T H E R R E A D I N G 10 Temple R., Ellenberg S.S. Placebo-controlled trials and active- 1 Blackwelder W.C. ‘‘Proving the null hypothesis’’ in clinical control trials in the evaluation of new treatments. Ann. Intern.
trials. Con. Clin. Trials (1982) 3 345–353.
2 Farrington C.P., Manning G. Test statistics and sample size formulae for comparative binomial trials with null hypothesis of Ó 2003 Blackwell Publishing Fundamental & Clinical Pharmacology 17 (2003) 483–490

Source: http://www.translumina.in/upload/non%20inferiority.pdf