Adaptive Designs In Clinical Trials Statistics
 Words: 8366
 Category: Statistics
 Pages: 31
Get Full Essay
Get access to this section to get all the help you need with your essay and educational goals.
Get AccessContents
 The grounds for this diminution in the drug blessing rates are due to:
 2. Under streamlining clinical tests, specific countries of involvement were identified. They include
 3. Definitions and Concept of Adaptive Designs
 4. Motivation for Using Adaptive Designs
 5. Concerns and Pitfalls Associated With Using Adaptive Design
 5.1 Inflation of Type I error
 5.2 Potential for Counterproductive impacts of Adaptive Designs
 5.3 Controlling Type II mistake
 5.4 Complexity of Adaptive designs:
 6. Study Design Changes That Are Not Considered Adaptive Designs
 7. Types of Adaptive Designs
 7.1 Adaptive Randomization Design
 7.1.1 Treatment adaptative randomisation or discrepancy adaptative randomisation
 7.1.2 Covariate Adaptive Randomization
 7.1.3 Response Adaptive Randomization
 7.2 Adaptive DoseRanging Design
 7.3 Adaptive Seamless Phase II/III Design
 7.4 BiomarkerAdaptive Design
 7.5 Adaptive TreatmentSwitching Design
 7.6 AdaptiveHypothesis Design
 7.7 MultipleAdaptive Design
 7.8 Adaptive Group Sequential Design
 7.8.1 Aims of GSD
 7.8.2 General Form
 7.8.3 Type of Boundaries and Alpha Spending Function
 7.8.4 Group Sequential Trial Inference
 7.9 Adaptive SampleSize Reestimation Design
 8. Hypothesis/Simulations/Case Study
 8.1 Fixed Sample Size Design
 8.2 Group Sequential Design
 8.3 Adaptive Design
 Adaptive design, Plan 4 specifications:
 8.3.1 Cui, Hung and Wang Method
 Interim Monitor
In recent old ages investing in pharmaceutical research has more than doubled when compared to the old decennary but the figure of drugs approved has non reflected this increased investing in R & A ; D.
Harmonizing to the pharmaceutical industry ‘s trade association, Pharmaceutical Research and Manufacturers of America ( PhRMA ) , reported disbursement of its member organisations rose more than nine crease between 1980 and 2006, from about $ 6 billion to $ 55 billion. Harmonizing to recent informations, Pharma companies have spent around $ 65.2 billion in 20082. With the recent promotions in basic scientific disciplines, molecular biological science and cistron engineering, Pharma companies are able to prosecute new countries of research thereby increasing the costs of R & A ; D. But this continued growing in R & A ; D disbursement appears to hold small consequence on the figure of drug blessings. From 1996 to 2009, drug blessing rates have dropped by over 53 % from around 56 blessings in 1996 to 26 blessings in 2009.
The grounds for this diminution in the drug blessing rates are due to:
 As FDA is forcing for more rigorous intervention consequence differences between the active and control groups, companies are happening it tough to formalize the drug benefits with the degree of trouble that comes with demoing a clinically relevant intervention differences.
 With the recent scientific promotions in biological science, medical scientific disciplines particularly with the continued development of Genbank, there are larger countries of chances that could take to a more robust portfolio of merchandises. But the industry has non yet to the full able to use their possible.
 The recent bicker of amalgamations and acquisitions has non helped with NDA blessings as they were focused on acquiring the immediate market portion instead than R & A ; D grapevine.
 Based on the 73 NDA ‘s approved in 2009, merely 16 can be considered as NME, the remainder of the 57 NDA ‘s are based on new preparation or new combination or new indicant or new manufacturer3. This leads to the decision that most of the companies are aiming easy marks instead than concentrate on complex and long term curative marks.
 The mean success rate of a drug from first human tests to acquiring approved is merely at 12 % 55
 As regulative bureaus are demanding more safety informations, with the recent extremely publicized backdown of Vioxx etc, it is making a riskadverse environment and the drug companies are playing it safe.
To turn to the falling drug blessing rates, the United Sates Food and Drug Administration ( FDA ) started a “Critical Path Initiative” to assist pharmaceutical companies better grasp the scientific challenges and supply alternate attacks for drug development. In 2004, FDA published a study “Challenge and Opportunity on the Critical Path to New Medical Products” as portion of a countrywide attempt to force the medical merchandise development and patient attention into twentyfirst century. In 2006, FDA released a followup papers “Critical Path Opportunities Report and List” that highlighted to utilize of interim analysis informations or anterior experience informations that can be used for clinical trail designing. The “Opportunities Report and List” presented 76 specific scientific chances. This list was updated in 20074. The six precedence countries identified by FDA are
 Better rating tools
 Streamlining clinical tests
 Harnessing bioinformatics
 Traveling fabricating into twentyfirst century
 Developing attacks to turn to pressing public demands
 Specific atrisk populations – paediatricss
2. Under streamlining clinical tests, specific countries of involvement were identified. They include
 Development and use of biomarkers or genomic markers
 Constitution of quantitative disease theoretical accounts
 Use of more enlightening designs such as adaptative or Bayesian designs
 Modeling and simulations
 Data gaining control and direction engineerings
 Development of enriched test designs
3. Definitions and Concept of Adaptive Designs
To increase the chance of success of clinical development, and besides to expeditiously place the clinical benefits of the trial intervention under probe, a clinical test may sometimes undergo a alteration either in the trail procedures that include eligibility standards, survey dosage, intervention continuance, survey end points, research lab testing processs, diagnostic processs, standards for evaluability, appraisal of clinical response and statistical processs that include randomisation, survey design, survey objectives/hypothesis, sample size, informations monitoring and interim analysis, statistical analysis program, and methods for informations analysis.
Harmonizing to EMEA, March 20067 published paper, “an survey design is called “adaptive” if statistical methodological analysis allows the alteration of a design component ( e.g. samplesize, randomisation ratio, figure of intervention weaponries ) at an interim analysis with full control of the type I error” . Harmonizing to PhRMA Working Group an adaptative design can be defined as “a clinical multistage survey design that uses roll uping informations to make up one’s mind on how to modify facets of the survey without sabotaging the cogency and unity of the trial”45. Harmonizing to the recent FDA guidance5 ‘An adaptative design clinical survey is defined as a survey that includes a prospectively planed chance for alteration of one or more specified facets of survey design and hypothesis based on analysis of informations ( normally interim informations ) from topics in study” An adaptative design can besides be called as a flexible design. “validity” refers to infering right statistical analysis without consisting the prejudice that might include protecting the Type I error. “integrity” refers to decently carry oning the survey design that might include and non limited to prespecified planning in the protocol, proper care of blinding processs, careful monitoring of the interim analysis consequences. The cogency includes both internal and external cogencies. Internal cogency is the grade to which we are successful in extinguishing confusing variables and set uping a causeeffect relationship within the intervention itself. A survey that readily allows its findings to generalise to the population at big has high external cogency. Integrity involves minimising operational prejudice ; making a scientifically sound protocol design ; adhering steadfastly to the survey protocol and criterion operating processs ( SOPs ) ; put to deathing the test systematically over clip and across sites or state ; supplying comprehensive analysis of test informations and indifferent readings of the consequences ; and keeping the confidentiality of the data5.
By definition, most of the clinical test adaptative designs can be slotted into either Prospective or Ongoing or Retrospective.
Adaptation 
Examples 
Prospective ( by design ) 
Interim analysis 
Early fillet of the clinical test either due to futility or efficaciousness. 

Sample size reestimation 

Group Sequential design 

Ongoing ( ad hoc ) 
Inclusion/exclusion standards 
Dose response or dose regimen 

Treatment continuance 

Retrospective 
Study end point 
Switch overing of the primary analysis from high quality to noninferiority etc 
Table 1. Different classs of adaptative designs in clinical tests.
Study design facets that are revised based on information obtained wholly from beginnings outside of the particular survey are non considered adaptative design clinical trials5.
Interim analysis: Any analysis that utilizes informations which is obtained at an certain clip point before the completion of a test can be attributed as interim analysis.
Blinded analyses: Any analysis that utilizes informations in which the topic intervention weaponries are non revealed so that consequences can non be used to foretell the evident intervention consequence can be called as blinded analysis.
Unblinded analyses: Any analysis that utilizes informations in which the topic intervention weaponries are revealed so that consequences can non be used to foretell the evident intervention consequence can be called as unblinded analysis.
Type I error: Type I error can be defined as chance of rejecting a true void hypothesis
Type II mistake: Type II mistake can be defined as the chance of neglecting to reject a false nothing hypothesis.
Power: Probability of rejecting a false nothing hypothesis.
Bias: Harmonizing to the statistical definition an prejudice can be defined as “a systematic deformation of a statistical value”8. One of the chief concerns associated with adaptative designs is presenting a prejudice after looking at the interim analysis informations and doing determinations that could blow up the Type I error.
4. Motivation for Using Adaptive Designs
 Adaptive designs by their interim analysis informations reappraisal can uncover more information about the survey intervention effects.
 Adaptive designs can increase the success rate of run intoing the primary end point by careful planning.
 Outputs improved apprehension of the interventions consequence.
 Adaptive design surveies might cut down the clip and resources needed to measure each particular pick within a scope of parametric quantity values.
 Adaptive designs can besides be used to extinguish the clip period that occurs between separate exploratory and effectivity surveies in conventional drug development plans.
· Adaptive designs by their flexible nature can be used to command the budget resources based on scientific and statistical evidences.
5. Concerns and Pitfalls Associated With Using Adaptive Design
5.1 Inflation of Type I error
Bias Associated with the Multiplicity of Options: When determinations are taken by looking at multiple consequences based on different end points, at different phases there is a possibility of choosing the most successful consequence that could demo the needed intervention. With each interim phase there is a prejudice that is introduced that could ensue in blow uping type I error. To forestall this all the interim phase versions should be prespecified in the protocol and SAP. Protocols should besides hold subdivisions for statistical methodological analysiss that can be used to antagonize the rising prices of Type I error.
Operational Bias: Operational prejudice is a nonstatistical prejudice that could be introduced into the survey as different forces involved in the unblinding of the interim informations might act upon the consequences by either information sharing or incorrect survey appraisals etc.
5.2 Potential for Counterproductive impacts of Adaptive Designs
Potential to present statistical prejudice while gauging intervention consequence: As adaptative designs are used to do determinations at each interim phase to choose the appropriate sample size and power so that an estimated intervention consequence is achieved, it introduces a statistical prejudice. Alternatively of utilizing information from merely one or two interim phases, attention should be taken so that hypothesis testing can be done utilizing each and every phase ‘s informations.
Elimination of Time to Thoughtfully Explore Study Results: Adaptive designs can be used to extinguish clip between different stages of survey by looking at interim consequences. Care should be taken to turn to any possible safety jobs found in the interim consequences before come oning to the following clinical stage. If non, these safety issues could ensue in more extended timelines for completion of stage 3 tests.
5.3 Controlling Type II mistake
There is a possible for increasing Type II mistake rate in adaptative designs. When based on interim consequences, multiple doses are non selected that were assuring in a dose response stage I surveies, based on limited informations, it could potentially neglect to show a needed statistical intervention consequence and blow uping the Type II mistake rate. Care should be taken while planing adaptative designs, that utilize early fillet regulations for futility and dose response designs to command the Type II mistake rate.
5.4 Complexity of Adaptive designs:
Although adaptative designs are attractive to clinical research workers and patrons due to their flexibleness their complexness should besides be considered while taking an adaptative design. The complexness could be significant depending upon the type of version employed and could ensue in complex calculations.
Figure 3 depicts an overview of the procedure associated with adaptative test design and analysis.
6. Study Design Changes That Are Not Considered Adaptive Designs
Revisions after unplanned findings in an interim analysis: During an interim analysis there could be findings bespeaking commissariats for design alterations that could heighten the opportunities of a successful statistical consequence. Any of such actions taken that were non defined in progress in the protocol or SAP can non be considered portion of adaptative designs. Furthermore, these unplanned alterations could ensue in the rising prices of Type I error rate.
Revisions based on information from a survey external beginning: Any information that is gathered from outside the survey and could be used to modify the current survey design does non represent an adaptative design.
7. Types of Adaptive Designs
The most common adaptative design methods in clinical tests include, but are non limited to:
 Adaptive randomisation design
 Adaptive doseresponse design
 Adaptive seamless stage II/III design
 Biomarker adaptative design
 Adaptive intervention exchanging design
 AdaptiveHypothesis design
 Multiple adaptative design
 Group consecutive design
 Sample size reestimation design
7.1 Adaptive Randomization Design
Simple or complete randomisation is likely one of the most commonly employed conventional randomisation processs in clinical tests. In a simple randomisation, each patient is indiscriminately assigned to each of the intervention groups with a fixed allotment chance. To guarantee intervention balance, most of the clip there is an equal allotment chance for any topic to be assigned to any intervention group. In some instances, where needed like in Oncology tests, unequal allotment of topics between intervention groups may be employed.
7.1.1 Treatment adaptative randomisation or discrepancy adaptative randomisation
This randomisation is usually used to accomplish a more balanced design seeks or to cut down the divergence organize the mark intervention allotment ration by using a varied allotment chance. Some of the more common intervention adaptative randomisation methods are:
Block randomisation: In block randomisation the allotment chance is a fixed changeless before any of the two intervention groups reach its mark figure. Once the mark figure is reached in one of the two intervention groups, any future registration will be assigned to the other intervention group. The minimal block size normally chosen in clinical tests is two16.
Efron ‘s Biased Coin Design: Efron19 ( 1971 ) was the first to suggest the usage of a colored coin design to command the instability between interventions in a clinical test. His method is rather similar to the Pocock and Simon method with hundred = 147.
Lachin ‘s Urn theoretical account: Lachin ‘s urn model36 is based on the patient result, and by adding balls of a peculiar colour to the urn, the allotment chances are changed. If a intervention is successful a ball of that colour is added. If a intervention fails a ball of the other colour is added. Using this method allows the allotment of topics to stay a random procedure.
7.1.2 Covariate Adaptive Randomization
One of the chief uses of covariate adaptative randomisation is to cut down the covariate instability between different intervention groups. It is besides referred to as adaptative stratification. At different clip points in a test, as more cumulative information becomes available about baseline covariates and intervention assignments, the allotment chance is modified to guarantee proper covariate balance between intervention groups. Some of the more normally used covariate adaptative methods are:
Zelen ‘s Rule: Zelen58 ( 1969 ) proposed utilizing a preassigned randomisation sequence without stratification. This sequence could be generated utilizing complete randomisation or restricted ( block ) randomisation.
Pocock and Simon ‘s Method: Pocock and Simon47 ( 1975 ) proposed the usage of a difference every bit good. For a new capable come ining the survey, the difference between interventions in all strata that the topic would be in, is calculated. If this difference is 0 so the topic is randomized to either intervention with equal chance. If, nevertheless, the difference is more or less than 0, the topic is randomized to the intervention that has been less favored at that point in the survey with a chance of ( c+1 ) /3, where degree Celsius can run between 0.5 and 2.
Wei ‘s Marginal Urn: Wei53 ( 1978 ) used an urn theoretical account to cover with the job of instability. Alternatively of utilizing an urn to stand for every crossdivision of the strata variables, he used an urn for each degree of each strata variable.
7.1.3 Response Adaptive Randomization
Responseadaptive randomisation is used when ethical considerations make it unwanted to hold an equal figure of patients assigned to each intervention. In this randomisation, patients are allocated to different intervention groups based on response or result of old patients. This helps in patients acquiring assigned to build up that is demoing considerable intervention consequence. Wellknown responseadaptive theoretical accounts include the randomised playthewinner theoretical account, the double adaptative biased coin design, and the droptheloser model14.
Playthewinner theoretical account: In Playthewinnermodel a patient gets assigned to a intervention group based on old patient ‘s result. This theoretical account is more likely to be used in trails with binary results or success or failure. This theoretical account assumes that before the following patient is randomized, the old patient ‘s result is known.
Doubly adaptative coin design: In this design patients are randomized non merely based on old patient ‘s response, but besides old patient ‘s assigned intervention group.
Dropthelosermodel: In this theoretical account, any intervention group that does non demo promising intervention consequence is dropped wholly and no patients are allocated to this group in future. This is accomplished by utilizing an urn theoretical account similar to the randomized playthewinner theoretical account but by taking balls alternatively of adding them. This basic droptheloser theoretical account was developed chiefly for a dichotomous response ( responder – nonresponder ) 15.
7.2 Adaptive DoseRanging Design
Insufficient geographic expedition of the doseresponse relationship frequently leads to a hapless pick of the optimum dosage used in the confirmatory test that, accordingly, might take to the failure of the test and the clinical plan. Twenty per centum of drugs approved by FDA between 1980 and 1989 had the initial dosage changed, in most instances take downing it. Even though the regulative bureaus prefer patrons to restrict the usage of adaptative designs in confirmatory tests, both the FDA and the EMEA are supportive and they encourage the deployment of suitably planned adaptative designs in doseranging surveies.
Adaptive doseranging designs can hold a figure of aims. For illustration, it can be used to set up the overall doseresponse relationship for an efficaciousness parametric quantity or efficaciousness and safety parametric quantities, estimate the curative window or aid with the choice of a individual mark dosage. Other applications of adaptative doseranging designs might be related to specific indicants. For illustration, an oncology survey might utilize an adaptative attack to happen a maximal tolerated dose7.
In Phase I clinical tests, most formal doseescalation processs have been developed in the oncology curative country. In this scene, the aim is normally to find the maximal tolerated dosage ( MTD ) . The traditional Phase I designs are algorithmbased. The widely used standard Phase I design is the so called “3+3 design” . Lin and Shih ( 2001 ) 42 discussed the belongingss of the traditional ( 3+3 ) and modified algorithmbased designs in a general scene ( A+B ) . Although the 3+3 method has been criticized for its inclination to include excessively many patients at suboptimal dosage degrees and give a hapless estimation of the MTD, it is still widely used in pattern because of its simpleness in logistics for the clinical research workers to transport out. Other algorithmbased designs include accelerated titration designs and group upanddown designs. The modelbased attacks include the continual reappraisal method ( CRM ) proposed by O’Quigley, Pepe, Fisher ( 1990 ) 43 and its alterations, the escalation with overdose control ( EWOC ) method proposed by Babb, Rogatko, Zacks ( 1998 ) 12 and Bayesian determination processs proposed by Whitehead and Brunier ( 1995 ) 57.
Traditionally, Phase II designs measure the efficaciousness of doses in the false acceptabletoxicity dosage scope. Both safety and efficaciousness informations can be used in certain instances, to choose the proper dose.
In 2005, the Pharmaceutical Innovation Steering Committee ( PISC ) of PhRMA formed several working groups “to look into different drivers of the decreasing success rates observed in drug development plans across the pharmaceutical industry.” Among those groups was the Adaptive DoseRanging Designs ( ab initio called Rolling Dose Studies ) working group. In 2007, the group published a white paper “Innovative Approaches for Designing and Analyzing Adaptive DoseRanging Trials”13 in Journal of Biopharmaceutical Statistics. The white paper summarizes the work of the group, including the consequences and decisions of a comprehensive simulation survey, and puts forward recommendations on how to better dosage runing tests, including, but non limited to, the usage of adaptative doseranging methods5.
Adaptive designs present an chance to derive more information expeditiously about the dose response at early phases of development. However, adaptative designs require flexibleness and tonss of planning, and might non ever be better than fixed designs. The added scientific and operational complexness should be justified. Trial simulations should be used for gauging operational features of the design and to compare public presentation across designs.
7.3 Adaptive Seamless Phase II/III Design
Adaptive seamless stage II/III test design is a design that can bring forth the same information in a individual survey instead than utilizing two separate stage IIb and III clinical test surveies. Most of the adaptative seamless designs are based on uniting Phase IIb ( “learning” ) and III ( “confirming” ) tests. One of the chief advantages of seamless stage II/III designs is it can cut down both sample size every bit good as continuance of test thereby holding a cost benefit for the sponsor17.
In an seamless design, as there is no lag clip between the acquisition and confirmatory tests, and since all the informations collected at the larning stage is used along with the informations from confirmatory trails, it can take to efficiencies in roll uping more safety or efficaciousness informations at a quicker gait. Adaptive seamless designs can be classified into four classs harmonizing to design characteristics and versions.
 Fixed figure of regimens, which includes halting early for futility, biomarkerinformed halting early for futility, and halting early for futility/efficacy with sample size reestimation.
 Flexible figure of regimens, which includes a flexible figure of regimens, adaptative hypotheses, and responseadaptive randomisation
 Population version, where the figure of patient groups can be changed from the larning stage to the confirmatory stage and
 Combination of 2 and 3.
As the usage of an adaptative seamless design is to shorten the clip of development, whether the adaptative seamless design would accomplish the survey aims within a decreased timeframe would be another of import factor for feasibleness consideration. Logistic and operational challenges besides should be accounted for while be aftering an adaptative seamless design to account for drug supply and packaging.
Due to the complexness of combing two different phases, calculations for pvalues and proper statistical illation demand to be carefully evaluated to command any prejudice.
7.4 BiomarkerAdaptive Design
Biomarkeradaptive design ( BAD ) refers to a design that allows for versions utilizing information obtained from biomarkers. A biomarker is a characteristic that is objectively measured and evaluated as an index of normal biologic or infective procedure or pharmacologic response to a curative intercession.
A classifier biomarker is a marker that normally does non alter over the class of the survey, like DNA markers. When the size of the selective population is excessively little to warrant the overall benefit to the patient population, a BAD may be used, where the biomarker response at interim analysis can be used to find which mark populations should be focused on. A prognostic biomarker informs the intervention consequence on the clinical end point and utilizing it could take to faster determination devising, although formalizing prognostic biomarkers itself a challenge. This can be addressed through BAD. In a BAD, ‘softly ‘ validated biomarkers are used at the interim analysis to help in determination devising, while the concluding determination can still be based on a goldstandard endpoint, such as endurance, to continue the typeI error7.
7.5 Adaptive TreatmentSwitching Design
An adaptative treatmentswitching design ( ATSD ) is a design in which patients can be switched over to intervention group that shows more promising consequences based on safety or efficaciousness.
To measure the efficaciousness and safety of a trial intervention for progressive diseases, such as malignant neoplastic diseases and HIV a parallelgroup, activecontrol, randomised clinical test is frequently conducted. In this type of test, qualified patients are indiscriminately assigned to have either an active control or a trial intervention under probe. Due to ethical considerations patients are allowed to exchange from one intervention to another if there is grounds of deficiency of efficaciousness or disease progression14.
7.6 AdaptiveHypothesis Design
An adaptivehypotheses design is a design where alterations of hypotheses during the behavior of a test occur due to some grounds ( e.g. , an research worker method has non been validated yet, information from other surveies is needed to be after the following phase of the survey, there is a demand to include extra doses, recommendations from the informations monitoring commission ) .
7.7 MultipleAdaptive Design
A multipleadaptive design is a combination of multiple adaptative designs. Some of the designs could be ( I ) a combination of adaptative GSD and adaptative seamless test design and ( two ) adaptative doseresponse design with adaptative randomisation design etc.
7.8 Adaptive Group Sequential Design
The Clinical Research Dictionary defines consecutive design as “a test design that allows a expression at the informations at peculiar clip points or after a defined figure of patients have been entered and followed up based on explicating a fillet regulation derived from repeated significance tests.” Another definition was given by Chang ( 2008 ) 15: “a group consecutive design is an adaptative design that allows for prematuretermination of a test due to efficacy or futility, based on the consequences of interim analyses” .
Consecutive techniques have sprung from a long history utilizing a methodological analysis that involves Brownian gesture, a phenomenon discovered in 1827 by an English phytologist Robert Brown, and continued by Albert Einstein in 1905 who developed the first mathematical theory of Brownian gesture, a part for which he has received the Nobel award. Using this mathematical base, group consecutive designs for clinical tests with interim analyses have been introduced for ethical, economical, and practicableness grounds ( Pocock, 1977 ) 46. Early work done by Pocock ( 1977 ) 46, and O’Brien & A ; Fleming ( 1979 ) 11 has been a strong base for recent development of group consecutive methodological analysiss. More recent work done by Lan & A ; DeMets ( 1983 ) 40 refering unequal group sizes, and Wang & A ; Tsiatis54 ( 1987 ) on methods of seting critical values to keep the overall false positive rate, Jenninson & A ; Turnbull ( 1990, 2000 ) 29,30 on repeated assurance intervals, Lan & A ; Wittes ( 1988 ) 38 on conditional power and stochastic curtailment, and Whitehead ( 1997 ) 56 on asymmetrical boundaries have led to broad credence, strength and rapid advancement in statistical methodological analysiss now available.
The Pharmaceutical Research and Manufacturers of America ( PhRMA ) Working Group on Adaptive Design was formed in March of 2005 to look into and ease chances for wider credence and use of adaptative designs and related methodological analysiss ( including adaptative group consecutive designs ) among statisticians, clinicians, regulators within the pharmaceutical industry, academe and wellness authorization bureaus. In November of 2006, the PhRMA Working Group published a full White Paper on adaptative designs in the Drug Information Journal ( Volume 40, Number 4 ) .
There has been much treatment on whether adaptative group consecutive designs or sample size reestimation ( SSR ) methods should be used in planning/conducting a clinical test. The general decision is that efficient adaptative designs for SSR have small to offer over efficient group consecutive designs in footings of sample size. The recommendation is to measure whether or non a group consecutive design ( GSD ) is equal before sing adaptative samplesize reestimation. GSDs by and large have good regulative credence, and offer wellunderstood methods that allow significant flexibleness, while adaptative sample size reestimation methods offer more flexibleness and are by and large more appealing to patrons. SSR and GSD methods could besides be combined, and we have seen many petitions from patrons where a combination method would react to their demands better than either GSD or SSR methods entirely. The proposed combination procedure is to transport out the reestimation of the entire information prior to the first interim testing of the intervention consequence and to later set the informationtime graduated table based on the updated entire information for all the group consecutive trials.
7.8.1 Aims of GSD
Group consecutive designs are utile for assorted principles including, but non limited to:
 The ethical demand to supervise the test to guarantee that topics are non administered insecure
or uneffective interventions ;
 Early scrutiny of the consequences at an interim analysis may sometimes uncover some
administrative jobs that can be resolved before test terminal ;
 Early on halting for effectivity could intend the possibility that the drug is marketed
Oklahoman, and for a test with a negative consequence, early fillet agencies cutting costs and salvaging resources.
However, if non used decently, group consecutive designs have the possible to present statistical and operational prejudices. It is of import that a test using these GSD methods will be conducted merely for clearly defined grounds and follow prespecified operational processs. This is particularly the instance for confirmatory tests since improperly conducted tests can impact the reading and regulative credence of the test results8.
7.8.2 General Form
Repeated expressions at the roll uping information additions the chance of declaring a intervention consequence even when there is none. How to take this into history through building of statistical boundaries is the topic of the statistical methodological analysiss behind GSDs.
In order to be able to depict the general signifier of the trials, we assume that we test the nothing hypotheses H0: ? = 0 with reversible Type I error chance ? and power 1 – ? at ? = ±? . We consider a group consecutive design in which up to K analyses are permitted ( the Kth analysis being the concluding analysis ) . We besides suppose that standardised trial statistics Z1, Z2, … , ZK are available at analyses 1,2, … , K and they follow a canonical articulation normal distribution.
A general group consecutive trial with possible early fillet to accept or reject the void hypothesis is defined by a brace of boundaries ( Alaska, berkelium ) , with 0 ? Alaska & lt ; berkelium and aK=bK, and it has the signifier:
After group k = 1, … , K 1
if Zk ? berkelium halt, and Reject H0
if Zk & lt ; ak halt, and Accept H0
otherwise continue to group k+1
After group K
if ZK ? berkelium halt, and Reject H0
if ZK & lt ; aK halt, and Accept H0
The upper ( berkelium ) and lower ( Alaska ) boundaries represent the difference between benefit and injury. Traversing the upper boundary demonstrates benefit, while traversing the lower boundary suggests injury. The undermentioned subdivision discusses assorted types of boundaries, among which Pocock ( 1977 ) 46, O’Brien and Fleming ( 1979 ) 11, Wang & A ; Tsiatis ( 1987 ) 54, Lan and DeMets ( 1983 ) 40, and Whitehead ( 1997 ) 56 are the most common.
7.8.3 Type of Boundaries and Alpha Spending Function
In 1977 Pocock46 adapted the thought of perennial significance trials to analyse roll uping informations a figure of times during the survey while commanding the prespecified Type I error. He suggested utilizing a unvarying boundary across the interim analyses and the concluding analysis. A twosome of old ages subsequently, O’Brien and Fleming ( 1979 ) 11 introduced a boundary which is more conservative at earlier phases and less conservative closer to the concluding analysis. O’BrienFleming ‘s attack uses really little significance degree early on and saves most of it for the concluding analysis while the Pocock attack applies a changeless set of halting standards over clip, doing it easier to end early. However, if the test returns to the ulterior analysis times ( closer to the concluding 1 ) , the Pocock trial program is less likely to allow early halting comparative to O’BrienFleming ‘s design. Another popular trial used largely in oncology tests is Haybittle – Peto trial introduced by Haybittle & A ; Peto ( 1976 ) 51.It uses a critical boundary value of 3 ( for standardized usually distributed trial statistics ) for all analyses but the concluding 1 ; the critical value for the concluding analysis is adjusted to give an overall Type I error precisely equal to the prespecified ? . The common denominator among these designs is that a sequence of trial statistics ( e.g. , standardised Zscores ) is selected in order to continue the overall false positive rate ( Type I error ) . Figure 4 shows how each of these fillet regulations are mapped as a map of the trial statistic versus the interim analysis figure.
The lines drawn on the figure represent halting boundaries, unambiguously determined by the sequence of trial statistics for each peculiar process. Therefore, if the ascertained Zscore at any analysis clip exceeds a certain boundary ( e.g. , is greater than the upper boundary value ) , so sufficient grounds exists that any trial conducted to that clip will be associated with a type I error rate of no greater than the prespecified ? . Therefore, the false positive rate is decently controlled, and based on the usage of the regulation, dependable grounds exists to urge early fillet of the test. The upper and lower fillet boundaries are used to measure the statistical grounds against the nothing and alternate hypotheses, severally.
In 1987, Wang and Tsiatis proposed a household of reversible trials which offers boundaries of different forms including the Pocock and O’BrienFleming boundaries as particular instances. Asymmetrical boundaries where the upper boundary is used for claiming efficaciousness and the lower boundary for safety monitoring have been proposed by Whitehead ( 1997 ) 56. This type of asymmetrical boundary can besides be used to at the same time turn to efficaciousness and futility in the interim analysis. Detailss of the mathematical background of methods will non be presented here, but can be found in several firstclass mentions ( e.g. , Jennison, Turnbull ( 2000 ) 29, and Whitehead ( 1997 ) 56 ) .
Both the Pocock and O’BrienFleming methods maintain the overall ?level, but require prespecified maximal figure of patients, the figure of interim analysis, and equal increases of information between interim phases. Lan and DeMets39 ( 1983 ) eliminated the equal increases of information demand by presenting a disbursement map attack. The attack spends the allowable Type I error rate over clip harmonizing to a chosen disbursement rule and the sum of information accrued. Several types of disbursement maps were proposed, including those that produce boundaries similar to the Pocock and the O’BrienFleming boundaries. The disbursement map attack allows dropping or adding an interim analysis. Since the boundaries determine how decisions will be drawn at the interim analysis and the concluding analysis, it is of import to prespecify which type of disbursement map and boundary will be employed.
The pick of statistical attack and the type of boundaries should depend on the aims of the test and the function of the test in the clinical plan. While sing halting for the overpowering efficaciousness, one should maintain in head the deduction of halting early on the safety profile of the drug.
7.8.4 Group Sequential Trial Inference
Groupsequential processs are designed to prove hypotheses. Simply reasoning that one intervention is better than another as demonstrated by rejecting a hypothesis tested at a Type 1 mistake rate of ? without an estimation of the size of the consequence leaves the reading of consequences of the test incomplete. Inference following a groupsequential test is debatable because pvalues, point estimations and assurance intervals are biased if consecutive monitoring is non taken into history.
There is no alone pvalue in the consecutive scene. When planing the test, one must predetermine non merely the consecutive boundaries, but besides the algorithm by which the pvalues will be calculated. Often the corrected and uncorrected pvalues are similar, but sometimes the difference can be significant and in such instances uncorrected pvalues green goods an excessively optimistic position of the strength of the grounds for a intervention consequence. The magnitude of the difference between corrected and uncorrected pvalues depends on the nature of the monitoring boundary and the phase at which the test is stopped. The pvalue depends on the order of the possible results, but there is no individual natural manner to make it. There are four ordinations that had received most attending: Stagewise ordination, MLE ordination, Likelihood ratio ordination, and Score trial ordination. Detailss of each method can be found in Jennison, Turnbull ( 2000 ) 29. Proschan, Lan and Wittes ( 2006 ) 48 strongly urge the stagewise ordination process because its pvalue depends merely on what has happened so far and non on future programs. The stagewise ordination is besides consistent in the sense that the pvalue ? ? if the boundary was crossed. Similarly, the ascertained consequence from a test that is stopped early overestimates the true value, but it still reported and used for simpleness. There are two state of affairss in which adjusted interval estimations might be required: after the trial statistic has crossed a fillet boundary and at an interim phase of the survey regardless of early expiration ( repeated assurance intervals ) .
Most of the techniques that are used for the groupsequential illations are implemented and available in commercial package like ADDPLAN, EAST, R, SAS etc.
7.9 Adaptive SampleSize Reestimation Design
Adaptive sample size reestimation ( SSR ) design is a design that allows a sample size reestimation during an predefined phase of a test based on the informations collected from the roll uping test informations. Sample size appraisal for a test is sensitive to the intervention consequence and associated variableness assumed for the calculations. One of the grounds to carry on an interim analysis is to look into the premises used to cipher the original sample size. It is particularly of import to look into these premises when unequal information about the parametric quantities was available when gauging the necessary sample size at the design phase.
Sample size reestimation can be based on blinded or unblinded informations. In the first scenario, the sample size accommodation is normally based on the observed pooled discrepancy or in instance of categorical response pooled even rate at the interim analysis and does non necessitate unblinding of interventions. Gould ( 1992, 1995 ) 21, 22, Gould and Shih ( 1992 ) 23, Kieser and Friede ( 2003 and 2004 ) 34, 35, Proschan, Lan, and Wittes ( 2006 ) 48 have more information on the statistical processs used for blinded sample size reestimation. In the 2nd scenario, the intervention consequence and its variableness are evaluated during the interim analysis on unblinded informations and the sample size is adjusted based on the updated information. Statistical methods for unblinded sample size reestimation could be based on consequence size ( intervention difference and/or its discrepancy ) or conditional power. The research has been done to measure how sample size accommodation based on unblinded discrepancy estimations affects the Type I error rate. Depending on the designs, the consequence on the Type I error rate varies from negligible to significant. Kieser and Friede ( 2000 ) 34, Proschan and Wittes ( 2000 ) 49, and Coffey and Muller ( 2001 ) 18 provide solutions on what to make to continue the overall Type I error rate.
There are occasions when the intervention difference assumed at the design phase is excessively optimistic. Subsequently, there is a desire to redesign the survey to observe a smaller but still clinically meaningful difference. Fisher ( 1998 ) 20 proposed a reestimating sample size method based on the construct of “variance spending” . Other attacks have besides been proposed by Cui, Hung, and Wang ( 1999 ) 16, Lehmacher and Wassmer ( 1999 ) 37, Lan and Trost ( 1997 ) 41, Muller and Shafer ( 2001 ) 27, Proschan and Hunsberger ( 1995, 2005 ) 50. Jennison and Turnbull ( 2003 ) 31 recommended a group consecutive process for state of affairss where a smaller, but still clinically meaningful, consequence size was to be considered based on interim analysis consequences. The attack proposed by Jennison and Turnbull has a smaller mean sample size compared to other consecutive processs designed for a similar intent.
Although the flexibleness to set the sample size of a test during an interim analysis is appealing when information is limited at the design phase, it does non come without a monetary value. Execution of adaptative processs for collateral tests demands to be carefully planned and executed. ICH Guideline E9 encourages seting sample size based on blinded information. When the accommodation is made on unblinded informations, it is of import to take stairss to continue the Type I error rate.
Type I error control is the critical issue related to SSR designs and is controlled by utilizing a combination of informations collected from each predefined phase. Based on its old phase ‘s consequences the current phases sample size is calculated and the protocol needs to turn to these statistical methodological analysiss. Cui, Hung and Wang Method ( CHW,1999 ) 16 paper, uses predefined weights at predefined phases to cipher the sample size while able to continue Type I error rate. Proschan and Hunsberger ( 1995, 2005 ) 50 proposed to used conditional power along with the first interim phase informations to increase the sample size at the 2nd phase. Most of the SSR designs use combined pvalues at different phases to gauge the concluding pvalue. Lehmacher and Wassmer ( 1999 ) 37, proposed to utilize the reverse criterion normal distribution to the pvalue at each phase for acquiring a standardized normal statistic. By uniting these normal statistics, utilizing a weighed amount attack, an overall normal statistic and the corresponding pvalue is achieved.
Adaptive Sample Size Reestimation design based on blinded informations, does non present prejudice or the concluding determination procedure and is recommended by FDA. But to analyze the true intervention size consequence, unblinded SSR should be used and it could present some concerns about type I error rate, statistical and operational prejudice and unity of the test. Use independent Data Monitoring Committee ( DMC ) to execute interim informations analysis for both safety and efficaciousness. Based on DMC recommendations, the test can be continued or stopped for either futility or efficaciousness, decrease/increase the sample size for the following phases to make a statistically important power for the survey
8. Hypothesis/Simulations/Case Study
When planing clinical tests, research workers frequently estimate the sample size from limited information about the discrepancy of the response and the size of the intervention consequence. However, the dependability of the sample size appraisal can be limited due to
 The anterior information may hold been derived from little tests that do non supply robust estimations of the intervention consequence
 The patient population in the current test may differ from that in earlier tests
 Effect size estimation may hold been derived from clinical tests of other drugs in the same category as the trial drug.
In such instances, we can utilize different attacks like Cui, Hung and Wang Method ( CHW,1999 ) 16 ; Chen, DeMets and Lan Method ( CDL,2004 ) 28 ; Fisher ( GSD,1998 ) 20 to cipher sample size during the class of survey to accomplish coveted power against an alternate hypothesis derived from the roll uping informations. For the intent of this paper, we will cipher sample sizes and power for the four methods, Fixed Sample, Fisher ‘s GSD Method, CHW Method and CDL Method and trial for their equivalency utilizing EastAdapt™ Software. All these four methods are based on unblinded analyses and are able to continue the conditional type I error.
Let NCHW, NCDL represent sample sizes for CHW and CDL Methods severally. We will prove for the equality in samples sizes and matching power at different values of ? , for these two methods.
Our hypothesis statement for this paper would be
H0N: NCHW = NCDL
h3N: NCHW ? NCDL
For the interest of this paper, I consider a randomized, twoarm survey design to find efficaciousness for the experimental intervention arm drug in comparing with the industry criterion intervention for bipolar I depression. The primary end point is average alteration from baseline to hebdomad 20 in the MontgomeryAsberg Depression Rating Scale ( MADRS ) for mensurating the badness of depression symptoms.
Let ?t = difference between the average MADRS at baseline and the average MADRS at hebdomad 32 for the intervention arm
?c = the corresponding difference of agencies for the control arm.
efficaciousness addition, ? = ?t ? ?c
? , the betweensubject criterion divergence, = 2.
the minimal clinical relevant difference, ? ? 0.6.
Standard divergence and minimal clinical relevant difference are based on the published material9
For this peculiar test, void hypothesis H0: ? = 0 and the reversible alternate hypothesis, H1: ? ?0.
Due to the uncertainness about the true value of ? , the patron wants to carry on the test at ? of 0.7
Consequence of Power with relation to try size and type I error can be summed up as
^ Sample Size = & gt ; ^ Study Power and vType I error
8.1 Fixed Sample Size Design
Plan 1: Create program 1 utilizing the minimal clinical relevant difference of ? = 0.6, ? = 0.05 ( reversible ) , power = 90 % , given ? = 2.
Plan 2: Create Plan 2, utilizing ? = 0.7, as per patron ‘s petition.
Plan 1 
Plan 2 

? 
sample size 
Power 
sample size 
Power 
0.7 
467 
96 % 
343 
90 % 
0.68 
467 
95 % 
343 
88 % 
0.65 
467 
93 % 
343 
86 % 
0.62 
467 
91 % 
343 
84 % 
0.6 
467 
90 % 
343 
79 % 
Table 2: Operating features of Plan 1 V Plan 2
Looking at table 2, under Plan 1, enroll 467 topics for a clinically important intervention consequence of ? = 0.6 with 90 % power or ? = 0.7 with 96 % power. Under Plan 2, enroll 343 topics for a clinically important intervention consequence of ? = 0.6 with 79 % power. To acquire to 90 % power for program 2, the intervention consequence needs to be at 0.7. So based on a patron ‘s position, if he can merely afford resources ( staffing and money ) for 343 topics, he can merely accomplish ? = 0.6 at 79 % power. Therefore there is a possibility of underpowered survey and might non acquire approved by the regulative governments. The operating features of Plan 1 and Plan 2 are displayed side by side in Table 2 for values of ? between 0.6 and 0.7. By looking at Table 2, if the patron had plentifulness of resources to back up the survey, Plan 1 would clearly be the best option.
Sing the variableness in foretelling the right intervention consequence, and the patron ‘s averseness to apportioning adequate resources for Plan 2, nor his committedness for an overpowered Plan 1, a flexible design with regard to try size can be used. Group Sequential Design and Adaptive Design will be considered as the two options.
8.2 Group Sequential Design
Let us presume that capable registration will be at the rate of 8 per hebdomad and the efficaciousness end point for this test will merely be observed at hebdomad 20 and we will build a GSD with one interim phase and power = 90 % to observe a ? = 0.6 such that if in fact ? = 0.7. , the test will halt early. We will name this Plan 3. Due to the registration rate, there will be a 20 hebdomad spread between registration and survey completion. When measuring the nest eggs achieved by using a GSD these 20 hebdomads should besides be taken into history.
As with GSD designs, with each interim phase, we need to account for the sum of type I error to be spent at that phase along with the timing of interim analysis. Keeping in head the safety profile considerations, allow us presume that we can end the trail for efficaciousness based on informations from 120 topics. Based on these premises, a suited clip point for the interim analysis is hebdomad 30, when we will hold enrolled 240subjects with informations on 120 completers. We will utilize the O’BrienFleming ( 1979 ) 28 halting boundary, to guarantee that type I error is spent cautiously at this phase.
Both Plan 1 and Plan 3 have 90 % power to observe ? = 0.6 with a reversible level0.05 trial and about equal sample size demands. The difference being for Plan 1, there are no early fillet regulations whereas Plan 3, can be used to halt the test early and thereby salvage on sample size.
But for a GSD, we need to take into history the job of ‘overruns ‘ . In our instance, even if the early stopped boundary is crossed at hebdomad 30 on the footing of informations from 120 completers, we still take into history the extra 120 patients that enrolled between hebdomad 20 and 30. The extra 120 patients are referred to as “overruns” . Table 2 shows the impact of overproductions on the expected sample size on GSD.
Plan 1 
Plan 2 
Plan 3 ( GSD ) 

? 
Sample size 
Power 
Sample size 
Power 
Sample SizeNo Overruns 
Sample Size with Overruns 
Power 
0.7 
467 
96 % 
343 
90 % 
457 
456 
96 % 
0.68 
467 
95 % 
343 
88 % 
458 
457 
95 % 
0.65 
467 
93 % 
343 
86 % 
460 
462 
93 % 
0.62 
467 
91 % 
343 
84 % 
461 
465 
91 % 
0.6 
467 
90 % 
343 
79 % 
462 
467 
90 % 
Table 3. Sample Size and Power comparing between GSD and fixed sample sizes.
Looking at table 3, it is seen that GSD offers a modest benefit relation to be after 1. After accounting for overproductions, the expected sample size under GSD ranges between 456467 for matching ? between 0.7 and 0.6, as compared to the fixed sample size of 465. so we can reason that for our current test there is non much advantage over utilizing an GSD in comparing with a fixed sample size design.
8.3 Adaptive Design
This design, which we call as Plan 4 starts out with a sample size of 343 topics as in Plan 2, but takes an interim expression after informations is available on 93 completers. As in GSD, there is no proviso for early fillet of the test, instead the interim analysis is used to analyse the interim informations and supply any recommendations for farther registration. Interim analysis should be planned when the sites are still inscribing topics. Based on our enrolment rate of 8/wk, by 24 hebdomads we will hold 192 topics, with 93 finishing the needed 20 hebdomads of followup for the primary end point. Merely the information from the 93 completers will be used in doing the determination to increase the sample size. After this determination is taken, registration will go on until the coveted sample size is attained. The concluding primary efficaciousness analysis is still based on the full 20 hebdomads of followup informations from all the enrolled topics. This is one of the major differences between GSD and Adaptive design SSR. In instance of GSD, the information from the 120 ‘overruns ‘ is non used in the concluding primary analysis whereas in Adaptive design SSR, every topic contributes to the concluding analysis.
Adaptive design, Plan 4 specifications:
 The initial sample size is 343 topics, with 90 % power to observe ?= 0.7, ? = 0.05 reversible degree.
 An interim analysis is performed after informations is available on 90 completers with 20 hebdomads of followup informations at hebdomad 30 ( 8 * 24 = 192 enrolled, out of which 90 will hold complete informations, with the remainder of 120 still in followup ) .
 At the interim analysis based on the computed conditional power ( i.e. chance of obtaining a
positive result at the terminal of the test, given the informations already observed ) any sample size
alteration is initiated. If the conditional power lies between 30 % and 90 % , the meantime
result is considered promising and the sample size is increased. If the conditional power is less
than 30 % or greater than 90 % no sample size alteration is needed. These are known as the
unfavourable ( & lt ; 30 % ) and favourable zones ( & gt ; 90 % ) .
 If the interim result falls into assuring zone, sample size is recomputed to accomplish the mark 90 % conditional power
 If the recomputed sample size is less than 343 ( unfavourable zone ) , the original sample size of 343 topics is used. If the recomputed sample size exceeds 343 ( favourable zone ) , the sample size is cutoff at 686 topics.
Looking at Figure 6, we can see that the lower limit and maximal samples sizes are 343 and 686 with the mark conditional power at 0.9. For 100,000 simulations, ? = 0.6, outputs a sample size of 410 at 84 % power.
Plan 2 
Plan 4 ( AD ) CHW 

? 
sample size 
Power 
sample size 
Power 
0.7 
343 
90 % 
403 
93 % 
0.65 
343 
86 % 
407 
88 % 
0.6 
343 
79 % 
410 
84 % 
Table 4: Operating features of Plan 2 and Plan 4 ( Adaptive Design ) method.
Looking at Figure 7, the simulations fall in the unfavourable zone, assuring zone and favourable zones 18 % , 28 % and 54 % of the clip severally. Table 4 and Figure 7, reveal that although the overall chance of obtaining a important consequence is merely 84 % at ? = 0.6, this chance jumps to 92 % conditional chance on falling in the promising zone. From table 4, although the power addition is merely 5 % at the disbursal of another 67 patients, for ? = 0.6, there is a important benefit of utilizing adaptative design that is revealed by Figure 7. Sample size is increased if merely when the interim consequences are assuring i.e. , when the conditional power at the interim analysis is greater than or equal to 30 % but less than 90 % . This is the really state of affairs in which it is advantageous to increase the sample size and thereby avoid an underpowered test.
8.3.1 Cui, Hung and Wang Method
We will next discourse the Cui, Hung and Wang or CHW Method ( 1999 ) 16 for adaptative sample size alteration of an ongoing twoarm, Klook group consecutive clinical test. The method is based on doing a sample size alteration, if required, each clip that an interim analysis is performed. The interim monitoring continues in this manner until either a boundary is crossed or the K expressions are exhausted. Since the alterations to the sample size may be based on unblinded analyses of the accruing informations, the trial statistic is non the usual Wald statistic utilized for supervising a conventional group consecutive design. Alternatively the trial statistic is comprised of a leaden amount of incremental Wald statistics with weights that are prespecified and chosen suitably so as to continue the type1 mistake. This trial statistic was proposed independently by Cui, Hung and Wang ( 1999 ) 16 and by Lehmacher and Wassmer ( 1999 ) 37. We shall mention to this trial statistic as the CHW statistic and to this method of doing adaptative sample size alterations as the CHW method. The CHW method is merely valid for adaptative designs affecting informations dependent changes in the sample size.