Clinical Trials andIntervention Studies

The Intervention Study

In a controlled observational cohort study, twogroups of subjects are selected from two populationsthat (hopefully) differ in only one characteristic atthe start. The groups of subjects are studied for aspecific period and contrasted at the end of the studyperiod. For instance, smokers and nonsmokers arestudied for a period of 10 years, and at the end theproportions of smokers and nonsmokers that died inthat period are compared. On the other hand, in anintervention study, the subjects are selected from onepopulation with a particular characteristic present;then, immediately after baseline, the total study groupis split up into a group that receives the interventionand a group that does not receive that intervention(control group). The comparison of the outcomes ofthe two groups at the end of the study period is anevaluation of the intervention. For instance, smokerscan be divided into those who will be subject to asmoking-cessation program and those who will notbe motivated to stop smoking.

Interventions have the intention to improve thecondition of an individual or a group of individu-als. Some examples of intervention studies in pub-lic health research are studies that evaluate theimpact of a program: (a) to promote a healthierlifestyle (avoiding smoking, reducing alcohol drink-ing, increasing physical activity, etc.), (b) to preventHIV-transmission, (c) to start brushing teeth early inbabies, and so on. Ample intervention studies can alsobe found in other disciplines; two examples illustratethis. First, Palmer, Brown, and Barrera [4] report onan intervention study that tests a short-term groupprogram for abusive husbands against a control pro-gram. The two groups are compared with respect tothe recidivism rates of the men regarding abuse oftheir female partners. Second, Moens et al. [1] eval-uated in a controlled intervention study the effect ofteaching of how to lift and transfer patients to nurs-ing students in a nursing school. After two years of

Reproduced from the Encyclopedia of Statistics inBehavioral Science. John Wiley & Sons, Ltd.ISBN: 0-470-86080-4.

follow-up, the incidence risk of one or more episodesof back pain was compared between the two groupsof nursing students.

Controlled clinical trials constitute a separate butimportant class of intervention studies. There, the aimis to compare the effectiveness and safety of two(or more) medical treatments or surgical operationsor combinations thereof. Clearly, now the targetpopulation constitutes patients with a specific diseaseor symptom. More aspects of clinical trials will behighlighted in section ‘Typical Aspects of ClinicalTrials’.

Intervention studies are often applied on an indi-vidual level but they can also be applied on a grouplevel. For instance, promoting better brushing habitsfor children could be done on an individual basis, forexample, by means of a personal advice to the par-ents of the child, or on a group basis, for example, byintroducing special courses on good brushing habitsin school. Intervention studies operating on a grouplevel need dedicated statistical methods. We will startwith the intervention studies on individual level butcome back to intervention studies on group level insection ‘Intervention Studies on Group Level’.

Basic Aspects of an Intervention Study

The first step in any intervention study is to spec-ify the target population, which is the populationto which the findings should be extrapolated. Thisrequires a specific definition of the subjects in thestudy prior to selection. In a clinical trial, this isachieved by specifying inclusion and exclusion crite-ria. In general, the inclusion criteria specify the typeof patients who need the treatment under examinationand the exclusion criteria exclude patients for whichthere will be most likely safety concerns or for whichthe treatment effect might not be clear, for exam-ple, because they are already on another, competing,treatment.

To obtain a clear idea about the effect of the inter-vention, the two groups (intervention and control)should be comparable at the start. More specifically,at baseline, the two groups should be selected fromthe same population – only in that case a differencebetween the two groups at the end of the study is asign of an effect of the intervention. Comparabilityor balance at baseline is achieved by randomly allo-cating subjects to the two groups; this is known as

2 Clinical Trials and Intervention Studies

randomization. Simple randomization correspondsto tossing a coin and when (say) heads, the sub-ject will receive the intervention and in the othercase (s)he will be in the control group. But otherrandomization schemes exist, like block- and strati-fied randomization (see Block Random Assignmentand Stratification). It is important to realize thatrandomization can only guarantee balance for largestudies and that random imbalance can often occurin small studies.

For several types of intervention studies, balanceat baseline is a sufficient condition for an interpretableresult at the end. However, in a clinical trial we needto be more careful. Indeed, while most interventionsaim to achieve a change in attitude (a psycholog-ical effect), medical treatments need to show theireffectiveness apart from their psychological impact,which is also called the placebo effect. The placeboeffect is the pure psychological effect that a medicaltreatment can have on a patient. This effect can bemeasured by administering placebo (inactive medi-cation with the same taste, texture, etc. as the activemedication) to patients who are blinded for the factthat they haven’t received active treatment. Placebo-controlled trials, that is, trials with a placebo groupas control, are quite common. When only the patientis unaware of the administered treatment, the studyis called single-blinded. Sometimes, also the treatingphysician needs to be blinded, if possible, in orderto avoid bias in scoring the effect and safety of themedication. When patients as well as physician(s) areblinded, we call it a double-blinded clinical trial. Sucha trial allows distinguishing the biological effect of adrug from its psychological effect.

The advantage of randomization (plus blinding ina clinical trial) is that the analysis of the results canoften be done with simple statistical techniques suchas an unpaired t Test for continuous measurementsor a chi-squared test for categorical variables. This isin contrast to the analysis of controlled observationalcohort studies where regression models are needed totake care of the imbalance at baseline since subjectsare often self-selected in the two groups.

To evaluate the effect of the intervention, a spe-cific outcome needs to be chosen. In the context ofclinical trials, this outcome is called the endpoint. Itis advisable to choose one endpoint, the primary end-point, to avoid multiple-testing issues. If this is notpossible, then a correction for multiple testing such

as a Bonferroni adjustment (see Multiple Compari-son Procedures) is needed. The choice of the primaryendpoint has a large impact on the design of the study,as will be exemplified in the section ‘Typical Aspectsof Clinical Trials’. Further, it is important that theintervention study is able to detect the anticipatedeffect of the intervention with a high probability. Tothis end, the necessary sample size needs to be deter-mined such that the Power is high enough (in clinicaltrials, the minimal value nowadays equals 0.80).

Although not a statistical issue, it is clear thatany intervention study should be ethically sound. Forinstance, an intervention study is being set up inSouth Africa where on the one hand adolescents aregiven guidelines of how to avoid HIV-transmissionand on the other hand, for ethical reasons, adoles-cents are given general guidelines to live a healthierlife (like no smoking, etc.). In clinical trials, ethicalconsiderations are even more of an issue. There-fore, patients are supposed to sign an informed con-sent document.

Typical Aspects of Clinical Trials

The majority of clinical trials are drug trials. Itis important to realize that it takes many years ofclinical research and often billions of dollars todevelop and register a new drug. In this context,clinical trials are essential, partly because regulatorybodies like the Food and Drug Administration (FDA)in the United States and the European MedicineAgency (EMEA) in Europe have imposed stringentcriteria on the pharmaceutical industry before a newdrug can be registered. Further, the development of anew drug involves different steps such that drug trialsare typically subdivided into phases. Four phases areoften distinguished. Phase I trials are small, ofteninvolve volunteers, and are designed to learn aboutthe drug, like establishing a safe dose of the drug,establishing the schedule of administration, and soon. Phase II trials build on the results of phase Itrials and study the characteristics of the medicationwith the purpose to examine if the treatment shouldbe used in large-scale randomized studies. PhaseII designs usually involve patients, are sometimesdouble blind and randomized, but most often notplacebo-controlled. When a drug shows a reasonableeffect, it is time to compare it to a placebo orstandard treatment; this is done in a phase III trial.

Clinical Trials and Intervention Studies 3

This phase is the most rigorous and extensive partof the investigation of the drug. Most often, phaseIII studies are double-blind, controlled, randomized,and involve many centers (often hospitals); it isthe typical controlled clinical trial as introducedabove. The size of a phase III trial will depend onthe anticipated effect of the drug. Such studies arethe basis for registration of the medication. Afterapproval of the drug, large-scale studies are neededto monitor for (rare) adverse effects; they belong tothe phase IV development stage.

The typical clinical trial design varies with thephase of the drug development. For instance, inphase I studies, an Analysis of variance designcomparing the different doses is often encountered. Inphase II studies, crossover designs, whereby patientsare randomly assigned to treatment sequences, arecommon. In phase III studies, the most commondesign is the simple parallel-group design wheretwo groups of patients are studied over time afterdrug administration. Occasionally, three or moregroups are compared; when two (or more) types oftreatments are combined, a factorial design is popularallowing the estimation of the effects of each typeof treatment.

Many phase III trials need a lot of patients andtake a long time to give a definite answer about theefficacy of the new drug. For economic as well asethical reasons, one might be interested in havingan idea of the effect of the new drug before theplanned number of patients is recruited and/or isstudied over time. For this reason, one might want tohave interim looks at the data, called interim analyses.A clinical trial with planned interim analyses hasa so-called group-sequential design indicating thatspecific statistical (correction for multiple testing)and practical (interim meetings and reports) actionsare planned. Usually, this is taken care of by anindependent committee, called the Data and SafetyMonitoring Board (DSMB). The DSMB consists ofclinicians and statisticians overlooking the efficacybut especially the safety of the new drug.

Most of the clinical trials are superiority trialswith the aim to show a better performance of thenew drug compared to the control drug. When thecontrol drug is not placebo but a standard activedrug, and it is conceived to be difficult to improveupon the efficacy of that standard drug, one mightconsider showing that the new drug has comparableefficacy. When the new drug is believed to have

comparable efficacy and has other advantages, forexample, a much cheaper cost, a noninferiority trialis an option. For a noninferiority trial, the aim is toshow that the new medication is not (much) worsethan the standard treatment (see Equivalence Trials).Currently, noninferiority trials are becoming quitefrequent due to the difficulty to improve upon existingtherapies.

The choice of the primary endpoint can have alarge impact on the design of the study. For instance,changing from a binary outcome evaluating short-term survival (say at 30 days) to survival time asendpoint not only changes the statistical test froma chi-square test to, say, a logrank test but canalso have a major practical impact on the trial.For instance, with long-term survival as endpoint, agroup-sequential design might become a necessity.

Despite the fact that most clinical trials are care-fully planned, many problems can occur during theconduct of the study. Some examples are as fol-lows: (a) patients who do not satisfy the inclusionand/or exclusion criteria are included in the trial; (b) apatient is randomized to treatment A but has beentreated with B; (c) some patients drop out from thestudy; (d) some patients are not compliant, that is,do not take their medication as instructed, and so on.Because of these problems, one might be temptedto restrict the comparison of the treatments to theideal patients, that is, those who adhered perfectlyto the clinical trial instructions as stipulated in theprotocol. This population is classically called the per-protocol population and the analysis is called the per-protocol analysis. A per-protocol analysis envisagesdetermining the biological effect of the new drug.However, by restricting the analysis to a selectedpatient population, it does not show the practicalvalue of the new drug. Therefore, regulatory bodiespush the intention-to-treat (ITT) analysis forward. Inthe ITT population, none of the patients is excludedand patients are analyzed according to the random-ization scheme. Although medical investigators haveoften difficulties in accepting the ITT analysis, it isthe pivotal analysis for FDA and EMEA.

Although the statistical techniques employed inclinical trials are often quite simple, recent statis-tical research tackled specific and difficult clinicaltrial issues, like dropouts, compliance, noninferior-ity studies, and so on. Probably the most importantproblem is the occurrence of dropout in a clinicaltrial. For instance, when patients drop out before a

4 Clinical Trials and Intervention Studies

response can be obtained, they cannot be includedin the analysis, even not in an ITT analysis. Whenpatients are examined on a regular basis, a series ofmeasurements is obtained. In that case, the measure-ments obtained before the patient dropped out canbe used to establish the unknown measurement atthe end of the study. FDA has been recommendingfor a long time the Last-Observation-Carried-Forward(LOCF) method. Recent research shows that thismethod gives a biased estimate of the treatment effectand underestimates the variability of the estimatedresult [6]. More sophisticated methods are reviewedin [7] (see Missing Data).

Intervention Studies on Group Level

Many intervention studies act on the group level; theyare called group-randomized studies. For instance,Murray et al. [2] describe an intervention study toevaluate four interventions to reduce the tobacco useamong adolescents. Forty-eight schools were random-ized to the four interventions. After two years offollow-up, the proportion of children using ‘smoke-less tobacco’ was compared. The proportions foundin two of the four treatment groups were 58/1341 =0.043 and 91/1479 = 0.062. A simple chi-squaretest gives a P value of 0.03. However, this testassumes independence of the subjects. When adoles-cents are motivated in a school context, there willbe a high interaction among adolescents of the sameclass/school, that is, the outcome of one adolescentwill depend on the outcome of another adolescent.Hence, the chi-square test is not appropriate. Anadjusted chi-square test taking the correlation amongthe adolescents into account (see [2]) gives a P valueof 0.18.

In general, the appropriate statistical techniquesfor group-randomized studies need to take the

correlation among subjects in the same group intoaccount (see Intraclass Correlation). This impliesthe use of techniques like Generalized EstimatingEquations (GEE) and random effects models; see,for example, [7] and Linear Multilevel Models andGeneralized Linear Models (GLM).

Further Reading

An excellent source for clinical trial methodology can be foundin [5]. Intervention studies operating on group level gained inimportance the last decade; for an overview of these designs,we refer to [3].

References

[1] Moens, G., Johannik, K., Dohogne, T. & Vandepoele, G.(2002). The effectiveness of teaching appropriate liftingand transfer techniques to nursing students: results aftertwo years of follow-up, Archives of Public Health 60,115–223.

[2] Murray, D.M., Hannan, P.J. & Zucker, D. (1998). Anal-ysis issues in school-based health promotion studies,Health Education Quarterly 16(2),, 315–330.

[3] Murray, D.M. (1998). Design and Analysis of Group-randomized Trials, Monographs in Epidemiology andBiostatistics, 27, Oxford University Press, New York.

[4] Palmer, S.E., Brown, R.A. & Barrera, M.E. (1992).Group treatment program for abusive husbands: long-termevaluation, American Journal of Orthopsychiatry 62(2),276–283.

[5] Piantadosi, S. (1997). Clinical Trials: A MethodologicalPerspective, John Wiley & Sons, New York.

[6] Verbeke, G. & Molenberghs, G. Linear Mixed Modelsin Practice: A SAS-oriented Approach, Lecture Notes inStatistics 126, Springer-Verlag, New York, 1997.

[7] Verbeke, G. & Molenberghs, G. (2000). Linear MixedModels for Longitudinal Data, Springer Series in Statis-tics, Springer-Verlag, New York.

EMMANUEL LESAFFRE and GEERT VERBEKE