These chapters were published with modifications by Oxford University Press, 2004

Integrating Genetics Into Randomized Controlled Trials
John P.A. Ioannidis and Joseph Lau

Tables | References

Introduction
The new era of human genetics poses significant challenges to randomized controlled trials (RCTs) for exploiting genetic information in the evaluation of preventive and therapeutic interventions. Genetic information may have two different, potentially complementary uses. First, a genetic parameter may be a predictor of outcomes, such as disease susceptibility, disease progression, target organ disease or toxicity. Second, genetic parameters may modify the postulated positive or negative effects of preventive or therapeutic interventions. A better predictive ability and a more detailed understanding of effect modification both lead to increasing our chances for rationally individualizing treatment (1, 2) to optimize health outcomes.

Individualized treatment is a goal that has been difficult to attain or even approach until recently. RCTs have typically been designed, conducted and analyzed with the view of obtaining average answers for the average patient with a given condition. Subgroup analyses (3) within the population of an RCT and predictive modeling (4) have been seen with skepticism - perhaps justifiably given the high probability of type I error. Multiple comparisons between different subgroups, e.g. those defined by genotype results may yield different results simply by chance. On the other hand, stringently defined genetic subgroups may have too few subjects to be adequately powered in a clinical trial to show the efficacy of interventions, let alone the differential efficacy in different subgroups. Finally, even simple predictive modeling is far from “simple” in any population, including the population of patients enrolled in a clinical trial.

Genetic information may change the design and conduct of clinical trials, and in their turn, clinical trials of genetic tests themselves may also change our appreciation about the importance and uses of this information. With the explosive development of human genetics, a challenge of clinical trials would be to evaluate whether the availability and use of genetic information is actually effective and improves outcomes in clinical practice. Trials aiming to study such effects face several challenges that are worthwhile understanding and overcoming, whenever possible.

The integration of genetic information in RCTs has potentially important implications for the practice of medicine. Guideline development for clinical practice is increasingly based on RCTs. Genetic issues should be given due weight not only when designing and analyzing RCTs, but also when the evidence from RCTs is being critically appraised for guideline development and clinical practice.

In this chapter, we shall try to examine the advent of various uses of genetic information in RCTs and the unique opportunities as well as problems that are generated in this very challenging interface. The chapter is divided in two parts. The first part deals with general issues of genetic factors in the design and evaluation of RCTs, and the second part deals with the application of RCTs to the evaluation of genetic tests.

PART A. Genetic Factors in the Design and Evaluation of Randomized Clinical Trials

Integrating Genetics in the Design of Randomized Clinical Trials
Selection of Study Populations and Eligibility Criteria. Genetic parameters may allow a more accurate definition of various diseases that have a strong genetic component. They may be helpful in defining disease conditions where genetic factors or gene-environment interactions may be equally important or even more important than environmental and acquired parameters. Accurate definition of a condition is essential and may improve the efficiency of a study design when the intervention is likely to work only in patients with that condition. While disease conditions in the past have been defined mostly on the basis of phenotype, definitions based on detailed genotyping information are likely to become more common. The selection of the genetic group to be targeted may be made based on prior evidence from predictive risk models or from biological rationale. Since the validity of genetic risk factors is often not fully studied and the underlying biology may be poorly understood, one should be prepared for surprises. For example, the PARIS study (5) was designed to include only patients who have the DD genotype for the angiotensin-I converting enzyme deletion/insertion (D/I) allele. The investigators had background evidence that this genotype may be a risk factor for restenosis after coronary stent implantation. Thus they targeted the DD high-risk group in order to assess whether quinapril, a drug that blocked angiotensin-I converting enzyme, would reduce the risk of angiographic restenosis. Coronary stenting was performed in 345 consecutive patients and genotyping showed that 115 of them had the DD genotype. Ninety-one of them were randomized to quinapril vs. placebo. Paradoxically, the trial showed that quinapril was associated with a significantly greater reduction in lumen diameter in this selected population.

Stratification and A Priori Adjusted Analyses of the Collected Data.
An alternative approach to the one presented above, would be to avoid limiting enrollment to subjects with specific genotypes, but design a trial with explicit a priori stratification according to genotype. Or, specify a priori that the final analyses will adjust for genotype and the comparison of the treatment effects between subgroups with different genotypes will be a key endpoint of the trial. Stratification has been a contentious and controversial issue in RCTs (6). In theory, randomization in very large RCTs should help obviate concerns about an imbalance between the compared groups. Nevertheless, even if overall imbalance is avoided, it is still conceivable that the absolute magnitude of the treatment effect may be substantially different in high vs. low risk subgroups (7), especially if these subgroups differ markedly in their baseline level of risk. Strong risk factors would thus need to be considered for stratification. Imbalance is more likely to occur in smaller RCTs, but in this case stratification during the randomization phase may be more cumbersome. The use of adjusted analyses and subgroup analyses may overcome the lack of stratification by evaluating differences between subgroups at the analysis stage.

With the exception of clear-cut monogenetic diseases where one gene is directly responsible for the disease outcome, most proposed associations between genetic polymorphisms and diseases to-date are represented by relatively small effect sizes (relative risks 0.2 to 5). Such risk factors may be less influential in creating imbalance during randomization or differentiating the baseline risk. Nevertheless, it may be useful to account for them upfront in the study design, when possible. If the genetic risk factors are known to be “strong” ones (i.e. they have large risk ratios associated with them and are not uncommon), at a minimum they should be included in the description of the baseline characteristics, to show that the compared arms are similar in that regard.

There is no absolute rule on what constitutes a “strong” risk factor at the population level. The attributable fraction is one potential approach to quantify the overall importance of a risk factor at the population level. Attributable fraction is given by PR(RR-1)/[1+PR(RR-1)], where PR is the prevalence of the risk factor and RR is the risk ratio associated with it. This means that for a prevalence of 40% and a risk ratio of 1.5, the attributable fraction is 0.167, suggesting that 16.7% of the disease or outcome of interest may be attributed to the risk factor. One might suggest that it may be unnecessary to take seriously into account genetic risk factors that have an attributable fraction associated with them of less than 0.05, while those with higher attributable fraction may require more attention. Other more formal approaches have also been proposed to model at the design stage the expected variability of risk of the population targeted by an RCT based on the prevalence and risk ratios of known risk factors (8).

Clinical trialists may often routinely stratify for parameters that are unlikely to reflect the disease risk, such as gender or clinical site. In such cases, it may be better to stratify or adjust for genetic or other parameters that are stronger determinants of risk.

Integrating Genetics in the Analysis of Randomized Clinical Trials
In most cases, genetic information is incorporated in clinical trials at the analysis stage, most often in the setting of secondary and exploratory analyses. Many times, the post hoc nature of the analyses is unavoidable, since new genetic polymorphisms may be identified or their role may be speculated after the commencement of an RCT or even after an RCT has been completed. Such post hoc research should be seen with the same caution that should accompany any kind of exploratory research. The essential issue is that finding need cautious replication since they are largely hypothesis generating. Even though the study population may carry the name of a “randomized trial”, in fact the study population is treated in a manner where the emphasis of the comparisons is not on the randomization process. While non-randomized designs may tend to agree on average with randomized studies (9-11), over-interpretation may be more common with non-randomized studies (12).

Disease Association Studies
The population of an RCT may well be used as a sample for performing disease association studies. The classic study design in this setting is a case-control experiment and since the study sample is derived from a large population, the terminology “nested case control” design is more accurate. For example, in an early application of this approach, one group of investigators (13) evaluated a sample of 619 of the 12,866 participants of the Multiple Risk Intervention Trial (93 with death from coronary artery disease, 113 with nonfatal myocardial infarction and 412 matched controls). Patients with and without coronary disease outcomes were compared in regards to the allele frequencies of the apolipoprotein E gene (epsilon 2, epsilon 3, epsilon 4).

Besides main genetic effects, case-control studies may also investigate situations where a specific polymorphism may be considered to be a modifier of the effect of an environmental or other risk factor. For example, a group of investigators (14) performed a nested case-control study using 141 cases of nonfatal myocardial infarction and 270 matched controls from the study population of the Helsinki Heart Study, a primary prevention trial. The selected subjects were genotyped for the 344C/T polymorphism of the gene encoding aldosterone synthase (CYP11B2). The investigators found that the polymorphism was not a strong risk factor for myocardial infarction. However, they suggested that there may be a strong interaction with the effect of smoking. Overall, smokers had a relative risk of myocardial infarction of 2.5 compared with non-smokers. In the presence of the 344CC genotype, this relative risk became 4.67, while in the presence of 344TT homozygosity, this relative risk dissipated to 1.09. Gene-gene interactions represent more convoluted “second-order” effects. They may provide insights to pathogenetic mechanisms such as the involvement of specific genes in modulating environmental risk factors. However, given their more complex nature, the risk of false positive findings is probably higher than for “first-order” relationships.

For randomized trials that possess a robust design and adequate archiving of clinical information and blood or tissue samples, a multitude of nested case-control studies may be performed. Such studies may extend the scientific value of the original randomized research effort. In this regard, RCTs are similar to prospective cohort studies that may also be utilized for nested case-control designs. This creates the need for improved archiving and adequate data banks to be supported during the conduct of clinical trials. At the other end, data banking may be expensive and meaningless if performed without purpose and if case-control studies are performed in a haphazard fashion without some underlying biological and clinical rationale. Such efforts may likely lead to false positive spurious, and clinically misleading, findings.

The ethics of performing subsequent case-control studies should also be considered. Ideally, at the time of consent for DNA storage as part of an RCT, subjects need to know what is going to be tested, thus the design of the RCT and its consent form should take this into account. Trying to obtain additional consent at a later stage is often difficult, since the RCT population may be difficult to reassemble, once the trial is completed. On the other hand, consent needs to be generic enough so that the study investigators could have the option of testing new genetic polymorphisms. Striking a balance between protecting the rights of trial participants and satisfying the growing needs of genetic association research may not be straightforward. Procedures should be improved, standardized, and, when possible, simplified, to protect patient confidentiality without hindering research.

Predictive Modeling of Disease Risk - Genetic Predictors of Study Outcomes
The population of the RCT may be used to evaluate specific polymorphisms as predictors of the outcome of interest. The outcomes of interest may be hard clinical endpoints, such as disease progression or death, or surrogate endpoints such as laboratory or genetic markers.

For example, one group of investigators (15) examined whether polymorphisms of the genes for apolipoprotein B, apo AIV, lipoprotein lipase and cholesterol ester transfer protein may be associated with a greater change in dense LDL cholesterol in a crossover RCT study population treated with two dietary interventions of 4 weeks each (high saturated fat diet vs. high polyunsaturated fat diet). Of the polymorphisms tested, only the Q360H polymorphism in the apo AIV gene was significantly predictive of the change in dense LDL cholesterol. In cases such as this, randomization is not really any longer the essential feature of the study design. The RCT design simply serves to provide a population with fairly standardized exposures to important parameters, in this case diet, that may be influencing also the outcome of interest. The RCT is treated as a cohort study.

Soft biological outcomes may sometimes be misleading for use in clinical decision-making (16). However, sometimes the interest of such analyses may be more focused on making pathophysiologic investigations rather than deriving clinical inferences. For example, basic research may be performed on tissue samples from randomized patients. One group of investigators (17) found that the angiotensin II type 1 receptor A1166C polymorphism is a significant predictor determining KCL-induced angiotensin II responses in excess segments of the internal mammary artery in both the experimental and the control group of a randomized study comparing an angiotensin converting enzyme inhibitor vs. placebo in patients undergoing bypass surgery. Such information may yield helpful pathophysiologic support for further hypothesis testing.

Effect Modification
Effect modification is probably the most challenging feature in the integration of genetics into RCTs. The old question is “can we find out which patients are likely to benefit more from a specific therapy.” If benefit is measured in an absolute scale (e.g. absolute risk reduction), then for a treatment that achieves a consistent relative risk reduction at all levels of baseline risk, the absolute benefit is likely to vary substantially across patients in different categories of risk. For example, a risk ratio of 0.7 for mortality translates to a 0.3% absolute risk reduction for death when the baseline risk of death is 1%, while it translates to a 6% absolute risk reduction for death when the baseline risk is 20%. Thus predictive modeling with individual patient data, including genetic and other predictors, may provide in essence evidence for effect modification in an absolute risk scale. However, usually effect modification as a term is reserved for cases where different subgroups (of different or even similar baseline risk) show significantly diverse relative responses to treatment. Such subgroups may be defined by genetic parameters. Specific genetic parameters may separate patients who benefit differentially from the same treatment, regardless of whether these parameters affect also the prognosis in the absence of treatment.

Most of the effect modification work to-date in genetics has not examined hard clinical endpoints, such as survival, but surrogate laboratory and biological parameters. The reason is probably that for hard endpoints such as survival, very large trials are required to show main effects and the sample sizes required to show effect modification are even larger - and thus largely prohibitive. There are several examples of postulated effect modification with surrogate markers however. For example, one group of investigators (18) found that in a sample of patients enrolled in a RCT of maintenance antiretroviral treatment, the haplotypes of the CCR5 and CCR5 promoter genes might be determinants of the magnitude of decrease of plasma human immunodeficiency virus RNA in response to potent antiretroviral therapy. In another example, in the Lipoprotein and Coronary Atherosclerosis Study (19), the investigators detected a strong significant genotype-by-treatment interaction in the relative response of total cholesterol, low-density lipoprotein cholesterol and apoliporotein B with fluvastatin vs. placebo. Patients with the DD genotype had greater reductions in these lipid parameters with fluvastatin than placebo-recipients in the same randomized arm. In a study of lisinopril vs. placebo in patients after renal transplantation, lisinopril had a beneficial effect on LV mass index reduction, and the effect was more prominent in patients with the DD genotype (8.4% vs. –7.2%) than in the other two genotypes [ID and II] (2.8% vs. –11.4%) (20).

Occasionally, effect modification may be seen predominantly in the control group rather than in the treatment group. For example, a RCT examined the influence of the PvuII polymorphism of the estrogen receptor alpha gene on the response of bone mineral density in post-menopausal women treated with hormone replacement therapy vs. placebo (21). Overall, bone mineral density fared better in women given hormonal replacement than in those who got no placebo. Moreover, in the group of patients receiving hormonal replacement therapy, the bone mineral density change was not affected by genotype. Conversely, in the control group, the loss of bone mineral density was larger in the PP and Pp genotypes (6.4% and 5.2%, respectively) than in the pp genotype (2.9%, p=0.002). This information is fairly similar to what can be obtained by predictive modeling in a uniformly untreated cohort of patients. It can be used to select the patients who might not have to be treated, especially when the available therapy is potentially toxic or controversial for other reasons.

Predicting Adverse Drug Reactions
A useful potential application of genetics in clinical trials is to identify genetic parameters that can be used to select patients who have the best or worse tolerance of toxic treatments. For example, it might be possible to detect patients who have a worse reaction to specific chemotherapeutic agents. A study (22) based on subjects from the St. Jude’s Children’s Research Hospital Protocol Total XII addressed whether mercaptopurine therapy intolerance is associated with polymorphisms within the thiopurine S-methyltransferase gene. 6-mercaptopurine causes accumulation of thiopurine nucleotides. The investigators found that dose reductions due to toxicity were ubiquitous in patients homozygous in thiopurine S-methyltransferase enzyme activity deficiency, occurred in 35% of those with heterozygosity and were very uncommon for wild-type patients (7%) (based on phenotyping and confirmed also with genotyping in a subset of patients). Thus such knowledge, if validated, could be used in selecting the starting dose of 6-mercaptopurine for individual patients. In a different field, another group of investigators (23) found that homozygosity for the insertion allele (II) of the angiotensin-I-converting enzyme gene affects the cough threshold in patients treated with an angiotensin converting enzyme inhibitor, cilazapril, in a crossover placebo-controlled trial. To maximize statistical efficiency in the comparison of interest, the investigators only recruited those subjects with II and DD genotypes (homozygous for the insertion allele or for the deletion allele). The knowledge of an increased susceptibility to cough among II individuals may be used to select whether an angiotensin converting enzyme inhibitor or an agent from a different drug class should be used in specific individual patients, when several alternative regimens of equal efficacy are available.

Assessment of Generalizability
An interesting frontier where genetic information may have applications in RCTs is the assessment of the generalizability of the trial results. For diseases where genetic factors are strong predictors of the disease outcome or an effect modifier for the impact of a treatment, determination of the genetic profile of the study population may provide insight on whether the results may be generalizable to other patient populations. Several genetic polymorphisms have explicit diversity in their distribution in different racial or ethnic subgroups. This means that effective treatments that depend on the presence of a specific genetic polymorphism may not be generalizable to ethnic or other subgroups where this polymorphism is missing or is encountered in low frequency. In a different approach, if a trial shows surprisingly no efficacy for an intervention, post hoc genetic testing of the study population might lead to hypotheses about specific parameters in the genetic profile of the study populations that might be more amenable to a new treatment. These are hypothesis-generating findings, and should always be interpreted with due caution. Nevertheless, genetic information could contribute further in the assessment of the external validity of treatment recommendations derived from the interpretation of randomized evidence.
Table 15-1 summarizes the aspects of RCT design and analysis where genetic information could be used, in a similar, although perhaps more informative fashion, as other more traditional parameters that have been used to-date for these purposes.

Caveats in the Use of Genetic Information in Clinical Trials
Several caveats must be pointed out in the use of genetic information in the design, analysis and interpretation of results of RCTs. Several of these caveats pertain to the use of genetic information in other settings as well, while others may be more specific to RCTs.

Validity of Genotyping. Genetic information may sometimes suffer from low accuracy, regardless of whether it is performed as part of an RCT or for other purposes. Reasons could include lack of internal validation, lack of blinding in the assessment of the genetic test, a large test failure rate and many gray measurements, and large observer variability (24).

Linkage Disequilibrium. Genetic markers in linkage disequilibrium may complicate the interpretation of genetic associations or effect modification that is observed in RCTs. The observed relationships may not reflect a true association with direct pathophysiologic consequences, but may result from linkage disequilibrium of the tested genetic marker with some other unknown or unprobed marker that is the one truly responsible for the association or effect modification (25).

Heterogeneity in Linkage. The strength of linkage between genetic markers may vary in different samples and patient populations. This may result in further heterogeneity in the strength of the detected genetic relationships in RCTs and may lead to a low ability to replicate the findings in other RCTs or generalize their interpretation for clinical use.

Definitions of Clinical Trial Endpoints and Outcomes. Unclear definitions of outcomes or “moving the goal posts” may generate spurious associations in genetic analyses involving RCTs. This problem occurs for both genetic and non-genetic parameters. One of the great design advantages of RCTs is the fact that outcomes should ideally be specified upfront in a specific and accurate manner. This is in contrast to hypothesis-generating epidemiologic research where outcomes may be (appropriately) manipulated in search of new associations. In some occasions, genetic research in RCT patient populations may examine new outcomes that may or may not be robustly defined. For newly conceived outcomes, a population of subjects derived from an RCT does not offer any clear advantage to a population derived from a well-designed epidemiologic non-randomized cohort (26).

Surrogate Outcomes. Surrogate outcome may give us clear insights about a pathophysiologic process. In some occasions they may clearly replicate the findings of hard clinical endpoints. However, RCTs have often been misled by surrogate endpoints that were not validated with corresponding clinical outcomes. As clinical trials become more linked with molecular medicine, use of biologic markers as endpoints is only likely to increase. Several of these biologic endpoints present problems of validation, replication, reproducibility in measurements, random error, and a complex correlation pattern with each other. Given their potential multiplicity, issues of multiple comparisons should also be considered as a potential problem.

Non-randomized Uses of Rrandomized Study Populations. As we stated above, most of the situations where genetics have interacted with RCT research to-date have involved the use of the RCT population or samples thereof in ways that the advantage of randomization is lost. Such research should be seen more in the context of non-randomized semi-experimental designs rather than randomized experiments and inferences should therefore be appropriately more cautious.

Multiple Comparisons. As discussed above, in genetic studies within RCTs, there may exist a multiplicity of outcomes, and a multiplicity of potential subgroup comparisons. To complicate matters, most diseases with a genetic background are likely to have very complex genetic patterns. Thus, there may be a multitude of potential putative genetic markers to be probed for disease association or effect modification. Some of the mutation sites are polymorphic, i.e. they may be several variations at the same site. Let us consider for example a very simple genetic polymorphism where there are only two different alleles, A and a. The number of potential genotypes is 3, i.e. AA, Aa, and aa. The number of potential genetic contrasts is 5: AA vs. others, Aa vs. others, aa vs. others, AA and aa vs. Aa, a allele vs. A allele. For a genetic polymorphism with 3 alleles, the number of potential genotypes is 6 and the number of potential contrasts increases exponentially. In exploratory analyses, all of these contrasts may be analyzed and one or more of them may show statistical significance that may simply reflect type I error. The situation is further compounded by the fact that there may be variations at multiple sites within the same gene. This creates a plethora of possible genetic comparisons. Large-scale testing in genetics (27), although exciting, may further increase the problem of type I error. Multiple comparisons with sparse data for many haplotypes may often lead to spurious results (28, 29).

Replication of Findings. Given the above caveats, replication of findings is essential in genetic epidemiology (30). This applies to all aspects of genetic associations including those derived from randomized studies. Empirical evidence suggests that the findings of subsequent research tends to have a greater likelihood of disagreeing with the results of the original research on a genetic polymorphism, when the first studies are of small sample size, and when more subsequent evidence accumulates. Functional data, evolutionary conservation and biological plausibility should also be considered in determining which polymorphisms should be tested first and are likely to be most important, but it is unclear how much they improve the validation potential of genetic association studies.

PART B. Randomized Trials Evaluating the Clinical Use and Impact of Genetic Information
Randomized trials are considered the reference standard for evaluating medical technologies. Genetic tests are a rapidly expanding area of biotechnology that is being rapidly introduced into clinical care. However, in most cases, the supporting evidence for the introduction of genetic tests into routine care may be lacking or suboptimal. It is estimated that currently more than 700 genetic tests are already available or in late research development (GeneTests; www.genetests.org). RCTs have hardly ever been performed to document that these tests are warranted and have beneficial consequences when applied in specific clinical setting.

Prerequisites for Randomized Trials
RCTs are likely to be performed for tests that are candidates in possessing some meaningful clinical utility. In order for a genetic test to have clinical utility it must meet several requirements. We discuss these requirements in the context of performing and interpreting RCTs that evaluate the clinical use and impact of genetic tests.
First, accurate and reproducible routine methods must be available for the determination of the genetic trait of interest (31). It is conceivable that highly experimental, novel methods may be used in hypothesis-generating studies of genetic disease association or effect modification. However, when a genetic test reaches the stage of clinical use, it must be standardized and routinely applicable with adequate accuracy and reproducibility. It would be difficult to make inferences about the use of a test in the general population, if the assays used during clinical development cannot yet be applied to the general population. Designing a clinical trial to assess the usefulness of a screening strategy that depends on a non-standardized test may result in low generalizability of the trial findings.

Second, the trial population must be readily identifiable and usually should be a rather limited/circumscribed group of subjects. Otherwise it is unlikely that the test would be cost-effective, unless the disease is very common in the general population. The frequency of the disease-related genotype(s) or allele(s) in the screened population is an important consideration. For a rare genotype, even a good test with high sensitivity and specificity may have relatively limited positive and/or negative predictive value. Thus, the eligibility criteria should be carefully selected in an RCT appraising a genetic test.

Third, the diagnostic test under study must be acceptable to the target population. Issues related to acceptability include costs, perceived and actual side effects, ease of administration and test accuracy, especially its false positive rate. There are substantial ethical and social issues involved in genetic testing. These issues are often latent and difficult to eliminate (32-35). Genetic testing may provoke anxiety and sometimes result in psychological harm, insurance and employment discrimination and worsening personal, family and social relationships. False negative results may also have grave consequences as they may convey a false sense of reassurance to the misled patient and this could result in postponement of diagnosis or of use of indicated therapies in the future. All of these “side-effects” are difficult to measure in a clinical trial or other study design, but they should not be neglected in the interpretation of the results.

Fourth, the genetic test should ideally be a strong determinant of the disease process or a potent effect modifier of the response to available treatment. Nevertheless, genetic markers with modest effects may still be worthwhile as screening targets, if they have a high prevalence in the screened population. In this setting, the attributable fraction associated with them may still be substantial. For weak and rare, silent genetic traits, clinical trials may not be feasible to perform since they would require the screening of very large number of subjects and a very large sample size of test-positive subjects in order to have adequate power to detect differences in outcomes with different approaches.

Fifth, effective and acceptable prevention or treatment options must be available for subjects where the test is positive and therapy should be possible to initiate promptly. Also, genetic effect modification may be more useful to know when there are several alternative preventive or therapeutic regimens and only some of them are affected by the genetic trait. Moreover, given the rapid change in therapeutics in many medical fields, one would have to be cautious about whether long-term trials would yield results that still hold true in a radically modified therapeutic environment by the time they are completed.

Sixth, preventive and therapeutic interventions must be accessible and affordable to the population identified to be at-risk and they should have a favorable cost-effectiveness ratio. Ideally, they should have both short-term and long-term benefits for major disease outcomes. Long-term benefits may be more important to document, but they are likely to be more difficult and expensive to study with an RCT design. Nevertheless, it is unclear whether observational research can ever supplement and cover the lack of long-term randomized data in this field (11,12).

Other Design Considerations
The design of RCTs to assess specific genetic tests is still at its infancy. Studies of primary prevention screening and early interventions are intuitively the most attractive, given the theoretically anticipated larger gains of primary prevention vs. late interventions. However, these studies are also the most challenging given the need for very large sample size and long-term follow-up. The theoretical promise of preventive medicine may not be justified when tested in real life. The design of such trials poses challenges similar to those faced in the conduct of long-term RCTs in nutritional chemoprevention (e.g. with various antioxidants) that have started appearing in the literature during the last decade. Moreover, additional problems may arise. For example, patient preferences may be an important obstacle to randomization. Or, large genetic heterogeneity may make guidance difficult to standardize. Finally, long-term follow-up may be problematic and associated with high rates of loss to follow up or voluntary crossover of subjects into the opposite study arms.

One may discuss some of the issues that arise in trying to implement screening for hereditary breast and ovarian cancer. There is some evidence that for BRCA1 and BRCA2 screening, subjects may have strong preferences both in regards to genetic testing and in regards to subsequent interventions (36). A study has found (37) that positive results in BRCA1 and BRCA2 screening tended to reinforce the intention towards prophylactic surgery among women who were already leaning towards this intervention; however, women who were reluctant to have surgery upon study entry, were still reluctant after testing and counseling. Consent for randomization might be difficult to obtain for testing the comparative merits of different preventive or therapeutic options. Differences between options may be subtle in the short-term, but more clinically meaningful in the long-term, when major events start accruing. However, maintaining a largely asymptomatic trial population under routine follow-up for very lengthy periods of time may be unrealistic.

Decision analysis has been used in order to model some of the decisions that may be involved in genetic testing and the actions derived from the genetic information. The inferences of such models may illustrate some of the problems that may be faced by RCTs in these areas. For example, a decision analysis compared prophylactic mastectomy, bilateral prophylactic oophorectomy, tamoxifen and no intervention for women with breast cancer and BRCA1 or BRCA2 mutations (38). It found that the three interventions increased life expectancy by 0.6-2.1, 0.2-1.8, and 0.4-1.3 years over a horizon of 10 years for the baseline scenario of a 30-year old woman with early breast cancer. However, the results were substantially sensitive on the penetrance rate of the mutation. The differences between the three interventions may be difficult to study unless one had a very large sample size. Even documenting the superiority of these interventions over no intervention at all in an RCT would still require a large sample size and long-term follow-up. In some cases, decision analysis may help decide whether an RCT is desirable at all in a specific population. For example, a different group of investigators (39) found that BRCA1 and BRCA2 screening would not benefit women without a family history or early breast cancer, because the pre-test probability is very low and surgical prophylaxis is largely undesirable. Conversely, up to 2 quality-adjusted life years may be gained in women with a family history or early breast or ovarian cancer.

RCTs may also be designed to examine what are the relative merits of genetic testing vs. using some other technology or a combination of various technologies. The same challenges apply here as in the case of testing vs. no testing comparisons. Examples include whether screening for familial adenomatous polyposis of the colon should use genetic testing for mutations in the implicated APC gene or colonoscopy; or whether screening for familial hemochromatosis should involve genetic analysis of the hemochromatosis HFE gene or iron studies. Both questions have been approached with decision analysis modeling that suggests the superiority of genetic testing for both examples (40, 41). Questions comparing technologies may be even more difficult to subject to the rigorous standards of randomized evaluation and may require even larger numbers of subjects, since the differences are likely to be smaller than in test vs. no test comparisons. While modeling approaches are a useful substitute in this setting, one is left with the wish that actual randomized evidence were available.

Trials of Educational and Counseling Approaches in Genetic Testing
We need to learn more about the proper implementation of genetic testing for different conditions and the value of adjunctive educational and counseling measures. Modern medical practice in many developed countries is moving away from physician-initiated prescriptions and towards a greater emphasis on patient-initiated choices. Patients have prompt access to vast amounts of medical information through various sources, including in particular the Internet. Such information may be loaded with errors (42). Genetic information may be difficult to comprehend. Health care consumers may often misunderstand genetic testing and there may be misconceptions about the actual implications of a genetic test. For example, patients may overestimate the diagnostic ability of a test. Or, they may perceive a positive test as a sign of irreparable “genetic doom”. Given this situation, it is important to study the optimal approaches towards enhancing the appreciation and use of genetic testing by health care consumers. This is a promising field that is suitable for randomized trials.

For example, a randomized trial (43) evaluated pretest education regarding BRCA1 testing vs. education plus counseling vs. a waiting-list (control) condition among women at low to moderate risk with a family history of breast or ovarian cancer. Both education and counseling led to increases in overall knowledge, but only counseling heightened the perception about the limitations and risks of BRCA1 testing. Neither intervention changed the intention of women to have BRCA1 testing and about half of the women eventually gave a blood sample. In another trial (44), written and video information was found to be equally effective in providing information about cystic fibrosis carrier screening and achieved high levels of subject-matter knowledge. This might suggest that information technologies may often substitute effectively face-to-face education and counseling, but this may not hold true in all circumstances and for all genetic tests.

RCTs may also study the setting where a genetic test should be recommended and/or implemented. Genetic tests often have implications that extend beyond the individual and affect also couples or whole families. This may generate differential reactions to genetic testing recommendations depending on whether information is conveyed to an individual, a couple or a family. For example, one group of investigators (45) randomized offering counseling and carrier testing for cystic fibrosis either to pregnant women in the first instance (stepwise screening) or to couples upfront (couple screening). The two groups differed significantly in transient and late anxiety levels and in the false reassurance rates among subjects testing negative.

Concluding Comments
Implementation of randomized research in the field of genetics is difficult and challenging, but not unfeasible. A genetic test needs to be evaluated rigorously as any other diagnostic technology. The cost-savings or the wasted expenses associated with the use of a genetic test may rival any other diagnostic technology, especially when one considers genetic tests that target the general population or large segments thereof. The introduction of genetic tests into clinical practice without some strong supporting evidence is worrisome. While regulatory actions should not strangle this exciting, rapidly expanding field, some more attention should be given towards materializing randomized experiments testing the usefulness of genetic tests. Such research may give us valuable lessons.

1. Glasziou PP, Irwig LM. An evidence based approach to individualising treatment. BMJ 1995;311:1356-9.
2. Ioannidis JP, Lau J. Uncontrolled pearls, controlled evidence, meta-analysis and the individual patient. J Clin Epidemiol 1998;51:709-11
3. Oxman AD, Guyatt GH. A consumer’s guide to subgroup analyses. Ann Intern Med 1992;116:78-84.
4. Altman DG, Royston P. What do we mean by validating a prognostic model? Stat Med 2000;19:453-73.
5. Meurice T, Bauters C, Hermant X, et al. Effect of ACE inhibitors on angiographic restenosis after coronary stenting (PARIS): a randomized, double-blind, placebo-controlled trial. Lancet 2001;357:1321-4.
6. Meinert CL. Design and conduct of clinical trials: course slides. Baltimore: Johns Hopkins University Center for Clinical Trials, 1994.
7. Ioannidis JP, Lau J. The impact of high-risk patients on the results of clinical trials. J Clin Epidemiol 1997;50:1089-98.
8. Ioannidis JP, Lau J. Heterogeneity of the baseline risk within clinical trial populations: a proposed evaluation algorithm. Am J Epidemiol 1998;148:1117-26.
9. Benson K, Hartz AJ. A comparison of observational studies and randomized, controlled trials. N Engl J Med 2000;342:1878-86.
10. Concato J, Shah N, Horowitz RI. Randomized, controlled trials, observational studies and the hierarchy of research designs. N Engl J Med 2000;342:1887-92.
11. Ioannidis JPA, Haidich A-B, Lau J. Any casualties in the clash between randomised and observational evidence? BMJ 2001;322:879-80.
12. Ioannidis JP, Haidich AB, Pappa M, et al. Comparison of evidence of treatment effects in randomized and non-randomized studies. JAMA 2001;286:821-30.
13. Eichner JE, Kuller LH, Orchard TJ, et al. Relation of apolipoprotein E phenotype to myocardial infarction and mortality from coronary artery disease. Am J Cardiol 1993;71:160-5.
14. Hautanen A, Toivanen P, Manttari M, et al. Joint effects of an aldosterone synthase (CYP11B2) gene polymorphism and classic risk factors on risk of myocardial infarction. Circulation 2000;100:2213-8.
15. Wallace AJ, Humphries SE, Fisher RM, Mann JI, Chisholm A, Sutherland WH. Genetic factors associated with response to LDL subfraction to change in the nature of dietary fat. Atherosclerosis 2000;149:387-94.
16. Fleming TR, DeMets DL. Surrogate endpoints in clinical trials: are we being misled? Ann Intern Med 1996;125:605-13.
17. van Geel PP, Pinto YM, Voors AA, et al. angiotensin II type 1 receptor A1166C gene polymorphism is associated with an increased response to angiotensin II in human arteries. Hypertension 2000;35:717-21.
18. O’Brien TR, McDermott DH, Ioannidis JP, et al. Effect of chemokine receptor gene polymorphisms on the response to potent antiretroviral therapy. AIDS 2000 ;14 :821-6.
19. Marian AJ, Safari F, Ferlic L, et al. Interactions between angiotensin-I converting enzyme insertion/deletion polymorphism and response of plasma lipids and coronary arterosclerosis to treatment with fluvastatin: the lipoprotein and coronary atherosclerosis study. J am Coll Cardiol 2000;35:89-95.
20. Hernandez D, Lacalzada J, Salido E, et al. Regression of left ventricular hypertrophy by lisinopril after renal transplantation: role of ACE gene polymorphism. Kidney Int 2000;58:889-97.
21. Salmen T, Heikkinen AM, Mahonen A, et al. Early postmenopausal bone loss is associated with PvuII estrogen receptor gene polymorphism in Finnish women: effect of hormone replacement therapy. J Bone Miner Res 2000;15:315-21.
22. Relling MV, Hancock ML, Rivera GK, et al. Mercaptopurine therapy intolerance and heterozygosity at the thiopurine S-methyltransferase gene locus. J Natl Cancer Inst 1999;91:2001-8.
23. Takahashi T, Yamaguchi E, Furuya K, Kawakami Y. The ACE gene polymorphism and cough threshold for capsaicin after cilazapril usage. Respir Med 2001;95:130-5.
24. Bogardus ST, Jr, Concato J. Feinstein, AR. Clinical epidemiological quality in molecular genetic research. The need for methodological standards. JAMA 1999;281:1919-26.
25. Reich DE, Cargill M, Bolk S., et al. Linkage disequilibrium in the human genome. Nature 20001;411:199-204.
26. Langholz B, Rothman N, Wacholder S, Thomas DC. Cohort studies for characterizing measured genes. J Natl Cancer Inst Monogr 1999;26:39-42.
27. Risch N, Merikangas K. The future of genetic studies of complex human diseases. Science 1996;273:1516-7.
28. Schork NJ, Fallin D, Lanchbury JS. Single nucleotide polymorphisms and the future of genetic epidemiology. Clin Genetics 2000;58:250-64.
29. Fallin D, Cohen A, Essioux L, et al. Genetic analysis of case/control data using estimated haplotype frequencies: application to APOE locus variation and Alzheimer's disease. Genome Research 2001;11:143-51.
30. Ioannidis JPA, Ntzani E, Trikalinos TA, Contopoulos-Ioannidis JPA. Replication validity of genetic association studies. Nature Genetics 2001;306-309.
31. Holtzman NA, Watson MS. Promoting safe and effective use of genetic testing in the United States: final report of the task force on genetic testing. Baltimore: Johns Hopkins University Press, 1998.
32. Billings P, Kohn MA, deCuevas M et al. Discrimination as a consequence of genetic testing Am J Hum Genet 1992;50:472-482.
33. Rothenberg KH. Genetic information and health insurance: state legislative approaches. J Law Med Ethics 1995;23:312-9.
34. Lapham EV, Kozma C, Weiss JO. Genetic discrimination: perspectives of consumers. Science 1996;274:621-624.
35. Khoury MJ, Thrasher JF, Burke W, Gettig EA, Fridinger F, Jackson R. Challenges in communication genetics: a public health approach. Genet Med 2000; 2:198-201.
36. Weber BL, Giusti RM, Liu ET. Developing strategies for intervention and prevention in hereditary breast cancer. J Natl cancer Inst Monogr 1995;(17):99-102.
37. Miron A, Schildkraut JM, Rimer BK, et al. Testing for hereditary breast and ovarian cancer in the southeastern United States. Ann Surg 2000;231:624-34.
38. Schrag D, Kuntz KM, Garber JE, Weeks JC. Life expectancy gains from cancer prevention strategies for women with breast cancer and BRCA1 or BRCA2 mutations. JAMA 2000;283:617-24.
39. Tengs TO, Winter EP, Paddock S, Agular-Chavez O, Berry DA. Testing for BRCA1 and BRCA2 breast-ovarian cancer susceptibility genes: a decision analysis. Med Dec Making 1998;18:365-75.
40. Bapat B, Noorani H, Cohen Z, et al. Cost comparison of predictive genetic testing vs. conventional clinical screening for familial adenomatous polyposis. Gut 1999;44:698-703.
41. El-Seray HB, Inadoni JM, Kowdley KV. Screening for hereditary hemochromatosis in siblings and children of affected patients. A cost-effectiveness analysis. Ann Intern Med 2000;132:261-9.
42. Jadad AR, Gagliardi A. Rating health information on the Internet: navigating to knowledge or to Babel? JAMA 1998;279:611-4.
43. Lerman C, Biessecker B, Benkendorf JL, et al. Controlled trial of pretest education approaches to enhance informed decision-making for BRA1 gene testing. J Natl Cancer Inst 1997;89:148-57.
44. Clayton EW, Hannig VL, Pfotenhauer JP, et al. Teaching about cystic fibrosis carrier screening by using written and video information. Am J Hum Genet 1995;57:171-81.
45. Miedzybrodzka ZH, Hall MH, Mollison J, et al. Antenatal screening for carriers of cystic fibrosis: randomised trial of stepwise v. couple screening. BMJ 1995;310:353-7.