Behavior change following risk assessment of AD was more likely the greater the numerical lifetime risk estimate and following disclosure of a genotype associated with increased risk of AD. We were not able to assess the effect of disclosing an ε4 genotype (relative to the control group) after controlling for numerical risk estimate. When numerical risk estimates were similar, as they were for those in the control group and for those in the intervention group informed they were ε4 negative, there was no difference in self-reported behavior change. This suggests that, contrary to predictions (Chen and Chaiken, 1999
; Marteau and Senior, 2004
; LaRusse et al., 2005
), disclosure of genotype status associated with lower disease risk has no impact upon behavior beyond the impact of any associated numerical risk estimate.
The findings from these analyses illustrate a general problem in trying to isolate the motivational impact of genotype disclosure and indeed other biomarker risk information. The estimates derived from genotype disclosure differ from those derived from other sources both in provenance and the range of magnitudes of the risk estimated. So, for example, randomizing groups to undergo any additional biomarker test, in this case an analysis of APOE genotype, will result in a greater segregation of risk in those subjected to the additional biomarker, leading to the generation of both lower and higher risk magnitudes although overall the risk in the population tested remains the same. The interest, however, is in being able to disentangle the effects of type of test from numerical risk estimates to test the hypothesis that the salience of genotype has an impact on motivation beyond that produced by feedback of the risk estimates generated from genotype. Communicating the results of predictive genetic testing for common complex conditions is difficult, involving the communication of genotype and numerical risk estimates. If communicating genotype status has no motivating effect upon risk-reducing behavior beyond the motivating effect of disclosing the associated numerical risk estimate or if it has a demotivating effect (for example, if it instils a sense of fatalism), then it may be more effective and efficient to not disclose genotype but only the resultant numerical risk.
A further, more general problem with randomized trials designed to assess the behavioral impact of DNA predictive testing (McBride et al., 2002
; Ito et al., 2006
; Sanderson et al., 2008
) is that the main comparison between the intervention and the control group is most often not informative. This is because the intervention group contains two subgroups—one of individuals receiving genotype-positive test results and one of individuals receiving genotype-negative test results. These different test results lead to higher and lower risk estimates that lead to higher and lower risk perceptions (Marteau et al., 2005
). There is, therefore, an expectation that genotype-positive and genotype-negative test results will have opposite effects on behavior, and thus when the two subgroups are pooled to form a genotype feedback intervention group, there is unlikely to be any difference between this and a control group. The solution most often applied to this problem is to conduct subgroup analyses (McBride et al., 2002
; Ito et al., 2006
; Chao et al., 2008
; Sanderson et al., 2008
). However, given that subgroup allocation is not randomly determined, such comparisons are at risk of confounding (Pocock, 1983
). They may also lack the statistical power to detect differences between subgroups.
The role of gender in explaining behavior change represents a further complexity in interpreting the results of the analyses presented in the current study. There is an association between gender and lifetime risk of AD (females tend to have higher risks than males) and, in this sample, between gender and genotype (a greater proportion of females were mutation positive). If behavior change is more likely in women undergoing AD risk assessment, then this may explain the higher rates of behavior change in those receiving higher lifetime risk estimates of AD and those who are mutation positive. While the effects of gender were controlled for in the analyses, the strengths of the gender variable’s associations with genotype and lifetime risk are such that we cannot be sure that some of the apparent effect of the latter pair of variables is not in part due to gender.
The solution to the problems of collinearity and subgroup analyses outlined above may be to consider alternative designs for studies of this type in preference to using increasingly complex methods of statistical analysis. We propose two possible designs. The first is to use explanatory as opposed to pragmatic trials (MacRae, 1989
) in which the risk estimates given in the two trial arms are equivalent, but in only one arm is the provenance of the test revealed as emanating from genotypes. In the other arm the test could be described as an unspecified biomarker test or a test of protein. This would mean that both groups received comparable risk estimates allowing the variable of interest—namely, the provision of risks that stem from an analysis of genotype—to be assessed. While conceptually neat, this design raises questions concerning acceptability and feasibility. The acceptability would critically depend upon the views of clinical ethics committees reviewing such a study. The feasibility would be influenced by how plausible an unspecified bio-marker test would be for study participants. This would require piloting. In addition, further clinical studies assessing the behavioral impact of genotype feedback will provide stronger evidence if they include measures of actual behavior change.
The second type of design for addressing the problems of collinearity involves the use of analog studies, that is, those in which individuals are asked to respond as though they were in a particular situation. This allows variables of interest to be experimentally manipulated either prior to a clinical study or instead of one. While the internal validity of such studies is high, there is also an evidence that their external validity can be acceptable, provided the study mirrors closely the situation it is intended to mimic (Holt and Mazzuca 1992
; Lanza et al., 1997
) with greater validity likely with the use of video-based technologies (Lievens and Sackett, 2006
). In the current context, an analog study might involve asking participants to imagine being given a lifetime risk for AD or indeed any other common complex disease. The risk estimate provided would then vary independently of the type of test and test result. So, for example, those given lifetime risk estimates of AD which were 35% would be randomly assigned to be told that this was based on their genotypes, or not. The outcome variables might include risk perceptions and intentions to engage in risk-reducing behaviors. Results from such studies provide an estimate of the extent to which cognitive and behavioral responses to risk information are predicted by risk estimations and type of test. The impact of describing test results as emanating from genotype-positive or genotype-negative test results could also be assessed by comparing the impact of presenting the risk estimate either with the genotype described as positive or negative, with the impact of leaving the genotype undisclosed. It should be noted, however, that even responses to hypothetical scenarios, however richly drawn, require validation in studies in which individuals respond to actual risk information, with behavior being measured objectively and not by self-report.
We are experimenting with both these designs in a series of studies investigating the motivational impact of DNA testing for common complex conditions (e.g., Wright et al., 2008
If neither of these design options is possible, we would advocate opting for a design tailored to allow a test of the hypothesis in a subset of participants who have similar enough risk distributions in each study group to avoid collinearity. Pilot work could be used to assess the frequency distribution of risk in study groups and therefore the extent of collinearity, and through repeated simulation of the pilot data to identify a range of risk within which choices in study sample size and in optimal allocation ratio to trial arms provide reasonable power for an answer to the question in the restricted risk range. In the REVEAL study, the subset of ε4-positive participants with risk less than 30% was small (eight participants), the collinearity was very high, and the event rate was zero in this subset; this did not affect the ability to answer the primary trial questions, but did prohibit reliable estimation required to address one of the current hypotheses. Design options in future trials could be modified further to allow an assessment of both sets of hypotheses.