Twins have participated in medical research for more than a century (Boomsma et al., Reference Boomsma, Busjahn and Peltonen2002; Craig et al., Reference Craig, Calais-Ferreira, Umstad and Buchwald2020; Hrubec & Robinette, Reference Hrubec and Robinette1984; Martin et al., Reference Martin, Boomsma and Machin1997). Previous reviews have highlighted the unique role twins play in understanding genetic and environmental contributions to disease and have provided many examples of the valuable contribution twin studies have made to medical research (Craig et al., Reference Craig, Calais-Ferreira, Umstad and Buchwald2020; Hrubec & Robinette, Reference Hrubec and Robinette1984). Studies in twins can be used to address questions of gene–environment interactions (Buil et al., Reference Buil, Brown, Lappalainen, Vinuela, Davies, Zheng, Richards, Glass, Small, Durbin, Spector and Dermitzakis2015) and cause versus association (Sjölander et al., Reference Sjölander, Frisell and Öberg2012) and have been used in modern specialist areas of medical science such as epigenetics (Bell & Saffery, Reference Bell and Saffery2012), stem cells (Hibaoui et al., Reference Hibaoui, Grad, Letourneau, Sailani, Dahoun, Santoni, Gimelli, Guipponi, Pelte, Bena, Antonarakis and Feki2014) and microbiome research (Smith et al., Reference Smith, Yatsunenko, Manary, Trehan, Mkakosya, Cheng, Kau, Rich, Concannon, Mychaleckyj, Liu, Houpt, Li, Holmes, Nicholson, Knights, Ursell, Knight and Gordon2013). However, clinical trials conducted entirely in twin populations, or ‘twin-only trials’, have been limited to date. A recent review on the participation of twins in clinical trials found that among 186,027 clinical trials registered in ClinicalTrials.gov, only six trials restricted participation to twins (Sumathipala et al., Reference Sumathipala, Yelland, Green, Shepherd, Jayaweera, Ferreira and Craig2018). This finding highlights the potential for conducting more clinical trials in this unique population.
The purpose of the current narrative review is to demonstrate the substantial benefits and address the key challenges of conducting twin-only trials. We consider the design, analysis, recruitment and ethical issues that arise in such trials, including how to randomize twins, the impact of twins on sample size calculations and statistical analysis methods for estimating treatment effects, the utility of monozygotic (MZ) and dizygotic (DZ) twins for studying variation in outcomes, factors affecting recruitment, and the unique ethical considerations. We also discuss the advantages and disadvantages of conducting twin-only trials. For simplicity, we focus on phase III, parallel group randomized trials with an intervention and a control group, although many of the issues discussed also relate to more complex designs.
Design Issues
Randomization of Twins
When trials are conducted in twin populations, participants from a twin pair can be randomized to the intervention or control group in one of three ways: using the co-twin control intervention design, cluster randomization or individual randomization (see Figure 1).
Different methods of randomization for twins in twin-only trials
The ‘co-twin control intervention design’ (Plomin & Haworth, Reference Plomin and Haworth2010) involves assigning one twin from each pair to the intervention group and the other twin to the control group. It is an example of the classical matched pairs design and has long been recognized as the most powerful design for twin-only trials and an efficient alternative to studies in unrelated individuals (Christian & Kang, Reference Christian and Kang1972) because treatments can be compared within twin pairs. It provides near-perfect control (for MZ twins, who share almost all their genetic material) or partial control (for DZ twins, who share 50% of their genetic variation on average) for key covariates, including those that are not measured. This removes some of the noise from the treatment effect estimate and hence this design requires the smallest sample size (Yelland et al., Reference Yelland, Sullivan, Price and Lee2017). The reduced sample size increases the feasibility of recruiting from the relatively small twin population and is most useful in settings where the trial is very expensive, or few twins meet the inclusion–exclusion criteria. While the co-twin control intervention design is clearly the ideal design for twin-only trials for the reasons mentioned above and has been chosen specifically for its efficiency (e.g., Pinheiro et al., Reference Pinheiro, Ho, Ferreira, Refshauge, Grunstein, Hopper, Maher, Koes, Ordoñana and Ferreira2016), it is also the least preferred approach among potential participants and their caregivers, primarily due to the potential psychological trauma associated with twins in different treatment groups achieving different outcomes (Bernardo et al., Reference Bernardo, Nowacki, Martin, Fanaroff and Hibbs2015). Recruitment may therefore prove difficult for trials using this design and alternative randomization methods may be needed.
Cluster randomization, where both twins in a pair are assigned to the same treatment group, is useful in settings where there is a high risk of treatment group contamination (Donner & Klar, Reference Donner and Klar2000). For example, if twins are assigned to different treatment groups in an unblinded trial, the twin assigned to the intervention group may discuss details or share some part of the intervention with the twin assigned to the control group. Treatment group contamination can also occur in blinded trials involving active and placebo treatments that are identical in appearance and taken in the home environment, since treatments could get mixed up for twins assigned to different groups who are living together. Both parents of twins and adult twins have a strong preference for cluster randomization due to the increased chance that both twins will achieve similar outcomes (Bernardo et al., Reference Bernardo, Nowacki, Martin, Fanaroff and Hibbs2015) and hence recruitment may be easier for trials that utilize this method of randomization. In some trials, cluster randomization is the only option, as both twins will necessarily receive the same treatment due to the nature of the intervention (e.g., a parental tool designed to increase children’s acceptance of vegetables; Fildes et al., Reference Fildes, van Jaarsveld, Wardle and Cooke2014).
Individual randomization, where each twin is randomized independently, is useful in settings where there is a need to balance the risk of treatment group contamination and the size of the trial. In such a design, approximately half of all twin pairs will be assigned to the same treatment group and hence treatment group contamination is not an issue among these pairs. The remaining twin pairs will be assigned to different treatment groups and hence treatments can be compared within these pairs. This design is more powerful than the cluster-randomized design but not as powerful as the co-twin control intervention design (Yelland et al., Reference Yelland, Sullivan, Price and Lee2017).
In practice, the best way to randomize twins depends on the individual trial, and the main advantages and disadvantages of each approach are summarized in Table 1. Among 50 twin-only trials identified in a recent review, only 6% used individual randomization, while randomizing twins to different treatment groups (66%) and cluster randomization (26%) were more common (Sumathipala et al., Reference Sumathipala, Yelland, Green, Shepherd, Jayaweera, Ferreira and Craig2018).
Advantages and disadvantages of different methods of randomization for twins in twin-only trials
Sample Size Calculations for Estimating Treatment Effects
At the trial design stage, sample size calculations are typically performed to estimate the number of participants needed to detect a particular treatment effect based on a specified level of power and significance, as well as a set of assumptions about the data (Kirby et al., Reference Kirby, Gebski and Keech2002). Standard sample size calculations assume that observations are independent, but observations collected in a twin-only trial are expected to be similar or correlated within twin pairs. Ignoring this correlation in sample size calculations can result in a trial that is underpowered (for a trial using cluster randomization) or overpowered (for a trial using the co-twin control intervention design; Yelland et al., Reference Yelland, Sullivan, Price and Lee2017), both of which are a poor use of resources and may therefore be considered unethical (Carlin & Doyle, Reference Carlin and Doyle2002).
Sample size calculations for twin-only trials can be performed using a two-step process. First, the sample size is calculated using standard methods that assume the outcomes of all trial participants are independent (Kirby et al., Reference Kirby, Gebski and Keech2002). Second, this sample size is multiplied by a quantity known as the design effect, which measures the degree of inflation required to account for dependence in the data. The design effect depends on trial-specific factors, including how twins will be randomized, how the data will be analyzed and the expected correlation between outcomes of twins. Equations for the design effect have been published elsewhere (Yelland et al., Reference Yelland, Sullivan, Price and Lee2017), and sample size calculators that account for the correlation between outcomes of twins are freely available (Yelland et al., Reference Yelland, Sullivan, Collins, Price, McPhee and Lee2018).
Including MZ and/or DZ Twins
An important decision to make when designing a twin-only trial is which type(s) of twins to recruit. MZ pairs are more closely matched than DZ pairs. In a co-twin control intervention design, each MZ pair may be considered an approximate counterfactual, as this is the closest we can get in real life to both treating and not treating the same individual at the same time. The strong correlation between MZ twins makes this a highly efficient design. Although twins in DZ pairs are not as closely matched as MZ pairs, and hence provide a less ideal real-life counterfactual, the matching may still be relatively close. Including DZ twins will increase the required sample size compared to a trial restricted to MZ twins due to the lower correlation (Yelland et al., Reference Yelland, Sullivan, Price and Lee2017) but will help with meeting recruitment targets. Twin-only trials most commonly involve MZ twins only, followed by MZ and DZ twins, with few trials restricted to DZ twins (Sumathipala et al., Reference Sumathipala, Yelland, Green, Shepherd, Jayaweera, Ferreira and Craig2018).
Analysis Issues
Analysis Methods for Estimating Treatment Effects
When analyzing data collected in a clinical trial, the primary aim is typically to estimate the effect of treatment on a set of prespecified outcomes. Standard methods of analysis for comparing these outcomes between treatment groups often rely on an assumption of independence between outcomes of all trial participants, which is violated in twin-only trials. Failure to account for the correlation between outcomes of twins in the analysis can result in confidence intervals for treatment effects that are too narrow or too wide, and false conclusions about the effectiveness of the intervention (Carlin et al., Reference Carlin, Gurrin, Sterne, Morley and Dwyer2005; Yelland et al., Reference Yelland, Salter, Ryan and Makrides2011).
Treatment effects can be estimated in twin-only trials using a range of analysis approaches, including regression models fitted using the generalized estimating equations method of estimation or mixed-effects models. Both approaches are appropriate for analyzing many forms of clustered data, not just data collected on twins, and are discussed in detail elsewhere. Briefly, generalized estimating equations account for the correlation between outcomes of twins implicitly through specification of an assumed pattern of correlation for the data, known as a working correlation structure (Hardin & Hilbe, Reference Hardin and Hilbe2013; Zeger & Liang, Reference Zeger and Liang1986). In contrast, mixed-effects models account for the correlation between outcomes of twins explicitly by including a random effect for each cluster in the analysis model that represents the unique effect of each twin pair on the outcome (Carlin et al., Reference Carlin, Gurrin, Sterne, Morley and Dwyer2005). Simpler analysis methods, such as the paired t test, may also be applicable in some settings, but do not provide the same flexibility to control for or explore the effects of other covariates on outcomes.
Analysis Methods for Studying Outcome Variation
A major benefit of conducting twin-only trials is that the data can be used to study the variation and covariation in outcomes, in addition to the effect of the treatment on outcomes. The estimated variances and covariances can be used to make inferences about random variation in outcomes, the influence of genes on outcomes, and gene–treatment interactions, as illustrated below. We begin by discussing the information that can be gained from the co-twin control intervention design. We then describe how the classic twin model can be applied with the co-twin control intervention design and cluster randomization. We conclude by considering individual randomization and the role of opposite-sex twins.
Individuals in a pair assigned to different treatment groups
If the co-twin control intervention design is used, the trial can provide information about variation in outcomes between similar individuals when treated and untreated. If residual variation differs between treated and untreated individuals after adjusting for mean differences due to treatment effects, this may provide useful information about the utility and effects of the treatment. For example, in an early co-twin control intervention design involving 20 pairs of MZ girls who were randomized to receive a 6-month daily calcium and vitamin D supplement or a placebo (Greene & Naughton, Reference Greene and Naughton2011), the standard deviation for some measurements was larger in the supplement group versus the placebo (e.g., 52.3 mm2 vs. 39.1 mm2 for trabecular area). If significantly different, this would be consistent with variation in the size of the treatment effect between individuals (although this was not formally tested).
The classic twin model for the co-twin control intervention design
The classic twin model (Fisher, Reference Fisher1951; Hopper & Mathews, Reference Hopper and Mathews1982) can be applied in twin-only trials, including both MZ and DZ twins, to make inferences about the causes of variability in an outcome. If there is evidence that the correlation between MZ twins is greater than the correlation between DZ twins, this is consistent with additive genetic effects (A) > 0 and genetic effects contributing to variation in outcomes under the equal environments assumption (Hopper, Reference Hopper, Spector, Snieder and MacGregor2000; Scurrah & Hopper, Reference Scurrah and Hopper2019). If the MZ and DZ correlations are similar, this is consistent with either A = 0 or unequal environmental covariance for MZ and DZ pairs, although there is no way to formally test which of these explanations is more appropriate. This type of analysis was used in a recent co-twin control intervention design involving 44 twin pairs (34 MZ and 10 DZ pairs) who were randomly assigned to an 8-week low-fat or high-fat diet (Costanzo et al., Reference Costanzo, Nowson, Orellana, Bolhuis, Duesing and Keast2018). The trial investigated the genetic and environmental factors influencing sensitivity to fatty acid taste at baseline. Baseline correlations for MZ and DZ pairs were similar (r MZ = .33, r DZ = .29, p value for difference = .41), suggesting that environment rather than genetic factors is the primary influencer of fat taste sensitivity.
The classic twin model for cluster-randomized trials
When cluster randomization is used, the classic twin model can be used to study variation in outcomes after adjustment for treatment and measured covariate effects, as well as variation in baseline measures. This approach was used in a recent cluster-randomized crossover trial (the STRUETH trial) to assess whether genetic effects influence variation in response to exercise, and whether the contribution of genetic effects depends on the type of exercise (Marsh et al., Reference Marsh, Thomas, Naylor, Scurrah and Green2020a, Reference Marsh, Thomas, Naylor, Scurrah and Green2020b; Thomas et al., Reference Thomas, Marsh, Maslen, Scurrah, Naylor and Green2021). Twin pairs (30 MZ and 12 DZ pairs) were randomized together to either resistance or endurance training, trained together for 3 months, underwent a washout period, then crossed over to the other training regime for 3 months. For resting heart rate in response to endurance exercise, there was low correlation between outcomes for MZ pairs (r MZ = .12), suggesting that within-subject variability (i.e., variation that is not due to shared twin factors, such as genetic or environmental factors) may be high and within-subject repeatability may be low (Marsh et al., Reference Marsh, Thomas, Naylor, Scurrah and Green2020b). If there is substantial between-subject variation for two individuals with identical genes and very similar environments, there is likely to be substantial variation in outcomes for the same individual experiencing the same treatment at different times. This finding is strengthened if DZ twins are included in the trial and the correlation between DZ twins is lower than the correlation between MZ twins, as seen in the STRUETH trial (r DZ < .01 for resting heart rate in response to endurance exercise). The other plausible explanation for such a finding is that a covariate that frequently differs between twins in a pair and that is associated with the outcome has not been adjusted for. In contrast, if both MZ and DZ correlations are high, this suggests that within-subject repeatability is high and within-subject variability is low. If the MZ correlation is higher than the DZ correlation, as it was for baseline V02Max (r MZ = .92 and r DZ = .78; Thomas et al., Reference Thomas, Marsh, Maslen, Scurrah, Naylor and Green2021), this is consistent with genetic influences on variation in outcome if equal environmental correlations are assumed for MZ and DZ twins. When the outcome is a change from baseline, if the MZ correlation is higher than the DZ correlation and the treatment can be considered an environmental (i.e., nongenetic) factor, this suggests a gene–environment interaction.
Twins randomized independently
Under this design, all of the above analysis approaches are theoretically possible, but in practice, power is only likely to be sufficient for large trials (see Hopper, Reference Hopper, Spector, Snieder and MacGregor2000, for discussion on the number of pairs needed to detect shared environmental effects). Few twin-only trials randomize twins independently (Sumathipala et al., Reference Sumathipala, Yelland, Green, Shepherd, Jayaweera, Ferreira and Craig2018) and we are not aware of any examples of this design being used to study outcome variation.
Role of opposite-sex twins
Inclusion of opposite-sex twin pairs in a trial enables assessment of whether the correlation between outcomes differs by sex. If the correlation for same-sex DZ twins is higher than the correlation for opposite-sex DZ twins, this suggests that an interaction between sex and unmeasured genetic or environmental effects may be present. If the correlation for female DZ pairs is different from the correlation for male DZ pairs, this may suggest that different environmental factors affect the outcomes in females and males. However, most twin-only trials have either excluded opposite-sex twins (e.g., Marsh et al., Reference Marsh, Thomas, Naylor, Scurrah and Green2020b) or had too few opposite-sex pairs to enable these types of analyses (e.g., Costanzo et al., Reference Costanzo, Nowson, Orellana, Bolhuis, Duesing and Keast2018).
Recruitment Issues
One challenge that arises when recruiting for twin-only trials is identifying potential participants, as twins are a relatively small subgroup of the population. This may be overcome through the involvement of individual twin registries such as Twins Research Australia (Murphy et al., Reference Murphy, Lam, Cutler, Tyler, Calais-Ferreira, Li, Little, Ferreira, Craig, Scurrah and Hopper2019), the International Network of Twin Registries (Buchwald et al., Reference Buchwald, Kaprio, Hopper, Sung, Goldberg, Fortier, Busjhan, Sumathipala, Cozen, Mack, Craig and Harris2014) or multiple birth associations. A recent review of 50 twin-only trials found that 16% used a twin registry to support recruitment (Sumathipala et al., Reference Sumathipala, Yelland, Green, Shepherd, Jayaweera, Ferreira and Craig2018), highlighting the utility of this recruitment strategy. Since both twins are required to participate in a twin-only trial, it is important to minimize the use of exclusion criteria that may render one twin ineligible.
Another recruitment challenge in twin-only trials relates to obtaining consent. Research suggests that consent of a twin is substantially more likely if their co-twin consents (Ullemar et al., Reference Ullemar, Lundholm, Ortqvist, Gumpert, Anckarsater, Lundstrom and Almqvist2015) and that twins or their caregivers are more likely to consent if both twins will receive the same treatment, at least in the neonatal setting (Bernardo et al., Reference Bernardo, Nowacki, Martin, Fanaroff and Hibbs2015). Consent is especially complex when the twins are minors and their caregivers are the decision-makers due to concerns over the potential outcomes of treatment and the vulnerability of the children. Further research is needed to explore the full range of factors that influence the involvement of twins from all age groups in clinical trials.
Ethical Issues
As with all research involving human participants, the ethical principles of informed consent and justice must guide the recruitment of twins into clinical trials (Beauchamp & Childress, Reference Beauchamp and Childress2019). However, there are additional issues unique to twins that researchers should be mindful of when planning a twin-only trial. First, researchers have an obligation to ensure that individuals are not unduly influenced in their decision to participate in a trial, and that there are no potentially coercive elements to the recruitment procedure. In the case of twin-only trials, if one twin wishes to enroll then this might exert undue pressure on the other twin to join, since the refusal of one twin necessarily excludes the other from participation. Second, twins or their caregivers may feel an unusual degree of pressure to enroll in a trial, given that many benefits of involving twins in research are likely known to the public.
Twin-only trials may raise justice concerns in recruitment. It is generally considered unethical for the burden of research to be borne by one population, while the benefits are conferred elsewhere. A classic example is the use of research participants in low- and middle-income countries to test drugs that are then sold to citizens in highly developed countries (Zumla & Costello, Reference Zumla and Costello2002). For research that is primarily aimed at benefitting twins, there is no issue in exclusively recruiting twins. However, for research that is intended to benefit the general population, it is necessary to ensure that twin populations are not being unfairly burdened. In some cases, this might entail some form of compensation for assuming the risks of research on behalf of the wider population, which may reduce the likelihood that twins feel their special similarity is being used as a commodity.
The ethical principles of beneficence and nonmaleficence aim to ensure that clinical trials are designed to yield the best possible benefit to society while avoiding harm to participants. Given the methodological strengths of twin-only trials, these studies clearly satisfy the first of these principles. The need to avoid harm is the same as in all clinical trials, but with additional considerations related to how the twins will be randomized. While all methods of randomization may be deemed ethical under an assumption of clinical equipoise, the psychological implications of assigning twins to different treatment groups should be taken into consideration, as some participants may experience psychological trauma if twins receive different treatments and achieve different outcomes (Bernardo et al., Reference Bernardo, Nowacki, Martin, Fanaroff and Hibbs2015). Choosing a method of randomization that is more acceptable to participants may be the more ethical approach (see Figure 1 and Table 1).
Advantages and Disadvantages of Twin-Only Trials
The most obvious reason for conducting a twin-only trial is that twins are the target population for the treatment of interest. For example, a new intervention designed to improve outcomes for infants affected by twin-to-twin transfusion syndrome would necessarily be conducted in twins. However, there are many known or perceived advantages of conducting twin-only trials when the target population extends beyond twins. First, enrolling twins may provide some economic advantage if it is faster to recruit and collect household-level characteristics for a set of twins than two unrelated individuals. Second, twins are generally representative of the broader population (Andrew et al., Reference Andrew, Hart, Snieder, de Lange, Spector and MacGregor2001), and hence treatment effects observed in twins may be extrapolated to other individuals. Third, compliance may be better in twins than singletons if twins encourage each other to follow their treatment schedule or complete outcome assessments, though conversely, noncompliance in one twin could increase the risk of noncompliance in the other twin. Fourth, in trials using the co-twin control intervention design, twins provide a natural control for many factors that influence outcomes and hence a smaller sample size is required to achieve the same power as a trial involving unrelated individuals (Yelland et al., Reference Yelland, Sullivan, Price and Lee2017). Finally, twin-only trials can provide useful information about variation in outcomes in addition to treatment group effects. The value of this information needs to be weighed against the increase in sample size required to conduct a twin-only trial if twins will be cluster randomized and the intervention is not specifically targeted at twins.
There are also some known or perceived disadvantages of conducting twin-only trials. First, withdrawals may be a larger problem when studying twins, since the withdrawal of one twin could lead to the withdrawal of the other twin. Second, treatment group contamination may occur within twin pairs assigned to different treatment groups, which can lead to underestimation of the treatment effect and hence an increased risk of missing an effective intervention (Torgerson, Reference Torgerson2001). Third, sample size calculations and analysis methods are more complex for trials conducted in twins rather than singletons due to the correlation between outcomes of twins (Carlin et al., Reference Carlin, Gurrin, Sterne, Morley and Dwyer2005; Yelland et al., Reference Yelland, Sullivan, Collins, Price, McPhee and Lee2018). Finally, identification and recruitment of twins may be more difficult and costly than singletons due to the low prevalence of twins in most disease groups and the potential for adult twins to live in different geographic regions.
Further research is needed to explore these advantages and disadvantages and to identify additional advantages and disadvantages of conducting twin-only trials. The benefits and challenges of conducting a twin-only trial should be carefully considered in the context of the individual trial when deciding whether recruitment should be restricted to twins.
Conclusions
Given the potential advantages of conducting clinical trials entirely in twin populations, we recommend that twin-only trials be considered to complement and contribute valuable additional information to trials involving singletons. We have outlined the design, analysis, recruitment and ethical issues that should be considered when conducting twin-only trials and discussed the advantages and disadvantages of undertaking such trials. In conclusion, conducting clinical trials entirely in twin populations can add important insights into the action of specific treatments and interventions for the benefit of all of society and should be considered more often.
Acknowledgments
The authors thank Professor John Hopper for his contribution to useful discussions relating to this review.
Financial Support
This work was supported by the Australian National Health and Medical Research Council (L.Y., Early Career Fellowship APP1052388), (K.S. and L.C., Centre of Research Excellence APP1079102), (P.F., Career Development Fellowship APP1144311) and (K.L., Career Development Fellowship APP1053609).
Conflict of Interest
None.
Open access
