Reactions to genetic causes often take one of two extremes: either genetic causes are aversive, or they are deterministic and put upon a pedestal. We believe that Madole & Harden's (M&H's) framework and reasoned argument could help to moderate these common reactions. Genetic variants are not uniquely powerful, nor uniquely flawed. The way M&H draw parallels with established social science methods helps to make this point.

We agree with M&H that the mechanism behind a cause can act through a complex network of biological, social, and psychological pathways. For most behavioural traits it is difficult to disentangle the purely biological from the complex interactions between those biological manifestations and environments, which in turn can feed into the causal mechanistic process. Hence, the authors define these traits as having “shallow” causes – which are non-unitary, non-uniform, and non-explanatory. We question whether this is a property of their nature, or rather a limit of our understanding. While we cannot currently fine-map every genetic locus and build a mechanistic model of how each genetic variant acts within a causal network, we shouldn't underestimate the potential for scientific advance. Do the authors think there is potential for genetic causes of complex traits to move from shallow to deep as our understanding improves?

Some traits have genetic causes which are clearly deep, for example, phenylketonuria or cystic fibrosis. Their causative genetic variants explain both mechanism and treatment. However, single-genetic variants with large explanatory effects are rare. Complex traits are usually highly polygenic, with each genetic variant exerting only a small effect. Consequently, M&H suggest that their causes are shallow. But occasionally these single variants can produce deep mechanistic insights, even within a complex causal network. For example, a novel genetic variant for Crohn's disease implicated autophagy as a mechanism (Rioux et al., Reference Rioux, Xavier, Taylor, Silverberg, Goyette, Huett and Brant2007), and individual genetic variants implicated metabolic functions in the development of anorexia nervosa (Watson et al., Reference Watson, Yilmaz, Thornton, Hübel, Coleman, Gaspar and Mattheisen2019). Consequently, novel mechanistic insight can come before we have identifed most of the components in a complex causal network. Therefore, we see the line between shallow and deep to change over time as our knowledge advances, rather than as an inherent property of a complex behavioural trait.

In addition, we propose extending the applications of this causal framework to Mendelian randomisation. If we accept the premise that a genetic variant could be causal for trait A then a natural extension is to test whether trait A causes trait B. Let us take M&H's example: Does education (trait A) reduce crime rates (trait B)? One alternative to a randomised education intervention would be Mendelian randomisation, using genetic variants as a proxy for levels of education. The Mendelian randomisation method operates under similar assumptions (and similar limitations) to those laid out in section 3.2 of M&H, following the principle of “genetic inheritance as a natural experiment” (Davey Smith & Hemani, Reference Davey Smith and Hemani2014). Mendelian randomisation studies (given their speed and relative low cost) are often posited to be a useful first step to guide future randomisation/intervention studies.

Mendelian randomisation estimates are average causal effects and consequently only constitute first-generation causal knowledge. But extensions of the method can help us towards second-generation causal knowledge. For example, multivariable Mendelian randomisation tests possible mediation pathways (Sanderson, Davey Smith, Windmeijer, & Bowden, Reference Sanderson, Davey Smith, Windmeijer and Bowden2019) and factorial Mendelian randomisation tests interaction effects (Rees, Foley, & Burgess, Reference Rees, Foley and Burgess2020). However, these methods can only test possible mediator/moderator variables from genome-wide association studies (GWASs). To obtain sufficient sample sizes for GWASs, the quality of phenotyping is often poor. This currently limits our ability to move beyond shallow causes using Mendelian randomisation methods alone. We must triangulate Mendelian randomisation results with other study designs (Lawlor, Tilling, & Davey Smith, Reference Lawlor, Tilling and Davey Smith2016), which can serve two purposes to help us towards second-generation causal knowledge: (1) replication in another sample/context can help us to understand the durability and consistency of the causal effect, and (2) exploring sources of heterogeneity can improve our understanding of the causal pathway, beyond limited GWAS phenotypes.

M&H also highlight heterogeneity in their framework. They discuss how investigating individual differences in treatment response (i.e., effect heterogeneity) can indicate that the causal relationship is dependent on other factors (moderators). If we can identify these moderators, it can inform us about modifiable intervention targets because it deepens our understanding of the causal mechanism. We strongly support this point, and further argue that the same is true for genetic associations: Understanding heterogeneity here can help to identify modifiable targets in the pathway between genes and behaviour that could inform behavioural (or pharmacological) interventions to draw out genetic strengths and mitigate genetic risks. We wish to highlight this as a crucial future direction for the field, which could be conceptualised as hypothesis-free gene–environment interaction (or indeed gene–gene interaction, but we focus our discussion on environments). Most gene–environment interaction studies focus on a specific environmental variable with a plausible mechanism because this is one way to protect against false positives (Moffitt, Caspi, & Rutter, Reference Moffitt, Caspi and Rutter2005). But with the increasing availability of large sample sizes and more robust statistical methods we could use the heterogeneity in genetic association as a method for identifying relevant (environmental) moderators. One way to do this is to explore why some people are resilient to genetic risk, that is, they have high genetic risk for an outcome but have not yet developed that outcome, akin to the wealth of research exploring protective factors (Armitage et al., Reference Armitage, Wang, Davis, Collard and Haworth2021).

These two suggested extensions of the causal framework are complementary and could be used straight away with available data and appropriate statistical care, to highlight potential modifying exposures for follow-up, either for more deep phenotyping or to include in intervention studies. Both extensions have the potential to move us closer to second-generation causal knowledge. We believe the field should aspire to this, as well as to moving from shallow to deep causes by pursuing mechanistic information. We commend the framework proposed by M&H for providing the foundation to push the field to pursue these ambitious aims.