Interpreting Behavior Genetic Models: Complexity, compression, and the gloomy prospect

Interpreting Behavior Genetic Models: Seven Developmental Processes to Understand. Daniel A. Briley et al. Behavior Genetics, March 2019, Volume 49, Issue 2, pp 196–210, November 22 2018. https://link.springer.com/article/10.1007/s10519-018-9939-6

Abstract: Behavior genetic findings figure in debates ranging from urgent public policy matters to perennial questions about the nature of human agency. Despite a common set of methodological tools, behavior genetic studies approach scientific questions with potentially divergent goals. Some studies may be interested in identifying a complete model of how individual differences come to be (e.g., identifying causal pathways among genotypes, environments, and phenotypes across development). Other studies place primary importance on developing models with predictive utility, in which case understanding of underlying causal processes is not necessarily required. Although certainly not mutually exclusive, these two goals often represent tradeoffs in terms of costs and benefits associated with various methodological approaches. In particular, given that most empirical behavior genetic research assumes that variance can be neatly decomposed into independent genetic and environmental components, violations of model assumptions have different consequences for interpretation, depending on the particular goals. Developmental behavior genetic theories postulate complex transactions between genetic variation and environmental experiences over time, meaning assumptions are routinely violated. Here, we consider two primary questions: (1) How might the simultaneous operation of several mechanisms of gene–environment (GE)-interplay affect behavioral genetic model estimates? (2) At what level of GE-interplay does the ‘gloomy prospect’ of unsystematic and non-replicable genetic associations with a phenotype become an unavoidable certainty?

Keywords: Gene–environment interplay Human agency Personality Cognitive ability Developmental genetics

Complexity, compression, and the gloomy prospect

As a field, behavior genetics has produced substantial knowledge concerning replicable patterns of genetic and environmental influences across the lifespan (Plomin et al. 2016). Heritability is substantial (Turkheimer 2000), but each SNP explains a tiny portion of variance (Chabris et al. 2015). There is some evidence of GE-interplay, even if the empirical data to this point have not identified many replicable examples for G × E. Genetic and environmental effects shift across the lifespan as phenotypes become more stable. Although the statistical and interpretational implications of GE-interplay processes are well-known, the magnitude of each process is not well-known. Worse still, the factors that affect behavior genetic estimates all occur potentially simultaneously and continuously across development, and they may even interact with one another in a nonlinear and highly complex fashion. Researchers can increase the reasonableness of their inferences from behavior genetic models by gaining clarity on what is known and unknown concerning processes that influence parameter estimates. Ruling out potential processes can substantially shrink the number of possible interpretations.

Some basic questions remain difficult to address: what processes led to an estimate of 40% heritability? Was it additive and independent genetic effects, rGE reinforcing initial differences associated with genotype, or some form of G × E? Would heritability have been 40% if the sample was 10 years younger? Would heritability actually be 50% if assortative mating was correctly handled? Numerous papers have been written on the interpretive problems of heritability (e.g., Johnson et al. 2011; Keller et al. 2010; Turkheimer 1998). Our point here is not to retread this ground, but instead to point out the number of considerations required. Each of these considerations can be deconstructed in isolation to infer what the impact would be on behavior genetic models. The real world combines them all simultaneously in different quantities for each phenotype.

In the face of such taxing complexity, a framework with which to visualize the impact of different combinations of structural inputs would be useful. A successful model could generate phenotype levels from the ground up, starting with partners producing offspring with synthetic genomes and environments. One goal could be to identify what sets of model parameters can fill in the gaps identified in this review. As noted, there are likely several plausible sets of developmental parameters that could lead to the empirical results found in the literature. It might be the case that several potential models could produce similar observed trends, such as increasing heritability with age. We view this as a useful demonstration of the potential for equifinality in behavior genetic models, a limitation of the models that could be overlooked due to implicit assumptions about the data-generating mechanisms. A simulation approach would force these assumptions to be explicit and would allow them to be contrasted with other plausible assumptions.

In this context, we may think of phenotype development or the task of individual-level prediction as falling along a continuum of complexity. At one end is perfect simplicity: a change in an input leads to a change in the output every time, and researchers are able to make accurate predictions with easily obtainable and cognizable information. At the other end, it may be the case that there is such complexity that a description of development requires the full history of all variables at all points in time; the data stream is incapable of any compression. Under this scenario, the best anyone can do is record what happens. There is no more efficient way to express the observations, and the observations do not support any interesting predictions. Although behavior geneticists widely acknowledge that the phenotypes under study are complex (i.e., not having a single cause or simple set of causes), less consideration has been given to the potential compressibility of the phenotypes across individuals relative to the set of available variables (e.g., Li and Vitányi 1997; Wallace and Freeman 1987). By "compression," we mean the ability to represent some large set of information in a more compact manner (Braddon-Mitchell 2001; Sayood 2005; Wheeler 2016). To what extent can behavior genetics move from thousands of genetic associations toward a cognizable and useful model of development (see Kendler 2008)? This type of question has emerged most clearly in the literature surrounding the "gloomy prospect."

The need to empirically evaluate the gloomy prospect

Under the limitations of empirical data collection, little behavior genetic research exists that explicitly considers the possibility of the gloomy prospect. Plomin and Daniels (1987, p. 8) described the gloomy prospect as a situation in which "the salient environment might be unsystematic, idiosyncratic, or serendipitous events," ultimately minimizing the possibility that much scientific progress can be made. Turkheimer and Gottesman (1996) used a simulation approach to illustrate the gloomy prospect; small shifts in environmental context completely removed all specific phenotype.environment associations. Turkheimer (2000, p. 163) applied the same gloomy outlook to molecular genetic associations in the real world due to the inherent complexity of development and noted that "the underlying complex causal processes would cause the apparent results [of molecular genetic studies] to be small, and to change unpredictably from one experiment to the next."

The gloomy prospect is discouraging from an empirical standpoint as it implies that the upper limit for scientific progress in predicting and explaining future behavior at the individual-level may already have been reached or be reached without substantially more meaningful progress. If phenotype development is driven by genetic effects that manifest differently across environments that are peculiar to a given individual, then identifying the effect that a genetic variant has on development will necessarily also be idiosyncratic. If true, the clinical utility of genetic or environmental information about individuals will be largely worthless, since a plethora of interdependent factors (many of which are inaccessible due to a failure of measurement over development) must be known before reasonable predictions can be made.

Gloominess falls on a continuum, and how gloomy the prospect of giving an informative behavior genetic account depends on the phenotype. For example, it may be that things are a bit gloomier for personality compared to cognitive ability or anthropometric traits (e.g., Cheesman et al. 2017). If there is no GE-interplay and no other potentially biasing factors, then molecular genetic associations will replicate and the prospects for giving an informative account is not gloomy at all. But if, on the other hand, GE-interplay is extremely large and the effects of any genetic variant are entirely dependent on the (potentially random) environmental context, then it is unlikely that any genetic effect will replicate. This situation would be maximally gloomy. However, most phenotypes likely fall somewhere between these extremes.

We suggest that a plausible starting point for identifying the "gloominess" of a phenotype is to investigate the seven developmental processes highlighted in this manuscript. Put differently, a greater understanding of phenotype processes (i.e., how the phenotype influences engagement with the environment), structure (i.e., how phenotypes covary), and development (i.e., how phenotypes respond to engagement with the environment in the context of other relevant phenotypes across the lifespan; see Baumert et al. 2017). Each of these questions can be addressed with behavior genetic methodology. For example, the field has established the genetic and environmental structure of many related phenotypes. We suggest that gains can be made in overcoming the gloomy prospect by better understanding our phenotypes, that is to say, gaining knowledge not only of genetic and environmental structure, but also of the processes that led to such a structure across developmental time. This work toward explanation is directly relevant to researchers interested primarily in prediction as the gloomy prospect may imply some upper limit on prediction. Evaluating simultaneous GE-interplay will be challenging, but such work could provide important insight into the mechanisms of phenotype growth.

Additionally, progress toward identifying the boundaries of the gloomy prospect could be made by drawing more heavily on animal models. Although the strength of animal models is typically seen as exerting control over environmental experiences, an increasing number of studies use designs in which GE-interplay is possible (Bell and Saltz 2017; Freund et al. 2013). For example, social niche construction refers to the tendency of certain organisms to form social groups partially based on genetic differences (i.e., rGE; Saltz and Foley 2011; Saltz and Nuzhdin 2014). This behavioral tendency has also been found to be context dependent (Saltz 2011) and influence development (Saltz 2013, 2014). More generally, animals exhibit repeatable behavioral syndromes (Bell et al. 2009; Sih et al. 2004), similar to human personality, and a host of tools are available to better explain and predict these patterns (Bengston et al. 2018). This work may be better situated to address major unanswered questions in human behavior genetics, such as potential sources of Gene × Environment interaction. Lee et al. (2018) found relatively few leads on why genetic associations with educational attainment might vary across contexts (although, see Tropf et al. 2017 for an analysis with individual-level data), but the animal literature may offer further clues (see Saltz et al. 2018). Of course, evidence from animal models may be difficult to extrapolate to a phenotype like educational attainment, but the ability to track the effect of GE-interplay on development dynamically and consistently across the lifespan is a major advantage of animal models.