Ross and Kendler argue that despite Nancy Krieger’s persuasive criticisms of the distal-versus-proximal framework, distinguishing between distal and proximal causes is still informative in medicine. They are especially concerned about viewing, by default, social causes of disease as distal and biological causes as proximal, showing that in some cases social causes occur in parallel with biological causes and in others social causes are more proximal than biological causes.

In one view, scientific explanation depends on laws, and generalizations in psychiatry let us explain the symptoms of individual patients. These generalizations might themselves be explained by being subsumed under further generalizations. Insofar as individual, idiosyncratic aspects of the patient’s behavior cannot be subsumed under laws, they cannot be scientifically explained. In contrast, though, the psychiatric interview, one of the basic data points for psychiatry, may include the therapist using idiosyncratic details of the patient’s delusions and working through the way in which one stage in the patient’s psychological history generated another. This involves understanding how one thing caused another, but there seems to be no subsumption under laws. Campbell proposes we think of causal explanation in psychiatry as not solely law based but as giving weight to the idea of the patient’s illness over time as a single construct, unfolding in accordance with general patterns. The role of generalizations in psychiatry is to define those constructs that develop over time. Nonetheless, the idiosyncratic detailed causal development of those progressions can be understood at a subjective level by the therapist following the patient’s line of thought and feeling as the disorder develops.

Since around 2010, genome-wide association studies (GWAS) have taken the field of psychiatric genetics by storm. Many of the time-honored methods – family studies, twin analyses, adoption studies – have been largely displaced by the rush toward the identification of genetic variants that could be considered causal of psychiatric illness. At the time of this writing in 2024, the field is awash with results. For key disorders such as schizophrenia, bipolar disorder, and major depression, hundreds of variants on the genome have been found to be associated with risk of illness. What we have learned so far is that for all psychiatric disorders investigated, many common variants in the human genome appear to reflect risk, and the effect size of all of them is small, sometimes very small. Furthermore, taken individually, most of these variants do not have any obvious biological meaning; they do not occur in the parts of the genome that codes for proteins. Often, they are just statistical signals for the block of the human genome (a “haplotype”) with which they have traveled through human – and, before then, primate – evolution.

Francis Galton invented the quincunx, or Galton board, to demonstrate that sums of large numbers of binary events could produce quasi-continuous normally distributed outcomes. The authors of this chapter suggest that the quincunx poses a deeper problem: whether the statistical devices we invent to model complex developmental processes represent the causal basis of development, as opposed to simply being statistical methods that model complex outcomes without explaining their etiology. Turkheimer and Kaplan simulate populations of large quincunx in which the pins have biases: deviations from equal probabilities of bouncing a ball left and right. The simulations reproduce many common results from complex quantitative genetics: quincunx with similar biases produce similar outcomes, and aggregate measures of biases in the pins are correlated with the final placement of the balls. A simulated GWAS, however, that attempted to infer the individual biases of the pins from variations in outcomes was completely unsuccessful. The authors conclude that the simulations are a proof of concept, demonstrating that there are causal structures that will not yield to reductive analysis.

This chapter considers the concept of proximal and distal causes in various health science fields and common reference to the “proximal–distal model” of disease causation. In many cases, this model is used to capture the interaction of social and biological factors, with social factors as distal causes and biological factors as proximal causes. The authors examine Krieger’s (2008) criticism of this framework and suggestion that proximal–distal language confuses causal thinking, reasoning, and distinctions in health science domains. While Krieger (2008) has recommended the elimination of language of proximal and distal causes, Ross and Kendler argue for another alternative, specifically that while the proximal–distal model captures some etiological scenarios, it should not be applied to all cases. They examine three ways in which social and biological causes lead to health outcomes and argue for the value (and possibility) of clear notions of proximal and distal causes.

The subject of Peter Zachar’s chapter is classification, with a focus on psychiatry (of course). He suggests that classificatory choices should be “understood from the perspective of scientific conventionalism.” I agree that there is an important conventional element in current systems for the classification of mental disorders. Rather than seeing this as necessarily a limitation that needs to be overcome, my view is that many scientific theories, including very successful ones, contain conventional elements. (I take Zachar to agree.) In what follows I try to distinguish or disentangle different ways in which a scientific theory can contain conventional elements and to assess the implications of the presence of conventions in psychiatry.

In their chapter, Eric Turkheimer and Jonathan Kaplan assert that Galton’s quincunx “unifies discrete and continuous, binary and normally distributed, random and deterministic, predetermined and epigenetic.” What starts as an anodyne recitation of antinomies, quickly raises questions about how they intend to use the quincunx model as they move from statistical couplets to the contrast of “predetermined and epigenetic.” In so doing, Turkheimer and Kaplan seem to reduce epigenetics, a diverse and complex set of biological processes, into a synecdoche standing in for all nongenetic factors (stochastic and environmental) in development (the pins of the quincunx). Epigenetics refers to a complex (and variably reversible) group of biological processes that regulate gene expression. Epigenetic mechanisms are involved in gametogenesis, the earliest stages of embryonic development, and throughout life. In early development, chromatin (complexes of DNA with histones and other proteins) packages DNA and forms the basis for epigenetic regulation mediated by chemical modifications of the chromatin. Early on, epigenetics establishes cell fate and cell identities; later in development and through postnatal life, epigenetic mechanisms transduce diverse physiological signals that may originate in the body or in the environment to reversibly regulate gene expression. Epigenetic regulation of gene expression represents one of many biological mechanisms (including such important processes as learning and memory) that utilize and build on genomic information to support an organism’s development, homeostasis, adaptation, and change. In so doing these processes, along with chance, increase the heterogeneity (“individuality”) of organisms and significantly degrade the possibility of making precise long-term predictions from DNA sequences. From genetics alone we cannot predict with certainty whether a person will suffer from a mental illness, finish graduate school, or get divorced. Even the effects of Mendelian alleles, such as the precise age at which a carrier will develop Huntington’s disease (Genetic Modifiers of Huntington’s Disease Consortium, 2015), are not deterministically predictable in the context of modifier genes, chance, and additional layers of regulation.

In his chapter Steven Hyman describes several paths forward for a more mechanistic understanding of psychiatric disorder, articulating the promise of such an approach as well as illustrating how it encompasses a variety of projects.

The author of the following chapter, Denny Borsboom, has played a leading role in the rise of the network model in psychopathology. This model has developed into a viable alternative to the older conventional common-factor model. These two models provide quite different explanations for why psychiatric disorders tend to cluster together in disorders. Let us illustrate this using two commonly co-occurring symptoms in individuals with major depression: insomnia and tiredness. The common-factor model assumes that the two symptoms co-occur because they are both caused by a latent construct: the psychobiological state of depression. By contrast, the network model implies a more common (and perhaps naïve) presentation, assuming a direct causal link between the two symptoms, especially in the direction of “I sleep poorly on night x” followed by “during day x + 1 I am tired.”

Campbell’s contribution investigates the possible function of scientific laws in psychopathology, specifically the idea that the standard notion of scientific law could function in the way it has traditionally done in scientific explanation. An important way of thinking about this, as outlined by Campbell, is to introduce scientific laws as general axioms that act as premises in a deductive argument through a covering law. Properties of an individual case then feature as the explanandum. Variants on this approach generally emulate Aristotle’s well-known modus ponens argument scheme that deduces the mortality of Socrates from the premises that all men are mortal and that Socrates is a man. Although many different theories of explanation have been produced over the past century, I think it is fair to say that most working scientists have something like this model in mind when they think about explanations of individual cases.

Symptom network analysis is now commonly used in psychopathology research. Network analysis results in networks with symptoms represented as nodes, while edges represent conditional associations between these. Direct causal relations between symptoms will produce nonzero conditional associations, but these associations can also be produced in other ways. In this chapter, Borsboom discusses six plausible mechanisms that could produce edges in symptom networks. The first is resource competition, where the presence of a symptom depletes resources, which causes another symptom to arise. The second is evidential overlap, in which judgments central to different symptoms involve a subjective assessment of the same evidence. The third is shared mechanisms, in which symptomatology involves processes that are shared among different symptoms. The fourth are consistency drives, which arise when individuals are prone to align their cognitions, affect states, and behavior. The fifth are statistical processes involved in research design and analysis (marginalization and conditioning). The sixth is the presence of unobserved common causes that affect multiple symptoms at the same time. The author argues that, in realistic situations, the mechanisms in question are not mutually exclusive, which preempts standard scientific approaches that pit one model against another to derive critically divergent predictions. Instead, making sense of symptom networks will require more advanced theory development and modeling.