Digital phenotyping of the mind: From biology to psychoinformatics

Abstract

Scientists and engineers in psychoinformatics are developing new ways to capture changes in mental conditions through data generated from people’s interaction with digital devices, especially smartphones. This new approach is called digital phenotyping. It draws on evolutionary biologist Richard Dawkins’ notion of the extended phenotype. However, there is surprisingly little biological thinking in the literature on digital phenotyping of the mind. This article pursues an epistemic critique of digital phenotyping of the mind through an ‘infrastructural inversion’ based on a material-discursive reconstruction. It rereads Dawkin’s work on the extended phenotype. It traces the emergence of a correlational psychology and of psychometric instruments used to validate extended digital phenotypes of mind and behaviour. Alternative ontologies of mental health and disorder are presented to challenge the medical model embedded in current international classifications of mental and behavioural diseases. Digital phenotyping of the mind may be in danger of just reproducing an already problematic medical model when the deeper theoretical background assumptions concerning causality and reverse inference are not properly addressed.

While the stuff from which our world picture is built is yielded exclusively from the sense organs as organs of the mind, so that every [person’s] world picture is and always remains a construct of [their] mind and cannot be proved to have any other existence, yet the conscious mind itself remains a stranger within that construct, it has no living space in it, you can spot it nowhere in space. We do not usually realize this fact, because we have entirely taken to thinking of the personality of the human being, or for that matter also that of an animal, as located in the interior of the body. To learn that it cannot really be found there is so amazing that it meets with doubt and hesitation, we are very loath to admit it. We have got used to localizing the conscious personality inside a person’s head – I should say an inch or two behind the midpoint of the eyes.

Schrödinger (1944/2018, pp. 122–123)

Scientists and engineers in psychoinformatics are developing new ways to capture changes in mental conditions through data generated from people’s interaction with digital devices, especially smartphones. This new approach is called digital phenotyping (Baumeister & Montag, 2019; Insell, 2017; Jain et al., 2015; Spinazze et al., 2019; Torous et al., 2016). It is a booming field, apparently vibrating with excitement about future promises and opportunities.

Given almost omnipresent human-machine interactions, people’s digital traces might allow for diagnosis, prognostic, and intervention activities in different areas of life [emphasis added] such as . . . estimating personality traits, attitudes and preferences (e.g. predicting people’s political orientations by social-network interactions), or improving patients health care (e.g.) estimating disease and treatment trajectories based on ecological momentary assessment data [emphasis added]. (Baumeister & Montag, 2019, p. xiii)

According to Onnela (2021):

Digital phenotyping combines the excitement of fast-paced technologies, smartphones, cloud computing and machine learning, with deep mathematical and statistical questions, and it does this in the service of a better understanding our own behavior in ways that are objective, scalable, and reproducible. (p. 45)

It is claimed that digital phenotyping captures or aims to capture disease phenotypes in an objective, continuous, passive (= more reliable) and ecological manner. Emerging after the completion of the Human Genome Project in the first decade of the 21st century, digital phenotyping is an extension of the genomics revolution that needs a similar revolution in phenomics to fulfil its promises for a personalized precision medicine. It will do so by being able to better correlate objective, measurement-based data on observable expressions of disease with variability in an individual’s genome in genome-wide association studies. The aim of this research effort in digital phenotyping is to overcome the barrier to progress that Onnela (2021) calls the “phenotyping problem,” that is, “our inability to precisely specify phenotypes, the observed manifestations of our genomes within our lived environments, in the individuals we study and treat” (p. 45). Diseases, including mental disorders, are here conceptualized as biological universals, universal across humankind and the globe. “[H]umanity is faced with a common set of diseases and our genes are written in a common alphabet,” Onnela (2021, p. 53) argues. “Mental disorders are global illnesses,” Insell (2017, p. 1215) writes.

Proponents of digital phenotyping draw on evolutionary biologist Richard Dawkins’ (1982/2016a) notion of the extended phenotype to argue for an extension of the notion of phenotype. That is,

the idea of phenotypes should not be limited just to biological processes, such as protein synthesis or tissue growth, but extended to include all effects that a gene has on its environment inside or outside of the body of the individual organism. (Jain et al., 2015, p. 462)

“Just as walking on a beach leaves a trail of footprints in the sand,” Onnela (2021, p. 45) writes:

the use of these [digital] devices generates, as a byproduct, digital trails of social, behavioral, and cognitive footprints . . . Given that [emphasis added] these trails reflect the lived experiences of people in their natural environments, it should be possible to leverage them to develop precise and temporally dynamic disease phenotypes and markers to diagnose, monitor and treat illness. (p. 45)

Given that . . .! It is towards this precondition that I propose we now turn our attention. Much of digital phenotyping’s promises and prospects hinge on it, not least where the fitness and disorders of the mental space—in which consciousness and experience are thought to reside—is concerned. What is it that proponents of digital phenotyping of the mind invite us to accept as a priori ontological and epistemological assumptions? About, for example, the nature and measurability of mental phenomena and disorders? About the endogenous generative matrix underlying external mental and behavioural phenotypes? Or about the allowability of reverse inferences from extended digital phenotypes to the existence, presence, and nature of endogenous mental disorders?

Theory and method

It is not uncommon for scientists pioneering a new approach to take a concept for its superficial associations, adapt it, and repurpose it, for example to brand-label a new field of research that they claim as their own. The danger of brand-labelling schools of thought is that we rarely go back to the original texts to discover the riches that are not captured by the summaries. There is surprisingly little biological thinking in the writings of proponents of digital phenotyping of the mind. If mind is the way the human body developed, then it is important to think more deeply about biological matters. This article is an invitation to do just that.

The aim of this article is to pursue an epistemic critique of ongoing developments in the digital phenotyping of the mind, by way of what Geoffrey Bowker and Susan Leigh Star have called an infrastructural inversion (Leigh Star, as cited in Bowker et al., 2015, p. 477). As a method, infrastructural inversion brings to the surface for critical scrutiny what usually remains hidden, sunk as it is into other structures, social arrangements, and technology. One way to do this is through the use of historical sources in a material-discursive reconstruction that traces changes in the materiality of investigative practices and associated shifts in ontological and epistemological assumptions. Reconstruction, as used here, makes use of the past to better understand the present, but has no ambition or claim to present a full history. It cuts between close shots and distant perspective, connecting and juxtaposing points or events that are disparate in time and space (Hacking, 2002, p. 6). As a form of critique, reconstruction does not produce tales of linear progress towards the present but is sensitive to historical patterns of diversification as well as patterns of conceptual blending or fusion. Like biological organisms, the tools of digital phenotyping of the mind are historical entities, in the sense that they are the result of developmental processes—in material and conceptual-discursive research practices—that allow us to enact (or perform) the world in specific ways rather than depict (or represent) it (for a more detailed methodological account, see Wackers & Schille-Rognmo, 2022).

The argument in this article comprises three parts. First, I propose we start by rereading Richard Dawkins on the extended phenotype. To establish the discursive context to which evolutionary biologist Dawkins contributed, I will briefly review the origins of the conceptual pair of phenotype and genotype. This will bring into focus the epigenetic gap between an observable phenotype and its causal, generative infrastructure. After rereading Dawkins, I will briefly review some of the biological processes that developmental biologists have proposed for the epigenotype, processes that fill in the epigenetic gap. Second, because digital phenotyping of the mind obviously depends on the validity of correlation, I will turn to the way in which correlational sciences in biometrics, psychology, and psychometrics have worked around this epigenetic gap by arguing that knowledge of causal mechanisms is not required for probabilistically inferred knowledge to be valid. Third, I will briefly present alternative ontologies proposed for understanding mental health and disorder. These alternative ontologies challenge and depart from some of the deep epistemic presumptions of the medical model that is entrenched in today’s diagnostic manuals and its associated psychiatric assessment tools. The multiplicity of alternative ontologies invokes additional questions concerning the validity of digital phenotyping and the inferences derived thereof.

Early 20th-century origins of the phenotype concept: Theories of heredity

For a biologically orientated reader, the notion of the phenotype unfailingly evokes its partner in a conceptual couple, the genotype. The Danish botanist and plant physiologist Johannsen (1911/2014) coined the term phenotype in the early 20^th century, in the context of a theory of heredity. Heredity is about transgenerational transmission. For Johannsen it was evident that biology borrowed and repurposed the terms heredity and inheritance from everyday language, “in which the meaning of these words is the ‘transmission’ of money or things, rights or duties—or even ideas and knowledge—from one person to another or some others: the ‘heirs’ or ‘inheritors’” (Johannsen, 1911/2014, p. 989).

In biology, scientists “have again and again tried to conceive or ‘explain’ the presumed transmission of general or peculiar characters and qualities ‘inherited’ from parents or more remote ancestors” (Johannsen, 1911/2014, p. 989). The British naturalist Charles Darwin’s pangenetic theory of heredity proposed that gemmules, hypothetical minute particles of inheritance, from all of the body’s cells and organs assembled in the egg cell that gave rise to the development of the embryo. The German biologist August Weismann argued for an immortal germ plasm that is transmitted from generation to generation by mortal disposable bodies. The formative agent (information?) travels in one direction only. The germ plasm informs the development of the body, but no information from the body travels to the germ plasm. The Dutch botanist Hugo de Vries and other Mendelians rediscovered Gregor Mendel’s work, explaining empirical laws of inheritance in terms of segregation and recombination of pairs of genes. These pairs of genes were called alleles, sitting in homologous places in pairs of chromosomes. In London, Francis Galton and Karl Pearson founded the “biometric school,” arguing that objective knowledge of hereditary could be achieved based on a purely statistical analysis of correlations between features of ancestral and offspring generations. No specific knowledge of the nature and mechanisms of heredity was required. This latter claim was the bone of contention in a heated controversy between the biometricians and the Mendelians (Mackenzie, 1981; Norton, 1978).

Against this backdrop of theories of heredity and based on his botanical work with pure line breeding of self-fertilizing barley and common peas, Johannsen (1911/2014) investigated how group identical seeds, sharing a single genotype, could be grown in different environmental conditions to yield a range of different phenotypes, also called the norm of reaction. Phenotypes, that is, the observable physical appearance of organisms, were, according to Johannsen, the result of interactions between the genotype—conceived as a genetic group identity—and the developing organisms’ environment.

Rereading Dawkins on the extended phenotype

In the 1940s, Darwin’s theory of natural selection was integrated with Mendelian genetics into what is known as the Modern Synthesis (Huxley, 1942). Watson and Crick discovered the double helix structure of DNA in 1953, providing the principles for the genetic code (Watson & Crick, 1953). Dawkins was propelled into fame by his early books on The Selfish Gene (Dawkins, 1976/2016b), and its sequel, The Extended Phenotype (Dawkins, 1982/2016a). In these books, Dawkins takes up position in the theoretical discourse on heredity, development, and evolution sketched so briefly above. His work is an attempt to theoretically reimagine possible answers to some of biology’s most fundamental questions. What is the nature and mechanism of the transgenerational transmission of characteristics (heredity); of evolution and the origin of new species; of the development of organisms from their earliest beginnings to their adult forms; of interactions and relationships between organisms and the environments in which they develop and live? The discourse on these fundamental issues continues today. What is the stuff transmitted from generation to generation in heredity? Germ plasm? Sequences of DNA that we call genes? Or is it the information—lacking an immortal or invariable material form—that is encoded in (relatively stable) DNA? What is the proper unit of Darwinian natural selection? The gene or the genotype? The whole organism or phenotypes? Or is it species? Is the material, behavioural, or mental form of the adult organism in some way already preformed in its genetic make-up, a blueprint instructing the development of the organism? A codified homunculus? Or is development shaped by epigenetic and environmental factors? Is evolution the result of cumulative effects of accidental genetic mutations? Or is it the result of a variety of selective pressures acting on variability in a population?

As an evolutionary biologist, Richard Dawkins is an ardent, card-carrying neo-Darwinian and a neo-Weismannian. As a Darwinian, he is “fundamentally interested in natural selection, therefore in the differential survival of replicating entities such as genes” (Dawkins, 1982/2016a, pp. 314–315). Echoing Weismann’s notions of germ plasm and disposable soma, Dawkins upholds a distinction between replicators and the vehicles that carry them. Although genes as we often understand them, as sequences of DNA coding for the synthesis of a protein, may be regarded as a kind of archetypical replicator, Dawkins’ concept of a replicator is wider, hence his proposal to consider memes as nongenetic replicators “which flourish only in the environment provided by complex communicating brains” (Dawkins, 1982/2016a, p. 165). Neither is, in Dawkins’ world, the notion of the phenotype tied to the idea of a vehicle or body. To the contrary, The Extended Phenotype (1982/2016a) constitutes a long, extended argument “to show that we can emancipate the concept of the phenotypic difference from that of the discrete vehicle [emphasis added] altogether, and this is the meaning of the title ‘extended phenotype’” (p. 298). Note that Dawkins writes about phenotypic differences; natural selection works on differences or variety in a population.

Genes have phenotypic effects that extend beyond the boundaries of the organism as traditionally defined by membrane or skin. This is the idea that digital phenotypers of the mind grabbed and repurposed. Let go of the idea that the phenotype is the set of observable characteristics of an individual’s body, often simplistically and misleadingly exemplified as eye, skin, or hair colour. Dawkins’ extended phenotype includes a spider’s web, a bird’s nest, and a beaver’s dam and lake. Not only physical structures, though. As a student of animal behaviour, an ethologist by training, having done his doctorate under the guidance of Nobel Prize-winning ethologist Niko Tinbergen, Dawkins also included behaviour under the notion of extended phenotypic effects. A gene can have phenotypic effects through the flesh and behaviour of organisms in which it takes residence. The replicators of the virus causing rabies, for example, have effects on the host’s biting behaviour, increasing their own chances of getting into the next body and of being replicated again.

The relationship between replicators and their phenotypic effects is, in Dawkins’ world, not a linear, deterministic, causal relationship. Genes neither determine nor control phenotypes:

If development were preformationistic, if DNA really were a “blueprint for a body”, really were a codified homunculus, reverse development—looking-glass embryology—would be conceivable. But the blueprint metaphor of textbooks is dreadfully misleading, for it implies a one-to-one mapping between bits of body and bits of genome. (Dawkins, 1982/2016a, p. 267)

In Dawkins’ world of the extended phenotype, the causal relationship between replicators and their phenotypic consequences is not linear and unidirectional but recursive.

Genes are allowed to exert their normal effects on development. Their developmental consequences—phenotypic effects—feed back on those genes’ chances of surviving, and as a result gene frequencies change in succeeding generations in adaptive directions. Selective theories of adaptation, but not instructive theories, can cope with the fact that the relationship between a gene and its phenotypic effect is not an intrinsic property of the gene, but a property of the forward developmental consequences of the gene when interacting with the consequences of many other genes and many external factors. (Dawkins, 1982/2016a, p. 269)

Dawkins employs a metaphorical language. He does so consciously and unhesitatingly. Dawkins admits to using metaphors of purpose and metaphors of power. Adaptations have a purpose; they are beneficial for something. What that some-thing is, is the question of the unit of selection in evolution (Dawkins, 1982/2016a, pp. 124–126). Genes that have phenotypic effects in the environment can be said, Dawkins argues, to be genes for that adaptation. Dawkins does not hesitate to say that there are genes for beaver lake size. Metaphors of power have to do with the question of whether the replicator achieves phenotypic effects that increase the probability of its own replication and survival. Phenotypic effects of genes are “levers of power to influence their respective worlds so as to survive” (Dawkins, 1982/2016a, p. 323):

An animal artefact, like any other phenotypic product whose variation is influenced by a gene, can be regarded as a phenotypic tool by which that gene could potentially lever itself into the next generation. . . . [I]n principle there is no reason why the phenotypic levers of gene power [emphasis added] should not reach for miles. (Dawkins, 1982/2016a, p. 304)

In Dawkins’ anti-instructive, selectionist view, the notion of the extended phenotype is inseparably tied to the notion of natural selection, which is “the process by which some alleles out-propagate their alternatives, and the instruments by which they achieve this are their phenotypic effects. It follows that phenotypic effects can always be thought of as relative to alternative phenotypic effects” (Dawkins, 1982/2016a, p. 298). “Most of the replicators in the world have won their place in it by defeating all available alternative alleles. The weapons with which they won, and the weapons with which their rivals lost, are their respective phenotypic consequences” (Dawkins, 1982/2016a, p. 361).

Although technically all environmental consequences could be called phenotypic effects, only those that influence the probability of their replicators’ survival, in competition with and relative to alternatives—natural selection works on differences!—are of interest to the evolutionary biologist. Onnela’s (2021) transient “trail of footprints in the sand” (p. 45) would, for the evolutionary biologist, not be an “extended phenotype” of much interest, unless it can be argued that the footprints in the sand have consequential effects on the survival and reproduction of the replicators that make up the organism that made them.

Inside the epigenetic black box: Canalization, pleiotropy, and degeneracy

Richard Dawkins was not the first nor the only biologist to argue that there is no linear one-to-one relationship between genes and their phenotypic effects. In the mid-20^th century, Waddington (1942/2012) brought together the study of genetics with experimental embryology in a new branch of developmental biology he called epigenetics. “Genetics,” Waddington argued, “has to observe the phenotypes, the adult characteristics of animals, in order to reach conclusions about the genotypes, the hereditary constitutions of which are its subject matter” (1942/2012, p. 10).

Contra Galton and Pearson’s biometric school, Waddington (1942/2012) pointed out the inadequacy of correlative approaches to the study of heredity that “merely . . . assume that changes in the genotype will produce correlated changes in the adult phenotype” and that “the mechanisms of this correlation need not concern us” (p. 10). The causal mechanisms at work were, in Waddington’s view, the kernel of the problem of development. “Between the genotype and phenotype, and connecting them to each other, there lies a whole complex of developmental processes” (Waddington, 1942/2012, p. 10). Waddington proposed epigenotype as a convenient name for this complex, and epigenetics as a name for its study.

In 1957 Waddington discussed some of the results of his epigenetic studies in his book on The Strategy of Genes (Waddington, 1957). He observed, for example, that developmental processes that are disturbed have a tendency to return to the track they were following and yield the same phenotypic end result. He called this phenomenon homeorhesis (Waddington, 1957, p. 32), and the buffering effect it produced canalization (Waddington, 1957, p. 13). Certain developmental processes are, on this view, buffered, to preserve favourable phenotypes on which natural selection works. Ergo, “epigenetic systems . . . can absorb some gene variation without producing any phenotypic effects” (Waddington, 1957, p. 122).

It was also well known to Waddington and his contemporaries that a single gene could influence multiple and divergent phenotypic characters that could influence an organism’s fitness in various ways. This feature was called pleiotropy. Microbiologists recognized that a strain of micro-organisms sharing a single genotype (group identity) could exhibit very different morphological shapes, a phenomenon called pleiomorphy. Obviously, this constituted a major challenge to the classification of micro-organisms (Sapp, 2009). The observation of two or more phenotypic forms in populations of isogenic cells or organisms—that is, cells or organisms with identical genomes—suggested the existence of switch genes, that when activated by internal or environmental factors could produce phenotypic switching—think of the tadpole and the frog, and of the caterpillar and the butterfly (Waddington, 1957, p. 175). Pleiotropy has also been described in protein synthesis.

The opposite of pleiotropy is the long-recognized biological phenomenon of degeneracy. That is the recognition that multiple alternative genetic systems may lead to equivalent phenotypic fitness. Waddington (1957) described degenerate characters as “genetically inhomogeneous” (p. 111). Degeneracy is a ubiquitous biological phenomenon. Edelman and Gally (2001) define degeneracy as “the ability of elements that are structurally different to perform the same function or yield the same output” (p. 13,763). Edelman and Gally (2001, p. 13,764) provide a list of 22 examples of degeneracy at various levels of biological organization, from subcellular organelles to interanimal communication. Degeneracy has been described as a prominent property of gene networks, neural networks, immune systems, and endocrine signalling pathways.

Pleiotropy and degeneracy are defining features of the complexity and adaptability of the dynamic networks (genetic, neural, immune, endocrine signalling pathways) that make up an organism and that allow the organism to tune and prune its relationships with its developmental niche. Of course, an organism’s genome is indispensable for its life, not as a deterministic blueprint or instructor for its phenotypic appearance but rather as a “parts catalogue” that opens up an immense space of possible configurations and activity patterns in body, brain, and mind. An organism does not simply interact with its environment, as if organism and environment were two separate and independently preexisting entities. An organism is “wired,” dynamically and recursively, into its world, exploiting the developmental resources and action affordances that constitute its niche (Barrett, 2020; Clark, 2016; Sterling, 2012, 2020).

Correlation in biology

How can these complex, internal and external, recursive cause–effect relationships underlying mental and behavioural phenotypes be disentangled? Science has not solved this problem, but it has found ways to work around it. The name of the workaround is correlation, or rather, statistical–mathematical expressions of the strength of association between variables. Correlation was developed in England, by Francis Galton and Karl Pearson, to study causal relationships of heredity in populations (Norton, 1978). To appreciate the conceptual shift from biology to information science in digital phenotyping of the mind, it is, however, important to consider first how the notion of correlation was used in biology.

The idea of correlation as it was used in biology came in two versions. The first version had to do with the correlation of organs within a single individual body. All higher organization of multicellular organisms was supposed to evolve through the principle of the physiological division of labour. This process occurred during the developmental process and resulted in a mutuality of the organism’s constituents. Organisms were conceived as co-operative assemblages with parts integrated into organs that live for and by another (Sapp, 2009, p. 20). These relations were studied with developmental methods, that is, by studying the mode and sequence of the formation of the parts of the cell and organism as they come into existence during the life history of the organism.

The second version of correlation in biology had to do with the “affinities” between organs observed by naturalists across a wide variety of groups of organisms, both higher organisms and bacterial types. These relationships were studied by comparative methods. Evolutionary biologists concerned with the development of a natural system of classification (taxonomy) for living organisms tried to distinguish between independently acquired adaptive characters on the one hand, and those which indicate real community of descent on the other. The former was called analogous, the latter homologous. Homology referred to “the same organ in different animals under every variety of form and function” (Sapp, 2009, p. 21). A bat’s wing, a cat’s paw, and a human hand would be homologous structures, indicating common ancestry (heredity), and proper natural relationships in an evolutionary, genealogical sense. Various appendages for flying, like the wings of bats, birds, and bees, would be analogous, mere adaptations to similar environmental conditions, and of secondary importance to the taxonomist. Such independent adaptations to similar circumstances could result in similar forms (phenotypes) by way of convergent evolution. The convergence of forms took place after divergence through speciation had occurred, like for example the (analogous) streamlined body forms of fish and mammals like dolphins and whales living in the ocean.

The shift from biological to statistical correlation

Francis Galton tried to find a way to work around the unknowability of intra-individual and transgenerational causal mechanism responsible for phenotypic expressions by calculating correlations between interindividual differences across populations. Galton’s eugenic mindset and his repurposing of the error theorists’ Gaussian curve as a frequency distribution for physical as well as mental features across a population are well known and do not have to be rehearsed here (Piovani, 2008; Stigler, 1989).

To demonstrate the principles of his statistical correlation methods, Galton (1888) latched onto the biological notion of correlation as mutuality between organs derived from a common cause: the developmental process. As an example, Galton calculated the correlation between stature and cubit, that is, between the height of a person and the length of their underarm, from the elbow to the tip of the middle finger. These are two physical and quantitative attributes of one object, one body. Galton found that he obtained the best statistical correlation when he converted the unit of the measurement scale from inches to a unit derived from the probable error of the anthropometric population data series for the attribute concerned. Galton plotted the frequency distributions for stature and cubit into a single graph to give the correlation a graphical representation. This could be done for any pair of attributes. Galton called the pair the “subject” and the “relative,” concepts that could serve as placeholders for any pair of attributes (variables). Furthermore, Galton found the subject and the relative to be interchangeable. The correlation obtained, independent of which attribute was plotted on the X-axis and which was plotted on the Y-axis.

The correlation between stature and cubit resulted from the developmental process, for which Galton now provided a statistical, numerical value. By calculating the correlation between the same attribute of parental and offspring populations, the statistical correlation provided a numerical value for the degree of heritability of the attribute. Knowledge of the nature of hidden, causative mechanisms of biological development in the individual, Waddington’s epigenotype, or of heredity between generations, like Mendelian genetics, was not required. The whole of causally effective mechanisms and processes interposing between two attributes could be black-boxed and replaced by a number, a mathematical abstraction, to represent them.

A few years later, in an 1890 paper, “Kinship and Correlation,” for the North American Review, Galton (1890/1989) likened the correlations calculated for “normal systems” to the regularities found by mathematicians in ideal geometric forms and their algebraic features, simultaneously suggesting that the method could be applied to solve social problems:

The numerical value of the scale of dispersion identifies a particular normal system just as completely as that of the length of a radius identifies a particular size of circle. Again, as circles have various properties and relations familiar to readers of Euclid, so normal systems of variables have their own peculiar properties, which enable numerous problems to be worked out concerning them, and make it possible to express in precise and definite language all that has been vaguely shadowed forth in the preceding pages about correlation. . . . There seems to be a wide field for the application of these methods to social problems. To take a possible example of such problems, I would mention the relation between pauperism and crime. (Galton, 1890/1989, pp. 85–86)

Galton’s work on statistical correlation was developed further by Karl Pearson (Norton, 1978). The mathematical tools they used derived from a branch of mathematics known as linear algebra, that is, the kind of mathematics that holds for geometrical forms and quantities. Galton’s and Pearson’s methods were in the first decade of the 20^th century adopted by Spearman (1904/1987, 1904) as the tools for a new and scientific correlative psychology. Spearman claimed boldly to have objectively determined and measured the innate and heritable faculty called intelligence, by measuring in school children from various socioeconomic backgrounds just-noticeable differences in the perceptual discrimination of weights, grey-scales, and pitch of sounds.

Naming the (f)actor: From mathematical abstraction to causative agent and probabilistically inferred causality

Spearman recognized that the essential, complex nature of the mental object was not amenable to direct measurement. Spearman rejected introspection as a viable road to the scientific knowledge of the mind. Whereas the mental object remained concealed to direct observation by the scientist, the mental object’s phenotypic attributes were not; these were observable. Spearman followed Francis Galton and Karl Pearson in claiming that neither insight into the nature of the object, nor the causal architecture underlying the pattern of observable, phenotypic attributes, was necessary for a scientific psychology to proceed. Hypothetical causes and causal relationships may be proposed, even argued to be plausible, but they are not necessary for the knowledge to be valid. Spearman strove for a “precise quantitative expression [emphasis added] derived impartially from the entire available data,” aiming for “a more complete acquaintance . . . concerning objective relations [emphasis added]” (1904, p. 225). In other words, objective knowledge could be obtained through the measurement of the strength of association (correlation) between multiple, observable, phenotypic attributes. Spearman rejected “introspection” as an avenue to scientific knowledge of the mind.

For Spearman, human observers were sources of error. He placed his trust in the calculations that mathematical statistics could provide. “The whole of our experimentally gained figures must without any selective treatment simply of themselves issue into one plain numerical value [emphasis added] varying conveniently from 1 for perfect correspondence down to 0 for perfect absence of correspondence” (Spearman, 1904, p. 225).

To reduce the complex system of calculations to fewer dimensions or factors, Spearman applied linear-algebraic tools known as “principle component analysis” or “factor analysis.” Principle component or factor analysis came in two in principle equivalent forms, the equivalence between the two reaching back to Pythagoras and Euclid. One form was algebraic and numerical, the other geometrical.

The geometrical form of factor analysis is perhaps more familiar than the algebraic one, namely as “line fitting in a scatterplot.” The scatterplot consists of data points representing individual measurements plotted as co-ordinates into a two-dimensional Cartesian space. The line fitted to a scatterplot is placed in such a way as to minimize the sum of the square of the distance of each point to the line (least squares method). In cases where the scatterplot is irregular in shape it may be impossible to represent the data by a single line. A second line may be necessary. In Spearman’s time, the second line had to be placed perpendicular to the first line. Each line is a factor. Although knowledge of the nature of causal mechanisms between the variables is no longer required, the lines fitted to the scatterplot are now, in a correlational world, said to explain the data.

The mathematical abstraction of a factor, compressing the multifarious unknown influences between cause and effect into a single number or line in a scatterplot, becomes a causative agent, an actor in the world. A concealed mental entity (faculty or disorder) inside a person becomes the generative cause of observable phenotypic effects. The correlations between externally observable, phenotypic attributes obtain, the argument goes, because they have a common cause, because they are attributes of the same object. Though hidden, or latent, the inferred common cause receives the status of a variable or attribute of the object itself. Although subject and relative can change places in Galton’s statistical correlations, the influence receives a direction by fixing the internal pole of the correlation as the independent variable. Loosing their imagined, hypothetical status, endogenous mental disorders now are the efficient cause of observable, phenotypic symptoms and their extended effects in the environment.

The efficient cause is the only one of Aristotle’s four causes that modern science has retained. Aristotle’s material, formal, and final (teleological) causes have been abandoned. In a cognitive-linguistic sense the efficient cause provides the basic, conceptual placeholder motif of the cause–effect relationship. This cause–effect motif has been blended with numerous other conceptual couples: physiology’s stimulus–response, linguistics’ object–attribute, Galton’s subject and relative, and correlative sciences’ independent–dependent variables. This hyperblend of causality has important entailments. Causes always precede their effects in time. Causes are the antecedents of their consequences. In a spatial dimension there is directionality, a movement from causes to effects. This movement is unidirectional. Furthermore, effects presuppose the existence and agency of antecedent causes, of some kind of causal infrastructure or matrix that generates the effects. When we reject introspection as a viable road to scientific knowledge of mental phenomena, and only have access to third-party observable phenotypes, the causal infrastructure must be inferred by moving in the opposite, reverse direction. We attempt to establish a referential connection to some concealed but causative entity. This is probabilistically inferred causality. If this were possible for the mind, it would constitute, paraphrasing Dawkins, a kind of “looking-glass psychology,” or “looking-glass psychiatry.”

(F)actors are given names. Spearman called his (f)actor for general intelligence, also known as Spearman’s-g. Spearman could present general intelligence numerically as a representation of something natural, innate, and heritable; reifying a hypothetical, imagined epistemic object into an ontological object. An individual’s externally observable performance on what came to be called tests came to serve as a proxy for a concealed internal entity (intelligence). These measures of intelligence could now be plotted in a graph against other variables that were assumed to stand in a causal relationship to the mental object, as either cause or effect: sex, race, ethnicity, socioeconomic status, level of parental education, and so forth. Relating individual performances on intelligence tests with age, for example, and comparing performances between age groups, it became possible to say that an individual of a certain calendar age had mental abilities typical for a younger age group. Hence the concept of mental age used to describe, qualify, and rank individuals at the lower end of the intelligence scale as retarded in their development.

The measurement of intelligence, and other mental features belonging to a domain internal to the subjects studied, became in this correlative psychology a matter of differentiating between individuals in terms of the position they occupied in a linear series. As Rose (1985) put it, “individual scores received their pertinence from the perspective of the population and their relation to its norms” (p. 120). If not shaped by the eugenic mindset of the time, the correlative psychology pioneered by Francis Galton, Karl Pearson, and Charles Spearman certainly had, to use Donald Mackenzie’s (1981) words, “a differential attractiveness and tactical appropriateness” (p. 50) for a eugenic strategy in politics in which a logic of social administration prevailed. Social control and administration required, as a kind of institutional demand, tools for the assessment of individuals in terms of a social norm to decide what to do with them. Correlative psychology provided these assessment tools.

Forced choice psychometrics

Spearman’s work on intelligence prefigured the development of a whole range of psychometric instruments, from which also psychiatry benefited a great deal. Together with a colleague, I have published elsewhere a reconstruction of psychometric instruments embedded in questionnaire-based mental health surveys (Wackers & Schille-Rognmo, 2022). There we note that psychophysicist Leon Louis Thurstone, working in Chicago, replaced Spearman’s psychophysical tests—the measurement of just-noticeable differences in weights, grey-scales, and pitch of sounds—with a rank-ordered list of verbal statements expressing opinions, with which he claimed the measurability of attitudes (Thurstone, 1928). Degrees of agreeability with the verbal statements were scored on a linear, digitized, numerical scale (today known as a Likert scale). Thurstone also ventured into the field of psychopathology with an explorative factor study of the insanities (Thurstone, 1934). In this study, he derived his psychometric instrument from symptom checklists that were used by practising psychiatrists to assess general improvement after treatment. With the introduction of the first psychopharmacological drugs in the 1950s (antipsychotics, antidepressants; Healy, 1997), these symptom checklists were adapted to shorter “effect measures” in clinical trials. Subsequently, following the introduction of the Diagnostic and Statistical Manual of Mental Disorders (DSM), and the International Classification of Diseases (ICD), they were adapted to serve as diagnostic psychiatric assessment scales (for anxiety, depression, etc.) in clinical settings, and as data-generating tools in mental health surveys of general populations.

Throughout this development, a commitment to the physicists’ ideal of “fundamental measurement” of extensive quantities has been retained, including a never-proven assumption of empirical additivity and scalability of mental phenomena that the notion of a quantity entails (Michell, 1999). The self-report versions of these psychiatric assessment scales and survey instruments have been argued to provide a hotline into the mental realm and experience of the individual. However, they do not allow participants to freely speak their own minds. They constitute “forced choice” methods (Barrett, 2018), in terms of both the wording of the items, derived from syndromic psychiatric disease categories, and the format of the response. Correlating patterns discerned by machine learning in data obtained from a person’s interaction with digital devices with scores on self-report psychiatric assessment scales, and the conclusions derived thereof in terms of the presence and severity of mental disorder, is at the very least problematic.

Multiple possible ontologies of mental disorder

The question of the causal infrastructure that generates mental phenotypic effects, extended or not, boils down to the question of how we conceive the nature and ontological status of mental disorders. Multiple, alternative ontologies have been offered.

The medical disease model

As operationalizations of the DSM’s and the ICD’s disease categories, psychiatric assessment scales are reflective of the way in which psychiatrists have conceptualized mental disorders, and of the way in which society thinks about normalcy and deviance, health and disease. In this medical model, symptoms are phenotypic effects of discrete disease entities (object–attribute relationship). Diseases can be distinguished from the normal/healthy state, described, and classified, so that different individuals can be said to have or suffer from the same disease. Symptoms of mental disorders are psychometrically construed as reflective indicators of what in the psychometric methods literature is called a latent variable (Borsboom, 2005). Across populations and the globe, people exhibiting the same set of correlating symptoms share a common determinant. At least, this is, McNally et al. (2014) argue, “the primary lens through which our field views psychopathology; and it motivates the endeavour to identify the underlying disease entities that produce the symptoms of mental disorders” (p. 2).

It is fair to say that the medical model outlined here has been the dominant way of thinking in psychiatric research and clinical practice since the 1970s. In many countries a diagnosis according to the “manual” is required to qualify for reimbursement of professional help and treatment. These requirements scaffold the medical model. Though some success has been claimed for this approach in psychiatric research, commentators deem the search for discrete mental disease entities to have failed. Half a century of massive research efforts since the early 1970s and DSM III have failed to “carve nature at its joints” where mental disorders are concerned (Danziger, 1997; Hacking, 1995a, 1995b, 1998, 2007; Rose, 2019). There are no “zones of rarity” that allow for the separation of discrete disease entities (Cooper, 2013). The causes of mental disorders appear massively multifactorial, thereby undermining the plausibility of a common cause explanation for the associations (correlations) between symptoms.

Are there alternative ways of conceptualizing mental health and disorder? The answer is a resounding yes. Space restrictions prevent their full elaboration. A brief review of two of them will have to suffice to point interested readers in the right direction. One emerges from the field of psychometrics itself, the other from neurobiological stress research.

A nonlinear network theory of mental disorder

Acknowledging the fundamental and unresolved problem (read: impossibility) of inferring hidden and unobservable (latent) causal mechanisms from phenotypic effects, Borsboom (2008, 2017), Cramer et al. (2016), McNally et al. (2014), and others (Fried et al., 2017; Wichers et al., 2017) propose a radically different conceptualization of mental disorders. Instead of carrying with them the baggage of fundamental measurement of quantities from classical Newtonian physics and linear algebra, they turn to the new physics and mathematics of nonlinear, complex adaptive systems (Kauffman, 2019; Strogatz, 2012, 2015). In their network model they no longer try to construct inferential causal connections that reach below the surface of the skin and of individual behaviour. Mental disorders arise from the interaction between symptoms in a network:

Recent work has put the hypothesis that we cannot find central disease mechanisms for mental disorders because no such mechanisms exist. In particular, instead of being effects of a common cause, psychiatric symptoms have been argued to cause each other. . . . Symptoms may form feedback loops that lead the person to spiral down into the state of prolonged symptom activation that we phenomenologically recognize as mental disorder. (Borsboom, 2017, pp. 5–6)

Mental disorders, their genesis, and the course they take can be thought of in terms of trajectories, tipping points, and attractors in an abstract mental state space. A whole new set of concepts comes into play. In their mathematical models these investigators have demonstrated hysteresis. In its most general formulation, hysteresis is the dependence of a system on its history. Hysteresis is common in biological systems (Noori, 2014). Cramer et al. (2016) found it in their model of major depression. Here it had to do with the threshold for tipping into another stable part of the mental state space. Bridge symptoms shared by multiple symptom networks allow for the spreading of activation from one network or cluster to another. In a network approach, bridge symptoms explain on the one hand the often-observed comorbidity of mental disorders (Fried et al., 2017, p. 2) and on the other hand the fact that research efforts have failed to find “zones of rarity” between mental disease categories (Cooper, 2013). Critical slowing down, referring to the increase in the time it takes for a complex adaptive system to return to its equilibrium state after a perturbation, is investigated as a predictive marker for approaching a tipping point (van de Leemput et al., 2014). In mental health care, the phenomenon is of interest to predict or prevent the onset of or relapse into, for example, a depressed state.

Allostasis, allostatic load, and mental distress

Another possible ontology of mental health and disorders starts from the notion of allostasis. Trying to understand the physiological basis of broad patterns of stress-related mortality, the term was coined in 1988 by neurobiologist Peter Sterling and epidemiologist Joseph Eyer. With this new term, they wished to mark the difference in perspective between homeostasis, conceived as “stability by constancy,” and their reconceptualization of the regulatory logic in organisms, namely predictive regulation, or “stability by change” (Sterling, 2012; Sterling & Eyer, 1988; see also Barrett & Simmons, 2015; Clark, 2016). Examples of allostasis abound. One will have to suffice here. In a behaviouristic framework, Pavlov’s work on classical conditioning is often used to illustrate that all behaviours can be explained by antecedent stimuli (incentives). However, Pavlov’s dogs, instead of exhibiting a dim-witted behaviourism, embody a prime example of predictive regulation (allostasis): seeing or smelling food, or hearing the tic-toc of a metronome that is predictively associated with the appearance of food, activates the body’s digestive systems before the demand arises (Sterling & Laughlin, 2015). In allostasis, predictive regulation replaces reactive time lags (delays) with proactive time leads. “Health,” as Sterling (2020) defines it, is “the capacity to respond optimally to fluctuations in demand” (p. 210).

In a predictive regulation perspective, prediction errors are the most newsworthy to the organism and brain. Prioritizing the processing of prediction errors, that is, seeking new information that can reduce uncertainty, is the brain’s most rational choice for the use of always-limited metabolic resources. Coupled with the brain’s neural reward system, which generates positive feelings of pleasure and negative feelings of anxiety, prediction errors constitute opportunities for learning: repeat behaviours that generate positive reward prediction errors; avoid behaviours that generate negative reward prediction errors.

From the early 1990s onwards, the allostatic perspective has been elaborated by other neuroscientists. This is a literature steeped in information theoretical and computational vocabulary. Biological calculation is costly. Even under optimal conditions of a relatively stable environment, predictive regulation is taxing for the brain/body’s constrained energy budget. Under rapidly changing circumstances, in hostile environments, or in societies where people live under the strain of socioeconomic, racial, or gender inequalities, negative reward prediction errors accumulate, forcing the organism in prolonged states of arousal, and generating many of the symptoms associated with mental illness:

A system becomes unhealthy when demand drives it to operate for long periods at high levels [of arousal] that were designed to serve only for brief excursions. . . . prolonged operation at high demand evokes adaptations that are slow to reverse during brief periods of lower demand. Consequently, damage accumulates. (Sterling, 2020, p. 210)

Over time, the body/brain/mind may sustain damage that may manifest itself across levels. Irreversibility of regulatory disruptions and damage amounts to “getting stuck” in a section of mental state space with prolonged symptom activation, which impairs social relationships and functioning in society.

Theoretically reworking Seyle’s notion of stress, McEwen introduced the concept of allostatic load and defined it “as the cost of chronic exposure to fluctuating or heightened neural or neuroendocrine response resulting from repeated or chronic environmental challenge that an individual reacts to as being particularly stressful” (McEwen & Stellar, 1993, p. 293; see also McEwen, 2003, 2006, 2009). This approach has engendered translational work that tries to identify allostatic biomarkers, assemble them into an “allostatic index,” and use these as tools for the measurement of allostatic load in epidemiological and public health studies (Beckie, 2012; Juster et al., 2010; Mauss et al., 2016; Schultz et al., 2012). In this literature, associations (correlations) are sought between morbidity and mortality patterns and socioeconomic and cultural inequalities. Where psychopharmacological research since the 1950s has orientated the search for causes of mental disorders to the inside of the individual brain, this allostatic load approach prefers environmental inequalities as the independent variables in the cause–effect motif. Treatment should focus on this system level, not on the lower level of synaptic neurochemistry. Sterling (2020) takes care to point out that while experiencing mental distress, “nothing is broken” (p. 210).

Conclusion

Where does all this leave the digital phenotypers of the mind that mine the digital environment for extended phenotypic footprints of mental disorders? With more questions rather than answers, no doubt. In summary of the previous sections, it is fair to say that we can take from Dawkins the idea that phenotypic effects extend beyond the boundaries of the body. We cannot take from Dawkins that there is a linear, unidirectional relationship between a concealed object (mental disorder?) inside a person’s mind or consciousness and a particular mental or behavioural phenotype, extended or not. On the one hand, there is the problem of the epigenotype, the network of recursive influences, interposing between the genome and phenotypic attributes. Correlationalists bracket that epigenotypic space and calculate probabilistically inferred causes from mathematical abstractions, and then grant them ontological status as real entities in the internal world of the human mind (reification). On the other hand, there is the problem of the phenotype being the result of the organism being wired into its environment. Part of the external influences converging on the organism may be recursive influences triggered by the organism’s own extended phenotypic effects into the environment, constituting a recursive loop where effects in the world cause effects on the source that generated them in the first place. Hence, we have three conceptual motifs of causality to choose from: the classic unidirectionally efficient kind of causality, probabilistically inferred causality, and recursive causality.

Drawing on Newtonian physics and linear mathematics, a branch of psychophysics emerged in the 19^th and early 20^th centuries, giving rise to a correlational psychology employing linear-algebraic mathematical tools for analysis. Present-day psychiatric assessment scales and psychometric instruments embedded in general population mental health surveys emerged from that psychophysical track. They inherited and retained unproven assumptions about the quantitative nature of mental phenomena (additivity and scalability) from their psychophysical ancestors. Is the psychoinformatics developed under the ambition of a digital phenotyping of the mind an extension of this tradition? Do present-day psychometric tools and adapted psychiatric assessment scales provide the best frame of reference for the correlational validation of new digital extended phenotypes?

Multiplying the number of alternative ontologies for mental disorder raises the question of what kind of ontological understanding of mental health and illness digital phenotypers of the mind choose to relate to. Theoretical reimaginations of mental entities and processes live by the grace of material investigative practices that support them. The medical model of mental distress as disease entities definitively has the advantage over alternative formulations, embedded as it is in the ICD and DSM manuals and their use in mental health care. Translational work for the alternatives mentioned here is underway, though. What does the decrease in typing speed (slowing down) captured by a key stroke logger in a mobile phone signify? An early-stage or low-grade depression? An approach to a critical tipping point in mental state space? Or does it signify a healthy strategy to reduce the allostatic burden engendered by the inequalities and challenges of one’s life situation?

No science has succeeded in obtaining privileged access to the mental realm of a human subject that we, to use Schrödinger’s (1944/2018) words, “got used to localizing . . . inside a person’s head . . . an inch or two behind the midpoint of the eyes” (pp. 122–123). It is not certain that extending the armoury of “sense organs of the scientific mind” (Schrödinger, 1944/2018, pp. 122–123) with digital phenotyping tools will bring us any closer. Digital phenotyping of the mind may be in danger of just reproducing an already problematic medical model of mental disorders when the deeper theoretical background assumptions are not properly addressed.