2. HPC kinds, projectability, and intrinsic properties

My discussion follows a tradition initiated by Goodman (Reference Goodman[1954] 1983) and expounded by Quine (Reference Quine1969, chap. 5), wherein a satisfactory account of natural kinds should explain how natural kind terms (or predicates) yield reliable and ampliative projectable inductive inferences. Essentialist accounts hold that natural kind terms (e.g., H₂0) yield ampliative projectable inferences (e.g., water will freeze at 0° C) because they individuate a set of necessary and sufficient (“essential”) intrinsic properties (e.g., the molecular structure of H₂0) shared by kind members, which participate in laws of nature (Khalidi Reference Khalidi2013). While essentialist accounts provide a straightforward explanation of the projectability of kind terms, they are not well-suited for describing the messier classes studied in the special sciences (e.g., biology, psychology) insofar as these classes are neither immutable nor participate in exceptionless laws of nature.

Boyd’s HPC theory of natural kinds is motivated to explain the projectability of classifications in the special sciences. HPC kind terms yield projectable inferences because members of a kind share a family of properties that cluster in a regular and nonaccidental (lawlike) manner. The key features of HPC kinds are the following (Boyd Reference Boyd1999a, 143–44):

1. (1) There is a family of properties (F) that are contingently clustered in nature.

2. (2) Their co-occurrence is the result of what may be described as “homeostasis”: Either the presence of some properties tends to favor the presence of others, or there are underlying mechanisms that tend to maintain the properties in F, or both.

3. (3) There is a kind term (t) that is applied to things in which the homeostatic clustering of most of the properties in F occurs.

In Boyd’s theory, the capacity of a kind term (t) to accurately represent (or accommodate) relevant causal structures—that is, the dispositional properties and mechanisms in (2) that cause properties to cluster in a regular and nonaccidental way—explains successful projectable inferences. According to this “accommodation thesis,” successful inductive inferences and explanatory generalizations that are generated by a paradigm offer (abductive) evidence that the posited kinds are accurately representing (“accommodating”) genuine causal regularities in nature. Boyd (Reference Boyd1991) writes: “Kinds useful for induction or explanation must always ‘cut the world at its joints’ in this sense” (139).

In the literature on HPC kinds, there is disagreement regarding whether any of the properties or mechanisms underwriting HPC kinds need to be intrinsic. Philosophers who discuss species as HPC kinds emphasize the importance of relational properties and mechanisms (e.g., phylogenetic relations, interbreeding with conspecifics, exposure to similar environmental pressures) that maintain the stability of the property clusters associated with species (Boyd Reference Boyd1999a; Wilson et al. Reference Wilson, Barker and Brigandt2007). Some (Griffiths Reference Griffiths1999; Millikan Reference Millikan1999) argue that species can be explained exclusively in terms of relational mechanisms (e.g., descent from a common ancestor). Others (Boyd Reference Boyd1999b; Wilson Reference Wilson1999) insist that some intrinsic (e.g., genetic, physiological) properties are necessary.

The central argument of this article is that HPC kind classifications—including classifications of human kinds—must individuate some intrinsic properties or mechanisms to produce robust and ampliative projectable inferences. As suggested by essentialists, intrinsic properties shared by kind members explain the projectability of their classifications. In this regard, it is important to distinguish two distinctive roles that mechanisms play in the theory of HPC kinds, which are sometimes conflated in the literature:

1. (1) From a metaphysical standpoint, they fix the stability of natural classes and explain the unity (“homeostasis”) of the observable properties that are used to identify kinds.

2. (2) From an epistemological standpoint, they ground the stability of projectable inferences made about such kinds.

While relational mechanisms (e.g., interbreeding with conspecifics, phylogenetic relationships) can account for (1), they are insufficient to address (2) because they are specified too generally to yield specific inferences about species members (cf. Devitt Reference Devitt2008, Reference Devitt2010). For projectable inferences about a particular species to be made, a kind term needs to individuate some intrinsic properties (e.g., genetic properties) or internal mechanisms (e.g., physiological mechanisms) common to kind members.

The argument that projectable HPC kind classifications must individuate intrinsic properties contrasts with Boyd’s approach. In response to an argument by Millikan (Reference Millikan1999) that species are “historical kinds” constituted entirely by relational mechanisms (e.g., a copying process), Boyd (Reference Boyd1999b) objects that species are HPC kinds underwritten by both intrinsic (e.g., genetic properties, phenotypic traits) and relational mechanisms. The importance of intrinsic properties is that these “(imperfectly) shared properties” explain the projectability of HPC terms (Boyd Reference Boyd1999b, 81–82). Elsewhere, Boyd (Reference Boyd1991, Reference Boyd1999a) suggests that some HPC kinds can be defined almost entirely in terms of relational and conventionally stipulated properties, and he presents “social roles” and “feudal economy” as examples. Boyd’s neutrality on the properties individuated by HPC kind terms reflects his a posteriori methodological stance that the “naturalness” of natural kinds is whatever reference to such kinds contributes to the accommodation of classificatory schemes to causal structures (Boyd Reference Boyd1999a, 158–9). But if the considerations he raises in response to Millikan are relevant for species, they should apply to HPC kinds more generally. Hence, the “naturalness” of HPC kinds should be identified—as in traditional essentialist accounts—with intrinsic natural properties.

3. Human kinds, stability, and projectability

Human kind classifications yield robust and ampliative projectable inferences when they individuate biological kinds, that is, HPC kinds constituted (in part) by a set of intrinsic biological properties and mechanisms. This perspective suggests limitations of arguments made by Ian Hacking and Ron Mallon. Contra-Hacking, human classifications that individuate biological kinds are sufficiently stable to yield robust projectable inferences despite their production of feedback effects. Contra-Mallon, social kinds constituted exclusively by relational mechanisms may yield transitory projectable inferences, but they will not yield the robust projectable inferences sought by human sciences (e.g., cognitive psychology, medicine) that aim to formulate predictive categories.

An issue raised in the Hacking–Boyd exchange at the 1989 Oberlin Colloquium (Hacking Reference Hacking1991a, Reference Hacking1991b; Boyd Reference Boyd1991) is whether human or social kinds can be natural kinds. While Boyd is optimistic that the HPC theory can apply to human kinds, Hacking (Reference Hacking1991b) is pessimistic, arguing that—in contrast with natural kind classifications—human kind classifications (e.g., “multiple personality,” “homosexuality”) are “made up” insofar as “the histories of human kinds are … different from those of natural kinds, for as [human] classifications are formed and moulded they loop back, interact with, and alter the individuals and the types of behaviour to which they apply” (152–3). This is a prototype of what Hacking (Reference Hacking, Sperber, Premack and Premack1995) subsequently formulates as the “looping effects of human kinds”: The meaning of human science classifications changes classified people such that the classifications need to be constantly revised. While the natural kinds (or “indifferent kinds”) classified in the natural sciences (e.g., electrons, H₂0) will not change based on how they are classified, the human kinds (or “interactive kinds”) classified in the human and social sciences (e.g., attention deficit hyperactivity disorder, autism) will change in response to how they are classified (Hacking Reference Hacking1999; cf. Khalidi Reference Khalidi2010). Because of looping effects, the targets of classification in the human sciences are “on the move” (Hacking Reference Hacking1999, 108). Hacking (Reference Hacking2007) writes:

Because human science classifications change the people they classify, classified people constantly change in response to how they are classified and their classifications need to be constantly revised to accommodate these changes. If human kinds and their classifications are perpetually unstable, human kinds cannot be natural kinds and their classifications cannot yield stable projectable inferences.

Hacking’s conclusion that objects of human science classifications are inherently unstable (“moving targets”) is a non sequitur that does not follow from his (correct) claim that human science classifications change classified people (Tsou Reference Tsou2021). In particular, Hacking conflates and fails to distinguish two separate theses:

1. (1) Classificatory feedback: Human science classifications change the experiences and behavior of classified people.

2. (2) Looping effects: Human science classifications must be constantly revised to accommodate changes they cause among classified people.

Classificatory feedback is a ubiquitous feature of human science classifications. If classified people are aware of and will change in response to how they are classified, (1) will be present. However, the presence of (1) does not entail (2), which is the key premise supporting Hacking’s conclusion that the targets of human science classifications are inherently unstable. If a human science classification refers to a biological kind, then (1) will not render that object of classification unstable. This applies to medical classifications (e.g., HIV, breast cancer) that cause classificatory feedback. While being diagnosed as HIV positive will inevitably change the experiences and behavior of classified individuals, these changes will not require revising the meaning of HIV nor the symptoms used to diagnose it. The HIV classification is stable because it refers to a biological kind (viz., an immunodeficiency virus). This also applies to psychiatric classifications (e.g., depression, schizophrenia) that individuate biological kinds. While being diagnosed with depression will invariably change classified individuals, these changes will not require revising the meaning of major depression nor the criteria used to define it (e.g., persistent feelings of sadness). Hence, biological kinds are stable targets of classification despite the presence of classificatory feedback. While there are instances of looping effects in the human sciences (e.g., hysteria, multiple personality), these cases are far less prevalent than Hacking suggests and involve classifications that primarily track social kinds, rather than biological ones.

The argument that projectable human classifications individuate biological kinds implies that classifications of social kinds cannot yield robust projectable inferences. Drawing on Boyd’s enthusiasm for extending HPC kinds to social kinds, Mallon (Reference Mallon2016) argues that socially constructed human classifications (e.g., race, gender) yield projectable inferences when they refer to HPC kinds constituted by social properties. Mallon defends a realist account of social construction, which explains the (social) reality of human kinds in terms of social roles. Social roles are structured by a variety of representations (e.g., attitudes, theories, narratives, models, norms, texts). A social role exists when: (1) representations pick out a category of persons and a set of beliefs and evaluations associated with the term, and (2) most of the beliefs and evaluations associated with the role are common knowledge (ibid., 6). Condition (2) is achieved through the public broadcasting of information (e.g., by scientific institutions or the media) about the social role. When a social role (e.g., woman) is entrenched by various social (relational) mechanisms (e.g., looping effects that stabilize a category), they are transformed into human kinds: social categories that support inductive inferences, predictions, and explanation. Mallon’s explanation of the projectability of socially constructed human categories appeals to the idea that entrenched social roles (or human kinds) are HPC kinds (ibid., 89). Mallon finds Boyd’s HPC theory apt for characterizing social roles given its liberalization of the properties that constitute kinds to include relational properties, in addition to intrinsic ones (cf. Boyd Reference Boyd1991, 140–3). This “opens the door to properties that result from human practices, norms, conventions, and so forth” (Mallon Reference Mallon2016, 92). If entrenched social roles are HPC kinds, then this could explain the projectability of socially constructed human kinds:

Mallon’s account of how social roles produce HPC kinds is somewhat murky, but the basic idea is that entrenched social roles acquire causal significance insofar as there are real and systematic causal consequences of being member of these constructed categories, which ground projectable inferences associated with their classifications.

While the social kinds invoked in Mallon’s social role account may yield transitory projectable inferences, human classifications that individuate biological kinds yield stable and robust projectable inferences.Footnote ² Consider Mallon’s paradigms of social kinds: race and gender. Mallon is interested in how these categories yield projectable inferences related to oppression. In this light, these categories currently yield some ampliative inferences (e.g., African American men are more likely to be arrested by the police in the United States, women will earn less money than men). These projectable inferences are transitory because the social mechanisms that ground them (e.g., racism, sexism) are subject to change. By contrast, classifications of biological kinds (e.g., PKU) yield robust projectable inferences (e.g., placing PKU infants on a phenylalanine-free diet during a critical period allows brain myelination to occur normally) over longer time-scales and across different social contexts. These projectable inferences are robust because they are causal consequences of stable intrinsic biological properties (e.g., PKU is caused by a lack of a liver enzyme, which interferes with brain myelination when phenylalanine enters the bloodstream). Hence, projectable inferences yielded by classifications of biological kinds will be as stable as the biological processes that constitute such kinds. For diseases, this assumes that there are species-specific (normal) functions that contribute to health and interferences to these functions result in systematic causal effects (Boorse Reference Boorse1977). While biological properties and mechanisms are subject to change due to evolutionary processes, they are sufficiently stable to ground robust projectable inferences. This perspective accords with Woodward’s view that explanatory generalizations in the special sciences should be invariant under various interventions and background conditions (Woodward Reference Woodward2000). While generalizations about social kinds will vary across different social contexts, generalizations about biological kinds have a much wider domain of invariance.

4. Psychiatric kinds, pluralism, and levels of classification

In the following section, I examine issues of psychiatric classification in light of the analysis of projectability defended in sections 2 and 3. While projectable human science classifications individuate biological kinds, there is no privileged or uniquely correct way of classifying biological kinds. Human science classifications, like all scientific classifications, involve elements of convention because they individuate classes that are relevant to the values and interests of classifiers. This perspective draws on Dupré’s stance on natural kinds (“promiscuous realism”) that accommodates aspects of realism and pluralism: “The realism derives from the fact that there are many sameness relations that serve to distinguish classes … in ways that are relevant to various concerns; the promiscuity derives from the fact that none of these relations is privileged” (Dupré Reference Dupré1981, 82). A pluralistic approach to kinds and classification implies that there are multiple legitimate levels of classification, so long as classifications pick out natural similarities that unify members of a class. My analysis of projectability implies that predictively useful human classifications individuate kinds constituted by intrinsic biological properties and mechanisms.

In contemporary debates about psychiatric classification, there is disagreement regarding the appropriate level to classify psychiatric constructs. The most influential system, the Diagnostic and Statistical Manual of Mental Disorder (DSM), classifies constructs at the level of mental disorders, which are assumed to be discrete disease syndromes identifiable at the individual and population level. In the DSM, mental disorders are defined operationally by a polythetic set of necessary and sufficient diagnostic criteria. The DSM has been criticized on numerous grounds, including its inability to formulate valid diagnostic categories, the high co-occurrence (comorbidity) of categories (e.g., depression and anxiety) that are assumed to be causally distinct, and its failure to adequately incorporate empirical findings (Tsou Reference Tsou2021). Herein, I focus on the methodological criticism that the DSM formulates its classifications at an inappropriate level of analysis.

Disenchantment with the failure of the DSM to provide valid diagnostic categories (i.e., classifications that accurately represent real phenomena) stimulated the formulation of alternative classifications systems that define psychiatric constructs at different levels of analysis. The Research Domain Criteria (RDoC) classifies psychiatric constructs (e.g., auditory perception, potential threat/anxiety, reward prediction error) at lower levels of analysis than the DSM and explicitly demands that there are biological correlates (e.g., genetic or neurobiological biomarkers) for these constructs (see NIMH 2018). In contrast with the DSM’s categorical approach to classification, the Hierarchical Taxonomy of Psychopathology (HiTOP) adopts a dimensional approach that organizes psychiatric constructs in a hierarchical manner with spectrum constructs (e.g., internalizing, externalizing) at the top of the hierarchy, which subsume more specific domain constructs (e.g., fear, distress), which subsume disorders (e.g., depression), which subsume behaviors and symptoms at the bottom of the hierarchy (Kotov et al. Reference Kotov, Krueger and Watson2017, 462).

Drawing on methodological criticisms of the DSM by advocates of alternative systems, Kathryn Tabb (Reference Tabb2015) argues against the DSM’s assumption of “diagnostic discrimination,” which maintains that “the operationalized criteria for diagnosing clinical types will … successfully pick out populations about which relevant biomedical facts can be discovered” (1048). Tabb’s argument emphasizes the polythetic nature of DSM categories, which generate an immense array of incongruent symptom-profiles. Tabb argues that the heterogeneous groups of patients produced by DSM definitions are unsuitable for discovering underlying mechanisms. She praises alternative systems for offering novel ways of formulating psychiatric constructs and compiling research populations that do not rely on DSM definitions. Moreover, Tabb (Reference Tabb2019) criticizes philosophers of psychiatry for implicitly accepting the DSM assumption that psychiatric constructs formulated at the level of mental disorders (diagnostic kinds) are the most important psychiatric kinds to investigate. Against this view, she defends a pluralistic view that encourages investigating a multiplicity of psychiatric kinds (e.g., risk factors, personality dimensions, phenomenological states, social conditions) other than diagnostic kinds. In Tabb’s ideal of integrative pluralism (cf. Mitchell Reference Mitchell2003), efforts to integrate a metaphysically diverse array of psychiatric kinds provides a more promising path for identifying valid and pragmatically useful psychiatric categories.

Tabb’s argument that DSM definitions have impeded the formulation of valid psychiatric categories in research contexts is compelling; however, she overstates the impossibility of diagnostic discrimination. For example, since the mid-1990s, the DSM has defined schizophrenia by the following symptoms (APA 2013, 99–100):

1. (1) Delusions,

2. (2) Hallucinations,

3. (3) Disorganized speech,

4. (4) Grossly disorganized or catatonic behavior, and

5. (5) Flat emotional expression, poverty of speech, or lack of motivation.

Despite content validity problems with this definition (e.g., the exclusion of cognitive impairments that are commonly observed in schizophrenia), it has been sufficiently valid to facilitate the discovery of novel neurobiological mechanisms (e.g., glutamate systems, impaired neuroplasticity) and genetic biomarkers (e.g., genetic anomalies on chromosomes 6 and 1) in genome-wide association studies (see Tsou Reference Tsou2021). This counters Tabb’s pessimism that diagnostic kinds cannot produce research samples suitable for discovering biomedical facts.

The analysis of this article suggests that both the validity and projectability of psychiatric classifications requires the accurate individuation of stable biological kinds. Elsewhere, I have argued that multiple lines of independent research (e.g., pharmacology, brain imaging studies, animal models) support the inference that the DSM’s schizophrenia classification tracks a stable and complex neurobiological kind (Tsou Reference Tsou2012, Reference Tsou2021). In this regard, the specific symptoms used to define schizophrenia correlate with distinctive neurobiological biomarkers (or criterion-related validators). For example, the positive (or psychotic) symptoms of schizophrenia (symptoms 1 and 2) correlate with excessive dopamine activity in the mesolimbic pathway, while negative symptoms (symptom 5) correlate with deficient dopamine activity in the mesocortical pathway and deficient glutamate activity in the prefrontal cortex. Moreover, negative and positive symptoms tend to co-occur because deficient activity of dopamine in the prefrontal cortex (the terminal region of the mesocortical pathway) correlates with excessive dopamine activity in mesolimbic pathway.^{Footnote 3} Robust and ampliative projectable inferences yielded by the schizophrenia classification (e.g., schizophrenia can be treated by dopamine antagonist drugs or glutamate agonist drugs) are explained by the fact that it reliably tracks these intrinsic biological properties.

While Tabb’s integrative pluralism is unobjectionable, her analysis should be qualified by the demand that psychiatric kinds formulated at various levels of analysis should individuate (or causally relate to) biological kinds. This prescription is consonant with the RDoC’s bottom-up strategy of formulating psychiatric constructs that correlate with biomarkers at different (e.g., molecular, cellular, neurobiological) levels. For example, the RDoC construct of “potential threat/anxiety” correlates—at the level of neurobiological circuitry—with activity in the bed nucleus of the stria terminalis (NIMH 2018). Compared to the diagnostic kind model adopted by the DSM (Tabb Reference Tabb2019), the strategy of correlating hypothesized psychiatric constructs with biomarkers provides a more direct and transparent approach for ensuring that psychiatric constructs possess (criterion-related) validity. The HiTOP strategy of classifying psychiatric constructs at different hierarchical levels, wherein higher-level constructs subsume lower-order constructs, suggests that some DSM categories (e.g., major depression, generalized anxiety, PTSD) can be grouped together at a higher level (e.g., distress disorders). Hyman (Reference Hyman2007) points out that the architects of DSM-III somewhat arbitrarily decided to split (rather than lump) disorders, which resulted in a proliferation of diagnostic categories in subsequent editions. My analysis suggests that disorders should be classified at a level that individuates a causally distinctive biological kind (cf. Craver Reference Craver2009), which implies that fine-grained DSM categories (e.g., obsessive-compulsive disorder, agoraphobia, social phobia) should be classified (“lumped”) at a higher level of analysis (e.g., fear disorders) if they are constituted by similar biological mechanisms. Elsewhere, I have argued that these fine-grained psychiatric categories (and culture-bound syndromes) are differentiated by the social mechanisms emphasized by Hacking and Mallon (e.g., broadcasting of social roles, imitation of stereotypes), which stabilize a culturally-specific expression of a biological kind (Tsou Reference Tsou2021).