Antirealism Will Not Save the DSM From Empirical Inadequacy
Antirealism Will Not Save the DSM From Empirical Inadequacy
What does the DSM say about the unobservable structure of psychopathology?
A 2022 paper by Ken Kendler in JAMA Psychiatry offers an opportunity for me to reflect on what, if anything, the DSM says about the unobservable structure of psychopathology beneath the surface descriptions of symptom clusters. Kendler’s paper is a brief viewpoint article on potential lessons for the DSM from contemporary philosophy of science. In particular, the paper comments on the status of psychiatric disorders in the DSM within the context of the scientific realism versus instrumentalism (antirealism) philosophical debate1, and offers some arguments in support of an instrumentalist approach to the DSM. Kendler uses the status of “unobservables” in scientific theories as the starting point for his discussion.
“Many scientific theories assume constructs that are not directly observable (muons, genetic drift) but whose existence is inferred. In mental health research, psychiatric diagnoses play such a role. We assume that constructs, such as schizophrenia or alcohol use disorder, exist but we can only observe the signs, symptoms, and course of illness that we postulate result from these disorders.”
“Philosophy of science has had a long-running debate about the status of such postulated entities. Two major positions have evolved: scientific realism and instrumentalism. Advocates of the former argue that these constructs truly exist. Instrumentalists are more modest and argue that such constructs should be treated as tools and evaluated on their empirical adequacy (i.e., do they predict what we want them to?) and not their truth status.”
What has bothered me about this is not the case for an instrumentalist approach towards the DSM, but the implication that the DSM actually says something meaningful about the unobservable aspects of its diagnostic constructs. I have been of the view that the DSM says very little in this regard, but I have been second guessing myself since reading the Kendler paper. Debating the realist status of unobservables in the DSM only makes sense to me if the DSM actually posits or infers unobservables to begin with.
The DSM project, from DSM-III onwards, has traditionally been characterized as descriptive. Descriptive frameworks are not theory-free, but they focus on accurately depicting and categorizing phenomena without providing a deeper understanding of the underlying mechanisms or causes. They aim to systematically describe the patterns and relationships observed, and the identification of such patterns can become the basis for creating explanatory hypotheses, which, if borne out, could potentially further refine the descriptions in an iterative process.
Descriptions have to be optimized using guiding principles. DSM seeks to be informed by considerations of
clinical utility (practical usefulness and relevance in clinical practice)
reliability (agreement on the diagnosis between different evaluators)
validity across a range of validators (things like genetics, temperament, biomarker associations, clinical course, response to treatment)2
In addition, the starting point for any revisions to the DSM is the existing schema, inherited from DSM-III (largely based on expert clinical consensus informed by available evidence, which was quite limited in the 1970s during the development of DSM-III). So while DSM does seek to optimize differentiation among validators, it is simultaneously constrained by considerations of clinical utility and by its history.
On a minimal interpretation, DSM only assumes the existence of an unspecified set of mechanisms that, through unspecified processes, generate clinically recognizable symptom clusters, and these clusters show differences from each other on a range of validators.
These assumptions are so weak that I hesitate to even say that something substantive is being asserted about unobservable entities. In this interpretation, all we are saying is that somehow the clinical patterns as described are produced. We are not saying that our categorization of clinical patterns reflects the underlying, unobservable structure of mechanisms or etiological processes.
In contrast, on a strong interpretation, DSM’s categorization is assumed to reflect the unobservable structure of psychopathology. That is, the existence of categorically distinct, discrete entities corresponding to different DSM disorders is being posited or inferred. The strong reading is tempting for the uninitiated, and this is how many critics and the general public seem to understand it, but the DSM text itself disputes the strong interpretation:
“Despite the categorical framework, it is important to recognize that in DSM-5 there is no assumption that each category of mental disorder is a completely discrete entity with absolute boundaries dividing it from other mental disorders or from no mental disorder. There is also no assumption that all individuals described as having the same mental disorder are alike in all important ways. The clinician using DSM-5 should therefore consider that individuals sharing a diagnosis are likely to be heterogeneous even in regard to the defining features of the diagnosis and that boundary cases will be difficult to diagnose in any but a probabilistic fashion.” (DSM-5-TR, Introduction)
Let’s step back and clarify what we mean by instrumentalism.
According to the Stanford Encyclopedia of Philosophy, instrumentalism is “the view that theories are merely instruments for predicting observable phenomena or systematizing observation reports.”
“According to the best known, traditional form of instrumentalism, terms for unobservables have no meaning all by themselves; construed literally, statements involving them are not even candidates for truth or falsity… The most influential advocates of this view were the logical empiricists (or logical positivists)… In order to rationalize the ubiquitous use of terms which might otherwise be taken to refer to unobservables in scientific discourse, they adopted a non-literal semantics according to which these terms acquire meaning by being associated with terms for observables (for example, “electron” might mean “white streak in a cloud chamber”), or with demonstrable laboratory procedures (a view called “operationalism”). Insuperable difficulties with this semantics led ultimately (in large measure) to the demise of logical empiricism and the growth of realism…”
“Van Fraassen (1980) reinvented empiricism in the scientific context, evading many of the challenges faced by logical empiricism by adopting a realist semantics. His position, “constructive empiricism”, holds that the aim of science is empirical adequacy, where “a theory is empirically adequate exactly if what it says about the observable things and events in the world, is true” (1980: 12; p. 64 gives a more technical definition in terms of the embedding of observable structures in scientific models). Crucially, unlike logical empiricism, constructive empiricism interprets theories in precisely the same manner as realism. The antirealism of the position is due entirely to its epistemology—it recommends belief in our best theories only insofar as they describe observable phenomena, and is satisfied with an agnostic attitude regarding anything unobservable.”
“Empirical adequacy” refers to the extent to which a scientific theory or model accurately represents and accounts for observed phenomena. Constructs within a theoretical framework are considered empirically adequate if the theory’s predictions and explanations align well with empirical data, meaning that the theory correctly describes what we observe in experiments. Kendler defines empirical adequacy of constructs very briefly as “do they predict what we want them to?” but prediction of what we may want in a given context is only a subset of aligning well with the totality of empirical data.
DSM is a theoretical framework, an instrument for systematizing clinical observation reports of psychopathology. As noted earlier, this project of systematization is guided by clinical utility, reliability, external validity, and historical contingency. According to one view, the framework makes minimal commitments with regards to unobservables, and maintains a focus on categorizations of observables informed by validator associations. According to another view, DSM posits unobservables that correspond to the schema of the manual (a view rejected by the manual itself).
If minimal commitments about unobservables are being made, then debating a realist vs antirealist interpretation of these unobservables is largely moot. The realism vs antirealism debate with regards to a descriptive framework only makes sense if something meaningful is actually being asserted about unobservables underlying the categorization.
If minimal commitments about unobservables are being made, then debating a realist vs antirealist interpretation of these unobservables is largely moot. Realism vs antirealism debate with regards to a descriptive framework only makes sense if something meaningful is actually being asserted about unobservables underlying the categorization.
Let’s assume for a moment that the DSM does posit the existence of unobservables that correspond to its categorization.
From the perspective of logical empiricism, the unobservable categories of the DSM have no meaning and any meaning is acquired by virtue of operationalized criteria. This collapses into the view that DSM categories are nothing more than their operationalized criteria (a position that Kendler rightly rejects elsewhere).
From the perspective of Van Fraassen’s constructive empiricism, the aim is empirical adequacy, and the theoretical framework of the DSM is empirically adequate if what it says about psychopathology can be supported by what we can observe. Which leads us to the question… what does the DSM say exactly?
If the DSM does not assume the existence of discrete units of psychopathology that give rise to observable characteristics, then its empirical adequacy is largely a matter of differentiation among validators. However, if the DSM does assume the existence of discrete units of psychopathology, then we have to accept that such categorization is not empirically adequate in the sense envisioned by Van Fraassen, even though it may have quite a bit of clinical utility. Scientific investigations into the quantitative structure of psychopathology as well as genetic and brain mechanisms do not support the existence of discrete units, let alone discrete units that correspond to the DSM categorization.
DSM, in its current form, retains some degree of scientific validity only if it restricts itself to making minimal commitments about unobservables and asserts no more than the existence of clinically recognizable clusters that are associated with differences in validators.3
DSM retains some degree of scientific validity only if it restricts itself to making minimal commitments about unobservables and asserts no more than the existence of clinically recognizable clusters that are associated with differences in validators.
This is recognized even by DSM-5-TR itself:
“Structural problems rooted in the categorical design of DSM have emerged in both clinical practice and research. Relevant evidence of such problems includes high rates of comorbidity among disorders, symptom heterogeneity within disorders, and the substantial need for other specified and unspecified diagnoses to classify the substantial number of clinical presentations that do not meet criteria for any of the specific DSM disorders. Studies of both genetic and environmental risk factors, whether based on twin designs, familial transmission, or molecular analyses, have also raised questions about whether a categorical approach is the optimal way to structure of the DSM system.” (DSM-5-TR, Introduction)
“There is broad recognition that a too-rigid categorical system does not capture clinical experience or important scientific observations. The results of numerous studies of comorbidity and disease transmission in families, including twin studies and molecular genetic studies, make strong arguments for what many astute clinicians have long observed: the boundaries between many disorder “categories” are more fluid over the life course than has been recognized, and many symptoms that make up the essential features of a particular disorder may occur, at varying levels of severity, in many other disorders.” (DSM-5-TR, Introduction)
DSM-5-TR sticks with its current schema for practical reasons, not because empirical adequacy demands it:
“For reasons of both clinical utility and compatibility with the categorical ICD classification required for coding, DSM-5 continues to be a primarily categorical classification with dimensional elements that divides mental disorders into types based on criteria sets with defining features.”
We can distinguish among the following scenarios with regards to empirical adequacy:
Unobservables that have strong scientific support, are robustly corroborated by empirical evidence, and have an essential presence in our best scientific theories (e.g. quarks).
Unobservables that are known to be false, but invoking them is scientifically or practically useful in many contexts, because within a certain domain of applicability their predictions closely approximate more accurate empirical theories (e.g. absolute space and absolute time in Newtonian physics).
Unobservables that are known to be false, and although they may have offered some utility in the past, they are neither empirically adequate nor practically useful at present (e.g. phlogiston)
Unobservables about which practically nothing is known but whose existence is hypothesized for purposes of explanation (e.g. when the concept of “gene,” discrete inherited units that give rise to observable characteristics, was first proposed, nothing was known about the nature of such units.)
Unobservables whose existence is not yet conclusively established but there is a range of empirical evidence in support and their existence is the subject of an active research program (e.g. the five factor model of personality)
Unobservables that are underspecified and vague to the point of vacuity.
Any instrumental invocation of unobservables must distinguish among such scenarios. My view is that the DSM’s commitment to unobservables is along the lines of #6. Some processes somehow generate these symptoms clusters with differentiation on validators. A neo-Kraepelinian view about the existence of discrete mental disorders started off as #4 (hypothesized explanatory entities) and briefly existed as #5 (an active research program), but now exists in the #2 and #3 territory (research into psychopathology doesn’t reveal existence of discrete units of psychopathology, but acting as if they exist could be practically useful in some contexts). Other contemporary theoretical frameworks in psychopathology with unobservables, such as network theories invoking self-sustaining feedback loops, may be closer to #5.
Take the categorical classification of personality disorders as it officially exists in the DSM. We know that any hypothesis about the existence of discrete categorical units of personality psychopathology is empirically inadequate, because the existence of such entities is not supported by existing research, nonetheless the categorical schema offers enough clinical utility that clinicians can make use of it for clinical assessment and management. We can treat the official DSM categorization of personality disorders as instances of either scenario #6 (minimal commitments) or scenario #2 (false but clinically useful constructs), depending on how we want to interpret the diagnostic categories, but we cannot treat them as #1 or #5 (unobservable entities with strong or partial empirical support). Taking an antirealist stance towards histrionic personality disorder is very different from taking an antirealist stance towards quarks because their empirical adequacy is very different.
We can’t have our cake and eat it too. We cannot treat the DSM as a descriptive project with minimal commitments about unobservables and yet also treat it as a project that posits unobservable discrete units of psychopathology; the latter is not only empirically inadequate but also unnecessary for the manual’s clinical utility.
See also:
Antirealism is a broader concept than instrumentalism. Kendler focuses only on (a particular strand of) instrumentalism, but there are other antirealist position as well. As described in the Stanford Encyclopedia of Philosophy, opposition to any one or more of the following aspects of realism can be construed as antirealism: “the metaphysical commitment to the existence of a mind-independent reality; the semantic commitment to interpret theories literally or at face value; and the epistemological commitment to regard theories as furnishing knowledge of both observables and unobservables.”
Antecedent validators (similar genetic markers, family traits, temperament, and environmental exposure), concurrent validators (similar neural substrates, biomarkers, emotional and cognitive processing, and symptom similarity), and predictive validators (similar clinical course and treatment response)
I am repeatedly emphasizing the “clinically recognizable” part because if we ditch the clinical tradition and start with psychometrics, the resulting picture is quite different. See: DSM Disorders Disappear in Statistical Clustering of Psychiatric Symptoms