Quote_TryonThe psychosystems project approaches psychological research, and especially the measurement problem, from a novel methodological angle. The project departs from the classical psychometric view, which takes observed differences between individuals to be a function of an underlying (‘latent’) variable. Instead, we think that observables often indicate elementary psychological properties that form a causal system. Thus, instead of viewing observables as measures of a latent variable, we take them to be part of a causal system.


Figure 1. A latent variable model for the relation between depression and its symptoms.

A good example of how this works out in practice comes from the field of clinical psychology. In that field, mental disorders are often taken to underlie sets of symptoms. For instance, Major Depressive Episode (MDE) is characterized by symptoms such as ‘depressed mood’, ‘lack of interest’, ‘fatigue’, ‘concentration problems’, and ‘sleep disturbance’ (e.g., insomnia or hypersomnia). In traditional psychometric research on depression, these symptoms are modeled as a function of a latent variable (Figure 1). That latent variable is thought to represent the attribute of interest, i.e., MDE.  In the case of depression symptoms, the model fits observed symptom endorsement data reasonably well (Aggen, Neale, & Kendler, 2005).

This model is statistically indistinguishable from what is known as a common cause model in the philosophical literature. That is, the model says that the symptoms of MDE are correlated because they all depend on the same (latent) property. A situation that would be analogous to this one is, for instance, one in which you had a number of noisy thermometers. Their recordings would be correlated, because all of them depend on the same variable (ambient temperature). The difference is that, in the case of psychology, we usually have very little idea of what the latent variable in our model represents, or of how the observed measures could depend on it. For this reason, it is often unclear whether we are really measuring anything, and, if so, what it may be. This problem is known as the problem of test validity.

The latent variable model has influenced psychological research tremendously. Many lines of thinking in psychology are explicitly based on the model; many more depend on it implicitly. However, it is not obvious that this is the right way of thinking about the relation between ‘symptoms’ and ‘syndrome’, or, more generally, about the relation between observables and theoretical terms in psychology. One of the primary reasons for this is that, in many different fields, we appear to be unable to give a more precise characterization of latent variables than ‘whatever underlies these and these measures’ – where the term ‘underlies’ hides the somewhat embarrassing fact that we really do not know what the relation between or theory and our measures is. In addition, there are no convincing candidates for performing common cause job, as done by the latent variable in Figure 1. Genetic and neuroscientific research, for instance, have yielded many interesting theoretical and empirical results but none that should cause us to be optimistic about finding a simple neural or genetic variable that we may substitute for a latent variable. On the available evidence, depression is a strongly heritable, but also highly polygenic phenomenon (i.e., is influenced by many genes), that is weakly related to a variety of neurological properties (e.g., the serotonin and dopamin systems), mostly for unknown reasons – in particular it is not clear whether the neurological properties are causal antecedents with respect to the psychological problems, or effects of these problems, or codevelop with these problems.


Figure 2. The best available idea on the relation between genes, neurology, and psychopathology in schizophrenia. Reprinted from Gottesman & Gould (2003).

This is significant, because tremendous amounts of money and time have been spent looking for the genetic and neurological antecedents of depression. This means that if depression were a function of a simple neurological property (e.g., ‘too little dopamin’) that was coded by one or a few ‘genes for depression’, we would have identified these by now. But that is not the case. A similar situation exists for most other psychological variables (e.g., personality, intelligence, etc.). For psychopathology, the best idea currently in existence that is still vaguely related to the notion that psychological properties have biological ‘antecedents’ (causal ones) is the endophenotype idea (Figure 2). This idea holds that groups of genes together cause a constellation of neuropsychological factors that together determine the liability to develop a condition such as depression in certain environmental circumstances. Now, if we look at Figure 1 and then compare it to Figure 2, it is clear that whatever our latent variables may be, they are not simple functions of the brain or of genes (and Figure 2 is still a considerable oversimplification of the situation).

In fact the situation is much worse than Figure 2 shows, since in actual fact it is likely that we are feedback mechanisms. For instance, if you learn something, you change your brain; as a result your responses to a situation may change; for instance you may seek out a novel situation; in that situation you may learn new things; these things change your brain; etc. This is just a basic example, but there are many cases in ordinary life that exhibit feedback as well. An example may be panic disorder, in which people may be so afraid to get a new panic attack that they actually get one, which causes them to be afraid to a new panic attack, which… Similarly, people who are addicted may engage in their addiction to mitigate effects of that very condition caused by their actions. The most obvious example is that of  gamblers who try to win back money they lost in gambling. There are also many drug addicts who use drugs to forget the problems that the drug use caused. So feedback is all around the psychological place. This greatly complicates the project of disentangling cause and effects because, as a matter of fact, cause and effect are not neatly separable. Everything affects everything else. And unfortunately, this isn’t just a metaphor or a figure of speech. In psychology, it’s actually true.


Figure 3. A graph of all the symptoms listed in the DSM-IV.

So: what should we do? The psychometric tools that we use presuppose that underlying the noisy complexes of our observations is a beautifully simple world of elegantly structured properties. But the actual research indicates that underlying the complexities of the observations are just more complexities. In contrast to, say, physics – where things get simpler and more uniform if you dig down from tables to molecules to atoms to strings – we just find ever more complexity. The psychosystems research project attacks this problem by abandoning the idea that there is a simple causal structure lurking behind the data. Instead, we just assume the psychological world is stunningly complicated. Yet, we also know there is some order: as a well-known developmental psychologist once said: we all get crazy in our own way, but apparenty we do tend to end up in roughly similar states. We attempt to uncover order in the organization of the complex relations between psychological variables by using network representations and dynamical systems models. Roughly, the idea is that psychological constructs – depression, neuroticism, intelligence – refer to networks of factors that are tightly related through functional, causal, and homeostatic mechanisms. The goal of our research is a) to uncover at at least a little bit of the way phenomena like depression are structured, and b) to provice methodological tools that allow others to do the same.


Aggen, S. H., Neale, M. C., & Kendler, K. S. (2005). DSM criteria for major depression: evaluating symptom patterns using latent-trait item response models. Psychological Medicine, 35(04), 475-487.