Major depression, <em>qu'est-ce que c'est?</em>

Attention conservation notice: 1100+ words on a speculative scientific paper, proposing yet another reformation of psychopathology. The post contains equations and amateur philosophy of science. Reading it will not make you feel better. — Largely written in 2011 and then forgotten in my drafts folder, dusted off now because I chanced across one of the authors making related points.

As long-time readers may recall, I am a big fan of Denny Borsboom's work on psychometrics, and measurement problems more generally, so I am very pleased to be able to plug this paper:

Denny Borsboom, Angélique O. J. Cramer, Verena D. Schmittmann, Sacha Epskamp and Lourens J. Waldorp, "The Small World of Psychopathology", PLOS ONE 6 (2011): e27407 [Data, code, etc., not verified by me]

Abstract: Mental disorders are highly comorbid: people having one disorder are likely to have another as well. We explain empirical comorbidity patterns based on a network model of psychiatric symptoms, derived from an analysis of symptom overlap in the Diagnostic and Statistical Manual of Mental Disorders-IV (DSM-IV).

We show that a) half of the symptoms in the DSM-IV network are connected, b) the architecture of these connections conforms to a small world structure, featuring a high degree of clustering but a short average path length, and c) distances between disorders in this structure predict empirical comorbidity rates. Network simulations of Major Depressive Episode and Generalized Anxiety Disorder show that the model faithfully reproduces empirical population statistics for these disorders.

In the network model, mental disorders are inherently complex. This explains the limited successes of genetic, neuroscientific, and etiological approaches to unravel their causes. We outline a psychosystems approach to investigate the structure and dynamics of mental disorders.

In the initial construction of the graph here, two symptoms are linked if they are mentioned in the DSM as criteria for the same disorder. That is, Borsboom et al. think of the DSM as a bipartite graph of symptoms and disorders, and project down to just symptoms. (There is some data-tidying involved in distinguishing symptoms and disorder.)

The small-world stuff leaves me cold — by this point it might be more interesting to run across a large-world network — but the model is intriguing. Each node (i.e., symptom) is a binary variable. The probability that node $i$ gets activated at time $t$, $p_{it}$, is a function of the number of activated neighbors, $A_{i(t-1)}$: \[ p_{it} = a + (1-a) \frac{e^{b_i A_{i(t-1)}-c_i}}{(1-a)+e^{b_i A_{i(t-1)}-c_i}} \] In words, the more linked symptoms are present, the more likely it is for symptom $i$ to be present to, but symptoms can just appear out of nowhere.

Statistically, this is a logistic regression: $b_i$ is how much symptom $i$ is activated by its neighbors in the graph, $c_i$ is a threshold specific to that symptom, and $a$ controls the over-all rate of spontaneous symptom appearance and disappearance. Using a very interesting data set (the National Comorbidity Survey Replication of about 9200 US adults), Borsboom et al. in fact fixed the $b_i$ and $c_i$ parameters by running logistic regressions. The $a$ parameter, which was kept the same across symptoms, was tweaked to make the rate of spontaneous occurrence not too unreasonable.

What Borsboom et al. did with this model was to run it forward for 365 steps (i.e., a year), and then look at whether, in the course of the previous year, it would have met the DSM criteria for major depression, and for generalized anxiety disorder, and then repeat across multiple people. It did a pretty good job of matching the prevalence of both disorders, and got their co-morbidity a bit too high but not crazily so.

Now, as a realistic model, this is rubbish, for a host of reasons. Lots of the edges have to be wrong; the edges should be directed rather than undirected; the edges should be weighted; the logistic form owes more to what psychologists are used to than any scientific plausibility. (Why should psychopathology be a spin glass?) The homogeneity of parameters across people could easily fail. And yet even so it comes within spitting distance of reproducing the observed frequencies of different conditions, and their co-morbidities.

Notice that despite this, there are no underlying disease variables in this network,