Revisioning Mental Disorder Causation to Improve Treatments

I first met Glenn Saxe at the St. Regis Hotel in New York City through an introduction from a senior mentor. From our first conversation, I recognized him as a rare genius—someone who has applied his intellect with unwavering focus to improving the lives of others. His contributions could be transformative at a time when mental illness is increasingly salient.

We bonded over a shared love of physics and mathematics, though his expertise far surpassed mine—he is a master, while I am an aficionado. With polythetic diagnostic systems—characterized by overlapping, indistinct symptom clusters—sometimes failing patient care, and the rise of machine intelligence creating new opportunities, his work offers a potential paradigm shift.

Here is part of what we discussed.

Grant H. Brenner: Why do we need to have a better understanding of mental illness?

Glenn Saxe: Recent estimates indicate that approximately 970 million people globally are living with mental disorders. Evidence shows that only limited progress has been achieved in reducing the considerable burdens they carry.

While many highlight evidence-based treatments as evidence of progress, there are crucial gaps. These treatments often can't be implemented with fidelity in usual care settings, and their efficacy may be less than initially reported. A comprehensive meta-analysis by Leichsenring and colleagues, examining over 3,700 clinical trials with 650,000+ participants, revealed that treatment effects were generally modest and often subject to biased reporting. Many were surprised by these findings, but I believe they were inevitable.

To effectively reduce the expression of a mental disorder, an intervention must target its underlying causes. Interventions that target noncausal factors cannot lead to improvement. Mental disorders are widely considered multicausal, with diverse contributing factors such as genetics, brain circuits, and social influences. However, interventions for mental disorders tend only to target one or two factors, like a psychotropic drug targeting serotonin or norepinephrine or a psychosocial intervention targeting emotional dysregulation. Consequently, many causal factors must have been left untargeted in any of the clinical trials. So, the only way any intervention studied could have yielded significant outcome changes would be if it precisely targeted the one or two causes of the disorder with disproportionally large effects.

However, I don’t see how such precise targeting would have been possible without guidance from a solid foundation of scientific knowledge about causes. Unfortunately, we don’t possess such a solid body of knowledge. The field acknowledged this in 1980 when it released DSM-III, shifting away from diagnosis by cause due to its belief that its causal knowledge was lacking. We have never returned to a causal diagnostic system because the field still believes its causal knowledge is inadequate. If we lack sufficient causal knowledge for diagnosis, how can we assert it's adequate for treatment?

GHB: How does multicausality affect treatment approaches?

GS: Treatment approaches can only be effective when guided by causal knowledge, regardless of the number of factors causing the disorder. However, just because a disorder has multiple causes doesn’t mean that each one has equally small effects. If this were the case, the situation would be hopeless. You wouldn’t want to need to deliver intervention for all of the disorder’s many causes for a patient to benefit. Fortunately, nature gives us a break here. Not all causes should be expected to have effects of equal size. In complex systems, a small number of causes typically have disproportionately large effects. The challenge lies in identifying those causes with large effects and developing treatments that target them.

Chemical models like the serotonin or norepinephrine hypotheses, from the 1960s and '70s, led to the development of various antidepressants, but results show they help some patients significantly, others modestly, and leave many still with a significant burden. This is consistent with the multicausal nature of mental disorders. Patients likely differ on those causes that drive their condition, so advancements will come when we know how to select interventions for patients based on the clinical data we collect on them.¹

Since 1980, the mental health field has been on record expressing a lack of trust in its causal knowledge. This is an astonishing fact, considering the tens of thousands of articles published in scientific journals over decades on a wide range of biopsychosocial factors associated with mental disorders. The field continues to tout these publications as evidence of its progress while also mistrusting them for diagnosis. Its progress always seems to stop short of patient benefit. Where are these findings in usual clinical practice? When do you think our patients can expect to read about the first individual with a mental disorder getting a blood test or a brain scan that was used to help alleviate their suffering?

Psychiatry Essential Reads

Involuntary Hospitalization vs. Involuntary Treatment

The Real Problem With Psychiatric Prescribing Is the Order

GHB: How do we move forward?

GS: We cannot escape the reality that our progress depends on discovering causes and incorporating these findings into clinical practice. Expecting progress to occur in other ways is to believe in miracles. Since conducting human randomized experiments on causes is not feasible for ethical or practical reasons, we must find ways to discover causes through observational data since this is just about the only data we have for this purpose. We need rigorous causal methods for this data because the risk of false discovery is much greater when disorders are multicausal. If a disorder has many causes, but only a few of them are included in a study, there is a considerable risk the studies’ findings are wrong because each of the unstudied causes is available to confound the study’s findings. Our usual statistical methods are woefully inadequate to tackle this problem. Intervening on a confounded factor cannot improve a mental disorder. This is why I have heavily invested in causal data science methods in my research.

These causal methods use conditional independence tests to exclude confounding, even from observational data with thousands of variables. They can even reveal signatures of confounding from unmeasured variables. Estimates of a factor's effect on an outcome become much more trustworthy when confounding is excluded in this way. Although these methods have been around for a while and have helped drive progress in other fields, they are still relatively new to mental health, so we must be cautious.

GHB: Can you give an example of how these methods work in practice?

GS: Our research teams at NYU and the University of Minnesota have been applying these methods to mental health data for over a decade. In studies of PTSD in injured children, we identified causal factors including acute pain, low heart rate variability, parental acute stress, and specific genetic variants. In another study of police academy recruits, we found that acoustic startle responses during training and certain genetic variants were causally related to PTSD development one year after graduation.

This is our key message: A medical field that is misaligned with causality is also misaligned with science and, ultimately, with its patients. Our true north is our patients, so failure is not an option. By maintaining our noncausal diagnostic system, we acknowledge that, after decades of effort, we still mistrust our causal evidence. Therefore, we must move beyond the current paradigm and adopt methods that can help us discover and target the true causes of mental disorders.

For more detailed information, readers can refer to our "manifesto": Mental Health Progress Requires Causal Diagnostic Nosology and Scalable Causal Discovery (Frontiers in Psychiatry, 2022)