Mastering Heterogeneity

Abstract

In the past 50 years, countless trials have shown that psychological treatments are effective for many mental disorders. However, their benefits remain limited, with no relevant improvements in treatment outcomes over the last decades. Dissemination of psychotherapies has also not led to a reduction in the prevalence of mental disorders. This challenges our conception of psychotherapy research as a cumulative science that improves mental health care.

In this dissertation, I argue that perceived lack of “progress” in psychotherapy research is connected to the context-sensitivity of mental disorders, their treatment, and our scientific inquiry itself. This context-sensitivity manifests itself as heterogeneity, characteristic of the presentation of mental disorders, and the way patients respond to treatment. Drawing on information theory, I present a formal definition of heterogeneity, and describe its connection to meta-analytic research.

The unpredictability of concrete treatment effects is already implied by the exchangeability assumption of random-effects models. I then revisit claims that this lack of generalizability, if taken seriously, renders much of current quantitative psychological research pointless.

I present the results of six papers, all of which are centered around technological innovations to improve the impact and precision of psychological treatment research. One paper presents the technical infrastructure of a “meta-analytic research domain”, a system of living databases that aims to compile and systematize all existing evidence on psychological treatment effects, allowing for rapid evidence generation by various stakeholders. This novel infrastructure is used to conduct a large-scale meta-analysis of psychotherapy effects across twelve mental disorders.

Other papers focus on the development of meta-analytic prediction models, prognosticating individualized treatment benefits across settings and contexts, and deriving precision treatment rules from them. Lastly, I present the results of a study leveraging personalized deep learning to forecast patients' symptom course during psychological treatment, using ecological assessment and digital phenotyping.

I discuss how these findings align with prevailing visions of “precision” in psychiatry. I contend that partial improvements are possible using the innovations presented in this dissertation. However, we should not believe that our field is truly tractable with a single quantitative approach, no matter how sophisticated.

Drawing from contemporary philosophy of science, I argue that psychological intervention should not be viewed as governed by a single elegant theory, but as a patchwork of capacities that we can learn to progressively identify and harness.