Aims of the Dissertation🔗

Yarkoni’s “generalizability crisis” is not only remarkable in its foundational critique of quantitative psychological science – it also offers a bleak outlook to aspiring researchers. Yarkoni contemplates, in all sincerity, if quantitative social science may be abandoned altogether; i.e., that young psychologists may simply “do something else” (Yarkoni, 2022, p. 8).

This dissertation is less defeatist, and operates under the notion that quantitative inquiry still presents a path forward in psychotherapy research. I do believe, however, that some conceptual rethinking is necessary. In particular, we need methodological innovations which treat heterogeneity not as a mere nuisance parameter, but allow to better diagnose, predict and generalize over the true variability underlying psychological treatment. To this end, I aim to explore technological innovations as a way to improve the impact and precision of psychological intervention in the prevention and treatment of mental disorders.

The preceding chapters have outlined intrinsic connections between “heterogeneity-as-deviance”, exchangeability, and the meta-analytic model. Indeed, many methods developed and implemented in the following articles are extensions of meta-analysis leveraging individual participant data (IPD). In “meta-analytic predictive models”, this approach is enhanced by algorithms drawn from machine learning, allowing to make personalized predictions while accounting for contextual variability. This results in meta-analytic “precision treatment rules”, which can be evaluated based on quantities derived from the Neyman-Rubin causal model. Using “targeted superlearning”, whole ensembles of algorithms can be combined with each other, allowing to estimate the true distribution of individualized treatment effects within and across studies.

I also explore how technology itself can be used to improve psychological treatment research. Using “digital phenotyping”, for instance, unprecedented amounts of fine-grained symptom and process data can be collected unobtrusively from sensors in patients' smartphones. In one article, we use such information to forecast symptom developments in digital therapeutics. In particular, we examine if deep learning architectures drawing on such data can be improved by including both shared and personalized layers, the latter encoding unique variability in symptom patterns across patients.

Lastly, I will also introduce the technical infrastructure of a “meta-analytic research domain” (MARD) developed as part of this dissertation. This MARD aims to provide open access to living meta-analytic databases of psychological treatment across all relevant indications, including tailored graphical user interfaces that allow for rapid, flexible and reproducible evidence generation. In one article, I use this infrastructure to conduct the largest meta-analysis of psychotherapy effects conducted to date, spanning all world regions and twelve mental health problems.