The history of the Diagnostic Statistical Manual for mental disorders (DSM) shows, from the 1960s onwards, a sharp turn towards the empirical facts of psychiatric illness, up to the point where psychiatrists nowadays use data-driven machine learning methods for predicting course of illness and steering clinical interventions. An often implicit assumption in this broadly empiricist view of psychiatry is that the data give a reliable and sufficiently precise representation of the reality of psychiatric illness. However, clinical practice involves a rich and nuanced variety of indicators, patient characteristics, symptoms, and the like. What eventually gets captured in psychiatric data? Do we register enough to validate our disease concepts and motivate treatment decisions in the clinic? In our paper we investigate what data conception can do justice to proper translations, both ways, between clinical applications and statistical research. In the background we employ a rich literature on the theory-ladenness of observations, as well as more recent literature on the measurement and definition of symptoms and phenomena. Our conclusion will be that we need to keep our focus on empirical fact, echoing Kraepelin's viewpoints, but that, for a causally complex science like psychiatry, we have to enrich our schemas for registering those facts, employing what Ryle called "thick descriptions".