Revisiting the role operational definitions in psychological science

Operational definitions are afforded a central role in psychology, taught in nearly every introductory methods class and textbook. However, this deference may be unjustified; as we consider changes to how we carry out scientific inference, it is worth reconsidering the role of operational definitions in psychological science. Defining a construct in terms of a specific measurement outcome cuts out important theoretical content about how the construct affects behavior on a specific task, sources of error, the measurement process, and how the construct affects other tasks. In this talk, I examine how contemporary modeling approaches violate basic assumptions of operational definitions and operationalism more generally -- foregoing assumptions about objectivity, repeatability, independence, and fixed elicitation procedures. Counterintuitively, these departures imbue model-based definitions of constructs with superior measurement properties, such as improved reliability and validity, when compared to their operational counterparts. Instead of relying on operational definitions of constructs, I instead suggest that psychology can adapt relational or computational definitions, representing constructs as latent variables in a multilevel generative model of behavior, self-report, or neuroimaging data. These model-based metrics can better reflect measurement error, account for the interactions between measurement devices (tasks, scales) and measurement objects (participants, processes), provide a holistic account of latent constructs across different measurements, and improve scientific communication by clarifying core psychological concepts. Relational definitions of important constructs should naturally emerge as we apply models more regularly, and these definitions and models will improve as we discover or invent mathematical approaches that are suited to describing psychological processes.