Models of data and the representation of phenomena

Since the 1960s, the distinction between data and phenomena has fueled debates in the philosophy of science, with scholars arguing that data must be modeled in order to serve as evidence for phenomena. We claim that the modeling of data to obtain evidence for phenomena involves four levels: data, sample structure, population structure and phenomena. Our analysis suggests that the notion of pattern is essential to fully grasp the inferential capacity of data models, where representation occurs through nested surrogative reasoning –typically in the form of an isomorphism that holds at different layers. We also explain how our taxonomy of pattern-based inferential steps could shed light on various aspects of nested data modeling, such as the risk of theoretical bias. To illustrate our proposal, we examine the Eddington experiment –which tested general relativity by observing the deflection of starlight near the Sun–, and show how patterns at different levels of the data modeling provide the basis for nested surrogative reasoning in this case. Transforming the data points identified on photographic plates into a representation of light deflection requires a multi-layered search for patterns, where each pattern takes us one step further in data modeling and one step closer to the target phenomenon.