Alisa Bokulich (Boston University)
Paul Edwards's (1999, 2010) notion of model-data symbiosis has two components: on the one hand models are data-laden, in that large amounts of data often go into the construction and calibration of theoretical models. On the other hand, data are also model-laden, or as Edwards puts it, 'model filtered.' While the data-ladenness of models is familiar and relatively uncontroversial, the model-ladenness of data has received far less philosophical attention and is potentially far more controversial, insofar as "model-tampered" data is often assumed to be "corrupted" data. However, Edwards's choice of the term "symbiosis" suggests instead that model-filtered data is in fact beneficial for science. This view is also defended by Stephen Norton and Frederick Suppe (2001), who write, “To be properly interpreted and deployed, data must be modeled” (p. 70). Apart from this important preliminary work, however, the issue of the model-ladenness of data has remained undertheorized in the philosophy of science. Before one can begin to assess the epistemological implications of the model-ladenness of data—and determine, for example, when it is beneficial or problematic—one must first have a clearer picture of exactly where and how models are entering in to the construction and correction of data products, and hence the various forms that this model-ladenness of data can take. The aim of this talk is thus to begin to develop such a taxonomy of the different ways in which data can be model-laden, and elucidate what role(s) the models are playing in each of these contexts. In this talk I shall identify several different ways in which data are "model-filtered" and briefly illustrate each with examples drawn from across the geosciences. These will include the following: Data conversion: models are used to convert the data from one quantity or type (e.g., changes in electric current) used as a proxy, to the data type of interest (e.g., amount of vibration in seismology). The path from proxy to quantity of interest can be either straightforward or quite complicated, depending on the nature of the data conversion. Data correction: models are used to "vicariously" (Norton and Suppe 2001) remove unwanted elements or "noise" from the data that were not physically shielded during data collection (e.g., in geophysical gravity measurements, one needs to model and subtract drift, tidal, latitude, free-air, Bouger, terrain, Eötvös, and isostic factors to obtain relevant data signal). Data interpolation: in the geosciences data are often sparse, hence models are used to "fill-in" missing data points. Data scaling: observations often are not (or cannot) made at the spatial or temporal scale required for relevant theoretical purposes; hence the data must be upscaled or downscaled before the data can be used. Additionally models are used for data integration, data assimilation, and even for producing a substitute or benchmark for data in the case of synthetic data. Such a taxonomy, even if preliminary, provides an important foundation for further research into the epistemological implications of the model-ladeness of data

Craig Fox (University of Western Ontario)
Are there areas of science that focus directly on the construction of narratives and, if so, what role do such narratives play in interpreting data and garnering evidence? This paper argues in the affirmative, and shows that in sciences that reconstruct long-past events, narratives play very similar roles as models and theories in more traditional areas of science do. In particular, I show that narratives serve to integrate data sets that are far less informative when considered independently. That is, the probative value of some observations in reconstructing the deep past is enhanced by the integration with other observations that is made possible by additional assumptions that are licensed by the way they fit together within a provisional narrative. This integration is not unlike the more familiar way in which theories based on lawlike regularities enable different measuring devices to be calibrated. What is different, however, is that, when reconstructing the past, regularities are insufficient because often the event of interest constitutes a singular case. So what enables the integration is the provisional narrative, or sequence of sub-events that led to a target system’s evolution from an initial to a final state. The purpose of the narrative, then, is a recounting of a target system’s trajectory through state space. And the way in which the narrative serves to integrate different data sets and, thereby, make them more informative, is in part by postulating a causal sequence that brings the data sets together, making them measurements of different phenomena brought together under a single causal sequence. I illustrate this with the case of the investigation into the early history of the Earth and the origin of the Moon. The consensus among scientists is that the Moon formed from the impact ejecta launched into orbit by a giant impact between a Mars-sized proto-planet and the Earth, within the first 100 million years of the origin of the solar system. And the evidence that is largely responsible for this consensus is the hafnium-tungsten isotopic ratios of the Earth, Moon, and meteorites. But, as I show, this evidence is not actually very informative unless it can be integrated with the uranium-lead isotopic ratios of these bodies. But what licenses this integration is not the ability to calibrate these independent measures in the familiar sense. We cannot calibrate them as we can, say, calibrate a constant-volume air thermometer and a thermocouple. That is, we cannot measure a state of some target system, then vary that state, and repeat the measurements to show that the two track the same quantity, such as with the calibration of thermometers. In the case of the integration of the isotopic ratios, the two data sets only happen to be measuring the same thing in virtue of the particular sequence of events, postulated to have occurred some 4.5 billion years ago. What makes the hafnium-tungsten measurements be about the giant impact and origin of the moon is their place in the narrative. Indeed, without this narrative—more specifically, the integration with uranium-lead afforded by the narrative—the hafnium-tungsten observations are compatible with the Moon having formed yesterday (Halliday, 2003, 520).

Teru Miyake (Nanyang Technological University)
Seismology exhibits a high degree of interdependence between models, theory, and data. Information about the objects of investigation that is contained in seismic waves must be extracted through the use of models. This use of models comes with an inherent circularity, which raises two questions. First, how did seismology get off the ground in the first place? Second, how have justificatory worries about the circularity been dealt with in seismology? With these questions in mind, I will examine the history of development of models of earth structure and seismic sources from around the 1930s through the 1970s.

Wendy Parker (Durham University)
Model-data symbiosis -- a mutually dependent, yet mutually beneficial relationship between models and data -- can obtain at the level of research programs or of fields as a whole. In this talk, however, I give examples of model-data symbiosis in meteorology and climate science that obtain much more locally: computer models are used in support of the development of datasets, which are subsequently used in support of modelling activities that involve the same, or closely related, computer models. I then consider why only some of these tight symbiotic relationships raise real concerns about circularity.