Data Science

Inconsistent and unstable results have become a characteristic feature of contemporary empirical sciences. The results of data-intensive disciplines (such as econometrics, epidemiology, (clinical) medicine and psychology) lack stability. The presence of inconsistent results has a detrimental effect on theoretical discourse, policy guidance, and clinical practice. My purpose is twofold. First, I argue that alternative methodological commitments producing inconsistent statistical results are often plausible in the sense that one cannot appraise inconsistent studies based on their methodological quality. Second, I support a perspectival view on inconsistent statistical models and argue that inconsistent results are relevant for different clinical or policy decisions.

Researchers generating big data by pooling many data sources must ensure they combine to form a trustworthy result with a net increase in value. The best solution might appear to be constructing a maximally comprehensive dataset based on universal standards for representing the data's empirical content and fitness for use. We argue this is not true in general, and that the benefits of data pooling can be realized through infrastructure enabling lateral exchange and customization of data among multiple sources. We illustrate data integration without unification using big biodiversity data, which aims to address rapid species declines across the globe.

Two criticisms of data science have recently emerged. The first argues that seemingly objective machine learning algorithms often reinforce a racist, or otherwise unjust, status quo (Benjamin, 2019) . The second chastises practitioners for failing to develop a science of causal inference, rather than a mere collection of techniques for exploiting associations (Pearl and Mackenzie, 2019). The first throws into relief an urgent social problem; the second seems like an internal methodological dispute. We argue that these two critiques are deeply related. There can be no answer to the social justice critique until data science adopts a more causal orientation.