Spring 2024

What constitutes "good enough" data for quantitative social science? Statistical learning is increasingly applied to substantive questions in social science, enabling researchers to work with complex, high-dimensional data structures ranging from open-ended survey responses to video data. However, persistent uncertainty about data requirements makes it difficult to design and pre-register these studies while making effective use of scarce resources. To address this issue, we introduce a novel method which allows applied researchers to calculate bounds on the sample size necessary to achieve a minimum level of accuracy and confidence for any discrete classification method, implemented in the companion scR R package. Although the method's scope is universal, we will implement an application with a highly unknown data-generating process -- an open-ended survey aimed at detecting the political impact of competing narratives of historical conflict in Nigeria -- as a hard test of its ability to generate a general lower bound.