Interdisciplinary Seminar in Quantitative Economics (ISQM): How Small Data Can Leverage Big Data

Abstract:
We are living at a time when the “Big Data” movement is raging across the world, revolutionizing and stretching our computational imagination, when being a data scientist is perhaps more attractive than being a statistician to the new generation of quantitative scientists. This lecture will aim to illustrate how classical statistical principles can be used to incorporate external auxiliary information available from large data sources in improving inference based on a current dataset of modest size. We will consider three examples from biomedical sciences. (1) A new assaying technology is replacing the current practice: we have a large dataset measured in the old platform and a small sub-sample measured on the new one; can the old one help in boosting prediction of patient outcomes? (2) A new biomarker/predictor is being proposed to be added to an existing prediction model: while we have abundant published data on the established prediction model, the new biomarker is available on a smaller sample; can the existing information be used in a principled way to improve prediction under the new model? (3) We have a convenience sample of patients in a health system with access to their complete electronic medical records and genomewide scans: can we use knowledge from large population-based genome-wide association studies to learn and discover in this biased sample? Through these three examples, I will try to identify a connecting theme advocating for timeless statistical principles and study designs to be applied to cutting-edge problems in biomedical sciences. I will try to convince you that one can gainfully combine information from massive non-probabilistic samples with a small well-designed study to perform a bias-variance tradeoff for efficient inference.