Building Better Models: Prediction, Replication, and Machine Learning in the Social Sciences. Issue 1 (May 2015)

• Record Type: Journal Article
• Authors: Hindman, Matthew
• Editors: Shah, Dhavan V.; Cappella, Joseph N.; Neuman, W. Russell
• Is Part Of: Annals of the American Academy of Political and Social Science. Volume 659:Issue 1(2015:May)
• Page Start: 48
• Page End: 62
• Publication Date: 2015-05
• Subjects: big data -- machine learning -- predictive modeling -- data science -- penalized regression -- ensemble learning -- the Lasso; Social sciences -- Periodicals; Social sciences -- United States -- Periodicals; Political science -- Periodicals; United States -- Politics and government -- Periodicals; 300
• DOI: 10.1177/0002716215570279
• Languages: English
• ISSNs: 0002-7162

Abstract

Analytic techniques developed for big data have much broader applications in the social sciences, outperforming standard regression models even—or rather especially—in smaller datasets. This article offers an overview of machine learning methods well-suited to social science problems, including decision trees, dimension reduction methods, nearest neighbor algorithms, support vector models, and penalized regression. In addition to novel algorithms, machine learning places great emphasis on model checking (through holdout samples and cross-validation) and model shrinkage (adjusting predictions toward the mean to reduce overfitting). This article advocates replacing typical regression analyses with two different sorts of models used in concert. A multi-algorithm ensemble approach should be used to determine the noise floor of a given dataset, while simpler methods such as penalized regression or decision trees should be used for theory building and hypothesis testing.