INTRODUCTION

For more than a decade, increased scrutiny has been placed on the administration and integrity of democratic elections throughout the world (Levin and Alvarez 2012). The surge of interest in electoral integrity seems to be fueled by a number of different factors: an increase in the number of nations conducting elections, more concerns about election administration and voting technology, the increased use of social media, and a growing number of scholars throughout the world who are interested in the study of integrity and the possible manipulation of elections (Alvarez, Hall, and Hyde 2008).

Although there are many ways that the integrity of elections can be assessed – for example, by studying the opinions of voters about their confidence in the conduct of elections (Alvarez, Atkeson, and Hall 2012) or through election monitoring (Bjornlund 2004; Hyde 2007, 2011; Kelley 2013) – many methodologists, statisticians and computer scientists have contributed to the new and growing literature on “election forensics”. This body of research involves the development of a growing suite of tools – some as simple as looking at the distributions of variables, such as turnout in an election, and others that use more complex multivariate statistical models – to sift through observational data from elections to detect anomalies or outliers as potential indicators for election fraud and manipulation (Levin et al. 2009; Alvarez et al. 2014).

The literature on election forensics now has advanced a somewhat dizzying array of methods for detecting election anomalies, without providing guidance for when particular methods might best be utilized by analysts. That is, when is it best to look for anomalies in distributions of voter turnout? When should digit tests (such as Benford's Law) be applied? What about the use of regression models to detect outliers, either in single or multiple contests? How much statistical power do distributional tests have in common settings where you want to try to detect election outliers? These questions have motivated some of our recent research and have led us to consider the use of new techniques, such as machine learning, for the detection of election manipulation in nations like Venezuela (Alvarez et al. 2014).

Machine learning procedures use statistical tools to find patterns in the data that reveal new and relevant information that may prove useful for performing an action or task.

This volume has its origins in the rapid and staggering changes occurring in computational social research. As one of the editors of Political Analysis (an academic journal that publishes research articles in political methodology) and Analytical Methods for Social Research (a book series), I know I am witnessing amajor shift in social science research methodology. Researchers have vast (and complex) arrays of data to work with; we have incredible tools to sift through the data and recognize patterns in that data; there are now many sophisticated models that we can use to make sense of those patterns; and we have extremely powerful computational systems that help us accomplish these tasks quickly.

When I was in graduate school in the late 1980s and early 1990s, those of us who worked with survey and public opinion polling data were considered “big-N” researchers in the social sciences. When I teach introductory research methods in my graduate seminars at Caltech, I will often have students read the 1978 American Political Science Review paper by Steven J. Rosenstone and Raymond E. Wolfinger, “The Effect of Registration Laws on Voter Turnout.” Today this paper seems straightforward to students: Rosenstone and Wolfinger simply collected information on state-by-state voter registration and administrative practices, and merged that with the November 1972 U.S. Census Bureau's Current Population voting supplement, which the authors report as having more than 93,000 respondents.They then tested, using a relatively simple binary probit model, for the effects of various registration and election administration procedures on whether the survey respondent reported having voted in the 1972 federal general elections.

Most students of statistics, methodology, or econometrics today are familiar with the binary probit model and its near-cousin, binary logit. These are techniques that model the probability that an outcome is met (here, did a voter turn out in an election) based on the covariates or regressors on the right-hand side of the model. The parameters of the probit and logit model are typically fit via maximum-likelihood optimization. Today a student could use an off-the-shelf statistics software package and replicate the original Rosenstone-Wolfinger analysis, literally in the blink of an eye, on his or her laptop computer.