Individual Submission Summary
Share...

Direct link:

Lemons in the Political Marketplace: A Big-Data Approach to Detect ‘Scam PACs’

Sat, October 2, 4:00 to 5:30pm PDT (4:00 to 5:30pm PDT), TBA

Abstract

Scam PACs are political action committees (PACs) in the United States that use the campaign donations they raise to enrich their creators instead of advancing the candidate campaigns or political causes that they purport to champion. Based on numerous media reports, scam PACs have successfully targeted supporters of the Tea Party movement (Lipton and Steinhauser 2015), Ken Cuccinelli (Janetsky 2018), Donald Trump (Severns 2019), the “Blue Lives Matter” movement (Ellis and Hicken 2020), and many more.

Because campaign donors face difficulty in distinguishing scam PACs from PACs that are backed by bona fide ideological or issue-advocacy groups, and that there exists minimal regulatory oversight under current campaign finance laws, the rise of scam PACs not only undermines PACs’ accountability to donors, but also generates a lemons problem in the broader political marketplace (Ravel and Weintraub 2016). Based on my calculation, in the 2018 election cycle alone alleged scam PACs collectively raised at least $57 million, which could have fully funded 74 average House campaigns in the same cycle.

To reduce the information asymmetry that donors face in discerning scam PACs, I first quantitatively assess how scam PACs that have been identified by media outlets, non-profit organizations, and government agencies differ from comparable non-scam PACs on fundraising and expenditure patterns (e.g., abnormally high fraction of spending on operating expenses rather than direct contributions or independent expenditures in elections), donor characteristics (e.g., conservative, elderly small donors), PAC treasurer and vendor networks (e.g., shared political consultants among scam PACs), and campaign solicitation and other communication materials distributed on social media and other online platforms (planned analysis to be carried out soon).

Building on these descriptive analyses, I construct a supervised machine learning algorithm that systematically detects scam PACs in U.S. federal elections, using a rich set of model features that include all of the above summary statistics of observable PAC attributes, itemized individual contributions to PACs, itemized PAC disbursements, and PACs’ links to individual treasurers and vendors. Preliminary results from my random forest model achieve an out-of-sample Area Under Curve (AUC) metric of 0.891, which is close to the gold standard of 0.90 for most applications of machine learning classifications (Hosmer, Lemeshow, and Sturdivant 2013, p. 177). Manual inspection of model predictions also reveal that my algorithm helps to identify scam PACs that have yet to receive widespread media coverage or government scrutiny. Moreover, through feature analysis of the estimated algorithm, I identify key model features that help to distinguish scam PACs from non-scam PACs, including not only summary statistics of PAC conduct (e.g., itemization ratio and proportion of campaign expenditures devoted to fundraising) but also transaction-level data (e.g., identities of individual campaign vendors and donors).

Author