Print
E-MailJonathan M. Karpoff is Professor of Finance at the University of Washington. This post is based on an article forthcoming in Accounting Review, authored by Professor Karpoff; Allison Koester, Assistant Professor of Accounting at Georgetown University; D. Scott Lee, Professor of Finance at University of Nevada, Las Vegas; and Gerald S. Martin, Associate Professor of Finance at American University.
Research on the causes and consequences of financial misconduct has exploded in recent years, due partly to the availability of electronic databases that make it easy to compile samples of misconduct events. We identify more than 150 papers that examine financial misconduct based on samples drawn from one or more of four electronically-available databases: the Government Accountability Office (GAO) and Audit Analytics (AA) databases of restatement announcements, the Stanford Securities Class Action Clearinghouse (SCAC) database of securities class action lawsuits, and (4) the Securities and Exchange Commission’s (SEC’s) Accounting and Auditing Enforcement Releases, most recently as compiled by the University of California-Berkeley’s Center for Financial Reporting and Management (CFRM).
We examine researchers’ use of the GAO, AA, SCAC, and CFRM data to identify instances of financial misrepresentation and fraud. It turns out that these databases are not close substitutes and empirical inferences frequently are sensitive to the database selected. In tests that examine changes in such outcome variables as return on assets and equity issues around the revelation of misconduct, the replication rate is only 39% when the misconduct events from one randomly chosen alternative database are replaced by the misconduct events identified by randomly chosen alternative database.
These discrepancies occur because of four features of each database that are essential for researchers to understand and manage when conducting research on financial misconduct:
- Scope of coverage. The first reason different databases yield different results is that each captures a different subset of information about an instance of misconduct. A firm’s financial misconduct typically is revealed to the public via a complex sequence of announcements that, on average, spread out over multiple years. Some instances of misconduct involve one or more earnings restatements, others involve class action lawsuits, others involve various regulatory actions, and still others have various combinations of these types of events. By design, each database captures only one of these many types of announcements. So samples constructed from restatement announcements do not have large overlap with samples constructed from, say, class action lawsuits. More importantly, a sample that consists of only one type of event, e.g., lawsuits, omits large amounts of information about the misconduct that is associated with other types of events, e.g., restatements. We show that such differences in scope have material effects on how a researcher classifies and uses the events in each database.
- Dates of initial revelation. The initial public revelation of financial misconduct occurs, on average, months before the initial coverage in these databases, leading to discrepancies in event study measures and pre/post comparison tests.
- Identification of fraud. Most of the events captured by these databases are unrelated to financial fraud, and efforts to cull non-fraud events yield heterogeneous results because they rely on the researcher’s subjective judgment.
- Omitted or effectively omitted events. Three of the four databases omit large numbers of events they are designed to capture. Large omission rates can have material impacts on the external validity of one’s inferences and on tests that rely on control samples of non-misconduct firms.
To calibrate the magnitude and economic importance of each feature for each database, we compare the events in each database to relatively comprehensive case histories for 1,243 instances in which firms were penalized by the SEC and DOJ for financial misrepresentation that includes charges under Section 13(b) of the 1934 Securities Exchange Act. These case histories include 12,905 events—10.4 events per case—that include restatement announcements, lawsuit filings, and SEC and DOJ enforcement releases, as well as firm press releases, regulatory filings, and media reports. Example results include:
- Each databases captures only a small portion of the information content related to the instance of misconduct, as measured by impacts on share values. The average amount of the information captured ranges from 14% to 26%, depending on the database.
- Each database misses the initial revelation of misconduct by an average that ranges from 109 to 1,008 days, depending on the database.
- Using regulators’ charges to indicate fraud, the fraction of each database’s events that are related to fraud range from 2.6% to 42.6%, depending on the database.
- Each database omits, or effectively omits, between 13% and 61% of the events it seeks to capture, depending on the database.
Given these database features, our results imply several strategies to guide researchers in choosing the database most appropriate for their applications. These strategies include: matching the data to the research question based on the specific database’s features; selecting appropriate test designs and research methods that accommodate missing data, selection bias, or late initial revelation dates; culling procedures to identify cases of fraud as opposed to less material infractions; the importance of supplemental hand-collected data; and the relevance of replication and external validity tests.
The full article is available for download here.
