Tone at the Bottom: Measuring Corporate Misconduct Risk from the Text of Employee Reviews

Dennis Campbell is Dwight P. Robinson Jr. Professor of Business Administration at Harvard Business School, and Ruidi Shang is Assistant Professor of Accountancy at Tilburg University. This post is based on their recent paper.

Numerous cases of corporate misconduct have emerged globally in recent years and caused large financial, reputational, and other damages for firms, their stakeholders, and even broader society. Managers’ and employees’ inherent tendency to commit such misconduct is deeply rooted in the operating and control environment of their organization. For example, Wells Fargo was caught in 2016 for opening two million fake accounts and selling products and services to customers under false pretenses to increase sales figures. Even though the misconduct in Wells Fargo was committed by lower-level employees, their motive to engage in such misconduct was derived from the aggressive sales targets and cross-selling strategies set by middle- and upper-level managers. Further, weaknesses in Wells Fargo’s internal control systems also provided employees with opportunities to engage in misconduct. Similar examples can be seen in other recent high profile cases such as those of Volkswagen, Theranos, and BP. In each of these cases, as with Wells Fargo, both the underlying acts of misconduct and the organizational cultures and management pressures that gave rise to them were likely to be observed by numerous employees well before they resulted in economic damage to the firm and its customers, employees, and investors.

If external stakeholders, such as investors, could obtain such information about firms’ internal operating and control environments, they might be able to assess the risk of future misconduct. However, since external stakeholders do not directly observe or participate in the daily operating and control practices within firms, they cannot easily obtain such inside information. In comparison, employees have the best access to the information on firms’ internal operations and controls simply as a by-product of their daily work.

Employee Inside Information. In our paper, we argue that the inside information on firm operating practices, control mechanisms, and broader organizational cultures that can contribute to misconduct are likely to be widespread among employees. Such information can include observations about various organizational features that give rise to misconduct risk, such as incentive and promotion systems, internal competition, or behavioral patterns of superiors and peers that create significant work pressures. In other words, employees are likely to possess information that is useful in assessing misconduct risk before specific acts of misconduct arise. Employees may also observe specific acts of misconduct before they become more widespread, externally observable, and lead to reputational or financial damage for firms.

How to get access to employees’ inside information about their firms has always been a challenge. The emergence of company review websites that collect and disclose employee evaluations, opinions, and recommendations of their firms, offers external stakeholders a potentially useful channel to observe this type of information both across firms and over time. In our study, we rely on data from to develop text-based measures of employee inside information about misconduct risk in their firms. Glassdoor is one of the most developed and popular company review websites. Employees leave anonymous comments about their firms on Glassdoor, and the website is not oriented towards misconduct or negative comments about firms. Due to these features, employees can comment their firms on Glassdoor without being subject to the high costs and uncertainties associated with other information channels such as whistleblowing. Employees also tend to share general information about the control and operating practices in their firms and describe personal experiences with superiors and peers that result from these practices. Although such general information is not directly related to misconduct, it could be reflective of firm features that contribute to future misconduct. 

Developing Measures for Misconduct Risk. The primary goal of our study is to develop useful indicators of employee inside information related to misconduct risk based on observable text-based commentary posted on about their firms. We start by learning which words tend to occur in employee reviews differentially for firm-years in which future violations occur versus otherwise similar firm-years with no such future violations. We use textual analysis approaches to map word counts from the rich vocabulary embedded in employee reviews to observable misconduct outcomes, including the presence or absence of any violation, the number of violations, the penalties imposed by the relevant regulatory or legal authorities on the perpetrating firm. We then aggregate these word counts into indices of firm-level misconduct risk.

Predicting Future Misconduct. We then use our misconduct indices along with machine learning techniques to predict future violations and violation related outcomes like assessed penalties. We address whether such text-based measures of employee inside information are useful indicators of misconduct risk by gauging the extent to which they are ultimately useful in predicting these violation outcomes in out-of-sample tests. Our results suggest that these text-based measures have potentially useful properties for measuring misconduct risk. Namely, they increase prior to realizations of misconduct risk, decrease during periods of enforcement and likely remediation, clearly discriminate between firms with higher and lower incidence of future misconduct, and they are not simply proxies for more readily available firm ratings on Glassdoor. Most importantly, they are useful in predicting misconduct beyond other readily observable characteristics such as firm size, performance, press coverage, industry risk, and prior violations. Further, our text-based measures of employee inside information appear to be most useful when predicting longer-term misconduct risk, particularly in samples of firms with little prior misconduct history. Finally, we find that our measures are leading indicators of the more intermediate risk outcome of employee whistleblower complaints.

Practical Implications. Our study highlights the possibility of obtaining broad and timely employee information from the detailed text of external reviews. Recent research shows that whistleblower programs and the media can play important roles in revealing corporate misconduct. However, these channels usually focus on specific acts of misconduct that have already occurred and may not be timely enough for firms and their stakeholders (e.g. investors, boards, auditors, and regulators) to mitigate misconduct risk before it materializes. The high social and economic costs faced by whistleblowers also hinder many employees from sharing their inside information. Our study identifies external reviews as an information channel, through which external stakeholders can obtain inside information on firms. Further, we develop empirical approaches through which such information can be effectively aggregated to develop better measures of misconduct risk.

Our findings have practical implications in light of the considerable financial, legal, and reputational risks that can arise due to corporate misconduct. Our results suggest that approaches like the one we take in this paper for measuring employees’ inside information about misconduct risk may be useful in developing leading indicators of the quality and state of a firm’s internal control environment or even broader organizational culture. Such information would otherwise be difficult for interested outsiders (e.g. investors, regulators, law enforcement) or even boards of directors to observe. Corporate managers may also use such information to detect any ongoing acts of misconduct and manage the risks of future misconduct.

The complete paper is available for download here.

Both comments and trackbacks are currently closed.