Groupthink tendencies in top management teams and financial reporting fraud

Abstract

I investigate the factors that contribute to financial reporting fraud in firms that are, ex ante, at a high risk of committing fraud. Using propensity score matching, I select a sample of firms with similar ex ante risk for committing fraud. I find that within this sample, interconnectedness among members of the top management team (TMT), specifically connections developed outside the firm, is significantly and positively associated with financial reporting fraud. The effect of TMT interconnectedness on fraud is more pronounced in firms with more powerful Chief Executive Officers (CEOs) and in firms in which non-CEO executives’ wealth is more sensitive to firm risk, as measured by their portfolio vega. In addition, I find that the fraud committed by more interconnected TMTs persists for longer periods of time and is more difficult to detect. Further investigations suggest that the intensity of the connections between team members influences the risk of financial reporting fraud. My findings suggest that TMT interconnectedness promotes ‘groupthink’, which is associated with dysfunctional decision-making processes.

Keywords:

• groupthink
• TMT interconnectedness
• social connections
• accounting fraud
• financial reporting risk

Notes

1 For simplicity, I refer to ‘interconnections developed outside the firm’ as ‘interconnections’ throughout the rest of the paper.

2 The descriptive statistics (untabulated) suggests that the firms in the PSM sample exhibit significantly higher risk for financial reporting fraud than firms in the general Compustat sample, as shown by a significantly higher mean F-score of the PSM sample firms.

3 In an untabulated descriptive analysis, I find that in 77% of the fraud cases in my sample, the SEC filed civil fraud charges against top executives of the fraud firms. Of these cases, 79% include charges against multiple executives.

4 I thank Dechow et al. (2011) for providing their data. The data source is the SEC website: http://www.sec.gov/divisions/enforce/friactions.shtml.

5 I address these issues in the robustness analyses.

6 The F-score calculation is described in detail in Appendix B.

7 The PSM matched control sample is similar to the fraud sample along the dimensions used for PSM and achieves reasonable covariate balancing, a condition necessary for the PSM approach to produce unbiased estimators. In untabulated analyses, I generate alternative control samples including 170 randomly selected non-fraud firms matched on industry and year, all non-fraud firms matched on industry and year only, and post-fraud firm-years of the fraud firms. Compared with these alternative control samples, the PSM matched sample is more similar to the fraud firm-year observations.

8 In this set of tests, I drop the variable BIGN because BIGN equals one for all firms in the sub-sample with higher-than-median vega.

9 In untabulated tests, I calculate the total number of days to detect a fraud and estimate Equation (2) at the firm level. My results are qualitatively similar.

10 In untabulated analyses, I use Cox regressions, which assess the probability of hazard and are often used to deal with tests that involve survival data. The findings that more interconnected management teams commit fraud for a longer period and their fraud is harder to detect hold.

11 Additional untabulated results suggest that the effect does not depend on the position the most connected executive holds in the firms.

12 The three conditions are (1) the selection bias is mostly due to observable variables, (2) both treatment and non-treatment selections are possible (i.e. the ‘common support’ condition) and (3) the distributions of covariates are approximately similar for the treated and control groups after matching (Tucker 2010).

13 The analysis calculates the significance level ‘p-critical’ on the odds ratio of differential treatment effects for a certain level of hidden bias (denoted as Γ). An untabulated analysis suggests that a value of Γ = 2.2 would be necessary in order for my inferences to change. This means that the hidden bias must result in control firms being at least 2.2 times more likely than fraud firms to have high TMT interconnectedness. Although this is not impossible, it provides some reassurance that a large source of bias would have to exist for my results to be completely invalidated.

14 To make sure the one dimension on which the fraud and non-fraud firms differ (i.e. change in cash sales) does not confound my results, I control for this factor in my main analysis. My results hold.

15 Although this value is still higher than in other studies that use large samples (e.g. 5.3% fraud firms in Kang and Lee (2017) and 3% in Bruynseels and Cardinaels (2014)), the sub-sample composition is much closer to those larger samples than the original 50% fraud firms. The 9% fraud level is necessary to maintain sufficient test power in my study.

16 Cram et al. (2009, p. 477) specify three potential threats to matched sample analysis: ‘Error 1, use of unconditional analysis, when analysis conditional on effects of matching variables is needed, Error 2, failure to control for effect of imperfectly matched variables, and Error 3, failure to reweight observations according to differing sampling rates’. Cram et al. suggest that using a conditional logistic model that employs stricter matching than a logistic model does can mitigate Error 1 and Error 3 concerns. Including variables used in the matching can address Error 2 concerns, as Shipman et al. (2017) also suggest.

17 As N varies over time, I am not able to specify a uniform N.