What are the methodology and findings of used in this article?

1204 Y. JIA, L. VAN LENT, AND Y. ZENG within a sample of men." Pound, Penton-Voak, and Surridge [2009] also report that circulating levels of testosterone are higher (lower) after a successful competitive event for men with more (less) masculine faces. Thus, the speculation is that testosterone plays a role in the development of both the facial structure and the neural structures, including the neural circuits relevant for the complex of behaviors already described. Neverthe- less, the exact causality of the relation is not yet fully understood. Indeed, alternative mechanisms have been proposed to link facial masculinity to masculine behavior and some of these mechanisms do not rely (directly) on testosterone. For example, evidence exists that individuals use facial structures to guide judgments of aggression in others (Carre, Mccormick, and Mondloch [2009a], Carre et al. [2010b]). Since this evidence has been documented in both children and adults, humans may be "hard-wired" to rapidly and accurately detect a propensity for threatening masculine behav- ior in other individuals (Short et al. [2012]). Facial structures may provide what biologists call an honest signal (i.e., a signal that reliably predicts in- formation useful to the receiver). To the extent that individuals respond to facial structures, it is also possible that the anticipated behavior becomes self-fulfilling even when the stereotype based on the face is wrong (Eckel [2007]). 2.3 FACIAL MASCULINITY AND FINANCIAL MISREPORTING Economic theory suggests that the decision to engage in misreporting (and ultimately to commit fraud) is the outcome of cost/benefit analyses (Becker [1968]). The link between masculine behaviors (measured by fa- cial masculinity) and misreporting is intuitive because these behaviors plau- sibly affect the costs and benefits associated with misreporting. Perhaps the most directly relevant research is studies that document an association between facial masculinity and the propensity to cheat (Stir- rat and Perrett [2010]). Individuals with a higher propensity to cheat are more likely to experience lower emotional costs from misreporting. Other behaviors associated with facial masculinity, such as with riskseeking (Api- cella et al. [2008], Zuckerman and Kuhlman [2000]), egocentric behavior [Wright et al. [2012]), and a willingness to exploit others for one's own personal financial gain (Wright et al. [2012]), may increase a given exec- utive's propensity to engage in misreporting by affecting the way in which executives respond to incentives (Baker [2000], Garen [1994], Indjejikian [1999]); incentive pay, in turn, has been linked to misreporting (Dechow, Ge, and Schrand [2010]). Moreover, CEOs with more masculine faces may possibly perform better on the job as an outcome of their masculine behav- iors (Wong, Ormiston, and Haselhuhn [2011]). Indeed, being aggressive and risk tolerant might propel the firm forward and push boundaries, withMASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1205 positive outcomes. If so, the desire of these CEOs to maintain the social status (Eisenegger et al. [2010]) associated with high performance could increase incentives to engage in accounting manipulation. Consistent with this conjecture, prior research has shown that "superstar" CEOs are more likely to engage in misreporting after they achieve superstar status to main- tain their performance record (Malmendier and Tate [2009]). More gener- ally, to the extent that status can be maintained by providing an artificially rosy view in financial statements, CEOs with more masculine faces are more likely to misreport. The facial width-to-height ratio has been positively associated with a clus- ter of masculine behaviors, and it is currently not feasible to tease out which specific behavior dominates in influencing the CEO's propensity to misre- port. Indeed, more likely than not, the whole cluster of masculine behaviors is involved in many concrete corporate settings. Thus, the described link- ages between masculine behaviors and misreporting are best thought of as illustrations of plausible scenarios rather than as separate mechanisms. Taking all of these arguments together, we formulate the following main hypothesis. HI: The facial width-to-height ratio of CEOs is positively associated with the risk of materially misstated financial statements of their firm. 3. The Facial Width-to-Height Ratio and the Construction of the Executive Pictorial Database In this section, we describe how we measure a CEO's propensity to mascu- line behaviors by using an unfakeable facial feature, and we discuss the va- lidity of this measurement. We then outline how we construct our pictorial database. Finally, we discuss how having a measure of masculine behavior helps to address endogeneity issues in the research design. 3.1 MEASUREMENT OF THE FACIAL WIDTH-TO-HEIGHT RATIO We measure a given CEO's facial width-to-height ratio (/WHR) as the distance between the two zygions (bizygomatic width) relative to the dis- tance between the upper lip and the highest point of the eyelid (height of upper face). Prior studies show that WHR is a valid cue to masculine behaviors in males (Carre, Putnam, and Mccormick [2009b], Short et al. [2012], Weston, Friday and Lio [2007], Wong, Ormiston, and Haselhuhn [2011]). Carre and Mccormick [2008] show that /WHR can be measured from photographs (instead of measuring the skull) and that this ratio pre- dicts aggressive behavior inside and outside the laboratory. Ultimately, we use this cue to masculine behavior to derive our predictions with regard to misreporting.1206 Y. JIA, L. VAN LENT, AND Y. ZENG /WHR is not the only facial masculinity measure used in prior literature. Our choice is driven by two considerations. First, we rely on the evidence in Lefevre et al. [2013], who document correlations between saliva-assayed testosterone and various facial masculinity measures. /WHR, in contrast to the alternative measures, has significant positive correlations with testosterone responses. Since we speculate that the masculine behaviors of interest are ultimately linked to testosterone exposure, this evidence is especially pertinent. Second, we conduct a pilot study in which fWHR, among the available facial masculinity measures, yields the highest inter- rater reliability scores, which is an especially important methodological criterion when using imperfect photographs. 3.2 COLLECTING THE PHOTOGRAPH OF EXECUTIVES We first identify the semblance of the executive from the company Web site, the Forbes Web site, or the company's annual report. We then use the executive's family name together with the company name to search for the best available picture of his face on Google Images. If Google Images re- turns more than one picture for the same executive, we identify the best photograph in terms of resolution, whether the executive is forward facing in the picture, and whether he has a neutral expression. If we find no pic- ture, we use Google video to search for a movie fragment in which the CEO is present. We then obtain the CEO's face from the movie and include this picture in our database. 3.3 OBTAINING THE FACIAL WIDTH-TO-HEIGHT RATIO Following the procedure outlined in Carre, Mccormick, and Mondloch [2009a], we convert each picture into eight-bit images with a standard height of 400 pixels before taking the measurement. We use the jpg file format, although it has been criticized for not maintaining spatial relation- ships as well as some other formats. The file format choice may introduce noise in our measurements, but it is unlikely to substantially bias our re- sults. Two research assistants independently measure every picture using the Image] software (Rasband [2008]) provided by the National Institutes of Health. If the difference between the two measurements of a CEO's /WHR is less than 5%, we use the average of the two measurements as our /WHR value. If the measurements diverge by more than 5%, a third re- search assistant measures the facial structure. If the difference between this third measurement and either of the first two measurements is less than 5%, we code the picture as "good quality" and use the average value of the two closest measurements for /WHR. If not, the picture is categorized as "lower quality," and we use the average of the two closest measurements for /WHR. We further code pictures as "lower quality" if they have low res- olution; if the picture is tilted by more than 45 (which compromises theMASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1207 accuracy of the bizygomatic width measurement); or, if the individual's fa- cial expression is not neutral." Appendix B illustrates the measurement procedure. To the extent that facial masculinity is largely the outcome of adolescence bone growth, an individual's fWHR value should not change after puberty. Thus, for the purpose of our study, it should not matter whether we use a recent or an older picture to obtain our data. We verify this statement for a small sample (n = 19) for which we have two pictures for each executive at younger and older ages, respectively. We find no significant differences (p = 0.443) in /WHR values between the pictures of the same executive at the different ages. 3.4 CORRELATED OMITTED VARIABLES, ENDOGENOUS MATCHING, AND REVERSE CAUSALITY While we do not claim to demonstrate causality, our tests mitigate potential concerns about biases stemming from correlated (unobservable) omitted variables or from endogenous matching (Ackerberg and Botticini [2002], Prendergast [2002]). It is plausible that firms hire CEOs with certain specific characteristics that make them a good "fit" for the job (Graham, Harvey, and Puri [2013]). In empirical work, accounting for this selection effect is challenging because the selection variables (i.e., the CEO behaviors that a firm uses to hire the right person) are unobservable to the researcher. In our case, we propose that the facial structure of individuals reliably signals a complex of behaviors that together describe salient CEO personal characteristics. Thus, many potentially relevant but previously "unobservable" characteristics should be captured by our measure of facial structure. The econometric issues that arise because a researcher cannot observe the variables that firms use to select their CEOs are reduced to the extent that facial masculinity is related to these selection variables. In ad- dition, by using information on both sides of the match between manager and firm in our tests, our estimates should suffer less from endogenous matching problems (Bandiera et al. [2010]). We control for a comprehen sive set of firm-level variables that range from contractual arrangements to corporate governance structures and the firm's competitive environment. We also control for a variety of CEO-level characteristics." To the extent that an individual's facial masculinity is shaped early in life, we are also less vulnerable to reverse causality threats. Finally, in some tests, we use data on both the CEO and the CFO from the same firm, which allows us to use firm-fixed effects. In particular, these tests speak to concerns about the possibility of unobserved heterogeneity driving our results.4. Sample, Misreporting Measures, and Summary Statistics 4.1 SAMPLES OF S&P1500, AAERS, AND NON-AAERS We include in our broad S&P1500 sample all firm-years between 1996 and 2010 for which we have a measurable picture of the CEO in 2009.16 Table 1, panel A, reports how each step in collecting CEO pictures affects the sample. From the original sample of 1,500 CEOs, 22 cases have missing Execucomp data. We exclude 47 female CEOs, as the facial width-to-height ratio (/WHR) is only a valid measure of masculine behaviors in men. We are unable to find (measurable) pictures of 295 CEOs, and we remove them from the sample. The final sample consists of 1,136 (763) CEOs with mea- surable (good quality) pictures. We draw data from Compustat, CRSP, Risk Metrics, Thomson Reuters (Insiders data), and Execucomp to construct various CEO- and firm-level variables for the broad S&P1500 sample analy- ses. The number of firm-years with nonmissing values for these variables is 3,909 (2,643) for the sample of measurable (good quality) pictures. Table 1, panel B, details this sample selection process. Table 2, panel A, presents the number of observations by year and by industry for the 3,909 firm-years. The number of firm-years in panel A ranges from a low of 8 (other services industry) to a high of 1,912 (manufacturing). 17 We derive our sample of firms subject to SEC enforcement action by considering AAERs No. 1937-3222 from the period 2004 to 2010 using a database maintained by the Center for Accounting Research and Educa- tion at the University of Notre Dame. We have an initial sample of 509 firms with GVKEYs, which enable us to merge the sample with Compustat. We exclude 207 firms with an enforcement action unrelated to a material misstatement or with misstatements that cannot be clearly linked to a spe- cific violation period. Another 66 firms have missing data on Compustat or CRSP, yielding a final sample of 141 firms. During the accounting vio- lation periods identified in the AAERs, these 141 firms had 194 CEOs. We are unable to find measurable pictures for 22 of these CEOs and another two CEOs are females. For each AAER firm, we select a firm from the same (two-digit SIC) industry and with a similar size (as measured by the total assets at the fiscal year-end preceding the accounting violation period) to form a matched sample. In six cases, we are unable to find a reasonable (size or industry) match that has a male CEO with a measurable picture. Thus, our final matched AAER sample consists of 164 CEOs from AAER firms and 164 CEOs from non-AAER firms. Table 1, panel C, summarizes the sample selection process. Due to the small sample size of AAER firms, we use all available (measureable) pictures in our tests to ensure that we 16 Thus, individuals can enter the database as the CEO of different firms (provided they were the CEO of one of the SEP1500 firms in 2009 when we took their name fromMASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1209 TABLE 1 Sample Selection Panel A: CEO-level sample for the S&P1500 firms in 2009 Sampling Procedure N Original sample (S&P1500) 1,500 Less: Observations with missing Execucomp data 22 Female CEOs 47 CEOs without measurable pictures 295 Male CEOs with measurable pictures 1,136 Less: Pictures with bad resolution 72 Observations with /WHR measurement error > 5% 82 Faces tilted by more than 45 136 Nonneutral faces 83 Pictures with good quality 763 Panel B: Firm-year level S&P1500 sample for F-score N Sampling Procedure Measurable Good-quality pictures pictures Total firm-year observations with the fWHR data 6,868 4,611 and in Execucomp from 1996 to 2010 Less: Missing data for variables required for the regressions 2.183 1,436 Missing data for F-score 776 532 Final S&P1500 sample 3.909 2,643 Panel C: Firm-CEO level AAER sample Sampling Procedure N Firms accused in AAER No. 1937-No. 3222 from 2004 to 2010 with GVKEY 509 Less: Enforcements that are unrelated to earnings misstatement (e-g., 207 bribes, disclosure) or firms with misstatements that cannot be linked to specific reporting periods Firms with missing data in Compustat or CRSP 66 Earnings misstatement firms 141 Number of firm-CEO observations that overlap with the accounting violation 194 period Less: AAER CEOs without measurable pictures 22 Female AAER CEOs 2 Data unavailability in the first- and second-round matching 6 procedures for non-AAER CEOs: female CEOs or CEOs without measurable pictures in the non-AAER sample Final AAER sample 164 This table reports the sample construction for the CEO-level sample (in panel A), the firm-year level S&P1500 sample for the Escore (in panel B), and the firm-CEO level AAER sample (in panel C). The CEO. level sample comprises 1,136 male CEOs with measurable pictures and 763 male CEOs with good-quality pictures. The firm-year level S&P1500 sample comprises 3,909 (2.643) observations for the analyses using all measurable pictures (good-quality pictures). The firm-CEO level AAER sample comprises 164 pairs of CEOs in the AAER and non-AAER firms.TABLE 2 Composition of Sample by Fiscal Year and Industry Panel A: Firm-year level S&P1500 sample with measurable pictures (/V = 3,909) Sample by Fiscal Year Sample by Industry Year 2 2-Digit NAICS Industry 2 1996 0.69 Agriculture, forestry, fishing, and hunting 0.23 1997 53 1.36 Mining, quarrying, and oil and gas extraction 207 5.30 1998 61 1.56 90 Utilities 249 6.37 1999 87 2.23 29 Construction 20 0.51 2000 113 2.89 31-83 Manufacturing 1,912 48.91 2001 144 3.68 Wholesale trade 173 4.43 2002 173 4.43 44-45 Retail trade 274 7.01 2003 209 5.85 Transportation 160 4.09 2004 249 6.37 Warehousing 0.61 2005 307 7.85 Information 259 6.63 2006 871 9.49 Finance and insurance 2.43 2007 468 11.97 Real estate and rental and leasing 28 0.72 2008 538 13.76 Professional, scientific, and technical services 174 4.45 2009 621 15.89 Admin. and support and waste memt. and remed. services 107 2.74 2010 488 12.48 Educational services 0.90 Health care and social assistance 1.13 Arts, entertainment, and recreation 0.46 Accommodation and food services 2.89 81. 99 Other services (except public administration) 0.20 Total 3,909 100 3,909 100.00 ( Continued)1212 Y. JIA, L. VAN LENT, AND Y. ZENG have sufficient power. Table 2, panel B, reports the distribution of firms by year and by industry for the AAER sample. While we examine AAERs from the period 2004-2010, the earliest violation period associated with these enforcement actions dates back to 1993. 4.2 MISREPORTING MEASURES USED IN THE S&P1500 AND IN THE AAER (NON-AAER) SAMPLES The F-score is a composite measure of the likelihood of accounting ma- nipulation based on the insight that financial statement information be- yond accruals is useful for identifying earnings manipulation (Dechow et al. [2011]). We use scores computed based on model 1 in Dechow et al. [2011], which relies on information about accruals quality, performance, and market-related incentives. In constructing our variables, we exploit the guidance in Dechow et al. [2011] about the critical values of the F-score in relation to identifying the risk of financial misreporting. We consider the (survival) time until the firm, within the CEO's tenure, first experiences an F-score that is greater than 1.85, indicating "substantial risk" of misstate- ment ( Time until F-score > 1.85). We also use an ordinal scale Frisk, which starts at zero for F-scores smaller than one (i.e., the critical value for "normal risk" of misreporting) and extends to the value three if the F-score is greater than 2.45 (i.e., the critical value for "high risk" of misreporting). The inter- mediate values of F-risk, namely one and two, are for F-scores that indicate "above normal risk" (greater than one but smaller than or equal to 1.85) and "substantial risk" (greater than 1.85 but smaller than or equal to 2.45), respectively. We define D(AAER) as an indicator variable that equals one if the firm is subject to an AAER and zero for matched non-AAER firms. When reading the enforcement releases, we collect the names of CEOs or CFOs who are accused by name as the perpetrator of the misstatement. We introduce an indicator variable D(EXECaccused), which equals one for those executives (either CEO or CFO) who are named in the AAER, and zero otherwise. 4.3 SUMMARY STATISTICS FOR THE S&P1500 SAMPLE Table 3, panel A, presents the summary statistics computed at the CEO level for the S&P1500 sample described in table 1. We first consider fWHR in both the measurable and good-quality picture samples. The distribution of /WHR is very similar in the two samples, with a mean value of 2.01. This finding is comparable to statistics reported in other studies (e.g., Lefevre et al. [2012]). Table 3, panel B, reports summary statistics for the financial misreport- ing variables of Time until F-score > 1.85 (computed at the CEO level) and F-risk (computed at the firm-year level) for CEOs with above-median (below- median) /WHR using the sample with all measurable pictures." We includewhere i is an index across firms, and / is an index across years. Misreporting is either a misreporting proxy based on the Dechow et al. [2011] prediction model or on the AAERs (in section 4.2). When we use the sample of all measurable pictures, we include in equation (1) three variables that cap- ture the potential measurement error associated with these pictures due to nonneutral facial expressions, facial tilt, and picture resolution (D(GriExp), D(GriPiff), and D(LowRes), respectively)." In this manner, we can increase the sample size while simultaneously controlling for measurement error in the variable of interest. Our variable of interest is D(/WHR> median), an in- dicator variable that takes the value of one if a CEO's /WHR is above the median and zero otherwise." Inferences are based on standard errors that are clustered at the CEO level throughout.?2 5.2 FINDINGS FOR /-SCORE BASED PROXIES IN THE S&P1500 SAMPLE We first report the results of a Cox proportional hazard model, which reflects that our interest centers in part on the time elapsed from the start of a CEO's tenure until a "substantial" risk of misreporting occurs (i.e., Time until F-score > 1.85). Since not all CEOs experience a "substantial" misreporting risk during their tenure, our data are right-censored. The dependent variable is the hazard rate, defined as the probability of a firm experiencing the event (i.e., F-score > 1.85) at time , given that the firm has survived without such event until then. The proportional hazard model allows a nonparametric estimation of the baseline hazard function (i.e., the hazard function in the absence of covariates). The explanatory variables shift the baseline hazard as follows: h (t) = ho(t) exp(Bix + . . . + BRxx). where x1 , . ... xx compose the set of explanatory variables described in equation (1), B1. . ... Bx are the slope parameters to be estimated, and ho(f) is the nonparametric baseline hazard function. We interpret the estimated parameters by subtracting 1 from the hazard ratios (i.e., the " D( LowRes) is equal to one if the resolution of a CEO's picture is low or if the interrater measurement of /WHR diverges by more than 5%. " We use the indicator variable transformation of /WHR for two reasons. First, Greene [2000, p. 379] shows that this method can be used to reduce measurement error in the variable of interest. Second, the transformation helps to present the economic significance of the estimated effects. We find very similar results (detailed in the online appendix) when using the untransformed /W//R variable in the tests. An online appendix to this paper can be downloaded at http://research.chicagobooth.edu/arc/journal-of-accounting- research/online-supplements. " Our annual data yield too few observations to also cluster standard errors by year (see, Angrist and Pischke [2009], Petersen [2009]). Indeed, Petersen [2009] reports that, when the number of clusters in a dimension is below 40, clustering the standard errors increases the bias, On the other hand, our variable of interest, D(/WHR > median), is fixed at the CEOMASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1217 exponentialed coefficients). This transformation yields the percentage change in the hazard rate caused by a unit change in the explanatory variable. We report the hazard ratios in the tables. We also estimate ordered logistic regressions in which the dependent variable is F-risk. Table 4, columns 1 and 2, report the Cox proportional hazard regres- sions that model the Time until F-score > 1.85. Columns 3 and 4 present the ordered logistic regressions using F-risk. Results in columns 1 and 3 (2 and 4) are based on the sample of all measurable (good-quality) pictures. We find consistent results across the two F-score-based proxies for financial misreporting. The exponentialed coefficient on D(/WHR > median) in the proportional hazard model in column 1 (column 2) is 1.833 (1.984), with p 1.85 F-Risk Measurable Variable Good-Quality Measurable Pictures (1) Good-Quality Pictures (2) Pictures (3) Pictures (4) Intercept 1 5.393' 20.217- Intercept 2 (2.908) (3.552) 8.580 23.360." (2.926) Intercept 3 (3.717) 10.067- 24.813." D(WHR > median) (2.936) 1.833.* 1.984- (3.651) 0.434." 0.368" (0.203) (0.256) (0.136) D(GriExp) (0.161) 0.452 -0.801. (0.512) (0.313) D( GriPrfl) 0.403" 0.084 (0.382) D( LowRes) (0.197) 0.813 0.033 (0.339) (0.220) Adjusted return 1.616' 1.694 0.124 0.217 (0.270) (0.346) (0.120) Volatility (0.147) 0.764 0.670 -1.257- -1.231 (0.750) (0.984) (0.498) ROA (0.574) 3.401 4.759 -1.778 -1.356 (2.547) (3.742) (1.171) (1.468) Book to market 2.776 5.254- 0.444 0.819 (0.677) (0.892) (0.421) (0.524) Price to earnings 0.998 1.001 -0.001 0.001 (0.003) (0.003) (0.001) (0.002) Loss 1.618 1.980 -0.899." -0.666" (0.489) (0.591) (0.264) Leverage 10.837" (0.324) 16.544." 0.295 0.786 (0.844) (1.083) (0.518) Ln( Market value) (0.670) 0.858 0.724 -0.182+ -0.201" (0.143) (0.183) (0.080) 11.012. (0.085) Sales growth 14.850. 1.947." 2.020" (0.496) (0.800) Free cash flow (0.271) (0.374) 0.532 0.397 -0.240 0.030 (0.802) (1.021) (0.439) RD 46.155" (0.561) 580.564." -2.553. -1.909 (1.822) (2.236) (1.132) (1.398) In(Firm age) 0.815 0.962 0.028 0.044 (0.212) (0.296) (0.122) In(CEO age) (0.145) 0.817 0.391 0.221 -0.332 (0.989) (1.374) (0.642) (0.756) In(CEO tenure) 0.994 0.974 -0.009 0.056 (0.080) (0.102) (0.038) (0.042) CEO power 1.302" 1.423 -0.126 -0.166 (0.137) (0.181) (0.079) (0.092)MASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1219 TABLE 4-Continued Hazard Rate (F-score > 1.85) F-Risk Measurable Good-Quality Measurable Variable Good-Quality Pictures (1) Pictures (2) Pictures (3) Pictures (4) Inside CEO 1.292 1.473 0.068 -0.015 (0.226) (0.279) (0.158) TopExDret (0.188) 0.959 0.744 0.091 0.072 (0.145) (0.196) (0.088) CEO appoint (0.106) 0.741 0.513 -0.495 -0.840* (0.497) (0.628) (0.305) Independent director (0.388) 0.456 0.189 0.200 -0.090 (0.727) (0.922) (0.466) (0.563) Board size 1.081 1.073 0.021 0.071 (0.050) (0.064) (0.034) Ln(Salary) (0.040) 0.557 0.647 0.134 0.193 (0.317) (0.267) (0.214) (0.178) In(Bonus) 1.398. 1.520" 0.293-- 0.271. (0.147) (0.198) (0.081) Ln( New option PPS) (0.091) 1.004 0.969 0.006 0.014 (0.026) (0.030) (0.013) (0.016) Ln( New stock PPS) 0.997 1.003 0.024 0.016 (0.030) (0.040) (0.016) In(Equity holdings) (0.019) 1.134 1.467.- 0.093- 0.126' (0.116) (0.115) % Misreporting industry members (0.055) (0.068) 950.51 1." 1,684.123. 5.494- 5.554+ (0.987) (1.170) (0.767) (0.905) Year fixed effects Yes Yes Yes Yes Industry fixed effects Yes Yes Yes Yes Clustered std errors CEO level CEO level CEO level Pr > x CEO level 0.001 0.001 0.001 Pseudo-R' 0.001 22.37% Log pseudo likelihood 23.99% -577.119 -378.223 -2,707.119 N -1,826.657 3,417 2.282 3,909 2,643 This table presents an analysis of the association between D(WHR > median) and misreporting proxies measured by (1) Hazard Rate (Fscore > 1.85), the probability of a firm's Ascore being greater than 1.85 (in- dicating substantial misreporting risk) in a given year, conditional upon the firm having survived to the beginning of the year, and (2) Frisk, an ordinal scale classified according to the level of misreporting risk (see appendix A), Columns I and 2 (columns 3 and 4) report hazard ratios (coefficient estimates) for Cox proportional hazard regressions (ordered logistic regressions) relating Hazard Rate (F-score > 1.85) ( Frisk) to D(/WHR > median) and a vector of control variables, Variable definitions are provided in appendix A. We winsorize the variables at 1% and 99%, with the exception of RD and Leverage for which only the top per- centile observations were winsorized. We run separate regressions for the sample of all measurable pictures (columns 1 and 3) and for the sample of good-quality pictures (columns 2 and 4). The Cox proportional hazard model estimates how soon in the CEO's tenure a substantial risk of financial misreporting occurs (i.e., Fscore > 1.85). Once this threshold of risk has been passed, we truncate all further observations for the corresponding CEO from the sample. Hence, the number of observations is reduced to 3,417 (2,282) for all measurable pictures (good-quality pictures). Two-tailed probability values are computed using standard errors clustered by CEO. *", ", and * represent statistical significance at 1%, 5%, and 10%. The standard errors are reported in parentheses.5.3 FINDINGS FOR THE AAER- AND NON-AAER-MATCHED SAMPLE So far, we have used a prediction model for misreporting. We can- not rule out the possibility, however, that some firms are falsely classi- fied as a misreporting firm (type-l error). We therefore also consider firms that are subject to SEC AAERs. These firms have unambiguously material accounting misstatements. At the same time, not all misreport- ing firms will be subject to SEC enforcement actions and by using this measure we likely underidentify the number of misreporting firms (type- II error). Thus, the F-score-based and AAER approaches complement each other in the sense that each makes a different tradeoff between type-I and type-II errors. In supplementary tests described in the online appendix (http://research.chicagobooth.edu/arc/journal-of-accounting- research/online-supplements), we use the financial restatements due to in- tentional misstatements in the S&P1500 sample as an alternative ex post measure of misreporting to address the small sample size issue of AAERs.24 In table 5, panel B, we report the results of an ordinary (a conditional) logit regression using D(AAER) as the dependent variable in column 1 (column 2). When estimating a conditional logit, we exploit the (within) variation between the AAER firm and its non-AAER match to explain the dependent variable. We find a significant positive coefficient on D(fWHR > median) both in the ordinary and conditional logit regressions (both with p x 0.010 0.001 Pseudo-R 20.74% 42.04% Log pseudo likelihood -180.197 -65.882 N 328 328 This table presents an analysis of the association between D(/WR> median) and intentional misreporting proxies based on the AAER sample. Panel A presents the paired test (two-sample test) for difference in mean of CEO /WHR between the AAER and non-AAER samples (between the CEOs who are accused and those who are not accused as a perpetrator by the SEC). Two-tailed probability values are reported. Panel B presents coefficient estimates and model summary statistics for an ordinary (conditional) logistic regression relating D(AAER) (i.e., an indicator variable coded as one for the AAER firms and zero for the non-AAER firms) to D(/WHR> median) and a vector of control variables. Variable definitions are provided in Appendix A. We winsorize the variables at 1% and 99%, with the exception of RD and Leverage for which only the top percentile observations were winsorized. Two-tailed probability values are computed using robust standard errors. "", ", and * represent statistical significance at 1%, 5%, and 10%. The standard errors are reported in parentheses.as the perpetrator (p median) 1.615- 2.464" (0.810) (0.879) D( GriExp) -0.144 -34.062- -28.874"" (0.909) (1.759) (5.495) -1.505 -16.284" -27.771- (1.076) (1.542) (6.862) D( Ladies) -1.258 14.926- 6.472 (0.833) (1.859) (5.491) DCFO 1.946 1.063 (1.329) (1.482) $5 CEO accused in the industry 11.143. (2.773) Firm-level controls Yes No No Managerial level controls Yes Yes Yes Year fixed effects Yes No No Firm fixed effects No Yes Yes Standard errors Robust Robust Robust 0.003 0.001 0.001 Pseudo-It 48.57% 64.21% 86.35% Log pseudo likelihood -49.558 -7.667 -2.924 N 164 58 58 This table uses the sample of AAER firms only and presents an analysis of the association between an executive's (either CEO's or CFO's) /WHR and his being accused by name in the AAERs. Column I limits the executives to CEOs only and reports corfficient estimates and model summary statistics for an ordinary logistic regression relating (XEXXCarrused) (i.e., an indicator variable coded as one for the executive bring accused in the AAERs and zero otherwise) to D(WAR > wodion) and a vector of control variables reported in table 5, panel B. Variable definitions are provided in appendix A. Column 2 (column 3) reports cocfficientTABLE 7 CEO /WHR Versus CEO Overconfidence Panel A: Pearson and Spearman correlation matrix of CEO fWH/R and CEO overconfi- dence proxies for the CEO-level sample of all measurable pictures (/ = 1,136) D(WHR > median) OC OPTIONS OC FIRM4 OC FIRMS D(/WHR > median) 0.011 0.002 0.018 (0.721) (0.944) OG_OPTIONS (0.536) 0.009 1 0.053 0.025 (0.752) (0.073) (0.402) OC_FIRM4 -0.008 0.042 0.723 (0.790) (0.153) (0.001) OC FIRMS 0.023 0.024 0.729 (0.434) (0.412) (0.001) Panel B: Hazard ratio, coefficient estimates, and summary statistics for Cox proportional hazard regressions and ordered logistic regressions relating misreporting proxies based on F-score to CEO /WHR and proxies of CEO overconfidence Hazard Rate (F-score > 1.85) Frisk Good Good- Measurable Quality Measurable Variable Quality Pictures (1) Pictures (2) Pictures (3) Pictures (4) Intercept I 5.530" 20.278. (2.901) Intercept 2 (2.430) 8.725-" 23.429" (2.919) Intercept 3 (2.496) 10.218" 24.886." (2.929) D(/WHR > median) 1.819. (2.543) 1.958. 0.433." 0.363- (0.203) (0.252) OG_OPTIONS (0.136) (0.161) 1.991" 2.024' 0.286- 0.278" (0.285) (0.362) (0.114) D( GriExp) (0.140) 0.465 -0.785" (0.500) (0.315) 0.384" 0.074 (0.391) D(LowRes) (0.197) 0.817 0.030 (0.345) (0.220) Controls Yes Yes Yes Year fixed effects Yes Yes Yes Yes Industry fixed effects Yes Yes Yes Yes Yes Clustered std errors CEO level CEO level CEO level Pr > X' CEO level 0.001 0.001 0.001 Pseudo-/ 0.001 22.526 Log pseudo likelihood -573.593 24.11% -375.761 -2.702.048 -1,823.600 N 3,417 2.282 3,909 2.643 D(WHR > median) 1.851. 2.016. 0.441. 0.389. (0.202) (0.257) (0.135) OC FIRM4 4.816. (0.160) 4.657." 0.840" (0.319) (0.371) (0.115) (0.141) D(WHR > median) 1.915. 2.151." 0,412. 0.387- (0.201) (0.254) (0.134) OC FIRMS (0.158) 4.509 4.419." 0.763- 0.717 (0.270) (0.326) (0.III) (0.131) (Continued)MASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1227 F-score proxies, consistent with the findings of prior work that overconfi- dent managers tend to engage more in earnings manipulation (Schrand and Zechman [2012]).25 We conclude that the facial width-to-height ratio measures a distinct man- agerial trait, different from CEO overconfidence. Compared with the over- confidence proxies, a CEO's facial structure is measured at the CEO level and is not the endogenous outcome of corporate governance mechanisms. 6. Insider Trading, Stock Option Backdating, and the Executive's Facial Width-to-Height Ratio The preceding analyses provide important evidence but are limited to one particular managerial action, namely misreporting. A valid question is whether our findings regarding the fWHR effect are limited to accounting choices only. To address this question, we examine the relation between the likelihood that a CEO engages in opportunistic insider trading (stock option backdating) and his /WHR to validate our facial masculinity measure further. 6.1 INSIDER TRADING Our tests rely on an algorithm developed by Cohen, Malloy, and Po- morski [2012], which separates "routine" from "opportunistic" trades by insiders. Routine traders are insiders who placed a trade in the same calen- dar month over the past three consecutive years. Opportunistic traders are those who have traded in the same years as the routine traders, but without a clear pattern in the past timing of the trades." Cohen et al. document that trades by opportunistic traders (but not those by routine traders) are informative. In particular, they show that opportunistic trading significantly increases the likelihood of future SEC enforcement action (linked to illegal insider trading). Based on this prior work, we predict that the fWHR of an executive should be positively associated with the likelihood of him being classified as an opportunistic trader. The insider trading setting is appeal- ing for two reasons. First, insider trading is an intentional individual action by the executive, and thus the setting facilitates better validating our /WHR measure by directly linking the effect of the individual characteristic on be- havior. Second, and related, the CEO is not the only insider who trades; we use a sample of male CFOs from the same firm, to allow the inclusion of firm-fixed effects. The tests based on this sample address concerns that correlated omitted firm-level factors complicate inferences.MASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1201 Adolescent Facial cranial growth masculinity (WHR) IV Behaviors Aggression, Testosterone VIII Egocentrism, Risk seeking, Status maintenance VI VII Neural circuitry (amygdala) V FIG. 1.-The relation between testosterone, craniofacial features, and behavior. Underlying studies: Mechanism [I] Testosterone - adolescent cranial growth Verdonck et al. [1999], Nie [2005], Thornhill and Moller [1997], Thornhill and Gangestad [1999]; Lindberg et al. [2005] Mechanism [II] Adolescent cranial growth + facial masculinity Weston, Friday and Lio [2007], Gomez-Valdez et al. [2013] Mechanism [III] (Circulating) testosterone - facial structure Pound, Penton-Voak, and Surridge [2009], Penton-Voak and Chen [2004], Lefevre et al. [2013] Mechanism [IV] Facial masculinity - behaviors Carre and Mccormick [2008], Haselhuhn and Wong [2012], Stirrat and Perrett [2010], Api- cella et al. [2008], Wong, Ormiston, and Haselhuhn [2011], Lewis, Lefevre, and Bates [2012], Campbell et al. [2010] Mechanism [V] Testosterone ++ behaviors Mehta, Jones and Josephs [2008], Pound, Penton-Voak, and Surridge [2009], Eisenegger et al. [2010], Wright et al. [2012], Mechanism [VI] Testosterone - neural circuitry (amygdala) Johnson and Breedlove [2010], Sisk, Schulz, and Zehr [2003], Morris, Jordan, and Breedlove [2004], Bos et al. [2012] Mechanism [VII] Neural circuitry (amygdala) - behaviors Adolphs, Tranel, and Damasio [1998], Klein, Shepherd, and Platt [2009], Mehta and Beer [2009], Bos et al. [2012] Mechanism [VIII] Neural circuitry (amygdala) + Facial masculinity Carre, Murphy, and Hariri [2013] that the linkages proposed in the figure are topics of ongoing investigation. Importantly, however, no matter what the exact (neuroendocrinologist) mechanisms are, our study relies most directly on research demonstrating that the male facial structure can predict behaviors. Based on laboratory ev- idence as well as on data from naturalistic settings, studies have shown thatfacial masculinity metrics predict aggressiveness (Carre and Mccormick [2008], Christiansen and Winkler [1992]), cheating and deception (Hasel- huhn and Wong [2012]), male fitness and military rank (Loehr and O'Hara [2013]), exploitation of others and cooperation with peers (Stirrat and Per- rett [2010]), sensation seeking (Campbell et al. [2010]), and a predilection for risktaking (Apicella et al. [2008]). Wong, Ormiston, and Haselhuhn [2011] report a positive association between a firm's financial performance and the CEO's facial width-to-height ratio, while Lewis, Lefevre, and Bates [2012] show that the same measure predicts achievement drive in U.S. pres- idents. A prominent conjecture forwarded in this literature is that facial mas- culinity and behavior are related because both are linked to testosterone. The exact mechanism linking androgens, such as testosterone, with be- havior is currently debated in the neurosciences (Carre, Mccormick, and Hariri [2011], Zethraeus et al. [2009]). Studies have documented that individuals with higher levels of circulating or baseline testosterone have an enhanced motivation for competition and dominance, display reduced fear, and are more likely to engage in extremely risky behavior such as gam- bling and alcohol use (Apicella, Dreber, and Mollerstrom [2014], Mehta, Jones, and Josephs [2008], Pound, Penton-Voak, and Surridge [2009], Zuckerman and Kuhlman [2000]). These individuals are also more ego- centric and have a higher propensity to cheat as well as a stronger desire to maintain social status (Eisenegger et al. [2010], Haselhuhn and Wong [2012], Wright et al. [2012]). One possible explanation for the correlation between testosterone and behavior suggests that this hormone exerts organizational effects on the brain both prenatally (i.e., during fetal sexual differentiation) and during puberty (Morris, Jordan, and Breedlove [2004], Sisk, Schulz, and Zehr [2003]). The amygdala, a cluster of nerve cells that play a role in the processing of memory and emotional reactions, is particularly seen as the mediator between testosterone and other brain regions involved in evaluat ing complex social interactions (Adolphs, Tranel, and Damasio [1998], Bos et al. [2012], Klein, Shepherd, and Platt [2009], Mehta and Beer [2009]). The amygdala has neurons expressing androgen receptors for testosterone (Johnson and Breedlove [2010]). These vasopressinergic neurons in the amygdala may respond to testosterone by reducing communication to the orbitofrontal cortex and increasing communication with brainstem systems (Bos, Terburg, and van Honk [2010])." The amygdala activity activatesMASCULINITY, TESTOSTERONE, AND FINANCIAL MISREPORTING 1203 defense responses to interpersonal provocation and status threats, re- duces interpersonal trust, and lowers the ability to self-regulate (van Honk, Terburg, and Bos [2011]), which in turn engender the masculine behaviors described earlier. 10 Testosterone also regulates the adolescent growth spurt in the presence of growth hormone (Johnston et al. [2001]). In particular, testosterone administration causes craniofacial growth in adolescents (Lindberg et al. [2005], Nie [2005], Thornhill and Gangestad [1999], Thornhill and Moller [1997], Vanderschueren and Bouillon [1995], Verdonck et al. [1999]). Ac- cording to some studies, the growth trajectories of men and women diverge for bizygomatic width, but not for upper facial height (from upper lip to the highest point of the eyelid), at puberty. Weston, Friday, and Lio [2007] pro- pose that this sexual dimorphism in facial width-to-height ratio may reflect sexual selection pressures. This idea is captured in the immunocompetence handicap hypothesis, which posits that testosterone is responsible for the de- velopment of male secondary sex traits (such as the facial structure) but has a negative impact on the immune system (Folstad and Karter [1992]). Only healthy men can afford to display masculine traits without suffering the costs of parasite loads (due to testosterone's deleterious effect on the immune system). Thus, when choosing a mate, women may prefer men with sexually dimorphic facial characteristics. An important caveat to this literature should be noted. More recent work tends not to find evidence that /WHR is sexually dimorphic." In turn, this casts doubt over the theoretical explanation that links fWHR and masculine behavior to the role of testosterone in the development of male secondary sex traits during adolescence. That said, the lack of evidence on whether /WHR is sexually dimorphic does not necessarily invalidate a potential role of testosterone in the link between /WHR and behavior within males (Alrajih and Ward [2014]). For example, direct evidence about an association between testosterone and the facial width-to-height ratio comes from a recent study by Lefevre et al. [2013]. These authors document a positive association between the facial width-to-height ratio and saliva-assayed testosterone before and after a (speed-dating) event. Specifically, both baseline and (postevent) circulating levels of testosterone are significantly associated with /WHR. Lefevre et al. [2013] theorize that, while the facial width-to-height ratio might not be sexually dimorphic, testosterone may possibly have a more direct effect on facial bone size