Do opaque financial reports increase future crash risk? Comparing empirical models in the U.S. and Brazilian markets

Transcription

1 Do opaque financial reports increase future crash risk? Comparing empirical models in the U.S. and Brazilian markets 2018 Abstract This research aims to discuss the effect of accounting opacity on future crash risk, evaluating different empirical models in two different markets: the U.S., in which most previous researches focus, and Brazil, as an example of an emerging market with lower levels of financial development and different accounting systems. Accounting opacity is expected to increase stocks crash risk because if managers manipulate results hiding bad news prices will react strongly and negatively when they are finally released, forming stock crashes. The empirical strategies to evaluate this relation are diverse in the literature, and the concern to properly isolate the effect of opacity varies. We discuss this concern under different estimations methods, which yields different results for the two markets. For the U.S. sample, OLS and random effects seems to subestimate the effect of opacity, but for the Brazilian sample, the effect for these models seems overestimated. When we control for fixed unobserved factors in the U.S. sample, the effect of opacity is positive and larger than the other models, while in Brazil, it loses significance. When we evaluate a sample of Brazilian firms matched with the U.S. firms we found the results approximate to the U.S. sample results, but the lack of effect in the dynamic panel data remains. This suggests firms characteristics are important to explain the differences between the two markets, but other factors are still influencing the relationship between accounting opacity and future crash risk. Keywords: Accounting opacity; Future crash risk; governance; models estimation. 1 Introduction The purpose of this research is to identify the effect of opaque financial reports on future cash risk of stock prices for a developed market, the United States, which has been 1

2 extensively previously studied, as well for an emerging market represented by Brazil. To do so, we evaluate different identification assumptions different papers have used and evaluate how the conclusions of such studies may vary accordingly. Crashes, or sudden decrease of stock prices, are more likely to happen on firms with larger agency risks (e.g., Callen & Fang, 2013) because managers tend to explore the information asymmetry in order to maximize short-term income to comply with personal interests instead of shareholders interests. Therefore, they usually hold or delay the disclosure of bad news; however, this practice is not sustainable on the long-term (Kothari, Shu, & Wysocki, 2009) and, eventually, when those news come to light, the stock prices will react strongly and negatively, making what is called crashes (Andreou, Antoniou, Horton, & Louca, 2016). Agency theory, which predicts this kind of behavior by managers, as can be seen at the classic paper of Jensen and Meckling (1976), gave room to the theoretical and practical development of Corporate Governance (e.g., Becht, Bolton, & Röell, 2003), which seeks to design mechanisms to restrict expropriating behavior from management and to improve investment environment. The literature regarding such mechanisms is wide (e.g., Eckbo, 2014; Edmans, 2014; La Porta, Lopez-de-Silanes, Shleifer, & Vishny, 2000; Shleifer & Vishny, 1997; Zingales, 1997), however, the difficulty in isolating its causal effects and in evaluating its effectiveness still is a problem to be surpassed (Roberts & Whited, 2013). One of the most important dimensions of corporate governance is related to the transparency of financial disclosure, once one expects more transparent financial statements to directly decrease information asymmetry between firm and its stakeholders (Armstrong, Guay, & Weber, 2010). The supply of transparent and high quality information is crucial for the good operation of the market because it allows managers and investors to better distinguish good and bad projects (Habib, 2008), generating more informative prices and better substantiated decisions. Nonetheless, while the financial reporting can be seen as a mechanism of governance in itself, it also interacts with other mechanisms such as compensation contracts, which often establish incentives according to both market and accounting goals and values. While the logics of this kind of compensation is to align managers (agent) interest with investors (principal), there is a literature (e.g., Peng & Röell, 2008) that argues this kind of incentive generates a dangerous side effect, the accounting income manipulation, in order to obtain advantages in stock negotiations by using insider information. This argument is consistent with a series of empirical studies, such as Bergstresser and Philippon (2006), Burns and Kedia (2006), Cheng and Warfield (2005), Efendi, Srivastava, and Swanson (2007) and Armstrong et al. (2010). Manipulated accounting and financial information are known as non-transparent or opaque. While the International Financial Reporting Standards (IFRS) are built so accounting information reflects the economic fundamentals of a firm, according to the

3 logics that faithful information leads to more sustainable investment decision and to a healthier financial environment (IFRS Foundation, 2010), its essence involves subjectivity and judgement, which gives space to manipulation. In accordance with Jin and Myers (2006), the more opaque a firm, the larger the amount of hidden bad news that could come to light in a determined time. Given that the capacity of insiders in absorbing bad news is limited, if the sufficient amount of bad news accumulates, they come to light all at once and are followed by a strong negative reaction on the stock prices. Thus, stock of more opaque firms are more likely to suffer crashes. As found by DeFond, Hung, Li, and Li (2013), countries with poor information environment face a decrease in the level of crash risk after the adoption of IFRS. Therefore, to allow comparison, we analyzed both Brazil and United States. While the first has already adopted the IFRS since 2010, it has a poor information environment, while the other still uses its local GAAP, its information environment is known to be strong. Thus, we expect differences between these to countries. Accordingly, we found that different estimation strategies yield different results for the two markets. For the U.S. sample, OLS and random effects apparently subestimate the effect of opacity, while the effect for these models seems overestimated for the Brazilian sample. When we control for fixed unobserved factors in the U.S. sample, the effect of opacity is positive and larger than the other models, while in Brazil, it loses significance. In order to understand whether these differences are due to firms different characteristics between the two markets, we also run the analysis for a sample of Brazilian firms matched with the U.S. firms, and found the results approximate to the U.S. sample results, but the lack of effect in the dynamic panel data remains. This suggests firms characteristics are important to explain the differences between the two markets, but other factors are still influencing the relationship between accounting opacity and future crash risk. The paper is structured as following. Section 2 presents the literature review on the subject, highlighting the empirical strategies each research relied on, section 3 describes the estimators and the underlying hypotheses, and section 4 presents our sample and our variables definitions. The results are discussed in section 5, and, finally, section 6 brings some final considerations regarding our paper. 2 Literature Review Some papers looked to identify the relation between the opacity of financial statements and the occurrence of crashes in stock prices. Table 1 summarizes the ones we thought to be the most important for our research, with its purpose, sample, variables and model used, and conclusions. Besides specifically studying the effect of accounting opacity, some papers focused on the relation of crash risk and other corporate governance dimensions, such as J. B. Kim

4 and Zhang (2014), who focused on compensation contracts of CEOs and CFOs. An and Zhang (2013) analyzed the role of institutional investors; and Y. Kim, Li, and Li (2014) studied the role of social responsibility. As briefly appointed on Table 1, the research designs of this literature are wide, but the concern in isolating the effect of what one intends to research is common in all papers, while the way it is addressed this difficulty varies. The endogeneity of corporate governance mechanisms is a well-known problem in this literature e.g., Roberts and Whited (2013), but difficult to solve. Nevertheless, Wintoki, Linck, and Netter (2012), focusing on board structure, argue that this classic issue involves simultaneity and unobserved heterogeneity and can be decreased in a dynamic panel model estimated by a Generalized Method of Moments (GMM), which is capable of capturing the dynamic nature of choices of corporate governance mechanisms. However, specifically on the crash risk literature, few authors presented much concern with the endogeneity issue. As seen in Table 1, some authors do not mention concerns on endogeneity, such as J. B. Kim and Zhang (2014) and DeFond et al. (2013), while others resort to simultaneous equations and instrumental variables, such as Dewally and Shao (2013) and Xu, Li, Yuan, and Chan (2014). However, neither of these two works focus on accounting opacity. Specifically dealing with this, Hutton, Marcus, and Tehranian (2009) argue that there are no obvious concerns related to the accounting opacity endogeneity. Nevertheless, the presence of unobserved factors which determines simultaneously the tendency to manipulate information and the return distributions, that end up biasing the traditional estimates, is clear and constitutes in a wide way the informational environment in which each firm is placed. The information environment includes a series of observed factors passive of control, such as firms size, profitability, capital structure, among others; and unobserved factors, like managers incentives and career perspectives and the trust level of investors. Some of those factors are constant throughout the time, like business risk and the tendency some stocks have to be more or less volatile. With those simple considerations, OLS estimates of crash risk against accounting opacity, even including observable controls will clearly be biased and it is not possible to make conclusions regarding causal effects. Beside the omitted variables, simultaneity is also a concern, once observed crashes in one year can incentive earnings management on that year and on the following years too. Andreou et al. (2016) argue that they controlled these issues by regressing crash risk in t against opacity in t 1 and including as an explanatory variable the crash risk also in t 1. While including accounting opacity in t 1 is important to avoid reverse causality once financial statements of one year are available only in the next year and, then, being able to affect only the crashes of this following year; the strategy of including a lagged dependent variable in a OLS model will only worsen the endogeneity issue, given that the omitted factors that are constant in time, for instance, will be included on

5 the error that will be correlated with the explanatory variables. However, searching for the accounting opacity effect in future crash risk makes important to control for current crash risk, because the occurrence of sudden decreases in prices serves as proxy to the information environment of each firm which includes unobservable factors, constant or not through time, which, if endogenous, can bias even more the estimates. Therefore, we fall again in the problem brought by Wintoki et al. (2012) of the dynamic nature of corporate governance mechanisms.

6 Table 1: Main References: Crash risk versus Opacity Sample Variables Model Instruments Endogeneity Concerns Conclusions Hutton et al. (2009): Relation between financial statements transparency and the distribution of stock returns. 40,882 firm-year obs., , United States. x: Opacity (earnings management s Modified Jones model) y: Idiosyncratic risk (logistic transformation of R 2 ; Crash (1 if one or more weekly returns falls 3.09 standard deviations below the mean weekly firm-specific return for that year, and it is zero, otherwise); Jump (1 if a firm faces one or more weekly return 3.09 standard deviations above the mean for that year, and zero otherwise). Controls: size, leverage, market to book, and ROE. OLS and logistic regressions according to y. DeFond et al. (2013): Test whether mandatory IFRS adoption impacts on crash risk. 27 countries that required IFRS in 2005, includes developed and emerging markets. x: Post IFRS period. y: Crash risk (skewness of firm-specific weekly returns); Controls: differences of opinion among investors, return skewness, stocks volatility, stocks past returns, size, market-to-book ratio, leverage, operating performance, earnings management, country, year, and industry. Difference-indifferences that compares IFRS adopters with three benchmark samples from 2003 to No No. The authors believe that there should be no concern with endogeneity. Used control variables. Not explicitly mentioned, but used Difference-in- Differences and control variables. Dewally and Shao (2013): Test the impact of financial derivatives on informational structure and future stock performance of banks Opaque firms are more likely to suffer stock price crashes, but positive jumps are not related to opacity. After IFRS Crash risk decreases for non-financial firms. It decreases for financial firms only who are less affected by fair value provision of IFRS, and increases for those from countries with weak banking regulations. Continued on next page

7 Sample Variables Model Instruments Endogeneity Concerns Conclusions Table 1 Continued from previous page 98 publicly traded bank holding companies, (1,568 firm-year obs.), United States. x: Opacity (notional amount of derivatives reported as contracts held for non-trading purposes divided by total assets). y: Crash risk (1 if one or more weekly returns are too negative, and zero otherwise; negative conditional return skewness; down-to-up volatility measure). Controls: size, return on equity, leverage, market-to-book ratio, gap ratio, crisis and post-crisis dummies and the lagged volatility of banks weekly returns. OLS; Simultaneous equation system. Yes: Hedging Intensity in interest rates derivatives (HI-IR) by the number of employees in each bank. J. B. Kim and Zhang (2014): investigate the relation between financial reporting opacity and crash risk. Varies according to the specification, United States. x: Opacity (calculated in three ways: a measure of earnings management, such as in Hutton et al. (2009), the presence of restatements, and the presence of material ICW. y: Crash risk (calculated as the annual measure of implied volatility smirk of options). Controls: ATM implied volatility level, size, leverage, market-to-book ratio, earnings, sales, and cash flow volatilities, average stock turnover, market beta, volatility and negative skewness of stock returns, and annual stock returns. OLS, Panel data with fixed effects Xu et al. (2014): Test whether the consumption of executives excess perk relates to crash risk. No. They used the simultaneous equation with that purpose and control variables. Not mentioned. Used control variables. They found that the increase of interest rate derivatives (their measure for opacity) is significantly and positively related to bank s future stock price crash risk. They found evidence that the steepness of volatility smirk (crash risk) increases with financial reporting opacity. Continued on next page

8 Sample Variables Model Instruments Endogeneity Concerns Conclusions Table 1 Continued from previous page 2,171 firm-years obs of Chinese state-owned enterprises (SOEs), x: residuals of a regression of Perk/Sales on employee s compensation, total assets, and the income per capita of the region in which the firm is located. y: Crash risk (negative conditional skewness and the down-to-up volatility of firm-specific weekly returns). Controls: opacity (firm-level earnings management), conditional conservatism (Cscore from Khan and Watts (2009)), investor heterogeneity, the negative skewness of past firm-specific stock returns, the standard deviation of past firm-specific stock returns, the average firm-specific weekly return over the past year, Size, MTB, ROA, CEO s equity ownership, and dummies for industry and year dummies. OLS. Two-Stage least squares (2SLS). Government spending in the province where the firm is located; Average consumption of excess perk from all other firms in the same industry. Andreou et al. (2016): investigate which corporate governance attributes explain stock prices crashes. 8,119 firm-year obs., from , United States. x: 21 corporate governance attributes of ownership structure, accounting opacity, board structure and processes, and managerial incentives. y: Crash (negative conditional skewness and the down-to-up volatility of firm-specific weekly returns). Controls: investor heterogeneity, past average and past volatility of firm-specific weekly returns lagged firm size lagged market value of equity to book value of equity, lagged financial leverage, and lagged return on equity. Principal component analysis (PCA) to reduce the number of governance attributes. Ordinary Least Squares. Panel data with fixed effects. No. They used the 2SLS estimation and control variables. Include three lags of the dependent variable to account for dynamic endogeneity. Believe that they handled the endogeneity caused by unobserved heterogeneity employing a broad set of corporate governance attributes and many control variables. Excess perks predict crash risk. Opacity increases the impact and the conditional conservatism decreases it. They found that opacity of financial reports does not increase stock crashes, while future crashes increase with transient institutional ownership, CEO stock option incentives and the percentage of outside directors that hold shares and decrease with percentage of stocks held by insiders, the level of accounting conservatism, board size and the existence of a formal governance policy in the companies mandate.

9 3 Model Estimation Our model of interest is: Crash it = β 0 + β 1 Opac i,t 1 + β 2 Crash i,t 1 + X i,t 1 β + γ i + u it, (1) where Opac it is the accounting opacity at t, Crash it is the occurrence of sudden decreases in stock prices in the following year, thus, representing the future crash risk, γ i represents the unobserved heterogeneity constant in time, such as the business risk of the firm, and u it represents the other unobserved factors that varies over time, representing the quality of the overall information environment of each firm at each year. This specification is similar to the Andreou et al. (2016), who include set the dependent variable in the future and include its lagged value as explanatory variables. The authors estimate this model by either OLS or panel data with fixed effects. OLS in equation (1) generates biased and inconsistent estimates because the error ν it = γ i + u it is naturally correlated with Crash i,t 1, as well as inefficient estimates, due to the autocorrelation of ν it generated by γ i. While the autocorrelation problem can be solved in a random effects estimation, the most serious issue of inconsistency remains. literature of panel data suggests that when γ i is correlated to the explanatory variables, consistent estimates can be obtained through a data transformation, by the within-groups average in a fixed effects structure of by the data differentiation, which eliminates all time constant terms of the model, observable or not. However, the fixed effect is invalid when the model presents a lagged dependent variable among the explanatory variables. The This happen because the identification hypothesis of the panel data models (random effects, fixed effects, and first-differences) is the strict exogeneity, which is the assumption that the explanatory variables X it are not correlated to the error term u is for all t and s. This assumption implies that, once X it and γ i are controlled for, past or future values of X cannot affect y (Wooldridge, 2010). However, in our model this is necessary wrong. If we use the within transformation in the fixed effects approach, we estimate the model with demeaned variables, that is, y it ȳ it, X it X it and u it ū it, however, ū it carries u i,t 1 so that it necessarily correlates with y i,t 1, generating inconsistent estimates. If we eliminate γ i via a first-differences transformation we have the same problem, since u it u i,t 1 is necessarily correlated with y i,t 1 y i,t 2. Nevertheless, models with lagged dependent variable among the explanatory variables, however, can be estimated consistently under a less restrictive and more natural hypothesis than the strict exogeneity, which is the sequential exogeneity. The sequential moment restrictions (Chamberlain, 1992): E[u it X it, X i,t 1, X i,t 2,, X i1, γ i ] = 0, (2)

10 implies that controlling by X it and γ i, past values of X it have no effect on y. Applied to equation (1), conditioning to γ i, only past shocks on the information environment affect the current opacity levels, future shocks do not have such effect: E[u it Opac i,t 1, Opac i,t 2,, Opac i0, X i,t 1, X i,t 2,, X i0, γ i ] = 0, e, E[u it Crash i,t 1, Crash i,t 2,, Crash i0, X i,t 1, X i,t 2,, X i0, γ i ] = 0, (3) in other words, E[X i,t 1 u it ] = 0, t s. While the strict exogeneity hypothesis is clearly too strong for the interest model of this research, the sequential exogeneity hypothesis seems adequate, as argued by Keane and Runkle (1992) for models of rational expectations to which potential instruments are functions of variables in t 1, in a way that they are never strictly exogenous. It is reasonable to say that while changes on the information environment of a firm in a certain year must affect earnings management practices and the future returns distribution on the following years, there is no reason to believe that the returns distribution and the current information manipulation practices would be affected by changes on the information environment still to come on the following years. Anderson and Hsiao (1982) proposed an estimator of instrumental variables firstly obtained by taking first-differences to control for γ i and then to include y t 3 or y i,t 3 as instruments to y i,t 1. Arellano and Bond (1991) suggest an overidentified model using all previous information available, i.e., using all lagged variables left as instruments, in a GMM approach, which generates a consistent and efficient estimator under the exogeneity of the instruments Z (y i,t 3, y i,t 4,, y i0 ). Therefore, in order to obtain consistent and efficient estimates to the parameters of equation (1), the procedure used is, initially, measure the first differences to eliminate the unobserved heterogeneity γ i and estimate equation (4) via GMM using Crash i,t 3,, Crash i0 as instruments for Crash i,t 1 : Crash it = β 0 + β 1 Opac i,t 1 + β 2 Crash i,t 1 + X i,t 1 β u it. (4) We believe this strategy results in more trustworthy estimates of the accounting opacity effect on future crash risk than the approaches taken by the other studies we discussed. 4 Data This papers uses, following the previous literature An and Zhang (2013), Andreou et al. (2016), Chen, Hong, and Stein (2001), Dewally and Shao (2013), J. B. Kim, Li, and Zhang (2011), Xu et al. (2014), the down-to-up volatility as a measure for crash risk, which is the ratio of the standard deviation of stock returns on the down weeks to the

11 standard deviation on the up weeks. To compute this measure, we estimate the weekly idiosyncratic stock returns ˆɛ is of each firm of the sample are observed in each year as ˆɛ is = r is α i + β i r m,s, (5) and then compute the total of down and up weeks, which are the weeks when the returns were below or above the average on that year, respectively. Thus, the variable Crash of the firm i at the year t is calculated as: Crash it = log ( (n up 1) s t ˆɛ down ) is (n down 1) s t ˆɛ up is where n up and n down correspond to the number of up and down weeks on year t, s t ˆɛ down is and s t ˆɛ up is are the sum of the squared weekly idiosyncratic returns below or above the average, respectively, on year t. The variable is then winsorized at 5%. Values close to zero to the variable Crash indicates that the return distribution in weeks of ups and downs is balanced, greater values (positive) indicate that the down week volatility is higher than up weeks volatility, indicating a greater occurrence of sudden fall in prices and smaller (negative) values indicate greater occurrence of up in prices. We restrict our sample to firms with weekly prices available in all wee ks of the year. We believe that including firms which are not frequently traded can lead to a measurement error in the crash risk variable, since firms less frequently traded are subject to higher ups and downs in their prices. As a measure of accounting opacity, also according to previous literature (Andreou et al., 2016; Hutton et al., 2009), this paper uses the level of earning management via discretionary accruals in accordance with Dechow, Sloan, and Sweeney (1995). Accruals can be understood as the difference between the economic income of a year, presented on the Income Statement, and the total cash flow of the same year, which results of the accrual basis of financial accounting. The idea Dechow et al. (1995) is that part of the total accruals of a period can be seen as natural, or not discretionary, due to the patterns of revenue and receivables and depreciation and amortization of fixed assets, while the portion that cannot be explained by these patterns is considered discretionary and, thus, can be a result of manipulation. The idea is that the greater the discretionary accruals proportion, the greater the chance of manipulated financial statement. Therefore, discretionary accruals of a firm i on year t are estimated as the residuals ξ it of the regression T Acc it = α 1 + α 2 ( Rev it Rec it ) + α 3 P P E it + ξ it, (7) (6) where T Acc are the total accruals, Rev is the total net revenues, Rec is the total amount of receivables, and P P E is the gross property, plan and equipment, all of them weighted

12 by total assets. Equation (7) (11) was estimated in a panel of fixed effects. Once earnings management may happen increasing or decreasing accruals, we have: Opac it = ξ it, (8) thus, higher discretionary accrual values indicate a higher level of earnings management and, therefore, greater opacity. It is important to note that this measure does not apply to firms in the financial industry, so we restrict our analysis to non-financial firms. Finally, representing the set of control variables X it in equation (4), we use some firms characteristics found in the previous literature to be correlated with financial statements quality (Ahmed, Neel, & Wang, 2013; Barth, Landsman, & Lang, 2008), namely firms Size (Size), Book-to-market index (BT M), profitability (return on assets, ROA), leverage (Lev) and the incidence of accounting losses (Loss). We chose to carry our analysis evaluating two different markets. First, we analyse the United States stock market, in order to compare our results to the previous literature, which mostly focused on this country. Then, we also conduct the same analysis in Brazil, as an example of an emerging market so that we can evaluate possible differences related to the different level of development of the two markets as well as to the differences of accounting standards used in each country (United States uses U.S. GAAP and Brazil uses IFRS since 2010). Table 2 presents the variables mean values for the U.S. sample by year, which sums 12,114 firm-year observations, while table 3 do this for the Brazilian sample, which sums a much smaller number of firm-year observations, 1,863. Table 2: Descriptive Statistics for the U.S. sample Year N Crash Opac Size BT M ROA Lev Loss , , , From tables 2 and 3 one can see that the incidence of up weeks is higher in Brazil, since Crash is negative on average in all years but 2007 (mostly likely due to the financial global crisis), while the distribution of up and down weeks in the U.S. is better balanced,

13 Table 3: Descriptive Statistics for the Brazilian sample Year N Crash Opac Size BT M ROA Lev Loss since Crash is nearer to zero in this sample, but seems to have increased in the most recent years. Further, Accounting Opacity is much higher in Brazil than in the U.S., book values tend to be higher than market values more in Brazil than in the United States, and Brazilian firms are more leveraged. The occurrence of accounting losses are slightly higher in Brazil, averaging 21% over the period, while in the U.S. the average is 15% of the cases. 5 Results Table 4 shows the estimation results for the U.S. sample, using OLS and panel data with random (RE) or fixed effects (FE), both without and including the lagged dependent variables, and the dynamic panel GMM estimation. In all models, the Crash risk is measured at t and all explanatory variables are measured at t 1. All models have clustered standard errors, except for the GMM estimation which is already efficient, and all of them include time dummies. From table 4 we can see that all models estimates a positive and significant effect for accounting opacity, except for the RE estimation with the lagged variable. The FE and the GMM estimations gauges the larger effects, while there is no much difference between these later two. Since these last two models are the ones who controls for the unobserved fixed effect γ i, we see that it seems to be negatively correlated with opacity, since not controlling for it seems to induce a downward bias in the Opac coefficient. However, the bias generated by the correlation between ū it and y i,t 1 does not seem to be large, since the difference between instrumentalizing it (GMM) or not (FE) is very small. However, when we analyze the Brazilian sample, we arrive at very different conclusions. as shown in table 5. The OLS and random effects estimations generate positive

14 and significant effects for accounting opacity, which disappears once we control for γ i in the fixed effects and GMM estimations. As in the U.S. case, the effect of the lagged Crash Risk is very different in these later two models. When the current crash risk is not instrumentalized (FE) it is negatively correlated with future crash risk, but once we balance it with the other lags (GMM) it no longer appears significant. However, this relationships are not important for moderating the effect of accounting opacity. Table 4: Estimates for the U.S. Sample Future Crash Risk OLS OLS RE RE FE FE GMM (1) (2) (3) (4) (5) (6) (7) Opac (0.098) (0.098) (0.100) (0.087) (0.136) (0.141) (0.148) Crash (0.013) (0.018) (0.016) (0.012) Size (0.004) (0.004) (0.004) (0.003) (0.028) (0.029) (0.045) BT M (0.014) (0.014) (0.014) (0.012) (0.014) (0.014) (0.016) ROA (0.134) (0.132) (0.136) (0.095) (0.141) (0.145) (0.141) Lev (0.0004) (0.0004) (0.0004) (0.0004) (0.0004) (0.0003) (0.0002) Loss (0.020) (0.020) (0.020) (0.026) (0.028) (0.028) (0.033) Constant (0.031) (0.031) (0.031) (0.030) N 12,114 12,043 12,114 12,043 12,114 12,043 10,797 R Adj. R F Stat Wald Stat Note: p<0.1; p<0.05; p<0.01 The results show that, in the U.S. most specifications tend to find a positive and significant effect of accounting opacity on the future crash risk. However, the magnitude of the effect is different as we consider different sources of bias and control for them. Simple OLS estimations always gauge positive effects, however, they cannot be trusted because unobserved time-constant factors, such as the business risk, are not controlled for, as in the case for random effects estimations. When we control for these factors

15 Table 5: Estimates for the Brazilian Sample Future Crash Risk OLS OLS RE RE FE FE GMM (1) (2) (3) (4) (5) (6) (7) Opac (0.010) (0.012) (0.011) (0.010) (0.120) (0.120) (0.114) Crash (0.030) (0.032) (0.027) (0.036) Size (0.014) (0.014) (0.015) (0.013) (0.058) (0.058) (0.135) BT M (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) (0.017) ROA (0.008) (0.150) (0.009) (0.166) (0.075) (0.286) (0.355) Lev (0.0002) (0.0003) (0.0003) (0.0002) (0.001) (0.001) (0.0005) Loss (0.043) (0.045) (0.045) (0.043) (0.068) (0.071) (0.076) Constant (0.097) (0.097) (0.107) (0.081) N 1,863 1,842 1,863 1,842 1,863 1,842 1,513 R Adj. R F Stat Wald Stat Note: p<0.1; p<0.05; p<0.01

16 we yield a much higher effect of accounting opacity in the U.S. market. The results are similar for the U.S. and the Brazilian market for the OLS and random effects estimation, but are strikingly different for the fixed effect and dynamic panel estimations, since the accounting opacity effect increases for this models in the U.S. market but disappears in the Brazilian market. As seen in table 2 and 3, the Brazilian and U.S. firms form two very distinct sets of firms, which leads us to hypothesize the differences in the results are due to these inherent differences in firms characteristics. In order to further investigate the differences of the results for the two markets we created a sample of Brazilian firms who are matched to the U.S. firms and re-run the analysis. To build the match samples we estimate logit regressions using both the variables of crash risk and accounting opacity as well as the controls (firms size, BTM, ROA, leverage and a dummy indicating accounting losses). Table 6 shows the results for this Brazilian matched sample. The models with lagged dependent variables have much less observations because the matching does not guarantee sequential data availability. The results indicate that it is not only firms characteristics which are driving the absence of results, if we consider the GMM estimation. While the GMM opacity coefficient is still not significant, all other models generate significant positive results for opacity. However, it is important to notice some differences between the results of tables 5 and 6. First, the coefficients of opacity for the matched sample are larger, approximating the results for the Brazilian market to the U.S. results. Second, in this matched sample, the fixed effects estimation yields significant effects for opacity, which disappears once lagged crash risk is accounted for and instrumentalized. This suggests that while firms inherent characteristics play an important role in the relationship between future crash risk and accounting opacity, other factors differentiating the Brazilian and the U.S. market are also responsible for the estimated differences. 6 Concluding Remarks In this research we revisited the problem of identifying the effect of accounting opacity in the stocks future crash risk. The previous literature relies on different estimations strategies for analyzing how opaque financial reports affects future crash risk, which yields different results. Here, we discuss this issue estimating this effect under different assumptions for two different markets: the United States, which is extensively studied in the previous literature, and Brazil, as an example of an emerging market with different levels of financial development and accounting systems. We estimated the effect of accounting opacity on future crash risk using OLS, random and fixed effects and dynamic panel GMM, using data from 2001 to 2015 for the firms listed in these two countries. We conduct a separate analysis for each country and also a matched sample analysis. The results indicate that for the U.S. sample, OLS and

17 Table 6: Estimates for the Brazilian sample matched with U.S. firms Future Crash Risk OLS OLS RE RE FE FE GMM (1) (2) (3) (4) (5) (6) (7) Opac (0.155) (0.191) (0.156) (0.185) (0.176) (0.359) (0.481) Crash (0.029) (0.029) (0.061) (0.092) Size (0.008) (0.019) (0.008) (0.019) (0.041) (0.062) (0.095) BT M (0.008) (0.072) (0.009) (0.070) (0.013) (0.113) (0.110) ROA (0.125) (0.117) (0.125) (0.108) (0.106) (0.217) (0.326) Lev (0.002) (0.004) (0.002) (0.004) (0.002) (0.007) (0.067) Loss (0.049) (0.082) (0.049) (0.082) (0.050) (0.128) (0.150) Constant (0.079) (0.146) (0.079) (0.150) N 1,844 1,508 1,844 1,508 1,844 1,508 1,245 R Adj. R F Stat Wald Stat Note: p<0.1; p<0.05; p<0.01

18 random effects subestimates the effect of opacity, while for the Brazilian sample, these methods seems to overestimate the effects. Since the two samples form very distinct sets of firms, with different levels of opacity and crash risk, for instance, we also run an analysis matching the Brazilian sample with the U.S. firms. Doing so, we see higher effects for the OLS, random and fixed effects estimates, which are more similar to the U.S. results, but the GMM results are still not significant, contrasting with the U.S. firms. References Ahmed, A. S., Neel, M., & Wang, D. (2013). Does mandatory adoption of IFRS improve accounting quality? Preliminary evidence. Contemporary Accounting Research, 30 (4), An, H., & Zhang, T. (2013). Stock price synchronicity, crash risk, and institutional investors. Journal of Corporate Finance, 21, Anderson, T. W., & Hsiao, C. (1982). Formulation and estimation of dynamic models using panel data. Journal of Econometrics, 18, Andreou, P. C., Antoniou, C., Horton, J., & Louca, C. (2016). Corporate governance and firm-specific stock price crashes. European Financial Management. Arellano, M., & Bond, S. R. (1991). Some specification tests for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies, 58, Armstrong, C. S., Guay, W. R., & Weber, J. P. (2010). The role of information and financial reporting in corporate governance and debt contracting. Journal of Financial Economics, 50, Barth, M. E., Landsman, W. R., & Lang, M. H. (2008). International accounting standards and accounting quality. Journal of accounting research, 46 (3), Becht, M., Bolton, P., & Röell, A. (2003). Corporate governance and control. Handbook of the Economics of Finance, 1, Bergstresser, D., & Philippon, T. (2006). CEO incentives and earnings management. Journal of Financial Economics, 80 (3), Burns, N., & Kedia, S. (2006). The impact of performance-based compensation on misreporting. Journal of Financial Economics, 79 (1), Callen, J. L., & Fang, X. (2013). Institutional investors and crashes: Monitoring or expropriation? Journal of Banking and Finance, 43, Chamberlain, G. (1992). Comment: Sequential moment restrictions in panel data. Journal of Business and Economic Statistics, 10, Chen, J., Hong, H., & Stein, J. C. (2001). Forecasting crashes: Trading volume, past returns, and conditional skewness in stock prices. Journal of Financial Economics, 61 (3),

19 Cheng, Q., & Warfield, T. D. (2005). Equity incentives and earnings management. The Accounting Review, 80 (2), Dechow, P. M., Sloan, R. G., & Sweeney, A. P. (1995). Detecting earnings management. The Accounting Review, 70 (2), DeFond, M., Hung, M., Li, S., & Li, Y. (2013). Does manadatory IFRS adoption affect crash risk? Dewally, M., & Shao, Y. (2013). Financial derivatives, opacity, and crash risk: Evidence from large us banks. Journal of Financial Stability, 9 (4), Eckbo, E. (2014). Corporate takeovers and economic efficiency. The Annual Review of Financial Economics, 6, Edmans, A. (2014). Blockholders and corporate governance. The Annual Review of Financial Economics, 6, Efendi, J., Srivastava, A., & Swanson, E. P. (2007). Why do corporate managers misstate financial statements? The role of option compensation and other factors. Journal of Financial Economics, 85 (3), Habib, A. (2008). Corporate transparency, financial development and the allocation of capital: Empirical evidence. Abacus, 44 (1), Hutton, A. P., Marcus, A. J., & Tehranian, H. (2009). Opaque financial reports, r-squared, and crash risk. Journal of financial Economics, 94 (1), IFRS Foundation. (2010). Basis for conclusions on Chapter 1: The objective of general purpose financial reporting. In Red Book Part A. Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behaviour, agency costs and ownership structure. Journal of Financial Economics, 3, Jin, L., & Myers, S. C. (2006). R-squared around the World: New theory and new tests. Journal of Financial Economics, 79, Keane, M. P., & Runkle, D. E. (1992). On the estimation of panel data models with serial correlation when instruments are not strictly exogenous. Journal of Business and Economic Statistics, 10, 1 9. Khan, M., & Watts, R. L. (2009). Estimation and empirical properties of a firm-year measure of accounting conservatism. Journal of accounting and Economics, 48 (2), Kim, J. B., Li, Y., & Zhang, L. (2011). CFOs versus CEOs: Equity incentives and crashes. Journal of Financial Economics, 101, Kim, J. B., & Zhang, L. (2014). Financial reporting opacity and expected crash risk: Evidence from implied volatility smirks. Contemporary Accounting Research, 31 (3), Kim, Y., Li, H., & Li, S. (2014). Corporate social responsibility and stock price crash risk. Journal of Banking and Finance, 43, 1 13.

20 Kothari, S. P., Shu, S., & Wysocki, P. (2009). Do managers withhold bad news? Journal of Accounting Research, 47 (1). La Porta, R., Lopez-de-Silanes, F., Shleifer, A., & Vishny, R. (2000). Investor protection and corporate governance. Journal of Financial Economics, 58 (1), Peng, L., & Röell, A. (2008). Manipulation and equity-based compensation. American Economic Review, 98 (2), Roberts, M. R., & Whited, T. M. (2013). Endogeneity in empirical corporate finance. Handbook of the Economics of Finance, 2, Shleifer, A., & Vishny, R. W. (1997). A survey of corporate governance. The Journal of Finance, 52 (2), Wintoki, M. B., Linck, J. S., & Netter, J. M. (2012). Endogeneity and the dynamics of internal corporate governance. Journal of Financial Economics, 105, Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd). Cambridge, Massachusetts: The MIT Press. Xu, N., Li, X., Yuan, Q., & Chan, K. C. (2014). Excess perks and stock price crash risk: Evidence from china. Journal of Corporate Finance, 25, Zingales, L. (1997). Corporate governance. The New Palgrave Dictionary of Economics and the Law.