HOW EXPECTATIONS OF ALGORITHM-BASED INFORMATION PROCESSING CAN AFFECT FIRM DISCLOSURES AND DECISIONS A Dissertation Presented to the Faculty of the Graduate School of Cornell University In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy by Patrick Debevec Witz May 2020 © 2020 Patrick Debevec Witz HOW EXPECTATIONS OF ALGORITHM-BASED INFORMATION PROCESSING CAN AFFECT FIRM DISCLOSURES AND DECISIONS Patrick Debevec Witz, Ph. D. Cornell University 2020 Sophisticated investors increasingly use algorithms to analyze firm disclosures. This change has implications for how managers think about their capital market users. In two studies, I examine how (1) expectations of capital market users can affect real and accruals-based earnings management and (2) how these expectations can be the result of systematic cognitive biases in addition to other factors. In the first study, I employ two experiments and a survey to demonstrate that the presence of algorithms can increase the likelihood of real earnings management through one expectation-based mechanism, yet decrease the likelihood of accruals-based earnings management through a second expectation-based mechanism. A survey of experienced managers demonstrates that the expectations that lead to these effects exist within the institutional environment. In the second study, I employ two experiments and a survey to examine how managers’ expectations of investors’ information processing can be systematically biased through a psychological process involving egocentric focus and insufficient adjustment. Both of these studies contribute to existing literature relating to financial disclosure and the emerging literature on how expectations of information processing costs and capabilities can affect manager decision making. BIOGRAPHICAL SKETCH Patrick Debevec Witz was born in East Longmeadow, MA. He pursued his undergraduate degree at the University of Massachusetts at Amherst, where he was recognized as a Jack Welch Scholar. Patrick graduated first in his class, with double majors in Economics and Accounting & Information Systems. Upon graduation, Patrick worked for three years as an auditor in the Boston Office of Ernst & Young. Patrick then pursued a Ph.D. at Cornell University, where he developed strong interests in academic research relating to the evolution of the financial disclosure landscape and the introduction of emerging technologies into the accounting institutional environment. iii ACKNOWLEDGMENTS I am grateful for the invaluable guidance and support I received from my dissertation committee members: Robert Libby, Kristina Rennekamp, Ryan Guggenmos, and J. Edward Russo. I thank Sanjeev Bhojraj, Robert Bloomfield, Kelsey Brasel, Nicole Cade, Mark Cecchini, Ben Commerford, Michael Durney, Travis Dyer, Harry Evans, Mac Festa, Vicki Hoffman, Scott Jackson, Eric Johnson, Marlys Lipe, Patrick Martin, Don Moser, Mark Nelson, Nadra Pencle, Jason Rasso, Mani Sethuraman, Blake Steenhoven, Chad Stefaniak, Scott Vandervelde, Xinyu Zhang, Aaron Zimbelman, Luo Zuo and workshop participants at Ball State University, Bucknell University, Cornell University, the University of Kentucky, the University of Pittsburgh, the University of South Carolina, and the University of Wyoming, for their helpful comments and suggestions. I am grateful for the financial support provided by Cornell University and the Deloitte Foundation. Any remaining errors are my own. TABLE OF CONTENTS BIOGRAPHICAL SKETCH ......................................................................................... iii TABLE OF CONTENTS .............................................................................................. iv CHAPTER 1 ................................................................................................................... 1 INTRODUCTION ...................................................................................................... 2 BACKGROUND AND HYPOTHESIS DEVELOPMENT ....................................... 6 EXPERIMENT 1 – METHOD ................................................................................. 12 EXPERIMENT 1 – RESULTS ................................................................................. 16 EXPERIMENT 2 – METHOD ................................................................................. 18 EXPERIMENT 2 – RESULTS ................................................................................. 20 SURVEY OF EXPERIENCED MANAGERS ......................................................... 22 CONCLUSION ......................................................................................................... 27 FIGURES .................................................................................................................. 32 TABLES ................................................................................................................... 34 APPENDICES .......................................................................................................... 43 CHAPTER 2 ................................................................................................................. 44 INTRODUCTION .................................................................................................... 45 BACKGROUND AND HYPOTHESIS DEVELOPMENT ..................................... 50 EXPERIMENT 1 – METHOD ................................................................................. 55 EXPERIMENT 1 – RESULTS ................................................................................. 58 EXPERIMENT 2 – METHOD ................................................................................. 61 iv EXPERIMENT 2 – RESULTS ................................................................................. 63 SURVEY OF EXPERIENCED MANAGERS ......................................................... 64 CONCLUSION ......................................................................................................... 65 FIGURES .................................................................................................................. 70 TABLES ................................................................................................................... 74 APPENDICES .......................................................................................................... 81 REFERENCES ............................................................................................................. 85 v CHAPTER 1 – How Expectations of Algorithm-Based Information Processing Can Affect Firm Disclosures and Decisions ABSTRACT Sophisticated investors increasingly use algorithms to analyze firm disclosures. This change has implications for how managers think about their capital market users. In two laboratory experiments and a survey, I examine how perceptions of capital market users can affect real and accruals-based earnings management. In both experiments, I manipulate managers’ expectations of whether their long-term institutional investors will rely upon algorithms or analysts to process information, while holding beliefs about algorithms and analysts constant. Results from the first experiment indicate that the presence of algorithms can increase the likelihood of real earnings management by limiting the extent to which managers feel they can effectively use narrative reporting to convey multi-period opportunities. Results from the second experiment indicate that the presence of algorithms can decrease the likelihood of accruals-based earnings management, in line with General Deterrence Theory. My survey confirms the experimental assumptions about experienced managers’ beliefs concerning the relative strengths and weaknesses of algorithms and analysts, and provides external validity and context to the experimental findings. This study contributes to existing literatures relating to financial disclosure and earnings management, and advances understanding of how expectations over how capital market users process information may affect manager behavior. 1 I. INTRODUCTION Algorithmic investing is reshaping capital markets. Five out of the top six largest hedge funds in the world are now structured around algorithmic approaches (Cantrell 2017) and even traditional funds now rely on hybrid approaches that emphasize algorithms. Prior literature recognizes audience knowledge as fundamental to effective communication (Fussell and Krauss 1992; Hirsh, Kang, and Bodenhausen 2012) and managers devote extensive resources to identifying and understanding their capital market users, and tailoring the firm’s message to these users (Brown, Call, Clement, and Sharp 2019). As a result, differences in managers’ perceptions of algorithms’ and traditional analysts’ information processing strengths and weaknesses could affect both operating and reporting decisions. In this study, I examine whether managers’ beliefs about the use of algorithms by long- term institutional investors can increase the extent to which managers engage in real earnings management through one mechanism, yet decrease the extent to which managers engage in accruals-based earnings management through a second mechanism. I examine these issues in two stages. In both Experiment 1 and Experiment 2, participants assume the role of managers and receive background information about a company. In both experiments, I hold constant participants’ beliefs about the relative strengths and weaknesses of algorithms and analysts, and manipulate whether participants believe the majority of their capital market users are long-term institutional investors who rely primarily on algorithms or analysts to process information. Then, in the survey, I verify my assumptions about experienced managers’ beliefs about the relative strengths and weaknesses of algorithms and analysts, and managers’ expected responses to those differences. In Experiment 1, participants are presented with a real earnings management scenario. 2 They indicate how likely they are to invest in an R&D project that results in higher income in the long term but lower income in the current period, and jointly choose whether to provide a voluntary narrative disclosure about this decision to investors. Participants are significantly more likely to engage in real earnings management (and not initiate the long-term profitable project) when long-term institutional investors rely primarily on algorithms rather than analysts to process information. Participants also choose to provide fewer narrative disclosures when long- term institutional investors rely upon algorithms, and are less confident in the persuasiveness of the narrative disclosures that they do provide. These findings are consistent with the idea that perceived information processing characteristics are judged by managers to impact the effectiveness with which they can convey narrative information. In Experiment 2, participants are presented with an accruals-based earnings management scenario and indicate how likely they are to adjust estimates to achieve an earnings target. Participants are significantly less likely to engage in accruals-based earnings management when long-term institutional investors rely primarily on algorithms rather than analysts. This appears to result from participants’ increased concerns of detection in the presence of investors who rely primarily on algorithms, despite managers’ significant discretion to adjust estimates to be less conservative, although this evidence is not conclusive. Finally, in the survey of experienced managers, I capture current beliefs about the relative information processing strengths and weaknesses of algorithms and analysts and experienced managers’ expected responses to those differences. Survey participants (1) report current beliefs about algorithms and analysts that are consistent with my experimental assumptions, (2) report that they would adjust their own reporting and operating behavior in response to expectations of how their audience will process the information, and (3) exhibit actual behavior in mini- 3 scenarios consistent with my experimental results. Survey participants also provide insights about future trends involving the use of algorithms within their firms. These findings have important implications for managers given ongoing changes to the capital markets. First, this study contributes to the literature on the effects of managers’ perceptions of how capital market users process information. Different capital market users may be expected to have different information processing strengths and weaknesses, and managers have been encouraged by regulators to think about and accommodate how users process information (SEC 1998; 2009).1 While the majority of this literature focuses on the determinants of ease of processing for different types of users of financial disclosures and their effects on capital market outcomes (Allee et al. 2018; Blankespoor, deHaan, and Marinovic 2019; Li 2008; Rennekamp 2012), this study’s results suggest two important ways that perceptions of these processing differences can affect managers’ judgments and decisions. Blankespoor et al. (2019) discuss the need for research to explore information processing-based feedback effects such as this, and this study leverages the comparative advantages of experimental and survey methodology to obtain strong causal inferences and insight into managers’ strategic intent, which has traditionally been unobservable. Second, this study adds to the literature on real earnings management. Prior literature suggests that managers are willing to choose less profitable longer-term investments to boost current earnings (Graham, Harvey, and Rajgopal 2005), and that one purpose of narrative explanations is to convince others about the effectiveness of strategies not easily seen from quantitative measures alone (Bentley 2019). This study builds upon these findings by 1 Regulators have done this with both “readability” (SEC 1998) and “scriptability” in the form of the XBRL mandate (SEC 2009). “Readability” refers to the ease with which a human reader can process and comprehend written text while “scriptability” (Allee, DeAngelis, and Moon 2018) refers to the ease in which a computer program can transform large amounts of unstructured data in firm disclosures into useable information. 4 demonstrating that the use of algorithms by long-term institutional investors to process information can reduce investment and increase real earnings management by changing manager perceptions of the effectiveness with which they feel they can convey various types of information through financial disclosures. This study suggests that even when managers have the opportunity to provide narrative disclosures, if they feel disclosures are less likely to be persuasive based on expected information processing characteristics, managers may preempt the need to use disclosures by engaging in fundamentally different investment behavior. This study also contributes to literature on accruals-based earnings management. Prior literature suggests that earnings management may be widespread and that managers may engage in it strategically (Dechow, Ge, and Schrand 2010). Prior literature also explores how manager psychology can affect the nature and extent of earnings management (Asay 2018; Brown 2014; Guggenmos 2019). This study suggests that expectations over how capital market users process information can reduce accruals-based earnings management through one mechanism, and it can increase real earnings management through another. This finding can extend prior work on factors which affect the utilization of different earnings management techniques (Gunny 2010; Zang 2012) and have important consequences going forward. By demonstrating that managers strategically choose to manage earnings more or less aggressively, and in different ways, based on expectations of how capital market users will process information, this study helps address an open question in the literature (Blankespoor et al. 2019). These findings also offer contributions to practice. By emphasizing to managers the need to consider how certain types of users process information (SEC 1998; 2009), regulators may inadvertently be affecting managers’ considerations around how they issue disclosures and make key investment decisions. While the processing of information by capital market users is 5 becoming more efficient, the changes in the way in which information is processed can affect the information that reaches these users in the first place. These unintended consequences could have ramifications that regulators may wish to consider. Together, my two experiments and survey offer perspective on how the effects of algorithmic investing can alter the landscape of well- documented areas of the literature. Expectations of information processing advantages and disadvantages of algorithms may be consequential, and intrinsically connected. In presenting both, I highlight the mechanisms by which two expectation-based effects can operate, and call attention to the consequences that the growth of algorithmic investing may have on aspects of financial reporting and decision making in the future. The remainder of this paper is organized as follows. Section II discusses the theories used to develop the experiments and contains my specific hypotheses. Sections III and IV discuss the method and results, respectively, for Experiment 1. Sections V and VI discuss the method and results, respectively, for Experiment 2. Section VII discusses the survey of experienced managers. Section VIII concludes the paper. II. BACKGROUND AND HYPOTHESIS DEVELOPMENT In this study, I demonstrate that algorithmic approaches to investing can have an effect beyond direct capital market outcomes. I focus on how managers’ expectations of how algorithm- versus analyst-focused capital market users will process information can affect their operating and disclosure decisions. Research has just begun to examine information-processing- based feedback effects such as this, and Blankespoor et al. (2019) note that there is considerable potential for new research in this area. Feedback Effects of Investor Information Processing on Manager Decision-Making 6 Prior literature demonstrates that managers and firms devote considerable resources to understanding their capital market users and tailoring their messaging to these users (Brown et al. 2019; Bushee and Miller 2012; Kirk and Vincent 2014; Trentmann 2019). Firms are individually attentive to important institutional investors (Brown et al. 2019) and will go out of their way to provide information to these users (Kirk and Markov 2016; Solomon and Soltes 2015). This literature suggests that firms take great care in how they communicate with users, and likely expect their efforts to influence how their capital market users process information and arrive at decisions. Yet expectations of how investors will process information may also affect managers’ operating and disclosure decisions, in what Blankespoor et al. (2019) term a “feedback effect.” Blankespoor et al. (2019) discuss feedback effects as an interesting and broad avenue for research, although they note that existing work is just beginning to examine the area. Early findings suggest that disclosure characteristics can be affected by expectations of information processing costs (Abramova, Core, and Sutherland 2019; Basu, Pierce, and Stephan 2019; Blankespoor 2019; Guay, Samuels, and Taylor 2016). A key challenge with studies in this area is that manager intent is unobservable and it is difficult to isolate managers’ discretionary responses to information processing expectations when broadly viewing both operating and disclosure outcomes (Blankespoor et al. 2019). My study leverages the comparative advantages of experimental and survey methods to control the information environment, manipulate and measure expectations over investor information processing, and obtain direct evidence into current information processing beliefs and actual strategic behavior in response to these beliefs. In doing so, I explore whether managers strategically choose to manage earnings more or less aggressively, and in different ways, based 7 on expectations of how capital market users will process information. This helps address an open question in the literature (Blankespoor et al. 2019). Expectations of Algorithm-based Information Processing Allee et al. (2018) observe a general trend where analysis of financial information in SEC filings has shifted towards programmers and computers and away from manual processing by individual analysts. This alleviates the problem where, historically, processing of quantitative data has been money- and time-intensive for investors to perform (Whalen 2004). The standardization of data into a machine-readable XBRL format has also increased the speed of information acquisition for many investors (Bhattacharya, Cho, and Kim 2018). In particular, machine-readable data seems well-suited to quantitative or algorithmic investment strategies, which have experienced a rapid rise that coincides with the rise of standardized quantitative data (Wigglesworth 2017; Zuckerman and Hope 2017). However, the SEC explicitly warns that a standardized reporting format “might affect a company’s ability to communicate its unique financial attributes to investors” (SEC 2009).2 Similarly, Sloan (2019) warns that quantitative investing strategies often overlook fundamental analysis which relies on qualitative factors in addition to quantitative ones. These changes in data availability and investing approach reflected in the above discussion offer insight into a common view of the strengths and weaknesses of algorithms relative to human analysts as they process information, where algorithms may be perceived as 2 In its implementation of XBRL, the SEC (2009) specifically weighs the benefits of XBRL in “facilitating [investors] automated parsing and analysis of performance information” and “facilitating easier comparability between companies” against the costs of XBRL in terms of “less precise information about a company.” In a similar discussion, Allee et al. (2018) observe that if disclosures can be standardized or made scriptable, algorithm-based technology can enable data analysis at scale through greatly enhanced processing capabilities and techniques. However, if inputs are non-comparable across firms, the pre-programmed decision rules that algorithms rely upon can perform poorly. Liberti and Petersen (2018) provide a discussion on the relative suitability of quantitative information to standardization, noting that valuable context can often be lost in the process of standardizing qualitative information. 8 having relative strengths when focusing on (typically more standardized) quantitative information and relative weaknesses when focusing on (typically less standardized) qualitative explanations. This view is supported by psychology literature on algorithm aversion and algorithm appreciation in a variety of decision making contexts. This literature demonstrates relative trust or distrust of information provided by algorithms and algorithm-based technology in a variety of contexts, in both personal (Logg, Minson, and Moore 2019) and professional (Commerford, Dennis, Joe, and Wang 2019) settings. One key distinction emerges between qualitative domains and quantitative domains. The algorithm aversion literature demonstrates that advice from algorithms is routinely discounted in domains of qualitative judgments (Promberger and Baron 2006; Sinha and Swearingen 2001; Yeomans et al. 2019). Notably, advice relating to these judgments is discounted even in circumstances where algorithmic advice may have superior accuracy relative to individual expert judgments (Dawes, Faust, and Meehl 1989; Grove et al. 2000; Promberger and Baron 2006; Yeomans et al. 2019). On the other hand, the algorithm appreciation literature demonstrates that algorithmic advice is often relied upon more than human advice across a variety of domains where quantitative judgments are required (Logg, Minson, and Moore 2019). Real Earnings Management and Persuasion Gunny (2010) and Zang (2012) suggest that factors that affect the relative costs of real earnings management and accruals-based earnings management can affect whether managers engage in one or the other. My study investigates whether expectations over how long-term institutional investors are expected to process information is one such factor. While prior literature suggests that managers are willing to choose less profitable longer-term investments to boost current earnings (Graham, Harvey, and Rajgopal 2005), very few papers have examined 9 real effects of managers’ expectations of disclosure processing costs (see Roychowdhury, Shroff, and Verdi 2019 for a review).3 Because real earnings management activities can reduce firm value, firms may wish to persuade investors of the value of the activities rather than present superficially higher income in the current period. However, effective persuasion often relies upon understanding the audience that is being communicated to (Hirsh, Kang, and Bodenhausen 2012; Noar, Harrington, and Aldrich 2009) and individuals frequently employ different communication strategies with different audiences (Banerjee 2002; Fussell and Krauss 1992; Gardner and Martinko 1988; Gibbins, McCracken, and Salterio 2010). Accordingly, different expected information processing characteristics of algorithms relative to analysts may affect expectations of the effectiveness of persuasive arguments in influencing investors. Ultimately, this may make managers more hesitant to make persuasion attempts through narrative explanations when they expect their institutional investors to rely primarily on algorithms to process financial disclosures. Prior work suggests that opportunities to provide narrative explanations can deter measure management (Bentley 2019). However, if managers expect reduced potential for effective persuasion, managers may engage in real earnings management and forgo a voluntary opportunity to provide an explanation. Stated formally, I hypothesize that: H1: Managers are more likely to not initiate the R&D investment project (and engage in real earnings management) when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. I further expect this effect to occur as a result of differences in expectations as to the perceived effectiveness of narrative disclosures. If managers do not expect capital market users 3 The use of real activities to manage earnings has been increasing in the post-Sarbanes-Oxley period (Cohen, Dey, and Lys 2008), and these actions can reduce firm value (Roychowdhury 2006). 10 to be as persuaded by narrative disclosures, they will be less likely to issue voluntary disclosures and instead let the impact of different operating decisions on the financial statements make their argument for them. Stated formally, I hypothesize that: H2: Managers are less likely to provide a voluntary disclosure relating to their choice to initiate or not initiate an R&D investment project when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. Accruals Based Earnings Management and Detection Risk While I expect greater reliance on algorithms can potentially have a negative effect in terms of increasing real earnings management through decreasing expectations over the perceived persuasiveness of managers’ narrative disclosures, I expect greater reliance on algorithms can also have a positive effect in deterring accruals-based earnings management through increasing expectations of investor attentiveness to quantitative information. Managers have been shown to engage in accruals-based earnings management behavior strategically (Dechow et al. 2010), with manager psychology influencing behavior (Asay 2018; Brown 2014). Additional aspects of psychology might affect this strategic balance, such as perceptions relating to the likelihood of detection, as affected by overarching expectations of audience information processing characteristics. I focus my study on types of accruals-based earnings management that are subtle and thus more difficult to detect and deter. As Lo (2008) observes, the fundamental hope of those managing earnings is that it should not be easy to detect.4 Healy and Wahlen (1999) discuss the technical difficulty of detection, however other work highlights that earnings management can be detected with advanced quantitative models and analysis (Dechow, Sloan, and Sweeney 1995; 4 Prior literature also identifies an association between earnings management and the complexity of the information environment (Lo, Ramos, and Rogo 2017). 11 Phillips, Pincus, and Rego 2003). If regulatory enforcement targeting opportunistic earnings management is low and strong board monitoring is not present (Hales, Koka, and Venkataraman 2018), direct perceptions over the threat of detection by capital market users can impact managers’ decision making. General Deterrence Theory suggests that reducing the certainty of avoiding negative consequences can shift behavior, and suggests that an increased sense of monitoring can both affect perceptions of certainty and severity of potential detection (D’Arcy, Hovav, and Galletta 2009; Nagin and Pogarsky 2001). In these circumstances, an increased sense of analytical and attentiveness capabilities of investors who rely upon algorithms may deter opportunistic behavior, even for subtle forms of earnings management that may require advanced quantitative models to detect. Stated formally, I hypothesize that: H3: Managers are less likely to adjust estimates (and engage in accruals-based earnings management) when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. H4: Managers will feel that the likelihood of detection of actions to adjust estimates (and engage in accruals-based earnings management) is higher when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. III. EXPERIMENT 1 – METHOD Participants Participants in the first experiment are 100 MBA students recruited from a top-rated MBA program. On average, participants are 29 years old and have 5.38 years of work experience. 71% of participants report involvement in preparing performance reports for firm use. 80% of participants report involvement in making strategic decisions within a firm. 12 Design Choices The two experiments and survey were designed in combination to provide complementary evidence on whether differences in expectations over how capital market users will process information can affect managers’ operating and financial reporting decisions. The experiments are designed to allow for clean inferences and to have strong internal validity within a controlled laboratory setting (Libby, Bloomfield, and Nelson 2002). The survey is designed to contextualize the findings (Bloomfield, Nelson, and Soltes 2016) and demonstrate external validity to the larger institutional environment. As a result, the experiments are designed with the goal of exploring whether expectations over how institutional investors will process information affect the extent to which managers engage in real and accruals-based earnings management, given certain beliefs about the algorithms and analysts that investors rely on. The beliefs are held constant, with all participants receiving the same description of the relative strengths and weaknesses of algorithms and analysts. The survey is designed with the goal of verifying that experienced managers have the beliefs about algorithms and analysts assumed by the experiments. It is also designed to establish the robustness of the experimental findings. The survey is designed to show that experienced managers will behave in a similar manner to experimental participants (in two mini-scenarios modeled after the two experiments), as a result of already holding the same beliefs assumed in the experiments. The role of both methods is necessary to understand the design choices made for each method individually, and both methods are essential in arriving at the overall contribution of the study. Task and Manipulation Participants are asked to assume they are involved in making operating and financial 13 reporting decisions at Becker International, a hypothetical firm. They are presented with background information about Becker and are informed that they will be presented with a key operating decision for Becker in the current period as well as report performance, which is affected by this decision, to investors. The experiment uses a 1 x 2 between-subjects design that manipulates whether long-term institutional investors are expected to rely primarily on (1) algorithms or (2) analysts to process information. Setting and Procedure After being introduced to the task, participants read background information and are presented with details relevant to the setting. Participants are informed that they will be presented with an important operating decision that will affect reported financial performance. Information is provided about Becker’s business operations, which involve developing medical devices to sell to customers worldwide.5 Participants are also given a brief description of their long-term institutional investors, who they are told may vary in their reliance on analysts versus algorithms. In a manipulation across conditions, an investor relations officer informs participants that: “Institutional investors differ in their reliance on algorithms versus analysts. In examining patterns of usage, your investor relations officer has obtained statistics from the company’s website and corporate servers indicating that the vast majority of downloads of SEC filings are made in file formats used by analysts [algorithms].”6 Next, participants are presented with an operating scenario that describes the key decision they must consider. Participants are informed that they “have the option to not initiate or initiate a new R&D project.” The project is described as being likely to generate significant future benefits, however initiating the project will require that the company report higher operating 5 Business operations are based on the model of Boston Scientific. 6 The bracketed item in the quote represents the manipulation between conditions. 14 expenses and lower net income in the current financial period. Next, participants proceed to a page containing the operating decision. Participants are asked how likely they are to initiate or not initiate the new R&D project. After providing a likelihood response, participants are informed that they made the operating decision that they indicated they were more likely to make and are given an opportunity to provide an optional voluntary disclosure relating to this operating decision. On a later screen, participants are also directly asked to what extent they felt they would be able to use narrative disclosures to persuade capital market users of the effectiveness of strategies not easily seen from the numbers alone. Finally, participants provide demographic information. Dependent Measures Likelihood of R&D Investment Measure In my setting, real earnings management involves reducing long-term investments in order to improve short-term financial performance. Participants in my study face a decision as to whether to initiate or not initiate a new research and development (R&D) project. Participants indicate how likely they are to initiate or not initiate this new R&D project. Participants are presented with a scale ranging from “1 – extremely unlikely to initiate the project” to “8 – extremely likely to initiate the project”. Voluntary Disclosure and Persuasiveness Measures After participants indicate the choice they are likely to make, they are shown the impact of this decision on the current financials and they are asked if they would like to provide an additional voluntary disclosure to capital market users. Participants are aware that they will have this optional opportunity to provide a voluntary disclosure prior to their initial R&D investment choice. Participants are asked “How likely would you be to issue an optional explanation of your 15 decision to initiate [not initiate] the R&D project?” Participants are presented with a scale ranging from “1 – extremely unlikely to provide an optional explanation” to “8 – extremely likely to provide an optional explanation”.7 On a subsequent screen, participants are then presented with a hypothetical scenario. They are told that “regardless of your decision on the prior screen, assume that you decided to initiate the R&D project” and “also assume that you decided to issue an optional explanation of your decision”. Participants draft a voluntary disclosure in a text box. After drafting their disclosure, participants are asked to what extent they believed that they would be able to use an optional explanation to persuade their capital market users about strategies not easily seen from the numbers alone. Participants are presented with a scale ranging from “1 – did not believe at all” to “8 – believed very much”. IV. EXPERIMENT 1 – RESULTS Manipulation Checks To determine the effectiveness of my knowledge of capital market user manipulation, I ask participants, “Which type of users did your investor relations officer discover represent the vast majority of the users of your disclosures?” 95% of participants answer correctly, suggesting a successful manipulation. Hypothesis 1 My expectation for H1 is that managers are more likely to engage in increased real earnings management when they expect long-term institutional investors to rely primarily on algorithms rather than analysts to process information. Descriptive statistics are presented in 7 The phrase “optional explanation” is used in place of “voluntary disclosure” in the instrument to reduce jargon and attempt to reinforce to participants that they had a choice as to whether or not to provide an additional text response. 16 panel A of Table 1. Figure 1a presents the results graphically. [Insert Table 1] [Insert Figure 1a] Results of hypothesis tests are presented in panel B of Table 1. I use participants’ R&D investment likelihood ratings as the dependent variable and whether participants expect long- term institutional investors to rely primarily on algorithms or analysts as the independent variable. I identify support for my predicted effect of capital market users. Participants are less likely to make R&D investments in the current period when long-term institutional investors rely primarily on algorithms compared to when long-term institutional investors rely primarily on analysts (t79.53 = 2.60, p = 0.006, one-tailed). This result supports my hypothesis that managers are more likely to engage in increased real earnings management when they expect long-term institutional investors to rely primarily on algorithms rather than analysts to process information. Hypothesis 2 My expectation for H2 is that managers are less likely to provide a voluntary disclosure relating to their choice to not initiate or initiate a new R&D project when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. Descriptive statistics are presented in panel A of Table 2. Figure 1b presents the results graphically. [Insert Table 2] [Insert Figure 1b] Results of hypothesis tests are presented in panel B of Table 2. Participants are less likely to provide a voluntary disclosure relating to their choice to not initiate or initiate an R&D project when long-term institutional investors rely primarily on algorithms relative to when long-term 17 institutional investors rely primarily on analysts to process information (t77.18 = 2.18, p = 0.016, one-tailed). This result provides support for H2. V. EXPERIMENT 2 – METHOD Participants Participants in the second experiment are 100 MBA students recruited from a top-rated MBA program. On average, participants are 29 years old and have 5.62 years of work experience. 64% of participants report involvement in preparing performance reports for firm use. 79% of participants report involvement in making strategic decisions within a firm. Design and Manipulations As in Experiment 1, participants are asked to assume they are involved in making operating and financial reporting decisions at Becker International, a hypothetical firm. They are presented with background information about Becker and are informed that they will be presented with a key decision for Becker in the current period. The experiment uses a 1 x 2 between-subjects design that manipulates whether participants expect long-term institutional investors to rely primarily on algorithms or primarily on analysts. The background information and manipulations contained in Experiment 2 are the same as in Experiment 1, with minor modifications to remove information relating to the Experiment 1 operating scenario setting. Task and Procedure After being introduced to the task, participants read background information and are presented with details relevant to the accruals-based earnings management setting. Participants are informed that they will be presented with an important decision. As in Experiment 1, in a manipulation across conditions, an investor relations officer informs participants that 18 “Institutional investors differ in their reliance on algorithms versus analysts. In examining patterns of usage, your investor relations officer has obtained statistics from the company’s website and corporate servers indicating that the vast majority of downloads of SEC filings are made in file formats used by analysts [algorithms].” Next, participants are presented with a scenario that describes the key decision they must consider. Participants are informed that Becker is projected to miss its earnings targets, however Becker has a number of estimates that can be favorably adjusted in order to achieve the earnings performance that was expected. Participants are informed that they can adjust these estimates as they see fit within a reasonable range, but should be careful to try to avoid detection, as users may react poorly if they perceive opportunistic adjustments. Participants are also told that there is a possibility that capital market users can detect opportunistic adjustments even if adjustments are subtle. Next, participants are asked how likely they would be to adjust estimates to achieve the expected earnings target. After providing a likelihood response, participants are asked how likely they feel capital market users would have been able to detect their actions if they adjusted estimates. Finally, participants provide demographic information. Dependent Measures Accruals-based Earnings Management Measure Participants indicate how likely they are to adjust estimates to achieve the earnings target. Participants are presented with a scale ranging from “1 – extremely unlikely to adjust estimates” to “8 – extremely likely to adjust estimates”. The likelihood of accruals-based earnings management measure is calculated as participants’ responses to this scale. Likelihood of Detection Measure 19 Participants are also asked how likely do they feel their capital market users would have been to detect actions adjust estimates. Participants are presented with a scale ranging from “1 – very unlikely to detect” to “8 – very likely to detect”. The likelihood of detection measure is calculated as participants’ responses to this scale. VI. EXPERIMENT 2 – RESULTS Manipulation Checks To determine the effectiveness of my knowledge of capital market user manipulation, I ask participants, “Which type of users did your investor relations officer discover represent the vast majority of the users of your disclosures?” 77% of participants answer correctly, suggesting a successful manipulation. Hypothesis 3 My expectation for H3 is that managers are less likely to engage in accruals-based earnings management when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. Descriptive statistics are presented in panel A of Table 3. Figure 2a presents the results graphically. [Insert Table 3] [Insert Figure 2a] Results of hypothesis tests are presented in panel B of Table 3. I use participants’ likelihood of adjusting estimates measure as the dependent variable and whether managers expect long-term institutional investors to rely primarily on algorithms or analysts as the independent variable. I identify support for my predicted effect of capital market users. Participants are less likely to adjust estimates (and engage in accruals-based earnings 20 management) when long-term institutional investors are expected to rely primarily on algorithms relative to when long-term institutional investors are expected to rely primarily on analysts to process information (t91.54 = 2.27, p = 0.013, one-tailed). This result provides support for H3. Hypothesis 4 My expectation for H4 is that managers will feel capital market users are more likely to detect actions indicating accruals-based earnings management when they expect long-term institutional investors to rely primarily on algorithms rather than individual analysts to process information. Descriptive statistics are presented in panel A of Table 4. Figure 2b presents the results graphically. [Insert Table 4] [Insert Figure 2b] Results of hypothesis tests are presented in panel B of Table 4. I use participants’ likelihood of detection measure as the dependent variable and whether participants expect long- term institutional investors to rely primarily on algorithms or analysts as the independent variable. I do not find support for my predicted effect of capital market users. Participants express a directionally greater likelihood of detection when long-term institutional investors are expected to rely primarily on algorithms relative to when long-term institutional investors rely primarily on analysts to process information, but the difference is not significant (t97.99 = 1.23, p = 0.110, one-tailed). In considering this result, it is possible that part of the effect that leads to finding support for H3 is nonconscious. However, it is also possible that the manner in which this measure was elicited allowed for a response heuristic where a minority of participants misread scale endpoints 21 on the second of two inverted scale measures.8 As multiple explanations are plausible, this finding should be interpreted with caution. Additionally, as findings are unclear, I attempt to gather additional support in my survey of experienced managers. VII. SURVEY OF EXPERIENCED MANAGERS The two laboratory experiments provide evidence that differences in expectations of how capital market users will process information can affect managers’ operating and financial reporting decisions, as well as the information that reaches these capital market users in the first place. The purpose of my survey is to provide complementary evidence on managers’ current beliefs about the information processing capabilities of algorithms versus individual analysts, and gather insights on whether experienced managers would strategically adjust their reporting and operating behavior in response to expectations of how their audience will process information. The survey also gathers insights about how firms currently use advanced data analysis techniques to analyze performance information and managers’ expectations of whether the use of algorithms will likely increase in the future. Survey participants are 116 experienced managers recruited through a highly-rated EMBA program. On average, participants are approximately 38 years old and have spent 6.54 years with their current employer. Participants report working across a wide variety of industries in firms that, on average, have a market capitalization of greater than $1 billion. Participants also report working across a variety of job functions including management, finance, marketing/sales, 8 In line with this possibility, I observe 17 participants who responded in an endpoint-consistent manner. These participants were 55% more likely to fail manipulation checks relative to other participants. 22 and engineering/R&D. Full demographic information is presented in Table 5. [Insert Table 5] Method The survey is comprised of 8 questions and includes 3 mini-scenarios. The questions are designed to elicit (1) beliefs about perceived information processing strengths and weaknesses of algorithms and analysts, (2) beliefs about how experienced managers’ own reporting and operating behavior would be affected by expectations of how an audience will process the information, and (3) insights about future trends involving the use of algorithms within firms. Two of the three mini-scenarios are briefer within-subjects versions of my two laboratory experiments. These scenarios are designed to provide indications of how experienced managers would behave as a result of their existing beliefs, and allow me to examine whether the findings identified in my laboratory experiments would generalize to the conscious choices of more experienced managers based on their beliefs. Results Beliefs about Algorithms and Analysts Table 6 presents descriptive evidence with respect to experienced managers’ beliefs about how algorithms and individuals process information. Consistent with the assumption that motivates my two laboratory experiments, (1) 96% of experienced managers believe that human analysts are better able to understand a qualitative explanation describing a unique situation and (2) 94% of experienced managers believe that algorithms are better able to perform large-scale quantitative analysis of information. Chi-squared tests confirm that these information processing expectations differ from chance (χ²1 = 96.86, p < 0.001; χ²1 = 89.69, p < 0.001; respectively, not tabulated). In addition, these findings are consistent with prior algorithm aversion and algorithm appreciation research (Logg et al. 2019). Together with the two laboratory experiments, this 23 finding suggests that additional effects on manager behavior may already be occurring as a result of existing beliefs around the use of algorithms or analysts to process information. [Insert Table 6] Beliefs about Audience Effects on Behavior (and Scenario Responses) Table 7 presents descriptive evidence regarding experienced managers’ beliefs about how their own reporting and operating behavior would change in response to different audience information processing expectations. It also presents results from mini-scenarios. [Insert Table 7] In Panel A, experienced managers are presented with a series of five statements, and indicate their agreement or disagreement with each statement (with 0 = “strongly disagree”, 100 = “strongly agree”). Experienced managers report that anticipating how others respond to behavior is an important skill to have and develop (mean = 86.66, t115 = 31.18, p < 0.001, one- tailed, not tabulated). Experienced managers also agree that they would adjust their reporting behavior in response to expectations of how different users will process information (mean = 74.68, t115 = 12.19, p < 0.001, one-tailed, not tabulated). Further, if they believed their audience was better able to analyze information relating to certain performance metrics, experienced managers agree that they would target their operating behavior to achieve better performance on that metric (mean = 75.17, t115 = 13.18, p < 0.001, one-tailed, not tabulated). Moreover, they believe it is important to be conscious of and considerate to the specific strengths and weaknesses of others when preparing reports for them (mean = 74.81, t115 = 11.78, p < 0.001, one-tailed, not tabulated). Additionally, they disagree that it is dishonest to tailor a particular message differently for different audiences (mean = 39.48, t115 = 3.40, p < 0.001, one-tailed, not tabulated). Together, these results suggest that experienced managers strategically respond to 24 expectations of how their audience will process information, and that these strategic responses affect their reporting and operating behavior. In the mini-scenarios that are presented in Panel B and Panel C, experienced managers also behave in a manner consistent with the experimental findings. In Panel B, experienced managers are more likely to initiate an R&D project and provide an optional explanation to investors if these investors relied primarily on human analysts to process information rather than algorithms. The results of this survey response are consistent with Experiment 1, as experienced managers’ responses are significantly greater than the midpoint of the scale (t115 = 5.14, p < 0.001, one-tailed, not tabulated). Panel C shows that experienced managers are more likely to adjust estimates, in the hope of avoiding detection, when they believe investors relied primarily on analysts rather than algorithms to process information. The results of this survey response are consistent with Experiment 2, as experienced managers’ responses are significantly greater than the midpoint of the scale (t114 = 3.55, p < 0.001, one-tailed, not tabulated). 9 Trends within Firms that Affect How Performance Information is Processed Table 8 reports descriptive evidence on how experienced managers’ organizations currently process performance information, and expectations over how trends will affect these processes in the future. In Panel A, experienced managers are presented with three statements, and indicate their agreement or disagreement with each statement (with 0 = “strongly disagree”, 100 = “strongly agree”). Experienced managers report agreement that their organization currently takes advantage of “big data” or advanced data analysis techniques to perform analysis of internal performance information (mean = 54.86). Experienced managers also generally agree 9 Similarly, the third mini-scenario (not tabulated) indicates experienced managers expect automated monitoring to be more effective than manual monitoring at deterring behavior that could negatively affect their organization (t115 = 2.96, p = 0.002, one-tailed). 25 that their organizations use quantitative performance metrics for determining bonuses and promotions (mean = 62.05). Additionally, if performance metrics are lower than they should be for whatever reason, experienced managers currently feel that they will have effective opportunities to provide qualitative explanations relating to the metrics before these metrics negatively affect them (mean = 62.78). [Insert Table 8] Experienced managers also offer insights relating to future trends at their organizations. In Panel B, experienced managers are presented with a series of five statements, and indicate their expectations for the future in regards to the likelihood of each statement (with 0 = “not at all likely”, 100 = “extremely likely”). Experienced managers expect that their organizations will increase the extent to which they take advantage of “big data” or advanced data analysis techniques to perform internal analysis of performance information (mean = 69.96). Experienced managers also expect external users of their organization’s financial disclosures to increase their use of algorithms and software to analyze company information (mean = 65.37). Further, experienced managers believe that automated monitoring techniques will increase within their firm (mean = 68.23). Finally, experienced managers expect that capital investment decisions within their organizations will be held to more scrutiny by outside stakeholders (mean = 67.22), and experienced managers expect that there will be increased pressure for individuals to support and justify their proposals and decisions using hard data relative to more qualitative explanations (mean = 75.76). Together, these findings suggest that experienced managers expect trends in the institutional environment relating to the use of algorithms and technology to process information to increase in the future, and this is likely to affect perceptions of the relative information 26 processing focus on different information types. Overall, this survey provides evidence to suggest that different beliefs about the capabilities of algorithms and individuals already exist in practice, and may be affecting processes relating to both internal and external reporting and operating activities. It also suggests that behavioral effects resulting from these differences are likely to grow in importance based on the trends experienced managers have observed within their firms. VIII. CONCLUSION The growth of algorithmic investing is already reshaping capital markets. Less is known about the consequences of this transformation on manager behavior. Regulators have made salient for managers how different types of capital market users process information (SEC 1998; SEC 2009), and prior literature indicates that managers invest considerable resources into understanding their users (Brown et al. 2019). This study investigates whether perceived differences in how capital market users will process financial disclosures can affect the very financial disclosures that reach users in the first place. In this study, I use two controlled experiments to demonstrate that expectations over how capital market users will process information affect whether managers use voluntary disclosures, engage in real earnings management, and engage in accruals-based earnings management. In a first experiment, I find that participants are more likely to not initiate an R&D investment project (and engage in increased real earnings management) when they expect long-term institutional investors to rely primarily on algorithms rather than analysts to process information. Consistent with theory regarding the process, participants choose to provide fewer narrative disclosures when they expect long-term institutional investors to rely primarily on algorithms, and are less 27 confident in the persuasiveness of the narrative disclosures that they do provide. In a second experiment, I find that participants are less likely to adjust estimates (and engage in accruals-based earnings management) when they expect long-term institutional investors to rely primarily on algorithms rather than analysts to process information. This may result from managers feeling capital market users are more likely to detect their actions when they adjust estimates when long-term institutional investors rely primarily on algorithms rather than analysts. Results in both experiments are consistent with theory on manager expectations over information processing characteristics affecting manager behavior, as well as prior literature on algorithm aversion and algorithm appreciation (Logg et al. 2019). Results in Experiment 1 are also consistent with theories relating to audience effects and persuasion (Fussell and Krauss 1992) and results in Experiment 2 are consistent with General Deterrence Theory (D’Arcy et al. 2009; Nagin and Pogarsky 2001). Together, the two experiments suggest that expectations over algorithmic investing can decrease accruals-based earnings management through one mechanism, and can increase real earnings management through another. In a complementary survey of experienced managers, I find that current beliefs about algorithms and individuals are consistent with my experimental assumptions. I also find that a majority of survey participants report that they would adjust their own reporting and operating behavior in response to expectations of how their audience will process the information. These survey participants’ existing beliefs lead them to behave in a manner consistent with the behavior of participants in the laboratory experiments. Finally, survey participants provide their opinions about future trends involving the use of algorithms within their firms. These insights suggest that the findings demonstrated in the two laboratory experiments will grow in importance as both firms and capital market users increase their use of algorithms to analyze performance 28 information. These findings offer a number of contributions. First, they add to the literature around capital market user ease of processing emphasis in terms of both “readability” and “scriptability”. While most of this literature focuses on direct effects on investors as they process information, (Allee et al. 2018; Blankespoor et al. 2019; Bloomfield 2008; Li 2008; Rennekamp 2012), this study investigates how manager judgments and decisions can be affected by competing processing emphases. Results suggest that these different emphases may impact manager psychology in ways which have not been fully unexplored. This study extends prior literature on feedback effects (Blankespoor et al. 2019), and leverages the comparative advantages of experimental and survey methods to obtain strong causal inferences and insights into managers’ decision making in response to expected information processing differences. Second, this study adds to literature on real earnings management. It adds to existing work on factors which affect the utilization of different earnings management techniques (Chan et al. 2015; Gunny 2010; Zang 2012), and suggests that managers can preemptively engage in real earnings management based on expectations of how users will process information, rather than through formal interactions. This study also emphasizes the importance of the expected persuasiveness of narrative disclosures to the decisions that managers make. Prior literature suggests that the absence of narrative disclosures can lead to measure management (Bentley 2019), and this study suggests that even if managers have the ability to provide disclosures, factors which affect how persuasive such disclosures are expected to be can impact whether they are utilized. This study also adds to literature on accruals-based earnings management. Prior literature suggests that earnings management may be widespread and that managers may engage in it 29 strategically. Prior literature also explores how manager psychology can affect the nature and extent of earnings management. This study suggests that a shift in expectations of how capital market users process information may affect the ways in which managers engage in earnings management. It suggests that managers strategically respond to expectations over information processing costs, and that these responses can affect both operating and reporting outcomes. This potential shift in manager behavior could have important consequences going forward. These findings also offer contributions to practice. Existing guidance highlights how different types of capital market users can process information differently. As differences in how capital market users process information become more salient, and more precise data becomes available to managers on what users are doing with information, managers may increasingly shape their behavior around perceived expectations of how their users will process information. Ultimately, this could affect manager psychology around how disclosures are issued and how key operating decisions are made. These unintended consequences could have meaningful effects that regulators may wish to consider, particularly as algorithmic investing continues to grow in importance. My study is subject to certain limitations. One limitation is that I am able to able to document existing beliefs about the strengths and weaknesses of algorithms and analysts for information processing, yet it is possible that beliefs may change in the future. Additionally, while I am able to obtain strong causal inferences within a controlled laboratory setting, it is possible that the effects observed within this setting will not generalize to all settings involving real or accruals-based earnings management. My survey findings are also subject to standard limitations, where managers may not have full self-insight into their own behavior and may have self-presentation considerations. 30 My study also provides many opportunities for future research. Future research may consider investigating whether different expectations over how capital market users process information can affect manager behavior in other areas than earnings management, narrative disclosures, or investment decisions. Future research can also consider exploring other systematic differences that may exist across how different institutional investors process information. Future research may also consider investigating whether other constructs of interest might moderate the effects I observe. Finally, my survey evidence suggests that existing practices and trends indicate that expectations over how users will process information can strongly impact both internal and external reporting, and future research may consider exploring the impact of these effects in internal processes in more depth. 31 Figure 1a – Likelihood of Engaging in Real Earnings Management (E1) Likelihood of Engaging in Real Earnings Management 8 7 6 5 4 3.28 3 2.58 2 1 Investors Rely Primarily on Analysts Investors Rely Primarily on Algorithms This figure graphically presents participants’ likelihood of engaging in real earnings management (which is equal to 9 – likelihood of initiating the R&D investment project). Figure 1b – Likelihood of Providing Voluntary Disclosure (E1) Likelihood of Providing Voluntary Disclosure 8 7.26 7 6.44 6 5 4 3 2 1 Investors Rely Primarily on Analysts Investors Rely Primarily on Algorithms This figure graphically presents participants’ likelihood of providing a voluntary disclosure. 32 Figure 2a – Likelihood of Engaging in Accruals-based Earnings Management (E2) Likelihood of Engaging in Accruals-based Earnings Management 8 7 6 5 3.8 4 2.94 3 2 1 Investors Rely Primarily on Analysts Investors Rely Primarily on Algorithms This figure graphically presents participants’ likelihood of adjusting estimates (and engaging in accruals-based earnings management). Figure 2b – Expressed Likelihood of Detection (E2) Expressed Likelihood of Detection 8 7 6 5.6 5.2 5 4 3 2 1 Investors Rely Primarily on Analysts Investors Rely Primarily on Algorithms This figure graphically presents participants’ expressed likelihood of detection of accruals-based earnings management actions. 33 Table 1 – Likelihood of Initiating the R&D Project TABLE 1 Panel A: Descriptive statistics: Mean and (standard deviation) Investors Rely Investors Rely Cho ice Primarily o n Analysts Primarily on Algorithms Ove rall Likelihood of 6.42 5.72 6.07 Initiating the (0.97) (1.64) (1.38) R&D Project n=50 n=50 n=100 Panel B: Welch’s t-test for capital market users Comparisons df t-statistic p-value Investors Rely Primarily on Analysts Versus Investors Rely Primarily on Algorithms 79.53 2.60 0.006* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to their likelihood to initiate the R&D project. All participants were provided with background information about the firm and the setting. A manipulation occurs over whether participants believe their long-term institutional investors will rely primarily on algorithms or analysts in analyzing financial disclosures. Participants are asked “How likely would you be to not initiate or initiate the new R&D project?” Participants are presented with a scale ranging from “1 – extremely unlikely to initiate the project” to “8 – extremely likely to initiate the project”. *One-tailed. 34 Table 2 – Likelihood of Providing Voluntary Disclosure TABLE 2 Panel A: Descriptive statistics: Mean and (standard deviation) Investors Rely Investors Rely Cho ice Primarily o n Analysts Primarily on Algorithms Ove rall Likelihood of 7.26 6.44 6.85 Providing (1.31) (2.32) (1.91) Voluntary Disclosure n=50 n=50 n=100 Panel B: Welch’s t-test for capital market users Comparisons df t-statistic p-value Investors Rely Primarily on Analysts Versus Investors Rely Primarily on Algorithms 77.18 2.18 0.016* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to their likelihood to provide a voluntary disclosure of their decision. All participants were provided with background information about the firm and the setting. Participants are aware that they would have an opportunity to provide a voluntary disclosure prior to their initial investment choice. A manipulation occurs over whether participants believe their long-term institutional investors will rely primarily on algorithms or analysts in analyzing financial disclosures. Participants are asked “How likely would you be to issue an optional explanation of your decision to initiate [not initiate] the R&D project?” Participants are presented with a scale ranging from “1 – extremely unlikely to provide an optional explanation” to “8 – extremely likely to provide an optional explanation”. *One-tailed. 35 Table 3 – Likelihood of Engaging in Accruals-based Earnings Management TABLE 3 Panel A: Descriptive statistics: Mean and (standard deviation) Investors Rely Investors Rely Cho ice Primarily o n Analysts Primarily on Algorithms Ove rall Likelihood of 3.80 2.94 3.37 Engaging in (2.11) (1.61) (1.92) Accruals-based Earnings n=50 n=50 n=100 Management Panel B: Welch’s t-test for capital market users Comparisons df t-statistic p-value Investors Rely Primarily on Analysts Versus Investors Rely Primarily on Algorithms 91.54 2.27 0.013* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to their likelihood of engaging in accruals-based earnings management. All participants were provided with background information about the firm and the setting. A manipulation occurs over whether participants believe their long-term institutional investors will rely primarily on algorithms or analysts in analyzing financial disclosures. Participants indicate how likely they are to adjust estimates to achieve the earnings target. Participants are presented with a scale ranging from “1 – extremely unlikely to adjust estimates” to “8 – extremely likely to adjust estimates”. *One-tailed. 36 Table 4 – Expressed Likelihood of Detection TABLE 4 Panel A: Descriptive statistics: Mean and (standard deviation) Investors Rely Investors Rely Primarily Cho ice Primarily o n Analysts on Algo rithms Ove rall Expressed 5.20 5.60 5.40 Likelihood of (1.60) (1.61) (1.62) Detection n=50 n=50 n=100 Panel B: Welch’s t-test for capital market users Comparisons df t-statistic p-value Investors Rely Primarily on Analysts Versus Investors Rely Primarily on Algorithms 97.99 1.23 0.110* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to how likely they feel their capital market users would have been to detect their actions if they engaged in accruals-based earnings management. All participants were provided with background information about the firm and the setting. A manipulation occurs over whether participants believe their long-term institutional investors will rely primarily on analysts or algorithms in analyzing financial disclosures. Participants are asked how likely do they feel their capital market users would have been to detect actions adjust estimates. Participants are presented with a scale ranging from “1 – very unlikely to detect” to “8 – very likely to detect”. *One-tailed. 37 Table 5 – Demographic Characteristics of Survey Respondents TABLE 5 Primary Industry % Years with Current Employer % Consumer Discretionary 4.31 <1 year 1.78 Consumer Staples 4.31 1-3 years 38.39 Energy 16.38 4-6 years 16.07 Financials 18.10 7-9 years 17.86 Health Care 7.76 10+ years 25.90 Industrials 6.90 Information Technology 13.79 Company market capitalization Materials 2.59 <$100 million 23.28 Other 25.86 $100 million - $249 million 7.75 $250 million - $499 million 10.34 Age $500 million - $999 million 2.59 <30 0.86 $1 billion - $10 billion 19.83 30-39 70.69 >$10 billion 36.21 40-49 26.72 50-59 1.73 Primary Job Function Engineering / R&D 17.24 Gender Finance 11.21 Male 73.28 Management 40.52 Female 26.72 Marketing / Sales 12.07 Other 18.96 This table presents demographic information for a survey conducted of 116 experienced managers recruited from a highly rated Executive MBA program. 38 Table 6 – Survey responses to Question 1 “Performance information can be evaluated by different types of users including (1) human analysts and (2) algorithms or software. Which of these users would you say is better at performing each of the following tasks?” TABLE 6 Human Analysts Algorithms Understanding a qualitative explanation describing a 111 5 unique situation Performing large-scale quantitative analysis of 7 109 information This table presents survey responses to the question: “Performance information can be evaluated by different types of users including (1) human analysts and (2) algorithms or software. Which of these users would you say is better at performing each of the following tasks?” The question is presented as a binary choice and data is recorded as the count of the number of participants selecting each response. 39 Table 7 – Survey responses to Questions 2, 3, 4, and 8 TABLE 7 Panel A – Question 2 Please rate your level of agreement with the following statements: Average ( 0 – Strongly disagree, 100 – Strongly agree) Anticipating how others will likely respond to certain information is an important skill to have and develop. 86.66 You would adjust your reporting behavior in response to expectations of how different users would process information. 74.68 If you believed your audience was better able to analyze information relating to certain performance metrics, you would target your performance to be better on these metrics. 75.17 It's important to be conscious of and considerate to the specific strengths and weaknesses of others when preparing reports for them. 74.81 It is dishonest to tailor a particular message differently for different audiences. 39.48 Panel B – Question 3 (1 – Much more likely…if investors relied primarily on algorithms to process information, 8 – Much more likely…if investors relied primarily on analysts to process information) Average Rating Imagine the following scenario. You are in charge of a firm that is considering initiating a new R&D project. Your firm will report earnings to the market soon. Initiating the R&D project will result in increases in immediate expenses and lower current earnings. However, initiating the project will likely result in even greater increases in future earnings that you can try to explain to investors in an optional explanation accompanying the current financial statements. Investors can rely upon analysts or algorithms to process the information 5.47 that firms disclose. If you feel you can persuade investors as to the R&D project's merits with a narrative explanation, they may appreciate your decision to initiate the project. Otherwise, if you don't feel you can persuade the investors, you may be punished for lower earnings in the current period. Would you be more likely to initiate the new R&D project and try to explain the decision to investors with a disclosure if investors relied primarily on (1) algorithms or (2) analysts to process the information that a firm discloses? 40 Panel C – Question 4 (1 – Much more likely…if investors relied primarily on algorithms to process information, 8 – Much more likely…if investors relied primarily on analysts to process information) Average Rating Imagine the following scenario. You are in charge of a firm that will report earnings to the market soon. Current earnings are expected to fall short of expectations. You can report this disappointing performance to the market as is, but you can also adjust estimates to make performance appear more favorable. However, if you are to adjust estimates you must consider how likely your capital market users would be to detect your actions to adjust estimates and plan accordingly. Investors can rely upon analysts or algorithms to process the information 5.14 that firms disclose. If adjustments towards more favorable estimates are not detected, capital market users will react positively to increased reported earnings in the current period. If they are detected, capital market users will react more negatively than if you simply reported numbers that missed current earnings targets. Adjustments are only likely to be detected with advanced quantitative analysis. Would you be more likely to adjust estimates in the hopes that you can avoid detection if investors relied primarily on (1) algorithms or (2) analysts to process the information that a firm discloses? This table presents survey responses to the questions 2, 3, 4, and 8. Question 2 asks participants about whether they believe their own reporting and operating behavior would be affected by audience expectations. Question 3 presents participants with a mini-scenario analogous to Experiment 1. A higher rating indicates that participants would be much more likely to initiate the R&D project (and avoid engaging in real earnings management) if investors relied primarily on analysts rather than algorithms to process information. Question 4 presents participants with a mini- scenario analogous to Experiment 2. A higher rating indicates that participants would be much more likely to adjust estimates and engage in accruals-based earnings management if investors relied primarily on analysts rather than algorithms to process information. Question 8, not tabulated, presents participants with a separate mini-scenario relating to detection risk where participants are asked whether they believe automated monitoring or manual monitoring would be more effective at deterring undesirable behavior that could increase the risk profile of a firm. Participants’ average response indicates that they believe automated monitoring would be more effective. 41 Table 8 – Survey responses to Questions 5 and 6 TABLE 8 Panel A – Question 5 Please rate your level of agreement with the following statements: Average ( 0 – Strongly disagree, 100 – Strongly agree) Your firm currently takes advantage of "big data" or advanced data analysis techniques to perform internal analysis of performance information. 54.86 In considering promotions and bonus compensation, your firm uses performance evaluations that incorporate quantitative metrics. 62.05 At your firm, if your performance metrics are lower than they should be for whatever reason, you will have an opportunity to explain any extenuating circumstances that might have caused this before the metrics negatively affect you. 62.78 Panel B – Question 6 Please rate your expectations for the future in regards to the likelihood of the Average f ollowing items: (0 – Not at all likely, 100 – Extremely likely) Your firm will increase the extent to which they take advantage of "big data" or advanced data analysis techniques to perform internal analysis of performance information. 69.96 External users of your firm's financial statements and financial disclosures will increase their use of algorithms and other software to analyze company information. 65.37 Automated monitoring techniques will increase as your firm fine-tunes its processes. 68.23 Capital investment decisions will be held to more scrutiny by outside stakeholders who will use expanded data analysis capabilities to evaluate information. 67.22 Because of increased availability of data and data analysis, there will be more pressure for individuals to support and justify their proposals and decisions using hard data relative to more qualitative explanations. 75.76 This table presents survey responses to the questions 5 and 6. Question 5 polls participants about current firm practices relevant to use of advanced data analysis techniques and quantitative metrics in performance evaluations. Question 6 asks participants about expectations over future trends which might be indicative as to whether the issues explored in this study are expected to grow in importance in the future. Additionally, participants are allowed to elaborate on their responses to Question 6 in free response form in Question 7, which is not tabulated. 42 Appendix A – Information about LT Institutional Investors – Manipulation and Assumed Beliefs (that are held constant) 43 CHAPTER 2 – Egocentrism, Human versus Algorithmic Information Processing, and Selection of Disclosure Medium ABSTRACT Recent literature demonstrates that expectations of how investors will process information can affect the information that reaches investors in the first place. Yet it is unclear whether the formation of managers’ expectations is entirely rational, or whether it can be systematically biased by certain cognitive factors. In two laboratory experiments and a survey, this study examines how one systematic bias can form, and can cause unintentional distortion in managers’ selection of disclosure mediums by which to release information. Results from the first experiment indicate participants’ utilization of more sensory (video) disclosure mediums is reduced in response to expectations of algorithm-based information processing. Yet results from the second experiment indicate that participants’ trust in assessments of sensory information in video disclosure mediums is greater in response to algorithm-based information processing. Participants thus behave in an inconsistent manner when their perspective is flipped from issuing disclosures to processing disclosures. This finding is consistent with an egocentric focus in perspective taking bias that extends prior psychology literature on the Spotlight Effect and the Illusion of Transparency. This bias can be represented through a two-stage psychological mechanism involving (1) the formation of a setting-specific egocentric default perspective and (2) insufficient adjustment away from this perspective for more similar vs. less similar others. This study contributes to existing literature on financial disclosure, and introduces the role of cognitive biases in driving certain feedback effects being explored in the emerging literature. 44 I. INTRODUCTION Firms are increasingly attentive to how investors process information, and managers’ reporting and operating decisions are being affected by expectations of investors’ information processing costs and capabilities. An emerging literature has begun to explore these “feedback effects” (Blankespoor, deHaan, and Marinovic 2019) and has demonstrated that they can influence a variety of reporting and operating decisions in consequential ways (Blankespoor 2019; Basu, Pierce, and Stephan 2019; Witz Chapter 1). Yet it is unclear in this literature whether the formation of managers’ expectations is entirely rational, or whether it can be systematically biased by certain cognitive factors. In this study, I examine how one systematic cognitive bias can form, and distort managers’ reporting behavior. The presence of this bias implies a potential inefficiency in managers’ utilization of different disclosure mediums. In two experiments, I demonstrate participants form inconsistent beliefs about the information processing strengths and weaknesses of individuals vs. algorithms depending on whether they are issuing or processing disclosures. These beliefs then influence how participants behave in selecting disclosure mediums and in relying on investment advice. In Experiment 1, participants assume the role of managers and decide whether to use a video- or text-based disclosure medium to issue a disclosure. In Experiment 2, participants are tasked with analyzing a video-based disclosure in the presence of information processing assistance. A survey of experienced managers adds context to the experimental findings and documents current institutional beliefs about information processing strengths and weaknesses of different parties across different disclosure mediums. In Experiment 1, participants are given background information about their firm, and are presented with a scenario where they have a choice as to how to communicate information to 45 investors. They indicate how likely they are to issue a video or text disclosure. I manipulate whether (1) participants are feeling anxious or confident in communicating information and (2) whether participants expect investors’ initial stage of disclosure processing to be performed by individuals or algorithms. A significant interaction emerges where participants are more likely to adjust their use of video disclosure mediums (in response to their internal emotional state) when individuals are expected to process their disclosures rather than algorithms. This finding is consistent with the idea that participants in the first experiment believe investors are more likely to be able to discern their internal state through verbal and non-verbal cues contained in video disclosures when investors process disclosures using individual analysts rather than algorithms. However, in Experiment 2, participants are alternatively tasked with analyzing a video disclosure. They are asked to discern whether the CEO presenting the disclosure is concealing information, and they receive additional advice in the form of a summary assessment that claims to detect (or not detect) signs of information concealment in the verbal and non-verbal behavior of the CEO. I manipulate whether this summary assessment is received from another individual or from an algorithm. In this setting, participants alternatively place greater trust in the algorithm-based advice in detecting (or not detecting) signs of information concealment in the verbal and non-verbal behavior of the individual presenting the disclosure. This result is inconsistent with the result from Experiment 1, as participants’ beliefs about the relative effectiveness of algorithms at processing verbal and non-verbal behavioral cues flip between the two settings. However, it occurs due to an egocentric focus in perspective taking bias. The psychological mechanism underlying this bias occurs in two stages. In one stage, participants first strongly experience and focus on their own egocentric perspective. Participants experience different firsthand perspectives in the two experiments. When communicating 46 information in Experiment 1, participants strongly feel their own internal emotions and believe others may be able to detect these emotions as well (more so than they actually can).10 When analyzing information in Experiment 2, however, participants alternatively experience and focus on their own difficulty in detecting cues from the verbal and non-verbal behavior of others. In the second stage, participants form judgments about the information processing abilities of others by adjusting from their own egocentric perspective. Insufficient adjustments from egocentric defaults are a well-known phenomenon in psychology, and literature suggests that individuals retain their egocentric default more when taking the perspective of more similar vs. less similar others (Epley 2008). In my setting, this results in smaller adjustments when participants consider the perspective of more similar individuals versus less similar algorithms. Thus, the setting-specific inconsistency in participants’ beliefs about the relative information processing abilities of individuals versus algorithms demonstrates cognitive bias in both settings, or at least one, as the experimental design flips participants’ perspectives while holding other information constant. Further, as the manager disclosure medium choice is my primary setting of interest, I employ process measures to confirm the presence of the predicted cognitive bias within this setting. In addition to direct measures, I utilize linguistic analysis to obtain an unobtrusive measure of residual egocentrism immediately following participants’ disclosure medium choice. Finally, in the survey of experienced managers, I capture current beliefs about the relative information processing strengths and weaknesses of algorithms and individual analysts. These beliefs are broadly consistent with the Experiment 1 findings, and demonstrate that the disclosure medium effects observed in the laboratory environment may also emerge out of sample as a 10 This is consistent with the Spotlight Effect (Gilovich and Savitsky 1999; Gilovich, Savitsky, and Medvec 2000) and the Illusion of Transparency (Gilovich, Savitsky, and Medvec 1998) in the psychology literature. 47 result of factors that lead to these beliefs in the institutional environment, such as the egocentric reporting perspective mechanism observed in Experiment 1. These findings have important implications for managers given ongoing changes to the capital markets. First, this study adds to the literature on disclosure mediums. Present-day firms have an expanding number of choices in how they issue disclosures (Libby and Emett 2014), and these choices can affect the information content that reaches the markets (Blankespoor, Hendricks, and Miller 2017; Mayew and Venkatachalam 2012). These choices can also affect how investors process information (Elliott, Hodge, and Sedor 2012). While prior literature examines how different disclosure mediums affect investors decision-making, managers’ choices of specific disclosure mediums are likely to be strategic and nonrandom. This study builds upon prior literature by demonstrating reasons why managers can elect to use one medium versus another. However, it also indicates that this behavior may be influenced by biases in the preceding formation of information processing expectations. The consequences of these effects may be increasingly impactful as algorithmic investing continues to grow in prominence (Cantrell 2017), and may prevent investors from taking full advantage of their increasing capabilities to analyze the verbal and non-verbal behavior of managers. Second, this study directly contributes to the emerging literature on feedback effects (Blankespoor et al. 2019), or how managers’ expectations of investors’ information processing costs can affect managers’ behavior. Prior literature demonstrates that managers’ responses to expectations of investors’ information processing costs and capabilities can be consequential (Blankespoor 2019; Basu et al. 2019; Witz Chapter 1). However, this literature has not yet deeply explored how managers’ expectations form. This may be an important evolution in this literature, as the mechanisms by which managers form expectations may affect how their 48 responses to these expectations are interpreted. This study offers a novel contribution to this literature by being the first to demonstrate that cognitive biases might distort manager behavior in consequential ways. Managers are likely unaware that their beliefs can be affected by systematic cognitive biases, and this study demonstrates that even strategic responses to these expectations may be misguided based on the way in which expectations form. These findings also offer contributions to practice and the existing stream of literature that suggests managers’ anticipation of investors information processing costs can affect their reporting and operating behavior. Findings in this literature suggest that managers may anticipate limits to investor attention when they disclose information, and adjust their behavior accordingly (Clor-Proell and Maines 2014). However, managers are unlikely to be aware of the specific cognitive bias explored in this study, and knowledge of this bias may affect their disclosure choices. This study also provides evidence that managers may opportunistically choose to utilize different disclosure mediums, and this knowledge may be informative as regulators and standard-setters consider how to legislate for the expanding disclosure landscape (Miller and Skinner 2015). Together, the two experiments and survey offer perspective into how managers may respond to investors’ increasing use of algorithms to analyze performance information, and call attention to the potential distortion that can occur in the information environment as a result of biases affecting managers’ disclosure behavior. The remainder of this paper is organized as follows. Section II discusses the theories used to develop the experiments and contains my specific hypotheses. Sections III and IV discuss the method and results, respectively, for Experiment 1. Sections V and VI discuss the method and results, respectively, for Experiment 2. Section VII discusses the survey of experienced managers. Section VIII concludes the paper. 49 II. BACKGROUND AND HYPOTHESIS DEVELOPMENT In this study, I offer a first look into the potential for information-processing-based feedback effects to be subject to cognitive bias in a systematic way. Research has just begun to explore feedback effects (Blankespoor et al. 2019), but early indications are that these effects can be highly consequential and widespread, while also difficult to observe (Blankespoor 2019; Basu et al. 2019; Witz Chapter 1). Feedback Effects – Managers Responses to Specific Information Processing Expectations The investor relations literature documents that firms are highly attentive to who the users of their financial statements are, and how these users process information (Brown et al. 2019; Bushee and Miller 2012; Trentmann 2019).11 More recent literature has started to causally explore how specific expectations of investors’ information processing costs can impact managers’ decisions. These effects are relatively challenging to explore, as it is difficult to isolate managers’ strategic intent and discretionary responses to only managers’ expectations of investors’ information processing costs when broadly viewing disclosure and operating outcomes (Blankespoor et al. 2019). However early research has indicated that managers’ direct responses to information processing expectations can be highly impactful (Blankespoor 2019; Basu et al. 2019; Witz Chapter 1). and as such, understanding the mechanisms that drive managers’ decision-making (and formation of information processing expectations) can be helpful. Egocentric Focus and Adjustment – The Spotlight Effect and the Illusion of Transparency 11 Prior work suggests that the choices managers make when providing disclosures can affect how costly disclosures are for investors to process (Allee, DeAngelis, and Moon 2018; Li 2008). Managers are presumably aware of the impact of their actions on how costly it is for investors to process information, as they often incur significant effort and expense to enable investors to more effectively process and extract information (Asay, Rennekamp, and Libby 2018b; Brown et al. 2019; Kirk and Markov 2016; Solomon and Soltes 2015). Correspondingly, prior literature also suggests that managers’ implicitly assume limits to investor attention, and demonstrate this assumption through a variety of reporting and operating behavior (Clor-Proell and Maines 2014; Graham et al. 2005; Libby and Emett 2014). 50 This study explores the role of one particular cognitive bias that can systematically influence the formation of managers’ information processing expectations. This bias is likely to be universal across all individuals, as the mechanism that underlies it is fundamental to human psychology. Moreover, it is generally believed to be nonconscious in such a way that most individuals likely have no idea they are subject to the bias. In psychology, the Spotlight Effect suggests that people overestimate how much others notice their actions (Gilovich and Savitsky 1999; Gilovich, Kruger, and Medvec 2002), and the Illusion of Transparency suggests that people tend to overestimate the extent to which others can discern their internal states (Gilovich, Savitsky, and Medvec 1998). In both of these findings, people do notice the actions of the presenter and can discern some information about their internal state, but not nearly as much information as the presenter expects them to. The mechanism underlying these effects can be represented in two stages. First, an egocentric focus occurs at the root of both effects, as individuals strongly feel and experience their own lived actions and emotions (Gilovich et al. 2002). This finding is consistent with a robust literature in psychology, including the longstanding literature on egocentrism (Dunning, Alicke, and Krueger 2005; Piaget 1932) and related literature on self-focus (Fenigstein and Abrams 1993; Gendolla and Wicklund 2009). In a financial reporting setting, this stage of the effect would manifest through managers strongly experiencing their own emotions as they prepare to communicate information to others. In a second stage, individuals then attempt to adjust away from this egocentric default when they consider the perspectives of others. Literature on perspective-taking suggests that people start from their egocentric default and subsequently adjust to accommodate differences 51 between themselves and others (Epley 2008).12 As a result, people rely on their egocentric defaults more when reasoning about others who they perceive as more similar to themselves (Clement and Krueger 2002; Mitchell, Macrae, and Banaji 2006; Robbins and Kruger 2005). In the accounting institutional environment, this is consequential as recent changes have shifted whether financial disclosures are expected to be processed by relatively more similar vs. relatively less similar others, especially at the initial stage of information processing. In the past, institutional investors relied primarily on individual analysts to process the majority of disclosures, however recent technological advances have enabled firms to increasingly rely upon algorithms (Allee, DeAngelis, and Moon 2018; Cantrell 2017).13 Choice of Disclosure Medium Expectations of how investors will process information can influence manager decision- making in a variety of contexts. One important context that has not yet been explored is in the choice of disclosure medium. Firms and managers can choose to issue disclosures in many different mediums, and firms have increasingly started to utilize new disclosure mediums to convey information (Elliott, Hodge, and Sedor 2012; Kirk and Markov 2016; Rennekamp and Witz 2018). Libby and Emett (2014) observe that the medium by which firms issue disclosures serves as a narrative attribute which can affect how investors process information as well as the nature of information that reaches investors. Managers’ choices of disclosure mediums are increasingly relevant as certain mediums allow investors’ greater insight into particular 12 Epley (2008) suggests that attentional and construal biases often leave individuals with very little or no conscious awareness of the ways in which their perception is influenced and constructed from their own egocentric default. This suggests the egocentric focus and (incomplete) adjustment mechanisms away from this egocentric default are very difficult to become aware of and to correct. 13 The notion that algorithms are perceived as having greater differences from individuals than other individuals is intuitive, and is also supported by work in related literatures. In the trust literature, more similar others are seen as more trustworthy (Farmer, McKay, and Tsakiris 2014). Correspondingly, in the trust in automation literature, findings suggest that anthropomorphizing automated (algorithm-based) systems can enhance trust in these systems (de Visser et al. 2012; Pak et al. 2012). 52 qualitative information, such as the verbal and non-verbal behavior of managers as they are issuing disclosures. This verbal and non-verbal behavior can be incorporated as valuable inputs to investors’ decision-making process (Mayew and Venkatachalam 2012; Blankespoor, Hendricks, and Miller 2017). While prior literature has looked at effects that occur on investor decision-making when disclosure mediums are independently manipulated, the choice of disclosure medium itself can be highly endogenous. Different mediums affect how much certain types of information are revealed, and they also affect the framing of information and how it is associated with the individuals who are conveying it (Asay, Libby, and Rennekamp 2018a). Prior literature in psychology suggests that individuals are very conscious of revealing verbal and non-verbal information as they are communicating (Gilovich et al. 2002). As a result, they may adjust how they communicate when they believe it is beneficial or detrimental to give off this information to their anticipated audience. Video disclosures may be used more when managers wish to reveal a more confident internal state, but less when managers wish to conceal a more anxious internal state. This is especially likely to be the case when managers feel investors will be able to detect their internal state. With the egocentric focus in perspective taking mechanism, I predict managers will feel this is more likely when investors are expected to process disclosures using relatively more similar individuals, rather than using relatively less similar algorithms. Stated formally, I hypothesize that: H1a: When managers wish to reveal (versus conceal) their internal emotional state, they will be more likely to provide video disclosures when they expect investors to process information using individuals rather than algorithms. H1b: Managers’ adjustments will be driven by whether they wish to reveal or conceal verbal and non-verbal information cues, and whether they think their investors will be able to detect these cues based on how they process information. 53 Flipping Perspective – A Bias is Revealed However, this behavior may be based on biased expectations due to the nature of the egocentric focus in perspective taking mechanism. When communicating information, individuals strongly feel their own lived (or recalled) experiences and emotions, and then assume others can easily observe these signals as well based on an incomplete adjustment from their egocentric default. Yet the Illusion of Transparency (Gilovich et al. 1998) suggests that others almost always have more difficulty detecting this information than expected. When these same individuals are alternatively attempting to analyze the verbal and non- verbal behavior of others, they may realize that this is a difficult task. They may then egocentrically focus on their own difficulty in performing this task, and incompletely adjust from this alternative perspective when evaluating other cognitively similar individuals. This may cause them to perceive that similar individuals will struggle to effectively analyze this information as well. However, they may more completely adjust for the perspectives of relatively less similar algorithms. In this alternative setting, I predict individuals to rely more on algorithm- provided advice than individual-provided advice when seeking to evaluate the verbal and non- verbal behavior of a CEO issuing disclosures. Stated formally, I hypothesize that: H2: Individuals are more likely to trust algorithm-provided advice compared to individual-provided advice when they are offered assistance in evaluating the verbal and non-verbal behavior of others. Together, the two experiments thus reveal an egocentric focus in perspective taking bias that results in inconsistent information processing expectations depending on whether participants are communicating or analyzing verbal and non-verbal cues of qualitative information. 54 III. EXPERIMENT 1 – METHOD Participants Participants in the first experiment are 110 MBA students recruited from a large northeastern university. On average, participants are 29.33 years old and have 5.55 years of work experience. 74.54% of participants report involvement in preparing performance reports for firm use. 81.82% of participants report involvement in making strategic decisions within a firm. Design Choices The two experiments were designed together to demonstrate the existence of an egocentric focus in perspective taking bias in regards to how expectations of information processing costs and capabilities are formed. The design of each experiment is intended to mirror the setting of the other experiment as closely as possible. This ensures that it is the predicted psychological mechanism, rather than other factors, that results in the formation of inconsistent and biased expectations. Investors in both experiments are described as sophisticated institutional investors, and both settings relate to potential information concealment that may or may not be revealed by verbal and non-verbal information cues. Process measures also directly capture expectations about the capabilities of individuals vs. algorithms in detecting verbal and non- verbal indicators in video disclosures. The inclusion of unobtrusive process measures in Experiment 1 provides further support for the predicted psychological bias within this setting. Task and Manipulation Participants in the first experiment are told that they are the CEO of Becker International, a hypothetical firm. They are presented with background information about Becker and are informed that they are preparing to issue a press release discussing performance in the most 55 recent financial period. The experiment uses a 2 x 2 between-subjects design that manipulates (1) whether participants believe their verbal and non-verbal behavior in communicating information will reflect anxiety or confidence, and (2) whether participants believe investors will process their disclosures using individuals or using algorithms. Setting and Procedure After being introduced to the setting and reading background information, participants prepare to make a decision about whether they would choose to issue the press release in a video or text format. In the first manipulation across conditions, participants are told that they are feeling anxious or confident about issuing the disclosure, and are asked to recall a time in their life when they were anxious or confident while communicating information. In the second manipulation, participants are told that their investor relations officer has discovered that investors perform the initial stage of processing their disclosures using individuals or using algorithms. Held constant across conditions, investors are described as sophisticated institutional investors who hold long-term investments in the company. Once participants receive the setting information, and the manipulations, participants provide an assessment of whether they would be likely to provide their disclosure in a video or text format. Immediately following this decision, participants provide a free response explaining, in their own words, why they made the decision that they did. Participants also provide judgments as to whether they think it would be beneficial or detrimental for investors to process verbal and non-verbal cues of their internal state, as well as judgments of how likely they believe investors would be able to detect these cues based on their process. Dependent Measures Likelihood of Video or Text Disclosure Measure 56 In my setting, the main dependent variable is how likely participants would be to choose to provide the press release in a text or video format. Participants are presented with a scale ranging from “1 – Extremely likely to choose text” to “8 – Extremely likely to choose video”. Free Response Process Measure Participants are also asked to provide the thinking that led to their decision. This is an unobtrusive free response measure that documents participants’ own stated reasons for choosing to provide the text or video disclosure to investors. This measure can also evaluate the predicted psychological mechanism of egocentric focus and insufficient adjustment, independent of the more direct process measures listed below. Likelihood and Direction of Processing Cues Measures Finally, participants are asked a series of three questions that directly elicit participant beliefs about their thinking as it relates to investors’ processing of verbal and non-verbal information cues. Participants are first asked “Do you think a text or video disclosure is likely to provide more verbal and non-verbal cues of your internal mental state?” Participants are presented with a scale ranging from “1 – Text disclosures much more likely to provide cues” to “8 – video disclosure much more likely to provide cues.” Participants are then asked whether these verbal and non-verbal cues are likely to help or hurt them in investors evaluation of disclosures. This variable is expected to be impacted by whether such cues are believed to reflect anxiety or confidence. Participants are presented with a scale ranging from “1 – Likely to hurt me” to “8 – Likely to help me.” Participants are then asked to assume they had provided a video disclosure, and asked “how likely do you believe your investors would have been to pick up on verbal and non-verbal cues of your mental state based on their process.” Participants respond on a scale ranging from 57 “1 – Not at all likely” to “8 – Extremely likely.” IV. EXPERIMENT 1 – RESULTS Manipulation Checks To determine the effectiveness of my manipulations, participants respond to whether they are feeling anxious or confident about information that may be revealed during the financial disclosure process. 95% of participants answer this manipulation check correctly. Participants are also asked whether their investors are expected to use processing involving individuals, or processing involving algorithms. 94% of participants answer this manipulation check correctly. Based on this, my manipulations appear to be effective. Hypothesis 1a My expectation for H1a is that when participants wish to reveal (versus conceal) their internal (emotional) state, they are more likely to provide video disclosures when they expect investors to process information using individuals rather than algorithms. Descriptive statistics are presented in panel A of Table 1. Figure 1 presents the results graphically. [Insert Table 1] [Insert Figure 1] Results of hypothesis tests are presented in panel B and panel C of Table 1. Participants likelihood of providing a text or video disclosure serves as the dependent variable and (1) whether participants internal state is anxiety or confidence, and (2) whether participants believe investors will process their disclosures using individuals or algorithms serve as the independent variables. I begin to evaluate my predicted result by running an ANOVA and examining the interaction term. I identify support for my predicted pattern of results. Participants are less likely 58 to use a video disclosure medium in response to their internal mental state when investors are believed to rely upon a processing approach involving algorithms rather than a processing approach involving individuals (F = 4.43, p = 0.019, one-tailed equivalent). Simple effects also confirm that this interaction is in the predicted direction. Participants are more likely to provide a video disclosure, when they are feeling confident relative to anxious, when they believe investors will rely upon processing by individuals (F = 15.98, p < 0.001, one-tailed equivalent). However, no difference is observed, when participants are feeling confident relative to anxious, when they believe investors will rely upon processing by algorithms (F = 1.04, p = 0.310, two-tailed). These results provide support for H1a. Hypothesis 1b Elicited Process Measures My expectation for H1b is that managers’ strategic adjustments will be driven by whether they wish to reveal or conceal verbal and non-verbal information cues, and whether they think their investors will be able to detect these cues based on how they process information. Descriptive statistics are presented in panel A of Table 2 and panel A of Table 3. Figure 2a and Figure 2b present the results graphically. [Insert Table 2] [Insert Table 3] [Insert Figure 2a] [Insert Figure 2b] Results of hypothesis tests are presented in panel B of Table 2 and Table 3. Participants generally believe that a video disclosure is much more likely than a text disclosure to provide verbal and non-verbal cues of their internal mental state. Participants also believe that giving off 59 verbal and non-verbal cues would be more likely to help them rather than hurt them when they are feeling confident relative to anxious (t107.18 = 7.00, p < 0.001, one-tailed). Finally, participants believe investors are more likely to pick up on these verbal and non-verbal information cues when investors are expected to use processing by individuals rather than processing by algorithms (t86.31 = 3.52, p = 0.001, one-tailed). These results provide support for H1b. Linguistic Analysis of Free Response Measure To provide additional support for the mechanism underlying H1b, I perform preliminary linguistic analysis over the free response measure which contains participants’ own stated reasons for choosing to provide the text or video disclosure to investors. I perform this linguistic analysis using a self-developed R package that (1) imports participants’ raw free response text from Qualtrics, (2) transforms each free response into a series of words that remain assigned to each condition, (3) checks for the presence of each word in an established psychometric dictionary, (4) calculates the average of dictionary-matched words per condition to total words per condition, and (5) performs statistical analysis to examine whether this average differs across conditions. Descriptive statistics are presented in panel A of Table 4. Figure 3 presents the results graphically. [Insert Table 4] [Insert Figure 3] As my predicted psychological process is expected to operate through greater residual egocentrism remaining for participants who expect investors to process their disclosures using individuals rather than using algorithms, the primary dictionary that I use in performing linguistic analysis is a known and established dictionary of first-person pronouns (Pennebaker et al. 2015). Prior literature has identified that first-person pronouns are reflective of an egocentric, 60 self-focused perspective (Vogeley and Fink 2003). Appendix C contains details of this psychometric dictionary used to evaluate residual egocentrism. Results of hypothesis tests are presented in Panel B and Panel C of Table 4. Participants’ free responses reflect significantly greater egocentrism when participants expect their disclosures to be processed by investors using individual analysts rather than using algorithms to perform the first stage of information processing (F = 13.75, p < 0.001, one-tailed equivalent).14 These results provide further support for H1b and the egocentric focus in perspective taking mechanism. V. EXPERIMENT 2 – METHOD Participants Participants in a pilot version of the second experiment are 100 individuals recruited from AMT. On average, participants are 37.06 years old and have 13.95 years of work experience. Design and Manipulations In this experiment, participants are tasked with analyzing a video disclosure made by a CEO of a hypothetical firm. The experiment uses a 1 x 2 between-subjects design that manipulates whether participants are provided advice in analyzing this disclosure from an individual hired by their investment firm, or from an algorithm designed by their investment firm. To mirror the setting from the first experiment, the investment firm is described as a sophisticated institutional investor. The main function of this design is to bolster inferences about the psychological mechanism identified in Experiment 1. This design seeks to accomplish this by 14 The second stage of the expected psychological mechanism, insufficient adjustments from an egocentric default, is expected to operate through differences in perceived similarity between individuals and algorithms. As such, I also perform analysis over another psychometric measure that is expected to be affected by similarity. Similarity is known to lead to greater trust is the psychology literature (Farmer et al. 2014), so I examine trust-oriented language across conditions using the NRC emotion lexicon (Mohammad and Turney 2013). Consistent with the effects of similarity on egocentrism, trust-oriented language is also used more frequently when information processing is expected to be performed using individuals rather than algorithms (F = 2.96, p = 0.085, two-tailed, not tabulated). 61 demonstrating inconsistent information processing expectations emerge between the two settings as a result of participants’ experiencing different egocentric default perspectives in the two experiments, with other setting characteristics held constant. Task and Procedure After being introduced to the task, participants are presented with details relevant to the setting. They are told that they are tasked with determining whether the CEO presenting information in a video disclosure is being honest or concealing information. They consider this task as they study the CEO’s behavior in presenting information in a video disclosure. After watching the video, participants are also provided with a manipulation between conditions where they receive advice, in the form of a summary assessment, from either an individual or an algorithm. This advice claims to have detected indicators (or no indicators) of concealment present in the verbal and non-verbal behavior of the CEO presenting information in the video. To avoid concerns that trust in advice is directional based on claims that indicators of concealment are detected, an additional control is implemented. An additional randomization occurs over whether the provided advice states that “there are indicators [no indicators] of deception present” in the behavior of the CEO in the video. Next, participants are asked how likely they would be to trust the assessment provided by the individual or algorithm. Finally, participants provide demographic information. Dependent Measures Trust in Advice Measure In my setting, the main dependent variable is how much participants would trust the accuracy of an assessment provided by the individual or the algorithm. Participants respond on a scale ranging from “1 – Not at all trust” to “8 – Completely trust.” 62 VI. EXPERIMENT 2 – RESULTS Manipulation Checks To determine the effectiveness of the manipulation, participants are asked “did you receive an assessment as to whether or not the individual in the video may be concealing information from an algorithm or from an individual”? 90% of participants answer correctly, indicating a successful manipulation. Hypothesis 2 My expectation for H2 is that participants are more likely to trust an algorithm-provided summary assessment compared to an individual-provided summary assessment when they are offered assistance in evaluating the verbal and non-verbal behavior of others. Descriptive statistics are presented in panel A of Table 5. Figure 4 presents the results graphically. [Insert Table 5] [Insert Figure 4] Results of hypothesis tests are presented in panel B of Table 5. Participants’ likelihood of trust in advice measure serves as the dependent variable and whether participants receive advice from an individual or an algorithm serves as the independent variable. Support is identified for the predicted effect. Participants are more likely to place trust in the accuracy of an assessment when this assessment is received from an algorithm rather than an individual (t97.04 = 2.46, p = 0.008, one-tailed).15 This result provides support for H2. 15 Further, inferences are consistent both when the advice claims to have detected indicators of concealment (F = 3.83, p = 0.026, one-tailed equivalent, not tabulated) or detected no indicators of concealment (F = 2.30, p = 0.066, one-tailed equivalent, not tabulated). There is also no main effect of indicators of concealment on trust in advice (F = 0.05, p = 0.600, two-tailed, not tabulated). This indicates that the effect is not conditional and occurs both when the algorithmic advice claims to detect and not detect revealing information indicators. 63 VII. SURVEY OF EXPERIENCED MANAGERS The two laboratory experiments demonstrate that an inconsistency emerges between participants’ expectations of the information processing costs and capabilities of individuals and algorithms, depending on whether participants are the ones communicating or analyzing verbal and non-verbal information. The purpose of my survey is to provide complementary evidence on existing beliefs of the capabilities of algorithms or individual analysts within the institutional setting, outside of the laboratory-induced perspective of either role. Survey participants are 116 experienced managers recruited through a highly-rated EMBA program. On average, participants are approximately 38 years old and have spent 6.54 years with their current employer. Participants report working across a wide variety of industries in firms that, on average, have a market capitalization of greater than $1 billion. Participants also report working across a variety of job functions including management, finance, marketing/sales, and engineering/R&D. Full demographic information is presented in Table 6. [Insert Table 6] Method The survey elicits beliefs about the perceived information processing strengths and weaknesses of algorithms and individual analysts. Specifically, experienced managers are presented with a prompt that says “performance information can be evaluated by different types of users including (1) human analysts and (2) algorithms or software. Which of these users would you say is better at performing each of the following tasks?” Results 64 Beliefs about Algorithms and Analysts Table 7 presents descriptive evidence with respect to experienced managers’ beliefs about how algorithms and individuals process information. A directional difference as to expected capabilities emerges across written data and video data. Although experienced managers are not explicitly put into a communication or analysis frame of mind while taking the survey, they directionally expect algorithms to be relatively better at processing written data (61 to 55) and individual analysts to be relatively better at processing video data (66 to 50). [Insert Table 7] Overall, this survey adds context to the experimental findings by demonstrating that the observed pattern of beliefs in the institutional environment about information processing capabilities across different communication mediums is somewhat balanced, although the video data analysis assessment is relatively closer to that observed from the communicator perspective in Experiment 1. Notably, this survey assesses a broader perspective than that examined in the experimental instruments. The survey asks about beliefs of overall information processing capabilities relating to text and video data, while the experiments focus more narrowly on beliefs about the detection of verbal and non-verbal information cues in video data. VIII. CONCLUSION Emerging literature demonstrates that managers’ reactions to expectations of investors’ information processing costs and capabilities can significantly affect financial reporting outcomes and the information that reaches the capital markets. Less is known about how manager expectations of investors’ information processing costs and capabilities are formed. This study investigates a fundamental cognitive mechanism that can cause information 65 processing expectations to form, and it demonstrates a systematic egocentric focus in perspective taking bias that can distort managers’ judgments. In this study, two controlled experiments reflect an inconsistency in beliefs about the expected capabilities of algorithm-based information processing (relative to individual-based processing) when participants’ perspectives are flipped from communicating to analyzing information. In a first experiment, I find that participants are less likely to adjust their use of a video disclosure medium when investors are believed to rely upon processing involving algorithms rather than processing involving individuals. This finding occurs as managers feel that individuals (relative to algorithms) are more likely to detect verbal and non-verbal information cues indicative of their internal mental state when they issue video disclosures. In a second experiment, I find that when participants are the ones actually analyzing video disclosures, they trust advice coming from an algorithm relatively more than advice coming from an individual when this advice claims to have detected (or not detected) information cues in the verbal and non-verbal behavior of a CEO presenting a disclosure. The results from the two experiments are logically inconsistent when viewed together, but reflect a systematic egocentric focus in perspective taking bias. Participants anchor on their own perspective, and adjust less from this egocentric default when considering the perspectives of other individuals relative to algorithms. In Experiment 1, participants strongly feel their own internal behavior when communicating disclosures, and in Experiment 2 participants have more difficulty detecting the internal behavior of others when analyzing disclosures. In both experiments, they adjust relatively less from their own perspective for more similar individuals as opposed to less similar algorithms. In an additional survey of experienced managers, I find that current beliefs about algorithms and individuals are more consistent with the communicator 66 perspective in the first experiment. These findings offer a number of contributions. First, this study adds to literature on disclosure mediums. This study provides information on how managers may respond to the expanding number of choices available for how they issue disclosures. This finding builds upon prior literature that demonstrates these disclosure mediums can affect the information content that reaches the markets (Blankespoor et al. 2017; Mayew and Venkatachalam 2012) and how this information is processed by investors (Libby and Emett 2014; Elliott et al. 2012). This study demonstrates that managers respond to expectations of investors in choosing different disclosure mediums, but moreover that these responses can be biased by fundamental cognitive mechanisms. The nature of this bias may be impactful as it affects the information that reaches investors in the capital market environment. It could inhibit investors’ ability to take advantage of expanding resources to analyze the “soft” qualitative information that may be contained in firm disclosures. Second, my findings contribute to the emerging literature on feedback effects of investors’ information processing costs on managers behavior (Blankespoor et al. 2019). This is the first study to demonstrate that cognitive biases relating to the formation of information processing expectations can distort manager behavior in consequential ways. Given that prior research on feedback effects demonstrates that these effects are likely to be very widespread, and have a significant influence on both reporting and operating decisions (Blankespoor et al. 2019; Witz Chapter 1), the notion that these effects may be influenced by unintended biases is an important evolution of this literature. Whereas prior work did not discuss the cognitive mechanisms that were likely to generate feedback effects, this study demonstrates that understanding these mechanisms can be a rich avenue of future research. This understanding 67 may help inform reactions to the discovery of different feedback effects, as managers’ decisions directly affect the nature information that reaches the capital markets. These findings also contribute to practice. It is unlikely that managers are aware their disclosure choices may be affected by cognitive biases, and this knowledge might help them correct any inadvertent behavior. This study also demonstrates that managers may opportunistically choose to utilize different disclosure mediums, and regulators and standard- setters may find this informative as they consider how to produce standards in the evolving disclosure landscape. Additionally, this study offers perspective into how managers may react to investors’ increasing use of algorithms, and calls attention to distortion that can occur in the information environment as a result. My study is subject to certain limitations. One limitation is that I explore my research questions in controlled and simplified reporting settings. The takeaways from these settings are likely to generalize to the broader institutional environment, given that the relevant theory reflects fundamental human psychology that all humans exhibit. However, the exploration of manager reporting biases is difficult to explore outside of a controlled laboratory context, given the complex and myriad ways in which expectations can form. In the institutional environment, it is likely that expectations are an aggregation of both logical thinking and cognitive biases, which are difficult to disentangle. An additional limitation is that managers and investors may have different goals, and these goals may influence decision-making beyond the mechanisms explored in this study. Although this study demonstrates beliefs in information processing capabilities of individuals and algorithms are inconsistent between different settings when participants are presenting and analyzing information, it is also possible that factors beyond processing beliefs can influence the choice of disclosure medium and the utilization of advice. 68 My study also provides many opportunities for future research. Future research may consider investigating other systematic biases that emerge in managers formation of investors’ information processing expectations. Future research can also explore whether interventions can correct these potential biases, and limit the distortion that might occur to the capital markets in the form of misguided feedback effects. 69 Figure 1 – Experiment 1 – Likelihood to Choose Text or Video Likelihood to Choose to Provide Text (1) or Video (8) Disclosure 8 7 6 5.7 5 4 3.43 4.04 3 3.32 2 1 Anxious Confident Individual Type Algorithm Type This figure graphically presents participants’ likelihood to choose to provide text or video disclosures to their investors. 70 Figure 2a – Experiment 1 – Would Giving Off Verbal and Non-Verbal Cues Hurt or Help Belief that Cues Would Hurt (1) or Help (8) 8 7 6 5.61 5 4 3.36 3 2 1 Anxious Confident This figure graphically presents participants’ beliefs as to whether giving off verbal and non-verbal cues would hurt them or help them in investors’ evaluation of their disclosure. Figure 2b – Experiment 1 – Would Investors Pick Up on Verbal and Non-Verbal Cues Likelihood of Investors Picking up on Video Disclosure Cues 8 7 6.45 6 5.58 5 4 3 2 1 Individual Type Algorithm Type This figure graphically presents participants’ beliefs as to whether, if they provided video disclosures, investors would be likely to pick up on verbal and non-verbal cues of their mental state. 71 Figure 3 – Experiment 1 – Linguistic Analysis of Residual Egocentrism Egocentric Words per Total Words 0.1 0.09 0.08 0.07 0.06 0.05 0.05 0.04 0.029 0.03 0.02 0.01 0 Individual Type Algorithm Type This figure graphically presents participants’ use of first-person pronouns as a percentage of total words in their free response measure explaining why they made their decision to provide text or video disclosures. This measure is designed to capture residual egocentrism reflective of the predicted psychological process. 72 Figure 4 – Experiment 2 – Trust in Accuracy of Assessment Provided Trust in Accuracy of Assessment Provided 8 7 6.12 6 5.33 5 4 3 2 1 Individual Advice Algorithm Advice This figure graphically presents participants’ trust in the accuracy of the summary assessment provided by an individual hired by or an algorithm designed by the investment firm participants work for. 73 Table 1 – Experiment 1 – Likelihood to Choose Text or Video Disclosure TABLE 1 Panel A: Descriptive statistics: Mean and (standard deviation) Anxious Confident Row Means A B 3.32 5.70 4.49 Individual (2.28) (1.68) (2.32) n=28 n=27 n=55 C D 3.43 4.04 3.73 Algorithm (2.33) (2.46) (2.39) n=28 n=27 n=55 3.37 4.87 4.11 Overall (2.29) (2.50) (2.38) n=56 n=54 n=110 Panel B: Analysis of variance results Source of Variation S.S. df M.S. F-statistic p-value State 61 1 61 12.59 0.000 Investor Type 17 1 17 3.42 0.067 State * Investor Type 22 1 22 4.43 0.019* Error 518 106 5 Panel C: Simple effects tests for text or video disclosure judgment Comparisons df F-Statistic p-value Effect of State given Individual-based Investor (A vs. B) 106 15.98 0.000 Effect of State given Algorithm-based Investor (C vs. D) 106 1.04 0.310 This table presents descriptive statistics, ANOVA results, and simple effects tests for participants’ likelihood to choose to provide text or video disclosures. Two manipulations occur over whether participants believe they have an anxious or confident internal state and whether participants expect their investors to process their disclosures using individuals or using algorithms. Participants are asked “Would you be more likely to choose text or video?” Participants are presented with a scale ranging from “1 – Extremely likely to choose text” to “8 – Extremely likely to choose video.” *One-tailed equivalent. 74 Table 2 – Experiment 1 – Belief that Cues Would Hurt or Help TABLE 2 Panel A: Descriptive statistics: Mean and (standard deviation) Anx ious Confi dent Ove rall 3.36 5.61 4.46 Would Cues (1.79) (1.58) (2.03) Help or Hurt? n=56 n=54 n=110 Panel B: Welch’s t-test for belief that cues would help or hurt Comparisons df t-statistic p-value Anxious vs. Confident Internal State 107.18 7.00 0.000* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to whether they believe giving off verbal and non-verbal cues be likely to hurt or help them in investors’ evaluation of their disclosure. Two manipulations occur over whether participants believe they have an anxious or confident internal state and whether participants expect their investors to process their disclosures using individuals or using algorithms. Participants are asked “Would giving off verbal and non-verbal cues be likely to hurt you or help you in investors evaluation of your disclosure?” Participants are presented with a scale ranging from “1 – Likely to hurt me” to “8 – Likely to help me.” *One-tailed. 75 Table 3 – Experiment 1 – Likelihood of Investors Picking up on Video Disclosure Cues TABLE 3 Panel A: Descriptive statistics: Mean and (standard deviation) Indiv idual Algor ithm Ove rall Likelihood of 6.45 5.58 6.02 Investors (0.92) (1.60) (1.39) Picking up on n=55 n=55 Cues n=110 Panel B: Welch’s t-test for likelihood of investors picking up on video disclosure cues Comparisons df t-statistic p-value Individual vs. Algorithm Investor Type 86.31 3.52 0.001* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to, if they provided a video disclosure, how likely do they believe their investors would have been to pick up on verbal and non-verbal cues of their mental state. Two manipulations occur over whether participants believe they have an anxious or confident internal state and whether participants expect their investors to process their disclosures using individuals or using algorithms. Participants are asked “If you provided a video disclosure, how likely do you believe your investors would have been to pick up on verbal and non-verbal cues of your mental state based on their process?” Participants are presented with a scale ranging from “1 – Not at all likely” to “8 – Extremely likely.” *One-tailed. 76 Table 4 – Experiment 1 – Linguistic Analysis of Residual Egocentrism (Unobtrusive) TABLE 4 Panel A: Descriptive statistics: Mean and (standard deviation) Anxious Confident Row Means A B 0.055 0.045 0.050 Individual (0.23) (0.21) (0.21) n=1089 n=1247 n=2336 C D 0.024 0.034 0.029 Algorithm (0.16) (0.18) (0.18) n=1267 n=1204 n=2471 0.39 0.040 0.039 Overall (0.19) (0.20) (0.19) n=2356 n=2451 n=4807 Panel B: Analysis of variance results Source of Variation S.S. df M.S. F-statistic p-value State 0.0 1 0.0 0.00 0.957 Investor Type 0.5 1 0.5 13.75 0.000* State * Investor Type 0.2 1 0.2 3.12 0.077 Error 180.0 4803 0.0 Panel C: Simple effects tests for residual egocentrism Comparisons df F-Statistic p-value Effect of Type given Anxious (A vs. B) 4803 14.66 0.000* Effect of Type given Confident (C vs. D) 4803 1.93 0.080* This table presents descriptive statistics, ANOVA results, and simple effects tests for participants’ use of first-person pronouns as a percentage of total words in their free response explaining why they chose to provide text or video disclosures. This measure is designed to capture residual egocentrism that participants are experiencing immediately after providing the main DV, but before responding to obtrusive process measures. Two manipulations occur over whether participants believe they have an anxious or confident internal state and whether participants expect their investors to process their disclosures using individuals or using algorithms. After participants indicate their likelihood of providing a text or video disclosure, participants are asked to “Please provide insight into the thinking that led to your decision:” *One-tailed equivalent. 77 Table 5 – Experiment 2 – Trust in Accuracy of Assessment Provided TABLE 5 Panel A: Descriptive statistics: Mean and (standard deviation) Indiv idual Algor ithm Ove rall Trust in 5.33 6.12 5.73 Accuracy of (1.49) (1.72) (1.65) Assessment n=49 n=51 n=100 Panel B: Welch’s t-test for trust in accuracy of assessment provided Comparisons df t-statistic p-value Individual vs. Algorithm 97.04 2.46 0.008* This table presents descriptive statistics and a Welch’s t-test for participants’ responses as to how much they would trust the accuracy of a summary assessment that claims to detect (or not detect) indicators in the verbal and nonverbal behavior of the CEO presenting information in a video disclosure. A manipulation occurs over whether participants receive this assessment from an algorithm designed by their investment firm or an individual hired by their investment firm. Participants are asked “How much would you trust the accuracy of this assessment provided by the individual [algorithm].” Participants are presented with a scale ranging from “1 – Not at all trust” to “8 – Completely trust.” *One-tailed. 78 Table 6 – Demographic Characteristics of Survey Respondents TABLE 6 Primary Industry % Years with Current Employer % Consumer Discretionary 4.31 <1 year 1.78 Consumer Staples 4.31 1-3 years 38.39 Energy 16.38 4-6 years 16.07 Financials 18.10 7-9 years 17.86 Health Care 7.76 10+ years 25.90 Industrials 6.90 Information Technology 13.79 Company market capitalization Materials 2.59 <$100 million 23.28 Other 25.86 $100 million - $249 million 7.75 $250 million - $499 million 10.34 Age $500 million - $999 million 2.59 <30 0.86 $1 billion - $10 billion 19.83 30-39 70.69 >$10 billion 36.21 40-49 26.72 50-59 1.73 Primary Job Function Engineering / R&D 17.24 Gender Finance 11.21 Male 73.28 Management 40.52 Female 26.72 Marketing / Sales 12.07 Other 18.96 This table presents demographic information for a survey conducted of 116 experienced managers recruited from a highly rated Executive MBA program. 79 Table 7 – Survey responses “Performance information can be evaluated by different types of users including (1) human analysts and (2) algorithms or software. Which of these users would you say is better at performing each of the following tasks?” TABLE 7 Human Analysts Algorithms Analyzing written data 55 61 Analyzing video data 66 50 This table presents survey responses to the question: “Performance information can be evaluated by different types of users including (1) human analysts and (2) algorithms or software. Which of these users would you say is better at performing each of the following tasks?” The question is presented as a binary choice and data is recorded as the count of the number of participants selecting each response. 80 Appendix A – Experimental Materials (Manipulations) Experiment 1 – Internal State Manipulation How You Are Feeling About Issuing Disclosures In the current period, you are feeling extremely anxious [confident] about issuing financial disclosures. You are feeling anxious [confident] about the nature of information that has not yet been revealed, but could be during the financial disclosure process. - - - - - - Think about a time in your life where being anxious [confident] while communicating affected your audience’s perception of what you were saying. Experiment 1 – Investor Type Manipulation How Investors Process Your Disclosures Your investor relations officer has found that the typical investor processes your disclosure directly [indirectly], with an individual [algorithm] preparing a summary assessment of the disclosure that the investor relies upon in its decision-making. The content of your disclosure and the way in which it is communicated will thus be viewed directly [indirectly] through an individual’s eyes [algorithm’s scan]. 81 Experiment 2 – Source of Summary Assessment Manipulation How Investors Process Your Disclosures You will receive a summary assessment of whether or not the CEO is concealing information provided by an individual [an algorithm] that Alpha Investments hired [designed]. - - - - - - The individual [algorithm] claims with 95% confidence that there are indicators [no indicators] of deception present in the verbal and nonverbal behavior displayed in the video. 82 Appendix B – Roadmap of Egocentric Focus in Perspective Taking Bias in Overall Process 83 Appendix C – Dictionary Used to Evaluate Residual Egocentrism (LIWC 2015 First Person Singular Pronouns) Words Included in Dictionary i i'd i'd've I'll I'm I've id idc idgaf idk idontknow idve ikr ily im ima imean imma ive me methinks mine my myself 84 REFERENCES Abramova, I., Core, J., & Sutherland, A. (2019). Institutional Investor Attention and Firm Disclosure. Working Paper. Allee, K. D., DeAngelis, M. D., & Moon Jr, J. R. (2018). Disclosure “Scriptability”. Journal of Accounting Research, 56 (2), 363-430. Asay, H.S. (2018). Horizon-Induced Optimism as a Gateway to Earnings Management. Contemporary Accounting Research, 35 (1), 7-30. Asay, H. S., Libby, R., & Rennekamp, K. M. (2018a). Do features that associate managers with a message magnify investors’ reactions to narrative disclosures?. Accounting, Organizations and Society, 68, 1-14. Asay, H. S., Libby, R., & Rennekamp, K. (2018b). Firm performance, reporting goals, and language choices in narrative disclosures. Journal of Accounting and Economics, 65 (2- 3), 380-398. Banerjee, R. (2002). Audience effects on self-presentation in childhood. Social Development, 11 (4), 487-507. Basu, R., Pierce, S., & Stephan, A. (2019). The effect of investor inattention on voluntary disclosure. Working Paper. Bentley, J. W. (2019). Decreasing operational distortion and surrogation through narrative reporting. The Accounting Review, 94 (3), 27-55. Bhattacharya, N., Cho, Y. J., & Kim, J. B. (2018). Leveling the playing field between large and small institutions: evidence from the SEC's XBRL mandate. The Accounting Review, 93 (5), 51-71. Blankespoor, E. (2019). The Impact of Information Processing Costs on Firm Disclosure Choice: Evidence from the XBRL Mandate. Journal of Accounting Research. Forthcoming. Blankespoor, E., deHaan, E, & Marinovic, I. (2019). Disclosure Processing Costs’, Investors’ Information Choice, and Equity Market Outcomes: A Review. Working Paper. Blankespoor, E., Hendricks, B. E., & Miller, G. S. (2017). Perceptions and price: Evidence from 85 CEO presentations at IPO roadshows. Journal of Accounting Research, 55 (2), 275-327. Bloomfield, R. (2008). Discussion of “annual report readability, current earnings, and earnings persistence”. Journal of Accounting and Economics, 45 (2), 248-252. Bloomfield, R., Nelson, M. W., & Soltes, E. (2016). Gathering data for archival, field, survey, and experimental accounting research. Journal of Accounting Research, 54 (2), 341-395. Brown, T.J. (2014). Advantageous comparison and rationalization of earnings management. Journal of Accounting Research, 52 (4), 849-876. Brown, L.D., Call, A.C., Clement, M.B., and Sharp, N.Y. (2019). Managing the narrative: investor relations officers and corporate disclosures. Journal of Accounting and Economics, 67 (1), 58-79. Bushee, B.J., & Miller, G.S. (2012). Investor relations, firm visibility, and investor following. The Accounting Review, 87 (3), 867-897. Cantrell, A. (2017). Quants Rule Alpha’s Hedge Fund 100 List. Retrieved from Institutional Investor’s Website: https://www.institutionalinvestor.com/article/b1505pmf2v2hg3/quants-rule-alphas- hedge-fund-100-list. Chan, L. H., Chen, K. C., Chen, T. Y., & Yu, Y. (2015). Substitution between real and accruals- based earnings management after voluntary adoption of compensation clawback provisions. The Accounting Review, 90 (1), 147-174. Clement, R. W., & Krueger, J. (2002). Social categorization moderates social projection. Journal of Experimental Social Psychology, 38 (3), 219-231. Clor‐Proell, S. M., & Maines, L. A. (2014). The impact of recognition versus disclosure on financial information: A preparer's perspective. Journal of Accounting Research, 52 (3), 671-701. Cohen, D.A., Dey, A., & Lys, T.Z. (2008). Real and accruals-based earnings management in the pre- and post-Sarbanes-Oxley periods. The Accounting Review, 83 (3), 757-787. Commerford, B. P., Dennis, S. A., Joe, J.R., & Wang, J. (2019). Complex estimates and auditor 86 reliance on artificial intelligence. Working Paper. D’Arcy, J., Hovav, A., & Galletta, D. (2009). User awareness of security countermeasures and its impact on information systems misuse: A deterrence approach. Information Systems Research, 20 (1), 79-98. Dawes, R. M., Faust, D., & Meehl, P. E. (1989). Clinical versus actuarial judgment. Science, 243 (4899), 1668-1674. Dechow, P., Ge, W., & Schrand, C. (2010). Understanding earnings quality: A review of the proxies, their determinants and their consequences. Journal of accounting and economics, 50 (2-3), 344-401. Dechow, P. M., Sloan, R. G., & Sweeney, A. P. (1995). Detecting earnings management. Accounting Review, 70 (2), 193-225. de Visser, E. J., Krueger, F., McKnight, P., Scheid, S., Smith, M., Chalk, S., & Parasuraman, R. (2012, September). The world is not enough: Trust in cognitive agents. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting (Vol. 56, No. 1, pp. 263- 267). Sage CA: Los Angeles, CA: Sage Publications. Dunning, D. A., Alicke, M. D., & Krueger, J. I. (Eds.). (2005). The self in social judgment. Psychology Press. Elliott, W. B., Hodge, F. D., & Sedor, L. M. (2012). Using online video to announce a restatement: Influences on investment decisions and the mediating role of trust. The Accounting Review, 87 (2), 513-535. Epley, N. (2008). Solving the (real) other minds problem. Social and personality psychology compass, 2 (3), 1455-1474. Farmer, H., McKay, R., & Tsakiris, M. (2014). Trust in me: Trustworthy others are seen as more physically similar to the self. Psychological science, 25 (1), 290-292. Fenigstein, A., & Abrams, D. (1993). Self-attention and the egocentric assumption of shared perspectives. Journal of experimental social psychology, 29 (4), 287-303. 87 Fussell, S.R., & Krauss, R.M. (1992). Coordination of knowledge in communication: Effects of speakers’ assumptions about what others know. Journal of Personality and Social Psychology, 62 (3), 378-391. Gardner, W.L., & Martinko. M.J. (1988). Impression management: An observational study linking audience characteristics with verbal self-presentations. Academy of Management Journal, 31 (1), 42-65. Gendolla, G. H., & Wicklund, R. A. (2009). Self-focused attention, perspective-taking, and false consensus. Social Psychology, 40 (2), 66-72. Gibbins, M., McCracken, S., & Salterio, S.E. (2010). The auditor’s strategy selection for negotiation with management: Flexibility of initial accounting position and nature of the relationship. Accounting, Organizations, and Society, 35, 579-595. Gilovich, T., & Savitsky, K. (1999). The Spotlight Effect and the Illusion of Transparency: Egocentric assessments of how we are seen by others. Current Directions in Psychological Science, 8 (6), 165-168. Gilovich, T., Savitsky, K., & Medvec, V.H. (1998). The illusion of transparency: Biased assessments of others ability to read our emotional states. Journal of Personality and Social Psychology, 75, 332-346. Gilovich, T., Savitsky, K., & Medvec, V.H. (2000). The Spotlight Effect in social judgment: An egocentric bias in estimates of the salience of one’s own actions and appearance. Journal of Personality and Social Psychology, 78 (2), 211-222. Gilovich, T., Kruger, J., & Medvec, V.H. (2002). The Spotlight Effect Revisited: Overestimating the manifest variability of our actions and appearance. Journal of Experimental Social Psychology, 38, 93-99. Graham, J.R., Harvey, C.R., & Rajgopal, S. (2005). The economic implications of corporate financial reporting. Journal of Accounting and Economics, 40, 3-73. Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical prediction: a meta-analysis. Psychological assessment, 12 (1), 19. Guay, W., Samuels, D., & Taylor, D. (2016). Guiding through the fog: Financial statement complexity and voluntary disclosure. Journal of Accounting and Economics, 62 (2), 234- 88 269. Guggenmos, R. (2019). The effects of a creative culture on real earnings management. Working paper. Gunny, K. A. (2010). The relation between earnings management using real activities manipulation and future performance: Evidence from meeting earnings benchmarks. Contemporary Accounting Research, 27 (3), 855-888. Hales, J., Koka, B., & Venkataraman, S. (2018). Curbing earnings management: Experimental evidence on how clawback provisions and board monitoring affect managers’ use of discretion. Working paper. Healy, P. M., & Wahlen, J. M. (1999). A review of the earnings management literature and its implications for standard setting. Accounting horizons, 13 (4), 365-383. Hirsh, J.B., Kang, S.K., & Bodenhausen, G.V. (2012). Personalized persuasion: Tailoring persuasive appeals to recipients’ personality traits. Psychological Science, 23 (6), 578- 581. Kirk, M.P., & Markov, S. (2016). Come on over: Analyst/investor days as a disclosure medium. The Accounting Review, 91 (6), 1725-1750. Kirk, M.P., & Vincent, J.D. (2014). Professional investor relations within the firm. The Accounting Review, 89 (4), 1421-1452. Li, F. (2008). Annual report readability, current earnings, and earnings persistence. Journal of Accounting and economics, 45 (2), 221-247. Libby, R., Bloomfield, R., & Nelson, M. (2002). Experimental research in financial accounting. Accounting, Organizations, and Society, (27), 775-810. Libby, R., & Emett, S. A. (2014). Earnings presentation effects on manager reporting choices and investor decisions. Accounting and Business Research, 44 (4), 410-438. Liberti, J. M., & Petersen, M.A. (2018). Information: Hard and Soft. Review of Corporate Finance Studies, 8 (1), 1-41. 89 Lo, K. (2008). Earnings management and earnings quality. Journal of Accounting and Economics, 45 (2-3), 350-357. Lo, K., Ramos, F., & Rogo, R. (2017). Earnings management and annual report readability. Journal of Accounting and Economics, 63 (1), 1-25. Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90-103. Mayew, W. J., & Venkatachalam, M. (2012). The power of voice: Managerial affective states and future firm performance. The Journal of Finance, 67 (1), 1-43. Miller, G. S., & Skinner, D. J. (2015). The evolving disclosure landscape: How changes in technology, the media, and capital markets are affecting disclosure. Journal of Accounting Research, 53 (2), 221-239. Mitchell, J. P., Macrae, C. N., & Banaji, M. R. (2006). Dissociable medial prefrontal contributions to judgments of similar and dissimilar others. Neuron, 50 (4), 655-663. Mohammad, S. M., & Turney, P. D. (2013). Nrc emotion lexicon. National Research Council, Canada. Nagin, D.S., & Pogarsky, G. (2001). Integrating celerity, impulsivity, and extralegal sanction threats into a model of general deterrence: Theory and evidence. Criminology, 39 (4), 865-891. Noar, S.M., Harrington, N.G., Aldrich, R.S. (2009). The role of message tailoring in the development of persuasive health communication messages. From C.S. Beck’s Communication Yearbook, 33, 73-133. Pak, R., Fink, N., Price, M., Bass, B., & Sturre, L. (2012). Decision support aids with anthropomorphic characteristics influence trust and performance in younger and older adults. Ergonomics, 55 (9), 1059-1072. Pennebaker, J. W., Boyd, R. L., Jordan, K., & Blackburn, K. (2015). The development and psychometric properties of LIWC2015. 90 Phillips, J., Pincus, M., & Rego, S. O. (2003). Earnings management: New evidence based on deferred tax expense. The Accounting Review, 78 (2), 491-521. Piaget, J. (1932). 1965. The moral judgment of the child. Promberger, M., & Baron, J. (2006). Do patients trust computers? Journal of Behavioral Decision Making, 19 (5), 455-468. Rennekamp, K. (2012). Processing fluency and investors’ reactions to disclosure readability. Journal of Accounting Research, 50 (5), 1319-1354. Robbins, J. M., & Krueger, J. I. (2005). Social projection to ingroups and outgroups: A review and meta-analysis. Personality and social psychology review, 9 (1), 32-47. Roychowdhury, S. (2006). Earnings management through real activities manipulation. Journal of Accounting and Economics, 42 (3), 335-370. Roychowdhury, S., Shroff, N., & Verdi, R. (2019). The effects of financial reporting and disclosure on corporate investment: A review. Working Paper. Securities and Exchange Commission (SEC). (1998). A Plain English Handbook: How to Create Clear SEC Disclosure. SEC Office of Investor Education and Assistance. Retrieved from http://www.sec.gov/pdf/handbook.pdf. Securities and Exchange Commission (SEC). (2009). Interactive Data to Improve Financial Reporting. Retrieved from https://www.sec.gov/rules/final/2009/33-9002.pdf. Sinha, R. R., & Swearingen, K. (2001). Comparing recommendations made by online systems and friends. In DELOS. Sloan, R. G. (2019). Fundamental analysis redux. The Accounting Review, 94 (2), 363-377. Solomon, D., & Soltes, E. (2015). What are we meeting for? The consequences of private meetings with investors. Journal of Law and Economics, 58: 325-355. Trentmann, N. (2019). Companies adjust financial communications to resonate with humans and robots. Wall Street Journal. Available at: https://www.wsj.com/articles/companies- adjust-financial-communications-to-resonate-with-humans-and-robots-11574193390. 91 Vogeley, K., & Fink, G. R. (2003). Neural correlates of the first-person-perspective. Trends in cognitive sciences, 7 (1), 38-42. Whalen, C. (2004). Opening the books: A new computer format for financial reporting could improve investment analysis. Barron’s Online. Wigglesworth, R. (2017). Fledgling quant funds seek to disrupt Wall Street. Financial Times. Available at: http://www.ft.com/content/83f4e562-65be-11e7-9a66-93fb352ba1fe. Yeomans, M., Shah, A., Mullainathan, S., & Kleinberg, J. (2019). Making sense of recommendations. Journal of Behavioral Decision Making, 32, 403-414. Zang, A. Y. (2012). Evidence on the trade-off between real activities manipulation and accrual- based earnings management. The Accounting Review, 87 (2), 675-703. Zuckerman, G., & Hope, B. (2017). The quants run Wall Street now. Wall Street Journal. Available at: https://www.wsj.com/articles/the-quants-run-wall-street-now-1495389108. 92