The Opioid Crisis, Employee Health Capital, and Corporate Information Production

Abstract

This study examines the effect of employee health on corporate information production. Utilizing exposure to opioid abuse as the proxy for employee health, we find that firms headquartered in counties of high opioid prescription rate produce significantly less accurate management earnings forecasts. This result is robust to controlling for other dimensions of human capital and to utilizing the effective anti-opioid legislation across states as a plausibly exogenous variation that limits the prescriptions of opioid. The negative effect of opioid abuse is stronger for firms facing higher forecast difficulty, and is mitigated for firms with easier access to opioid treatment, for firms with superior employee welfare policies, and for firms with a corporate social responsibility (CSR) committee. We also show that managers delay earnings announcements and reduce forecast precision amidst high local opioid activity. Finally, we show that investors react less strongly to news in forecasts issued by firms located in high opioid areas, consistent with their recognition of the potential adverse effect of opioid abuse on information production within the firm.

Keywords:

• Opioid crisis
• Management forecasts
• Health capital

JEL codes:

• G02
• J01
• J30
• M51

1. Introduction

Corporate information production is a key mechanism employed by firms to inform the market, reduce information asymmetry, and lower information uncertainty (Beyer et al., 2010; Ferracuti & Stubben, 2019), making it crucial to investigate the factors that influence the quality of information produced. The impact of human capital, comprised of the knowledge, skills and health that people invest in and accumulate throughout their lives (Schultz, 1961), is of particular interest because each of these attributes conceivably impacts employees’ ability to collect, process and communicate complex information about their employer and its prospects. While existing literature has highlighted the significance of the education and skills of a firm’s workforce in generating high-quality corporate information (Call et al., 2017; Chen, Wu, et al., 2021), little is known about how human health capital affects corporate information production. Among the various aspects of human capital, health capital stands out as the most fundamental and critical (Grossman, 1972; Becker, 2007). Unlike other components of human capital such as education and skills, individuals’ health capital can vary swiftly over time, directly affecting their ability to effectively apply existing knowledge and skills to work tasks (Grossman, 1972; Helliwell, 2003; Becker, 2007). Indeed, CEOs’ growing recognition of the role of employee health during the COVID-19 pandemic, highlights the importance of this aspect of human capital.¹ Thus, to enhance our understanding of the impact of health capital on corporate information production, we examine how variation in employee health affects the quality of management earnings forecasts, the production of which is labor-intensive and relies heavily on the input of employees located at a firm’s headquarters who are required to collaborate under pressure to meet reporting deadlines (Suphap, 2003; Sherin, 2010; Chourou et al., 2020; Chen et al., 2023).

Examining the effect of employee health on corporate information production empirically is challenging. Firstly, previous research indicates that poor employee health can result in reduced performance due to both ‘absenteeism’ and ‘presenteeism’, making it difficult to identify a proxy that captures the complete effect of employee health on forecast quality (Caverley et al., 2007; Johns, 2010). Secondly, both employee health and forecast quality may be influenced by unobserved factors (e.g., management policies), complicating efforts to establish causal relationships. To address these concerns, we utilize the health crisis deriving from opioid abuse in the US to examine the impact of employee health on the quality of information production.² Previous studies document the impact of opioids on an individual’s job performance, encompassing both ‘absenteeism’ and ‘presenteeism’ factors (Ruetsch, 2010; Henke et al., 2020),³ suggesting that employing opioid abuse as a measure for employee well-being holds the potential to identify health-related consequences that may go unnoticed or do not lead to work absences. Moreover, a significant literature has revealed that the opioid crisis emerged as a result of extensive misinformation campaigns conducted by pharmaceutical companies in the 1990s. These campaigns specifically targeted the medical community and downplayed the risk of addiction associated with prescription opioids (Christie et al., 2017; Kolodny et al., 2015; Alpert et al., 2022; Arteaga & Barone, 2022). Despite the severe consequences associated with opioid abuse, the public and even healthcare professionals had limited awareness of the full extent of the risks associated with opioid use until recently (Van Zee, 2009; Keefe, 2017). For example, it was not until 2017, when the Department of Health and Human Services declared it a public health emergency that regulators began acknowledging the magnitude of the threat to public health. Prior to that, the Food and Drug Administration's attempts to remove false claims from OxyContin labeling in 2001 were met with reluctance. This lack of oversight and awareness may suggest that the extent of opioid abuse may be difficult for managers to confidently detect and has been driven largely by factors over which managers have limited influence, reducing reverse causality concerns.

Recent pharmaceutical studies find that employees affected by opioids tend to suffer drowsiness, mental confusion, difficulty concentrating, and impaired cognitive skill and judgement, which can lead to poor work performance and substantial work absenteeism (Ruetsch, 2010; Goplerud et al., 2017). Because the development of management forecasts requires intensive coordination of employees at a firm’s headquarter to gather, process and analyze information, the quality of this work is likely to be impaired if employees are under the influence of opioids or are absent from their work. Moreover, because affected firms are unlikely to find perfect short-term substitutes for absent employees without cost, absenteeism by opioid-affected employees is likely to increase the workload of the ‘healthy’ employees, reducing the quality of their work (Krol et al., 2012; Morris & Hoitash, 2018). We therefore expect that employees of firms headquartered in high opioid abuse areas are on average less capable of exercising their professional expertise in producing the information that underpins management forecasts, thereby decreasing the quality of forecasts.

Opioid abuse in a firm’s headquarter county is proxied by the county-level opioid prescription rate, which has been shown to be a direct driver of the opioid abuse epidemic (Cicero et al., 2007; Wisniewski et al., 2008; Modarai et al., 2013; Harris et al., 2020; Rietveld & Patel, 2021; Finkelstein & Gentzkow, 2021). For example, around a third of patients treated with opioids become dependent or addicted to their prescriptions (Juurlink & Dhalla, 2012). Among the 10.1 million people estimated to have misused opioid-related drugs in 2018, 9.7 million misused prescription opioids (SAMHSA, 2019). Moreover, research shows that a substantial proportion of prescribed opioids are diverted to friends, family, or illegal markets (Surratt et al., 2014; Compton et al., 2015; Shei et al., 2015; Lipari & Hughes, 2017; Harris et al., 2020). These diverted pharmaceuticals are typically consumed in the local community, leading to a positive relationship between opioid prescription rates and opioid abuse in each geographic area.

We analyze a sample of 50,314 annual management earnings forecasts from 2006 to 2019. After controlling for an array of variables reported by previous studies to affect firms’ forecast quality, we find that local opioid abuse is associated with significantly lower management earnings forecast accuracy. Our results are both statistically and economically significant. A one-standard-deviation increase in local opioid prescription rate is, on average, associated with about an 8.22% standard deviation decrease of management earnings forecast accuracy. Our results continue to hold after using three alternative measures of opioid abuse, or using various approaches to address omitted variable bias including the procedure proposed by Oster (2019), placebo analysis, propensity score matching, and entropy balancing (Hainmueller, 2012). To further strengthen the identification, we exploit the staggered implementation of state-level Prescription Drug Monitoring Programs (PDMPs) as an exogenous shock that decreases the opioid prescriptions in the affected states. We find consistent evidence that the adoption of these programs significantly reduces local opioid abuse and increases firms’ forecasting quality.

To deepen our analysis, we conduct several cross-sectional tests. First, we expect and find that the negative impact of opioids on forecast quality is more significant for firms facing higher forecast difficulty, which we proxied by analyst forecast dispersion and stock return volatility. Second, improved access to opioid treatment facilities can mitigate the adverse effects of opioid abuse (Clark et al., 2011; Bondurant et al., 2018; Haley et al., 2019). We expect that employees with easier access to such centers recover more quickly, reducing the impact on their contribution to firms’ forecast production. In line with this, we find that firms located in areas with fewer opioid treatment facilities experience a more pronounced negative impact on forecast quality due to the opioid crisis. Third, firms with stronger environmental, social, and governance (ESG) commitments are known to mitigate the depreciation of employee health capital (Grossman, 1972; Holland, 2017; Gubler et al., 2018). Given the negative impact of opioid abuse on employee health, we expect that firms with stronger ESG commitments would be less susceptible to the effects of local opioid abuse. Our results support this claim, indicating that the effect of employee opioid abuse on forecast quality is less pronounced in firms demonstrating a greater commitment to employee health and having established a corporate social responsibility (CSR) committee.⁴

To investigate alternative explanations for our findings, we perform a battery of additional tests. One potential concern is that the effect of opioid abuse on firms’ fundamentals might be driving the observed relationship between opioid abuse and forecast accuracy (Billings et al., 2015; Nagar et al., 2019). To investigate this, we follow Chen et al. (2023) and analyze cases where most employees are located outside the headquarters region because the threat to our main results would be greatest in cases where most of the firm’s employees are located in the headquarters region. We further restrict this sub-sample by requiring the cases to have a high difference in the level of opioid prescriptions between the headquarters county and all other counties that the firm operates. We find that the negative impact of opioids on forecast quality persists for this subsample, suggesting that the effect of opioid abuse on firms’ fundamentals is unlikely to be driving our main results. To address concerns about potential confounding factors such as employee education or the firm’s corporate governance quality, we conduct two sets of tests. First, we find that our main results remain robust after controlling for employee education, internal control weaknesses, and board independence in our regression analysis. Second, these variables do not significantly affect the impact of opioid abuse on management forecast accuracy. Additionally, we consider the potential impact of opioid abuse on individuals’ emotions (Scherrer et al., 2016; Nummenmaa & Tuominen, 2018) by including controls related to moods and sentiment based on prior studies (Dehaan et al., 2017; Chen, Cheng, et al., 2021; Cuculiza et al., 2021; Chen et al., 2022). Our results remain unchanged.

We also explore how employee opioid abuse impacts other aspects of corporate information production. We find no evidence suggesting that opioid abuse affects the likelihood of management forecast issuance. Meanwhile, we find a significant positive association between opioid abuse and firms’ financial reporting lag (i.e., the time elapsed between each fiscal quarter end and the subsequent earnings announcement). This increase in reporting delay may reflect the less efficient processing of historic data needed to forecast future earnings or to confirm forecasting assumptions to allow the continued provision of forecasts, which are typically bundled with earnings announcements and the cessation of which may impose high costs on managers (e.g., Houston et al., 2010). We further find that employee opioid abuse decreases firms’ forecasting precision, suggesting that managers perceive the information production constraints induced by opioids and lower the specificity of the forecasts to mitigate the adverse effect associated with inaccurate forecasts. Finally, we explore whether investors recognize the adverse effect of employee opioid abuse on the reliability of firms’ forecasts. We find that the management forecasts issued by firms located in high opioid areas trigger weaker stock market reactions, suggesting that investors assess management forecasts in high opioid areas as less informative.

We contribute to the literature in several ways. First, our study contributes to the literature on the determinants of corporate information production, more specifically, management forecast quality. Prior studies find that other aspects of employee human capital, such as education and experience, affect the firms’ ability to produce high-quality information (Call et al., 2017; Chen, Wu, et al., 2021). We show that employee health exerts a unique and economically significant influence on the quality of firms’ information production, incremental to other dimensions of human capital.

Second, we contribute to the literature investigating the association between ESG-related issues and financial reporting outcomes. Prior research suggests that firms’ commitment to ESG can improve the integrity and soundness of corporate accounting practices. For example, Kim et al. (2012) finds that socially responsible firms are less likely to manage earnings, thereby delivering more transparent and reliable financial information to investors. Guo et al. (2016) provides evidence that employee-friendly policies limits material weaknesses in internal control and reduce financial restatements. In line with this, we show that firms’ commitment to ESG may help alleviate the negative effect of opioid abuse on corporate information production. Our findings suggest that firms should place greater emphasis on employee health in managing human resources.

Finally, we contribute to the emerging health economics literature that examines the economic consequences of opioid abuse. Research in this area mainly focuses on the significant adverse macroeconomic consequences of opioid abuse, including labor participation, unemployment rates and municipal finance (Krueger, 2017; Harris et al., 2020; Cornaggia et al., 2022). We add to this literature by showing that opioid abuse can also have significant economic consequences by impairing the quality of the forward-looking information provided by firms. Given the prolonged epidemic of opioid abuse in the US, our study is particularly relevant because management forecast is a key mechanism through which managers communicate with capital market.

The remainder of this paper is organized as follows. Section 2 discusses prior literature and develops hypothesis. Section 3 describes the research method and sample. Section 4 presents the main results of the study. Section 5 presents the results of cross-sectional tests. Section 6 presents the analyses addressing alternative explanations. Section 7 presents tests addressing measurement error. Section 8 presents the additional analyses, and Section 9 concludes the paper.

2. Literature Review and Hypothesis Development

2.1. Employee Health Capital

With the advent of the knowledge economy, human capital has become an essential strategic resource for businesses (McGregor, 1960; Schultz, 1961; Rajan & Zingales, 1998). Prior literature has focused on employee education and work experience when explaining the influence of human capital on corporate financial reporting. For example, Call et al. (2017) finds that the average workforce education level in the region where firms’ headquarters are located is associated with better financial reporting outcomes. Similarly, Chen, Wu, et al. (2021) shows a positive association between corporate accounting employees’ prior Big N work experience and firms’ financial reporting quality.

The literature recognizes health as the most critical and fundamental component of human capital (e.g., Mushkin, 1962; Grossman, 1972; Becker, 2007), because health directly affects an individual’s ability to engage in activities that require physical or mental exertion. For employees who experience poor health, the total amount of time they spend working and their ability to productively apply their existing knowledge and skills to employment tasks is limited (e.g., Mushkin, 1962; Grossman, 1972; Becker, 2007; Bleakley, 2010). Research suggests that poor health can have a significant negative impact on performance, regardless of an individual’s level of education or skill (Grossman, 1972; Helliwell, 2003; Becker, 2007).

Prior studies suggest that employee health capital can have significant impacts on labor market outcomes and employee performance. For example, García-Gómez et al. (2013) finds that adverse health shocks significantly decrease labor market participation and labor earnings. Morris and Hoitash (2018) finds that reductions in audit firm employee health capital caused by influenza outbreaks are associated with clients’ adverse audit outcomes. Li et al. (2020) documents that poor air quality impairs analysts’ health capital, jeopardizing their ability to provide information to capital market. Overall, these studies suggest that employee health can significantly affect their participation and productivity in the labor market.

2.2. Opioid Abuse

Opioid abuse has reached a crisis level in the US, to the extent that approximately 80 percent of the global opioid supply is consumed in this country. The Centers for Disease Control and Prevention (CDC) reports that almost 500,000 people died from opioid overdoses from 1999 to 2019.⁵ Prior studies suggest that the supply of prescription opioids is a key contributor to the crisis (Christie et al., 2017; Kolodny et al., 2015; Alpert et al., 2022; Arteaga & Barone, 2022). From the 1980s, the medical community in the US began to push for a more aggressive approach to pain treatment. Pharmaceutical companies, such as Purdue Pharma which developed OxyContin, marketed aggressively and neglected to disclose the true potential for addiction as a possible negative side effect of using opioid to treat pain. For example, Purdue widely spread the message that OxyContin was safe and non-addictive, and it spent hundreds of millions of dollars targeting hospitals and medical schools and sponsored medical education seminars to assure doctors repeatedly and without evidence that ‘fewer than one per cent’ of patients who took OxyContin became addicted (Van Zee, 2009; Keefe, 2017). Driven by pharmaceutical companies’ misleading marketing of opioid pain relievers and irresponsible drug distributors,⁶ doctors began to prescribe opioids at an astonishing rate. The number of opioid prescriptions written and dispensed grew more than threefold from 76 million in 1991 to almost 245 million in 2014, roughly equivalent to one opioid prescription per adult in the US (Volkow et al., 2014; Volkow & McLellan, 2016). This subsequently led to widespread diversion of opioid prescriptions, and abuse, overdose, addictions, and death relating to these products (Modarai et al., 2013; NIH, 2019). In fact, until recently, the general public and even healthcare professionals had limited awareness of the full extent of the risks associated with opioid use. For example, despite the Food and Drug Administration's efforts to remove false claims from OxyContin labeling in 2001, regulators were reluctant to acknowledge the magnitude of the threat to public health until 2017 when the Department of Health and Human Services declared it a public health emergency.

Extant evidence suggests that opioid abuse can have a significant impact on employees in the workplace. For example, in a recent survey by the National Safety Council (2019), 75% of employers reported that they have been affected by opioid misuse and abuse.⁷ Prior studies suggest that opioid abuse significantly impairs employee health, decreasing employees’ abilities to discharge their work responsibilities (Ruetsch, 2010; Goplerud et al., 2017). Consistent with this, Harris et al. (2020) finds that opioid abuse has significant adverse effects on labor force participation and unemployment rates. Ouimet et al. (2019) finds a negative effect of opioid prescriptions on the subsequent local employment. They also suggest that firms respond to opioid crisis by making production processes more capital-intensive.

2.3. Opioid Abuse and Firms’ Forecasting Quality

The literature recognizes that opioid abuse can significantly reduce employee health capital, by inducing drowsiness, mental confusion, and difficulty in concentrating, and impairing cognitive skill and judgement (Birnbaum et al., 2011; Kuhn, 2017). This can diminish employees’ capability to work, resulting in poor vocational performance and substantial work absenteeism (e.g., Ruetsch, 2010; Goplerud et al., 2017; Harris et al., 2020).

Earnings forecasting is a labor-intensive exercise, which requires information planning, analysis and compilation by employees in a firm’s headquarter, such that the input of any one employee is likely to affect the quality of work performed by others in a complementary fashion (Suphap, 2003; Sherin, 2010; Chourou et al., 2020; Chen et al., 2023). To the extent that forecasting team members are affected by opioids, the quality of the earnings forecasts produced is likely to be impaired. Moreover, as firms may lack the resources or the capability to find suitable replacements for the absent employees, the remaining healthy employees may need to take on extra work. This extra workload can increase employee fatigue, resulting in lower quality of work outputs (Pauly et al., 2002; Nicholson et al., 2006). We thus expect that in regions where opioid abuse is more prevalent, employees engaged in firms’ forecasting process are more likely to suffer from this affliction than similar employees in areas where opioid abuse is less common, which in turn decreases the information quality in the forecasting process. We thus state our main hypothesis as follows.

HYPOTHESIS: The intensity of opioid abuse in a firm’s headquarter county is negatively associated with management earnings forecast quality.

3. Research Design

3.1. Sample Construction

Our initial sample consists of all annual earnings per share (EPS) management forecasts in the I/B/E/S Guidance database from 2006 to 2019.⁸ We exclude forecasts issued on or after the corresponding fiscal period-end to eliminate earnings preannouncements. We then merge the forecasts with opioid prescription data from the CDC website, using historical county of firm headquarters information from 10 K header data provided by Bill McDonald.⁹ Finally, we merge all control variables calculated using firm financial information, stock returns, analyst and institutional shareholder information, and geographic demographic information from Compustat, CRSP, I/B/E/S, Thomson Reuters 13F, and US Census Bureau, respectively. All control variables are measured at the latest fiscal quarter end prior to the forecast date. These procedures leave us with a final sample of 50,314 forecast-level observations.

3.2. Measuring Opioid Abuse

The extent of opioid abuse in a firm’s workforce is not directly observable. Therefore, we follow prior studies (Cicero et al., 2007; Harris et al., 2020; Ouimet et al., 2019) and rely on county-level annual opioid prescription data provided by the CDC to proxy for the level of employee opioid abuse. Our variable of interest, Opioid, is measured as the total number of opioid prescriptions per 100 people in the county. A higher value of Opioid indicates more severe opioid abuse in a firm’s headquarters county, such that employees are more likely to be affected when they are generating financial information.

3.3. Measuring Management Earnings Forecast Quality

The primary outcome variable is the accuracy of firms’ annual management earnings forecasts (Accuracy), which is defined as the absolute value of the difference between the earnings forecast (either the point estimate or the midpoint of the range estimate) and actual earnings per share, multiplied by minus 100 and scaled by the stock price at the end of prior fiscal quarter (Ke et al., 2019). Larger (smaller) values of Accuracy correspond to more (less) accurate forecasts.

3.4. Regression Model

To test the effect of opioid abuse on the quality of management earnings forecasts, we estimate the following regression model: (1) $\mathrm{Accurac}{y}_{i,t}={\text{β}}_1\mathrm{Opioi}{d}_{c,t}+\mathrm{Control}{s}_{i,t}+\mathrm{Fixed}\mathrm{Effects}+{ϵ}_{i,t}$(1) where i, c and t denote firm, county of headquarter location, and year-quarter, respectively. We employ Accuracy as our dependent variable. A negative and significant estimate of β₁ would support our main hypothesis that opioid abuse decreases management earnings forecast quality.

Following prior research, we include a set of controls that affect management forecast quality, including firm size (Size) (Kasznik & Lev, 1995), financial leverage (Leverage) (Ali et al., 2014), market-to-book ratio (MB) (Bamber & Cheon, 1998; Chen et al., 2010), research and development expenditures (R&D) (Huang et al., 2021), an indicator for M&A in the prior quarter (Acquisition) (Ge & Lennox, 2011), quarterly stock return (Return) (Baginski et al., 2002; Miller, 2002), analyst forecast dispersion (Dispersion) (Heflin et al., 2016), analyst coverage (Analyst) (Ajinkya et al., 2005; Kross et al., 2011), institutional shareholding (IOR) (Ajinkya et al., 2005), and forecast horizon (Horizon), which is the natural logarithm of the number of days between the forecast date and the ending date of the fiscal period that the management earnings forecast pertains to (Ajinkya et al., 2005). We also include controls for local demographic variables, including Population (the natural log of the county population), Unemployment (the unemployment rate at the county level), and Income (median household income at the county level). All continuous variables are winsorized at the 1st and 99th percentiles. Appendix A presents detailed definitions for the variables. We include firm fixed effects to control for time-invariant firm characteristics that may affect forecast behavior. Following a recent study on opioid crisis by Cornaggia et al. (2022), we also include headquarter state × year-quarter fixed effects to control for unobserved time-varying heterogeneity within the states of headquarters that could affect both opioid abuse level and management forecasts.¹⁰ We adjust standard errors for clustering at the headquarter state × year-quarter level.

4. Empirical Results

4.1. Descriptive Statistics

Table 1 presents descriptive statistics for our sample. On average, the absolute value of management forecast errors is 1.097% of lagged share price. Our sample firms have mean (median) total assets of $2,847 million ($2,757 million), a mean (median) market-to-book ratio of 3.587 (2.597), and a mean (median) leverage ratio of 0.219 (0.209). The average number of analysts covering a firm is 9.2 and the institutions own 78.3% of the shares in an average sample firm. Around 45.5% of firms have acquisition-related cash flows. The average quarterly return is 3.2%.

Table 1. Descriptive statistics.

Descriptive Statistics (n = 50,314)
Variable	Mean	SD	P25	Median	P75
Accuracy	−1.097	3.099	−0.777	−0.277	−0.095
Opioid	63.008	25.484	44.100	57.300	77.500
Size	7.954	1.715	6.746	7.922	9.090
Leverage	0.219	0.174	0.074	0.209	0.325
MB	3.587	4.709	1.648	2.597	4.200
R&D	0.007	0.013	0.000	0.000	0.008
Acquisition	0.455	0.498	0.000	0.000	1.000
Return	0.032	0.176	−0.064	0.032	0.125
Dispersion	0.068	0.214	0.012	0.023	0.050
Analyst	2.217	0.616	1.792	2.303	2.708
IOR	0.783	0.208	0.693	0.834	0.931
Horizon	5.126	0.723	4.700	5.403	5.724
Population	13.051	1.013	12.533	13.074	13.659
Unemployment	5.949	2.248	4.200	5.300	7.400
Income	57390.700	22164.450	43161.000	51349.000	63853.000

Notes: This table reports the distribution of variables in the final sample used in our main test. See Appendix A for variable definitions.

The average opioid prescription rate for the county in which firms are headquartered in is 63.008, suggesting that there were around 63 opioid prescriptions issued to every 100 people each year during our sample period. We find considerable variation in opioid prescription levels across counties over time. For example, opioid prescription rates in Middlesex, MA (comprising most of Boston) fell from 53.1 (per 100 people) in 2006–24.1 in 2019, while New York County, NY (i.e., Manhattan) saw an increase in opioid prescriptions rates from 40.7 to 63.2 over the same period. Figure 1 shows the spatial distribution of the opioid crisis across counties throughout the US Because the dispersion in opioid within state can be large, our observations are at the county level, thus preserving the granularity which is a hallmark of this crisis.

Figure 1. Spatial distribution of opioid.

4.2. Main Regression Results

The results of regressions based on Eq. (1) are reported in Table 2. In column (1), we exclude all time-varying and firm-specific control variables to alleviate the concern of inconsistent estimates (Gormley & Matsa, 2014), while in column (2) we include all the firm-quarter level control variables as defined in Section 3.4. In both columns, the coefficient for Opioid is negative and significant (p < 0.01). In terms of economic significance, a one-standard-deviation increase in Opioid is, on average, associated with an 8.22% standard deviation decrease in Accuracy.¹¹ In terms of control variables, Leverage, Horizon and Dispersion are negatively associated with Accuracy, while the coefficient on MB is positive. These results are consistent with findings in prior studies (e.g., Call et al., 2017). Overall, these findings are consistent with our prediction that opioid abuse impedes employees’ forecasting ability, decreasing management earnings forecasts quality.¹²

Table 2. Baseline results.

	(1)	(2)
Dep. Var. =	Accuracy	Accuracy
Opioid	−0.008***	−0.010***
	(−3.05)	(−3.72)
Size		−0.024
		(−0.50)
Leverage		−1.114***
		(−6.38)
MB		0.021***
		(7.38)
R&D		−0.708
		(−0.36)
Acquisition		−0.002
		(−0.08)
Return		0.905***
		(8.98)
Dispersion		−1.015***
		(−7.54)
Analyst		0.044
		(1.14)
IOR		0.484***
		(3.63)
Horizon		−0.538***
		(−19.71)
Population		−0.179**
		(−2.04)
Unemployment		−0.013
		(−0.42)
Income		−0.000***
		(−3.93)
Firm FE	Yes	Yes
State × YQ FE	Yes	Yes
Adj. R^2	0.583	0.597
N	50,314	50,314

Notes: This table presents the regression results of main tests. All continuous variables are winsorized at the top and bottom percentile. *, **, and *** indicate that the estimated coefficient is statistically significant at the 10%, 5% and 1% levels, respectively. See Appendix A for detailed variable definitions.

We further explore whether our main findings could be explained by alternative explanations. First, one may be concerned that our main results could be driven by the effect of opioid abuse on actual firm performance. We follow Chen et al. (2023) and partition our sample into two groups based on whether the majority of a firm’s business units or employees are located outside its headquarter county. Our main findings hold regardless of the proportion of employees based in the headquarters region. We further restrict the sub-sample with lower proportion of employees based in headquarters region by requiring the cases to have a high difference in the level of opioid prescriptions between the headquarters county and all other counties that the firm operates. We find that the negative impact of opioids on forecast quality persists for this subsample, suggesting that the effect of opioid abuse on firms’ fundamentals is unlikely to be driving our main results. Second, we show that our main results are robust to controlling for the effect of employee education, internal control weaknesses, and board independence. Third, we show that our main results are unlikely to be driven by the effect of opioid abuse on individuals’ emotions. Further discussion and tabulation of results of these additional tests are provided in Online Appendix A1.

4.3. Difference-in-differences Analyses

Next, to address the concern that firms with high forecasts’ quality may self-select to base themselves in counties with low opioid abuse, we exploit the staggered implementation of state-level PDMPs designed to combat opioid abuse. These state-level PDMP databases collect and track controlled opioid prescriptions and provide this data to prescribers and dispensers. Prior studies suggest that PDMPs can help to identify individuals who may be misusing prescription opioids, resulting in lower opioid diversions and fewer opioid pills being prescribed (e.g., Surratt et al., 2014; Buchmueller & Carey, 2018; Winstanley et al., 2018). We expect that, by limiting the supply of opioids available for nonmedical diversion, the adoption of PDMP should reduce opioid abuse and, in turn, improve management forecast quality.

We first establish whether PDMP adoptions affect opioid abuse before testing the potential effect of PDMPs on management forecast quality. We follow the methodology suggested by Baker et al. (2022) and implement a stacked difference-in-differences regression to test the impact of PDMP implementation. Specifically, for each event year with PDMP adoption, we obtain a cohort of counties that are within the state adopting the PDMP in that year (treated counties) and those that are not within any states adopting the PDMP over a six-year time window around the event year (control counties). We then pool the data across different cohorts to calculate an average effect across all events using a single set of treatment indicators. Specifically, we estimate the treatment effect of PDMPs on the county-level opioid prescription rate with the following model: (2) $\mathrm{Opioi}{d}_{\mathrm{cyg}}={\text{β}}_1\mathrm{PDM}{P_{\mathrm{sy}}}_g+\mathrm{Control}{s}_{\mathrm{cyg}}+\mathrm{Fixed}\mathrm{Effect}{s}_g+{ϵ}_{\mathrm{cyg}}$(2) where s, c, y and g denote state, county, year and cohort, respectively. The dependent variable is the county-level opioid prescription rate defined in the main test. PDMP is an indicator variable equal to one for the 3-year period after the PDMP adoption year in state s (i.e., years t to t + 2), and zero for the 3-year period before the adoption year (i.e., years t − 3 to t − 1) and for the control states that have not adopted the PDMP during these six years surrounding the adoption year in state s (i.e., years t − 3 to t + 2).¹³ We include several county-level demographic controls, including Population (natural log of county population), Unemployment (the unemployment rate at the county level), and Income (median household income at the county level). We include state and year fixed effects to absorb variations across state and year. Standard errors are clustered at the state × year level. Column (1) of Table 3 shows the results. We find that the coefficient on PDMP is negative and significant (p < 0.01), suggesting that the adoption of PDMPs significantly reduce county-level opioid abuse, which is consistent with prior pharmaceutical studies (e.g., Ellyson et al., 2022).

Table 3. Difference-in-differences tests.

	Opioid	Accuracy
Dep. Var. =	(1)	(2)
PDMP	−4.177***	0.066*
	(−4.53)	(1.73)
Population	7.859***	0.030
	(45.61)	(0.67)
Unemployment	1.307***	0.023**
	(11.08)	(2.10)
Income	−0.001***	−0.000
	(−35.75)	(−1.15)
Size		−0.006
		(−0.14)
Leverage		−0.944***
		(−7.51)
MB		0.020***
		(10.79)
R&D		−1.686
		(−1.37)
Acquisition		−0.014
		(−0.98)
Return		0.805***
		(15.30)
Dispersion		−0.658***
		(−8.83)
Analyst		0.034
		(1.46)
IOR		0.647***
		(6.95)
Horizon		−0.556***
		(−35.06)
Cohort FE	Yes	Yes
State FE	Yes	No
Year FE	Yes	No
Firm FE	No	Yes
YQ FE	No	Yes
Adj. R^2	0.251	0.636
N	130,664	148,388

Notes: This table presents the regression results of difference-in-differences regressions. All continuous variables are winsorized at the top and bottom percentile. *, **, and *** indicate that the estimated coefficient is statistically significant at the 10%, 5% and 1% levels, respectively. See Appendix A for detailed variable definitions.

After establishing the link between PDMP implementation and the subsequent reduction in opioid abuse, we next explore the effect of PDMPs on management forecasts quality. Specifically, we estimate the following model using a stacked difference-in-differences regression: (3) $\mathrm{Accurac}{y}_{\mathrm{itg}}={\text{β}}_1\mathrm{PDM}{P_{\mathrm{st}}}_g+\mathrm{Control}{s}_{\mathrm{itg}}+\mathrm{Fixed}\mathrm{Effect}{s}_g+{ϵ}_{\mathrm{itg}}$(3) where s, i, t and g denote state, firm, year-quarter and cohort, respectively. We employ management forecasts accuracy as our dependent variable. We include the same set of controls we used in our main test as defined in Eq. (1). In column (2) of Table 3, the coefficient for PDMP indicates a significant increase in management forecast accuracy after the adoption of PDMP in the firm headquarter state. These results support our earlier attribution of decreased forecasts quality to opioid abuse.

5. Cross-sectional Analyses

Below we conduct an array of cross-sectional tests designed to sharpen identification and to improve understanding of our main findings.

5.1. Firm Forecast Difficulty

The uncertainty of firms’ operating environment has been found to be associated with a greater need for internal information processing and production (Gordon & Narayanan, 1984). Information production in firms with greater uncertainty requires more employee time and effort than is the case for other firms. Consequently, forecasting future earnings becomes more challenging for such firms. Thus, we expect that the negative impact of opioid abuse should be more pronounced for those firms with higher forecast difficulty.

To test this prediction, we use analyst forecast dispersion (Dispersion) and stock return volatility (RetSd) to proxy for firms’ forecast difficulty. We partition our sample into two sub-samples, based on the median value of Dispersion and RetSd, respectively. We re-estimate our baseline regression separately for the two subsamples. Table 4 reports the estimation results. The coefficient on Opioid is only negative and significant in the subsample with higher than median analysts’ dispersion and the subsample with higher than median return volatility (column (2) and (4)). We do not find any significant association between the extent of the opioid abuse and forecasts quality in the subsample with lower-than-median analysts’ dispersion and the subsample with lower-than-median return volatility (column (1) and (3), respectively). In addition, the difference between the coefficients for Opioid between column (1) and (2), as well as column (3) and (4), are significant at the one percent level. Consistent with our prediction, the results indicate that the negative impact of opioid abuse on forecast quality is more pronounced in firms subject to higher forecast difficulty.

Table 4. Cross-sectional tests.

Forecast Difficulty
	Accuracy
	(1)	(2)	(3)	(4)
Dep. Var. =	Low Analysts’ Dispersion	High Analysts’ Dispersion	Low Return Volatility	High Return Volatility
Opioid	−0.001	−0.018***	−0.001	−0.021***
	(−0.56)	(−3.41)	(−0.44)	(−3.75)
Difference in Opioid	0.017***		0.020***
	(p value =0.005)		(p value =0.000)
Size	−0.094**	−0.002	−0.122***	0.057
	(−2.21)	(−0.03)	(−3.44)	(0.68)
Leverage	−0.467**	−1.689***	0.049	−1.948***
	(−2.36)	(−5.98)	(0.37)	(−6.72)
MB	0.007***	0.037***	0.004***	0.039***
	(3.93)	(5.91)	(3.46)	(6.37)
R&D	−0.761	0.456	−1.055	−0.677
	(−0.33)	(0.14)	(−0.78)	(−0.19)
Acquisition	0.013	−0.012	0.016	−0.038
	(0.49)	(−0.26)	(0.87)	(−0.80)
Return	0.660***	1.004***	0.468***	0.926***
	(5.31)	(7.45)	(5.16)	(7.27)
Dispersion	0.316	−0.900***	−0.576	−0.981***
	(0.17)	(−6.41)	(−1.15)	(−6.67)
Analyst	0.062*	0.027	−0.044	0.123
	(1.83)	(0.32)	(−1.64)	(1.59)
IOR	0.281***	0.654***	0.120	0.697***
	(2.88)	(2.88)	(1.15)	(3.10)
Horizon	−0.342***	−0.763***	−0.319***	−0.775***
	(−14.58)	(−14.79)	(−15.47)	(−13.68)
Population	−0.048	−0.340**	−0.054	−0.266*
	(−0.40)	(−2.37)	(−0.44)	(−1.95)
Unemployment	−0.018	−0.028	−0.018	0.001
	(−0.69)	(−0.46)	(−0.88)	(0.02)
Income	−0.000**	−0.000***	−0.000**	−0.000**
	(−2.09)	(−2.70)	(−2.30)	(−2.21)
Firm FE	Yes	Yes	Yes	Yes
State × YQ FE	Yes	Yes	Yes	Yes
Adj. R^2	0.612	0.598	0.657	0.592
N	24,718	24,815	24,764	24,778

Notes: This table presents the regression results of cross-sectional tests. The sample size varies because singleton observations are dropped when estimating the regression separately for the subsamples. All continuous variables are winsorized at the top and bottom percentile. *, **, and *** indicate that the estimated coefficient is statistically significant at the 10%, 5% and 1% levels, respectively. See Appendix A for detailed variable definitions.

5.2. Mitigating the Adverse Effects of Opioid Abuse

5.2.1. Treatment center access

Opioid abuse involves both physical and psychological dependence. As such, recovery from opioid abuse disorder usually requires intensive treatment. There are different treatment modalities, including methadone treatment, cognitive behavioral therapy and group counselling, but all typically involve frequent visits to a facility, which suggests that access to treatment facilities may be especially important for the success of opioid abuse treatment. Recent work indicates that enhancing access to treatment facilities can reduce the adverse effect associated with drug use disorder (Clark et al., 2011; Swensen, 2015; Bondurant et al., 2018; Haley et al., 2019; Corredor-Waldron & Currie, 2022). Thus, we expect that employees with easier access to opioid treatment centers will be affected less by the opioid abuse.

To test this prediction, we use the number of opioid treatment centers within a 10 km radius of corporate headquarters (Access) as a proxy for treatment center access. There are 1,761 opioid treatment centers across the US. The distance between the opioid treatment centers and corporate headquarters is calculated as a straight-line from the firm headquarters address provided by Bill McDonald’s 10 K header data to the exact location of treatment facilities obtained from SAMHSA website.¹⁴ We partition our sample into ‘High Treatment’ and ‘Low Treatment’ subsamples, based on the median value of Access in our sample. We re-estimate our baseline regression separately for the two subsamples. As shown in column (1) and (2) of Table 5, the coefficient on Opioid is only negative and significant in the ‘Low Treatment’ subsample (p < 0.01). We find no evidence of a significant association between opioid abuse and forecasts quality in the ‘High Treatment’ subsample. The difference between the coefficients for Opioid across the subsamples is significant at the five percent level. Consistent with our prediction, the results indicate that the negative impact of opioid abuse on forecast quality is more pronounced in firms whose employees have less access to treatment facilities.

Table 5. Cross-sectional tests.

Mitigating Adverse Effect of Opioid
	Accuracy
	(1)	(2)	(3)	(4)	(5)	(6)
Dep. Var. =	Low Treatment	High Treatment	Low Employee Welfare	High Employee Welfare	Without CSR Committee	With CSR Committee
Opioid	−0.010***	0.002	−0.025**	−0.003	−0.006*	0.004
	(−2.90)	(0.68)	(−2.25)	(−0.59)	(−1.73)	(1.51)
Difference in Opioid	−0.012**		−0.022*		−0.010**
	(p value =0.013)		(p value =0.083)		(p value =0.022)
Size	0.005	−0.009	−0.061	0.051	0.065	0.046
	(0.07)	(−0.15)	(−0.45)	(0.54)	(0.78)	(0.45)
Leverage	−1.309***	−0.730***	−0.835*	−0.303	−0.505	0.171
	(−5.61)	(−3.02)	(−1.96)	(−1.28)	(−1.65)	(0.68)
MB	0.029***	0.015***	0.015***	0.000	0.011**	0.005
	(6.24)	(3.84)	(3.01)	(0.07)	(2.17)	(1.08)
R&D	−2.139	−3.199	0.313	1.805	−0.398	0.817
	(−0.78)	(−1.30)	(0.09)	(0.65)	(−0.17)	(0.31)
Acquisition	0.044	−0.068*	0.042	−0.024	0.055	−0.085**
	(1.26)	(−1.81)	(0.87)	(−0.51)	(1.42)	(−2.57)
Return	0.996***	0.630***	0.509***	−0.019	0.534***	0.394***
	(7.20)	(4.41)	(3.77)	(−0.08)	(3.55)	(2.64)
Dispersion	−1.033***	−0.956***	−1.601***	−0.545	−1.232***	−0.275
	(−6.05)	(−4.57)	(−3.03)	(−1.38)	(−3.70)	(−0.75)
Analyst	0.109**	0.059	0.045	−0.150**	0.023	−0.074
	(2.03)	(1.02)	(0.87)	(−2.22)	(0.48)	(−1.46)
IOR	0.436**	0.552**	−0.104	1.145**	−0.090	1.210***
	(2.53)	(2.35)	(−0.75)	(2.35)	(−0.97)	(2.82)
Horizon	−0.550***	−0.522***	−0.325***	−0.384***	−0.339***	−0.358***
	(−14.33)	(−12.61)	(−12.75)	(−9.55)	(−12.41)	(−11.32)
Population	−0.025	0.679***	−0.148	0.235	−0.080	0.441**
	(−0.28)	(3.13)	(−0.85)	(1.56)	(−0.80)	(2.41)
Unemployment	−0.045	0.038	0.092	−0.009	−0.013	0.017
	(−0.95)	(0.88)	(1.59)	(−0.19)	(−0.26)	(0.53)
Income	−0.000***	0.000	−0.000	−0.000	−0.000*	0.000*
	(−4.15)	(0.93)	(−0.43)	(−0.70)	(−1.74)	(1.73)
Firm FE	Yes	Yes	Yes	Yes	Yes	Yes
State × YQ FE	Yes	Yes	Yes	Yes	Yes	Yes
Adj. R^2	0.613	0.598	0.505	0.470	0.425	0.443
N	29,449	20,547	11,100	11,160	15,939	12,380

Notes: This table presents the regression results of cross-sectional tests. The sample size varies because we remove singleton observations and data on workforce score and CSR committee information is not available for all observations. All continuous variables are winsorized at the top and bottom percentile. *, **, and *** indicate that the estimated coefficient is statistically significant at the 10%, 5% and 1% levels, respectively. See Appendix A for detailed variable definitions.

5.2.2. Firms’ commitment to ESG

Employee health capital is a depreciating asset. Opioid abuse can have a significant negative impact of employee health capital (Birnbaum et al., 2011; Kuhn, 2017). When employees are under the influence of opioids, they may be less productive or even absent, which leads to lower quality of forecasts information. Prior studies suggest that firms’ commitments to ESG may mitigate the depreciation in employee health capital (Grossman, 1972; Holland, 2017; Gubler et al., 2018). We thus expect firms that are more committed to ESG are less likely to be affected by opioid abuse.

To test this prediction, we use the workforce score (which captures a firm’s effectiveness in providing employee job satisfaction and a healthy and safe workplace) reported by Thomson Reuters ASSET4 database (Employee_Welfare) and firms’ establishment of a CSR committee (CSR_Committee) as proxies for a firm’s commitments to ESG. We partition our sample based on the median value of Employee_Welfare and CSR_Committee, respectively. We re-estimate our baseline regression separately for the two set of subsamples. As shown in column (3) – column (6) of Table 5, the coefficient on Opioid is only negative and significant in the subsample with lower than median Employee_Welfare score and the subsample without a CSR committee. There is no significant association between local opioid abuse and managerial forecast quality for the subsample with higher than median Employee_Welfare score and the subsample with a CSR committee. The difference between the coefficients for Opioid between column (3) and (4), as well as column (5) and (6) are significant at the five percent level. Consistent with our prediction, the results indicate that the negative impact of opioid abuse on forecast quality is more pronounced in firms that are less committed to ESG.

6. Measurement Error

6.1. Alternative Measures of Opioid Abuse

Like most empirical proxies for unobservable underlying constructs, our proxy of opioid abuse is subject to measurement error. First, large measurement errors could exist because the county-level opioid abuse of a firm’s headquarters may deviate from the actual firm-level employee opioid abuse activities. To alleviate this concern, we employ a more granular measure of opioid based on the ZIP-level opioid pills distribution data for each year quarter from 2006 to 2014. We obtain the data from the DEA database, which tracks every opioid pill sold in the US and reports the ZIP code of each ultimate retail distributor.¹⁵ We define our first alternative measure, Opioid_Pill, as the total opioid pills sold in each ZIP code and year quarter scaled by population.¹⁶ We merge Opioid_Pill to historical ZIP code of firm headquarters information from 10K header data provided by Bill McDonald to estimate the effect of opioid abuse on firms’ forecasts quality. Column (1) of Table 6 Panel A shows that the results remain unchanged using this alternative measure of opioid abuse. Second, one may concern that our primary measure (the local opioid prescription rate) may not fully capture the realized (ex post) adverse effect of opioid abuse on employees. To address this concern, we use year-quarterly opioid-related emergency department visit rates in a firm’s headquarters state, Opioid_EDvisits, as an ex post measure of opioid abuse. We obtain the data from the Healthcare Cost and Utilization Project (HCUP) administered by the Agency for Healthcare Research and Quality. Consistent with the main results, we continue to observe that exposure to opioid abuse results in lower management forecasts accuracy in column (2) of Table 6 Panel A. These results lend further support to our main findings of the negative impact of opioid abuse on management forecasts quality. Third, it is possible that firms conduct operations across several counties and employees outside the firms’ headquarters may also affect corporate information production. We follow Call et al. (2017) and develop a combined measure of opioid abuse that considers all the counties in which the firm operates. Specifically, Combined_Opioid is measured as the average of opioid abuse level across all counties mentioned in the firm’s 10-K filings. We replace Opioid with Combined_Opioid and re-estimate our baseline regression. As shown in column (3) of Table 6 Panel A, we find consistent results using the combined measure of opioid abuse.

Table 6. Measurement errors.

Panel A. Alternative measures of opioid
	Accuracy
Dep. Var. =	(1)	(2)	(3)
Opioid_Pill	−0.183**
	(−2.33)
Opioid_EDvisits		−0.001***
		(−3.05)
Opioid_Combined			−0.003***
			(−2.92)
Controls	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes
State × YQ FE	Yes	No	Yes
YQ FE	No	Yes	No
Adj. R^2	0.644	0.593	0.597
N	24,871	38,737	50,314
Panel B. Omitted variables threats
	Accuracy
	(1)	(2)	(3)
Dep. Var. =	Placebo Opioid	PSM	Entropy Balancing
Opioid_Placebo	0.000
	(1.18)
Opioid_High		−0.155**
		(−2.25)
Opioid			−0.011***
			(−3.90)
Controls	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes
State × YQ FE	Yes	Yes	Yes
Adj. R^2	0.597	0.580	0.598
N	50,314	29,882	50,314

Notes: This table presents the results of tests addressing measurement errors of opioid. Panel A reports the results using three alternative measures of opioid abuse. Panel B reports the results for analyses addressing omitted variable threats. Column (1) of Panel B reports the results obtained using the opioid abuse level in a randomly selected placebo county outside the focal firm's headquarters county. Column (2) of Panel B displays results derived from a propensity score matched sample. Column (3) in Panel B demonstrates findings generated through the entropy balancing approach. All continuous variables are winsorized at the top and bottom percentile. *, **, and *** indicate that the estimated coefficient is statistically significant at the 10%, 5% and 1% levels, respectively. See Appendix A for detailed variable definitions.

6.2. Omitted Variable Threats

We conduct a number of analyses to address potential omitted variable threats. We first assess the sensitivity of our baseline results to an unobserved correlated variable by applying the procedure in Oster (2019). Specifically, Oster (2019) evaluates the potential for the selection on unobservables by testing the sensitivity of the coefficient estimate (Opioid in our case) to the inclusion of additional controls through measuring the extent of change in R² across regression models. Oster (2019) develops a test statistic (δ*) for stability of the coefficient estimate under reasonable assumptions about the maximum attainable R², whereby the value of δ* denotes the degree of selection on unobservables relative to observables that would be necessary to explain the estimated coefficient. An estimate of δ* greater than 1 suggests that it is unlikely for the coefficient estimate to be confounded by selection on unobservables. Untabulated results show that the coefficient on Opioid passes this test of coefficient stability (δ* = 3.012), suggesting that the relation between Opioid and Accuracy in our baseline model is unlikely to be fully driven by omitted variable bias.¹⁷ Nevertheless, we also conduct multiple empirical approaches in the following subsections to further address the omitted variable issue.

6.2.1. Placebo tests

Another potential concern is that our results could be driven by events occurring outside a firm’s headquarters county, but are correlated with the opioid abuse in the firm headquarters county (e.g., a common shock across the counties). To address this concern, we conduct a placebo test by randomly selecting a county outside a focal firm’s headquarters county to calculate Opioid_Placebo. If our results are driven by unobservable social and economic characteristics in counties outside a firm’s headquarters county, we expect to observe the effect of Opioid_Placebo similar to that of Opioid. We replace Opioid in Eq. (1) with Opioid_Placebo and re-estimate the analyses. Column (1) of Table 6 Panel B reports the results. We find the coefficient on Opioid_Placebo is statistically insignificant. We further find in untabulated results that our estimation of the effect of Opioid in our main results (−0.010 in column (2) of Table 2) is smaller than the 1st percentile of the distributions from simulations on Opioid_Placebo if we repeat the above analysis 1,000 times. The results of the placebo tests further suggest that our findings are unlikely to be driven by other county-level factors.

6.2.2. Propensity score matching

We next analyze the differences in forecast quality between firms located in high opioid counties and those located in low opioid counties by employing a propensity score matching procedure. This method allows us to identify a control sample of firms that are located in low opioid areas, and exhibit no significant differences in observable characteristics relative to the treatment sample (firms located in high opioid areas). Thus, each pair of matched firms is distinguished from one another in terms of just one key characteristic: opioid abuse intensity.

To implement this methodology, we rank county-years included in our sample into two groups based on the median value of Opioid by year and match each treatment firm headquartered in the high opioid group (Opioid_High = 1) to a control firm headquartered in the low opioid group (Opioid_High = 0) that has the closest score in the same year within a distance of 0.05 from the treatment firm's propensity score (with no replacement). The propensity score is estimated within-year as a function of all firm-level characteristics employed in our baseline model. Column (2) of Table 6 Panel B reports the results using propensity score matched sample. The sample size varies because of our one-to-one matching of treatment firm to control firm with no replacement. The coefficient for Opioid_High is negative and significant, suggesting that the differences in firm level characteristics do not drive the previously documented effects of opioid on management forecasts quality.

6.2.3. Entropy balancing

We use an entropy balancing approach to further assess the effect of potential misspecification (e.g., omitted variables) in the estimation of treatment effects (Abadie & Imbens, 2011; Hainmueller, 2012). We execute this approach by first partitioning our sample into two groups, based on the median value of Opioid within each sample year (high opioid and low opioid samples), and then use the entropy balancing method to balance the first three moments (i.e., mean, variance, and skewness) of the control variables used in Eq. (1) across the two samples. Results from the estimation of our main regressions based on the reweighted variables, reported in Column (3) of Table 6 Panel B, show that the coefficient on Opioid remains negative and significant, with the magnitude of the coefficient also not deviating from that reported in our main tests in Table 2.

7. Additional Tests

7.1. Other Aspects of Corporate Information Production

In this section, we investigate how managers react to the reduction in employee health capital, and thus forecast quality, induced by the local opioid abuse. We first look at the propensity to issue annual earnings forecasts. On the one hand, if opioid abuse affects firms’ operating performance and uncertainty, firms may issue more forecasts to provide timely updates to investors due to litigation risks or greater demand from analysts. Also, ceasing the provision of earnings forecasts may harm the firm’s information environment and increase the managers’ litigation risk and reputation (e.g., Houston et al., 2010). On the other hand, opioid abuse may constrain firms’ ability to produce high quality information, reducing their incentives to issue voluntary disclosure (Call et al., 2017). We investigate this issue by regressing Issue, which is an indicator variable for issuing at least one annual earnings forecast during each fiscal quarter, on Opioid and the set of control variable as in Eq. (1). In untabulated tests, we find no evidence that opioid abuse affects the likelihood of management forecast issuance. These results mitigate the concern that our main results reported in Table 2 are subject to self-selection bias.

Next, we test whether firms take longer to report quarterly earnings and earnings forecasts when their financial employees are more likely to be affected by opioid abuse. We construct the variable Lag, defined as the natural logarithm of the number of days between the end of fiscal quarter and its upcoming actual earnings announcement date. We then replace the dependent variable in Eq. (1) with Lag, and estimate an OLS regression. Columns (1) of Table 7 reports the results. The sample size varies because we report the results of the firm-quarter-level regressions rather than forecast-level regressions we used for our main tests. The coefficient for Opioid is positive and significant, suggesting that firms take longer to report earnings when the employees are more likely to be under the influence of opioid abuse. This finding is consistent with opioid abuse constrains firms’ information processing capacity. Given that over 75% of the management forecasts in our sample period are bundled forecasts, this result also suggests that managers may choose to spend more time preparing for the earnings forecasts amidst high opioid activity, to avoid ceasing the provision of such forecasts.

Table 7. Other aspects of corporate information production.

	(1)	(2)
Dep. Var. =	Lag	Precision
Opioid	0.020***	−0.001*
	(2.98)	(−1.67)
Size	0.195	0.050***
	(1.62)	(4.01)
Leverage	1.921***	−0.459***
	(4.69)	(−10.34)
MB	−0.030***	0.009***
	(−3.38)	(9.59)
R&D	−47.694***	−2.016***
	(−6.41)	(−3.30)
Acquisition	0.162**	0.033***
	(2.01)	(4.88)
Return	−0.232	0.389***
	(−1.20)	(14.52)
Dispersion	0.671***	−0.259***
	(3.96)	(−7.84)
Analyst	−0.706***	0.004
	(−6.57)	(0.40)
IOR	−1.419***	0.308***
	(−5.28)	(9.08)
Population	−0.004	0.028
	(−0.03)	(1.59)
Unemployment	−0.001	−0.052***
	(−0.01)	(−6.40)
Income	0.000	−0.000***
	(0.45)	(−4.39)
Horizon		−0.105***
		(−15.25)
Firm FE	Yes	Yes
State × YQ FE	Yes	Yes
Adj. R^2	0.713	0.729
N	38,217	50,314

Notes: This table presents the regression results of reporting lag and forecast precision. All continuous variables are winsorized at the top and bottom percentile. *, **, and *** indicate that the estimated coefficient is statistically significant at the 10%, 5% and 1% levels, respectively. See Appendix A for detailed variable definitions.

We next examine whether high local opioid activity is associated with the precision of management earnings forecasts. Prior studies suggest that managers have strong incentives to provide accurate forecasts because labor market would penalize managers for inaccurate forecasts (Graham et al., 2005; Lee et al., 2012). As more precise forecasts are more likely be subsequently proven inaccurate, we expect managers tend to widen their forecast range and lower the forecast’s precision if they realize the forecasts are more likely to be inaccurate. To test this prediction, we estimate Eq. (1) with the dependent variable replaced by Precision, which is defined as forecast range (zero for point estimates) multiplied by minus 100 and scaled by the stock price at the beginning of the fiscal quarter. The regression results are shown in column (2) of Table 7. Consistent with our prediction, the coefficients on Opioid are negative and significant, suggesting that the opioid abuse also induces managers to decrease the precision of earnings forecasts.

7.2. Market Reaction to Management Forecasts

Finally, we examine whether investors realize the reduced quality of management forecast and become less responsive to the forecasts issued by firms located in high-opioid areas. Following prior studies (Guan et al., 2020), we regress the absolute value of the three-day cumulative market-adjusted abnormal return during the [−1, + 1] window around the issuance date of the forecasts multiplied by 100 (ABS_Car) on opioid abuse of a firm’s headquarters. In addition to the control variables employed in the main tests, we include the absolute value of forecasts surprise (Surprise) in the model to control for the magnitude of the news conveyed in the forecasts. We measure Surprise as the difference between management’s earnings forecast and analysts’ consensus pre-management forecast, multiplied by 100 and scaled by the stock price at the beginning of the fiscal quarter. We report the regression results in Online Appendix (Table A2). We find the coefficient on Opioid is negative and significant (p < 0.01), suggesting that investors recognize that opioid abuse may distort the information quality of earnings forecasts.

8. Conclusion

This study examines the effect of employee health on corporate information production. Using the opioid prescription rate in a firm’s headquarters county as the proxy for a firm’s employee health, we find that management earnings forecast accuracy significantly declines when the intensity of opioid abuse is higher. These results are consistent with our conjecture that opioid abuse represents a threat to the health capital of employees who play a role in the firm’s forecasting functions and thereby impairs management forecast quality. We also find that the negative effect of opioid abuse on forecast quality is stronger for firms facing higher forecast difficulty and is mitigated for firms with easier access to opioid treatment, for firms investing more in employee welfare, and for firms with a CSR committee. Further investigation reveals that investors react less to forecasts issued by firms located in high opioid areas. Overall, our evidence suggests that opioid abuse constitutes a serious shock to a firm’s human capital resources, adversely affecting firms’ information production.

This study contributes to the growing literature exploring the role of employee human capital in producing firms’ information (Call et al., 2017; Chen, Wu, et al., 2021). Our results suggest that employee health, a fundamental attribute of human capital, plays a unique and economically significant role in the forecast generation process. This study also contributes to literature investigating the relation between firms’ ESG commitments and financial reporting quality. We extend this line of research by showing that firms’ commitment to ESG can help alleviate the adverse effect of opioid abuse on corporate information production. Finally, we add to the literature on the impact of opioid crisis. We extend this line of research by showing that opioid can negatively affect firm-level management forecasts quality. Given the ongoing opioid epidemics across the US and the rest of the world, this finding is particularly important and relevant because the management forecast is a key mechanism through which firms communicate with investors in the capital market.

Supplemental Data and Research Materials

Supplemental data for this article can be accessed on the Taylor & Francis website, doi:10.1080/09638180.2023.2272622.

Appendix OA: Discussion of alternate explanations for our main findings.

• Table A1: Tests of Alternative Explanations

• Table A2: Tests of the Market Reaction to Management Forecasts

Acknowledgements

We thank Chen Chen, Neil Fargher, Jeroen Koenraadt (discussant), Gary Monroe, Rencheng Wang, Yangxin Yu and participants in EAA 2022 Annual Conference for their useful comments.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data Availability Statement

The data used in this study are available from the public sources identified in the paper.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Notes

1 See: https://www2.deloitte.com/us/en/insights/topics/leadership/ceo-role-employee-health-wellness.html.

2 The widespread opioid abuse has reached crisis levels in the US, with more than 10 million Americans estimated to misuse opioids annually, resulting in $50 billion in healthcare and criminal justice costs, as well as $32 billion in productivity losses (Florence et al., 2021).

3 For example, American Addiction Centers (2022) suggests that opioids (e.g., OxyContin, Vicodin, Fentanyl, Xanax and sleeping pills like Ambien) have made significant marks to white-collar professionals as these employees often resort to these pills for stress relief, relaxation, or to get some sleep at night. See: https://americanaddictioncenters.org/workforce-addiction/white-collar.

4 In untabulated results, we further explore whether the attributes of the management team and/or workforce attenuate or exacerbate the negative effect of opioid abuse on managerial forecast accuracy. We find that the effect of opioid abuse is more pronounced when the workforce is more ethnically diverse and younger. This finding aligns with prior literature, which suggests that pharmaceutical companies tend to target counties with younger and more diverse populations in their marketing of opioids (Hadland et al., 2019).

5 See: https://www.cdc.gov/opioids/basics/epidemic.html/.

6 In 2019, Johnson & Johnson was ordered to pay $572 million for its deceptive conducts that contributes to opioid crisis. Purdue Pharma offered $10 billion to $12 billion to settle more than 2,000 opioid lawsuits against the company for starting and sustaining the opioid crisis. See detailed discussion via https://time.com/5662827/johnson-opioid-crisis-lawsuits/ and https://www.nbcnews.com/news/us-news/purdue-pharma-offers-10-12-billion-settle-opioid-claims-n1046526.

7 Deborah Hersman, the president and CEO of the National Safety Council, said ‘Employers must understand that the most dangerously misused drug today may be sitting in employees’ medicine cabinets. Even when they are taken as prescribed, prescription drugs and opioids can impair workers’ performance and create hazards on the job’. Consistent with this, Greg Zoeller, the Indiana Attorney General, said ‘prescription opioid abuse can cause impairment, injury and may lead employees to bad choices, such as theft or embezzlement from their employers’. See detailed discussion via https://www.nsc.org/in-the-newsroom/prescription-drug-abuse-impacts-80-of-indiana-workplaces-says-national-safety-council-poll.

8 CDC data on opioid prescriptions is available from 2006 onwards.

9 10K header data was downloaded from https://sraf.nd.edu/data/augmented-10-x-header-data.

10 We include firm fixed effects to mitigate the concern that our results can be attributed to unobservable time-invariant factors (e.g., firms’ long-run disclosure policies and inherent forecast difficulty). By doing so, we exploit the variation within firms over time. Our results are robust to controlling the following fixed effects: 1) firm and year-quarter fixed effects, 2) state, firm and year-quarter fixed effects, 3) county and year-quarter fixed effects, and 4) firm and industry × year-quarter fixed effects.

11 To calculate this effect on accuracy, we apply the following equation: (25.484 [standard deviation of Opioid in Table 1] × (−0.010) [coeff. on Opioid in column 2 of Table 2] ÷ 3.099 [standard deviation of Accuracy in Table 1].

12 Our results may suffer from sample selection bias if local rates of opioid abuse can affect the likelihood of a firm issuing management earnings forecasts. However, we find in Section 8.1 that opioid abuse does not affect the likelihood of management forecast issuance. Our results are also robust to using a Heckman selection model or using only bundled annual forecasts.

13 The timing of state-level PDMPs is acquired from http://pdaps.org/.

14 The location of opioid treatment facilities is obtained from: https://dpt2.samhsa.gov/treatment/directory.aspx.

15 The database is accessible at https://www.slcg.com.

16 We follow prior literature to standardize opioid strength using the morphine milligram equivalent (MME) value for each pill (e.g., oxycodone is 50% stronger than hydrocodone, so it has an MME multiplier of 1.5).

17 We first regress Accuracy on Opioid, absent of controls and fixed effects, and obtain the R² of baseline effect. We next regress Accuracy on Opioid, control variables and fixed effects as Eq. (1) and obtain the R² of controlled effect. Computing δ* requires setting a value for R²_max. Following prior studies (Oster, 2019), we set R²_max as 1.3×R² of controlled effect.

Appendix A:

Variable Definitions

Table

Variable	Definition and Measurement
Accuracy	Negative one multiplied by the absolute value of the forecast error. The forecast error is calculated as the difference between management earnings forecast and actual EPS, multiplied by 100 and scaled by the stock price at the beginning of the fiscal quarter.
Precision	Forecast range (high estimate minus low estimate) multiplied by minus 100 and scaled by the stock price at the beginning of the fiscal quarter.
Issue	An indicator variable for issuing at least one annual earnings forecast during each fiscal quarter, and zero otherwise.
Surprise	The difference between management’s earnings forecast and analysts’ consensus pre-management forecast, multiplied by 100 and scaled by the stock price at the beginning of the fiscal quarter.
Opioid	The total count of opioid prescriptions per 100 people in the county reported by CDC.
Opioid_Pill	The total count of opioid pills per capita in the ZIP code reported by DEA.
Opioid_EDvisits	The opioid-related emergency department visit rates in the state reported every year-quarter by HCUP.
Opioid_Placebo	The total count of opioid prescriptions per 100 people in a random county other than a firm’s headquarter county.
Opioid-High	An indicator variable equals one if the opioid level of the county is above median in a year, and zero otherwise.
Opioid_Combined	The average of opioid abuse level across all counties mentioned in the firm’s 10-K filings. We use a Python script to search for city-state locations that mentioned in any portion of firms’ 10-K filings and then match the corresponding counties (FIPS code) to these city-state combinations.
PDMP	An indicator variable equal to one for the 3-year period after the PDMP adoption year in the state and zero for the 3-year period before the adoption year or the control states that are not treated during the six years surrounding the event year.
Size	The natural logarithm of total assets (atq) at the end of fiscal quarter t.
Leverage	Long-term debt (dlttq) divided by total assets (atq) at the end of fiscal quarter t.
MB	The sum of the market value of equity and long-term debt (prccq × cshoq + dlttq), divided by the book value of total assets (atq) at the end of fiscal quarter t.
R&D	R&D expenditures (xrdq, zero if missing) divided by total assets (atq) of fiscal quarter t.
Acquisition	An indicator variable that equals one if the firm reports non-zero, non- missing cash flows related to year-to-date acquisitions for fiscal quarter t (i.e., aqcy is positive), and zero otherwise.
Return	Buy-and-hold tock return over fiscal quarter t.
Dispersion	The standard deviation of one-year-ahead analyst forecasts divided by the absolute value of the median analyst forecast during fiscal quarter t. For analysts who issued multiple forecasts for the same fiscal period during the quarter, we use the value of the first forecast.
Analyst	The natural logarithm of the number of analysts issuing one-year ahead earnings forecasts for the firm during fiscal quarter t.
IOR	The percentage of shares owned by institutional investors in the fiscal quarter t.
Horizon	The natural logarithm of the number of days between the bundled forecast issuance date and the corresponding fiscal period end of the forecasted item.
Population	The natural logarithm of population in the county from US Census Bureau.
Unemployment	Unemployment rate in a county from Bureau of Labor Statistics.
Income	Median household income at the county level from US Census Bureau.
RetSd	The standard deviation of daily stock returns during the fiscal quarter t.
Access	The number of opioid treatment centers within 10 km radius surrounding the corporate headquarter.
Employee_Welfare	The workforce score reported by Thomson Reuters ASSET4 database.
CSR_Committee	An indicator variable equals one if the firm has a corporate social responsibility committee in a year, and zero otherwise.
Education	The average education level of respondents to the American Community Survey (ACS) in the county where the firm is headquartered. Source: the University of Minnesota’s Integrated Public Use Microdata Series (IPUMS–USA).
Lag	The natural logarithm of the number of days between the end of fiscal quarter t and its upcoming actual earnings announcement date.
ABS_Car	The absolute value of the three-day cumulative market-adjusted abnormal return during the [−1, + 1] window around the issuance date of the forecasts multiplied by 100.