Rev Quant Finan Acc DOI 10.1007/s11156-016-0608-7 ORIGINAL RESEARCH
Improvement in clinical trial disclosures and analysts’ forecast accuracy: evidence from the pharmaceutical industry Maggie Hao1 • Dana A. Forgione2 • Liang Guo3 Hongxian Zhang4
•
Springer Science+Business Media New York 2016
Abstract This paper examines whether financial analysts use the information contained in clinical trial disclosures to improve their forecast accuracy for pharmaceutical companies. Findings indicate that the improved clinical trial disclosures due to a quasi-regulation issued by the International Committee of Medical Journal Editors (ICMJE) significantly reduce analysts’ long-term forecast error. In addition, a propensity-score matching analysis provides additional strong evidence that issuance of the 2005 ICMJE’s regulation is accompanied by an average 45 % decrease in long-term forecast error, and a more than 50 % decrease in long-term forecast dispersion. This study contributes to the accounting literature regarding nonfinancial disclosures by providing the first insights into financial analysts’ use of clinical trial disclosures in their forecasts of future earnings. In addition,
& Maggie Hao
[email protected]; http://www.uhcl.edu Dana A. Forgione
[email protected]; http://www.utsa.edu Liang Guo
[email protected]; http://www.csusb.edu Hongxian Zhang
[email protected]; http://www.mst.edu 1
College of Business, University of Houston Clear Lake, 2700 Bay Area Boulevard, Houston, TX 77058, USA
2
College of Business, University of Texas at San Antonio, One UTSA Circle, San Antonio, TX 78249-0632, USA
3
Jack H. Brown College of Business and Public Administration, California State University, San Bernardino, 5500 University Parkway, San Bernardino, CA 92407, USA
4
College of Arts, Sciences, and Business, Missouri University of Science and Technology, 1201 North State Street, Rolla, MO 65409, USA
123
M. Hao et al.
because the major event examined in this study is a quasi-regulation issued by the ICMJE, we provide additional insights on the effectiveness of industry-initiated regulations (or quasi-regulations) on nonfinancial disclosure practice. Keywords Non-financial disclosure Analyst forecast accuracy Clinical trial disclosures Pharmaceutical companies JEL Classification G14 G18 M41
1 Introduction In this paper, we investigate whether increased clinical trial disclosures help to improve financial analysts’ forecast accuracy for pharmaceutical companies. Analysts’ earnings forecasts have been shown to influence investors’ expectations (e.g. Givoly and Lakonishok 1979; Lys and Sohn 1990; Francis and Soffer 1997) and are often used as a proxy for the market’s beliefs (Simpson 2010). Clinical trial status is considered useful in assessing pharmaceutical companies’ R&D pipeline and consequent future revenue and cash flow generation capability. However, what is less clear is whether and how financial analysts use this information. In particular, whether increased clinical trial disclosures improve analysts forecast accuracy. As opposed to required financial disclosures, nonfinancial disclosures are largely voluntary. As Dye (2001) points out, the theory about voluntary disclosure is a special case of game theory, which means that firms will voluntarily disclose good news and withhold bad news. Simpson (2010) argues that the strategic disclosure of nonfinancial information may potentially impair its usefulness to investors and analysts. Pharmaceutical researchers have traditionally reported pre- and post-market clinical trial results in peer-reviewed medical journals (Williams 2007). In recent years, however, there has been a growing awareness of selective publication of clinical trial information (‘‘publication bias’’) as well as the selective reporting of favorable trial outcomes (‘‘outcome reporting bias’’) in the pharmaceutical industry. A series of allegations, focusing on selective reporting of clinical data by the pharmaceutical companies—especially the instance in which sponsors disclosed positive results from clinical trials while leaving out negative trial results—have caused a loss of public trust in the pharmaceutical industry (Thomas and Tesch 2007). The growing awareness of reporting bias has led policy makers to call for increased clinical trial transparency through the public disclosure of key informational items about clinical trials (Tse et al. 2009). In 2005, the International Committee of Medical Journal Editors (ICMJE)—which offers guidance to authors in its uniform requirements for manuscripts submitted to peerreviewed biomedical journals—issued an editorial policy that requires registration of clinical trials at or before the onset of patient enrollment, as a condition for considering a manuscript for publication of clinical trials in its member journals. According to the data of ICMJE, there were 651 participating biomedical journals following ICMJE’s Uniform Requirement as of 2006.1 After the issuance of the ICMJE’s editorial policy, the average number of clinical trials registered in ClinicalTrials.gov, the national public clinical trial
1
The list of participating biomedical journals is available at www.icmje.org/journals.html.
123
Improvement in clinical trial disclosures and analysts’…
registry maintained by the National Institute of Health, increased 500 %—from less than 2000 studies per year to more than 10,000 studies per year in 2005.2 Prior accounting literature suggests that the level of financial disclosure is positively associated with analyst forecast accuracy (i.e. Brown et al. 1987; Lang and Lundholm 1996; Abarbanell and Bushee 1997; and Behn et al. 2008). Does this benefit apply to improved nonfinancial disclosures as well? To shed some light on the nonfinancial disclosure-forecast accuracy relationship, we investigate the association between improved clinical trial disclosures and analyst forecast accuracy for the period between 2001 and 2006. Following Ely et al. (2003) and Xu et al. (2007), we construct a disclosed drug portfolio (DISC) as the sum of the weighted number of registered clinical trials for each clinical trial phase (Phase I, II, and III) deflated by the beginning balance of total assets. To measure the improvement in clinical trial disclosure due to the ICMJE’s editorial policy, we use a dummy variable REG to indicate the quasi-regulated period from year 2005 to 2006. If the increased clinical trial disclosures due to ICMJE’s editorial policy improved analyst forecast accuracy, we expect a negative association between our two forecast quality proxies (absolute forecast error and forecast dispersion) and the joint variable of disclosed drug portfolio and regulated period (DISC*REG). We find that the improved disclosure on clinical trials due to the quasi-regulation issued by ICMJE is significantly and negatively associated with analysts’ forecast errors for three-year ahead earnings forecasts. In addition, when using a propensity-score matching (PSM) analysis to match firm characteristics between pre-2005 and post-2005 observations, we find that issuance of the 2005 ICMJE’s editorial policy is accompanied by an average 45 % decrease in long-term forecast errors and more than 50 % decrease in long-term forecast dispersion. This result indicates significant improvement in analyst forecast accuracy following the improved clinical trial disclosures. Our study contributes to the literature in the following two ways: First, Eccles et al. (2011) suggest that market interests in nonfinancial information are not consistent. Therefore, it is important for companies to identify the specific interests of market participants (i.e. financial analysts) in nonfinancial information and ensure information is provided accordingly. Although there are accounting studies (i.e. Shortridge 2001; Hirschey et al. 2002; Joos 2002; Ely et al. 2003; Dedman et al. 2008) showing that clinical trial disclosures are value-relevant, the documented value-relevance does not necessarily imply that the market fully appreciates its implication for future earnings (Simpson 2010). To the best of our knowledge, this study is the first to provide insights into financial analysts’ use of clinical trial disclosures in their forecasts of future earnings for pharmaceutical companies. Second, because the major event examined in this study is a quasiregulation issued by the ICMJE, we provide additional insights on the effectiveness of industry-initiated regulations (or quasi-regulations) on nonfinancial disclosure practice. ‘‘Appendix 1’’ provides a summary of acronyms used in this paper.
2 Background Disclosure is a communication of the firms’ information to the public. In addition to the call for greater financial disclosures, various individuals and groups have called for greater disclosure of nonfinancial information by corporations (AICPA 1994; Boulton et al. 2000; 2
The ClinicalTrials.gov is an online clinical trial registry developed by the National Library of Medicine for the National Institute of Health. It was first available to the public on February 29, 2000.
123
M. Hao et al.
Norton 2000; Eccles et al. 2001; Lev 2001; and Maines and McDaniel 2000). Such a demand rises from the awareness that the mandatory financial disclosures omit salient information about the company. Traditional financial reporting focuses on the historical performance of a company. Lev (2001) argues that ‘‘in the world where the market value of the firm is decoupled from the value of its underlying assets, nonfinancial information offers a tool for measuring the value arising from intangibles and future cash flows that is missing from traditional financial reports.’’ (as also cited in Cohen et al. 2011 p. 3). This argument is particularly true for companies with high Research and Development (R&D) investments, given the fact that R&D investment is generally expensed as incurred under the U.S. Generally Accepted Accounting Principles (GAAP). Amir and Lev (1996) suggest that for high R&D companies, the reported accounting information (i.e., earnings, book values, and cash flows) by itself is largely value-irrelevant, unless it is combined with nonfinancial information. Pharmaceutical companies are in a high-R&D industry. According to a fact sheet prepared by the Pharmaceutical Research and Manufacturers of America (PhRMA 2007), it takes pharmaceutical companies on average 10 to 15 years and about $800 million to $1 billion to get a new drug from the research lab to the shelf of the retail pharmacy. For pharmaceutical companies, their success depends not only on the profitability and market share of existing products, but also on whether the companies can maintain a successful R&D pipeline to replace the current products once their patents expire. In the U.S., drug patents provide 20 years of protection for pharmaceutical companies to manufacture and market their branded drugs. Once the patent expires, a generic version will enter the market at a much lower price, and quickly siphon off as much as 90 % of the sales (DeRuiter and Holston 2012). In addition, for a majority of branded drugs, the patent clock starts even before clinical trials begin, leaving an effective life of the drug patent at less than 12 years (Grabowski 2002) after the drug is approved for large-scale manufacturing and marketing.3 In the drug development and approval process, a clinical trial is the test of the new drug on human subjects to assess its safety and effectiveness. It has four different phases (Phase I, II, III, and IV) and generally takes between six to seven years. Clinical trial status is one of the most important nonfinancial disclosures associated with the pharmaceutical companies’ R&D pipeline, since the legal requirement for the approval of a new drug is largely based on the substantial evidence of efficacy demonstrated through controlled clinical trials (Food, Drug, and Cosmetic Act 2006).4 Thus, a balanced, complete and accurate disclosure of clinical trial status, especially the results from the latter phases of the drug development cycle, will enhance the market’s assessment of the pharmaceutical companies’ R&D pipeline, and accordingly the assessment of its future cash flows and revenue generation capabilities.
3
The first three phases of clinical trials generally take six to seven years, and the final review and approval by the Food and Drug Administration (FDA) takes another one to two years.
4
In 2005, the State of Maine passed a public law, Maine Sec 1.22 MRSA C.605, which requires pharmaceutical companies to publicly disclose clinical trial information for products that are or have been FDA approved for marketing and are or have been dispensed, administered, delivered or promoted in Maine. A drug may not be advertised in the State of Maine unless the manufacturer has disclosed key information about the trial, including potential or actual adverse effects of the drug. According to this law, clinical trial means a clinical investigation as defined by the FDA that involves any trial to test the safety or efficacy of a drug or biological product with one or more human subjects and that is intended to be submitted to, or held for inspection by, the FDA as part of an application for a research or marketing permit (Maine Public Law 2005).
123
Improvement in clinical trial disclosures and analysts’…
According to the mosaic theory of security analysis, analysts can obtain public, nonpublic, and non-material information about a firm and then use that information to build a mosaic for valuing the company’s worth and provide recommendations to investors. The theory has been recognized by the Chartered Financial Analyst (CFA) Institute as a valid method of analysis. There are different methods analysts can adopt to gather information about a firm. To try and get an edge on other professional investors and analysts, as Sorkin (2010) explained, every day, analysts ‘‘work the phones to ferret out information about companies that can’t be found by simply reading news releases. Some will walk through shopping malls interviewing store managers at Gap, for example, to gauge how sales are going. Others might monitor sales of certain component parts in Asia to determine how many iPads Apple might sell this quarter. For better or for worse, that is what passes as ‘‘research’’ in the finance world.’’ However, the increasing insider-trading scandals have caused the legality of mosaic theory to be questioned and a particular target is the so-called ‘‘expert network’’.5 Though the expert network can help analysts and institutional investors fill the gaps in their knowledge, the provision of expert services can create legal issues if the information provided by experts is considered proprietary or material and nonpublic. The improved disclosure of clinical trial status provides more publicly available information about pharmaceutical companies. In this sense, the improved disclosure of clinical trials can not only reduce information asymmetry, but also reduce the risk of becoming implicated in illegal insider trading through conversations with experts.6 ‘‘Appendix 2’’ provides an overview of the new drug development and approval process.
3 Related research and hypothesis development Several studies have examined the valuation role of nonfinancial disclosures. Among those, several have documented the value-relevance of drug development status for pharmaceutical and biotech companies. For example, Shortridge (2001) finds that the stock market values R&D higher for companies with relatively more drug approvals. Joos (2002) argues that nonfinancial information does not enter the valuation model as a separate term, but rather affects the coefficients of the accounting variables in the valuation equation. He shows that the number of patent applications significantly affects the market valuation of R&D for pharmaceutical firms. Using start-up biotech firms, Ely et al. (2003) demonstrates that drug development status, such as the number of drugs developed by a firm that are in clinical trials and the number of applications submitted, provides significant explanatory power beyond that recognized in accounting data. Dedman et al. (2008) even suggests that for the UK biotechnology sector, earnings announcements have a much lower price impact than drug development announcements. However, the documented value-relevance of nonfinancial information does not necessarily imply that the market fully appreciates its implications for future earnings (Simpson 2010). Different from financial disclosures, nonfinancial disclosures are largely voluntary. This voluntary nature leads to concerns about reliability, comparability, and 5
Expert network is a group of professionals with specialized expertise in the client’s area of interest. For example, investors in the pharmaceutical field may gain valuable insights from certain doctors. These research outfits pool together vast networks of experts who can provide unique perspectives and insights to various industries (Davidowitz 2015).
6
The authors would like to thank an anonymous referee for this comment.
123
M. Hao et al.
persistence of nonfinancial information, and thus may influence analysts’ use of such information. Prior evidence on the association between nonfinancial disclosures and analyst forecast accuracy is scant and yields mixed results. For instance, using data from three continental European countries (Belgium, Germany and the Netherlands), Vanstraelen et al. (2003) show that the level of forward looking nonfinancial disclosures is positively associated with lower dispersion and higher accuracy in financial analysts’ earnings forecasts. Chandra et al. (1999), however, find an insignificant association between the change of book-to-bill ratio (a forward-looking industry-wide indicator disclosed by the trade association of the semiconductor industry) and the revision of analysts’ sales forecasts. Chandra et al. (1999) argue that one possible explanation for this insignificant result is that the nonfinancial indicators voluntarily disclosed to the industry trade association by the companies may be perceived as unreliable, since companies will not share critical information with their competitors if such disclosure is not enforceable. More recently, Nichols and Wieland (2009) examines how analysts react to the disclosures of business and product expansion plans through press releases. They find that not only does the forecasting activity nearly double at the disclosure date, but also the forecasts associated with those disclosures become more accurate and less dispersed across analysts. In an international setting, Dhaliwal et al. (2012) examine the relationship between analyst forecast accuracy and the issuance of stand-alone corporate social responsibility (CSR) reports. They argue that the CSR reports contain information about how well companies handle issues related to stakeholders, thus they are likely to be a useful input for analyst forecasts. Their empirical analysis reveals that the issuance of CSR reports is associated with lower forecast errors. In addition, their results suggest that the negative association between forecast accuracy and financial opacity is tempered by such disclosures. In contrast to the evidence of a positive association between the level of nonfinancial disclosure and analyst forecast accuracy, Simpson (2010) finds that analysts appear to under-react to the disclosure of nonfinancial information items such as customer acquisition cost, average revenue per user, and number of subscribers, by companies in the wireless industry. She suggests that the non-persistent pattern of nonfinancial disclosures may contribute to the observed analysts’ under-reaction to nonfinancial disclosures. The ICMJE’s quasi-regulation of clinical trial registration provides us with an opportunity to further examine the association between nonfinancial disclosures and analyst forecast accuracy. Though not expressed in accounting numbers, clinical trial disclosures likely have important implications for future earnings and fair value estimates. For example, in an analyst’s report for Pfizer, the analyst discussed the impact of a pipeline failure on fair value estimates: ‘‘Removing torcetrapib from our model caused us to knock about $1 off our fair value estimate. We thought torcetrapib had the potential to bring in $10 billion in peak sales for Pfizer, but we had assumed sales from the drug would not approach that level until around 2014. Further, because the drug was still in Phase III, we had assumed a 60 % probability it would eventually receive Food and Drug Administration approval’’.7 Thus, if analysts effectively observe the link between clinical trial status and future earnings, we should expect to observe an improvement in analyst forecast accuracy when 7
Heather Brilliant wrote the report entitled, ‘‘Sticking with Pfizer’s Fair Value,’’ on December 4, 2006. The analyst’s report is available on the MorningStar database, accessed on Feb 9, 2012.
123
Improvement in clinical trial disclosures and analysts’…
the number of clinical trials disclosed in ClinicalTrials.gov significantly increased after the ICMJE’s regulation. Based on this, we hypothesize that: Ho The improved clinical trial disclosures due to ICMJE’s editorial policy are positively associated with analyst forecast accuracy.
4 Methodology We first draw our clinical trial disclosures data from the ClinicalTrials.gov database. Following Ely et al. (2003), for each pharmaceutical company, we construct a disclosed drug portfolio (DISC). The DISC is calculated as the sum of the number of trials registered for each phase times the pre-assigned weight for that phase, deflated by total assets at the beginning of the year.8 DiMasi (2001) suggests that the eventual approval rate for a Phase I clinical trial is 0.24, for a Phase II clinical trial is 0.32, and for a Phase III clinical trial is 0.75. Thus, the disclosed drug portfolio is calculated as: Pn NDRUGSIPi Pi ð1Þ DISCt ¼ i¼1 ATt1 where: DISCt NDRUGSIPi Qi ATt-1
the disclosed drug portfolio in year t; the number of drugs registered as a Phase i (i = 1, 2, or 3) clinical trial in ClinicalTrials.gov as of the end of the year; the conditional success rate of stage i (Q1 = 0.24, Q2 = 0.32, and Q3 = 0.75); and the total assets in year t-1
Then, we obtain financial data from Compustat and analyst forecast data from the Institutional Brokers’ Estimate System (I/B/E/S). We use two proxies to measure analyst forecast accuracy. We first use analyst forecast error as an inverse proxy for forecast accuracy. Following Dhaliwal et al. (2012), we define forecast error (FERROR) as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year: FERRORðY Þi;t ¼
N 1X Y jFCi;t;j EPSYi;t j=Pi;t N j¼1
ð2Þ
where: FERROR (Y)i,t FCi,t,j EPSi,t Pi,t
8
analyst forecast error of firm i, in year t; analyst earnings forecast for firm i, year t, and forecast j; actual earnings per share for firm i in year t; stock price at the beginning of the year for firm i in year t; and Y = Y year ahead, for Y = 1, 2, or 3
The number of trails registered is zero if no record is identified on ClinicalTrails.gov.
123
M. Hao et al.
Since the clinical trial disclosures are likely to link with future earnings, the indicator Y takes three values, 1, 2, or 3, to denote the target earnings and forecasts for the Y year ahead. Our second proxy is the analyst forecast dispersion (FDISP) which is measured as the logarithm of the standard deviation of analysts’ estimates for firm i in year t divided by the stock price at the beginning of the year for firm i in year t: Y =Pi;t ð3Þ FDISPðY Þi;t ¼ Log STDFi;t where: FDISP(Y)i,t STDFi,t Pi,t
Analyst forecast dispersion of firm i, in year t; standard deviation of analyst earnings forecast for firm i, year t; stock price at the beginning of the year for firm i in year t; and Y = Y year ahead, for Y = 1, 2, or 3
Our main dependent variables of interest are analyst forecast error (FERROR) and analyst forecast dispersion (FDISP). We include a dummy variable, REG, to indicate the regulated period of the ICMJE’s trial registration policy from year 2005 to year 2006. Following prior research, we include a number of control variables in our regression. Hope (2003) suggests that the level of a firm’s financial transparency likely influences analyst forecast accuracy. Following Hope (2003) and Dhaliwal et al. (2012), we use total scaled accruals, measured by the Bhattacharya et al. (2003) model, to control for this factor. Lys and Soo (1995) suggest that a greater analyst following indicates more intense competition among analysts, and therefore greater incentive for analysts to enhance their forecast accuracy (Dhaliwal et al. 2012). Hence, we include the number of analysts who provide a forecast in our regression. To control for the firm’s general information environment, we use the natural logarithm of total assets to proxy for firm size (SIZE). Further, to control the effect of earnings volatility on forecast accuracy, we include the natural logarithm of the time-series standard deviation of earnings per share (EPS), (VAREAN), and a dummy variable indicating negative earnings (LOSS).9 Finally, we use the median number of days between the earnings announcement date and the forecast date (FHORIZON) to control for the amount of information available to analysts, given different forecast horizons. We test Ho by estimating: FERRORðY Þi;t ¼ b0 þ b1 DISCi;t þ b2 REGi;t þ b3 DISCi;t REGi;t þ b4 FTRANi;t þ b5 ANANOi;t þ b6 SIZEi;t þ b7 LOSSi;t þ b8 VAREANi;t
ð4Þ
þ b9 FHORIZONi;t þ ei;t FDISPðY Þi;t ¼ b0 þ b1 DISCi;t þ b2 REGi;t þ b3 DISCi;t REGi;t þ b4 FTRANi;t þ b5 ANANOi;t þ b6 SIZEi;t þ b7 LOSSi;t þ b8 VAREANi;t þ b9 FHORIZONi;t þ ei;t ð5Þ where for firm i, and year t: Dependent variables:
9
See, for example, Kadapakkam and Zhang (2014) and Guo, Lien, and Dai (2016).
123
Improvement in clinical trial disclosures and analysts’…
FERROR
FDISP
Analyst forecast error measured as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year; Analyst forecast dispersion measured as the logarithm of the standard deviation of analyst estimates divided by the stock price at the beginning of the year;
Experimental variables: DISC REG DISC*REG
The disclosed drug portfolio divided by total assets at the beginning of the year; 1 if fiscal year equals 2005 or 2006, and 0 otherwise; The disclosed drug portfolio in year 2005 or 2006 (quasi-regulated period);
Control variables: FTRAN ANANO SIZE
VAREAN FHORIZON
Financial transparency measured by total scaled accruals, based on the model of Bhattacharya et al. (2003)10; The number of analysts who provide a forecast for Y year ahead earnings; The natural logarithm of firm’s market value at the beginning of the year; LOSS = 1 if the company reports negative earnings for the year, and 0 otherwise; The natural logarithm of the time-series standard deviation of EPS11; The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst.
The test period is from year 2001 to year 2006.12 If the increased clinical trial disclosures, due to the ICMJE’s regulation, improved analysts’ forecast accuracy, then we expect the sum of the coefficients for the drug portfolio DISC (b1) and the joint variable DISC*REG (b3) to be negative and significant.
5 Results and analysis The pharmaceutical industry includes firms that are engaged in the manufacture, extraction, processing, and packaging of chemical materials used as medicine for human beings or animals. The sample of pharmaceutical firms used in this analysis are drawn from those 10 ACCRUALi,t = (DCAi,t – DCLi,t – DCASHi,t ? DSTDi,t – DEPi,t ? DTPi,t)/TAi,t. Where ACCRUALi,t is the scaled accruals of firm i in year t; DCAi,t is the change in current assets of firm i from year t–1 to year t; DCLi,t is the change in current liabilities of firm i from year t–1 to year t; DCASHi,t is the change in cash of firm i from year t–1 to year t; DSTDi,t is the change in current portion of long-term debt of firm i from year t– 1 to year t; DEP is the depreciation and amortization expense of firm i in year t; DTPi,t is the change in income tax payable of firm i from year t–1 to year t; and TAi,t is the total assets of firm i in year t. To reduce measurement error, we follow Dhaliwal et al. (2012) to convert the absolute value of ACCRUAL into an indicator variable that takes a value of 1 if a firm’s average absolute accruals during the testing period is greater than the median of the sample, and 0 otherwise. The result is consistent with that of using the absolute value of ACCRUAL. 11 Following Dhaliwal et al. (2012), we use a rolling window of 10 years and require a minimum of three years of EPS to calculate the standard deviation. 12 The six-year testing period covers the two-year post-regulation period (2005 and 2006), and four-year pre-regulation period.
123
M. Hao et al. Table 1 Descriptive statistics of variables One-year ahead forecasts (Y1)
Two-year ahead forecasts (Y2)
Three-year ahead forecasts (Y3)
Mean
Mean
Mean
Median
Median
Variables
n = 573
FERROR
0.024
0.008
0.041
0.018
0.058
0.023
FDISP
0.003
0.001
0.002
0.001
0.019
0.004
DISC
0.006
0.000
0.006
0.000
0.006
0.000
FTRAN
0.211
0.114
0.203
0.106
0.194
0.094
ANANO
6.876
4.750
6.485
4.000
3.229
2.174
SIZE
5.777
5.116
5.937
5.315
6.435
5.835
LOSS
0.558
1.000
0.571
1.000
0.516
–1.644
-1.527
-1.635
-1.514
-1.595
-1.466
180.050
182.000
528.918
544.000
878.455
903.000
VAREAN FHORIZON
n = 515
Median
n = 364
1.000
Variable definitions FERROR = Analyst forecast error measured as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year FDISP = Analyst forecast dispersion measured as the logarithm of the standard deviation of analyst estimates divided by the stock price at the beginning of the year DISC = The disclosed drug portfolio divided by total assets at the beginning of the year FTRAN = Financial transparency measured by total scaled accruals based on the model of Bhattacharya et al. (2003) ANANO = The number of analysts who provide a forecast for Y year ahead earnings SIZE = The natural logarithm of firm’s market value at the beginning of the year LOSS = 1 if the company reports negative earnings for the year, and 0 otherwise VAREAN = The natural logarithm of the time-series standard deviation of earnings per share (EPS) FORIZON = The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst
in SIC code 2834.13 After eliminating firms that do not have sufficient data for constructing key regression variables, and firms that primarily manufacture generic drugs, we obtain a final sample of 573 firm-year observations for one-year ahead analysis, 515 firm-year observations for two-year ahead analysis, and 364 firm-year observations for three-year ahead analysis. Table 1 presents the mean and median of our key variables by analysis (one-year ahead, two-year ahead, and three-year ahead). Notably, the mean of analyst forecast error (FERROR) increases from 0.024 for one-year ahead forecasts, to 0.041 for two-year ahead forecasts, then to 0.058 for three-year ahead forecasts. This pattern is consistent with that of the analyst forecast dispersion (FDISP) where FDISP increases from 0.003 for one-year ahead forecasts to 0.019 for three-year ahead forecasts. We also note that the number of analysts who provide three-year ahead forecasts drops in half, compared to the number of analysts who provide one-year ahead or two-year ahead forecasts.
13 According to the description provided by the U.S. Department of Commerce, firms with an SIC code of 2834 are pharmaceutical preparation firms engaged in manufacturing, fabricating, or processing drugs in pharmaceutical preparations for human or veterinary use.
123
Improvement in clinical trial disclosures and analysts’… Table 2 Pearson correlation coefficient matrix A
B
C
D
E
F
G
H
I
Panel A: One-year ahead (Y1) forecasts: n = 573 FERROR
A
1.00
FDISP
B
0.51
1.00
DISC
C
-0.01
-0.03
1.00
FTRAN
D
0.23
0.14
0.16
1.00
ANANO
E
-0.20
-0.20
-0.08
-0.17
1.00
SIZE
F
-0.24
-0.28
-0.14
-0.31
0.73
LOSS
G
0.20
0.21
0.10
0.28
-0.39
-0.56
1.00
VAREAN
H
0.08
-0.08
-0.09
-0.03
0.19
0.31
-0.11
1.00
FHORIZON
I
0.05
0.06
0.12
0.07
0.05
-0.03
0.06
-0.02
1.00
1.00
Panel B: Two-year ahead (Y2) forecasts: n = 515 FERROR
A
1.00
FDISP
B
0.61
1.00
DISC
C
-0.03
-0.01
1.00
FTRAN
D
0.17
0.05
0.18
1.00
ANANO
E
-0.21
-0.12
-0.07
-0.18
1.00
SIZE
F
-0.33
-0.19
-0.14
-0.31
0.75
LOSS
G
0.33
0.16
0.10
0.28
-0.40
-0.59
VAREAN
H
0.07
-0.04
-0.07
-0.03
0.19
0.28
-0.12
1.00
FHORIZON
I
0.04
0.10
-0.01
0.03
0.24
0.09
-0.03
0.08
1.00 1.00 1.00
Panel C: three-year ahead (Y3) forecasts: n = 364 FERROR
A
1.00
FDISP
B
0.63
1.00
DISC
C
-0.09
-0.07
1.00
FTRAN
D
0.12
0.08
0.20
1.00
ANANO
E
-0.20
-0.11
-0.13
-0.17
1.00
SIZE
F
-0.40
-0.20
-0.04
-0.33
0.60
LOSS
G
0.36
0.18
0.09
0.25
-0.33
-0.63
1.00
VAREAN
H
-0.11
-0.11
-0.09
-0.04
0.15
0.26
-0.12
1.00
FHORIZON
I
0.13
0.07
0.06
0.04
0.21
0.13
0.02
0.16
1.00
1.00
Variable definitions FERROR = Analyst forecast error measured as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year FDISP = Analyst forecast dispersion measured as the logarithm of the standard deviation of analyst estimates divided by the stock price at the beginning of the year DISC = The disclosed drug portfolio divided by total assets at the beginning of the year FTRAN = Financial transparency measured by total scaled accruals based on the model of Bhattacharya et al. (2003) ANANO = The number of analysts who provide a forecast for Y year ahead earnings SIZE = The natural logarithm of firm’s market value at the beginning of the year LOSS = 1 if the company reports negative earnings for the year, and 0 otherwise VAREAN = The natural logarithm of the time-series standard deviation of earnings per share (EPS) FORIZON = The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst
123
M. Hao et al. Table 3 One-year ahead forecast (Y1) regression analysis Model (4): FERROR(1) = b0 ? b1 DISC ? b2 REG ? b3 DISC*REG ? b4 FTRAN ? b5 ANANO ? b6 SIZE ? b7 LOSS ? b8 VAREAN ? b9 FHORIZON ? eit Model (5): FDISP(1) = b0 ? b1 DISC ? b2 REG ? b3 DISC*REG ? b4 FTRAN ? b5 ANANO ? b6 SIZE ? b7 LOSS ? b8 VAREAN ? b9 FHORIZON ? eit Model (4) FERROR(1) Coeff. (p)
Model (5) FDISP(1) Coeff. (p)
Experimental variables DISC
b1
-0.139 (0.184)
-0.008 (0.570)
REG
b2
-0.005 (0.365)
0.000 (0.861)
DISC*REG
b3
0.084 (0.451)
0.002 (0.917)
FTRAN
b4
0.038 (0.065)*
-0.001 (0.860)
ANANO
b5
-0.001 (0.048)**
-0.001 (0.465)
SIZE
b6
-0.006 (0.002)***
-0.002 (0.001)***
LOSS
b7
0.004 (0.501)
-0.001 (0.797)
VAREAN
b8
0.008 (0.003)***
0.001 (0.602)
FHORIZON
b9
0.001 (0.869)
0.000 (0.451)
b0
0.068 (0.110)
0.012 (0.001)***
Adj. R2
11.27 %
8.01 %
Prob [ Chi Sq. p
(0.000)***
(0.000)***
Control variables
Intercept
Variable definitions FERROR = Analyst forecast error measured as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year FDISP = Analyst forecast dispersion measured as the logarithm of the standard deviation of analyst estimates divided by the stock price at the beginning of the year DISC = The disclosed drug portfolio divided by total assets at the beginning of the year FTRAN = Financial transparency measured by total scaled accruals based on the model of Bhattacharya et al. (2003) ANANO = The number of analysts who provide a forecast for Y year ahead earnings SIZE = The natural logarithm of firm’s market value at the beginning of the year LOSS = 1 if the company reports negative earnings for the year, and 0 otherwise VAREAN = The natural logarithm of the time-series standard deviation of earnings per share (EPS) FORIZON = The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst *, **, *** Significantly different from 0 at the 0.10, 0.05, and 0.01 levels, respectively
123
Improvement in clinical trial disclosures and analysts’…
Table 2 provides the Pearson correlation matrix: Panel A is the matrix for one-year ahead forecast analysis, Panel B is the matrix for two-year ahead forecast analysis, and Panel C is the matrix for three-year ahead forecast analysis. Consistent with H0, the disclosed drug portfolio (DISC) is negatively associated with analyst forecast error (FERROR) and analyst forecast dispersion (FDISP) for all three forecasting horizons. Table 3 presents the empirical results of the test of H0 for the one-year ahead forecasts. We find that the one-year ahead forecast errors are positively associated financial transparency as proxied by the scale of total accruals, and variation of future earnings, but negatively associated with firm size and number of analysts following the company. We also find that the one-year ahead forecast dispersion is negatively associated with firm size. However, for our main variables of interest DISC, REG, DISC*REG, we did not find a significant association with either the one-year ahead forecast errors or the forecast dispersion. One possible explanation for the insignificant result is that clinical trial disclosures have more impact on future earnings forecasts than on current earnings forecasts, given the long life-cycle of the drug development process. Table 4 documents the regression results of the test of H0 for the two-year ahead forecasts. We find that the two-year ahead forecast errors are significantly lower during the quasi-regulated period (REG = -0.011, p B 0.068). The coefficient for the joint variable DISC*REG is negative as expected, but not significant. Consistent with the regression result for the one-year ahead forecast, the forecast errors are positively associated with financial transparency as proxied by the scale of total accruals, likelihood of negative earnings, variation of future earnings, and the median forecast horizon, but negatively associated with firm size and number of analysts following the company. For the regression on forecast dispersion, we find that the two-year ahead forecast dispersion is negatively associated with disclosed drug portfolio (DISC = -0.016, p B 0.059) and the dummy variable indicating the quasi-regulated period (REG = -0.002, p B 0.008), but positively associated with the joint variable of DISC*REG (DISC*REG = 0.013, p B 0.090). Taken together, this result suggests that improved disclosures of the drug portfolio likely reduce the dispersion of analysts’ earnings forecasts, but such effects may be attenuated by the requirements of mandatory disclosures. Table 5 shows the empirical results of the test of H0 for the three-year ahead forecasts. For the three-year ahead analysis, though we do not find a significant association between forecast dispersion and our main variables of interest, we do find that the three-year ahead forecast errors are significantly and negatively associated with the dummy variable indicating the quasi-regulated period (REG = -0.024, p B 0.012), and the joint variable DISC*REG (DISC*REG = -0.067, p B 0.057). Taken together, these results suggest that the improved drug disclosures due to ICMJE quasi-regulation significantly lower analysts’ earnings forecast errors in their three-year ahead analysis. In estimating models 4 and 5, we cluster the observations by firm to control for the effects of repeated measures. In our panel data analysis, we use both fixed-effects and general-effects models. The results are statistically similar. We then run a Hausman test to choose between the fixed-effects and random-effects models. The p value of the Hausman test is greater than 0.05, indicating that the random-effects model provides a more efficient estimator.
123
M. Hao et al. Table 4 Two-year ahead forecast (Y2) regression analysis Model (4): FERROR(2) = b0 ? b1 DISC ? b2 REG ? b3 DISC*REG ? b4 FTRAN ? b5 ANANO ? b6 SIZE ? b7 LOSS ? b8 VAREAN ? b9 FHORIZON ? eit Model (5): FDISP(2) = b0 ? b1 DISC ? b2 REG ? b3 DISC*REG ? b4 FTRAN ? b5 ANANO ? b6 SIZE ? b7 LOSS ? b8 VAREAN ? b9 FHORIZON ? eit Model (4) FERROR(2) Coeff. (p)
Model (5) FDISP(2) Coeff. (p)
Experimental variables DISC
b1
0.058 (0.632)
-0.016 (0.059)*
REG
b2
-0.011 (0.068)*
-0.002 (0.008)***
DISC*REG
b3
-0.151 (0.220)
0.013 (0.090)*
FTRAN
b4
-0.001 (0.888)
-0.001 (0.571)
ANANO
b5
-0.002 (0.053)*
0.000 (0.744)
SIZE
b6
-0.014 (0.001)***
-0.001 (0.029)**
LOSS
b7
0.027 (0.010)***
0.001 (0.277)
VAREAN
b8
0.009 (0.000)***
0.001 (0.928)
FHORIZON
b9
0.001 (0.001)***
0.001 (0.008)***
b0
0.032 (0.388)
-0.005 (0.236)
Adj. R2
16.00 %
6.84 %
Prob [ Chi Sq. p
(0.000)***
(0.000)***
Control variables
Intercept
Variable definitions FERROR = Analyst forecast error measured as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year FDISP = Analyst forecast dispersion measured as the logarithm of the standard deviation of analyst estimates divided by the stock price at the beginning of the year DISC = The disclosed drug portfolio divided by total assets at the beginning of the year FTRAN = Financial transparency measured by total scaled accruals based on the model of Bhattacharya et al. (2003) ANANO = The number of analysts who provide a forecast for Y year ahead earnings SIZE = The natural logarithm of firm’s market value at the beginning of the year LOSS = 1 if the company reports negative earnings for the year, and 0 otherwise VAREAN = The natural logarithm of the time-series standard deviation of earnings per share (EPS) FORIZON = The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst *, **, *** Significantly different from 0 at the 0.10, 0.05, and 0.01 levels, respectively
123
Improvement in clinical trial disclosures and analysts’… Table 5 Three-year ahead forecast (Y3) regression analysis Model (4): FERROR(3) = b0 ? b1 DISC ? b2 REG ? b3 DISC*REG ? b4 FTRAN ? b5 ANANO ? b6 SIZE ? b7 LOSS ? b8 VAREAN ? b9 FHORIZON ? eit Model (5): FDISP(3) = b0 ? b1 DISC ? b2 REG ? b3 DISC*REG ? b4 FTRAN ? b5 ANANO ? b6 SIZE ? b7 LOSS ? b8 VAREAN ? b9 FHORIZON ? eit Model (4) FERROR(3) Coeff. (p)
Model (5) FDISP(3) Coeff. (p)
Experimental variables DISC
b1
0.606 (0.118)
-0.438 (0.159)
REG
b2
-0.024 (0.012)**
0.001 (0.856)
DISC*REG
b3
-0.672 (0.057)*
0.437 (0.114)
FTRAN
b4
0.040 (0.084)*
-0.005 (0.645)
ANANO
b5
-0.002 (0.093)*
-0.001 (0.412)
SIZE
b6
-0.023 (0.001)***
-0.007 (0.100)*
LOSS
b7
0.025 (0.018)**
0.001 (0.834)
VAREAN
b8
-0.001 (0.816)
-0.004 (0.338)
FHORIZON
b9
0.001 (0.001)***
0.001 (0.018)**
b0
0.113 (0.117)
-0.025 (0.413)
Control variables
Intercept Adj. R2
22.63 %
Prob [ Chi Sq. p
(0.000)***
5.45 % (0.000)***
Variable definitions FERROR = Analyst forecast error measured as the average of the absolute error of all forecasts made in the year for target earnings, scaled by the stock price at the beginning of the year FDISP = Analyst forecast dispersion measured as the logarithm of the standard deviation of analyst estimates divided by the stock price at the beginning of the year DISC = The disclosed drug portfolio divided by total assets at the beginning of the year FTRAN = Financial transparency measured by total scaled accruals based on the model of Bhattacharya et al. (2003) ANANO = The number of analysts who provide a forecast for Y year ahead earnings SIZE = The natural logarithm of firm’s market value at the beginning of the year LOSS = 1 if the company reports negative earnings for the year, and 0 otherwise VAREAN = The natural logarithm of the time-series standard deviation of earnings per share (EPS) FORIZON = The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst *, **, *** Significantly different from 0 at the 0.10, 0.05, and 0.01 levels, respectively
123
123 s.e.
0.024
0.048
0.065
0.039
0.043
0.024
0.003
0.016
0.020
0.002
0.003
0.003
-0.004
-0.002**
-0.000
-0.017
-0.005
-0.000
0.008
0.001
0.001
0.011
0.007
0.006
0.033
0.015
0.038
0.002
0.004
0.003
0.003
0.049
0.089
0.039
0.051
0.024
-0.023**
-0.002**
-0.000
-0.040***
-0.013
-0.009
0.010
0.001
0.001
0.013
0.008
0.006
s.e.
0.015
0.035
0.002
0.004
0.002
0.003
0.050
0.093
0.039
0.046
0.023
0.026
Mean
-0.019**
-0.002***
-0.000
-0.043***
-0.008
-0.004
ATT
Radius matching
0.008
0.001
0.001
0.012
0.007
0.006
s.e.
0.015
0.030
0.002
0.003
0.003
0.003
0.049
0.090
0.039
0.045
0.024
0.026
Mean
-0.015**
-0.002***
-0.000
-0.041***
-0.006
-0.002
ATT
Kernel matching
0.008
0.001
0.001
0.012
0.006
0.005
s.e.
This table examines the impact of the 2005 ICMJE regulatory changes on analyst forecast accuracy after accounting for firm characteristic differences. Panel A and Panel B report the propensity score matching estimate for analyst forecast errors and forecast dispersion, respectively. Superscripts *, **, and *** indicate significance at 10, 5, and 1 % levels, respectively. ATT is defined as the average treatment effect of the ICMJE’s editorial policy on the treated deals
222
142
Pre-2005 (controls)
Post-2005 (treated)
Three-year ahead forecast (Y3)
319
196
Pre-2005 (controls)
Post-2005 (Treated)
Two-year ahead forecast (Y2)
359
214
Pre-2005 (controls)
Post-2005 (treated)
One-year ahead forecast (Y1)
Panel B: Analyst forecast dispersion
222
142
Pre-2005 (controls)
Post-2005 (treated)
Three-year ahead forecast (Y3)
319
196
Pre-2005 (controls)
Post-2005 (treated)
Two-year ahead forecast (Y2)
359
214
Pre-2005 (Controls)
Post-2005 (Treated)
One-year ahead forecast (Y1)
Panel A: Analyst forecast errors
ATT
Mean
Diff. of mean
n
Mean
Nearest neighbor matching
Matching algorithms
Unmatched sample
Table 6 Propensity-score matching analysis
M. Hao et al.
Improvement in clinical trial disclosures and analysts’…
6 Robustness check Simpson (2010) argues that voluntary disclosure involves a trade-off between costs and benefits. Therefore, the pattern of disclosures may be determined by managerial discretion, which may have substantial influence on analyst forecast accuracy as well. Given this possibility, our regression test to examine the effects of improved disclosures on forecast accuracy may be exposed to endogeneity problems, leading to a biased and inconsistent estimate of the disclosure effects.14 A propensity-score matching analysis is commonly used in the accounting and finance literature to remedy potential endogeneity bias.15 As a robustness check, we employ propensity-score matching methods in this study to address the difference in firm characteristics between pre-2005 and post-2005 observations. We estimate the propensity score of each observation as the probability of receiving the treatment, given a set of firm characteristics. Based on the estimated propensity scores, we match each observation in the post-2005 period (treatment) with those in the pre-2005 period (control).16 In Table 6, Panel A, we report the effects of the 2005 ICMJE’s editorial policy on analyst forecast errors under the nearest-neighbor matching, radius matching, and kernel matching methods.17 In Panel B we present a similar analysis for analyst forecast dispersion. Although there is no significant difference between the treatment and control groups in our original unmatched samples (with the exception of the two-year ahead forecast dispersion), the propensity-score matching analysis provides interesting additional insights. Consistent with the regression analysis in the previous section, our matching results demonstrate that the improved clinical trial disclosures due to the 2005 ICMJE’s editorial policy significantly reduce three-year ahead forecast errors. We find that the average treatment effects on the treated (ATTs), under three different matching algorithms, are all significantly negative and economically large. As presented in Table 6, Panel A, the three-year ahead forecast errors in the post-2005 period are narrowed by about 45 % when compared to the matched ones in the pre-2005 period, regardless of which matching algorithm is used. For example, under the nearest-neighbor matching method, the mean of three-year ahead analyst forecast errors for the matching samples (after controlling the firm characteristics) is 0.089 in the pre-2005 period as opposed to 0.049 in the post-2005 period. For three-year ahead forecast errors, the differences between the pre-2005 and post-2005 observations are between -0.040 and -0.043 among these three matching methods, and all are statistically significant at the 1 % level. As shown in Table 6, Panel B, fairly similar results are observed for analyst forecast dispersion. More interestingly, all three matching methods unanimously show that both the two-year ahead and three-year ahead forecast dispersion in post-2015 observations shrink by more than 50 % when compared to those in matched pre-2005 observations. For example, under the nearest-neighbor matching method, the two-year (three-year) ahead 14
The authors wish to thank an anonymous referee for this comment.
15
See, for example, Pana et al. (2015); Guo (2016); Wu, Shen, and Chen (2016).
16
Appendix C presents details about our implementation of the propensity score matching analysis and balancing tests. 17 That is, the average treatment effect on the treated (ATT). See further explanation of ATT in Appendix C. Several matching algorithms have been proposed in the literature. However, there is no clear rule to determine which one is more appropriate (Heinrich et al. 2010). To make sure that our findings are not driven by the selection of a particular strategy, we use three of the most commonly employed matching algorithms (i.e., nearest-neighbor matching, radius matching, and kernel matching) to estimate an average treatment effect on the treated (ATT) in this study.
123
M. Hao et al.
forecast dispersion for post-2005 observations is 0.002 (0.015) as opposed to 0.004 (0.038) for matched pre-2005 observations. Overall, our results provide strong evidence that improved clinical trial disclosures due to the 2005 ICMJE’s editorial policy tend to remarkably improve earnings forecast accuracy by narrowing analyst forecast errors and forecast dispersion, particularly for the three-year ahead forecasts. Collectively, the evidence provides valuable insights into financial analysts’ use of clinical trial disclosures in their forecasts of future earnings.
7 Conclusions In this study, we examine whether improved disclosure of clinical trials, as proxied by a disclosed drug portfolio, is associated with improved earnings forecast accuracy by financial analysts. After controlling for various potential confounding factors, we find that the improved disclosure of clinical trials due to a quasi-regulation issued by the ICMJE is significantly and negatively associated with analysts’ earnings forecast errors for long-term forecasts. This finding provides evidence of financial analysts’ use of clinical trial disclosures in their forecasts of long-term earnings for pharmaceutical companies. In addition, recent research on accounting disclosures suggests that the market reacts differently to disclosure rules, depending upon whether the government or a professional self-regulating organization developed the rules (Jamal et al. 2003, 2005; Bertomeu and Cheynel 2013; Bertomeu and Magee 2015). Politicians compete for votes and thus favor proposals that appear popular, but are not necessarily in the best interests of diversified shareholders. Under self-regulation, standard-setting organizations are responsibly to their constituency, with board members chosen by trustees. Self-regulation encourages active participation by issuers, but may induce strategic manipulation of new agenda items. In our setting, rules for clinical trial disclosures were proposed and implemented by the ICMJE, a self-regulated and respected professional organization. This organization is composed of prestegious medical journal editors whose participants meet annually to develop and maintain Recommendations for the Conduct, Reporting, Editing and Publication of Scholarly Work in Medical Journals. Thus, another contribution of this study is that our findings may also have implications for academics and practitioners in their understanding of the role played by quasi-regulation (or self-regulation) in the disclosure process. Overall, we conclude that the demonstrated association between improved clinical trial disclosures due to the ICMJE editorial policy, and improvements in long-term forecast accuracy, is an important first step that we hope will spur future research into disclosure of nonfinancial information. For example, future research might investigate the type of clinical trial disclosures (trial registration vs. results disclosure) that is most useful in forecasting. In addition, the 2007 Food and Drug Administration Amendment Act (FDAAA) require mandatory disclosure of trial registration and trial results. It provides a unique opportunity for accounting researchers to examine the effects of mandatory nonfinancial disclosures, given the fact that the majority of nonfinancial disclosures are voluntary. Acknowledgements The authors would like to thank Drs. Jeff Boone, James Groff, Emeka Nweaze, and the research workshop participants at The University of Texas at San Antonio, and participants at the 2014 American Accounting Association Annual Meeting and Midwest Region Meeting for their helpful comments. Also, we would like to thank Drs. Cheryl Linthicum and Pamela Smith for their valuable comments and suggestions on an earlier draft. Finally we would like to thank the editor and anonymous reviewers for their valuable feedback throughout the peer review process.
123
Improvement in clinical trial disclosures and analysts’… Data availability Data are available from public sources identified in the paper.
Appendix 1 Acronym
Term
AICPA
American Institute of Certified Public Accountants
CSR
Corporate social responsibility
FDA
Food and drug administration
FDAAA
Food and drug administration amendments act
GAAP
Generally accepted accounting principles
ICMJE
International Committee of Medical Journal Editors
IND
Investigational new drug
NDA
New drug application
NIH
National Institutes of Health
PhRMA
Pharmaceutical Research and Manufacturer of America
R&D
Research and development
SIC
Standard industry classification
Appendix 2: New drug development and approval process The new drug development and approval process generally takes six steps, and two formal applications—the investigational new drug (IND) and new drug application (NDA)18— which need to be filed with and approved by the FDA before a new drug is approved for large-scale manufacture and marketing in the U.S. The six steps include, (1) pre-discovery, (2) new drug discovery, (3) pre-clinical trial and IND application, (4) clinical trials Phases I, II and III, (5) FDA review and NDA approval, and (6) post-marketing clinical trials Phase IV. The new drug development is an enormously time consuming process. According to the data of U.S. PhRMA, steps 2 through 5 generally take about 10 to 15 years, within which the first three phases of clinical trials (step 4) alone will take about six to seven years. Moreover, only one out of every 5000 compounds identified during the initial drug discovery process will eventually be approved for marketing by the FDA after this long and costly R&D process (PhRMA 2007). The new drug development and approval process is summarized in Fig. 1, as discussed below.
Pre-discovery and new drug discovery The new drug development process starts with gaining an understanding of the diseases to be treated and the underlying causes of the health condition. During the pre-discovery phase, pharmaceutical researchers will select a target, likely a gene or protein involved in a particular disease, which can potentially interact with and be affected by a drug molecule 18 If the promising compound is determined to be reasonably safe based on animal testing during the preclinical trials, an IND is filed to request approval for clinical trials to test the compound on humans. If the compound can successfully pass the first three phases of clinical trials on human beings, an NDA is filed to request approval for marketing and large-scale manufacturing, which is then followed by the Phase IV clinical trials.
123
M. Hao et al.
Fig. 1 New drug development and approval process. Source Drug discovery and development by the pharmaceutical research and manufacturers of America (PhRMA) (2007)
(PhRMA 2007). Once the target is selected, the second step is to identify a promising drug molecule, or a lead compound, that may act on the target to alter the disease course. This stage is defined as the new drug discovery process, during which a series of lab tests are performed on several promising compounds, to test for safety and help researchers prioritize lead compounds and identify optimal ones.
Pre-clinical trials and IND application Once a set of optimal compounds is identified, pre-clinical trials are performed to determine whether the optimal compounds identified from the drug discovery process are safe enough to be tested on human beings. During this phase, pharmaceutical researchers will perform thorough laboratory and animal tests, as requested by the FDA, to discover how the drug works, and whether it is likely to be safe and work well in humans. After preclinical testing on animals is completed, an IND is filed with the FDA to request approval for clinical trials on human beings. The new drug discovery and pre-clinical trials, together, take an average three to six years. After starting with approximately 5000 to 10,000 compounds, scientists will have narrowed the group down to between one and five ‘‘candidate drugs’’ which will be studied in clinical trials (PhRMA 2007). New drug discovery research and pre-clinical testing results are often voluntarily and selectively published in scientific journals, and companies only sporadically and selectively disclose pre-clinical research and patent approvals in their financial reports (Joos 2003).
Clinical trials (phase I, II, and III) A clinical trial is a research study in human volunteers. The clinical trial process is both expensive and time-consuming, and ends more often in failure than success (PhRMA 2007). There are three phases of clinical trials the candidate drugs must pass through before an NDA can be filed with the FDA. In Phase I of the clinical trial, a candidate drug is tested for the first time on human beings with the disease or condition under study. The
123
Improvement in clinical trial disclosures and analysts’…
trial is conducted on a small group of patients, normally ranging from 20 to 100, to assess its safety, determine a safe dosage range, and identify any side effects. In Phase II, a candidate drug is provided to a larger group of patients with the disease or condition under study, to evaluate its effectiveness and further assess its safety. The usual sample size during Phase II clinical trials is between 100 and 500 patients. In a Phase III clinical trial, the sample size is increased to between 1000 and 5000 patients. Phase III trials are both the costliest and longest trials, during which hundreds of sites around the U.S. and the world participate. This enables the study to obtain a large and diverse group of patients, and generate statically significant data about the safety, effectiveness and overall benefit-risk relationship of the drug (PhRMA 2007). The first three phases of clinical trials take an average of six to seven years to complete. During any of these three phases, either the FDA, or the pharmaceutical company who sponsors the study, can stop the testing, due to significant adverse effects or lack of effectiveness.
FDA final review and NDA approval If the candidate drug successfully completes the first three phases of clinical trials, the pharmaceutical company will file an NDA with the FDA to request approval for marketing and large-scale manufacture. The FDA will review the application from the perspective of safety and effectiveness. Based on the review, the FDA can either: (1) approve the new drug, (2) request additional information or testing before final approval is granted, or (3) deny the application. The FDA review and approval normally takes between six months to two years. It is estimated that only one out of 5000 to 10,000 compounds identified during the initial drug discovery process will ultimately be approved by the FDA for large-scale manufacture and marketing (PhRMA 2007).
Post-marketing clinical trial (phase IV) Even after the approval for marketing and large-scale manufacture, the testing of the new drug does not stop. The FDA requires pharmaceutical companies to continue monitoring the drug for any unexpected serious adverse effects after it is used in the larger population. These studies on approved drugs are Phase IV clinical trials. Failure to promptly conduct any mandatory Phase IV clinical trials or failure of the product during the post-market testing could result in withdrawal of approval or other legal sanctions (Optimer 2011). For example, one of Merck’s blockbuster drugs, Vioxx, was approved by the FDA for relief of the pain associated with arthritis in 1999. Before the FDA’s approval, Vioxx had been tested by 8076 patients during the Phase III clinical trials (Greener 2005). However, on September 30, 2004, Vioxx was withdrawn from the market due to the increased risk of heart disease. The withdraw of Vioxx is an example of failure on post-marketing clinical trials, and it also shows the significant consequence of negative results from clinical trials. In summary, conducting clinical trials is a lengthy and expensive process, and it ends more often in failure than success. Therefore, a balanced, complete, and accurate disclosure of clinical trial status, especially the results from the latter phases of the drug development cycle, will enhance the market’s assessment of the pharmaceutical companies’ R&D pipeline and accordingly its revenue generating capabilities.
123
M. Hao et al.
Appendix 3: Propensity score matching analysis and balance testing Our purpose is to match pre-2005 and post-2005 observations with similar ex ante firm characteristics, to isolate the effects of the 2005 ICMJE’s editorial policy on analyst forecast accuracy. Rosenbaum and Rubin (1983, 1984) suggest that propensity score matching analysis offers a good control group design and reduces potential bias by largely eliminating pre-treatment characteristic differences when matching treatment and control groups. Let Yi be the outcome of interest (i.e. forecast error and forecast dispersion), where i represents each firm-year observation. We want to estimate the causal effect of treatment Ti on various outcome variables. In our case, the treatment is the issuance of the 2005 ICMJE’s editorial policy (Ti = 1). We define the treatment effect for an observation i as Di = Yi1 Yi0 , where Yi1 represents the outcome of observation i with treatment (Ti = 1, post-2005) and Yi0 represents the outcome without treatment (Ti = 0, pre-2005). However, for any observation i we can only observe its outcome in one of the two possible regimes. We do not have the counter-factual evidence (i.e., the outcome in the unobserved regime). The most common evaluation parameter of interest is the average treatment effect on the treated, ¼ E Yi1 Yi0 jTi ¼ 1 ¼ E½Yi1 jTi ¼ 1 E½Yi0 jTi ¼ 1, which measures the average treatment effect of the ICMJE’s editorial policy on analyst forecast accuracy. In this case, however, the counter-factual, E½Yi0 jTi ¼ 1, is unobserved. Under specific identifying assumptions (e.g., the conditional independence assumption and common support condition), we adopt the propensity-score matching techniques pioneered by Rosenbaum and Rubin (1983) to construct a valid counterfactual estimator for the average outcome of treated observations. To implement this, we define the propensity score as the probability of receiving the treatment given a set X = xi of firm characteristics: pð X Þ ¼ pr ðTi ¼ 1jX ¼ xi Þ. We estimate the propensity score using a Logit model with various firm characteristics, as listed for model 4 and model 5 (control variables). The STATA report shows that the balancing property holds within all five blocks for all covariates across treatment and control groups, which indicates that the propensity score’s distribution is similar across groups within each block and that the propensity score is properly specified.19 Table 7 presents the Logit regression results. Then, we match each treatment deal with a control deal on the basis of similar propensity scores. The average treatment effect of the ICMJE’s editorial policy on the treated deals (ATT) is estimated by computing the expected value of the difference in the outcome variable between each of the treated observations and the matched control observations. To ensure that our matching results are robust, we choose three of the most commonly used matching methods: nearest-neighbor matching, radius matching, and kernel matching. As to radius matching, we follow the general rule of thumb and use a caliper of 0.20 of the standard deviation of the propensity score (Austin 2011; Guo 2013). In the computation we also impose common support, to ensure that the matching estimation is taken in the region of common support. Finally, we evaluate the balance of our matched groups with a paired t-test between the treatment and control observations, which tests whether there are significant differences in the covariate means for both groups. Table 8 reports an example of the balancing test results after kernel matching. In 19 To ensure balance of the propensity score across treatment and comparison groups, we obtain an estimate of the propensity score’s distribution by splitting the sample by quintiles (blocks) of the propensity score. A starting test of balance is to ensure that the mean propensity score is equivalent in the treatment and comparison groups within each of the five quintiles (See, e.g., Imbens 2004).
123
Improvement in clinical trial disclosures and analysts’… Table 7 Propensity score estimation
p values are reported in parentheses below the coefficients. *** p \ 0.01, ** p \ 0.05, * p \ 0.1
Variable
Coeff. (p)
FTRAN
0.027 (0.95)
ANANO
0.309*** (0.00)
SIZE
-0.276*** (0.00)
LOSS
0.095 (0.75)
VAREAN
0.057 (0.53)
FHORIZON
0.005** (0.01)
Adj. R2
10.0 %
Table 8 Balancing tests from kernel matching estimators Variable
Sample
Mean Treatment
FTRAN ANANO SIZE LOSS
t test Control
t
p
Unmatched
0.204
0.187
0.66
0.51
Matched
0.212
0.205
0.25
0.80
Unmatched
4.148
2.641
4.79
0.00
Matched
3.331
3.483
-0.46
0.65
Unmatched
6.407
6.452
-0.18
0.85
Matched
6.157
6.339
-0.63
0.53
Unmatched
0.549
0.496
1.00
0.32
Matched
0.583
0.512
1.17
0.24
-1.527
-1.640
0.79
0.43
VAREAN
Unmatched Matched
-1.594
1.800
0.15
0.88
FHORIZON
Unmatched
895.000
867.870
3.50
0.00
Matched
893.740
894.530
-0.11
0.92
Variable definitions FTRAN Financial transparency measured by total scaled accruals based on the model of Bhattacharya et al. (2003); ANANO The number of analysts who provide a forecast for Y year ahead earnings SIZE The natural logarithm of firm’s market value at the beginning of the year LOSS 1 if the company reports negative earnings for the year, and 0 otherwise VAREAN The natural logarithm of the time-series standard deviation of earnings per share (EPS) FORIZON The median forecast horizon calculated as the number of days between the earnings announcement date and the forecast date for each company, year, and analyst
unmatched samples, a few covariates (e.g., ANANO and FHORIZON) present significant differences between the treated and control groups. However, after matching, the tstatistics demonstrate that we fail to reject the hypothesis that the mean differences for all
123
M. Hao et al.
covariates between treated and control deals are equal to zero. Overall, we ensure the effectiveness of our chosen propensity score specification by accounting for selection bias in our sample.
References Abarbanell JS, Bushee BJ (1997) Fundamental analysis, future earnings, and stock prices. J Account Res 35(Spring):1–24 American Institute of Certified Public Accountants (AICPA) (1994) Improving business reporting—a customer focus. AICPA, New York Amir E, Lev B (1996) Value relevance of nonfinancial information: the wireless communication industry. J Account Econ 22(August–December):3–30 Austin PC (2011) Optimal caliper widths for propensity-score matching when estimating differences in means and differences in proportions in observational studies. Pharm Stat 10:150–161 Behn BK, Choi J, Kang T (2008) Audit quality and properties of analyst earnings forecasts. Account Rev 83(2):327–349 Bertomeu J, Cheynel E (2013) Toward a positive theory of disclosure regulation: in search of institutional foundations. Account Rev 88(3):789–824 Bertomeu J, Magee RP (2015) Political pressures and the evolution of disclosure regulation. Rev Account Stud 20(2):775–802 Bhattacharya U, Daouk H, Welker M (2003) The world price of earnings opacity. Account Rev 78(3):641–678 Boulton R, Libert B, Samek S (2000) Cracking the value code. HarperBusiness, New York Brown LD, Richardson G, Schwager S (1987) An information interpretation of financial analyst superiority in forecasting earnings. J Acc Res 25(Spring):49–67 Chandra U, Procassini A, Waynire G (1999) The use of trade association disclosures by investors and analysts: evidence from the semiconductor industry. Contemp Account Res 16(4):643–670 Cohen JR, Holder-Webb LL, Nath L, Wood D (2011) Corporate reporting of nonfinancial leading indicators of economic performance and sustainability. Account Horiz 26(1):65–90 Davidowitz AS (2015) Abandoning the ‘mosaic theory’: Why the ‘mosaic theory’ of security analysis constitutes illegal insider trading and what to do about it. Wash U J L Pol’y 46:281–303 Dedman E, Lin SW, Prakash AJ, Chang C (2008) Voluntary disclosure and its impact on share prices: evidence from the UK biotechnology sector. J Account Public Policy 27:195–216 DeRuiter J, Holston PL (2012) Drug patent expirations and patent cliff. US Pharm 37(6):12–20 Dhaliwal DS, Radhakrishnan S, Tsang A, Yang Y (2012) Nonfinancial disclosure and analyst forecast accuracy: international evidence on corporate social responsibility disclosure. Acc Rev 87(3):723–759 DiMasi JA (2001) Risks in new drug development: approval success rates for investigational drugs. Clin Pharmacol Ther 2001(May):297–307 Dye RA (2001) An evaluation of ‘essays on disclosure’ and the disclosure literature in accounting. J Account Econ 32:181–235 Eccles RG, Herz R, Keegan E, Phillips DMH (2001) The value reporting revolution. Wiley, New York Eccles RG, Serafeim G, Krzus MP (2011) Market interest in nonfinancial information. J Appl Corp Finance 23(4):113–127 Ely P, Simko PJ, Thomas LG (2003) The usefulness of biotechnology firms’ drug development status in the evaluation of research and development costs. J Account Audit Finance 18(1):163–196 Food, Drug, and Cosmetic Act. Section 505; 21 USC 355. 2006. www.fda.gov/regulatoryinformation/ legislation/federalfooddrugandcosmeticactfdcact/default.htm. Accessed 9 Nov 2012 Francis J, Soffer L (1997) The relative informativeness of analysts’ stock recommendations and earnings forecast revisions. J Account Res 35(2):193–211 Givoly D, Lakonishok J (1979) The information content of financial analysts’ forecasts of earnings: some evidence on semi-strong inefficiency. J Account Econ 1(3):165–185 Grabowski H (2002) Patents, innovation and access to new pharmaceuticals. http://levine.sscnet.ucla.edu/ archive/grabow-patents_innov.pdf. Accessed 8 Dec 2012 Greener M (2005) Drug safety on trial. EMBO Rep 6(3):202–204 Guo L (2013) Determinants of credit spreads: the role of ambiguity and information uncertainty. N Am J Econ Finance 24(C):279–297 Guo L (2016) Are U.S. investors blindly chasing returns in foreign countries? Int Rev Econ Finance 41:309–334
123
Improvement in clinical trial disclosures and analysts’… Guo L, Lien D, Dai Y (2016) The effects of China’s split-share reform on firms’ capital structure choice. Appl Econ 48(27):2530–2549 Heinrich C, Maffioli A, Va´zquez G (2010) A primer for applying propensity score matching: impactevaluation guidelines. Inter-American Development Bank Hirschey M, Richardson V, Scholz S (2002) Value relevance of nonfinancial information: the case of patent data. Rev Quant Financ Account 17:223–235 Hope O (2003) Disclosure practice, enforcement of accounting standards, and analysts’ forecast accuracy: an international study. J Account Res 41(May):235–272 ICMJE (2005) Clinical trial registration: a statement from the International Committee of Medical Journal Editors. www.icmje.org/clin_trial.pdf. Accessed 9 Nov 2012 Imbens GW (2004) Nonparametric estimation of average treatment effects under exogeneity: a review. Rev Econ Stat 86(1):4–29 Jamal K, Maier M, Sunder S (2003) Privacy in E-commerce: development of reporting standards, disclosure, and assurance services in an unregulated market. J Account Res 41:285–309 Jamal K, Maier M, Sunder S (2005) Enforced standards versus evolution by general acceptance: a comparative study of e-commerce privacy disclosure and practice in the United States and the United Kingdom. J Account Res 43:73–96 Joos P (2002) Explaining cross-sectional difference in market-to-book ratio in the pharmaceutical industry. www.simon.rochester.edu/fac/philjoos/pharm2002.pdf. Accessed 9 Nov 2012 Joos P (2003) Discussion—the usefulness of biotechnology firms’ drug development status in the evaluation of research and development costs. J Account Audit Finance 18(1):197–205 Kadapakkam PR, Zhang H (2014) Investor ignorance in markets for worthless stocks. J Financ Mark 19:197–218 Lang M, Lundholm R (1996) Corporate disclosure policy and analyst behavior. Account Rev 71:467–493 Lev B (2001) Intangibles—management, measurement, and reporting. Brookings Institution Press, Washington, DC Lys T, Sohn S (1990) The associations between revisions of financial analysts’ earnings forecasts and security price changes. J Account Econ 13(4):341–363 Lys T, Soo L (1995) Analysts’ forecast precision as a response to competition. J Account Audit Finance 10(4):751–765 Maine Public Law (2005) Title 22: health and welfare; subtitle 2: health; part 5: food and drugs; chapter 605: prescription drug advertising heading: PL 2005, C. 392, §1. www.mainelegislature.org/legis/statutes/ 22/title22sec2700-a.html. Accessed 9 Nov 2012 Maines LA, McDaniel LS (2000) Effects of comprehensive-income characteristics on nonprofessional investors’ judgments: the role of financial-statement presentation format. Account Rev 75(2):179–207 Nichols C, Wieland M (2009) Do firms’ non-financial disclosures enhance the value of analyst services? Working paper, Cornell University. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1463005. Accessed 9 Nov 2012 Norton DP (2000) Should balanced scorecards be required? Balanced scorecard report July–August, pp 14–15 Optimer (2011) Annual report. http://216.139.227.101/interactive/optr2011/optr2011_download.pdf?print_ pages=true. Accessed 9 Nov 2012 Pana E, Vitzthum S, Willis D (2015) The impact of internet-based services on credit unions: a propensity score matching approach. Rev Quant Financ Acc 44:329–352 PhRMA (2007) Drug discovery and development: understanding the R&D process. www.innovation.org/ drug_discovery/objects/pdf/RD_Brochure.pdf. Accessed 9 Nov 2012 Rosenbaum PR, Rubin DB (1983) The central role of the propensity score in observational studies for causal effects. Biometrika 70:41–55 Rosenbaum PR, Rubin DB (1984) Reducing bias in observational studies using subclassification on the propensity score. J Am Stat Assoc 79:516–524 Shortridge R (2001) Measuring R&D success: a study of new drug approvals in the pharmaceutical industry. Working paper, Ball State University Simpson A (2010) Analyst’s use of nonfinancial information disclosures. Contemp Acc Res 27(1):249–288 Sorkin AR (2010) Just tidbits or material facts for insider trading? New York Times (November 29). http:// dealbook.nytimes.com/2010/11/29/just-tidbits-or-material-facts-for-insider-trading/. Accessed 9 Nov 2012 Thomas KB, Tesch C (2007) Clinical trial disclosure: the ongoing debate on public registers for clinical trials. Write Staff 16(2):67–72 Tse T, Williams RJ, Zarin DA (2009) Reporting ‘‘basic results’’ in ClinicalTrials.gov. Chest 136(1):295–303
123
M. Hao et al. Vanstraelen A, Zarzeski MT, Robb SWG (2003) Corporate nonfinancial disclosure practices and financial analyst forecast ability across three European countries. J Int Financ Manag Account 14(3):249–278 Williams ED (2007) CRS report for congress: clinical trial reporting and publication. http://biotech.law.lsu. edu/crs/RL32832.pdf. Accessed 9 Nov 2012 Wu M, Shen C, Chen T (2016) Application of multi-level matching between financial performance and corporate social responsibility in the banking industry. Rev Quant Financ Account. doi:10.1007/ S11156-016-0582-0 Xu B, Magnan ML, Essec PA (2007) The stock market valuation of R&D information in biotech firms. Contemp Account Res 24(4):1291–1318
123