Internet Appendix for The dark side of analyst coverage: The case of innovation

Transcription

1 Inerne Appendix for The dark side of analys coverage: The case of innovaion This inerne appendix provides robusness ess and supplemenal analyses o he main resuls presened in The Dark Side of Analys Coverage: The Case of Innovaion. Secion A presens he robusness checks for he baseline resuls repored in Secion 4 of he paper. Secion B explores a possible raionale for he direc effec of analys coverage on firm innovaion: he adverse consequences of missing earnings arges. A. Robusness checks for baseline resuls We conduc a rich se of robusness ess for our baseline resuls and repor hem in Table A1. To save space, we only repor he coefficien esimaes for he key variable of ineres, LnCoverage, and suppress hose of all oher conrols. Firs of all, we repea he baseline OLS regressions by using various subsamples. We sar wih he subsample of firms ha are in he S&P 500 index o check if our resuls are driven by small and young firms ha are ypically covered by fewer analyss bu are more innovaive. We repor he resuls in Panel A. The coefficien esimaes of LnCoverage are negaive in boh columns and significan a he 10% and 1% level, respecively, suggesing ha our resuls hold for large and maure firms. We hen focus on he subsamples defined by firms paening, namely, firms wih a leas N (N = 1, 2, 3, or 5) paens in he sample period, and repor he resuls in Panel B. We observe ha he coefficien esimaes of LnCoverage are all negaive and significan a eiher 5% or 1% level, excep for he case when we require N 5 and LnPaen is he dependen variable. Nex, we impose sricer resricions on he subsamples. In columns (1) and (2) of Panel C, we require he firms o have non-zero paens in mos (more han 50%) of he years hey appear in our sample. The coefficien esimaes of LnCoverage are negaive in boh columns and significan a he 5% level in column (2). Third, in columns (3) and (4) of Panel C, we impose he srices resricion by requiring he firms o have non-zero paens in every single year in he sample period. The coefficien esimaes of LnCoverage are sill negaive in boh columns bu become insignifican. This loss in saisical significance mainly comes from a decrease in he power of he es due o he dramaic drop in sample size from 25,860 o 1,321. 1

2 We furher examine subsamples defined by differen hresholds of analys coverage. We impose various hreshold values for he average annual number of analyss a firm needs o have o appear in he subsample: N = 5, 15, or 25, and repor he resuls in Panel D. In columns (1) and (2), we require N 5 analyss. The coefficien esimaes of LnCoverage are negaive and significan a he 1% level in boh columns. In columns (3) and (4), we require N 15 analyss. The coefficien esimaes of LnCoverage are negaive in boh columns and significan a he 1% level in column (4). In columns (5) and (6), we require N 25 analyss. The coefficien esimaes of LnCoverage are negaive in boh columns bu no saisically significan. I is consisen wih he findings in he Difference-in-Differences (DiD) framework repored in Table 3 Panel C of he paper, suggesing ha he negaive effec of analys coverage on innovaion disappears when firms are covered by many analyss. Second, we check wheher our baseline resuls are robus o alernaive proxies for analys coverage and innovaion. We sar wih using alernaive proxies for analys coverage following Yu (2008) o address he concern ha analys coverage is associaed wih many facors ha could also affec firms innovaion produciviy. Specifically, we consruc he residual coverage measure o remove he confounding effecs of hese facors. Following Yu (2008), we firs esimae he model below: LnCoverage LnAsses CFVolailiy Year 5 i 1 PasPerf 2 Growh 3 ExernalFinancing 4 (A1) where i indexes firm and indexes ime. Firm size is measured by he naural logarihm of oal asses; pas performance is measured by he lagged ROA; growh is measured by he growh rae of oal asses; exernal financing aciviies are measured by he ne cash proceeds from equiy and deb financing scaled by oal asses; and cash flow volailiy is measured by he sandard deviaion of cash flows of a firm in he enire sample period, scaled by lagged asses. We hen ake he residual from he above regression, label i ResCoverage, and use i as an alernaive analys coverage measure in he robusness ess. Panel E repors he resuls. In columns (1) and (2), we repor he OLS resuls. The coefficien esimaes of ResCoverage are negaive and significan a he 1% level in boh columns. In columns (3) and (4), we repor he 2SLS resuls wih ResCoverage insrumened by ExpCoverage, discussed in Secion 5.2 of he paper, o address he endogeneiy concern. We coninue o observe negaive and saisically significan coefficien esimaes. 2

3 Nex, we check wheher our resuls are robus o alernaive proxies for innovaion aciviies. While, as we discussed in he paper, R&D is a noisy measure for only one of he many observable innovaion inpus, we use i as an alernaive innovaion proxy and repor he resuls in Panel F. We find ha analys coverage is negaively relaed o a firm s R&D expendiures only in he conemporaneous year. However, we do no find analys coverage o be significanly associaed wih a firm s R&D expendiures in subsequen years. These findings are consisen wih he noion ha he shor-erm pressure imposed by analyss ends o affec innovaion inpu (e.g., R&D expendiures) quickly, bu i akes a longer ime o observe is oal effec on innovaion oupu (e.g., paens and ciaions). Third, besides he pooled OLS specificaion, we use a few alernaive economeric models o check he robusness of he baseline resuls. We sar wih he Poisson model when he dependen variable is he number of paens o ake care of he discree naure of paen couns and repor he resuls in Panel G. The coefficien esimae of LnCoverage is negaive and significan a he 1% level in column (1) when he number of hree-year-ahead paens is he dependen variable. We also use he zero-inflaed Poisson model (repored in column (2)) and find consisen resuls. If we look a he number of paens generaed five years down he road, we obain similar resuls, as shown in columns (3) and (4). Nex, since our dependen variables, paens and non-self ciaions, are righ skewed (e.g., less han half of our sample have a non-zero number of paens), we adop he quanile regression model a he 80 h percenile. 1 We repor he resuls in Panel H. The baseline resuls coninue o hold. We obain similar findings if we run he quanile regressions a he 75 h, he 85 h, or he 95 h percenile. Furher, given he non-negaive naure of paen and ciaion daa, we use he censored quanile regression (CQR) model, which places no requiremen on he disribuion of he errors and produces consisen esimaes in he presence of heeroskedasic errors for censored innovaion variables. The resuls are repored in Panel I and are robus. The coefficien esimaes of LnCoverage are negaive and significan. Furhermore, a reasonable concern is ha large firms ofen enhance heir innovaion hrough acquisiions of innovaive arge firms (Sevilir and Tian, 2012). In he meanime, analys coverage for such firms may also change afer heir acquisiions are compleed. This is because 1 Since he quanile regression model is non-linear and does no converge if firm fixed effecs are included, we demean all variables a he firm level o absorb any ime-invarian firm characerisics before running he quanile regressions. Similarly, we only conrolled for indusry fixed effecs when running he Poisson model because including firm fixed effecs will adversely affec model convergence. Due o he coun daa naure for paens, we could no demean i a he firm level for he Poisson model. 3

4 he analyss who covered he arge firm, now as a new subsidiary of he acquirer, may choose o cover he acquirer afer he ransacions (Tehranian, Zhao, and Zhu, 2010). Therefore, our baseline findings may be affeced by firms acquisiions. To address his concern, we consruc a variable, AcqAsses, which equals a firm s acquisiion expendiures normalized by is oal asses, and include i in our se of conrols. We repor he resuls in Panel J and observe consisen resuls wih our baseline findings. Finally, Panel K repors he baseline OLS resuls afer excluding R&D from he se of conrol variables. Panel L examines he effec of firm innovaion on is marke valuaion (oneyear-ahead Tobin s Q). We discuss he findings repored in hese wo panels in he paper. Overall, he comprehensive robusness checks presened in his secion for he baseline resuls repored in Secion 4 of he paper sugges ha he negaive relaion beween analys coverage and firm innovaion is robus o alernaive subsamples, alernaive proxies for analys coverage and innovaion, alernaive economeric models, and alernaive empirical specificaions. B. A possible raionale for he direc effec of analyss on innovaion While we show a direc (residual) effec of analys coverage on innovaion in he main ex of our paper, a naural quesion ha comes up is how financial analyss have he abiliy o direcly aler managers incenives of invesing in long-erm innovaive projecs. In his secion, we aemp o explore a possible raionale for he direc effec of analys coverage on firm innovaion: he adverse consequences of missing earnings arges. A large body of lieraure has shown ha conrolling for firm size and oher relevan facors, more analys coverage is associaed wih a faser and more complee price adjusmen o boh marke-wide common informaion (Brennan, Jegadeesh, and Swaminahan, 1993) and firmspecific informaion (Hong, Lim, and Sein, 2000; Gleason and Lee, 2003). These findings imply ha if a firm is followed by a larger number of analyss, is managers will have a sronger incenive o avoid missing earnings arge because such bad news will be more rapidly and fully incorporaed ino is sock price, which ulimaely reduces he managers compensaion and hurs heir repuaion and fuure career. Such adverse consequences o managers include significan declines in he firms sock prices (Barov, Givoly, and Hayn, 2002) and herefore managers sock-based compensaion, reduced CEO bonuses (Masunaga and Park, 2001), and an increased 4

5 probabiliy of managemen urnover (Mergenhaler, Rajgopal, and Srinivasan, 2011). Therefore, he adverse consequence of missing earnings arges when followed by a large number of analyss is a subsanive hrea o managers, who will in urn respond o such pressure by cuing back invesmen in innovaive projecs. To es his conjecure, we examine how marke reacions o negaive earnings surprises (i.e., when firms miss heir earnings arges) are relaed o he number of analyss following he firms. We firs consruc a sample of firms ha repor negaive quarerly earnings surprises in our sample period, i.e., he repored earnings fall shor of he consensus forecas ousanding a he earnings announcemen. We define he consensus forecas as he median earnings per share forecased by analyss in he hree monhs prior o he announcemen dae. We compue cumulaive abnormal reurns (CARs) over a wo-day [-1, 0] window around he earnings announcemen dae, as well as hree-day [-1, +1], four-day [-1, +2], and five-day [-2, +2] even windows. We use he CRSP value-weighed reurn as he marke reurn and esimae he marke model parameers over 200 rading days ending 50 rading days before he announcemen dae. 2 In an unabulaed univariae analysis, we observe srong and significanly adverse marke reacions o negaive earnings surprises (boh for mean and median values of CARs and for all four even windows considered here), which is consisen wih he exising lieraure. In Table A2, we repor he regression resuls esimaing he following model: CARs LnCoverage X Firmi Year Quarer (A2) where i indexes firm and indexes ime. The observaional uni of our analysis is firm-quarer. The dependen variables are he CARs calculaed based on he four differen even windows. LnCoverage is he naural logarihm of one plus he average number of analyss following he firm over he hree monhs prior o he earnings announcemen dae. We follow Hochkiss and Srickland (2003) o consruc a vecor of conrol variables, X, which includes: (1) he magniude of unexpeced earnings (ForecasError); (2) he price-earnings raio based on he sock price 30 days prior o he announcemen (PEraio); (3) he naural logarihm of he quarerly marke value of equiy (LnMV); (4) he quarerly Tobin s Q (TobinQ); (5) he average of annual sales growh over he prior hree years (SalesGrowh); (6) he quarerly dividend yield (DivYield); (7) he average marke-adjused sock reurn for he 12 monhs prior o he announcemen (Runup); and (8) he insiuional ownership a he calendar quarer end prior o he curren quarer (InsOwn). 2 Our resuls are robus if we use CRSP equal-weighed reurns as he benchmark reurn. 5

6 We obain quarerly firm accouning informaion from Compusa Quarerly daabase. Firm i, Year, and Quarer capure firm, year, and quarer fixed effecs, respecively. We cluser sandard errors a he firm level. The coefficien esimaes of LnCoverage are negaive and significan in all four columns, suggesing ha firms followed by more analyss experience a larger (more negaive) sock price reacion when hey fail o mee consensus earnings arges. Based on he coefficien esimaes repored in column (2), increasing analys coverage by one from is sample median (4.58) is associaed wih a 0.12% drop in cumulaive abnormal sock reurns during he hree-day period [- 1, +1], which is economically nonrivial, given ha he mean of CARs for he same window in our sample is -1.1%. 3 Overall, he evidence suggess ha more analys coverage is relaed o a larger decline in sock reurns if a firm misses is earnings arge. To avoid poenial severe consequences such as a reducion in sock-based compensaion, a loss of repuaion, as well as damage o heir fuure career prospecs, he firm managers, when followed by a large number of analyss, may cu down long-erm invesmens in innovaion o avoid missing near-erm earnings arges, which could parially explain he direc effec of financial analyss on firm innovaion. 3 For compleeness, we also examined how he CARs upon posiive earnings surprises (i.e., when a firm beas is earnings arge) are relaed o he number of analyss following he firm. We do no find analys coverage significanly affecs CARs. This finding is consisen wih he implicaion of he evidence repored in Hong, Lim, and Sein (2000) who conclude ha analys coverage makes a difference mainly for negaive news bu no for posiive news. 6

7 REFERENCES Barov, E., Givoly, D., Hayn, C., The rewards o meeing or beaing earnings expecaions. Journal of Accouning and Economics 37, Brennan, M., Jegadeesh, N., Swaminahan, B., Invesmen analysis and he adjusmen of sock prices o common informaion. Review of Financial Sudies 6, Gleason, C., Lee, C., Analys forecas revisions and marke price discovery. The Accouning Review 78, Hong, H., Lim, T., Sein, J., Bad news ravels slowly: Size, analys coverage, and he profiabiliy of momenum sraegies. Journal of Finance 55, Hochkiss, E., Srickland, D., Does shareholder composiion maer? Evidence from he marke reacion o corporae earnings announcemens. Journal of Finance 58, Masunaga, S., Park, C., The effec of missing a quarerly earnings benchmark on he CEO s annual bonus. The Accouning Review 76, Mergenhaler, R., Rajgopal, S., Srinivasan, S., CEO and CFO career penalies o missing quarerly analys forecass, Unpublished working paper. Sevilir, M., Tian, X., Acquiring innovaion. Unpublished working paper. Tehranian, H., Zhao, M., Zhu, J., Can analyss analyze mergers? Unpublished working paper. Yu, F., 2008, Analys coverage and earnings managemen. Journal of Financial Economics 88,

8 Table A1 Addiional robusness ess for baseline regressions. This able repors various robusness ess for he baseline regression. Panel A repors he resuls using he S&P 500 firms. Panels B and C repor he ess using subsamples defined by firm paening. Panel D repors he resuls using subsamples defined by differen hresholds of analys coverage. Panel E repors he resuls using an alernaive proxy for analys coverage: ResCoverage ha is consruced as he residuals from he following model: LnCoverage = α + β 1 LnAsses + β 2 PasPerf + β 3 Growh + β 4 ExernalFinancing + β 5 CFVolailiy i + Year + e, where LnAsses is measured by he naural logarihm of oal asses, PasPerf is measured by he lagged ROA, Growh is measured by he growh rae of oal asses, ExernalFinancing is measured by he ne cash proceeds from equiy and deb financing scaled by oal asses, and CFVolailiy i is measured by he sandard deviaion of cash flows of a firm in he enire sample period, scaled by lagged asses. Panel F repors he resuls using an alernaive proxy for innovaion: R&D expendiures per share. Panel G repors he resuls using he Poisson model. Panel H repors he resuls using he quanile regression model. Panel I repors he resuls using he censored quanile regression model. Panel J repors he resuls afer conrolling for a firm s acquisiion expendiures. Panel K repors he resuls afer excluding R&D from he se of conrol variables. Panel L examines he effec of firm innovaion on is marke valuaion. ***, **, and * indicae significance a he 1, 5, and 10 percen levels, respecively. Panel A: S&P 500 firms (1) (2) Dep. Var. LnPaen +3 LnCiePa +3 LnCoverage * *** (0.041) (0.039) Conrols Firm fixed effecs Year fixed effecs Observaions 5,349 5,349 R Panel B: Firms wih a leas N paens over he sample period (N = 1, 2, 3, 5) Paen 1 Paen 2 (1) (2) (3) (4) Dep. Var. LnPaen +3 LnCiePa +3 LnPaen +3 LnCiePa +3 LnCoverage ** *** ** *** (0.020) (0.021) (0.022) (0.022) Conrols Firm fixed effecs Year fixed effecs Observaions 14,141 14,141 12,699 12,699 R

9 Paen 3 Paen 5 (5) (6) (7) (8) Dep. Var. LnPaen +3 LnCiePa +3 LnPaen +3 LnCiePa +3 LnCoverage ** *** *** (0.024) (0.023) (0.025) (0.024) Conrols Firm fixed effecs Year fixed effecs Observaions 11,863 11,863 10,793 10,793 R Panel C: Firms generaing paens more han 50% of years or every year in our sample Paen Mos Year Paen Every Year (1) (2) (3) (4) Dep. Var. LnPaen +3 LnCiePa +3 LnPaen +3 LnCiePa +3 LnCoverage ** (0.031) (0.025) (0.057) (0.065) Conrols Firm fixed effecs Year fixed effecs Observaions 8,403 8,403 1,321 1,321 R Panel D: Firms wih an average of a leas N analyss over he sample period (N = 5, 15, 25) Coverage 5 Coverage 15 (1) (2) (3) (4) Dep. Var. LnPaen +3 LnCiePa +3 LnPaen +3 LnCiePa +3 LnCoverage *** *** *** (0.021) (0.021) (0.065) (0.060) Conrols Firm fixed effecs Year fixed effecs Observaions 12,456 12,456 3,422 3,422 R

10 Coverage 25 (5) (6) Dep. Var. LnPaen +3 LnPaen +3 LnCoverage (0.424) (0.319) Conrols Firm fixed effecs Year fixed effecs Observaions R Panel E: Residual coverage OLS 2SLS (2 nd sage) (1) (2) (3) (4) Dep. Var. LnPaen +3 LnCiePa +3 LnPaen +3 LnCiePa +3 ResCoverage *** *** *** *** (or insrumened) (0.016) (0.018) (0.024) (0.026) Conrols Firm fixed effecs Year fixed effecs Observaions 25,750 25,750 13,389 13,389 R Panel F: R&D expendiures per share as he dependen variable (1) (2) (3) (4) (5) Dep. Var. R&D R&D +1 R&D +2 R&D +3 R&D +5 LnCoverage *** (0.008) (0.008) (0.008) (0.008) (0.009) Conrols Firm fixed effecs Year fixed effecs Observaions 35,936 33,477 30,717 25,829 18,224 R

11 Panel G: Poisson regressions Dep. Var. Paen +3 Paen +3 Paen +5 Paen +5 (1) (2) (3) (4) Ordinary Poisson Zero-inflaed Poisson Ordinary Poisson Zero-inflaed Poisson LnCoverage *** *** * * (0.047) (0.047) (0.060) (0.062) Conrols Ind fixed effecs Year fixed effecs Observaions 25,195 25,195 22,037 22,037 Panel H: Quanile regressions (1) (2) Dep. Var. LnPaen +3 LnCiePa +3 LnCoverage ** *** (0.007) (0.010) Conrols Ind fixed effecs Year fixed effecs Observaions 25,691 25,691 Pseudo R Panel I: Censored quanile regressions (1) (2) Dep. Var. LnPaen +3 LnCiePa +3 LnCoverage ** *** (0.005) (0.007) Conrols Year fixed effecs Observaions 25,860 25,860 Pseudo R

12 Panel J: Conrolling for acquisiion expendiures (1) (2) Dep. Var. LnPaen +3 LnCiePa +3 LnCoverage *** *** (0.017) (0.019) Conrols Firm fixed effecs Year fixed effecs Observaions 24,470 24,470 R Panel K: Baseline regressions wihou R&D as a conrol (1) (2) Dep. Var. LnPaen +3 LnCiePa +3 LnCoverage *** *** (0.017) (0.019) Conrols Firm fixed effecs Year fixed effecs Observaions 25,860 25,860 R Panel L: The effec of paen and ciaions on Tobin s Q (1) (2) Dep. Var. TobinQ +1 TobinQ +1 LnPaen 0.022* (0.013) LnCiePa 0.061*** (0.008) Conrols Firm fixed effecs Year fixed effecs Observaions 42,771 42,771 R

13 Table A2 Regressions of CARs around earnings announcemens on analys coverage. This able repors regressions of he cumulaive marke model abnormal reurns (CARs) over differen windows around quarerly earnings announcemen on analys coverage when he repored earnings fall shor of he consensus forecas ousanding a he earnings announcemen dae. LnCoverage is he naural logarihm of one plus he average number of analyss following he firm over he hree monhs prior o he earnings announcemen dae. The oher explanaory variables include: forecas error (ForecasError), which is he difference beween he repored quarerly earnings and he consensus analys forecas (he median analys forecas over he hree monhs prior o he earnings announcemen dae), deflaed by he sock price 30 days prior o he announcemen; he price-earnings raio based on he sock price 30 days prior o he announcemen (PEraio); he naural logarihm of he quarerly marke value of equiy (LnMV); he quarerly Tobin s Q (TobinQ); he average of annual sales growh over he prior hree years (SalesGrowh); he quarerly dividend yield (DivYield); and he average marke-adjused sock reurn for he 12 monhs prior o he announcemen (Runup). Each regression includes a separae inercep as well as year, quarer, and firm fixed effecs. Robus sandard errors clusered by firm are displayed in parenheses. ***, **, and * indicae significance a he 1, 5, and 10 percen levels, respecively. (1) (2) (3) (4) Dep. Var. CAR -1,0 CAR -1,+1 CAR -1,+2 CAR -2,+2 LnCoverage ** *** *** ** (0.001) (0.001) (0.001) (0.002) ForecasError 0.109*** 0.207*** 0.235*** 0.256*** (0.013) (0.018) (0.019) (0.021) PEraio * ** *** *** (0.000) (0.000) (0.000) (0.000) LnMV *** *** *** *** (0.001) (0.001) (0.001) (0.001) TobinQ (0.001) (0.001) (0.001) (0.001) SalesGrowh (0.001) (0.001) (0.001) (0.001) DivYield (0.147) (0.193) (0.219) (0.225) Runup *** *** *** *** (0.012) (0.017) (0.018) (0.020) InsOwn *** *** *** *** (0.003) (0.004) (0.005) (0.005) Firm fixed effecs Year fixed effecs Quarer fixed effecs Observaions 65,830 65,830 65,831 65,831 R