NBER WORKING PAPER SERIES ESTIMATES OF THE MAGNITUDE OF FINANCIAL AND TAX REPORTING CONFLICTS. George A. Plesko

Transcription

1 NBER WORKING PAPER SERIES ESTIMATES OF THE MAGNITUDE OF FINANCIAL AND TAX REPORTING CONFLICTS George A. Plesko Workng Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambrdge, MA August 2007 I have beneftted from comments and/or dscussons wth Jennfer Bloun, Mchael Calegar, Raj Chetty, Peter Joos, Gl Manzon, Ll Mlls, Tom Omer, John Phllps, Sonja Olhoft Rego, Davd Wesbach, Pete Wlson, and semnar partcpants at Boston College, Boston Unversty, MIT, the Unversty of Connectcut, the Unversty of Illnos Tax Symposum, the Congress of the European Accountng Assocaton, the Annual Meetng of the Amercan Accountng Assocaton, and the Amercan Tax Polcy Insttute and Natonal Bureau of Economc Research's 2006 Conferences on Fnancal Reportng and Taxaton. Ths research was performed whle I was on temporary assgnment wth the Internal Revenue Servce. Partcular thanks to Karen Cys, Ken Szeflnsk, Nna Shumofsky, and Carol Mller of the Statstcs of Income Dvson. Research support was generously provded by the Amercan Tax Polcy Insttute n conjuncton wth the Conference on the Intersecton of Fnancal Accountng and Tax Polcy. The vews expressed n ths paper are mne alone and do not necessarly reflect the vews of the Natonal Bureau of Economc Research by George A. Plesko. All rghts reserved. Short sectons of text, not to exceed two paragraphs, may be quoted wthout explct permsson provded that full credt, ncludng notce, s gven to the source.

2 Estmates of the Magntude of Fnancal and Tax Reportng Conflcts George A. Plesko NBER Workng Paper No August 2007 JEL No. C50,H25,M41 ABSTRACT Ths study examnes the tax reportng consequences of fnancal reportng dscreton. Usng a matched sample of fnancal statements wth tax returns, I provde estmates of the accuracy of tax return nformaton nferred from fnancal statements. To examne the tradeoffs between fnancal and tax reportng, I model the relaton dscretonary fnancal accountng accruals have to dscretonary federal tax accruals. The methodology takes advantage of the contemporaneous nature of reportng to mtgate econometrc problems dentfed n prevous research. I fnd the extent tax reportng reflects dscretonary fnancal reportng vares dramatcally by ndustry, proftablty, and the sgn of dscretonary accruals. I also fnd managers are able to undertake tax reducng actvtes wth less of an effect on fnancal reportng than tax ncreasng accruals, consstent wth recent evdence on the dfferental growth of book and tax ncome, and wth tax avodance actvtes. George A. Plesko Department of Accountng Unversty of Connectcut School of Busness 2100 Hllsde Road, Unt 1041A Storrs, CT gplesko@uconn.edu

3 Estmates of the Magntude of Fnancal and Tax Reportng Conflcts 1. Introducton Companes face nherently conflctng nterests n ther reportng for fnancal and tax purposes. Whle hgher fnancal reportng earnngs are generally vewed as favorable, hgher taxable ncome can result n addtonal tax labltes. Although dfferences n the amounts of ncome reported under each system have exsted snce the ncepton of the corporate ncome tax (Smth and Butters, 1949), over the past decade an ncreasng dsparty has developed between both the levels, and growth rates, of each ncome measure. In an analyss of the tax returns of large corporatons, the U.S. Treasury (1999) reported ncome for fnancal and tax accountng purposes dverged sharply durng the latter part of the 1990s, a pattern also found n analyses of publcly-avalable data by Plesko (2000b), Manzon and Plesko (2002), and Desa (2003), and n comparsons of natonal ncome account data wth tax collectons (Sullvan 1999). Plesko and Shumofsky (2004) document that aggregate pretax book ncome reported by U.S. corporatons on ther 1995 tax returns was 22 percent hgher than ts tax accountng equvalent - a dfference that ncreased to more than 59 percent n 1999, approxmately $318 bllon n the aggregate. Ths dvergence n the amount of ncome reported under each system has been offered as prma face evdence of ncreases n tax avodance actvtes of frms and of a deteroraton n the effcacy of the tax system. 1 Increasng book tax dfferences, coupled wth concerns about corporate tax complance generally, have also prompted hearng by the Senate Fnance Commttee whch ncluded a dscusson on the merts of requrng the dsclosure of corporatons 1 See, generally, Bankman (1999, 2003), Schler (2002), Talsman (2000), Wesbach (2002), and Yn (2001). 1

4 tax return nformaton to mprove the publc s knowledge of frms tax postons. Such an argument s mplctly predcated on the assumpton that the tax nformaton that would be useful to the publc cannot be nferred from the publshed fnancal statements. 2 Ths study emprcally addresses the extent of the tradeoffs between fnancal and tax reportng n an earnngs management settng. I begn by analyzng the ablty of tax nformaton nferred from fnancal statements to accurately characterze both the tax poston and tax return values. Then, utlzng a standard model of dscretonary accruals, I estmate the extent the tax reportng system reflects dscretonary fnancal reportng decsons. Dfferences n accountng rules wll yeld dfferences n both the levels, and patterns, of accruals under each reportng system, even though each s drven by the same underlyng economc actvtes. I employ an emprcal approach to explot these dfferences to obtan more precse estmates of dscretonary accruals, mtgatng concerns over omtted varables and ms-specfcaton n accruals models. Further, the correlaton of the estmated dscretonary accruals from each system provdes a measure of the extent to whch dscretonary actons n one system, such as those taken to ncrease fnancal reportng ncome, s reflected n the accruals of the other. The ntutve motvaton for ths paper s as follows. If tax and fnancal reportng were based on dentcal rules, any change n ncome or expense under one system would be perfectly reflected n the other. Under dentcal rules, tax reports could not escape beng affected by fnancal reportng-motvated actvtes and fnancal accountng choces would have mmedate tax consequences. At the other extreme, f tax and fnancal measures of ncome were unrelated (e.g., both random draws from some underlyng dstrbuton) then tax accountng ncome 2 See Grassley (2006) for an overvew of the hearng and Klenbard (2006) for a dscusson of the merts of partal dsclosure of the corporate ncome tax return. 2

5 measures would not be affected by the fnancal reportng amounts. Any acton taken to change reported book ncome would not be contemporaneously reflected n the frm s tax accounts, and would yeld no tax reportng cost. 3 In practce, and desgn, the U.S. tax system falls between these two extremes. Although both accountng systems report ncome measures derved from the same underlyng economc transactons, the patterns, and magntudes, of ncome and expenses recognzed under each system vary. Some aspects of reportng may be dentcal under both (e.g., cash sales wth no rght of return), whle others are dsparate (e.g., non-qualfed stock optons). Ths latter category s nclusve of numerous transactons that have generated concern over companes abltes to greatly reduce tax labltes wthout any fnancal reportng consequences. I fnd that many mportant corporate tax attrbutes cannot be nferred from publclyavalable fnancal reportng nformaton, a result that potentally supports arguments for greater publc dsclosure of tax nformaton by frms. Wth respect to frms ablty to dfferentally report ncome to tax authortes and n fnancal reports, I fnd the degree to whch tax reportng reflects dscretonary fnancal reportng decsons vares sgnfcantly by ndustry, proftablty, and by the sgn of dscretonary accruals. These results mply that some managers are able to recognze sgnfcant fnancal reportng ncome wthout tax consequences and undertake substantal tax reducng actvtes wthout beng subject to fnancal reportng costs. Both of these results are consstent wth recent evdence on the dfferental growth of book and tax ncome and wth tax shelterng actvtes. The paper also provdes two methodologcal contrbutons. Frst, t demonstrates the 3 Deferred tax accountng s gnored for the sake of ths example. So long as a system based upon random draws treated all dfferences as permanent the example would stll hold. 3

6 effcacy of an alternatve approach to the estmaton of accruals models that utlzes tax nformaton to obtan better estmates of nondscretonary accruals. Second, the ndustry level estmates provded n the paper can be used n future emprcal studes to provde cross-sectonal controls for the extent of book-tax conformty, smlar to country controls utlzed n nternatonal comparsons. The next secton of the paper descrbes pror research on the tenson between fnancal and tax reportng. The thrd secton presents the methodology of the paper, followed n the fourth secton by a descrpton of the data and the constructon of the sample. Secton fve presents an analyss of the matched tax return to fnancal statement data, and tests the ablty of tax nformaton nferred from fnancal statements to assess tax characterstcs. The estmaton results for the model are provded n secton sx presents, along wth a dscusson of both aggregate and ndustry-level results, and a comparson of estmaton results usng dfferent datasets and specfcatons. The fnal secton presents the conclusons. 2. Pror research The tenson between fnancal reportng and other frm objectves s a common research theme n accountng choce generally (Felds et al. 2001) and n tax (Shackelford and Shevln 2001, Maydew, 2001). In examnng the fnancal-tax reportng tradeoff specfcally, the emprcal lterature has been nconsstent n ts assumptons regardng the lnkage of the two, and studes have made conflctng assumptons. 4 For example, n examnng the effects of varous 4 Certan transactons have transparent effects that appear to nfluence behavor, such as nventory methods (e.g. Hunt et al. 1996), the management of stock optons (Matsunaga et al. 1992) and fnancng methods (Engel et al. 1999). Shackelford and Shevln (2002), Maydew (2002) and Scholes, et al. (2002) chapter 6, all dscuss the mportance of non-tax consderatons n the tax plannng process. 4

7 aspects of the Tax Reform Act of 1986, Scholes, Wlson, and Wolfson (1992) test whether frms deferred ncome to take advantage of declnng tax rates, assumng such deferrals are reflected n taxable ncome n each perod. By contrast, the earnngs management lterature surroundng the adopton of the book ncome preference of the corporate alternatve mnmum tax (Gramlch 1991, Boynton et al. 1992, Manzon 1992) assumes reductons n book ncome were generally ndependent of the amount of contemporaneous taxable ncome, snce only asymmetrc reductons would reduce the book-tax dfference of a frm. In an attempt to drectly measure the lnk between accountng choces, Erckson et al. (2002) analyzed a small sample of frms subject to SEC acton for overstatng revenue. They conclude these frms concurrently ncreased reported taxable ncome, wth frms apparently wllng to pay approxmately 11 cents n tax per each dollar of overstated earnngs, based on revsons to the tax expense. There s no ndcaton these results can be generalzed beyond ther small and unque sample, gven these frms not only aggressvely overstated earnngs but dd so n a way so as to attract regulatory scrutny. Further, gven ther relance on the tax expense, they cannot quantfy the extent to whch addtonal taxes were actually pad beyond some dsclosures of expected tax refunds. Nonetheless, that frms found t necessary to pay addtonal taxes mples that not all changes n reported book ncome can be made ndependent of taxable ncome. Other evdence of potental costs s found n Mlls (1998), who reports frms wth greater book-tax dfferences are subject to greater scrutny by the IRS and to greater proposed adjustments n ther returns. The full extent of the cost of dfferental reportng cannot be nferred from her results snce the ultmate resoluton of the adjustments was not observed and 5

8 the fnal change n taxes pad beng only a small percentage of such adjustments. However, her results demonstrate that larger book tax dfferences wll draw the attenton of tax authortes to the extent that they are observable n the tax return, and therefore a dvergence between reported book and taxable ncome may mpose addtonal tax costs on the frm. Potentally reducng ths cost, Mlls and Plesko (2003) document sgnfcant shortcomngs n the reportng of book ncome for tax purposes, leavng open the possblty that many reportng dfferences reman unreported to, and undetected by, tax authortes. Even f explct tax shelterng s not the goal other accountng practces, such as the treatment of non-qualfed stock optons, have caused substantal dfferences n the amount of ncome reported for fnancal and tax purposes - dfferences not necessarly reflected n the tax accounts of a company s fnancal reports and supported by recent analyses of aggregate fnancal and tax return data. 5 In addton to these accountng dfferences, methodologcal concerns have been rased over the emprcal approaches used to test for tax and non-tax costs (Shackelford and Shevln, 2001). 3. Methodology To examne the reportng relaton between the two accountng systems I focus on dfferences n accruals under each as the amount of cash collected by a company n any gven perod s ndependent of the accountng method. 5 Hanlon and Shevln (2000) and Manzon and Plesko (2001) both dscuss the accountng treatment of stock optons and ther dstortonary effects on the tax accounts of fnancal statements. Boynton et al. (2004), Plesko (2002, 2003), and Plesko and Shumofsky (2004) present tax-return tabulatons on dentfable sources of book-tax dfferences. 6

9 Pretax fnancal reportng total accruals for company n perod t ( F TA t, ) can be wrtten as: F TA = RBI CF (1) t, t, t, where RBI t, s pretax reported book ncome and CF t, s the frm s pretax cash flow. Smlarly, tax reportng total accruals for company n perod t can be defned as: T TA = TI CF (2) t, t, t, where TI t, equals reported taxable ncome. Assumng cash flows are the same under each accountng system, tax accruals wll dffer from pretax book accruals to the extent pretax book ncome dffers from taxable ncome, that s F T TA TA = RBI TI (3) t, t, t, t, As a result, tax accountng accruals can be estmated from fnancal accountng accruals f one knows the dfference between reported book and taxable ncome. Such dfferences wll depend on both tmng dfferences (such as the dfferences n deprecaton patterns) and scope (such as the ncluson or excluson of unrepatrated foregn ncome). 6 Followng McNchols and Wlson (1988) total accruals under each accountng system 6 Tax accountng accruals s a bt of a msnomer n that frms are not as easly able to affect taxable ncome through non-cash means. However, actons a frm mght engage n to manpulate taxable ncome wthout affectng book ncome would explot the dfferental treatment of a transacton by the accountng systems. For example, repatratng foregn earnngs wll not affect the total amount of pre-tax cash a frm has, but wll ncrease tax net ncome wthout affectng pretax book ncome. 7

10 can be decomposed nto ther non-dscretonary (NDA) and dscretonary (DA) components: F F F TA = NDA + DA (4) t, t, t, T T T TA = NDA + DA (5) t, t, t, where the estmate of NDA s defned as the predcted value from an accruals equaton, and DA s defned as the equaton s resdual. The specfcaton and estmaton of accruals models has been the subject of sgnfcant dscusson, wth crtcal revews of earnngs management methodologes performed by Dechow et al. (1995), Thomas and Zhang (1999), McNchols (2000), Felds (2001), and Kothar (2001). Thomas and Zhang (1999) compare a number of emprcal approaches to estmatng accruals models, nclusve of the specfcatons tested by Dechow et al. (1995). Whle they conclude none of the models are partcularly strong, an ndustry level verson of the Jones (1991) model was found to outperform the orgnal Jones (1991) model. As a result, a modfcaton of the Jones (1991) specfcaton s used to estmate fnancal accruals: TA A F t, 1 ΔSalest, ( t 1) PPEt TA = αf0 + αf1 + βf + γ F + ρf A A A A t 1 F, t, 2 (6) where )Sales s the change n sales from the past year (adjusted for changes n accounts recevables), PPE s gross property, plant, and equpment, and A t-1 s pror year total assets. A lag of total accruals s ncluded to capture mean reverson. For tax accruals, the emprcal 8

11 model s specfed as T T TAt, 1 ΔGrossReceptst, PPEt, TAt, 1 = αt0 + αt1 + βt + γ T + ρt (7) A A A A A t, 2 wth the explanatory varables the same as those used n equaton (6) wth the excepton of )Gross Recepts, whch s the tax accountng equvalent of sales. The predcted values from these models are taken as estmates of nondscretonary accruals, and the resduals as estmates of dscretonary accruals (DA). An mportant concern hghlghted n the revews of emprcal accruals models has been the extent to whch the estmaton of total accruals equatons, such as (6) and (7), usng ordnary least squares (OLS) are affected by omtted varables and the smultanety of the explanatory varables wth the methods avalable to manage earnngs. 7 In ths settng, pretax total accruals under each accountng system are drven by the same underlyng economc actvty of the frm. Assumng total accruals are determned contemporaneously, rather than smultaneously, an econometrcally effcent approach s to jontly estmate the two equatons usng a system of seemngly unrelated regressons. 8 In such a settng, cross-correlatons and the omtted varables affectng each equaton wll be captured n the covarance matrx, and the ndependence of the errors across the two equatons can be explctly tested. Further, unlke the resduals from OLS, the resduals from FGLS wll not be jontly correlated wth the omtted varables affectng both 7 Kang and Svaramakrshnan (1995), n partcular, propose an nstrumental varables approach to address these ssues. McNchols (1999) suggests that addtonal varables be ncluded to control for long-term growth. 8 Greene (1997) dscusses Zellner s method of estmatng a feasble generalzed least squares regresson (FGLS) of seemngly unrelated regressons. 9

12 equatons, elmnatng any nduced correlaton of concern to prevous authors. Fnally, any correlated regularty n the two dfferent accruals processes wll lead to more effcent estmates of the parameters. Thus, the use of FGLS resduals n place of OLS resduals addresses the same set of econometrc ssues as Kang and Svaramakrshnan (1995), but potentally provdes an easer to mplement alternatve to both ther nstrumental varables approach, and the need to search for addtonal covarates to nclude n Jones (1991) based models. 9 After obtanng estmates of dscretonary accruals from the jont estmaton of (6) and (7) the relaton between dscretonary accruals across the two accountng systems can be estmated as The coeffcent N wll capture mean dfferences n accruals across the two systems and T DA$ T DA$ = φ+ ω A A F (8) provdes an estmate of the amount by whch estmated dscretonary accruals for fnancal reportng are contemporaneously related to dscretonary accruals related to taxable ncome. If T=0, dscretonary tax accruals are unrelated to dscretonary fnancal accruals. If T 0, the magntude of the coeffcent wll capture the change n the tax accrual assocated wth a change n pretax book accruals, and provde a measure of the extent of tax reportng costs assocated wth earnngs management. Addtonal covarates can be added to equaton (8) to test for dfference due to other characterstcs. In order to test for asymmetry n the effects of the sgn of dscretonary accruals 9 Calegar (2000) estmates the effect of captal structure changes on dscretonary accruals through the use of a smultaneous-equaton approach. However, to derve hs estmates of dscretonary accruals he estmates hs total accruals equatons usng OLS, mplctly assumng that OLS s applcable. 10

13 and of the tax status of the frm the equaton s specfed as: DA$ T F F F F DA$ DA$ DA$ DA$ = α + + > + < + > * NEGTI + < * NEGTI 0 α1 β1 0 β2 0 γ1 0 γ 2 0 (9) A A A A A where the explanatory dscretonary accruals varable s splt nto two separate varable dependng on the sgn of DA, and two addtonal varables are created by multplyng each DA varable by a bnary varable equal to one f the frm has negatve tax net ncome n the current year. The addton of an nteracton for negatve tax net ncome allows for testng whether dscretonary accruals are affected by the current perod s tax cost (or beneft) of a change n reported taxable ncome. 4. Data I use a matched sample of frms fnancal statement and tax return data for tax years 1994 to The advantage of the data used n ths paper s the ablty to observe actual reported taxable ncome rather than an estmate based on fnancal reportng. To construct the sample, I begn wth the Internal Revenue Servce s Statstcs of Income s (SOI) annual corporaton fle, contanng tax return nformaton for more than 80,000 corporatons. 10 Tax return data are recorded as fled, and valdated for accuracy, but do not reflect any subsequent amendments or audt adjustments. Frms flng 1120-A, the corporate short form, as well as passthrough enttes, such as subchapter S corporatons, REITs, and RICs,were dropped. 10 A descrpton of the IRS sample and samplng methodology can be found n U.S. Internal Revenue Servce (2004). 11

14 Fnancal statement nformaton was drawn from Compustat 11 and matched to the tax return data by employer dentfcaton number. Non-matched frms were deleted, as were frms wth mssng or zero assets, yeldng 43,320 frm-year observatons. Of ths group, 37,853 had suffcent fnancal statement data to allow a smple comparson to the matched tax return. Summary statstcs for ths set of frms are provded n Table 1, Column (A). To reduce the effects of consoldaton, and focus solely on accountng dfferences, the value of total assets on the tax return balance sheet was compared to that reported n the fnancal statement. The balance sheet of the tax return (Schedule L) should reflect the assets and labltes of the tax flng entty, regardless of GAAP consoldaton. 12 If the dfference between the two values exceeded 0.01 of the smaller value the record was deleted, reducng the sample to 17,617 frm-years. Summary statstcs for these frms are reported n Table 1, Column ( C). The estmaton of equatons (6) and (7) requres the mposton of addtonal restrctons on the sample, yeldng a fnal sample of 6,062 frm-year observatons for 1996 to Observatons for 1994 and 1995 were lost owng to the need to have lagged values n the constructon of the regresson varables and the presence of a lagged dependent varable n the regresson. The tax return contans two dfferent measures of ncome. Tax net ncome (LINE28) s 11 Scaled pretax total accruals were calculated as pretax book ncome (data123+data16) less cash flows from operatons adjusted for deferred taxes (data308-data126). Other Compustat varables used n the regressons are assets (6), the change n sales (data 12) adjusted for the change n accounts recevables (data302), and gross property plant and equpment (7). 12 See Dworn (1985), Manzon and Plesko (2002), Mlls (1998), Hanlon (2003), Plesko (2003), Mlls and Plesko (2003), and Boynton et al. (2004) for dscussons of consoldaton ssues. The effect of mposng a consoldaton rule should be to bas the estmaton towards fndng close lnks between the two systems. Wth more complcated frms, partcularly those wth multple tax and fnancal enttes, the ablty of ether set of data to capture changes n the other wll be reduced. 12

15 conceptually equvalent to pretax book ncome; ncome subject to tax (IST) s determned after the deducton for net operatng losses and dvdends receved, but, unlke LINE28, s constraned to be nonnegatve for purposes of calculatng tax lablty. I modfy LINE28 by subtractng specal deductons (dvdends receved and net operatng losses) to obtan a measure of taxable ncome that can be negatve. 13 Because tax return nformaton s not publcly avalable, I nclude estmates of tax lablty (data 63, Current Tax Expense) and taxable ncome, GROSS63, defned as the grossed-up amount of data 63, plus the change n net operatng loss carryforwards (tem 52) and test ts ablty to proxy for tax return nformaton. 14 Gven the unque nature of the data, the next secton presents an analyss of the matched sample of tax and fnancal statement data, and s followed by a dscusson of the results of the estmaton of the accruals models. 5. The fnancal statement as a source of tax nformaton The summary statstcs provded n Table 1 allow for a test of an essental queston of nterest to a broad group of fnancal statement users: how well does the fnancal statement represent the tax characterstcs of a frm? For the sample of 37,853 matched frms, Column (B) of Table 1 provdes the results of a t-test for the hypothess that the dfference n the two seres 13 Arguably, any change n IST as defned n the tax code wll be the result of a change n tax net ncome. However, the opposte need not be true as any decrease n net ncome for a company wth negatve net ncome would not affect IST, and an ncrease n net ncome would only change IST f t were large enough to make t postve. The modfed defnton allows for the possblty that frms could ncrease the amount of ncome reported to the IRS wthout affectng ther taxable ncome. 14 Plesko (2000a, 2003) dscusses the lmts of fnancal statement nformaton to capture dfferent attrbutes of the tax return and the effcacy of alternatve fnancal statement constructs. Mlls, et al. (2003) evaluate the ablty of Compustat to capture the correct amount of net operatng loss carryforwards. 13

16 (book mnus tax) s statstcally dfferent from zero. The frst cell n Column (B) reports that the test of whether book assets less tax assets s equal to zero cannot be accepted (t = -9.67). Whle ths result s not surprsng gven the ssues of consoldaton dscussed earler, t s surprsng that mean assets reported on the tax return are greater than those reported on fnancal statements. 15 The next cell presents the results of the same test between reported recepts, and agan rejects that ther dfference s zero. Mean sales reported on fnancal statements are more than 10 percent larger than those reported on the tax return. The thrd cell n Column (B) shows the comparson of the amount of book ncome reported to shareholders and the amount reported to tax authortes on the Schedule M-1. In contrast to the asset and gross recepts results, the values for pretax book ncome are not statstcally dfferent from each other (t=0.71). The next two cells report a smlar result n comparsons of the fnancal statement estmate of taxable ncome, GROSSS63, to both Tax Net Income and Income Subject to Tax - n nether case can the hypothess be rejected. Whle these results suggest that, n large samples, the fnancal statement provdes nformatve data regardng frms taxable ncome, ths fndng s puzzlng, gven that both assets and revenues are statstcally dfferent. Further, whle grossng-up the current tax expense appears to approxmate taxable ncome, the current tax expense does not provde a statstcally relable estmate of frms actual tax lablty, ether before or after credts. The mean of the current tax expense (18.23) falls between the reported amount of tax before credts (20.16) and tax after credts (14.88), and t-tests reject the hypotheses that the dfference between the book amount and 15 Whle surprsng, ths result s not anomalous, and has been prevously reported by Mlls et al. (2002) and Manzon and Plesko (2007). Boynton et al. (2004) provde a detaled examnaton of consoldaton dfferences for tax and fnancal reportng and the effects on the nterpretaton of reported book-tax reportng dfferences. 14

17 the tax amounts are zero. Column (D) reports the analyss for the sample of 17,617 frm-year observatons retaned because of the smlarty n reported assets. The smaller sample sze - more than half of the observatons n Column (A) are dropped - hghlghts the dffculty n makng proper nferences about frms tax attrbutes from publcly avalable data, snce the reportng enttes are generally dfferent. Whle the mean amount of revenues are statstcally equvalent (t = 0.81), the average amount of pre-tax book ncome reported to shareholders s statstcally greater than the amount reported on the tax return (t=3.47). Smlarly, the amount of taxable ncome nferred from the fnancal statements s statstcally greater than ether reported tax net ncome (t=13.83) or ncome subject to tax (t=6.63), suggestng that nferences of taxable ncome systematcally overstate the actual amount of taxable ncome for smlarly consoldated enttes. Worth notng here s that ncome subject to tax s larger than tax net ncome snce tax net ncome can be negatve whle ncome subject s bounded below by zero. Ths pattern contnues nto the estmates of actual tax lablty, ether before or after credts, wth the average amount nferred from fnancal statements statstcally greater than actual tax before, or after, credts (t= 2.71 and 6.45, respectvely). Fnally, Column (F) provdes a smlar analyss for the smaller sample of 6,062 frm-year observatons retaned because of the smlarty n reported assets and the avalablty of data to be used n testng the central questons of ths paper. The results for ths sample are smlar to those reported n Column (D), wth only gross recepts reported under the two systems not statstcally dfferent from each other (t=0.38). Book ncome reported to shareholders s greater than reported on the tax return s Schedule M-1 (t=1.87), as s estmated taxable ncome (GROSSS63 s 15

18 statstcally larger than both tax net ncome (t=9.84) and ncome subject to tax (t=6.00)). The values of tax lablty, ether before or after credts, are both smaller than the amount reported as the current tax expense, and each s statstcally dfferent from the value nferred from the fnancal statements. Tables 2 and 3 provde further evdence on the effcacy of nferrng the tax attrbutes of frms from fnancal statements by comparng the sgns, rather than the levels, of taxable ncome and tax labltes. Table 2, for each of the three samples, presents a cross tabulaton of the sgn of tax net ncome (LINE28) wth the sgn of taxable ncome that s nferred from the fnancal statements. Arguably, smlarty n the sgns of these varables s a lower standard than the equvalence n magntude, but a frm s current tax poston s an mportant determnant of behavor. The results are smlar for all three comparsons: between 12 and 13 percent of frms n the sample are ms-classfed, and n all cases the majorty of the ms-classfcatons usng fnancal statement nformaton mply that a frm has taxable ncome when, based on the tax return, t does not. The greatest mbalances n ms-classfcaton occur n the two cases where condtons have been mposed to ensure that the enttes are the most smlar, wth more than twce as many frms ncorrectly classfed as taxable than ms-classfed as non-taxable. Table 3 repeats the comparsons of Table 2, but focuses on whether the sgn of frms tax lablty (ether before or after credts) s correctly nferred from the fnancal statements. Ths addtonal comparson s necessary because a frm may have no lablty even when taxable ncome s postve (prmarly because of credts), or because a tax lablty may exst even when taxable ncome s zero (owng, for example, to non-ncome, recapture, or the mnmum tax). The pattern of ms-classfcaton for a current tax expense s the smlar to that reported n Table 16

19 2: ms-classfcatons n each of the sx cases prmarly label frms as taxable when they are not. The ms-classfcaton s less pronounced n tax after credts than tax before credts, but remans above 10 percent n each of the three samples. The data presented n Tables 1 through 3 provde strong evdence that fnancal statement nformaton yelds msleadng nferences about frms tax attrbutes. These results, encompassng many more years and observatons than prevous research, are consstent wth pror conceptual (Hanlon, 2003, Mlls and Plesko 2003) and emprcal (Dworn, 1985, Plesko 2001, Mlls et al. 2002, McGll and Outslay, 2004) analyses. At a mnmum, these results suggest that fnancal statement-based tax nferences must be nterpreted and used wth cauton. Wth respect to th current polcy debate over the magntude, and trend, of book-tax dfferences, the results of Tables 1 through 3 show that measures constructed from fnancal statements wll understate the magntude of reportng dfferences. The extent to whch fnancal statement constructs may adversely affect the results of emprcal research s addressed n the next secton. 6. Model estmaton A. General results Estmaton of the accrual models requres addtonal fnancal statement nformaton from each frm and the mposton of constrants on the data, leadng to a fnal sample of 6,062 frmyear observatons. Whle some of ths data has been descrbed n Column (E) of Table 1 and further n Tables 2 and 3, summary statstcs for the addtonal varables are presented n Table 4, Panel A. To obtan total accruals under tax accountng the dfference between pretax book and taxable ncome was subtracted from book total accruals. For tax accruals based on 17

20 Compustat tem 63, GROSS63, the dfference between estmated taxable ncome and book ncome was subtracted from book total accruals. Panel B of Table 4 provdes the correlatons between pretax book ncome (Compustat data tem 170), tax net ncome (LINE28) and GROSS63, the fnancal statement estmate of taxable ncome. The frst column provdes the correlatons between pretax book ncome and the two measures of taxable ncome. Whle the equalty of means between these varables was rejected n Table 1, Panel B shows that all of the correlatons are statstcally sgnfcant at one percent. The correlaton for GROSS63 s hgher than for modfed taxable ncome (lne 28), wth correlatons of 0.78 and 0.70, respectvely, a non-surprsng result snce GROSS63 and pretax book ncome are drawn from the same fnancal statement. The second column of the panel shows that the correlaton between tax net ncome and GROSS63, the fnancal statement measure of the same tem s The thrd panel of Table 4 provdes the correlatons of scaled total accruals for each defnton. Of the two tax ncome accruals, GROSSS63 s more hghly correlated wth pretax book ncome accruals - agan not a surprsng result gven that both are drawn from the same fnancal statements. Table 5 provdes a descrpton of the ndustry dstrbuton of the sample along wth the correlatons of the resduals from jontly estmatng equatons (6) and (7). 16 Two sets of systems of equatons are estmated: System 1 utlzes tax return nformaton to measure the effects on tax reportng whle System 2 substtutes tax return nformaton nferred from fnancal statements. Whle t s usual for both regulated ndustres and fnancal servces to be omtted from accruals models, I nclude both as benchmarks. Regulated ndustres are often excluded because the 16 Coeffcent estmates from the accruals equatons have been omtted n the nterest of space. 18

21 nature of the regulatory process makes them less nterestng to study. In ths settng, however, regulated ndustres should have relatvely hgh correlatons, and ther ncluson allows for ths hypothess to be examned. Smlarly, whle the Jones (1992) model s vewed as less representatve for non-manufacturng ndustres, the applcablty of the model can be drectly tested. Further, any ms-specfcaton due to omtted varables n the accruals model wll be mtgated through the use of FGLS. The correlaton of the resduals from the frst-stage OLS regresson provdes nformaton on whether the assumptons of the OLS model are volated and an mprovement can be made va jont estmaton. A Breusch-Pagan test of the ndependence of each set of resduals was performed for each set of equatons, and ndependence of the resduals can be rejected n all but one case (Publc Admnstraton for GROSS63), confrmng the approprateness of the FGLS approach over OLS wth ether set of data. 17 For the remanng tests, the resduals from the FGLS estmates are used as estmates of dscretonary accruals, where BOOKDA represents dscretonary book accruals, and LINE28DA and GROSS63DA represent taxable ncome accruals estmated, respectvely, usng tax return or fnancal statement nformaton. Table 6 provdes the Pearson correlaton coeffcents of the FGLS dscretonary accruals estmated at the ndustry level, and then pooled. Coeffcents n bold were estmated wthn the same system of equatons. Wthn the frst system, whch uses tax return nformaton, the correlaton of fnancal and tax dscretonary accruals s and statstcally sgnfcant. GROSS63, however, yelds a much lower correlaton (0.071) wthn the sample. The estmates of BOOKDA across the two systems are hghly correlated (0.993), and LINE28DA s nearly equally 17 Here and throughout the rest of the text, acceptng or rejectng statstcal tests are relatve to a 1 percent sgnfcance level. 19

22 correlated wth BOOKDA n each model (0.472 and 0.466). Of note, the correlaton between LINE28DA and GROSS63DA s only Table 7 presents the full sample results from estmatng equaton (8), the effect of book dscretonary accruals on tax dscretonary accruals. Because frms potentally face mmedate costs n recognzng ncome ncreasng accruals for tax and mmedate benefts n recognzng ncome decreasng accruals, the specfcaton of equaton (8) allows for the explct testng of ths asymmetry. Whle " 0 provdes the ntercept for the entre sample, " 1 s the coeffcent on a separate bnary varable equal to one f the observaton has negatve taxable ncome, and controls for any dfference n mean dscretonary accruals of negatve taxable ncome frms. To allow for dfferences n the pattern of ncome ncreasng and ncome decreasng accruals the dscretonary accruals varable, BOOK28DA s splt nto two based on sgn, BOOK28DA>0 and BOOK28DA<0, and potental dfferences n ther effects can be tested by comparng the coeffcents on each varable, $ 1 and $ 2. Snce the tax costs of a taxable ncome ncreasng accrual are smaller for frms wth negatve tax net ncome an addtonal varable s ncluded that nteracts the dscretonary accrual varables (BOOK28DA>0 and BOOK28DA<0) wth a bnary varable equal to one f the frm has negatve tax net ncome. The coeffcents on these varables, ( 1 and ( 2, are estmates of the addtonal effect tax status has on the reflecton of book dscretonary accruals n tax dscretonary accruals. The combned effects of $ 1 + ( 1 and $ 2 +( 2 can be compared to test f tax status has a statstcally dfferent effect on accruals behavor. Gven the systematc dfferences n the data taken from tax returns and that from 18 Gven the dfference n total accruals s derved from the amount of the book-tax ncome dfference t s possble that the dfference n dscretonary accruals s also determned by the magntude of reported ncome dfference. For System 1, the correlaton of the dfference n dscretonary accruals between book and tax to the aggregate book-tax ncome dfference s only 0.140, and for System 2. 20

23 fnancal statements, Equaton (8) s estmated for two sets of data: n Column (1) the estmaton uses matched tax and fnancal statement data whle n Column (2) the estmaton s performed usng publcly-avalable data only. Ths sde-by-sde comparson provdes an easy was to assess the effects of the dfferences n data on the outcome of the estmaton. 19 In both equatons " 0 s postve and sgnfcant, and " 1 s negatve and sgnfcant, mplyng that frms that are not taxable have, on average, lower (and negatve) dscretonary tax accruals. The test of a central queston of ths paper, whether ncome ncreasng and decreasng dscretonary book accruals are reflected to the same extent n taxable ncome, s found n the coeffcent estmates for $ 1 and $ 2. Both sets of data yeld the same result: dscretonary accruals that ncrease book ncome are reflected to a lesser extent n taxable ncome than dscretonary accruals that reduce book ncome ($ 1 < $ 2 ). For the tax return sample n Column (1), the estmated coeffcent for book-ncreasng dscretonary accruals mples that for each dollar of ncome ncreasng dscretonary accrual recognzed for fnancal reportng purposes, taxable ncome s ncreased by dollars. By contrast, frms wth ncome decreasng accruals are estmated to reduce taxable ncome by of the amount. The relatve magntude of these coeffcents s consstent wth frms ether takng advantage of opportuntes to recognze greater fnancal reportng ncome when the tax reportng effects are small, or frms mnmzng the tax effects of ncreased fnancal reportng ncome through other mechansms Whle the data used n Column (2) are collected from publc sources, ncluson n the sample s based on propretary knowledge of the frm's tax consoldaton. The smlarty n consoldaton should bas any emprcal results to be more smlar than would be expected wthout the constrant, and as such provdes an upper bound on the smlarty n results that a researcher would obtan f restrcted to usng publcly avalable data. 20 The mpled amount of addtonal tax pad, calculated as the estmated coeffcent for taxable frms (0.341) multpled by the statutory tax rate (0.35), s 0.119, smlar to the value found by Erckson et al. (2002). Ths calculaton s also an aggregate estmate of Shackelford et al. (2007) parameter of nterest, (M 2 - *M 3 )/M 1, representng the cash flow effects owng to taxaton of reportng addtonal book ncome. The results here, and n 21

24 Smlarly, to the extent that frms recognze ncome decreasng accruals, they appear to be able to take advantage of the tax system to concurrently reduce taxable ncome. For the fnancal statement-only data, reported n Column (2) the results are statstcally dentcal. The next two coeffcents, ( 1 and ( 2, capture the ncremental effects of a frm havng negatve taxable ncome on the magntude of dscretonary accruals. For the tax return sample, the coeffcent on book ncome ncreasng accruals (( 1 ) s 0.552, mplyng that a much greater share of each dollar of addtonal book ncome, a total of ( ), s reflected n taxable ncome when a frm s able to recognze the addtonal taxable ncome wth lttle or no change n ther tax lablty. In the case of ncome decreasng accruals the estmate of ( 2 s , mplyng that a smaller porton of the decrease n book ncome s reflected on a frm s tax return when they are unable to beneft from a concurrent reducton n taxable ncome. The F- tests of the equalty of coeffcents are provded at the bottom of the table, and show the dfferences between the coeffcents on the accruals varables, $ 1 and $ 2, and the accruals varables nteracted wth tax status, $ 1 + ( 1 and $ 2 +( 2, are statstcally sgnfcant. Reflectng the dffcultes n dentfyng non-taxable frms documented n Tables 1 through 3, the nferences from ( 1 and ( 2 dffer across the two samples. In partcular, ( 2, whch captures the effect of an ncome decreasng book accrual on taxable ncome for a loss frm s 2.5 tmes the magntude reported n Column (1) ( v ), and the sum of $ 2 and ( 2 s , mplyng that book decreasng dscretonary accruals are assocated wth ncreases n taxable ncome n the same perod. Further, the adjusted R 2 n the publc data sample s substantally lower than n the tax return matched sample, 0.04 versus the next secton, show that ths trade-off vares by both tax status and ndustry. 22

25 In summary, the aggregate results provde evdence that ncome-decreasng accruals are reflected to a greater degree on the tax return than ncome ncreasng accruals when the tax savngs from such decreases are the greatest ($ 2 > $ 1 ). However, when tax costs are reduced owng to a current perod loss, ncome-ncreasng accruals are contemporaneously reflected to a greater degree n current taxable ncome ($ 1 + ( 1 > $ 2 +( 2 ). 21 B. Industry effects Heterogenety n accountng rules and/or practces suggests that the aggregate results descrbed above may not be representatve of every ndustry. To explore possble dfferences across ndustres Table 8 provdes separate results for the non-manufacturng and manufacturng sectors along wth a number of subndustres (nne subndustres for nonmanufacturng and nneteen for manufacturng). The frst column of Table 8, Panel A, provdes aggregate results for non-manufacturng frms, and are smlar to the results for the entre sample. The estmated coeffcent for book ncome ncreasng dscretonary accruals (BOOK28DA>0) mples that for each dollar of ncome ncreasng dscretonary accrual recognzed for fnancal reportng purposes, taxable ncome s ncreased by only dollars. By contrast, frms wth ncome decreasng accruals are estmated to reduce taxable ncome by of the amount. Smlar to the results reported n Table 7, the coeffcent estmates are consstent wth ncome ncreasng and decreasng dscretonary accruals beng reflected dfferently n taxable ncome, and that frms recognze 21 Ths result s consstent wth, but more general, than Calegar (2000), who found that frms use accruals wth low book-tax conformty to manage book earnngs and hgh conformty to manage tax. The results here show the degree of conformty between book and tax accruals depends on the tax status of the frm. 23

26 greater fnancal reportng ncome when the tax reportng effects are small. Smlarly, when frms recognze ncome decreasng accruals, they appear to be able to take advantage of the tax system to concurrently reduce taxable ncome. The effects of tax status (( 1 and ( 2 ) are also smlar to those reported for the entre sample, wth ncome ncreasng accruals beng reflected to a much greater extent on the tax return when the tax costs are low (0.926 versus 0.337). Columns 2 through 10 of Panel A provde the results for each two-dgt SIC nonmanufacturng ndustry group. In all cases, $ 1, the coeffcent on BOOK28DA>0 s postve and statstcally sgnfcant, and ranges from (servces) to (wholesale trade). Consstent wth expectatons, regulated ndustres has a coeffcent statstcally ndstngushable from one (0.985), consstent wth a close regulatory tes between fnancal and tax reportng of postve accruals. For $ 2, the coeffcent on ncome decreasng accruals, all of the coeffcents are postve and all but two (for Mneral Industres and Publc Admnstraton) are sgnfcant. In four of the nne ndustres the hypothess that $ 1 = $ 2 cannot be accepted, and n three of those four ndustres (Agrculture, Mnerals, Regulated Industres) $ 1 > $ 2. Only n Servces s $ 2 estmated to be sgnfcantly greater than $ 1. The effects of negatve taxable ncome also dffer across the ndustres. For ncome ncreasng accrual frms ( 1 s sgnfcant n fve of the nne ndustres, but postve n only one, Servces, suggestng that the aggregate result may be strongly nfluenced by ths ndustry (Servces comprse 30 percent of the sample). The coeffcent on the nteracton of negatve taxable ncome wth ncome decreasng accruals s negatve and sgnfcant n two of the ndustres (Constructon and Servces) and postve and sgnfcant n Regulated Industres. Wth respect to the combned coeffcents, $ 1 + ( 1 and $ 2 +( 2, only two of the ndustres 24

27 combned coeffcents are estmated to be statstcally dfferent, Regulated Industres and Servces. For Regulated Industres, ncome ncreasng accruals of frms wth postve taxable ncome ($ 1 = 0.985), and ncome decreasng accruals of frms wth negatve taxable ncome ($ 2 +( 2 = 1.025) are statstcally ndstngushable. Panels B and C of Table 5 provde estmates for the manufacturng sector n total, and separately for each of the 19 sub-ndustres. Whle both $ 1 and $ 2 for the manufacturng sample (Column (1)) are postve, as n pror tables, they are not statstcally dstngushable from each other. However, the sum of the coeffcents of $ 1 + ( 1 and $ 2 +( 2 are sgnfcantly dfferent, and are consstent wth both the aggregate results and those reported for non-manufacturng frms. In sum, for the manufacturng sector as a whole, as for the aggregate non-manufacturng sector, ncome decreasng accruals are reflected dfferently on the tax return than ncome ncreasng accruals, wth ncome ncreasng accruals reflected on the tax return to a greater extent when the tax cost of dong so s small. As n the non-manufacturng ndustres, ncome decreasng accruals are reflected less on the tax return when the frm has a tax loss than when t s currently taxable. Examnng the ndvdual ndustres, all but one of the estmates of $ 1 are postve and sgnfcant, wth the excepton of Leather whch s not sgnfcant, and has only 13 observatons. All but three of the estmates of $ 2 (Leather, Instruments, and Paper) are sgnfcant, and all are postve. Comparng $ 1 to $ 2, n eleven of the nneteen ndustres the coeffcents are not statstcally dfferent from each other. In the remanng eght, n all but one ndustry (Electrcal Equpment) $ 1 s larger than $ 2, opposte the result found n the aggregate, and mplyng that for these ndustres ncreases n book ncome are reflected to a greater degree n taxable ncome than 25

28 decreases. Examnng the effects of the nteracton wth tax status, $ 1 + ( 1 and $ 2 +( 2, the sums of the coeffcents are estmated to be statstcally dfferent n 5 of the 19 ndustres, and n all but one of the fve the results are consstent wth the aggregate results for manufacturng: book ncome ncreasng accruals are reflected n taxable ncome to a greater extent when the tax cost s small. The excepton, Transportaton Equpment, yelds a large negatve coeffcent on ( 1, mplyng that ncreases n book accruals are assocated wth contemporaneous decreases n taxable ncome of frms that already have negatve taxable ncome. Taken together, the results of Table 8 provde evdence that fnancal reportng dscretonary accruals and tax dscretonary accruals are related, that the relaton vares sgnfcantly by ndustry, and that ncome ncreasng book accruals are reflected less n taxable ncome than ncome decreasng accruals unless the tax costs are small. The extent of the nterndustry dfferences can be sgnfcant, and suggest that emprcal analyses should allow for dfferences n the conformty of book and tax reportng beyond the common practce of omttng regulated ndustres (because they are asserted as beng hghly conformng) and fnancal nsttutons (because of ther general accountng dfferences). Smlar to controllng for country dfferences n book-tax conformty n the nternatonal lterature (e.g., Hung, 2000), the coeffcent estmates n Table 8 provde a set of varables that can be used to control for crosssectonal dfferences n conformty n emprcal accountng research usng Compustat or other U.S. data. Elmnatng the use of propretary data, Table 9 presents the estmated relaton between book and tax accruals when a fnancal statement-based measure of taxable ncome s used n 26