Information technology and its impact on productivity: firm-level evidence from government and private data sources,

Size: px
Start display at page:

Download "Information technology and its impact on productivity: firm-level evidence from government and private data sources,"

Transcription

1 Informtion technology nd its impct on productivity: firm-level evidence from government nd privte dt sources, B I L L L E H R nd F R A N K L I C H T E N B E R G Columbi University. Introduction Robert Solow s quip tht we see computers everywhere but in the productivity sttistics hs prompted lrge nd growing literture exmining the Informtion Productivity Prdox. Some ggregte nd industry-level studies hve filed to detect positive contribution to productivity growth from investments in computer technology. More recent studies utilizing firm-level dt, however, hve detected significnt contribution. This pper confirms the results of these ltter studies using U.S. firm-level computer sset nd finncil dt for nongriculturl firms during the period from vriety of dt sources including the U.S. Census Bureu s Enterprise nd Auxiliry Estblishment Surveys, Compustt, nd Computer Intelligence Infocorp, mrket reserch firm. Our principl finding, tht computers--especilly personl computers--do contribute positively to productivity growth, suggests tht the trditionl Informtion Productivity Prdox is lrgely mesurement problem. This is closely relted to the more generl problem of ssessing service sector productivity, becuse computers re used most intensively in the service sector nd in the service functions of non-service sector firms (e.g., pyroll nd purchsing). As Zvi Griliches (994) noted, these ctivities pose the gretest problems for output nd productivity mesurement. Moreover, lthough we my see computers everywhere, they represent only tiny frction of cpitl stock (on bout % of plnt, property nd equipment), so ggregte effects my be hrd to detect (Oliner nd Sichel 993). Our dt indicte tht computers not only contributed to productivity growth during the period , but they yielded excess returns reltive to other types of cpitl. The dt lso suggest tht computer productivity peked round 3 986/987, tht computers re complementry with skilled lbour (Krueger 993 nd Autor, Ktz nd Krueger 997), nd tht use of computers my permit reductions in inventory levels. Finlly, nd perhps most interesting, the evidence Cndin Journl of Economics Revue cndienne d Economique, Vol. 3, No. April / vril 999. Printed in Cnd Imprimé u Cnd / 99 / $.50 Cndin Economics Assocition

2 336 Bill Lehr nd Frnk Lichtenberg indictes tht the types of computers nd how they re used mkes n importnt difference (Lnduer 995): personl computers pper to hve n especilly lrge impct on productivity. This finding highlights the importnce nd difficulty of mesuring computer inputs ccurtely. We interpret dt on the number of PCs or PCs per employee s providing n indiction of the extent of usge of computers within the firm. Firms tht hd more PCs per employee would be expected to be using computers more brodly cross the orgniztion (nd, presumbly, hve higher percentge of computerliterte employees), which my ccount for their higher productivity. In the ner future, once virtully everyone hs computer on his or her desk, dt on the number of PCs per employee will be less informtive nd we will need to collect even more detiled firm- or business-unit-level dt in order to ccurtely mesure the contribution of computers. The rest of this pper is divided into four sections. In section we present our econometric (production function) model. Section 3 describes our dt sources nd summrizes brod ptterns of computer usge. Estimtes of the model re reported in section 4, nd section 5 gives summry nd conclusions.. The model The essence of the productivity prdox is, tht while we seem to hve been 4 investing hevily in computers for quite number of yers, the rte of mesured productivity growth hs filed to increse, nd my hve even decresed. Since productivity is defined s output per unit of input, nd computers re n input, we should strt by sking under wht conditions one would expect growth in computer intensity to rise productivity. The contribution of computers to productivity growth my be disembodied or embodied. The embodied pproch hypothesizes tht output (Y) is n exponentil function of fctor inputs of cpitl (K) nd lbour (L) times multiplictive technology prmeter (A), which yields the following Cobb-Dougls production á -á function: Y = A K L. In this formultion, totl fctor productivity (TFP) is defined s follows: Y TFP / = A. () K á L &á According to this view, computers contribute to productivity by rising A, which mkes ll fctor inputs proportiontely more productive. Computers might hve this effect if their principl function were to improve coordintion. It is lso possible tht computers my contribute to technicl progress directly becuse they re more productive thn other types of fctor inputs. One wy to represent embodied technicl progress is to model production s: á -á Y = A[K 0 + ( + è)k ] L ()

3 Informtion technology 337 where totl cpitl (K) is decomposed into computer cpitl (K ) nd non- computer cpitl (K / K - K ), á is the elsticity of output with respect to the 0 effective cpitl stock [K + ( + è)k ], nd è is prmeter tht mesures the 0 excess productivity of computer cpitl (K ) reltive to non-computer cpitl (K ). After re-rrnging nd tking logs, eqution () cn be expressed s: 0 ln Y = ln A + á ln K + á ln( + è ( IT%) + ( - á) ln L (3) where IT% / (K /K) is the shre of computer cpitl in the totl cpitl stock. This implies the following forms for TFP nd Lbour Productivity (Y/L): 5 ln TFP ln A + áè ( IT%, nd (4) ln Y L lna% áln K L % áè(it%. (5) Eqution 4 revels tht incresed computer-intensity (IT%) would be expected to increse totl fctor productivity only if computers re more productive thn 6 other types of cpitl. Under the null hypothesis of zero excess returns to 7 computer cpitl, the first order conditions for profit mximiztion require tht the rtio of the mrginl products of computer to non-computer cpitl be equl to the rtio of the user costs of cpitl for computer to non-computer cpitl, or: MP K ' (% è) ' R ' r% ä & E(p ) P (6) MP K0 R 0 r% ä 0 & E(p 0 ) P 0 where MP is the mrginl product, R is the user cost of cpitl, r is the riskdjusted discount rte, ä is the deprecition rte, P is the purchse price per unit of cpitl, nd E(p) is the expected rte of price pprecition. A sub indictes tht the vrible is for computer cpitl while sub 0 indictes tht it is for noncomputer cpitl. The rtio of the user cost of computers to other types of 8 cpitl ought to be in the rnge of from 3 to 6. To be conservtive, we use figure towrd the outside limit of this rnge in formulting our null hypothesis tht there re no excess returns ssocited with computers: H : No excess returns ] è = 5 (7) 0 Under this null hypothesis, TFP nd (Y/L) will not depend on the shre of computer cpitl, except perhps becuse of its effect on disembodied technicl progress (vi A, which will be cptured by the fixed yer nd industry/firm-level effects in our regression nlyses). If computers re more productive then other types of cpitl, TFP nd (Y/L) will increse with the shre of computer cpitl, but the effect will be ttenuted by cpitl's overll expenditure shre, á, which is typiclly estimted to be on the order of 0-30%. The smll size of á implies tht totl fctor productivity nd lbour productivity will be reltively insensitive even to chnges in overll cpitl intensity.

4 338 Bill Lehr nd Frnk Lichtenberg Mny populr discussions of productivity focus on lbour productivity rther thn TFP. One interprettion of the productivity prdox is tht, in the lst two decdes, IT% hs ccelerted but (Y/L) hs declined. But eqution (5) indictes tht lbour productivity depends on overll cpitl intensity (K/L) s well s on the composition of cpitl (IT%), so this finding would not be prdoxicl if cpitl deepening hd decelerted. This is indeed the cse: the growth rte of (K/L) declined from 3.0% in to.0% in to.3% in The stock of computer cpitl my hve been incresing rpidly (lthough high gross investment is lrgely offset by rpid deprecition of computers), but the growth in the stock of other cpitl hs been quite sluggish. The true structure of production is much more complicted thn is indicted by the bove. For exmple, lbour is heterogeneous nd output is lso function of intngible cpitl (generted by pst R&D investment). This implies tht the right-hnd-side vribles included in eqution (3) re very incomplete subset of the entire list of determinnts of productivity. This increses the probbility tht the so-clled productivity prdox is n ecologicl fllcy: the pprent lck of simple correltion between computers nd productivity in ggregte dt should not led us to infer tht computers hve not pid off. To ccurtely ssess the mrginl productivity of computers, it is necessry to nlyze microeconomic dt, especilly firm-level dt, s Brynjolffson nd Hitt (993), Lichtenberg (995), nd others hve done. There is nother reson (emphsized by Oliner nd Sichel, 993) to believe tht using ggregte productivity dt to ttempt to ssess the returns to computer investment my be like serching for needle in hystck: even tody, computer cpitl is smll shre of totl cpitl. To illustrte this point, it is useful to consider slightly different version of the production function: lny = á lnk + á lnk + ( - á - á ) lnl. (8) In growth rtes, this becomes, YN = á KN + á KN + ( - á - á )LN (9) where YN denotes the growth rte in Y, etc. The contribution of computer growth to output growth is ákn. Moreover, in equilibrium, the elsticity of output with respect to computers should be equl to the mrginl productivity of computers times the computer-to-output rtio, or á = MP (K /Y). Even if the mrginl productivity of computers is very high nd computer cpitl hs grown rpidly, 9 K /Y is still smll (on the order of %) nd so á is quite smll. Therefore, the contribution to ggregte output growth would be smll. Another reson why we my fil to mesure productivity gins from computers is tht there my be substntil time lgs before gins re relized. Pul Dvid (990) rgues tht computers my require substntil chnges in complementry infrstructure (e.g., humn nd knowledge cpitl, globl communictions infrstructure, etc.) before the gins to them my be relized.

5 Informtion technology 339 The longitudinl nd cross-sectionl depth of the dt presented here offers vible method for ddressing these concerns. Moreover, by exmining dt in five yer increments, we reduce problems ssocited with trnsient fluctutions. Filure to dequtely cpture qulity improvements is nother importnt source of mesurement error tht tends to bis downwrds estimtes of returns to computer investment (Siegel 994). If prices ccurtely reflect qulity chnges, using sles s mesure of output will help correct the problem, but typiclly, prices do not fully reflect qulity improvements. A lrge shre of the benefits ccrue to consumers without being mesured either in higher unit sles or industry revenues. Bresnhn (986) ttempted to ddress this problem by estimting the totl socil returns to computer investment. This pproch llowed him to impute substntil socil returns to computer investment. Ceteris pribus, the impliction of this effect for our results is to bis them downwrds since we do not ttempt to mesure the effect on consumer surplus. Perhps offsetting the bove bis is the dnger tht n incresed shre of computer cpitl is positively correlted with n unobserved input tht is more directly responsible for the incresed output. Computer cpitl my be positively correlted with lbour qulity (i.e., the shre of skilled workers). Suppose output vries not only with cpitl qulity (IT%), but lso lbour qulity (y) s follows: lntfp lna + áè( IT% + ( - á)ðy (0) where y is the shre of employment tht is skilled. If we fil to tke ccount of the dependency of output on y, then we will obtin bised estimtes of á if IT% nd y re correlted. In section 4, we find evidence of positive correltion between computer use nd eduction (nd wges), which suggests tht (K /L) nd y re positively correlted. The correltion between IT% nd y will depend lso on the level of cpitl intensity since (K /L) = (K /K)(K/L) = (IT%)(K/L). If we hypothesize tht the correltion between IT% nd y is given by ã (i.e., IT% = ãy + å), then we cn estimte the upwrd bis s: * plimá = á + ã(-á)ð. () The most direct solution to this problem is to include mesure of lbour qulity mong the regressors. We follow this pproch when possible. Finlly, there is dnger tht our mesure of computer cpitl (i.e., the replcement vlue of computer hrdwre) systemticlly underestimtes computer inputs becuse it fils to reflect investments in softwre, trining, or other computer-relted expenditures. Suppose tht eqution 4 is the correct model but our estimte of the shre of computer cpitl (IT%) is too low by hlf. In tht cse, our estimte of áè would be too high by fctor of two. 0 Alterntively, one might question whether mesuring computer inputs in current prices dequtely reflects the investment in embodied technicl progress. We ttempt to control for mesurement nd omitted vrible problems in four wys. First, firm fixed effects control for time-invrint (or slowly-chnging)

6 340 Bill Lehr nd Frnk Lichtenberg unobserved vribles. Second, wherever possible, we hve ttempted to include dditionl regressors to control for potentil determinnts of productivity tht my be correlted with computer intensity. This includes the shre of uxiliry employment in totl employment, the number of estblishments, nd lterntive mesures of the composition of computer ssets (e.g., the shre of lrge systems tht re minfrmes, nd the totl number of personl computers). Third, by utilizing dt from multiple, diverse sources, we re ble to prtilly crossvlidte our results. Fourth, we dopt conservtive specifiction for the null hypothesis regrding excess returns which will help offset ny errors due to n understtement of the pproprite shre of computer cpitl in totl ssets. 3. Dt description nd trends In the following two subsections we describe our dt nd then explore trends nd other indictors of how computer usge hs chnged over our smple period. 3. Dt Description This study utilizes mixture of public nd privte dt on the diffusion nd utiliztion of computers by lrge firms ssembled from four mjor sources nd vriety of ncillry sources covering the period (see tble ). Unlike number of erlier firm-level studies which considered only mnufcturing firms, over 50% of the totl employment for our smple is in non-mnufcturing firms (see tbles 3 nd 4). Our dt re split into two longitudinl pnels, the first covering the period nd the second covering the period For the period, we use U.S. Bureu of the Census dt from the Enterprise Survey (ES) nd the Auxiliry Estblishment Survey (AUX) for 3 to 4 thousnd firms, ccounting for over 0 million employees. The ES provides enterprise-wide finncil dt, while the AUX provides similr estblishmentlevel dt for ll of firm's uxiliry estblishments. Auxiliry estblishments 3 re non-production fcilities housing wht my be thought of s the servicesector functions of the firm. These include dministrtive hedqurters, R&D fcilities, sles offices, wrehouses, etc. Ech of these surveys is conducted every five yers nd our smple includes dt for 977, 98 nd 987. The AUX dt re especilly interesting becuse of our focus on business computing (rther thn fctory utomtion). Most of firm s support services-- which contribute to wht is generlly referred to s corporte overhed --re likely to be housed in uxiliry estblishments. These dt llow us to exmine the effect of computers on firm orgniztion, s mesured by the distribution of ctivity nd employment. Becuse the uxiliry estblishments perform servicesector functions for the firm, they offer n opportunity to investigte how computers ffect service sector productivity. Additionlly, the AUX dt provide detiled informtion bering on the composition of non-production stff workers in six ctegories of employment: () centrl dministrtive nd clericl, ()

7 Informtion technology 34 reserch nd development (R&D), (3) wrehousing, (4) sles nd sles support, (5) electronic dt processing (EDP), nd (6) other uxiliry employment. This llows us to prtilly control for lbour qulity. TABLE List of mjor dt sources Enterprise Survey (Census Bureu) Enterprise Surveys for 977, 98 nd 987 offer dt on computer investments (but not computer ssets) nd other blnce sheet nd income sttement vribles for 6,000 to 8,000 firms. Auxiliry Estblishment Survey (Census Bureu) Auxiliry (i.e., non-production) Estblishment Surveys for 977, 98 nd 987 offer dt on computer investments (but not computer ssets), other blnce sheet nd income sttement vribles, nd employment by occuption for 3,000 to 38,000 uxiliry estblishments. Computer Intelligence Infocorp Compny-level dt (derived from site-level survey) for 986, 99 nd 993 on computer ssets, by type of computer, for,000 -,400 lrge U.S. firms. Compustt Enterprise dt for 986, 99 nd 993 for income sttement, blnce sheet nd other finncil vribles for firms covered by Enterprise Survey nd Computer Intelligence Infocorp dt. For ech firm, we collected dt on totl sles (Y), the book vlue of plnt, property nd equipment (PPE or K), totl investment (I), totl investment in computers (I ), nd totl employment (L). In ddition, we collected dt on the shre of investment ssocited with uxiliry estblishments, the shre of PPE in uxiliry estblishments, nd the composition of uxiliry employment. For the period , we use dt from the mrketing reserch firm Computer Intelligence (CI) nd Compustt. The CI dt include detiled informtion bout the composition of computer ssets t Fortune 000 nd 4 Forbes 400 firms for the yers 986, 99 nd 993. The computer informtion includes the estimted replcement cost of ll computer ssets (K ), s well s detiled informtion on the composition of computer cpitl (e.g., counts of the number of systems of different types such s minfrmes, minis nd PCs; the totl number of MIPS; the totl volume of DASD; etc.). We linked the CI dt with Compustt finncil dt, resulting in dt set with pproximtely,500 observtions for 500 firms ccounting for totl employment of over 6 million. Our use of both the Census nd CI dt llows us to consider longer time period thn would be possible with either source lone. Moreover, the prtil overlp in llows us to investigte the reltionship between computer

8 34 Bill Lehr nd Frnk Lichtenberg investment nd computer cpitl dt nd the reltionship between both of these nd output. Unfortuntely, the only mesure of computer inputs included in the Census dt is the level of computer investment, which provides noisy estimte of the computer shre of totl ssets. We mtched the Census nd CI dt for the 5 yers 986/987 for 757 firms in order to estimte the reltionship between the computer shres of investment nd cpitl: K K ' á%â I I. () The estimted coefficients from this regression (see tble 6) re used to bckcst the computer shre of cpitl for the unmtched Census firms in 987 nd ll of the Census firms in 977 nd Dt trends Over the 6 yers covered by our dt, the diffusion of computers into the fbric of Americn business hs been drmtic. According to the Current Popultion Survey (CPS), the likelihood tht n employee is using computer in the workplce hs nerly doubled from in 4 in 984 to lmost in by 993 (see tble ). A similr pce of diffusion is evident in the firm-level dt from the Census Bureu nd CI. From 977 to 987, computer investment per employee incresed from pproximtely $63 to $67 in nominl terms, representing nominl growth rte of 6% per yer. Due to the more rpid deprecition rte for computer cpitl (0-30% per yer), the growth in computer ssets would hve 7 been slower. A better indiction of the diffusion of computer usge is provided by noting tht in 977, only 38% of the firms in our smple reported ny computer investment, wheres 8% reported computer investments by The CI dt offer clerer picture of these diffusion trends. From 986 to 993, the men replcement vlue of computer ssets per employee incresed from $995 to $56 (see tble 4). However, during this sme period, computer intensity incresed substntilly if one considers performnce-bsed mesures. For exmple, MIPS per employee incresed -fold, DASD cpcity per employee incresed lmost 3 fold, nd the men number of PCs nd terminls per 9 employee incresed from /4 to /3. It is comforting to note how close these firm-reported figures re to those implied by the household dt on work-bsed computer usge reported in tble. Even with this substntil growth, however, it is not surprising tht computers pper to hve filed to contribute to ggregte output growth. According to the

9 Informtion technology 343 TABLE Probbility of using computer t work (Source: Current Popultion Survey) Overll: By eduction: <9th grde By household income: <$0K 8 8 $0-5K 0 4 $5-0K 8 3 $0-5K $5-35K $35-50K 48 5 $50-75K 53 6 >$75K By occuption: Mng. & Professionl Tech Sles Admin Service 0 5 Prec Prod Crft 5 3 Opertors, Lbour 0 5 Frm, Forest, Fish 4 9 By industry: Agriculture 4 Mining 3 46 Construction 3 7 Mnufcturing Trns, Comm, Util Wholesle/retil 8 37 Finnce, Insurnce 8 79 Services Forest/Fisheries Public Admin 6 74 The Current Popultion Survey (CPS) of Census Bureu is household survey for yers 84, 89, 9, 93 nd includes responses to the question, Did you use computer t work? for 55,000 households.

10 344 Bill Lehr nd Frnk Lichtenberg CI dt, the computer shre of totl ssets is tiny (pproximtely 0.3% in 993), nd the shre of Plnt, Property nd Equipment (PPE) ws only % in In ddition to the trend towrds greter computer intensity in terms of both qulity (s mesured by the increses in computing power) nd the level of finncil commitment by the firm (s mesured both by the levels of investment nd computing cpitl), there ws movement towrds more distributed rchitectures s evidenced by the substntil growth in smller systems (PCs nd minicomputers), while the numbers of minfrme computers declined. This ppers to be relted to, nd my hve fcilitted, the incresed geogrphicl dispersion of firms. According to the CI dt, between 986 nd 993, the medin growth rte in the number of sites per firm ws 46% nd the medin number of employees per site declined 7%. The Census dt provides dditionl indirect evidence of the move towrds incresed decentrliztion nd distributed systems. In 977, 6% of the employment but % of the computer investment occurred in uxiliry estblishments, wheres in 993, 7% of the employment nd only 6% of the computer investment occurred in uxiliry 3 estblishments. Computer investments re more evenly distributed cross the firm in the ltter period (see tble 3). When firms re rnked by employment size, the lrger firms tend to invest 4 more per employee in computers, lthough the difference is smll. Moreover, the gp between lrger nd smller firms ppers to hve nrrowed over time. 5 Cross-industry comprisons indicte consistent trends, lthough nonmnufcturing sectors (e.g., Services, FIRE) re more computer intensive thn mnufcturing. The cross-industry differences re consistent with computer usge dt from the CPS (tble ). Similrly, within the firm, uxiliry estblishments (the non-mnufcturing, service-sector rms of firms) re more computer intensive. For exmple, uxiliries ccount for 0% of employment but 33% of the computer investment in 987. This is not surprising becuse the first uses for business computers were for R&D nd for such bck office support services s pyroll nd ccounting -- ctivities ssocited with AUX estblishments in our dt. While ll types of workers re more likely to use computers tody, there is strong skills-bis towrd the better educted, higher pid, mngeril nd 6 professionl workers (tble ). The revolution in informtion technology my offer prtil explntion for the widening wge gp between skilled nd unskilled workers. Computers nd skilled lbour re complementry (Krueger 993; Autor, Ktz nd Krueger 997). Our dt on the reltive computer intensity of uxiliry estblishments, lrge versus smll firms, nd cross-industry comprisons (with more knowledge-intensive industries such s FIRE being more computer intensive) re consistent with these results. As finl vlidtion check, we exmined the CI dt on the composition of computer ssets by regressing the totl vlue of computer ssets (K ) ginst the

11 Informtion technology 345 TABLE 3A Census dt on computer investment Number of firms 3,38 3,734 3,74 % which report: Computer investment>0 38% 70% 8% Computer investment in uxiliry>0 3% 43% 48% Medin Vlues Employment per firm,60,535,74 $0 $8 $68 % employment in uxiliry 6% 6% 7% % totl investment in computers 0% % 3% % computer investment in uxiliries % % 6% Men Vlues (unweighted) Employment per firm 6,80 5,865 6,339 Computer investment per employee $63 $5 $67 % employment in uxiliry 8% 0% 0% % totl investment in computers 3% 6% 8% % computer investment in uxiliries 46% 39% 33% Computer investment per employee $9 $98 $339 (weighted by employment) TABLE 3B Census Firms in Smple, by Industry SIC Percent of Men Code observtions employment per (-digit) in smple firm (000s) 0 Agriculture 0.5% 8 Mining, Construction 4.0% 4 Mnufcturing 3.4% 5 3 Mnufcturing 3.3% 7 4 Trnsport, Comm, Utilities 0.8% 6 5 Wholesle nd Retil Trde 33.4% 6 6 FIRE Services.4% 9 8 Helthcre, Legl, Eduction 3.% 5 Totl 00.0% 6 For subset of firms in Enterprise Survey which report hving Auxiliry Estblishments. Becuse not ll firms were present in ll three yers, this is only pproximtely equl to the shre of firms in the smple.

12 346 Bill Lehr nd Frnk Lichtenberg TABLE 4A Corporte computer utiliztion Number of Firms Totl Employment (000s) 6,9 6,30 6,747 Medin Employment per firm 4,783 3,85 3,05 Computer Assets per employee $680 $56 $736 Minfrmes per firm 5 3 MIPS per 000 employees DASD per employee PCs nd Terminls per employee PCs per employee Computer shre of PPE.4% 0.8%.0% Computer shre of Totl Assets 0.5% 0.3% 0.3% Men Employment per firm 35,47 3,537 3,40 Computer Assets per employee $995 $87 $,56 Minfrmes per firm 8 6 MIPS per 000 employees DASD per employee 6 3 PCs nd Terminls per employee PCs per employee Computer shre of PPE.5%.6%.% Computer shre of Totl Assets 0.7% 0.5% 0.6% TABLE 4B Computer intelligence firms in smple, by industry SIC Percent of Men Code observtions employment per (-digit) in smple firm (000s) 0 Agriculture - - Mining, Construction 4.9% 8 Mnufcturing 7.8% 3 3 Mnufcturing 8.5% 40 4 Trnsport, Comm, Utilities.% 3 5 Wholesle nd Retil Trde 0.4% 65 6 FIRE 4.7% 0 7 Services.7% 34 8 Helthcre, Legl, Eduction 0.7% 50 Totl 00.0% 33 SOURCE: Computer Intelligence Enterprise-level dt for Fortune 000 firms. Becuse not ll firms were present in ll three yers, this is only pproximtely equl to the shre of firms in the smple.

13 Informtion technology 347 number of minfrmes, minicomputers, nd PCs nd terminls for 986, 99 nd 993 (see tble 5). The coefficients for these regressions provide estimtes of the men replcement vlue for ech type of equipment. First, note tht the R of these regressions declines over time, reflecting the fct tht other types of equipment (e.g., LANs nd other types of dt communictions equipment) comprise growing shre of totl computer investments. Second, note tht while the medin number of minfrmes per firm hs declined nd the number of PCs nd terminls hs incresed, there hs been little chnge in the reltive vlue shres of these types of equipment. Firms were not replcing minfrmes with PCs, but rther replcing severl older minfrmes with smller number of more powerful minfrmes nd investing in PCs nd terminls. Third, note tht the computer sset shres nd prices which we estimte with our CI dt re similr to the vlue shres nd prices reported in industry dt for domestic 7 shipments. 4. Production function estimtes In the preceding section, we documented the drmtic increse in computer usge cross ll types of firms in ll industries. We now sk whether these chnges hve contributed to productivity growth. To test this, we estimte Cobb-Dougls production functions in two bsic forms: nd lny = ã + ë + á lnk + áèx + â lnl + µ (3) it t i it it it it lny = ã + ë + á lnk + á lnk + â lnl + µ. (4) it t i 0 0,it,it it it The prmeter ã t mesures disembodied technicl chnge, ë i is fixed firm-effect (or in some cses, fixed industry-effect) tht cptures stble, unobserved firm- (or industry-) specific determinnts of productivity, nd µ it is disturbnce 8 term. The first of these equtions follows from eqution (3), while the second is stndrd Cobb-Dougls production function generlized to include two types of cpitl: computer (K ) nd non-computer (K 0) cpitl. In the following three sub-sections, we present our estimtes. 4. Estimting computer sset shre from census dt As we noted erlier, the only mesure of computer inputs included in the Census dt is the level of computer investment. We therefore estimted the reltionship between the computer shre of Plnt, Property nd Equipment nd the computer shre of investment for mtched subset of firms tht re included in both the CI 9 nd Census smples for 986/987 (tble 6). This lso llowed us to ssess how noisy proxy the computer shre of investment is for the computer shre of ssets. Computers contribute to productivity growth in (6.) nd (6.), but the

14 TABLE 5 Composition of computer ssets regressions, Replcement Vlue Computers = á (# Minfrmes) + á (# Minicomputers) + á 3(# PCs + Terminls) Regression Coefficients Est. Shre of Totl Vlue Domestic Shipments 4 (Stndrd Errors) 3 of Computer Assets Shre Avg. Unit Price Minfrmes $97,98 $9,0 $,873,866 40% 3% 38% 8% $,73,878 (65,76) (94,50) (4,06) Minicomputers $66,948 $4,346 $3,664 % 6% 6% 5% $54,5 (,708) (,70) (,89) PCs nd Terminls $,795 $849 $90 49% 5% 46% 47% $,546 () (6) (69) Men Vlue of Computer Assets ($000s) $0,605 $9,783 $7,086 00% 00% 00% 00% Number of observtions R In the regressions, the dependent vrible is the replcement vlue of computer cpitl mesured in current dollrs. The independent vribles re the number of minfrmes, minicomputers, nd PCs nd terminls. The minfrme nd minicomputer ctegories include diverse rnge of mchine types, but we did not hve dt on the composition of these ctegories. The dt re from Computer Intelligence. Stndrd errors re in prentheses below estimtes. All coefficients re significnt t the % level. 3 The shre of computer sset vlue is computed by multiplying the men count of ech type of computer by its corresponding regression coefficient. 4 SOURCE: Informtion Technology Industry Council, Informtion Technology Industry Dt Book, , tble 4-3.

15 Informtion technology t-sttistic ssocited with (I /I) in eqution (6.) is much lower. When both mesures re included in eqution (6.3), the coefficient of (I /I) is insignificnt 3 nd the coefficient on (K /K) is essentilly unchnged. TABLE 6 Census regressions, Reltionship between K /K nd I /I with (K /K) = á + â (I /I) + e á â log(sles) = á log(k) + â log(l) + ä(i /I) + ã(k /K) + e Regression eqution log(k) log(l) K /K I /I *3 è b c denotes significnt t % level, denotes significnt t 5% level, denotes significnt t 0% level. The computtions were crried out using mtched sub-smple for 987 with N = 757. K is PPE, L is employment, I is totl investment, K is replcement vlue of computer ssets, nd I is computer investment. 3 * è is the estimte of the excess productivity of computer cpitl, computed s the rtio of the estimted coefficient on K /K divided by the estimted coefficient on logk. 4. Computer productivity regressions Tble 7 includes our principl productivity regression results. First, notice tht the computer vrible is significntly positive in ll of the pooled time-series regressions. Moreover, the mgnitude of these regression coefficients demonstrtes excess returns to computer cpitl using the test described in 3 eqution (7) in ll of the pooled regressions except eqution (7.4). These findings suggest tht the productivity prdox is n rtifct of econometric mesurement error which disppers with suitbly detiled, firm-specific dt. Second, notice tht the coefficients on K nd L re close to the typicl expenditure shres nd re resonbly stble cross ll of the regressions. We cnnot reject the hypothesis of constnt returns to scle for the first two 33 regressions with industry effects. Moreover, the coefficients for the computer vribles in regressions (7.) nd (7.) re remrkbly close, despite the fct tht these re estimted from two completely different dt sets covering two different time periods. c

16 350 Bill Lehr nd Frnk Lichtenberg TABLE 7,,3 Production function regressions for the model Regression log(sles) = á0 log(k) + á IT% + â0 log(l) + e: Fixed effects Census Computer Census Computer for: Industry Intelligence Industry Firm Intelligence Firm Yers: ln(k) ln(l) b IT% (= K /K) è N R Estimted Coefficients for Shre of Computer Assets (IT%), by Yer : Estimted Coefficient on Shre of Computer Assets N R * 4 è These computtions were crried out using Census nd Computer Intelligence Dt for All regressions include fixed yer effects. Industry effects re 3-digit SIC b codes for nd 4-digit for denotes significnt t % level, c denotes significnt t 5% level, denotes significnt t 0% level. The dependent vrible is log of sles. 3 K is Plnt, Property nd Equipment (PPE); L is Totl Employment; nd IT% is the shre of computer ssets in totl PPE. Becuse we do not observe K directly for the Census regressions, IT% is imputed or predicted using the shre of investment in computers nd the regression of the computer shre of ssets (dependent vrible) ginst the computer shre of investment (independent vrible) for the mtched smple of firms which pper in both the Census nd the CI dt in 986/ * è is the estimte of the excess productivity of computer cpitl, computed s the rtio of the estimted coefficient on K /K divided by the estimted coefficient on logk. 5 Includes industry fixed effects.

17 Informtion technology 35 Third, when we introduce firm effects in (7.3) nd (7.4), the computer coefficients re reduced, but still re sufficiently lrge to support finding of excess returns. This suggests tht there re omitted vribles tht re positively correlted with computer inputs nd tht lso contribute to productivity growth. Obvious cndidtes include knowledge cpitl nd higher qulity lbour force. We explore these possibilities further below. Although we find the bove results compelling, we might be flsely interpreting the direction of cuslity, nmely, tht productivity growth drives 34 investment in computers. Brynjolfsson nd Hitt (966) tried instrumentl 35 vribles (IV) s well s ordinry lest squres (OLS). They report n even lrger productivity contribution from computer cpitl with IV thn with OLS. Also, the Husmn specifiction test filed to reject the null hypothesis tht the error term ws uncorrelted with the regressors. In the second hlf of tble 7, we show the estimte for the coefficient on the shre of computer cpitl when the model with industry effects is estimted 36 seprtely for ech yer. These results suggest tht computer productivity incresed from 977, reched pek in 986/987, nd then begn to decline. 37 This would be consistent with high djustment costs initilly followed by rpid expnsion of computer ssets which would exhust opportunities to relize excess returns from further computeriztion s firms pproched the optiml level of computer cpitl. Alterntively, the insignificnt coefficients in 977, 99 nd 993 my be due to incresed mesurement error. We hve lredy discussed how the need to estimte (K /K) for 977 using coefficients computed for 987 is likely to hve contributed to mesurement error. The potentil for incresed mesurement error during the ltter period is less obvious, but might be ttributble to n incresing shre of unmesured computer-relted purchses (e.g., investments in dt communictions equipment, softwre, nd vriety of computer services such s mintennce, etc.) in totl IT expenditure. Tble 8 presents vrious sensitivity tests using the Census dt. Regression (8.) replces (K /K) with (I /I), yielding similr conclusions but less significnt results. Regression (8.) dds the shre of employment in uxiliries. The coefficient on (L /L) is significnt, which suggests tht uxiliry employees re more productive. However, their excess productivity is less thn their wge 38 differentil, suggesting tht they yield below norml returns. Similr results re provided by regressions (8.5) nd (8.6), which further decompose uxiliry employment into seprte ctegories. Only Electronic Dt Processing (EDP) 39 employees yield excess returns reltive to other types of workers. Inclusion of these proxies for lbour qulity differences does not significntly ffect the computer coefficient estimtes, suggesting tht computers re not simply 40 proxying for unobserved lbour qulity differentils.

18 35 Bill Lehr nd Frnk Lichtenberg TABLE 8 Production function regressions, Fixed effects for: Industry Firm Industry Industry Industry Industry Industry Industry ln(k) ln(k ) ln(k ) ln(l) IT% (=K /K) IT%.7.0 (=K /K ) y (=L /L) IT3% (I /I) 0.3 b %EDP of L.849. %CAO of L %WHS of L %R&D of L 0.64 %Other of L 0.0 c %Sles of L * 3 è N 0,69 0,69 0,69 0,69 0,09 0,09 0,69 0,69 R All regressions re computed using census dt for nd with Log(Sles) s the dependent vrible. The eqution estimted include fixed yer effects. Industry effects re 3- b c digit SIC codes. denotes significnt t % level, denotes significnt t 5% level, denotes significnt t 0% level. K is Plnt, Property nd Equipment (PPE); K 3 is non-mchinery nd equipment PPE; K is mchinery nd equipment; K is computer ssets; L is Totl Employment; L is Auxiliry Employment; I is Totl Investment; I is computer investment; %EDP is electronic dt processing employment shre of L; %CAO is centrl office dministrtion employment shre of L; %WHS is wrehouse employment shre of L; %R&D is reserch nd development shre of L; %Sles is sles nd customer support employment shre of L; nd, %Other is reminder of uxiliry employment shre of L. 3 * è is the estimte of the excess productivity of computer cpitl, computed s the rtio of the estimted coefficient on K /K divided by the estimted coefficient on logk.

19 Informtion technology 353 Finlly, regression (8.3) nd (8.4) decompose cpitl into mchinery nd equipment (K ) nd structures (K 0 ). These show tht computers yield excess 4 returns reltive to other types of mchinery nd equipment. Tble 9 presents nlogous sensitivity results for the CI dt s well s productivity estimtes bsed on eqution (4). Regressions (9.) to (9.7) experiment with different wys of mesuring computer ssets. This new form for the production function requires slightly different hypothesis test, but this new test lso indictes tht there re significnt excess returns to computer cpitl, indicting robustness of our principl findings to lterntive econometric 4 specifictions. Regressions (9.) through (9.7) substitute counts of vrious computer types for the replcement vlue of computers; in ll cses we find significnt contribution from computers. Wht is perhps most interesting is the mgnitude of the coefficient on the number of PCs nd terminls. This coefficient is huge nd highly significnt. Moreover, the coefficient is unffected by inclusion of MIPS nd DASD (mesures of computer cpcity) nd is much lrger thn the coefficient on minfrmes. This suggests tht rw computing power mtters less 43 thn how computers re used. More PCs mens tht computers re distributed more widely throughout the firm nd tht users re more likely to be on networks which llow them to tke dvntge of such pplictions s electronic mil. Tble 0 repets regressions (9.) nd (9.7) by yer. Regressions (0.) through (0.3) provide further support for our finding tht computer productivity seems to hve peked in 986/987 nd declined therefter. While the coefficients re ll significnt, excess returns re erned only in the first yer. Regressions (0.4) through (0.6) show tht the coefficient on PCs nd terminls remins firly constnt nd significnt over the entire smple period. This suggests tht the reduced productivity gins from computers re not ssocited with further deployment of PCs but my be due to excessive investments in mintining legcy systems. 4.3 Inventory Regressions Computers fcilitte outsourcing nd cn enble just-in-time inventories. Computers cn lso permit firms to design, mnufcture, distribute nd inventory much wider selection of goods. If the first effect domintes, we would expect computers to reduce inventory levels. The second effect would tend to increse inventory levels. In tble we present regressions of the inventory-to-sles rtio ginst the computer shre of PPE, controlling for firm size by including totl PPE. In ll of the regressions, cross both the Census nd CI smples, the point estimtes on (K /K) re negtive lthough not lwys significnt. Perhps the most interesting of these regressions re those with fixed firm 44 effects for which results re shown in the columns lbelled.3 nd.4. In both cses, we find computers hve significnt negtive impct on inventory

20 354 Bill Lehr nd Frnk Lichtenberg TABLE 9 Production function regressions, Regression Log(L) Log(K) 0.93 Log(K 0) IT% (=K /K).6 Log(K ) Log(SYSTEMS ) Log(MAIN) Log(MINIS) Log(PCTERM) Log(MIPS) Log(DASD) R The computtions were crried out using Computer Intelligence dt for with N=,487. The dependent vrible for ll of the estimted equtions is Log(Sles). All regressions include fixed yer nd industry effects. Industry effects re 4-digit SIC codes. b c denotes significnt t % level, denotes significnt t 5% level, denotes significnt t 0% level. K is Plnt, Property nd Equipment (PPE, $millions); K 0 is PPE which is not computers ($millions); K is vlue of computer ssets ($millions); L is Totl Employment (000s); SYSTEMS is the number of minfrmes (MAIN) plus minicomputers (MINIS); PCTERM is the number of PCs nd terminls (000s); MIPS is the number of MIPS; nd DASD re the megbytes of disk storge (000s). levels. The estimted coefficient on computer cpitl in regression (.4) suggests tht n dditionl dollr of computer cpitl would llow the firm to hve bout $0.38 less in totl inventories (which represents svings of bout $0.03 per yer ssuming n interest rte of 7%). The impct of computers is likely to be more drmtic in terms of how inventories re orgnized nd mnged, rther thn on the dollr cost of those inventories Conclusions This pper hs exmined trends in computer usge nd the effect on productivity growth for cross-industry pnel of firms during the period We linked firm-level finncil nd computer sset dt for non-griculturl firms from vriety of public nd privte dt sources, including Census Bureu dt from the Enterprise nd Auxiliry Estblishment Surveys, Compustt, nd the

21 Informtion technology 355 TABLE 0 Production function regressions, Yer Log(L) Log(K 0) b c Log(K ) Log(MAIN) c Log(MINI) b Log(PCTERM) b Log(MIPS) Log(DASD) N R All computtions were crried out using Computer Intelligence Dt for The dependent vrible for ll of the estimted equtions is Log(Sles). All regression include fixed industry effects. Industry effects re4-digit SIC codes. denotes significnt t % b c level, denotes significnt t 5% level, denotes significnt t 0% level. K is Plnt, Property nd Equipment (PPE, $millions); K 0 is PPE which is not computers ($millions); K is vlue of computer ssets ($millions); L is Totl Employment (000s); SYSTEMS is the number of minfrmes (MAIN) plus minicomputers (MINIS); PCTERM is the number of PCs nd terminls (000s); MIPS is the number of MIPS; nd DASD re the megbytes of disk storge (000s). TABLE, Inventory regressions for the model log (totl inventory/sles) = á0 log(k) + á IT% Yers: Fixed effects for: Industry Industry Firm Firm ln(k) c b IT% (=K/K) N 0,7,34 0,7,34 R The computtions re bsed on Census nd Computer Intelligence Dt for All regressions include fixed yer effects. Industry effects re 3-digit SIC codes for 977- b 987 nd 4-digit for denotes significnt t % level, denotes significnt t c 5% level, denotes significnt t 0% level. K is Plnt, Property nd Equipment (PPE), K is computer ssets; nd IT% is the shre of computers in totl cpitl.

22 356 Bill Lehr nd Frnk Lichtenberg mrket reserch firm, Computer Intelligence. The Census Bureu dt cover the yers 977, 98 nd 987, while the Computer Intelligence dt cover the yers 986, 99 nd 993. The former source offers reltively rich informtion bout the composition of employment, nd dt on computer investment (but not ssets); the ltter source includes only totl employment, but rich dt on the composition of computer ssets. After linking the Census dt for 987 nd the Computer Intelligence dt for 986, we estimted the reltionship between computer investment nd the level of computer ssets, nd used this to estimte the vlue of computer ssets in the rest of the Census smple. We then estimted production functions for both the Census nd Computer Intelligence dt with both fixed industry nd firm effects. While shifting to fixed firm effects significntly reduces the mgnitude of the elsticity of computer cpitl, we still observe excess returns to computers. The reduction in the estimted elsticity is consistent with the interprettion tht computer ssets re positively correlted with unobserved firm-specific fetures tht contribute to productivity growth. These results re robust cross both dt sub-smples. Moreover, becuse the mgnitude of the prmeter estimtes ws not ffected by the inclusion of regressors intended to control for differences in lbour composition, we do not believe tht the reduction in the estimted computer elsticity with firm effects is due to unobserved differences in lbour qulity. The Census Bureu dt on uxiliry estblishments (i.e., support, hedqurters nd other non-operting business units) llowed us to explore the reltionship between firm structure, overhed nd computers. We found tht computers re complementry with uxiliry estblishment employment, but the dt pper to be too noisy to enble us to detect significnt effects of computeriztion on the composition of employment within uxiliry estblishments (which my, perhps, be regrded s comprising the within-firm service sector ). Although our firm-level nlysis found excess returns to computer investment for both mnufcturing nd non-mnufcturing sub-smples, it lso illustrted the difficulties of overcoming dt limittions when seeking to investigte the effects of computer investment on service productivity. The Computer Intelligence dt for the ltter period llowed us to investigte the reltionship between productivity (nd other operting chrcteristics of the firm, e.g., inventory-to-sles rtios) nd the composition of computer ssets. This reveled tht productivity is strongly relted to the number of personl computers used by firm, nd tht rw computing power mtters less thn how computers re used. More PCs mens tht computers re distributed more widely throughout the firm, nd tht users re more likely to be on networks tht llow them to tke dvntge of such pplictions s electronic mil. Our comprison of Census nd Computer Intelligence dt demonstrted the superiority of using dt on stocks of computer sset vlues rther thn flows of computer investment. Our nlysis lso indicted tht mere counts of minfrmes

23 Informtion technology 357 nd minicomputers do not dequtely ccount for qulity differences. The overll conclusion from this reserch is tht computers do contribute positively to productivity growth, yielding excess returns. These excess returns, however, pper to hve peked in bout 986 or 987, nd pper in both the service nd non-service sectors. Further improvements will require nlysis of firm-level nd business-unit dt, especilly since huge shre of service-sector ctivity tkes plce in service units within non-service sector firms. Computers pper to be chnging the wy in which firms re orgnized nd operted, llowing firms to become more decentrlized nd ltering employment composition. Although demonstrting tht computers yield excess returns serves to cst doubt on the trditionl version of the productivity prdox, there re still number of interesting questions. For exmple, why is there such vribility cross firms in the productivity of computers nd how computers re used? Or, why is it the cse tht productivity gins which re clerly relized t the business-unit level (e.g., when computers permit significnt hedcount reductions) often seem to fil to flow through to the firm's bottom line? Notes We would like to thnk the following: NBER Slon project on Industril Productivity for finncil support; the stff of the Center for Economic Studies t the Census Bureu for help with dt; the mrketing reserch firm Computer Intelligence for help with dt; Timothy Bresnhn nd two nonymous referees; nd workshop prticipnts t the CSLS nd the NBER for useful comments. Any errors tht remin re our own. See for exmple Biley nd Gordon (988), Lovemn (990), Morrison nd Berndt (994), Roch (987), Strssmn (990), or Wolf (997) for ppers tht fil to detect positive contribution of computers to productivity growth. See for exmple, Brynjolfsson nd Hitt (993), Lichtenberg (995), or Lehr nd Lichtenberg (997). 3 This is prtilly consistent with the findings of Morrison nd Berndt (994), who found over-investment in computers up until the 980s, but increses in the mrginl benefit-cost rtio by Investment in Office Computing nd Accounting Equipment (OCA) s shre of totl investment in non-residentil producer durbles, incresed from 5.9% to 3.% in nominl terms from 977 to 993. (Source: Bureu of Economic Anlysis, Deprtment of Commerce, Tble 5.8). 5 The equtions re pproximte becuse we re substituting è*it% for ln(+è*éô%); the two re quite close s long s è*éô% is smll. As subsequent discussion will show, since IT% is on the order of -%, è my be quite lrge nd this pproximtion will still be resonble. 6 The hypothesis tht the mrginl product for computers is positive implies tht á > 0 nd è > -. 7 The test for excess returns is much stronger thn the test of whether computers re productive. A firm operting on its production frontier ought to employ inputs up to the level where the mrginl output from n dditionl dollr of input is blnced cross ll inputs. Computers yield excess returns if dollr invested in computers