PRELIMINARY DRAFT Please do not quote

Transcription

1 PRELIMINARY DRAFT Please do not quote Market Structure, Competon and Productvy Growth: Evdence from Canadan Manufacturng Industres Ram C Acharya Industry Canada May 2005 Abstract Usng data for 86 Canadan manufacturng ndustres from 1990 to 2002, ths paper estmates total factor productvy (TFP) growth allowng for non-constant returns to scale and mperfect competon. Somewhat surprsngly, the results show that we cannot reject the null hypothess of constant returns to scale for 80 ndustres and that of perfect competon for 55 ndustres. Although at the aggregate level the standard Solow resdual underestmates the actual TFP growth by only 0.2 to 0.4 percentage ponts, these two seres vary substantally at the ndustry level. The paper also shows that competon, as measured by lower level of rents and markup and hgher level of mport penetraton, s assocated wh a sgnfcantly hgher rate of TFP growth. Smlarly, the fall n the Herfndahl ndex and concentraton rato (proxes for a rse n competon) lead to a hgher productvy level effect. Correspondence: Ram Acharya Industry Canada 235 Queen Street, 10 th Floor East Ottawa, ON K1A 0H5 Emal: acharya.ram@c.gc.ca Vews expressed n ths paper are those of the author and do not necessarly reflect those of Industry Canada. In the paper, I have benefed from dscussons wh Serge Coulombe, for whch I am thankful. Of course, all errors are mne.

2 2 Market Structure, Competon and Productvy Growth: Evdence from Canadan Manufacturng Industres Usng data for 86 Canadan manufacturng ndustres from 1990 to 2002, ths paper estmates total factor productvy (TFP) growth allowng for non-constant returns to scale and mperfect competon. Somewhat surprsngly, the results show that we cannot reject the null hypothess of constant returns to scale for 80 ndustres and that of perfect competon for 55 ndustres. Although at the aggregate level the standard Solow resdual underestmates the actual TFP growth by only 0.2 to 0.4 percentage ponts, these two seres vary substantally at the ndustry level. The paper also shows that competon, as measured by lower level of rents and markup and hgher level of mport penetraton, s assocated wh a sgnfcantly hgher rate of TFP growth. Smlarly, the fall n the Herfndahl ndex and concentraton rato (proxes for a rse n competon) lead to a hgher productvy level effect. 1. Introducton The frst theorem of welfare economcs states that competve economy generates an effcent allocaton of resources. As stated n Nckell (1996), the expectaton regardng the effect of competon s not lmed to ths well-known result; has gone much further. The general belef s that competon rases effcency by exertng a downward pressure on producton costs and enhances nnovaton by provdng ncentves for frms to nnovate. In a smlar manner, Baly and Gersbach (1995) state that vgorous global competon urges allocatve effcency and forces structural change n ndustres. In contrast, there s also lerature whch argues that competon may lead to lower productvy growth. Ths reasonng s more n lne wh Schumpeteran hypothess that monopoly rents enable frms to spend more on nnovatve actves and rase productvy. For example, Kemen and Schwartz (2002) argue that a frm wh monopoly power s more lkely to strve for persstent domnance than a frm whout non-monopoly power. 1 1 The studes lnkng competon and nnovaton started wh the work of Schumpeter (1942) where he stated that the possbly of future monopoly rents s requred to motvate frms to nnovate. On the contrary, Boldrn and Levne (2004) argue that monopoly s neher needed for, nor a necessary consequence of, nnovaton. In a recent study,

3 3 Whch way does the relatonshp between competon and productvy go? Snce mprovement n the productvy s the only long-term soluton for a hgher standard of lvng, f more competon leads to hgher productvy, then fosterng competon would be a desred economc polcy. 2 In general, competon can affect productvy n three ways. Frst, can make exstng frms more competve. Second, as suspected by Nckell (1996) competon may work not by forcng effcency on ndvdual frms but by lettng many flowers bloom and ensurng only the best survve. Thrd, wh more frms competng n the market, the elastcy of substuton among products may ncrease thereby rasng the prce elastcy of demand and hence ncreasng the potental benef for wnners. The potentally larger profs nduce hgher manageral effort, creatng ncentves for effcency-enhancng actves such as nnovaton. There are two man strands of lerature on competon and productvy. The frst examnes the effects of competon on the effcency of frms managed by agents under mperfect contracts. 3 Ths lerature deals mostly wh frm s management aspects. The second strand of the lerature examnes the effect of competon on total factor productvy (TFP). Nckell (1996) s the frst paper to study the relatonshp between competon and TFP drectly. Usng around 700 U.K. manufacturng companes over the perod , he shows that competon measured by ncreased numbers of competors (nformaton collected by surveys) or by lower levels of rents s assocated wh a sgnfcantly hgher rate of TFP growth. Recently, Zzewz (2003) usng data on U.K. and U.S. tobacco ndustres show that technologcal nnovaton and consoldaton of producton were more rapd durng competve perods. Aghon et al. (2002) fnd that the relatonshp between competon and nnovaton s of nverted U shape; nnovaton rses wh the rse n competon untl a certan range; f competon rses further, nnovaton starts to fall. 2 Although productvy growth s not the same thng as welfare, f the resources released by productvy growth fnd adequate employment n other economc valuable sectors ncludng lesure, then productvy growth leads to hgher welfare (Grlches, 1994). 3 For ths type of studes, see Aghon et al. (1997), Schmdt (1997), Horn et al. (1994) and lerature ced n these papers.

4 4 Usng mcro data for U.K. manufacturng establshments for , Dsney et al. (2003) study the mpact of competon on TFP growth based on the framework developed by Nckell (1996). They fnd that a fall n market share and rents, whch they nterpret as a rse n competon, rases productvy levels. They also fnd that frms wh a hgh level of rents wll have low productvy growth. In the Canadan context, usng frm s perceptons of ther competve envronment from Statstcs Canada s 1999 Survey of Innovaton, Tang and Wang (2004) show that product market competon has a posve mpact on the performance of medum-szed and large-szed frms. These are the only papers we came across n ths feld. Even though may not be the exhaustve lst, s que evdent that more research s requred n ths area. Furthermore, the prevous studes estmate the relatonshp between productvy growth and competon varables whout explcly solatng the effects of potental non-constant returns and mperfect competon from the TFP growth seres. The mgaton of the mpact of non-constant returns and mperfect competon from TFP growth s mportant as we know from the semnal contrbuton of Hall (1988) that n ther presence, the usual Solow resdual (TFP growth) s a bas estmate of actual unobserved productvy growth. The basc framework n Nckell (1996) starts wh the assumpton of constant returns to scale. In the case of Dsney et al. (2003), although they do not make any assumpton regardng competon and scale, ther framework s not desgned to take out the effects of markups and nonconstant returns from each economc un. To address ths subtle ssue, n our estmaton of productvy and competon relatonshp, along wh other two TFP growth seres, we also use TFP growth that s estmated by takng out the effects of non-constant returns and mperfect competon from all ndustres separately.

5 5 Ths paper uses data based on North Amercan Industral Classfcaton Systems (NAICS) for 86 Canadan manufacturng ndustres from 1990 to Wh ths data we have three man objectves. As a frst step, we estmate the markup (whether prce s hgher than margnal cost) and scale effects (whether ndustry operates under ncreasng, constant or decreasng returns) for each ndustry. To our knowledge, ths s the frst tme such a study s done n such a dsaggregate level of ndustres. 4 Hence, the endeavor of explcly measurng markup and scale effects at ths level of detal s a novelty of the paper. To our surprse, we fnd that out of 86 ndustres we cannot reject the null hypothess of constant returns to scale to 80 ndustres and that of perfect competon to 55 ndustres. Ths fndng ndcates that the Canadan manufacturng sector s overwhelmngly characterzed by constant returns to scale, and one thrd of manufacturng ndustres operate under mperfect competon, prce hgher than margnal costs. Our second objectve s to estmate TFP growth by solatng the markup and scale effects and compare wh usual calculated TFP growth (based on the assumptons of perfect competon and constant returns) to assess the latter s basness n measurng unobserved productvy growth. Even though for few ndustres, these two seres do not dverge wdely; for some others, they do and even are oppose n ther sgns. At the aggregate, however, appears that the TFP growth measured under the assumptons of perfect competon and constant returns wll be a reasonable ndcator of unobserved productvy growth. The former underestmates the latter only by a range of 0.2 to 0.4 percentage ponts (from 5 to 10 percent of TFP growth) annually. Hence, even though at the aggregate level the calculated TFP provde a reasonable measure of productvy growth, for ndvdual ndustres these two seres may dverge wdely. 4 For U.S. ndustres, Hall (1988) measures the markup and Basu and Fernald (1997) measure the scale coeffcents. Hall (1988) s based on SIC 2-dg 26 ndustres, whereas Basu and Fernald (1997) s based on 34 ndustres. Kee (2004) estmate both markup and scale effects for nne Sngapore s ndustres.

6 6 Therefore, one should be cautous n relyng solely on calculated TFP to study the relatonshp between productvy growth and other economc varables at the ndustry level. Thrd, we estmate the relatonshp between TFP growth and competon, usng three types of TFP growth seres. They are: (1) seres allowng to vary the coeffcents of all nputs but restrctng them to be the same across ndustres (whch we consder as takng out the common effects of non-constant returns and mperfect competon from TFP growth) as done n Dsney et al. (2003), (2) seres based on the assumptons of constant returns and perfect competon (called calculated TFP), and (3) seres allowng to vary coeffcents of nputs by ndustry (whch s takng out the effect of non-constant returns and mperfect competon from each ndustres separately). In measurng competon, we use Herfndahl ndex, concentraton rato at 4, 8, 12, 16, 20 and 50 leadng enterprses, rent share n revenue and markup. We consder them as reverse measures of competon. On the other hand, we also use varables such as frms entry rate, ex rates and mport penetraton as drect measure of competon. Our fndng s that the competon, as measured by lower level of rents and markup and hgher level of mport penetraton, s assocated wh a sgnfcantly hgher rate of TFP growth. Smlarly, the fall n the Herfndahl ndex and the concentraton rato (proxes for a rse n competon) lead to hgher productvy level effect. Ths result s more or less robust whether we use TFP growth computed under the assumptons of constant returns to scale and perfect competon or estmated by relaxng them. The hgher level of entry rate may also lead to hgher rate of productvy growth, however, the ex rate does not have any mpact on productvy growth. Furthermore, an ncrease n mport penetraton wll lower productvy level. The rest of the paper s organzed as follows. In the next secton, we develop a theoretcal framework for testng markup and returns to scale. In Secton 3, we brefly dscuss the data used

7 7 n the paper. In Secton 4, we test the valdy of perfect competon and constant returns n Canadan manufacturng, usng the model developed n Secton 2. In Secton 5, based on the estmates n Secton 4, we estmate the TFP growth whch are free of the effect of non-constant returns and mperfect competon. In ths secton, we also compare these TFP measures wh the usual TFP growth that would arse under the assumptons of constant returns and perfect competon. In Secton 6, we develop an equaton for estmatng productvy-competon relaton and presents results of ths relatonshp n Secton 7. And lastly n Secton 8, we conclude the paper. 2. Framework for Testng Market Structure and Returns to Scale In ths secton, we develop an equaton to test the market structure and returns to scale, whch can also be reduced to a framework that could be used to calculate standard Solow resdual. We begn wh ndustry level producton functon where ndustry s gross output Q s produced usng producton workers ( E U ), non-producton workers ( S ), fuel and energy ( ), other nonenergy materals ( producton functon s gven by 5 M ), and capal stock ( K ). Hence for each ndustry at tme perod t, the ( (1) Q = A F U, S, E, M, K ), = 1,, 86; t = 1,, 12. Note that the productvy s represented by a Hcks Neutral augmentaton A, of aggregate output. 6 F s assumed to be homogenous of degree γ. Takng logarhm and then totally dfferentatng equaton (1) wh respect to tme yelds: 5 There s lerature treatng producton workers as unsklled and non-producton workers as sklled. In the same tradon, we denote producton workers by U (ndex for unsklled) and non-producton workers by S (ndex for sklled). 6 Although the Hcksan A measures the shft n producton functon at gven levels of nputs, s almost always dentfed wh techncal change. But ths s not generally an approprate nterpretaton. The dfference between the

8 8 ln Q = ln A F + M M F F + U U ln M F lnu F + K K F F + S ln K S ln S F. F + E E F ln E Notng that F J J F Q = J J Q, and denotng by α J for J = U, S, E, M and K, we have (2) ln Q = ln A + α U lnu + α S ln S + α E ln E + α M ln M + α K ln K, where α J s the average of two perods that we consder n estmatng growth n the output and nputs n the h ndustry. Note that the sum of elastcy of output wh respect to all nputs measures the elastcy of scale. Gven the assumpton that F s homogenous of degree γ, accordng to Euler s theorem, α U + α S + α E + α M + α K = γ. If γ s larger than 1, we have ncreasng returns to scale, equal to 1 yelds constant returns to scale and smaller than 1 mples decreasng returns to scale. Unless we are ready to make assumptons on market structure, the value of αs, the elastcy of output wh respect to nputs, are not observable from the data. Let us denote the log dfference of the varable by correspondng lower case letter, such as q = ln, and ncorporate γ nto Equaton (2) such that: Q q = a + u + s + e (3) α U α S α E α M ( γ α U α S α E α M ) + m + k. Hcksan shft parameter, A, and the rate of techncal change arses for many reasons. The most mportant s that the shft parameter captures only costless mprovements n the way an economy s nputs are transformed nto real GDP. For example, techncal change that results from R&D spendng wll not be captured by A unless R&D s excluded from nputs whch generally s not (Hulten, 2002, p. 9). Another ssue s mpled nature of techncal change. In general, the Hcksan formulaton of the producton functon s vald f nnovaton mproves the margnal productvy of all nputs equally. However, f (costless) mprovements n technology augment the margnal productvy of each nput separately (Hcks based technologcal change), then the productvy growth depends on the nput shares as well as the parameters of nnovaton. A change n ncome share can cause output per un nput (TFP) to ncrease, even f the underlyng rate of techncal change remans unchanged. Ths enforces the basc pont that productvy growth s not the same thng as techncal change (Hulten, 2002, p. 13).

9 9 If the frms n ndustry are prce takers n the nput markets, condons for cost mnmzaton yelds that Q J = p λ J, where λ s the Lagrange multpler, whch by Envelop theorem s the margnal cost, and p J s the prce of nput J. 7 Throughout the paper, we mantan that frms face perfectly competve factor markets. Hence, α J takes the followng form: (4) p J J α J = J = U, S, E, M, K, c Q where c s the margnal cost. Multply and dvde the rght-hand sde expresson of Equaton (4) by, the prce of output, to obtan p Q pj J p Q (5) α J = = µ rj J = U, S, E, M, K, p Q c Q where µ s the markup rato, µ = c, and r s the share of nput J n total revenue, Q, and p Q J p Q s observable from the data. Replacng nformaton n Equaton (5) nto Equaton (3) and rearrangng we have: (6) q = a + µ [ ru ( u k ) + rs ( s k ) + re ( e k ) + rm ( m k )] + γ k We wll be usng Equaton (6) to test market structure and returns to scale. However before proceedng further, let us see how ths expresson s dfferent from the regular expresson that s used n TFP growth accountng lerature, whch s based on the assumptons of constant returns and perfect competon. In that lerature, snce µ and γ are assumed to be one, Equaton (6) can be rearranged such that the TFP growth, whch s equal to unobserved productvy growth, a, s defned as. 7 Instead, f the frms are monopsonsts, the frst order condon of cost mnmzaton leads to the suaton that p J Q 1 = 1 +, where ξ s the prce elastcy of factor supply. J ξ λ

10 10 ˆ, (7) TFP = q ru ( u k ) rs ( s k ) re ( e k ) rm ( m k ) k where TF ˆ P s substuted for a. 8 Next, to see what happens when neher µ nor γ s equal to one, let us replace q n Equaton (7) by expresson n Equaton (6), such that Equaton (7) becomes (8) TFˆ P = a + ( µ 1) [ r U ( u k ) + r S ( s k ) + r E ( e k ) + r M ( m k )] + ( γ 1) k Ths equaton s smlar to that of Hall (1988) except that Hall does not consder the scale parameter. Klette (1999) have estmated somewhat smlar equaton for manufacturng frm level data for Norway from 1980 to Recently, Kee (2004) uses smlar equaton to compare prmal and dual productvy growth n Sngapore s manufacturng sector. Lookng at Equaton (8), we see that the TFP growth rate and unobserved actual productvy growth may dffer due to eher mperfect competon ( µ 1) or non-constant returns to scale ( γ 1), or any combnaton of these two. The usual TFP measure may bas the unobserved productvy growth n eher drecton, dependng on types of market mperfecton, nature of returns to scale and growth rates of factors of producton. By lookng at Equaton (8), one could fnd the drecton of basness easly. To calculate TFP growth under the assumptons of perfect competon and constant returns, we can use Equaton (7) or Equaton (3). To estmate TFP growth relaxng these two assumptons, we can use Equaton (6). Whout loss of generaly, we assume that the Hcks neutral technologcal progress s a random varable such that the growth rate of ndustry n perod t conssts of an ndustry-specfc growth rate, η, and a perod specfc growth rate, θ t, whch captures the macroeconomc shock that s common across ndustres n the same perod, plus a whe nose, u. Hence, for each ndustry at tme t, Equaton (6) takes the followng form:. 8 Solow (1957) recommended evaluatng the rght-hand sde n order to measure the Hcks-neutral techncal progress. Ths measure has come to be known as TFP growth because accounts for output and all nputs.

11 11 (9) q η + θ t + β 1 [ ru ( u k ) + rs ( s k ) + re ( e k ) + rm ( m k )] + β k + u = 2 Lookng at Equatons (6) and (9), s clear that the estmated value of β 1 wll be ndustry specfc markups, and the estmated value of β 2 wll be ndustry specfc returns to scale coeffcents. The estmaton of TFP growth usng Equaton (9) serves our four purposes. Frst, ths wll enable us to fnd out how many of these ndustres have market power and how many of them operate under non-constant returns to scale. Second, by comparng the TFP growth estmated based on ths equaton wh TFP calculaton under perfect competon and constant return to scale, we can assess the magnude of basness of Solow resdual (the latter) to unobserved productvy growth (the former). Thrd, we could use the estmated ndustry markup, β 1, as a measure of nverse competon n the next secton. Fourth, the TFP seres obtaned from (9) can be used for the robust check n the relaton between TFP growth and competon. The estmaton results wll be provded n Secton 4 after a dscusson on the data n the followng secton.. 3. Data Ths paper studes the NAICS 4-dg level manufacturng ndustres from 1990 to 2002, and there are 86 ndustres at ths level. The data used for the estmaton of Equaton (9), except capal stock, come from the Annual Survey of Manufactures (ASM), a sample survey of manufacturng establshments carred out by Statstcs Canada. Ths survey, whch covers manufacturng ndustres at a very detal level, has data on shpments, number of producton workers and hours, number of admnstratve employees, wages to producton workers, salary to admnstratve employees, cost of fuel and electrcy, expendure on materals (non-energy) and value added. 9 9 In ASM, value added fgure deduct only purchased commodes and products, ncludng energy but do not exclude purchased servces snce nformaton on purchase of servces (except subcontractng) s not collected as a regular part

12 12 The data from ASM are at current prce. To convert these current prce seres nto constant prce, we use deflators based on Lnk-Level rectangular nput-output tables at current and Paasche constant prces. Snce these tables have only 45 manufacturng ndustres, we have only 45 deflators. For computng these deflators, we aggregated all output produced by each of these 45 ndustres at both current and Paasche constant prce. We dd the same for nput used by these ndustres. Then we computed the rato of current to constant for both output produced and nput used for each of these 45 ndustres. Fnally, we adjusted output seres n ASM by output prce rato and nput seres n ASM by nput prce rato to convert ASM seres nto constant prce. For 24 ASM ndustres, the deflators have one-to-one correspondence. Four of the 45 ndustres fell nto a sngle 86 ASM group of ndustres. The remanng 17 ndustres prce deflators were used for the remanng 61 ASM ndustres. Some of these 17 ndustres were at NAICS 3-dg and some of them were at NAICS 2-dg. For 57 ndustres, prce deflator at 3- dg was used and for the remanng 5 ndustres, prce deflator at 2-dg level was used. Snce the nput-output tables for year 2002 were not avalable, we have used the 2001 deflators for 2002 as well. The capal stock data are obtaned from the Investment and Capal Stock Dvson of Statstcs Canada. Among dfferent seres of capal stock based on dfferent deprecaton methods, we use the geometrc nfne end-year net stock data seres n 1997 prce. Even though would be preferable to use data on capal nput rather than capal stock, we use the latter seres, as the former s not avalable. By usng capal stock data we are assumng that capal nput s proportonal to the measured capal stock, so capal stock growth rate s used to measure capal nput growth. Hence, there s no adjustment for capacy utlzaton. of the annual census. Value added also ncludes any ndrect taxes (such as property taxes). Therefore, the value added s overestmated.

13 13 In the second part of the paper, besdes ASM and capal stock data, we use other data to measure competon. Among them, Canada s mports from the world and exports to the world are taken from Industry Canada s webse. However, snce the trade data based on NAICS do not go pror to 1992 and those based on SIC do not extend beyond 2000, for years 1990 and 1991 we use the data on SIC 4-dg ndustres and convert them nto NAICS 4-dg (86) ndustres usng Statstcs Canada s converson table from SIC to NAICS. 10 The other data used n the paper such as Herfndahl ndex, frms entry rate, frms ex rate and shpment concentraton rato at leadng 4, 8, 12, 16, 20 and 50 enterprses come from Statstcs Canada. 4. All about Markups and Returns to Scale To estmate Equaton (9), for q, we use the ASM data seres called total value of shpments and other revenue at constant prce; for u, we use total number of producton workers; for s, we use total number of non-producton workers; for e, we use cost of fuel and electrcy at constant prce; for m, we use total costs of materals and supples at constant prce. For k, we use capal stock data at constant prce. 11 For the revenue share parameters, we use rato of each nput cost to total value of shpment and revenue at current prce. Snce we are computng TFP growth not 10 In case when one SIC ndustry fell nto more than one NAICS 4-dg ndustres, we dstrbuted the trade data n that SIC ndustry to all NAICS ndustres that the partcular SIC ndustres s spread out usng the shares of NAICS ndustres that constute the total output of that gven SIC ndustres as gven n the Statstcs Canada s concordance table. 11 The measure of ndustry output (total value of shpments and other revenue) s actually a sales measure, not a producton measure. Ths seres conssts of shpment of goods of own manufacturer and non-manufacturng shpments and revenue. Producton workers are those who works n producton n manufacturng actvy only. Admnstratve employees nclude all employees not ncluded n producton and related workers. Thus the producton and related workers engaged n non-manufacturng actvy are ncluded n admnstratve employment component. The seres total cost of materal and supples and cost of fuel and electrcy nclude cost of both manufacturng and non-manufacturng sectors.

14 14 level, we take the log dfference between two consecutve years for the nput use and average of two years for share parameters. A summary of the aggregate annual growth of all nputs and output along wh nputs shares n revenue s presented n Table 1. Durng , the gross output, value added and materal nput grew almost at the same rate of about 4.5 percent annually. Over ths perod, the fastest growng nputs have been materal use followed by fuel and energy whch grew by 4 percent annually n real terms. Annual capal growth rate was hgher than growth rates of both types of labor mplyng ncreased capal ntensy over tme. The slowest growng nput n manufacturng over ths perod was non-producton worker (dubbed as skll labor n the lerature), whose annual growth was less than 1 percent, and the salary bll of these employees occuped less than 6 percent of total revenue. Intermedate nput constutes more than half of gross output (52 percent), and the remanng 48 percent s shared almost equally by labor and capal. Table 1. Average Annual Growth Rates and Shares of Revenue (n Percent), Q Y U S E M K ( Y L) ln Share Note: All numbers are average annual percentage growth rates estmated by takng the log dfference and multplyng them by 100. The numbers are computed usng nformaton for the entre perod rather than the perod end years. Wh 13 years of data, we have 12 growth rates. Q s the real value of shpment; Y s the real value added; U s total number of producton workers; S s total number of non-producton workers; E s cost of fuel and energy, and M s the total cost of materal nputs. The sum of U and S s L, total number of employees. Labor productvy s gven by ( Y L). We have used employment share of each ndustry as weght for all other varables. For estmaton of Equaton (9), we pool all 86 ndustres for 12 years and use pool least square and feasble generalzed least square (weghted pool least square correctng for crosssecton heteroskedastcy). 12 We allow the coeffcents for each ndustry to vary along wh 12 Wh 86 2 coeffcents, 11 years of dummy and 86 ndustry fxed effects, there wll be a total of 253 parameters to estmate.

15 15 ndustry fxed effects and tme dummes. For 86 ndustres, the estmated coeffcents and ther t- test under the null hypothess that the coeffcents are equal to uny usng feasble generalzed least square method are reported n columns 1 through 4 n Table 5. Results show that for 73 out of 86 ndustres, the estmated markup coeffcent s greater than one ndcatng the possbly of prce hgher than margnal cost (Column 1). The magnude of these coeffcents ranges from as low as 0.56 for fru and vegetable preservng (NAICS-3114) ndustry to 3.5 for aerospace product and parts NAICS-3364) ndustry. On the other hand, 44 ndustres have scale coeffcents greater than one (Column 3). The magnude of ths coeffcent ranges from 0.12 for navgatonal and medcal nstruments (NAICS-3345) ndustry to 3.85 for pharmaceutcal and medcne (NAICS- 3254) ndustry. 13 However, n terms of statstcal sgnfcance, the results are somewhat surprsng. Out of 73 ndustres for whch the markup coeffcents are greater than one, only for 30 the coeffcents are sgnfcantly greater than one. Among them, the t-test s sgnfcant at 1 percent level for 10 ndustres, at 5 percent level for another 13 ndustres, and at 10 percent level for addonal 7 ndustres (Column 2). Thus for 30 ndustres, the null hypothess of perfect competon can be rejected. Among the 13 ndustres whose markup pont estmates are less than one, only one ndustry has markup sgnfcantly less than one. That ndustry s beverage (NAICS-3121), and s markup coeffcent s 0.57, mplyng that prce s almost half of s margnal cost. Among the 44 ndustres whose scale coeffcents are greater than one, s sgnfcantly greater than one for only one, the textle and fabrc fnshng (NAICS-3133) ndustry (Column 4). And among the remanng 42 ndustres for whch the scale estmates are less than one, they are sgnfcantly less than one only for three ndustres. These three ndustres are: (1) other textle and 13 The scale coeffcents are negatve for basc chemcal manufacturng (NAICS-3251) and aerospace product and part. For the prevous ndustry s 1.5; for the latter, s 7.8. We consder these two ndustres as outlers.

16 16 product mlls (NAICS-3149), (2) clothng accessores and others (NAICS-3151), and (3) navgatonal and medcal nstrument (NAICS-3345). To sum up, for 30 ndustres the null hypothess of perfect competon can be rejected n favor of prcng hgher than margnal cost and for one ndustry could be rejected n favor of prce lower than margnal cost. For the remanng 55 ndustres, the null hypothess of perfect competon cannot be rejected. In terms of scale effect, only for one ndustry, the null hypothess of constant returns can be rejected n favor of ncreasng returns to scale; for another three ndustres, the null hypotheses of constant returns can be rejected n favor of decreasng returns. Wh the excepton of two outler ndustres, for the remanng 80 ndustres the null hypothess of constant returns cannot be rejected. Hence, could be concluded that the Canadan manufacturng sector s overwhelmngly characterzed by constant returns to scale, and one thrd of Canadan manufacturng ndustres enjoy market power (prce greater than margnal cost) TFP Growth: Estmated vs. Calculated As mentoned n Secton 2, we assume that TFP growth can be gven by the estmates of ndustry fxed effects ( η ), tme dummes ( θ t ) and resduals ( ) u. Hence, based on Equaton (9) TFP growth for ndustry n year t wll be gven by the estmated value η + θ f resduals are not ncluded and by η θ + u f they are ncluded as part of technologcal change. Usng ths + t method, we have TFP growth measures for 86 ndustres for 12 years. The average (across years) TFP growth for all ndustres (based on generalzed weghted pool and pool regressons) s gven t 14 Usng markup and scale parameters, we can estmate the mpled share of rents by ndustres. Cost mnmzaton condon mples that γ (gven by β 2 n Equaton 9) s equal to the rato of average to margnal costs, and can be expressed as γ σ c = µ ( 1 r π ), where σ s the average cost; µ s gven by β1 n Equaton (9); and rπ s the share of economc profs n total revenue. Rearrangng we have r = 1 γ µ. Our calculaton based on ths π ( )

17 17 n Columns 5 and 6 of Table 4. Note that snce 12 u t= 1 = 0, the average TFP growth across tme wll be the same whether resdual s taken as part of TFP growth or not. Based on generalzed least square regresson, the annual TFP growth s posve for 70 ndustres; based on pool regresson, s posve for 65 ndustres. The annual growth rates of most of the ndustres range between 0 to 2 percent. Next we would lke to compare these estmated TFP growth seres wh the one that s calculated under assumptons of constant returns and perfect competon usng Equaton (7), whose varables on the rght-hand sde are observable, except capal share n total revenue. Based on the assumpton of constant returns to scale, we take capal share as resdual. The calculated TFP growth s presented n Column 7 n Table 4, whch s posve for 74 ndustres and negatve for the remanng 12 ndustres. The annual growth rate vared from a negatve of 3.2 n communcaton equpment (NAICS-3342) to a posve of 5.4 percent n aerospace product and parts (NAICS-3364). For most of the ndustres the annual growth rate s n the range of 1 to 2 percent. Comparng the estmated and computed TFP growth seres (Columns 5, 6 vs. Column 7), we fnd that for few ndustres the dfference between the two seres s small, whereas for majory of ndustres, s rather hgh; n some cases even the sgns are oppose. We have plotted the average estmated (wh pool regresson) and calculated TFP growth for 84 ndustres n Fgure 1. The wde dfference between two TFP growth seres for some ndustres make us cautous n usng only a sngle seres f the objectve s to relate TFP growth and other economc varable. expresson shows that out of 86 ndustres, 62 ndustres had average posve profs durng 1990 to 2002, whereas for the remanng 24 ndustres, the average prof was negatve.

18 18 Fgure 1. TFP Growth Accountng vs. Estmated Productvy Growth 6 4 Calculated TFP Growth Estmated TFP Growth Industry Note: we have plotted ths table for 84 ndustres. Two ndustres, pharmaceutcal and medcne manufacturng (NAICS-3254) and aerospace product and parts manufacturng (NAICS-3264) are taken out as outlers as ther annual TFP growth s respectvely unusually low (n negatve) and hgh. The estmated TFP growth s based on pooled regresson. Interestngly however, at the aggregate level, the calculated TFP growth s not very dfferent from the estmated one, whch we show n Table 2. To arrve nto ths table, frst we obtaned Domar-weghted TFP growth of each seres for each year and then took the average of ths value across 12 years. In other words, we obtaned annual TFP growth usng a t = ( v ). a, where 86 = 1 v = 1 p 2 p Q yt Q Y t + p Q 1 p yt 1 Q Y 1 t 1 s the Domar weght, p Q Q s the current dollar gross output n ndustry, and p yt Y t s current dollar aggregate manufacturng value added. Snce n ths case, the average growth depends on whether we use resduals as TFP growth or not, we have reported both seres under two columns, whout resduals and wh resduals Contrary to the ndustry level, the average of aggregate TFP dffer whether we use resdual as part of TFP or not. It happens because + θ η + θ + u v η + θ v η + θ + u. η for each year. It means that ( ) ( ) t t t t

19 19 Over the duraton of 13 years, the computed TFP grew by 3.48 percent per annum. Ths rate s somewhat hgher than reported changes n offcal statstcs. For example, for manufacturng, the TFP growth durng was 4 percent annually and for , was 1.8 percent. 16 Even though our measure of TFP growth should not be taken at face value and used as substute for Statstcs Canada s measure, as a relatve measures across tme and ndustres they reman very useful. Indeed, the estmaton of TFP growth along wh markups and scale effects at ths level of ndustry detal s a novel aspect of ths paper. Table 2. Calculated vs. Estmated Average Annual TFP Growth Rates, Whout Resduals Wh Resduals Calculated TFP Estmated TFP wh pool regresson Estmated TFP wh weghted pool regresson Average growth s obtaned by weghtng 86 ndustres TFP growth estmates by Domar weghts. Based on our results, f one adopts the TFP growth whout ncludng resduals, then the estmated TFP growth s hgher than computed TFP by about 0.2 percentage ponts n pool regresson and by 0.3 percentage ponts n generalzed pool regresson. On the other hand, f one takes the TFP growth that also ncludes resduals, then the estmated TFP growth s hgher than computed one by a range of 0.3 to 0.4 percentage ponts. In any case, the range of dfference between these two seres s 0.2 to 0.4 percentage ponts. That s, the estmated TFP growth s only 5 to 10 percent hgher than the calculated annual TFP growth of 3.48 percent per annum. At the aggregate, the calculated TFP s lower than the estmated TFP, no matter whchever method we adopt. Hence, the usual measure of TFP underestmates actual unobserved TFP 16 In ths respect, a cautonary note s n order sayng that a TFP number for a partcular ndustry n a partcular year from our seres may not be precse, as data lmaton precludes us from accountng several factors such as qualy dfferences of nputs, capacy utlzaton, use of more detal deflators and most mportantly, the subtracton of servces nput. Snce the ASM data does not nclude cost of servces nput except the cost nvolved n servces subcontractng, our TFP measure wll be larger than otherwse they would have been. Furthermore, we use shpment data whout adjustng for nventory change for output.

20 20 growth, but does so by a very small margn. In terms of Equaton (8), ths means that TFP < a. Hence, wh roughly γ = 0 and µ 0, ths relaton more or less mples that the expresson nsde the square bracket n Equaton (8) should have, on average, been negatve for Canadan manufacturng ndustres durng Measurng Competon Now we have two seres of TFP growth, one computed and the other estmated allowng each ndustry s coeffcent to vary. The next task s to obtan measurements of competon so that we can fnally estmate the relatonshp between productvy and competon. Snce there s no sngle ndcator for competon, we use a few of them. As n Nckell (1996) and Dsney at al. (2003), we categorze the competon varables nto two groups, those that affect the level of productvy and those that affect the productvy growth. We estmate the productvy and competon relatonshp n the followng form: (11) ln Q = α lnu 1 + β Z α ln S β Z 2 + α ln E 3 t + α ln M t + ψ + φ + ρ t + e 4, + α ln K 5 where, the frst fve varables on the rght-hand sdes are factors of producton. We are mplyng that determnes the level of TFP, and Z determnes the growth of TFP. All measures of Z1 2 competon that affect the level of TFP wll enter under Z 1, and those that affect the growth of TFP wll enter under Z 2. Unobserved factors that nfluence the level of productvy are covered by ndustry fxed effects ( ψ ) and tme effects ( φ ) t t, and those that affect TFP growth are covered by ρ. The problem of omted varables wll arse only f the unobserved shocks are not captured by these three varables.

21 21 Snce we are nterested n estmatng the TFP growth equaton (whch wll also take out the ndustry fxed effect), the productvy-competon relaton n (11) becomes (12) ln Q = α 1 lnu + 2 ln S + 3 ln E + 4 ln M + 5 ln + β Z 1 1 α + β Z ρ + φ + e α t α α K Note that those varables that affect the level of TFP wll appear wh frst dfference and those that affect the growth of TFP wll appear n level forms. If a certan varable could affect both the level and the growth of TFP, wll appear n both forms. To measure competon, we use Herfndahl ndex ( h ), concentraton ndex for leadng 4, 1 enterprses ( C ), leadng 5 th to 8 th 2 enterprses ( C ), leadng 9 th to 12 th 3 enterprses ( C ) 13 th to 16 th 4 enterprses ( C ), leadng 17 th to 20 th 5 enterprses ( C ) 6 enterprses ( C ), and leadng 21 st to 50 th, leadng. The sum of all these sx categores of concentraton rato s the concentraton rato for the 50 bggest enterprses. We also use frms entry rate ( ) mport penetraton ( Ω ), estmated markups ( ) µ and rents ( ) κ, frms ex rate ( ), x r π to measure competon. In a way, Herfndahl ndex and concentraton rato are measures of market power. In Appendx A, we have shown that wh homogenous products, and Cournot conjecture about the output change by other frms n response to s decson to change output, and Cobb-Douglas preferences, Herfndahl ndex s equal to prce-margnal cost margn, the Learner s ndex. Hence, s generally beleved that hgher Herfndahl ndex and concentraton rato means lower competon. However, there are a number of problems assocated wh the use of Herfndahl ndex and concentraton rato as a measure of market power (an nverse measure of competon). The most notably s that the share of output n an ndustry does not really capture the true market sze. Furthermore, the same level of Herfndahl ndex and concentraton rato mght mean dfferent levels of market power n dfferent ndustres.

22 22 In ths context, these measures are unlkely to be a cross secton measure of the nverse competon. However, changes n these measures are lkely to be a good tme seres measure of nverse competon. Hence, we wll regard the change n the concentraton rato and the Herfndahl ndex as a measure of change n market power. In Table 3, we see that there s que a varaton of Herfndahl ndex and concentraton rato across ndustres. For example, the average Herfndahl ndex vared from as low as n an ndustry to as hgh as The smlar varaton s found n all groups of four-enterprse concentraton ratos. At four leadng enterprses levels, there are ndustres rangng from as hgh as 97.5 percent output share to as low as less than 10 percent. Table 3. Average Herfndahl Index and Concentraton Rato, Herfndahl Concentraton Rato at Leadng Enterprses Index Maxmum Mnmum Average Note: the average of Herfndahl ndex s the smple average of 86 ndustres over 12 years. The other varables that we use as control are frms entry and ex rates. Here, we wll be lookng at cross secton relaton of entry and ex rates wh productvy growth. As presented n Table 4, the data shows that there s que a varaton of these rates across ndustres. The average ex rate durng vared from a mnmum of 3.6 percent to a maxmum of 18 percent. Smlarly, the entry rate vared from a mnmum of 5.2 percent to a maxmum of 21.8 percent. Table 4. Frms Ex Rate and Entry Rate Ex Rate Entry Rate Maxmum Mnmum Average Maxmum Mnmum Average Average ( ) Entry rate s defned as the percent of new establshments n total establshments, and ex rate s defned as the share of enterprses that exed n a gven year n total number of establshments n that year.

23 23 Another varable that we would want to use as an nverse measure of competon s the rent or the prce-average cost margn. Rent s defned as a share of total revenue, where rent s revenue less materal, labor and capal costs, as gven by the rght-hand sde expresson below: (13) r p σ p Q p J Q Q J π =, pq pqq where σ s the average cost; Q s total revenue and J s total payment to factors of p Q producton. We have data for ths expresson except for the cost of capal, whch we compute, as derved n the Appendx B, usng the followng expresson. (14) Cost of capal = ( r + δ ) p K, p J where p K s the prce of capal (value of capal stock), r s the real nterest rate and δ s the deprecaton rate. For prce of capal, we use total stock of capal at real prce; we assume r to be equal to 4 percent for all ndustres throughout the perod and compute the deprecaton rate for each ndustry and year usng capal stock and nvestment data usng formula gven n (B2) n Appendx B. In the lerature, rent has been used both as factor affectng productvy growth and productvy level. However, snce rents are expressed n terms of revenue, we feel that s more sensble to use the level rather than change as a regressor. For competon and lack of, what matters s the level of rents (normalzed by revenue) not the change n rents over tme. We expect the sgn of level of rents to be negatve. The another varable that we use as an nverse measure of competon s markup. We expand the markup for 12 years usng Equaton (9) as follows: (15) ˆ β 1 = [ r ( u k ) + r ( s k ) + r ( m k ) + r ( m k )] U S q ˆ η ˆ θ ˆ β k t E 2 M

24 24 where a hat above the varable means estmated value from the regresson. Note that βˆ 2 has no tme dmenson; s a vector of scale coeffcents from Equaton (9). By multplyng βˆ 2 by capal stock for 12 years, we obtan the seres of ˆβ 1 (markup) for 86 ndustres over 12 years. So the tme-varyng markup s derved under the assumptons that the scale coeffcents s the same across all 12 years. To control for foregn competon, we use mport penetraton rato whch s defned as the share of mports to total consumpton, where consumpton s obtaned as shpments plus mports mnus exports. We consder mport penetraton as a factor affectng both the level and growth of productvy. We have completed the dscusson on the varables that we wll be usng to measure competon. Now a word on the possbly of endogeney (the reverse causaly problem) of TFP growth wh most of the measures of competon such as concentraton rato, rents and Herfndahl ndex. Industres wh hgher productvy growth may, n the long run, end up wh hgher concentraton ratos and Herfndahl ndces than those wh lower productvy growth. However, note that n ths suaton the relatonshp from TFP growth to concentraton, Herfndahl ndex and rents should be posve. For example, f there are more productve frms n one ndustry than n others, then these productve frms wll strengthen ther market share and as a result (ceters parbus) the concentraton ratos and Herfndahl ndces would be hgher n the ndustry wh more productve frms compared to others, thereby mplyng a posve relatonshp from TFP growth to nverse measure of competon. Thus, as argued n Nckell (1996), f we fnd a negatve relatonshp between these nverse measures of competon and productvy growth, we mght argue that the true relatonshp s even stronger. In any case, to elmnate reverse causaly we lag

25 25 most of the explanatory varables by one year as n Nckell (1996) and Dsney at al. (2003), where they lag these varables by 2 years. Incorporatng all these varables nto Equaton (14), our estmatng equaton s: (16) ln Q = α + 6 j= 1 + β r 1 ln + α 2 ln + α 3 ln + α 4 ln + α 5 ln β C j 11 U j 1 + β Ω + + β h π S + β κ + β E + β Ω 13 x + ρ + φ + e M t K Ths equaton s smlar to Equaton (9), except the competon varables were not present n that equaton. Also, here we do not allow the coeffcents to vary by ndustry. In concentraton rato, we have used sx groups of enterprses whose sum provdes the concentraton rato for the 50 th largest enterprses. 7. Results and Interpretaton The man results are presented n Tables 6, 7 and 8. We have a sequence of estmated equatons to nvestgate the robustness of the results, wh regard to both changes n specfcaton of the ndependent and dependent varables. In Table 6, we start wh output growth as a dependent varable. The frst specfcaton s a basc model where we have all explanatory varables, except the Herfndahl ndex, entry rate and ex rate. We use rent as defned n Equaton (13), called rent 1 and s denoted by ( 1) r n the table. In specfcaton (2), we replace the sx π groups of concentraton rato by one sngle varable whch n effect s the concentraton rato of leadng 50 frms. In specfcaton (3), we use alternatve measure of rents defned as total revenue mnus cost of all other nputs, except capal, and normalzed by total revenue. In other words, the numerator ncludes all revenue that s left after payng for labor, fuel, energy and materal nputs. We called rent 2 and s denoted by ( 2) r n the table. Specfcaton (4) adopts Specfcaton (3), π

26 26 but replaces the sx groups of concentraton by one aggregate seres. Specfcaton (5) uses the markup estmated usng Equaton (9) as a proxy for rents. Specfcaton (6) s smlar to Specfcatons (1) and (2) but replaces the change n concentraton varable by the change n the Herfndahl ndex. In Specfcaton (7) we use entry and ex rates as measure of competon, and consder hgher level of both rates as hgher levels of competon. In ths specfcaton, we take out all other domestc measures of competon ncludng rent, as s consdered a result of entry and ex rate. 17 As expected, the rent varable n Specfcaton (1) s hghly sgnfcant, ndcatng that an ndustry wh hgh level of prof to revenue share n tme (t 1) wll have low productvy growth n tme t. The coeffcents of all sx groups of enterprses are negatve but only those for the frst and fourth four largest enterprses are statstcally sgnfcant. Ths means that fall n concentraton rato one year ago (whch we nterpret as rse n competon) wll rase productvy level ths year. Ths effect s sgnfcant for two groups of enterprses (the frst and the fourth groups). In Specfcaton (2), the coeffcent of 50 leadng enterprses concentraton rato s negatvely sgnfcant, renforcng that the relatonshp between change n market share and TFP level s negatve. All the other results of Specfcaton (1) contnue to hold n Specfcaton (2). By replacng rent 1 by rent 2 (comparng Specfcatons 1 vs. 3 and 2 vs. 4), almost all results reman the same. In Specfcaton 5, the markup s negatvely sgnfcant and shows that an ndustry wh low level of markup (hgh competon) a year ago wll have hgher productvy growth ths year. In ths test, the change n concentraton rato frst tme turns nsgnfcant. When we estmated ths specfcaton for all sx groups of concentraton ratos, the result was that the 17 Note that n the long run prof mght be zero, but stll there could be dfference between prce and margnal costs. Testng whether long run profs are posve s a test of free entry, not of (perfect) competon. Free entry guarantees that long-run profs equal zero, but not that prce equals margnal cost. Frms n a monopolstc competve ndustry

27 27 coeffcents were negatve for three and posve for three, but none of them were sgnfcant (not reported). The Herfndahl ndex, a measure of nverse competon, s negatvely sgnfcant (Specfcaton 6), confrmng the posve relatonshp between competon and productvy level. In Specfcaton (7), we see the entry rate affectng the TFP growth posvely, mplyng an ndustry wh hgh level of entry rate n year (t-1) wll have hgh productvy growth n year t. Ex rate s not sgnfcant. In neher of the specfcatons, the foregn competon varable the mport penetraton rate (level or change) s found sgnfcant whch means that Canadan manufacturng ndustres do not have an mpact of mports on ther productvy level and growth. The TFP growth seres we have used n Table 6 were obtaned whout makng any assumpton on market structure and returns to scale, but wh the assumpton that the coeffcents of all ndustres n Equaton (16) are the same. In that sense, ths estmaton takes out only the common (across ndustres) effect of potental markup and scale effects. In Table 7 (specfcatons 1 through 4), we present results by replacng ths TFP seres by calculated TFP under the assumptons of constant returns and perfect competon. And n Specfcatons 5 and 6, we use the TFP growth seres estmated n the prevous secton as a dependent varable, n contrast to the seres n Table 6, where we have allowed all ndustres to have dfferent slope coeffcents, and hence take out the markups and scale effects for each ndustry separately. The explanatory varables n Specfcatons 1, 2, 3 and 4 n Table 7 are the same as Specfcatons 2, 4, 5 and 6 respectvely n Table 6. The coeffcents for rents, concentraton ratos and the Herfndahl ndex are all negatvely sgnfcant as n Table 6, ensurng the posve relatonshp between level of competon and productvy growth and change n level of may earn zero prof even though prce s above margnal cost. Short-run profs reveal very ltle about the degree of competon n an ndustry because, n all market structure, short-run profs can be eher posve or negatve.