Credit, Misallocation and TFP: Mexico in the 21 st Century

Transcription

1 Credt, Msallocaton and TFP: Mexco n the 21 st Century Sere de documentos de trabajo Felpe Meza, ITAM Sangeeta Pratap, CUNY Carlos Urruta, ITAM

2 Credt, Msallocaton and TFP: Mexco n the 21st Century Felpe Meza CIE and Dept. of Economcs ITAM Sangeeta Pratap Hunter College & Graduate Center Cty Unversty of New York Carlos Urruta CIE and Dept. of Economcs ITAM September 30, 2014 Abstract We study the relaton between msallocaton of resources, TFP and credt condtons n Mexcan manufacturng at the ndustry level. We develop a tractable theoretcal framework to account for TFP changes n the Mexcan manufacturng sector due to changes n dstortons n the use of captal, labor and ntermedate goods. We show that these dstortons account for a large fracton of aggregate TFP changes n the perod. Reduced form estmatons reveal that changes n dstortons n the data are drven by changes n the avalablty and the cost of credt. We buld a general equlbrum model that maps nancal frctons nto dstortons. The model accounts for a large fracton of observed TFP growth between The contrbuton of nancal factors vares from over 90% of model predcted TFP growth n the expansonary years of to about 15% of the downturn n The recovery of 2010 was largely fuelled by a reducton n sectoral nancal frctons, partcularly a fall n nterest rates and the resultng reallocaton of credt to less dstorted sectors. We are grateful for the nancal support of a grant from the Fundacon de Estudos Fnanceros. We also thank the Insttuto Naconal de Estadstca, Geogra a e Informatca (INEGI), and especally Gerardo Leyva and Natala Volkow for ther help wth the survey on manufacturng establshments. We are ndebted to the Comson Naconal Bancara y de Valores for makng aggregate nformaton on bank loans avalable to us. All data provded to us was anonymzed by the respectve source. Comments from Hugo Hopenhayn, Pete Klenow, Dego Restucca, Mark Wrght, and partcpants at the Socety for Economc Dynamcs meetngs at Toronto and the North Amercan Summer meetngs of the Econometrc Socety at Mnneapols are gratefully acknowledged. Alonso de Gortar and Hoda Nour Khajav provded outstandng research assstance. Any errors are our own. 1

3 1 Introducton The relatonshp between output growth and total factor productvty s well establshed. The earlest calculatons of Solow (1957) attrbuted only 12% of the growth n the Unted States between 1900 and 1949 to the accumulaton of factors of producton, and the remanng 88% to the resdual". Subsequent work enlargng the scope and tme perod of these estmates reduced the sze of the resdual to about 50% (see for example Klenow and Rodrguez Clare 1997), however the prmacy of the contrbuton of TFP to growth s undenable. Movements n TFP have played an mportant role n recent growth mracles n Inda (Bollard et. al 2013) and Chna (Dekle and Vandenbroucke 2010). The reverse has also been true, large drops n output durng sudden stops have been accompaned by a large fall n TFP (Calvo 2006). Some recent studes have hghlghted the role of nput msallocaton as an mportant factor behnd these aggregate TFP changes. Pratap and Urruta (2012) show that nancal frctons can propagate nterest rate shocks nto measured TFP uctuatons by dstortng the use of nputs n the economy. Benjamn and Meza (2009) analyze the real e ects of Korea s 1997 crss and nd that a reallocaton of resources towards low productvty sectors at ths tme generates a fall n TFP. A recent strand of lterature seeks to explan the d erences n TFP levels between countres by emphaszng the role of rm-spec c dstortons,.e. mplct taxes, barrers and constrants whch result n the suboptmal allocaton of resources and lower TFP levels (see, for nstance, Restucca and Rogerson 2008, Hseh and Klenow 2009, Bartelsman et. al 2013). Sandlers and Wrght (2014) use these same nsghts to understand changes n TFP over tme, usng rm level data from Argentna n the perod around the 2001 crss. Chen and Irarrazabal (2013) perform a smlar analyss for the Chlean manufacturng sector after the 1982 debt crss. These papers however do not nvestgate the reasons behnd these dstortons, as ther focus s on quantfyng ther e ects. The goal of ths paper s to understand the role of credt and nancal frctons n accountng for the msallocaton of resources and the changes n TFP over tme. As n Pratap and Urruta (2012), our focus s on nancal frctons dstortng rms decsons to purchase nputs. However, as n the lterature revewed n the prevous paragraph, we analyze the relaton between credt and dstortons at the mcro level, nstead of the aggregate macro level. In ths sense, our work relates to Buera and Moll (2012), who extend the busness cycle accountng methodology to economes wth heterogeneous rms and show that credt shocks and nancal frctons can be mapped nto measured TFP and other aggregate wedges or dstortons. 1 As an emprcal contrbuton, we construct a novel data set by lnkng manufacturng actvty n Mexco wth credt ows at a dsaggregated level. Our data encompasses 82 sectors of actvty for the perod from 2003 to Ths s a partcularly nterestng tme frame to study, snce t ncludes both the perod of rapd growth of , the economc crss of and the subsequent recovery. We use ths data to construct two measures of nput dstortons: () a dstorton to the use of ntermedate goods () a dstorton to the captal to output rato. We construct a smple framework to show how these two dstortons a ect aggregate TFP. 1 Char, Kehoe and McGrattan (2007) ntroduced the methodology of busness cycle accountng as a way of nferrng the aggregate levels of dstortons that an economy faces and how they evolve over tme. The dea s to nfer these dstortons from the wedges obtaned from the optmalty condtons of a standard neoclasscal model. For example, the labor wedge s de ned as the rato between the margnal rate of substtuton of consumpton for lesure and the margnal product of labor, whch should be equal n an undstorted economy. 2

4 Reduced form analyss shows that these dstortons are lnked to d erences n credt avalablty and nterest rates across sectors and over tme. We therefore buld a general equlbrum model that maps nancal frctons nto dstortons to see how changes n nancal frctons a ect aggregate TFP. Frms face a workng captal constrant and a borrowng constrant. The former mples that rms have to nance ther nput purchases usng bank credt or the more expensve trade credt whle the latter constrant dctates the amount of bank credt avalable to rms. Taken together, these constrants dctate that the avalablty and costs of credt a ect rm purchases of nputs. We feed the model a sequence of nterest rate shocks, nancal avalablty shocks (as ndcated by the tghtness of the borrowng constrant) and productvty shocks, all calbrated form the real and nancal data at the sectoral level. Our TFP decomposton suggests that devatons from the optmal use of ntermedate goods and from the optmal captal labor rato account for a substantal amount of TFP varaton n ths perod. These dstortons are strongly lnked to varatons n credt and nterest rates across sectors. Our model whch ncorporates these varatons n nancal frctons can account, n large part for the observed movements n TFP n the perod. In partcular, we can account for over 90% of the TFP growth n The model slghtly overpredcts the fall n TFP n 2009 and the subsequent recovery of Fnancal factors and the reallocaton of credt across sectors play a large n role explanng changes n TFP. The contrbuton of nancal factors ranges from a modest 20% of TFP varaton n , to about 96% of the growth n TFP n the expansonary years of Interest rates and credt avalablty dd not play a large role n explanng the downturn of (about 15% of TFP growth), but the subsequent recovery was largely fuelled by a fall n nterest rates and more mportantly, an mproved allocaton of credt across sectors. A common factor explanng TFP growth throughout the perod s the reallocaton of credt n a manner that reduces dstortons and mproves productvty. The paper s organzed as follows. In Secton 2 we descrbe the data used to analyze the relatonshp between economc actvty and credt and present aggregate statstcs. Secton 3 sets out an analytcal framework to account for aggregate TFP changes through changes n technology, reallocaton of labor and changes n sectoral dstortons. We apply ths methodology to our dataset for Mexcan manufacturng and nd that changes n dstortons play a large role n TFP uctuatons. Secton 4 presents a smple model wth nancal frctons lnkng credt varables and sectoral dstortons and the results of panel regressons suggestng a robust relaton between these varables n our dataset. Fnally, n Secton 5 we calbrate and smulate the full equlbrum model subject to shocks to credt condtons taken from the data and dscuss ts mplcatons for aggregate TFP. 2 Data and Stylzed Facts A major contrbuton of our paper s the constructon of a data set that lnks manufacturng actvty wth bank lendng to rms. In ths secton we descrbe our two man data sources and our procedure to merge them. We also descrbe some stylzed facts for the manufacturng sector durng the perod obtaned from our combned database. In partcular, we show a large ncrease n total factor productvty (TFP) from , a contracton n 2009 related to the world nancal crss, and a small recovery n At the same tme real short term credt to manufacturng ncreased durng the rapd growth perod, especally from 2005 to 2008, and collapsed n 2009 and 2010, wth no observed recovery at the end of the perod. 3

5 2.1 Dataset Constructon We have two man data sources: The rst s the annual ndustral survey (EIA for ts acronym n Spansh) collected by the Mexcan statstcal agency INEGI. The second source s the loan portfolo of all commercal banks, known as the R04C, mantaned by the bankng regulatory authorty, the Comson Naconal Bancara y de Valores (CNBV). Con dentalty restrctons prohbt us from analyzng the data at the establshment or loan level. We therefore work at the lowest level of aggregaton currently feasble, namely at the 4-dgt ndustry level, followng the 2007 North Amercan Industral Class caton System (NAICS). The bankng data are at a monthly frequency, whereas the establshment level data are collected yearly. We therefore aggregate the loan data to the annual level. A further complcaton s that whle the EIA uses the NAICS throughout the sample, the R04C data s class ed accordng to an nternal class caton system for the perod up to July 2009 and to the NAICS thereafter. We construct a exble, probablstc crosswalk between these two systems, detals of whch are gven n Appendx A.1. The EIA data s a representatve sample of nearly 7000 manufacturng establshments. We were able to get data on 86 sectors at the 4-dgt level and 231 subsectors at the 6-dgt level. 2 We use data on gross output and expendture on ntermedate goods to construct measures of value added. The labor nput s measured as the number of people hred drectly and ndrectly by the establshment. The captal nput s constructed usng the perpetual nventory method, usng nformaton on nvestment. Appendx A.2 provdes a more detaled descrpton of the varables used from EIA. The R04C dataset s the unverse of loans by commercal banks to rms. Ths data s outstandng loan balances, collected at the loan level, and s avalable at a monthly frequency. We construct a measure of credt ow by lookng at the debt outstandng on all new loans (.e. loans wth dates of dsbursement n the month n whch the data s collected) n a partcular sector n a partcular month. Ths gves us nformaton on how much credt was dsbursed to each sector n each perod. Our measure of credt ncludes outstandng nterest payments. 3. We focus on data for short term credt, de ned as credt whch matures n a perod of 12 months or less. 4 Ths accounts for, on average, 85% of all bank credt. Fnally, we construct cost of credt measures by lookng at average real nterest rates pad by sector, weghted by the sze of the loan n total credt ow n the correspondng perod and de ated by the change n the manufacturng prce level. All nomnal varables, wth the excepton of ntermedate goods are de ated by the producers prce ndex for manufacturng publshed by the INEGI. Intermedate goods are de ated by an ntermedate goods ndex. 2.2 Aggregate Stylzed Facts: Output and TFP We calculate aggregates by addng up ndustry level varables. We focus on , the perod n whch our two databases overlap. As mentoned earler, ths perod allows us to study two phenomena: hgh growth n output and TFP n the rst ve years and a contracton n the last two perods as the nternatonal nancal crss a ected the Mexcan economy. 2 Aggregates of subsectors contanng 4 or fewer rms were not provded to us n the nterests of con dentalty. 3 We also constructed a measure of credt that excludes nterest payments whch behaves very smlarly to the other measure at the aggregate and sectoral level. 4 As we wll see n the next sectons, we consder nancal frctons whch a ect the amount of workng captal to whch rms have access. The data counterpart of that s short term credt. 4

6 GDP Captal Labor TFP Fgure 1: Real Output, Inputs and Total Factor Productvty (2003=100) Fgure 1 shows real manufacturng GDP, captal and labor nputs from our sample, relatve to the year The former s calculated by addng up real value added n all sectors, whereas the nputs are the stock of captal and personnel employed aggregated over sectors. The gure llustrates the manufacturng boom between 2003 and 2008, when output grew at an annual average of 2.5%. Interestngly labor nput was relatvely stagnant n ths perod, suggestng that the sources of growth lay n productvty. 5 Ths perod of expanson came to an abrupt halt n 2009, when output contracted by almost 11%. Ths fall n output was accompaned by a smaller drop n labor of about 7%. Investment slowed down substantally between 2008 and 2009, leadng to a decceleraton n the growth of captal stock. The recovery n the followng year was modest, as output and labor nputs grew by less than 4%. Fgure 1 also shows the evoluton of aggregate total factor productvty (TFP) over the perod, relatve to Assumng a Cobb Douglas producton functon, aggregate TFP s de ned as A = Y K L 1 where Y represents real value added (manufacturng GDP) and K and L measure aggregate nputs of captal and labor respectvely. The aggregate labor share 1 share of ncome n each sector. 6 s constructed as a weghted average of the labor The weghts are the share of each sector n total output. Fgure 1 shows that TFP ncreased between 2004 and 2008, mrrorng the ncrease n output. TFP fell n 2009 by 7.3% 5 We ver ed ths fact by lookng at employment n manufacturng n a completely d erent database, the Natonal Survey of Occupaton and Employment (ENOE for ts acronym n Spansh). The employment seres from EIA and from ENOE exhbt very smlar behavor. 6 For reasons elaborated n Secton 3, these ncome shares are taken from the correspondng sectors n the U.S. 5

7 Real Credt Credt/Output Fgure 2: Real Credt Flow and Credt Intensty (2003=100) and recovered slghtly n the followng year. Ths gure shows the mportance of the role of total factor productvty n output uctuatons. 2.3 Aggregate Stylzed Facts: Credt Flows and Interest Rates Fgure 2 shows the measure of short term credt ow descrbed earler. The aggregate ow of credt, whch was declnng n the rst few years of the sample, ncreased between 2005 and 2008, and dramatcally n the perod. 7 The fall n output n 2009 was also re ected n a fall n aggregate credt. Interestngly, the recovery n aggregate output was not accompaned by a recovery n credt. One natural queston s whether these changes n credt ows re ect supply or demand factors. Probably both. Stll, some evdence suggests that the supply of credt by banks has played a leadng role. We look at the behavor of credt ntensty, measured as the rato of credt ows to gross output. As also shown n Fgure 2, the expanson n real credt s accompaned by an ncrease n credt ntensty, and the credt contracton n also shows as a fall n ths rato. Therefore, credt dd not just respond to the cyclcal behavor of gross output. We addtonally provde nformaton on the behavor of weghted-average, real nterest rates. Fgure 3 reports the average across sectors, for each year. From a hgh of 4% n 2004 the real nterest rate fell 7 At ths pont t s mportant to note that, snce the nancal crss of 1994, Mexco has a low degree of nancal ntermedaton. The prvate credt to GDP rato was 14% n 2003, at the begnnng of our perod of analyss (see Kehoe and Meza 2011). Ths number s much lower than n smlar Latn Amercan economes (61% n Chle, for nstance). It s possble that, startng from the hstorcal low levels of credt n Mexco, the expanson of credt mght reduce msallocaton, as rms had more access to credt and were possbly able to come closer to e cent levels of factor use. 6

8 percent Fgure 3: Average Real Interest Rate tll 2008 when t was close to zero. The nternatonal nancal crss was re ected n a sharp ncrease n nterest rates to 3%, whch fell by a percentage pont n the next year of subsequent recovery. Notce that the negatve correlaton between the aggregate credt to output rato and the average nterest rate suggests agan an mportant role for supply factors drvng credt avalablty n ths perod. The aggregate pcture, whle nformatve, obscures sectoral allocaton ssues. In the followng secton we set up a smple framework to measure dstortons at the sectoral level from optmal nput use and show how they relate to aggregate TFP. 3 Sectoral Dstortons and Total Factor Productvty We consder a smple, partal equlbrum model of mult-sector producton wth ntermedate goods and sector spec c dstortons. These dstortons ntroduce wedges between factor prces and margnal products. We consder: () a dstorton a ectng the captal to labor rato; and () a dstorton a ectng the rato of ntermedate goods to output. In ths framework, we obtan an expresson relatng changes n aggregate TFP to the level of dstortons and ther growth rate. We perform an accountng exercse usng our dataset and nd a strong role for sectoral dstortons n explanng changes n TFP over tme for Mexcan manufacturng. 3.1 Producton and Frm s Optmzaton Manufacturng actvty, whch we refer to as the aggregate n what follows, s composed of a dscrete number n of sectors operatng under perfect competton. Each sector s characterzed by a representatve rm operatng under constant returns to scale, whch produces a d erentated good. Frms rent captal 7

9 K and labor L and combne them wth ntermedate goods M to produce (gross) output, accordng to the producton functon Y t = A t Kt L 1 t " Mt 1 " = 1; :::; n: (1) Parameter A s a sector spec c technology parameter, possbly changng over tme. Factor shares and " are assumed to be constant over tme, but are also sector spec c. Each sector faces two statc, sector-spec c dstortons: () a dstorton a ectng the rato of ntermedate goods to output, that we model as a tax on ntermedate goods; and () a dstorton a ectng the captal to labor rato, that we model as a tax on captal. 8 problem are de ned as p Y K rk In each perod, the pro ts of sector s representatve rm wl M p M M omttng the tme subscrpt from now on. Maxmzng pro ts subject to the producton functon constrant gves us the followng rst order condtons: K;L K = w L 1 r K M;Y M = 1 " p Y p M M In a world wthout nancal frctons these two dstortons dsappear ( K;L = M;Y = 1) and the standard rst order condtons equatng factor prces to margnal products would be recovered. For now, we wll take these dstortons as exogenous, as well as the sequences for factor prces (w t, r t, p M t prces (p t). ) and sectoral output 3.2 Aggregatng Sectors We de ne aggregate output (at constant prces) as the sum across sectors of the value of output: and, smlarly, aggregate nputs as Y X n =1 p 0Y K X n =1 K ; L X n =1 L ; M X n M : =1 We also de ne (value-added) aggregate total factor productvty as T F P Y pm 0 M K L 1 usng as the aggregate share of captal an average of the sectoral shares weghted by value added = X n =1! ;! p 0Y0 p M 0 M0 Y 0 p M 0 M : 0 Assumng that factor shares and relatve prces reman constant over tme, we can wrte aggregate TFP n growth rates, where ex t = log xt g T F P = x 0 xt x0 x 0 Y 0 Y 0 p M 0 M ey 0 p M 0 M 0 Y 0 p M 0 M 0 fm K e (1 ) L e (2). The assumpton of constant relatve prces wll be relaxed n Secton 5. For now, t just allow us to focus on the role of sectoral dstortons and reallocaton of factors across sectors. 8 The model would allow us to dentfy a thrd dstorton, for example one a ectng the rato of employment to output. However, t s easy to show that only changes n the two dstortons that we are consderng a ect the growth rate of aggregate total factor productvty once we take nto account reallocaton e ects. 8

10 3.3 Decomposng Aggregate TFP changes In Appendx A.3 we show how we can combne rms optmzaton rst order condtons wth the prevous equaton (2) to obtan Tg F P = X n =1 1 fa! " + f LL + g K;L K;L + g M;Y M;Y where! s the share of the sector n aggregate value added, and K L! 0 L (1 ) 0 ; K K;L 0 K 0 K 0 L 0! ; M;Y pm 0 M 0 Y 0 p M 0 M 0 1 "! " : Equaton (3) provdes a useful expresson decomposng aggregate TFP changes n four components: 1. Changes n sectoral technologes: P n =1! 1 " fa. 2. Reallocaton of labor across sectors: P n =1 L f L. 3. Changes n sectoral dstortons to the captal to labor rato: P n g =1 K;L K;L. 4. Changes n sectoral dstortons to the ntermedates to output rato: P n =1 M;Y g M;Y. To gan some ntuton, notce rst that n an undstorted economy the three coe cents ( L, K;L, M;Y ) are zero for all sectors and aggregate TFP s drven only by changes n sectoral technologes.9 We can also show usng the de nton of the aggregate labor share that P n =1 L = 0. Ths mples that changes n average employment n the same proporton for all sectors do not a ect TFP, only labor reallocaton across sectors does. A sector whch ncreases ts relatve employment sze would contrbute to ncrease aggregate TFP f and only f L > 0,.e., f and only f K! > 0 L + (1 ) 0 K 0 The condton mples that a sector that ncreases ts employment contrbutes postvely to TFP f ts share of aggregate value s larger than ts nput shares, relatve to the other sectors. Ths could occur because the sector s hghly productve (.e. t has a hgher value of A ) or purchases a suboptmal level of nputs. In other words, as expected, reallocaton of labor towards more productve and less dstorted sectors ncreases TFP. Smlarly, we can show that P n =1 K;L = 0. Therefore, a change n the average level of the captal dstorton whch a ects all sector equally does not mpact aggregate TFP. Only changes n dstortons whch are heterogeneous across sectors do because of ther mpact n the dsperson of captal to labor ratos. A sector whch ncreases ts captal dstorton K;L (and therefore decreases ts captal to labor rato) would reduce aggregate TFP as long as K;L < 0,.e., as long as K0 <! K 0 9 We wll not have anythng nterestng to say about the technology component, that we take as exogenous. In a more dsaggregated framework ths component could be also a functon of the dstrbuton of rm-level dstortons wthn a sector, whch we are omttng n our analyss. L 0 (3) 9

11 or that ts ntal captal stock s smaller than ts relatve returns to captal parameter would warrant. Ths means that an ncrease n the dstorton to the captal labor rato wll depress TFP for sectors whch were already dstorted ntally. Movements n the dstortons to the use of ntermedate goods a ects aggregate TFP d erently. The reason s that we are consderng value-added TFP, as opposed to gross output TFP, so n general P n =1 M;Y 6= 0. In our setup, an ncrease n the ntermedates dstorton M;Y reduces aggregate TFP f and only f M;Y < 0,.e., f and only f M;Y > 1 so that the ntal dstorton s postve n ths sector. It s worth emphaszng agan that a reducton n dstortons s not enough to ncrease TFP. Ths reducton must come from sectors whch are hghly dstorted ntally. 3.4 Sectoral Dstortons and Aggregate TFP n the Data In the rst step of our emprcal analyss we take the two dstortons n the use of nputs descrbed before as prmtves and use the establshment level data from the EIA to measure them. The data s annual and aggregated to the 4 dgt NAICS class caton. Once we exclude sectors wth mssng nformaton, we have a total of 82 sectors wthn Manufacturng. Each of these sectors s mapped nto a sector n the model (so n = 82). The result s a panel of sectoral dstortons and other varables for the perod, whose statstcal propertes we report. Usng ths nformaton, we perform the accountng exercse followng the decomposton n equaton (3) to assess the quanttatve contrbuton of sectoral dstortons to aggregate TFP changes. The results show that the components assocated to dstortons account for a large fracton of the varaton of aggregate TFP over tme Measurng Dstortons For each sector, we have data on gross output, employment, the wage bll, ntermedate goods purchased, nvestment and deprecaton. Except for employment, all these varables are measured at current prces. When necessary, we use an aggregate PPI for manufacturng (to de ate gross output, captal and nvestment) and an aggregate ndex for the prce of ntermedate goods to de ate the purchase of ntermedates. 10 construct our measure of the captal stock we use the perpetual nventory method. We use ntal nvestment and a steady-state assumpton to calculate the ntal captal stock. We then update the captal stock usng nvestment ows and a sector spec c deprecaton rate. The constructon of our measure of dstortons follows the formulas n (5). Notce that snce factor shares are not ndependent of dstortons, we cannot dentfy the producton functon coe cents wth these shares. Our strategy s to take the factor shares from the correspondng sectors the U.S., as an example of an undstorted economy. For each sector and at each perod t, dstortons are then computed as: (K;L);t = ;us nomnal wage bll t 1 ;us 0:14 nomnal captal stock t 10 Ideally, we would lke to use sector spec c de ators for output and captal goods de ators for captal stock, but the lack of su cently dsaggregated prce ndces prevents us from dong so. To 10

12 Percent of Sectors Percent of Sectors Intal K,L dstorton Intal M,Y dstorton Fgure 4: Intal Dstrbuton of Dstortons (2003) assumng a 14 percent rate of return (obtaned of the sum of the average real nterest rate on loans and the average deprecaton rate) and (M;Y ); t = 1 " ;us nomnal gross output t nomnal ntermedates purchases : t The producton functon (1) allows us to compute the sectoral technology level A t as A t = h Kt ;us Y t L t 1 ;us " ;us M t 1 " ;us where all varables (except employment) are de ated as descrbed above. Aggregaton of sectors s performed as n the model Sectoral Dstrbutons Fgure 4 plots the hstogram for each type of dstortons n the ntal year. The ntermedates dstortons exhbt a large mass of sectors around the level of one (no dstorton) and a fat rght tal, ndcatng a majorty of sectors wth dstortons that reduce ntermedates to output ratos. The captal-labor dstorton s more unformly dstrbuted, wth a large majorty of sectors (about 80 percent) wth dstortons that reduce the captal to labor ratos. Fgures 5 and 6 plot the hstograms for the changes n the two types of dstortons, n the subperods and As observed, most sectors reduced ther levels of dstortons n the perod prevous to the crss, although there s substantal heterogenety across sectors. In the crss perod the dstrbuton of 11

13 Percent of Sectors Percent of Sectors Percent of Sectors Percent of Sectors Change n K,L dstorton (%) Change n K,L dstorton (%) Fgure 5: Dstrbuton of Changes n Captal-Labor Dstortons Change n M,Y dstorton (%) Change n M,Y dstorton (%) Fgure 6: Dstrbuton of Changes n Intermedates Dstortons 12

14 Average K;L Std/Mean K;L Correl ( K;L,L) Correl ( K;L,A) Average M;Y Std/Mean M;Y Correl ( M;Y,L) Correl ( M;Y,A) Correl ( K;L, M;Y ) Table 1: Descrptve Statstcs for Sectoral Dstortons n Mexco s Manufacturng (yearly % changes) TFP growth due to technology (a) due to reallocaton (b) due to K;L (c) due to M;Y (a) + (b) + (c) resdual Table 2: TFP Growth Decomposton changes shfts to the rght, towards more sectors ncreasng ther level of dstortons or reducng them at a slower pace. An alternatve characterzaton of the dstrbutons of dstortons s provded as a set of summary statstcs n Table 1. The data shows a decrease n the average level of dstortons but an ncrease n ther dsperson across sectors, more marked for K;L than for M;Y:. Both dstortons seem to be correlated wth sectoral productvty, suggestng that more productve rms are constraned n ther use of nputs, and ths correlaton seems to be ncreasng over tme Dstortons and Aggregate TFP Table 2 reports the results of the decomposton of aggregate TFP growth for several perods. The resdual s computed as the actual TFP growth mnus the predcted TFP growth due to the four factors mentoned before, accordng to equatons (2) and (3). Omtted factors such as changes n relatve prces and factor shares, as well as errors n the approxmaton, are ncluded n ths resdual. 11 As expected, sectoral technologes account for a large fracton of aggregate TFP growth. Excludng the resdual, about one-thrd of the growth n , and 38% n comes from growth n sectoral technologes. However, changes n the two dstortons contrbute more than 70% of the growth n TFP n 11 As a check the robustness of the results, we repeated the exercse from the prevous subsecton usng a more dsaggregated sample from Mexcan manufacturng (EIA), at the 6-dgt NAICS level. Ths sample allows us to ncrease the number of sectors from 82 to 215. The results of the TFP growth decomposton are almost dentcal. 13

15 and 40% n In other words the reducton of dstortons and, to a smaller extent the reallocaton of labor, contrbute very sgn cantly to the productvty gans n manufacturng n ths perod. The crss of s re ected n the sharp drop n aggregate TFP. Sectoral technologes themselves fell by even more than the aggregate, however, the fall n aggregate TFP would have been much larger had t not been mtgated by the reallocaton of labor. The last column of the table shows the contrbutons of the sectoral dstortons to the recovery. Improvements n K;L and M;Y accounted for all the productvty gans n ths perod, whle mprovements n sectoral technologes were neglgble. Interestngly, labor reallocaton functoned as a drag on the recovery process. 4 Fnancal Frctons and Sectoral Dstortons We now consder an extenson of the smple partal equlbrum model n Secton 3, but wth nancal frctons nstead of exogenous dstortons. As before, nancal frctons ntroduce wedges between factor prces and margnal products, that can be mapped n our two prevous sector spec c dstortons to the captal to labor rato and to the rato of ntermedate goods to output. Ths new framework establshes then a lnk between credt condtons and dstortons. We present some exploratory panel regressons to test for ths lnk n our dataset. 4.1 A Smple Model wth Workng Captal and Borrowng Constrants As before, the representatve rm n sector purchases captal and labor servces and ntermedate goods to produce (gross) output, accordng to the producton functon spec ed n (1). We now ntroduce two types of nancal frctons to rms optmzaton problem: a smple workng captal constrant and a borrowng lmt. Frms have to nance ther workng captal (ncludng all captal servces and ntermedates purchases) ether through loans from the nancal system or loans from supplers (trade credt). Both types of loans are repad at the end of the perod ncludng a sector spec c, wthn perod, nterest rate t that we take as exogenous. In addton, loans from supplers nclude an exogenous nterest rate premum, re ectng the hgher cost of trade credt. On the other hand, loans from the nancal system are constraned to be no larger than a fracton t of the value of the rm s sales. A large lterature on trade credt nforms our modellng choces. Two emprcal facts about ths form of nancng are well establshed. Frst, trade credt s used more ntensvely by rms who are credt constraned, both n the US 12 and n other countres. 13 It also seems that the use of trade credt s more prevalent n countres wth poor enforcement and less developed nancal systems. Demrguc-Kunt and Maksmovc (2001) nd that the magntude of bank credt relatve to trade credt s hgher n countres wth more e cent legal systems whle Fsman and Love (2003) nd that ndustres wth hgher dependence on trade credt nance grow faster n countres wth weaker nancal nsttutons. Second, trade credt tends to be substantally more expensve than bank credt. A study of more than 30,000 trade credt transactons by Klapper et. al. (2012) shows that the medan annual nterest rate on these transactons s 54%. Cotler (2013) documents the mportance of trade credt for Mexco usng a survey 12 See for example, Petersen and Rajan (1994) and (1997). 13 Gama et. al (2010) nd ths to be true for a panel of Portugese and Spansh rms, whle Couppey-Soubeyran and Hercourt (2011) nd that rms n the MENA regon that have d culty n ganng access to bank credt use trade credt nstead. Atanasova and Wlson (2003) nd that the use of trade credt ncreases durng perods of monetary contractons n the UK. 14

16 carred out by the Central Bank of Mexco. He nds that 82% of all busnesses report usng trade credt to nance ther expenses, and most of them use smultaneously bank credt as workng captal and trade credt. Our calculatons usng the same survey (ENAFIN, for ts acronym n Spansh) reveals that the cost of ths credt was about 5% per month or an annualzed rate of close to 80%. In each perod, the problem of the representatve rm n each sector s to solve max p ty L t ;K t ;M t t ; t s:to: Y t w t L t = A t t + (1 t)) r t K t + p M t M t Kt L 1 t " Mt 1 t r t Kt + p M t Mt tp tyt t 0 t 1; where t s the endogenous fracton of workng captal nanced through the nancal system (we call t smply credt from now onwards). Together wth the nterest rate t, the sequence of parameters t captures credt condtons by governng the tghtness of the borrowng constrant. These credt condtons a ect sectors d erently and can change over tme. For now, the sequences for factor prces (w t, r t, p M t output prces (p t) are all exogenous. " (4) ) and sectoral 4.2 Mappng Credt Condtons nto Sectoral Dstortons Snce the problem for each representatve rm s statc, from now on we omt the subscrpt t. The rst order condtons for pro t maxmzaton mply wl 1 rk = + (1 ) + and p Y (1 ") p M M = + (1 ) + + where s the Lagrange multpler assocated to the borrowng constrant. From these rst order condtons, we can map the sectoral dstortons ntroduced n Secton 3 to the new model as K;L = + (1 ) + M;Y = + (1 ) + + : (5) Notce that dstortons now are endogenous and depend on the sector-spec c credt condtons (, ) together wth two endogenous varables ( and ). These dstortons arse because the shadow cost of credt ncreases the e ectve cost of captal relatve to labor, and of ntermedates relatve to output, dstortng the optmal mx of nputs. In a world wthout nancal frctons these two dstortons wll dsappear. In Appendx A.4 we further characterze the soluton to the rm s optmzaton problem and show how, dependng on the tghtness of the borrowng constrant ( ) relatve to other parameters, one of the followng three cases can arse: 15

17 1. For hgh tghtness the borrowng constrant bnds but < 1. Avalable bank credt s not su cent to purchase all nputs and the rm has to resort to the more expensve trade credt. Then, K;L = + M;Y = + + : 2. For ntermedate tghtness the borrowng constrant bnds and = 1. Avalable bank credt allows rms to avod trade credt, but stll ther purchase of nputs s constraned. In ths case, K;L = 1 1 " 1 " (1 ) M;Y = 1 " 1 : 3. For low tghtness the borrowng constrant s not bndng and = 1. There s no constrant to the rms purchase of nputs and all of t s nanced thorugh bank credt. Therefore, K;L = M;Y = : 4.3 Credt Condtons and Sectoral Dstortons n the Data In ths model, t s easy to show that an ncrease n the sector spec c nterest rate always ncreases both types of dstortons. Also, an ncrease n the borrowng lmt decreases the two dstortons untl the pont n whch the rm becomes unconstraned. We now test these predctons n our dataset, matchng bank credt data aggregated to the 4 dgt sector level to measured dstortons at the same level of aggregaton. As a measure of credt, we use the short term ow of credt to the sector nclusve of nterest labltes. 14 We measure real nterest rates as the nterest rate on the medan loan n the sector n the current year, de ated by the change n the manufacturng prce level Dstortons n the Intermedates to Output Rato Table 3 shows the results for the dstortons on ntermedate goods. Each row of the table represents a seres of regressons, each wth the ndependent varable n the left hand sde column. The rst three columns show a smple OLS wth tme dummes, whle the next three show xed e ects regressons. Columns (7) to (9) show xed e ects augmented wth tme e ects and the last three columns show regressons of each varable nteracted wth tme dummes. For brevty, only the nteractons for the perod are shown. Heteroscedastcty consstent standard errors are gven below the estmates. The rst panel shows regressons wth the credt to output ratos, where the denomnator s obtaned from the sectoral data of the EIA, whle the next two panels are the regressons wth measures of credt ow and nterest rates respectvely. In each case we also present estmates wth addtonal controls for sector sze or productvty. 14 We also use an alternatve measure of short term credt net of nterest labltes wth almost dentcal results. 15 We also used an alternatve measure of nterest rate whch s the average nterest rate, weghted by the sze of the loan. However, ths measure s lkely to be based downwards snce larger loans are assocated wth lower nterest rates n the data. 16

18 Table 3: The Intermedate Goods Dstorton Dependent Varable M;Y Panel A: (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Credt/Output ** * Credt/Output Labor y ** A y y y y Panel B: Credt Flow y y ** * * * * * * Credt Flow ** ** ** ** ** ** Labor y ** * A y y y y Panel C: Interest Rate y y y y y y Interest Rate Labor y y A y y y y Sector E ects No No No Yes Yes Yes Yes Yes Yes Yes Yes Yes Tme Dummes Yes Yes Yes No No No Yes Yes Yes Note: p < 0:15, p < 0:05, y p < 0:01 Heteroscedastcty consstent standard errors below.

19 Columns (1) to (3) of Panel A shows that credt ntensty and M:Y are negatvely related. Ths seems to suggest that the use of credt s an mportant source of mnmzng nput dstortons. Concerns about the endogenety of credt ntensty could arse f more productve sectors, or sectors wth larger collateral have smaller dstortons to nput use and also have more access to credt. However, as columns (2) and (3) show, our results are robust to the ncluson of addtonal controls such as sector sze (measured by number of employees) or sectoral productvty. Interestngly, more productve sectors also have larger dstortons, suggestng that the removal of these dstortons would have large e ects on output. Columns (4) to (6) and (10) to (12) show that the sgn of the estmates s not altered f we nclude sectoral heterogenety and tme varyng coe cents. In all cases the coe cent s not sgn cant, but gven that the numerator and the denomnator of the credt to output rato come from d erent sources, we expect a certan degree of varablty, re ected n the large standard errors. Panel B studes the e ects of actual short term credt ow n the perod. Whle the coe cent s postve when we don t nclude sectoral e ects (Columns (1) to (3)), t s negatve once sectoral heterogenety s taken nto account. In the last three columns, where we consder tme varyng coe cents, we nd that the avalablty of credt matters for the sze of the dstortons n both the crss and the recovery years. Sectors whch were able to secure credt were able to brng ther ntermedate goods usage closer to the optmal level. 16 In all cases, results are robust to addtonal controls of productvty and labor. Fnally, panel C consders the cost of the credt, as measured by the medan nterest rate on short term credt n the sector. The rst three columns show that ths s postvely and sgn cantly related to the dstorton, suggestng that hgher nterest rates prevent rms from usng ther optmal nput mx. The estmates reman postve when we ntroduce sectoral heterogenety (Columns (7) through (9)) but they are no longer sgn cant. The last three columns show that whle the coe cent was postve even n the crss and recovery years, t s not sgn cant. One concern here of course s how to create a representatve nterest rate for a group of loans n a sector. We have expermented wth alternate nterest rate measures such as averages of the nterest rates on all short term loans n the perod, weghted by loan sze, but ths measure has the drawback that t may over-represent low nterest loans and be based downwards. We also plan to experment wth averages weghted by loan maturty n further research Dstortons n the Captal Labor Rato Table 4 shows the relatonshp between the captal labor dstorton and ndcators of credt ntensty, avalablty and cost. The table s organzed n a smlar fashon as Table 3. In general, panel A shows that sectors wth a greater credt ntensty have lower dstortons. In partcular, the last three columns of ths panel ndcate that greater credt ntensty n the crss and recovery mpled that rms could get closer to ther undstorted optmal captal labor rato. As n the prevous estmatons, more productve sectors and sectors wth more employees have bgger dstortons n ther captal labor rato. 16 Column (10)-(12) n all tables s estmated wth a full set of nteractons ( ) but only the last three years are shown for compactness. None of the nteractons n the prevous years are sgn cant. Tme varyng coe cents were estmated for all other ndependent varables but yelded no results of nterest. All results avalable on request from authors. 18

20 Table 4: The Captal Labor Dstorton Dependent Varable K;L Panel A: (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Credt/Output ** ** ** Credt/Output * * * y y y ** ** ** Labor y y A y y y y Panel B: Credt Flow y y y Credt Flow * * ** y y ** y y y Labor * y y y A y y y y Panel C: Interest Rate y y y * * Interest Rate ** * ** Labor y y y y A y y y y Sector E ects No No No Yes Yes Yes Yes Yes Yes Yes Yes Yes Tme Dummes Yes Yes Yes No No No Yes Yes Yes Note: p < 0:15, p < 0:05, y p < 0:01. Heteroscedastcty consstent standard errors below.

21 The relatonshp between dstortons and credt ow s presented n panel B. The negatve relaton between these two ndcates that nancal constrants are an mportant factor underlyng dstortons. As columns (10) to (12) show, the avalablty of credt was partcularly mportant n the recovery from the crss. In other words, sectors wth greater avalablty of credt were able to reduce ther nput dstortons durng the crss. Gven the contrbuton of the dstortons to aggregate TFP, as evdenced n table 2 ths result ponts to the relatonshp between nancal frctons and TFP, through the e ect of the former on nput allocaton. Fnally the lowest panel shows the e ect of nterest rates on the captal labor dstorton. An ncrease n the cost of borrowng s assocated wth an ncrease n dstortons and s sgn cant n several cases. However, the cost of credt as measured by the medan nterest rate n the sector does not seem to play an mportant role n the recovery, suggestng that t was the avalablty of credt, rather than ts cost, whch was mportant n reducng dstortons. As mentoned earler, t s probably worth expermentng wth alternatve measures of the cost of credt to establsh the robustness of these results. It may also be that credt ratonng occurs by quantty, not prce, n the market for loans and so the cost of credt s not as mportant as ts avalablty In Summary The two sets of estmates n tables 3 and 4 taken together, hghlght the role of credt n explanng the dstortons on nput use. Both the ow of credt and the credt ntensty are negatvely related to the sze of dstortons n the optmal use of ntermedates and the captal labor rato. As mentoned earler, we have already seen the mportance of these dstortons n explanng aggregate TFP. Whle that was an nformatve accountng exercse, the estmates presented n ths secton gve some content to what les behnd the dstortons. Our results suggest that the amount of credt avalable to the sector s an mportant determnant of ts ablty to acheve ts best nput mx. We also nd that credt plays an especally mportant role n the recovery of the economy from the 2008 crss. Ths s noteworthy because ths recovery takes place wthout a correspondng ncrease n aggregate real credt and aggregate credt ntensty, as shown by Fgure 2. Our results therefore shed lght on an mportant puzzle n economcs,.e. the phenomenon of credtless recoveres. As Calvo et. al. (2006) document, output n many emergng economes recovered after nancal crses wthout a correspondng recovery n credt. However, our estmates show that recoveres can take place as long as credt goes to sectors that can reduce ther dstortons, regardless of the aggregate level of credt. 5 Fnancal Frctons and Aggregate TFP The prevous two sectons have shown an ndrect connecton between credt condtons and aggregate TFP, va the e ect of nancal frctons on sectoral dstortons and nputs use. In ths nal secton we put the peces together and analyze the mpact of exogenous credt shocks calbrated from the data on aggregate TFP. In partcular, we ask the queston of how much can we account of the movements of TFP n the data usng these credt shocks as prmtves, nstead of the sectoral dstortons (as n Secton 3). For ths, we embed the model of rms optmzaton wth workng captal and borrowng constrants from Secton 4 n a general equlbrum 20

22 structure. 17 The model s calbrated usng our dataset and used to perform counterfactual experments. 5.1 The General Equlbrum Model We consder an economy wth n sectors producng manufacturng goods under perfect competton. Each of these producers faces the sequence of statc optmzaton problems descrbed n equaton (4) above. The output of each sector s used ether for the producton of the composte nal good or the composte ntermedate good (Yt the constant returns to scale aggregators = Y ;F t + Y ;M t ). These composte goods are also produced under perfect competton usng Y t = nx =1! (Y ;F t ) 1! 1 for the nal good, wth elastcty of substtuton > 1 and nx! = 1, and =1 M t = nx =1! (Y ;M t ) 1! 1 for the ntermedate good. From standard rst order condtons we construct the prce of the nal good p t = nx =1! 1 1 1! (p t) 1 = 1 from the prces of the ntermedate goods produced by each sector. 18 The prce of the nal good s the numerare. A representatve consumer makes all savngs and nvestment decsons n a small open economy setup. The consumer maxmzes the standard ntertemporal utlty 1X t u(c t ) subject to the budget constrant t=0 C t + I t + B t+1 B t = nx =1 w t L t + r t K t + r t B t + T t : The consumer receves ncome from rentng captal and labor to the manufacturng sectors. An addtonal source of ncome comes from the nterest payments on trade credt, whch are rebated as a lump sum transfer 17 We need a general equlbrum model of the economy beacuse, as shown n Secton 3, the reallocaton of factors across sectors s key for ths accountng exercse. Wth constant returns to scale, the relatve sze of sectors s determned from the demand sde. Also, changes n factor prces would a ect the relaton between sectoral shocks and aggregate TFP. 18 From the same set of rst order condtons we obtan the demand for each sector s output n the nal good producton Y ;F 1 t =! p t Yt and n the ntermedate good producton Y ;M 1 t =! p t t M t: Under constant returns to scale, these demand functons allow us to pn down the sze of each sector and ts change over tme. 21