How much does management affect productive performance? New insights from a semi-nonparametric analysis of the World Management Survey

How much does management affect productve performance? New nsghts from a sem-nonparametrc analyss of the World Management Survey Andrew L. Johnson Texas A&M Unversty, College Staton TX
of 18
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
How much does management affect productve performance? New nsghts from a sem-nonparametrc analyss of the World Management Survey Andrew L. Johnson Texas A&M Unversty, College Staton TX , USA Aalto Unversty School of Busness, P.O. Box 21210, FI AALTO, FINLAND Tmo Kuosmanen Aalto Unversty School of Busness, P.O. Box 21210, FI AALTO, FINLAND PRELIMINARY AND INCOMPLETE: PLEASE DO NOT CITE WITHOUT AUTHORS Abstract PERMISSION It remans a challenge to demonstrate the effects of management for mprovng performance beyond smply relyng on case studes and anecdotal evdence. The World Management Survey presents a unque opportunty to look more closely at the relatonshp between management and performance. Ths paper crtques pror research and offers alternatve sem-nonparametrc estmaton technques. Fndngs reveal that the effect of management vary sgnfcantly across countres, that some management practces are more mportant than others, and that management has a sgnfcant effect on output, even n a cross-sectonal analyss. Keywords: Management practces, productvty; Competton; Effcency; Nonlnear Programmng; Productve Effcency Analyss; Benchmarkng JEL Classfcaton Codes: L2, M2, O32, O33 1. Introducton News headlnes about the salares of CEOs, Gabax and Lander (2008) and the latest scandals pont to the contnung debate concernng the role of management n global socety. But how much does management really matter to the world at large? The World Management Survey descrbed n Bloom and Van Reenen (2007) s the frst data set to nclude survey nformaton for multple countres across multple sectors extensve enough to allow researchers to begn quantfyng the effects of management. 1 Snce ts ncepton n 2004, the World Management Survey (WMS) has gathered data through drect ntervews wth company managers. The survey s executed by a management consultng frm under gudance from Ncholas Bloom and John Van Reenen. A double blnd approach s used to mnmze bas n data collecton. The survey focuses on four specfc areas: operatons (3 questons), montorng (5 questons), targets (5 questons), and ncentves (5 questons). The survey questons are weghted equally, standardzed and aggregated to a sngle management score. Whle perhaps addng to the robustness of the measure of management, aggregaton also contrbutes to the loss of nformaton. Therefore, ths paper nvestgates the dsaggregated management survey results to determne f further nsghts can be ganed. Whle the survey nformaton gathered has the potental to provde ground-breakng emprcal evdence that management matters, a careful nvestgaton of more plausble model specfcatons usng sem-nonparametrc estmators has the potental to strengthen the credblty of the survey results. Ths paper examnes the 2004 World Management Survey (hereafter WMS) to assess the robustness of Bloom and Van Reenen s concluson that better management s assocated wth hgher output. Specfcally, we nvestgate four ssues: 1) the assumpton of constant practces over tme based on survey results gathered n 2004 and performance nformaton from ; 2) the assumpton of constant management effects over countres whle allowng the productvty of other nput resources to change; 3) the use of a Cobb-Douglas functonal form that mposes the elastcty of substtuton among nputs and between nputs and management to be 1; and 4) the use of an aggregate measure of management. Our analytcal objectve s to determne f better management leads to hgher output condton on nput resources and the effect 1 We focus on the 2004 survey data results. Bloom et al. (2012) present an extended data set ncludng surveys taken n 2006 and 2009; however even n ths extended data, the rato of ntervews to frms s 10,161 /8,117 = 1.25, whch ndcates that the majorty of frms were only ntervewed once. Thus, the economc and econometrc concerns we rase also apply to ths larger data set. of competton. We dscuss these four modelng and estmaton ssues whch we pose as premses below. Our frst premse s that f management matters, then changes n management over tme should lead to dfferences n performance. Thus, we need to measure management and performance at the same tme. If there s a lag n measurng management, then t may be possble to attrbute the poor performance resultng from pror poor management to current management that has adopted, better, or best, management practces. Thus, we consder a crosssectonal analyss n whch we only analyze the performance data for Our second premse s that the effectveness of management practces depends upon the envronment n whch they are appled. For example, the Japanese culture s often cted as a reason for the superorty of Toyota s manufacturng system or lean manufacturng prncples Hno (2005). Whle Bloom and Van Reenen s orgnal analyss of the WMS data allows the effects of nputs to vary across countres by estmatng country specfc slope parameters for the Cobb-Douglas producton functon, the authors estmaton of a common effect of management across countres mples that the effects of labor, captal or raw materals may dffer across countres, but the effect of management s modeled as constant across countres. Thus, we consder country specfc management effects n order to compare the effects of management across countres. Our thrd premse s the possble msspecfcaton of the functonal form for the producton functon. Clearly, the varables quantfyng management and nputs could be correlated f more effectve managers are found n larger frms; however f the functonal form of the producton functon s also msspecfed, ths would lead to the effects of the nputs beng attrbuted to the management varables, thus overstatng the effect of management. Further, the Cobb-Douglas producton functon n Bloom and Van Reenen has the very restrctve propertes that the elastcty of substtuton between nputs s equal to one for all nput pars and the elastcty of substtuton between management and nputs s also one. Relatve to a less restrctve assumpton, t s not clear f the restrcton on the elastcty of substtuton would bas the analyst toward over or under estmatng the effect of management. Our fourth premse s that certan dmensons of management may matter more. Thus, we consder the dsaggregate WMS results. Our paper unfolds as follows. Secton 2 models a producton functon and the effects of management and other controls, such as competton, by usng the Convex Nonparametrc Least Squares (CNLS). Secton 3 dscusses the emprcal results of applyng the model we develop. Secton 4 suggests alternatve conclusons drawn from nvestgatng the WMS and why management matters. 2. Modelng and estmators We consder the jont estmaton of a producton functon aggregatng multple nputs and the effect of management varables and competton n a cross-sectonal producton model usng the followng sem-nonparametrc, partal log-lnear equaton lny = ln f( L, K, N ) + δm + γz + ε (1) where Y denotes the output of frm, L = labor, K = captal and N = ntermedate nputs m (materals) collectvely are the nputs to the producton functon, φ : R+ R+ s a classc (ncreasng and concave) producton functon, M R r denotes a row vector of contextual varables that characterze the measured management, and Z R t denotes a row vector of controls. The vector δ = ( δ... δ ) s a column vector of unknown parameters (to be estmated) representng the average effect of the management varables 1 M on performance, the vector γ = ( γ,..., γ ) s a column vector of unknown parameters representng the average effect of the control varables t Z on performance, and ε s a random dsturbance term representng the effects of omtted factors, measurement errors, and other stochastc nose. Here, we assume varable ε s an ndependently dstrbuted random varable that s uncorrelated wth the nput varables X = ( L, K, N), the management varables M, and the control varables We characterze the emprcal data of a sample of n frm by the output vector Y, n m nput matrx X, n r matrx of management varables M, n t matrx of control varables Z, and use 0 = (0 0) and = (1 1) Z. 1 wth approprate dmensons. Denotng = [ ] B X M Z, we assume that matrx B has full column rank. However, we allow any two columns of B to be correlated. 1 r For the parametrc part of the model (.e. δm + γz ) we assume a lnear functonal form, notng that the elements of M and transformaton (e.g., exponental or logarthmc). Further, vectors Z can frst be transformed by usng sutable data M and Z may well nclude quadratc, cubc or hgher-order polynomal transformatons of the orgnal management data or controls data. The elements of M n the WMS are prmarly bnary dummy varables representng categorcal and/or ordnal data descrbng management practces. The jont estmaton of the producton functon and the effect of management varables s crtcal. Often the relatonshp between productvty and management or other control varables s nvestgated va two-stage methods. In the frst stage a producton functon s estmated and the resdual s used to represent total factor productvty. And n the second stage productvty s regressed on management scores, Black and Lynch (2001) and Bloom and Van Reenen (2007). Smar and Wlson (2007) dscuss the lmtatons of two-stage methods, emphaszng the separablty assumptons necessary. Specfcally, f management does affect the output level and t s excluded from the frst stage producton functon estmaton, then the frst stage suffers from an omtted varable bas. We propose to avod ths ssue by jontly estmatng the producton functon and the effect of management as stated n model (1). 2.1.Sem-Nonparametrc Estmators wth Management Varables We estmate (1) va convex nonparemetrc least squares (CNLS), Hldreth (1954) and Kuosmanen (2008). We note that CNLS s partcularly suted for producton analyss because the typcal functons of nterest have axomatc propertes, such as monotoncty, convexty, or homogenety, that are easy to mpose n a CNLS formulaton. Kuosmanen (2008) has shown that the set of functons satsfyng the set of producton functon axoms (contnuous, monotonc ncreasng and globally concave) can be equvalently represented by a famly of pece-wse lnear functons that are characterzed usng the Afrat nequaltes (Afrat, 1967, 1972). Therefore, we state the Afrat nequaltes for a monotonc concave functon mappng multple regressors X to a sngle output Y as Y α + βx α + β X h, h h = α + βx + ε β 0 (2) where α and β defne the ntercept and slope parameters of the tangent hyperplanes that characterze the estmated producton functon. Symbol ε denotes the CNLS resdual. Note that n (2) the Greek letters are varables and the Latn letters are parameters (.e., (X, Y ) s observed data). We mpose the Afrat constrants on a standard least squares regresson estmaton, where rearrangng (1) allows us to state the resdual as ( f ) ε = lny ln ( X ) + δm + γz (3) Convenently, the varables M and Z appear only n the resdual whch wll appear n the objectve functon of CNLS regresson, wth no effect on Afrat nequaltes. We use the a varant of the estmator descrbed n Johnson and Kuosmanen (2011) whch we wll refer to as CNLS-M to estmate model (1) n mn (lny ln ˆ φ δm γz ) αβδγ,,,, ˆ φ = 1 st.. ˆ φ = α + βx = 1,..., n α + βx αh + β hx h, β 0 2 where agan α and β defne the ntercept and slope parameters of tangent hyperplanes that characterze the estmated pece-wse lnear fronter (note that βx = βll + βk K + βn N ). Symbol ˆ φ s the estmated output level assocated wth nput level (4) X. Because (4) s a nonlnear programmng (NLP) problem havng a nonlnear objectve functon and a system of lnear nequalty constrants, we solve t by usng standard NLP algorthms and solvers such as MINOS, CONOPT, KNITRO or PATHNLP. In problem (4), parameter vector δ and γ are common to all observatons. The CNLS-M estmator can be seen as a restrcted specal case of the models presented n Kuosmanen and Johnson (2010) and Kuosmanen and Kortelanen (2012), where Z and M are subsets of X for whch β = β, j = 1,..., n. j The CNLS-M estmator (4) has several desrable statstcal propertes. Johnson and Kuosmanen (2011), who examned the statstcal propertes of ths estmator n detal, have shown ts unbasedness, consstency, and asymptotc effcency. Most mportant for our analyss, these results mply that conventonal methods of statstcal nference from lnear regresson analyss (e.g., t-tests, confdence ntervals) can be used for asymptotc nferences regardng coeffcents δ and γ. In other words, the standard result of asymptotc normalty of the OLS coeffcents extends to the CNLS-M estmator of management and competton varables. We note that even though estmator (4) ncludes a nonparametrc functon n addton to a lnear regresson functon, the presence of the nonparametrc functon does not affect the lmtng dstrbuton of the parameter estmator n the lnear part. Johnson and Kuosmanen (2011) have shown that the estmates of δ and γ converge at the standard parametrc rate despte the presence of the nonparametrc producton functon n the regresson equaton. Snce performng statstcal nference on δ and γ may not appear obvous due to the complexty of the NLP formulaton (4), Kuosmanen et al. (2014) have proposed to run an OLS regresson, lny ln( ˆ φ + 1) = δm + γz + ε, n order to yeld the same coeffcents ˆδ obtaned as the optmal soluton to problem (4). Runnng such a regresson also returns the standard errors and other standard dagnostc statstcs, such as t-ratos, p-values, and confdence ntervals. 3. Emprcal Results To valdate that management effects output, we apply the CNLS-M estmator to the WMS 2004 data. 2 The survey data was gathered va a prelmnary ntervew and a frst and second ntervew from 709 medum-szed publc manufacturng frms excludng clents of the partnerng consultancy descrbed n more detal n Bloom and Van Reenen (2007). The performance data was gathered from Amadeus for European frms and Compustat for U.S. frms. We estmate (1) separately for each country usng a cross-sectonal analyss ncludng the survey data and 2003 performance data. 3 The ntervews were double blnd and ncluded 20 survey questons whereby the respondent could evaluate the frm as best practce (5) or worst practce (1). The 20 survey questons and ther mappng to varable names and performance areas are summarzed n table For the U.S s used, for Germany 2003 s used, and for U.K. and France 2002 s used. Table 1: Mappng of survey questons to areas and varable names Survey Questons Summary Area Varables Modern manufacturng, ntroducton Operatons alean1 Modern manufacturng, ratonale Operatons alean2 Success of modern manufacturng technques alean3 Process documentaton Operatons aperf1 Performance trackng Montorng aperf2 Performance revew Montorng aperf3 Performance dalogue Montorng aperf4 Consequence management Montorng aperf5 Target breadth Targets aperf6 Target nterconnecton Targets aperf7 Target tme horzon Targets aperf8 Stretch targets Targets aperf9 Performance clarty and comparablty Montorng aperf10 Instllng a talent mndset atalent1 Managng human captal Targets atalent2 Rewardng hgh performance Incentves atalent3 Removng poor performers Incentves atalent4 Promotng hgh performers Incentves atalent5 Attractng human captal Incentves atalent6 Retanng human captal Incentves atalent7 Table 2 presents some descrptve statstcs of the nputs, outputs, and competton varables for the four countres n 2003 along wth the aggregate management ndcator. Table 2: Descrptve statstcs of the outputs, nputs, management and competton varables Sales Labor Captal Materals Aggregate management ndcator Imperfect competton dummy USA Mean 513,114 2, , , St. Dev. 648,556 2, , , Skewness , Mn 8, , Max 3,995,902 16,000 1,522,006 2,913, UK Mean 173, , , St. Dev. 417, , , Skewness Mn 7, Max 4,048,836 7, ,171 7,942, GER Mean 513,482 2, , , St. Dev. 627,211 2, , , Skewness Mn 18, ,522 5, Max 2,936,797 10, ,142 2,150, FRA Mean 146, ,495 63, St. Dev. 217, ,095 89, Skewness Mn 17, Max 1,717,843 8, , , Secton 2 addresses our four ssues regardng prevous analyss of the WMS. The frst s that whle the survey results were gathered n 2004, Bloom and Van Reenen assocate the survey data wth performance data from , mplctly assumng constant practces over ths 11- model. 4 Second, the prmary specfcaton used n Bloom and Van Reenen allows the margnal survey. 5 Thrd, the orgnal analyss of the WMS data uses the Cobb-Douglas functonal form to year tme perod. Clearly, t s not plausble that the frms management practces dd not change over ths tme. Snce a cross-sectonal analyss assocatng a frm s performance wth the management practces at the partcular pont n tme the management survey was gathered would relax ths assumpton, we gather the performance data for 2003 and estmate a cross-sectonal productvty of nputs to change across countres, and country specfc slope parameters are ncluded n the Cobb-Douglas specfcaton. However, the management effect s held constant over countres. We relax ths assumpton and estmate (1) for each country separately whch allows us to estmate the effect of management n each country. Our more flexble approach would be partcularly crtcal f there were cultural characterstcs of the country that lend themselves to beneftng more or less from the management practces measured by the WMS defne the functonal relatonshp between nputs and output. Whle extensvely used n a wde varety of economc applcatons for nearly 100 years, the Cobb-Douglas functon puts consderable restrctons on the flexblty of the functon, specfcally the elastcty of substtuton among nputs and between nputs and management to be 1. Therefore, we estmate (1) for each country usng the CNLS-M estmator (4), whch allows us to mpose the axoms of monotoncty, convexty, and constant returns-to-scale on the producton functon wthout restrctng to a specfc functonal form. Note that f nputs are correlated wth the management varables whch would be expected f mproved management ncreased the probablty that a frm contnued n operatons and grows, then a msspecfcaton of the producton functon causes the effect of the nputs to show up n the estmates of the management parameters. Table A1 n the appendx reports the correlatons between nputs and the management varables and all correlatons are postve for the 2003 performance data and the WMS data. Ths would lead to over-estmatng the effect of management. 4 We would have preferred to use 2004 data; however, there were only 9, 2 and 21 observatons for the French, German, and U.K. frms n the performance data; thus we used 2003 data. However, we report the results of the analyss of 2004 U.S. data. 5 Bloom and Van Reenen suggest one example,.e. that the survey may be based toward an Anglo-Saxon defnton of good management and thus countres wth ths cultural background mght beneft more from these management practces. Ths s just one example, but many others are possble. Fourth, some management practces may have a detrmental effect on output. Whle Bloom and Van Reenen present some lmted results regardng the analyss of the aggregate data, we focus on the dsaggregate data. Ths wll allow specfc advce to be gven regardng the management practces that are most effectve. We begn by frst estmatng country specfc model (1) usng CNLS-M va the sngle aggregated management ndcator reported n Table 3. The estmated management coeffcents conform to the results reported by Bloom and Van Reenen (2007). Management has a postve effect on productvty n all countres. Management s statstcally sgnfcant for the U.S. and France, but not for the U.K. and Germany. The three nputs (labor, captal, and materals) explan more than 95% of the varance of sales across frms, whereas the aggregated management ndcator has mnmal explanatory power. For the U.S. n 2003 and 2004 management predcts approxmately 2% and 5% of the varaton n sales respectvely, but for the U.K., Germany and France t predcts less than 1% of sales. For the U.S. frms, we estmate the effect of aggregate management ndcator usng the nput-output data from years 2003 and 2004 as separate cross sectons. In both years, the management has a sgnfcant po
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!