"Measuring and Analyzing Aggregate Fluctuations - St. Louis Fed

The views expressed are theauthor’s and do not necessarilyreflect the opinions of the U.S.Census Bureau. The author isgrateful for many helpful dis-cussions with Martin Baily, BobBechtold, Eric Bartelsman,Ricardo Caballero, AndrewFilardo, Pu Shen, Steve Davis,Tim Dunne, Chuck Hulten,Frederick Knickerbocker,Sharon Kozicki, C.J. Krizan,Bob McGuckin, Al Nucci, ArnieReznek, Scott Schuh, JohnShea, Ken Troske, and JohnWallis. Special thanks to dis-cussants Jeff Campbell andAlan Heston for their verythoughtful comments.Financial support from theNational Science Foundation isgratefully acknowledged. Thefirst draft of this paper waswritten while the author was avisiting scholar at the KansasCity Federal Reserve Bank inSummer 1996. The authorthanks the research staff at theKansas City Fed for providing astimulating environment inwhich to think about theseissues.

1 References and detailed discus-sion of selected studies are pro-vided in the section entitled“Micro Heterogeneity andAggregate Fluctuations: A BriefReview of Recent Evidence.”

F E D E R A L R E S E RV E BA N K O F S T. L O U I S

55

MAY/JU N E 1997

Measuring andAnalyzing AggregateFluctuations:The Importance of Building fromMicroeconomicEvidence

John C. Haltiwanger

A

pervasive finding in recent researchusing longitudinal establishment-level data is that idiosyncratic factors

dominate the distribution of output,employment, investment, and productivitygrowth rates across establishments.1

Seemingly similar plants within the sameindustry exhibit substantially differentbehavior on a variety of measures of realactivity at cyclical and longer-run frequen-cies. In the fastest-growing industries, alarge fraction of establishments experiencesubstantial declines, whereas in the slow-est growing industries, a large fraction ofestablishments exhibit dramatic growth.During severe recessions virtually allindustries decline, but within each indus-try a substantial fraction of establishmentsexhibit substantial growth. Likewise, dur-ing robust recoveries, a substantial frac-tion of establishments are contracting.Simply put, the underlying gross micro-economic changes in activity dwarf thenet changes we observe, based on pub-lished aggregates.

Table 1 provides a simple characteriza-tion of the dominance of within-sectorfactors in accounting for variation ingrowth rates across establishments. Table1 is based on the computation of establish-ment-level growth rates during a 10-yearperiod for employment, capital stocks,output, labor productivity, and total factorproductivity for plants that appear in boththe 1977 and 1987 Census of Manufact-

ures. As indicated in Table 1, four-digitindustry effects account for less than 10percent of the cross-sectional variation ingrowth rates across continuing establish-ments for each of these measures.

The observed tremendous within-sector heterogeneity raises a variety ofquestions for our understanding and mea-surement of key macro aggregates. Muchof macroeconomic research and our mea-surement of aggregates is predicated onthe view that building macro aggregatesfrom industry-level data is sufficient for

understanding the behavior of the macro-economy. The implicit argument is that, atleast at the level of detailed industry, theassumption of a representative firm orestablishment is reasonable.

The finding of tremendous within-industry heterogeneity is not by itselfsufficient to justify abandoning this usefulassumption. As Lucas (1977) eloquently

John Haltiwanger is professor of economics at the University of Maryland and chief economist at the Bureau of the Census. Lucia Foster andAndrew Figura provided research assistance.

Fraction of Variance ofEstablishment-LevelGrowth Rates: Four-DigitIndustry EffectsDependent Variable

R 2

Employment growth 0.057Capital equipment growth 0.062Capital structures growth 0.052Output (gross) growth 0.089Labor productivity growth 0.086(gross output per hour)Total factor productivity growth 0.095

SOURCE: Tabulations from Longitudinal Research Database(LRD). Reported results are based on computed 10-year growth rates for establishments present inboth the 1977 and 1987 Census of Manufactures(CM). For such continuing establishments, the reported

R 2 is based on the regression of the esta-blishment-level growth rate for the indicated mea-sure on four-digit industry fixed effects. See Ap-pendix for discussion of the measurement of each of these indexes at the establishment level.

Table 1


56

MAY/JU N E 1997

argues in defense of representative agentmodels, there is undoubtedly considerablecancelling out of the impact of idiosyn-cratic shocks (e.g., taste, cost, and tech-nology) that underlie the heterogeneousfortunes across individual producers.However, the accumulating evidence fromrecent establishment-level studies ofemployment, investment, and productivitygrowth suggests that this canceling out isfar from complete. It is becoming increas-ingly apparent that changes in the keymacro aggregates at cyclical and secularfrequencies are best understood bytracking the evolution of the cross-sectionaldistribution of activity and changes at themicro level.

A number of different factors arepotentially important in this context. Theobserved heterogeneity in output, employ-ment, and investment growth rates withinsectors implies a large, continuous pace ofreallocation of real activity across produc-tion sites. Such reallocation inherentlyinvolves substantial frictions. An obviousand important friction is that it is time-and resource-consuming for workers (andfor other inputs) to reallocate acrossproduction sites. High- and low-frequencychanges in key macro aggregates are likelyassociated with the interaction of thesefrictions and the pace of reallocation. Thelevel of unemployment, as well as thegrowth rate of aggregate measures of realactivity (e.g., real output or productivity),will reflect the efficiency of the economyin accommodating the pace of reallocation.Changes in institutions, regulation, thepace of technological change, and the sec-torial mix of activity are all factors that mayalter the intensity of reallocative activityand the economy’s ability to accommodatethe reallocation.

In a related manner, it is important toconsider the nature of the adjustment costsat individual production sites in changingthe scale and scope of activity. Accumu-lating empirical evidence of lumpymicroeconomic adjustment of inputs likeemployment and capital suggests the pres-ence of nonconvexities in micro-adjustmentcosts or, at the minimum, it implies highly

nonlinear adjustment at the micro level.Nonlinear micro adjustment in combi-nation with micro heterogeneity haveimportant implications for aggregate fluc-tuations. One key implication is time-varying elasticities of aggregates withrespect to aggregate shocks. Roughlyspeaking, time-varying elasticities arise inthis context because the impact of anaggregate shock depends on the distribu-tion of where individual producers arewith respect to their adjustment thresh-olds. Viewed from this perspective,characterizing aggregate fluctuationsrequires tracking the evolution of the his-tory of the distribution of shocks andadjustments.

In the context of these heterogeneityand aggregation issues for aggregate fluct-uations, this article has two related objec-tives. The first objective is to quantify andassess the empirical importance of theseheterogeneity and aggregation issues. Theempirical questions to be evaluated include:Where, when, and how much do theseheterogeneity and aggregation issuesmatter for aggregate fluctuations? Iaddress these questions by summarizingand extending the recent empiricalevidence, using establishment-level data.This evidence is primarily based onresearch which uses the LongitudinalResearch Database (LRD), which is basedon the longitudinal linkage of establish-ment-level data from the Annual Survey ofManufactures (ASM) and Census of Manu-factures (CM). As such, the evidencepresented is primarily restricted to the U.S.manufacturing sector.

My second objective is, in light of thisevidence, to provide some guidanceregarding the collecting and processing ofdata on real activity by the U.S. statisticalagencies. In considering the second objec-tive, it is important to emphasize that themeasurement of key aggregates like realoutput growth and productivity growth aregenerated from myriad data sources linkedat an aggregate level (e.g., commodity orindustry). The individual ingredientsunderlying these measures (i.e., nominalreceipts or shipments, inventories, prices,


57

2 See, for example, Griliches(1994) and Gordon (1996).

3 In the discussion that follows,the distinction between estab-lishments and companies (adistinction that macroecono-mists do not typically empha-size or appreciate) is vital.Establishments are economicunits at a single physical loca-tion where business is conduct-ed or where services or indus-trial operations are performed.Companies are one or moreestablishments (e.g., GeneralMotors) owned by the samelegal entity or group of affiliat-ed entities.

MAY/JU N E 1997

intermediate inputs, capital stocks, capitalexpenditures, labor, wages, rental prices)are derived from a variety of statistical andfederal agencies’ surveys and economiccensuses of establishments and companies.As emphasized in the recent literature, thequality of the measurement varies widelyacross industries.2 The variation in qualityis partly a result of the substantial differ-ences in the nature and coverage of thesurveys across sectors and partly becauseof a number of unresolved conceptualissues in the measurement of output andinputs for some sectors. However, evenfor the best-measured sectors (e.g., manu-facturing) the information underlyingpublished aggregates (e.g., real output orproductivity growth) are based onmatching information from a variety of dif-ferent sources at an industry level. TheUnited States does not currently collectdata on the activities of the business popu-lation in a comprehensive, integratedmanner. The implication is that buildingthe requisite micro databases necessary toincorporate these heterogeneity and aggre-gation issues in the analysis of aggregatefluctuations is for most sectors currentlydifficult, if not impossible. With this inmind, I consider the possibilities and prac-ticalities of the data development requiredto pursue these objectives.

MICRO HETEROGENEITYAND AGGREGATE FLUCTUA-TIONS: RECENT EVIDENCE

Employment Dynamics

Job Creation and Destruction. Much ofthe recent empirical analysis documentingand analyzing the connection betweenmicro heterogeneity and aggregate fluctua-tions has focused on employment dynamics.One reason for this is that many of thefrictions involving establishment-leveladjustment and the reallocation of realactivity across production sites involvesworkers. A second reason is based on dataconstraints. Establishment-level surveys,censuses, and administrative record data-bases typically include employment.3

Furthermore, employment at the establish-ment is measured reasonably accuratelyand typically not imputed. Thus, in termsof coverage across sectors and time, estab-lishment-level employment data are themost plentiful and are of reasonable quality.

The evidence summarized here isbased primarily on the decomposition ofnet employment growth into job creationand destruction. Job creation is defined asthe sum of employment gains at expandingand new establishments. Job destructionis defined as the sum of employment lossesat contracting and closing establishments.Table 2 provides some summary statisticsfrom studies tabulating job creation anddestruction rates at annual and quarterlyfrequencies from a variety of differentsources. In manufacturing (the sectorwith the most readily available establish-ment-level data for the longest period),annual job creation and destruction rates

Estimates of Average Job Creation and Destruction Rates*

AnnualDataset (Sector) Period Job Creation Job DestructionLRD (mfg) 1972-93 8.7 10.1CWBH (all) 1979-83 11.4 9.9CWBH (mfg) 1979-83 10.2 11.5CWBH (services) 1979-83 10.6 8.7UI-Michigan (all) 1978-88 10.0 9.6UI-Michigan (mfg) 1978-88 6.2 8.5UI-Michigan (services) 1978-88 15.6 11.0

QuarterlyDataset (Sector) Period Job Creation Job DestructionLRD/MTD (mfg) 1947:1-88:4 6.0 6.0CWBH (all) 1978:3-84:1 7.1 6.4CWBH (mfg) 1978:3-1984:1 5.8 6.2CWBH (services) 1978:3-1984:1 7.9 6.7

*As Percentages of Employment

SOURCE: LRD tabulations from Longitudinal Research Database based on method-ology in Davis, Haltiwanger, and Schuh (1996). CWBH tabulations from the Continuous Worker and Benefit History files for six states, reported inAnderson and Meyer (1994). UI-Michigan tabulations of unemploymentinsurance record database reported in Foote (1995). LRD/MTD tabula-tions from spliced data from LRD and BLS Manufacturing Turnover Data(MTD) as reported in Davis and Haltiwanger (1996).

Table 2


58

4 By comparing the quarterlyrates with the annual rates, itis clear that many workersexperience repeated transitionsduring the year or transitionsthat are reversed within theyear. Davis, Haltiwanger, andSchuh (1996) characterize therelative persistence of job cre-ation and destruction rates.

5 This calculation is based on adecomposition of excess jobreallocation that is measuredas total job reallocation lessthe absolute value of net

MAY/JU N E 1997

are large in absolute magnitude. Roughly1 in 10 manufacturing jobs is created in atypical year, and 1 in 10 jobs is destroyedin a typical year. In nonmanufacturing(with spottier information based on tabu-lations from selected states for relativelyshort sample periods), job creation and jobdestruction rates are on average slightlyhigher. Quarterly job creation and jobdestruction rates average around 6 percentin manufacturing and somewhat higher innonmanufaturing.4

The large pace of implied job realloca-tion (measured as the sum of job creation

and job destruction) in both manufactur-ing and nonmanufacturing sectors highlightsthe remarkable fluidity in the distributionof job opportunities across locations in theU.S. economy. Much of this fluidityreflects shifts within narrowly defined sec-tors, rather than between sectors. Forexample, Davis, Haltiwanger, and Schuh(1996) calculate that only 13 percent ofjob reallocation in manufacturing reflectsshifts of employment opportunitiesbetween four-digit sectors.5

One important issue for the relevanceof these statistics for aggregate fluctuationsis the nature of time-series variation in thepace of job reallocation. The top panel ofFigure 1 depicts the annual rates of jobcreation (POS), job destruction (NEG), netemployment growth (NET), and job reallo-cation (REALLOC) for the U.S. manufact-uring sector for the period 1973-93. Thebottom panel depicts quarterly rates of jobcreation, job destruction, net employmentgrowth, and job reallocation for the U.S.manufacturing sector for the period 1947:1-88:4.6 In U.S. manufacturing, thepace of job reallocation varies systemati-cally throughout the cycle at annual andquarterly frequencies. During downturns,job reallocation in manufacturing rises.The countercyclical job reallocationreflects the asymmetric patterns of job cre-ation and destruction throughout thecycle. Although job creation is procyclicaland job destruction is countercyclical, muchof the cyclical variation in net employmentgrowth is driven by the greater cyclicalvolatility of job destruction. The lowerpanel of Figure 1 indicates that thispattern holds for the U.S. manufacturingsector for the entire post-World War II(WWII) period, although the pattern ismore pronounced in the 1970s and the1980s.7

In terms of secular changes, theannual 1973-93 data reveal no obvioustrend in the pace of job reallocation. Thisin itself is striking, given recent concernsin the popular press about rising job inse-curity.8 The quarterly data do not yetextend into the 1990s but offer a depictionof job-flow dynamics during a much

Quarter47 49 51 53 55 57 59 61 63 65 67 69 71 73 75 77 79 81 83 85 87 89

Perce

nt of

Emplo

ymen

t

20

15

10

5

0

-5

-10

SOURCE: Davis and Haltiwanger (1986).POS

Quarterly, 1947:1 to1988:4

NEG NET REALLOC

25

20

15

10

5

0

-5

-10

-1573 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93

Perce

nt of

Emplo

ymen

t

Net and Gross Flow Ratesin Manufacturing

Year

SOURCE: Tabulations from the LRD.POS NEG NET REALLOC

Annual, 1973 to 1993

Figure 1


59

longer period. The pace of job reallocationshows a mild downward trend throughout1947:1-88:4. An interesting aspect of thistrend is that it is primarily accounted forby a mild downward trend in the pace ofjob creation. Thus, in contrast to thecyclical changes in net employmentgrowth, the very low-frequency changes innet employment growth in U.S. manufac-turing appear to be driven more by changesin the rate of job creation than in the rateof job destruction.

Even modest frictions in the face ofthe observed magnitude and time-seriesvariation of job reallocation are likely toyield important implications for aggregatefluctuations. The aggregate implicationsof these job flows for unemployment-ratedynamics have recently been investigatedby Hall (1995). Hall develops a frame-work in which a burst of job destructionbegets further job separations. Workerswhose jobs are destroyed seek new matches,and, by their very nature, new matches aresubject to higher match termination ratesthan the typical match. Hall’s analysisprovides some quite striking empirical evi-dence on these dynamics, showing that animpulse in job destruction yields persistentrebuilding of employment relationships forseveral periods. For example, he finds thatan impulse in quarterly job destructionyields persistently higher inflows intounemployment via permanent layoffs foreight quarters.

Hall’s findings suggest that the processwhereby permanent job destruction begetsfurther employment losses for severalquarters may be an important part of thepersistence that we observe for aggregatefluctuations; there is no shortage of expla-nations for this persistence. But theseexplanations have been generally viewedas unsuccessful or incomplete becausethey can only account quantitatively forrelative short recessions. Although thisapproach looks promising in terms ofaccounting for recessions that persist forsignificant periods, a number of questionsremain. Of particular interest here is whywe observe the burst of permanent (and itis important to emphasize the permanent

component for Hall’s story) job destructionat the onset of recessions. In recent years,some economists have begun developingtheories to explain the magnitude andcyclical behavior of job (and worker) flowsand the connection to aggregate fluctua-tions. Two types of theories have receivedthe most attention.9 One type treats fluc-tuations over time in the intensity ofallocative shocks as an important drivingforce behind aggregate fluctuations. Thesecond type maintains that aggregate shocksare the primary driving forces underlyingbusiness cycles but that the propagation ofaggregate shocks involves intertemporalsubstitution effects changing the incent-ives for the timing of reallocation. For thisarticle’s purposes, the important debateabout the direction of causality and thusthe relative contribution of aggregate andallocative disturbances are not important.10

The relevant point here is that understand-ing aggregate fluctuations requires track-ing the evolution of the distribution ofmicroeconomic changes.

Nonlinear Micro Adjustment. The discus-sion thus far has focused on the aggregateconsequences generated by the resourceand time-consuming nature of realloca-tion. A closely related issue is that theadjustment at the individual producerlevel may be nonlinear. For example,Davis, Haltiwanger, and Schuh (1996)report that about two-thirds of annual jobcreation and destruction are accounted forby establishments with growth rates inexcess of 25 percent in absolute magni-tude. Of this, plant start-ups account for12 percent of annual job creation, whileplant shutdowns account for about 23 per-cent of annual job destruction. Thus, thedistribution of establishment-level employ-ment changes exhibits both considerableheterogeneity and fat tails. The lumpychanges at the micro level in combinationwith the heterogeneity in turn have conse-quences beyond those discussed earlier.

Building on the literature about theaggregation of (

S,s) models, a useful meansof organizing micro data to characterizethe interaction of nonlinear micro adjust-

employment growth. SeeTable 3.8 in Davis,Haltiwanger, and Schuh(1996), for further details anddiscussion.

6 The top panel of Figure 1 isbased on tabulations from theLRD, extending the methodolo-gy for generating annual jobflow statistics developed byDavis, Haltiwanger, and Schuh(1996). The lower panel ofFigure 1 is based on splicingjob creation and job destructionstatistics generated from theBureau of Labor Statistics(BLS) manufacturing turnoverdata along with the quarterlystatistics from Davis,Haltiwanger, and Schuh tabu-lated from the LRD. Themethodology for using the BLSmanufacturing data and splic-ing with the LRD series is dis-cussed in detail in Davis andHaltiwanger (1996).

7 The recent analysis by Foote(1995) suggests that the cycli-cal patterns of job creation anddestruction in nonmanufactur-ing sectors are different thanthose observed in manufactur-ing.

8 More direct evidence that overalljob security has not diminishedin recent years can be found inFarber (1995).

9 See Blanchard and Diamond(1989, 1990), Caballero(1992), Caballero andHammour (1994), Campbell(1995), Davis andHaltiwanger (1990), Hall(1991), and Mortensen andPissarides (1994) for exam-ples of recent models investi-gating these issues.

10 See Blanchard and Diamond(1989, 1990), Caballero,Engel, and Haltiwanger(1997), Campbell and Kuttner(1996), and Davis andHaltiwanger (1990, 1996) for

MAY/JU N E 1997


60

studies that attempt to useworker and job flows to quanti-fy the relative contributionof aggregate and allocativedisturbances for aggregate fluc-tuations.

11 The measure of aggregateemployment growth here is infact average (across establish-ments) employment growth.See Cabellero, Engel, andHaltiwanger (1997) for furtherdiscussion of this issue.

12 This specification is silent onthe distribution of micro adjust-ments underlying the averagerate of adjustment across pro-ducers at a given value of

x.For example, a rate of adjust-ment of 0.5 at a given valueof x could be driven by allestablishments adjusting by 50percent of x or by half of theestablishments exhibiting com-plete adjustment and half zeroadjustment. The results inCaballero, Engel, andHaltiwanger (1997) suggestthat the latter bimodal distribu-tion of adjustments is a closerapproximation to reality foremployment adjustment.

13 An alternative approach takenby Caballero and Engel (1993)for employment growth andCaballero and Engel (1994)for investment dynamics is tospecify functional forms for theadjustment-rate function andthe distributions of the shocksunderlying time-series variationin the cross-sectional distribu-tion. Under suitable restric-tions, the parameters of therelevant adjustment functionand distributions can be identi-fied and estimated using aggre-gate data.

14 A balanced panel of approxi-mately 10,000 large continu-ing establishments is used forthis analysis.

MAY/JU N E 1997

ment and heterogeneity has recently beendeveloped by Caballero and Engel (1992,1993). For employment (or as we will seelater in a similar specification for capitaladjustment), the relationship betweenmicro adjustment, micro heterogeneity,and aggregate fluctuations can be summa-rized in the following simple equation:

(1)

DEt =

exA(x,t) f(x,t)dx ,

where in this case the left side measuresaggregate employment growth, x is thedeviation between desired and actualemployment for an individual producer(i.e., “shortages” that can be positive ornegative), f(x,t) is the cross-sectionaldistribution of shortages across producers,and A(x,t) is the adjustment-rate func-tion.11 The latter measures the fraction ofthe shortage that producers with shortagex close on average in period t.12

This simple specification accommo-dates a wide variety of alternative charac-terizations of adjustment. The partialadjustment model is characterized by set-ting A(x,t) to a constant. In this specialbut well-known case, only first momentsof the distribution of shortages matter foraggregate employment growth. This caseyields the familiar expression that aggre-gate employment growth can be expressedas a function of the deviation between anaggregate measure of desired and actualemployment. More generally, a nonlinearA(x,t) [e.g., that generated by an (S,s)model] yields that higher moments of x areimportant for characterizing aggregate fluc-tuations. For example, suppose A(x,t) = λ 0

+ λ 2xt2, with λ 0 > 0 and λ 2 > 0 so that

Equation 1 can be expressed as:

(2) DEt =λ 0mx(t)+ 3λ 2mx(t)sx2(t)

+ λ 2mx3(t) + λ 2sx

3(t)g x(t),

where mx(t), and sx(t), and g x(t) denotethe mean, standard deviation, andskewness coefficients of the cross-sectionaldistribution of employment shortages attime t. In this simple example, higher

moments of the cross-sectional distribu-tion of shortages affect the evolution ofaggregate employment through mean-vari-ance and variance-skewness interactionterms. It is also the case that the firstmoment affects aggregate dynamics in anonlinear fashion, which is at the heart ofthe time-varying elasticities with respect toaggregate shocks (discussed below).

The attractive feature of this semi-reduced form specification is that, condi-tional on generating a measure for x (quitean exercise), this specification permits acompletely flexible characterization of theempirical microeconomic adjustmentfunction with micro data, permitting mea-surement of x. In addition, conditional ona measure of x, the above specificationyields an exact decomposition of aggregateemployment fluctuations into the contribu-tion of changes in the adjustment-ratefunction and the cross-sectional distribu-tion of shortages.13

This specification has been recentlyused with the establishment-level quarterlydata on hours and employment from theLRD by Caballero, Engel, and Haltiwanger(1997) (hereafter CEH, 1997) for theperiod 1972:1-1980:4.14 An importantissue is how to measure the deviationbetween desired and actual employmentfor an individual establishment. CEH(1997) implement a specification in whichthe deviation between desired and actualemployment is proportional to thedeviation between actual hours per workerand “normal” hours per worker at theestablishment. “Normal” hours ismeasured as the average hours per workerduring the sample period at each plant.The specification is motivated by modelsdeveloped by Bils (1987) and Caballeroand Engel (1993) in which it is assumedthat technology and wage schedules aresuch that if plants did not face costs ofadjusting their level of employment, theywould keep the same number of hours perworker. However, if costs of adjustingemployment are larger—at least in theshort run—then hours per worker will bepositively correlated with the degree of aplant’s employment shortage.


61

MAY/JU N E 1997

Although this specification can be the-oretically justified, the 1972:1-1980:4sample period and the use of hours perworker to construct the measure of x aredictated by data limitations. As Hamermesh(1993) emphasizes, for analysis of thenature of lumpy microeconomic adjust-ment of employment, it is important to usehigh-frequency data since employmentdecisions are undoubtedly made moreoften than annually. In the LRD the onlyquarterly variables collected are hours andemployment and, as of 1980:4, thequarterly hours data were no longercollected. In the section entitled “Implica-tions for Data Collection, Processing, andMeasurement,” I return to the issuesrelating to the fact that this approach toaggregation and heterogeneity naturallyrequires establishment-level informationon more than one variable.

The key finding in CEH (1997) is thatthe adjustment-rate function is highlynonlinear. Figure 2 depicts the average(over time) adjustment-rate function,along with the average (over time) crosssectional distribution from this analysis.15

Establishments are more likely to react (orreact by more) to large employment short-ages than to small ones. For example, onaverage, about 70 percent of a 10 percentshortage remains one quarter later, whileonly 50 percent of a 60 percent shortageremains one quarter later. The averagecross-sectional distribution providesanother view of the tremendous hetero-geneity in the fortunes across individualproducers. It does show, however, thatestablishments spend a large fraction oftheir time within plus or minus 30 percentof their desired employment level.

The micro nonlinear adjustment func-tion implies that higher moments matterfor aggregate fluctuations. Two resultsfrom CEH (1997) help quantify the aggre-gate significance of this microeconomicnonlinearity. First, CEH consider theimpact of adding higher moments of mea-sures of x, relative to a standard aggregateequation based on a partial adjustmentspecification. They find that adding onlytwo higher moments to a standard partial

adjustment specification improves the R-bar squared from 0.647 to 0.793.16 Second,CEH characterize and quantify the time-varying responsiveness of aggregateemployment to aggregate shocks thatemerges in a model with micro nonlineari-ties. Based on Equation 1, it is easy toshow that the marginal responsiveness toan aggregate shock will, in general, begiven by:

(3) Marginal Responsiveness=eA(x,t)[1+ a(x,t)] f(x,t)dx ,

where a (x,t) is the elasticity of the adjust-ment rate at time t, with respect to x.Standard linear models involve the firstterm on the right side of this equation(although without x as an argument) butonly nonlinear models include the secondterm, which involves a weighted average ofelasticities evaluated at different values ofx. With estimates of A(x,t) and f(x,t), thismarginal responsiveness can be estimated.CEH (1997) find that the marginal respon-siveness for employment varies as much as70 percent over time. Furthermore, theyfind that the impact of the time-varyingmarginal response is especially large inrecessions: The decline in the 1974-75recession was 59 percent larger than it

15 The adjustment rate functiondepicted corresponds to a cubicspline fit over a fine grid. CEH(1997) characterize the adjust-ment-rate function in eachquarter (and the correspondingcross-sectional distribution) andfind that the adjustment-ratefunction is relatively stable overtime so that most fluctuationsin aggregate employment areaccounted for by fluctuations inthe cross-sectional distribution.

16 To capture potential asymme-tries between positive and neg-ative adjustment at the microlevel, the two moments addedin CEH (1997) are the secondmoment of x, conditional on xbeing positive, and the secondmoment of x, conditional on xbeing negative.

1.00.90.80.70.60.50.40.30.20.10.0

-1.5 -1.0 0 1.51-0.5 0.5

Relationship Between EmploymentChange and Employment Deviation

Average Adjustment Rate Fraction of Observations0.030

0.025

0.020

0.015

0.010

0.005

0.000

Employment Deviation

Black line is average adjustment (A (x, t)). Blue line is average density (f (x, t)).SOURCE: Caballero, Engel, and Haltiwanger (1996).

Figure 2


62

would have been in the absence ofnonlinear adjustment.

Investment DynamicsNonlinearities in the adjustment

dynamics of capital, driven by irreversibili-ties and related nonconvexities in theadjustment costs of capital, have analogousimplications for aggregate investmentdynamics. Several recent studies of estab-lishment-level investment dynamicsprovide support for the view that microinvestment dynamics exhibit lumpyadjustment. Two recent papers charac-terize plant-level investment as beingdominated by large, scale-investmentepisodes—denoted investment spikes.Doms and Dunne (1994) find that, duringa 17-year horizon, the largest annualchange at an individual plant accounts forapproximately 25 percent of cumulativeinvestment over this period for the plant.Following on this work, Cooper,Haltiwanger, and Power (1995) find thatthe probability of an investment spike isincreasing in the time since the previousspike, lending additional support to theview of a microeconomic environmentwith nonconvexities in the adjustmenttechnology. Using the analogue ofEquation 1 but with x now measured as

the deviation between desired and actualcapital, Caballero, Engel, and Haltiwanger(1995) (hereafter CEH, 1995) characterizethe relationship between plant-levelinvestment dynamics and aggregate invest-ment.

Studies of establishment-level invest-ment dynamics must confront the difficultmeasurement issues in generating estimatesof real investment flows and capital stocks.Individual establishments purchase newand used capital and sell and retire capital.Through 1988, the LRD includes informa-tion on new expenditures, used expend-itures, and retirements (including sales ofassets). These series are exploited in CEH(1995) using an appropriately modifiedperpetual inventory method. The plant-level investment rate that CEH calculatein each period is based on the estimateddifference between real expenditures andreal retirements for the period, divided bythe estimated beginning-of-period realcapital stock.

This approach also requires measuringdesired capital at the plant level. Thespecification in CEH (1995) assumes thatdesired capital is proportional to friction-less optimal capital.17 The latter is asimpler construct and can be derived fromthe standard neoclassical expression. Thisyields a specification in which the devi-ation between desired and actual capital isa function of the output-capital ratio forthe plant, as well as of the cost of capital.

The average adjustment-rate functionA(x,t) and the average cross-sectional dis-tribution, estimated on the basis of thisspecification using annual LRD data forthe period 1972-88, are depicted in Figure3.18 For plants with positive excess capital,the left arm of the adjustment-rate func-tion (to the left of zero) is quite flat andclose to zero, which is consistent with irre-versibilities in investment. In contrast, theright arm of the adjustment-rate functionis highly nonlinear. Plants with large short-ages of capital adjust proportionally morethan plants with small shortages of capital.

As with employment dynamics, thenonlinear adjustment-rate function yieldstime-varying elasticities of aggregate invest-

MAY/JU N E 1997

17 A plant-specific constant is usedto capture the distinctionbetween desired capital (thelevel of capital that would bechosen if adjustment costswere momentarily removed)and frictionless capital (thelevel of capital consistent witha simple neoclassical model inthe absence of frictions).

18 A balanced panel of large, con-tinuing plants (about 7,000) isused in this analysis.

-5

-4

-3

-2

-1

0

0.014

0.012

0.010

0.008

0.006

0.004

0.002

0.000-2 -1 0 21

Relationship Between Investmentand Investment Deviation

Average Adjustment Rate Fraction of Observations

Investment Deviation

Black line is average adjustment (A (x, t )). Blue line is average density (f (x, t )).SOURCE: Caballero, Engel, and Haltiwanger (1995).

Figure 3


63

MAY/JU N E 1997

19 See, for example, Jovanovicand MacDonald (1994) andAndolfatto and MacDonald(1993).

ment with respect to aggregate shocks.For investment, the marginal responsive-ness generated by the analogue of Equa-tion 2 exhibits a procyclical pattern andvaries by as much as 70 percent. Thetime-varying elasticities suggest a possibleexplanation for the often puzzling responseof aggregate investment to cost of capitaland other shocks. The basic idea is thatthe difficulties the empirical aggregateinvestment literature has had in quantify-ing the relationship between aggregateinvestment and the cost of capital are aresult of the failure to incorporate thetime-varying responsiveness generated bythe interaction of nonlinear micro adjust-ment and heterogeneity.

Productivity DynamicsThe heterogeneous fortunes of indivi-

dual producers raises a variety of questionsabout the underlying forces generating theheterogeneity. Several strands of the theo-retical literature on firm dynamics andheterogeneity are helpful in providing guid-ance and in turn suggest that the ongoingprocess of reallocation is likely to be impor-tant for understanding both micro andaggregate productivity growth.

Models of selection (as in Jovanovic,1982, and Ericson and Pakes, 1994)suggest that individual producers faceuncertainty about either their initial condi-tions that determine the level of product-ivity at a particular production site orabout the productivity consequences ofretooling and reorganizing their product-ion processes. The learning processimplies dynamic selection as producerslearn about the success of their start-upsand their attempts at retooling. Vintagemodels of technological change as inSolow (1960); Cooley, Greenwood, andYorukoglu (1994); and Cooper, Halti-wanger, and Power (1995) stress the ideathat new technology is often embodied innew capital. In this environment, the pres-ence of idiosyncratic shocks, along withfixed costs of adjusting capital, helpexplain productivity differences acrossproduction sites. Related models by

Caballero and Hammour (1994), Morten-sen and Pissarides (1994), and Campbell(1995) stress the idea that new technologyis embodied in the creation of new plants,which in turn displace outmoded, olderplants. Sunk costs limit entry so that new,high-productivity plants coexist withlower productivity, older plants. Anotherrelated but distinct class of models charac-terizes the adoption of new technologiesvia the endogenous innovation and imita-tion process.19 In these latter models,producers must incur costs (both directand indirect) to acquire and implementnew technology. In addition, individualproducers are subject to idiosyncraticshocks (e.g., demand, cost, and product-ivity). The presence of these adoptioncosts, along with idiosyncratic shocks,implies variation in technology adoptionand productivity across producers.

The picture that emerges from thisgrowing literature is one in which techno-logical change is a noisy, complex processwith considerable experimentation (interms of entry and retooling) and failure(in terms of contraction and exit) playingintegral roles. The evidence on large-scale,within-sector job reallocation providesindirect support for this perspective. Moredirect empirical analysis of the implicationsof the pace of reallocation and restruct-uring for productivity dynamics has beenrecently provided by Baily, Hulten, andCampbell (1992); Olley and Pakes (1992);and Bartelsman and Dhrymes (1994) for aselected number of industries using theLRD. These studies find that the realloca-tion of output from less-productive tomore-productive plants within industriesplays an important role in the observedpatterns of industry-level total factor pro-ductivity (TFP) growth. In the balance ofthis subsection, some of the analysis ofBaily, Hulten, and Campbell (1992) isextended to all manufacturing industries.In addition, the decomposition I use hereprovides a more comprehensive anddetailed examination of the contribution ofwithin-plant, between-plant and net-entrychanges to industry-level TFP.

Following Baily, Hulten, and Camp-


64

MAY/JU N E 1997

bell, plant-level productivity for plant i inperiod t is measured as:

(4) lnTFPit=ln

Qit]αKlnKit

]αLlnLit] αMlnMit,

where Qit is real gross output at plant i inperiod t and Kit, Lit, and Mit are capital,labor and intermediate inputs, respectively.The capital input includes structures andequipment treated separately, and the inter-mediate inputs includes material inputsand energy purchases treated separately.The index of industry-level productivity inyear t used by Baily, Hulten, and Campbellis given by:

(5) lnTFPt =Si

uit lnTFPit ,

where uit it is the share of gross output forplant i in period t for the industry. Themeasure of industry productivity growthbetween periods t] k and t is thenmeasured as:

(6) D lnTFPt = lnTFPt ] lnTFPt -k ,

In what follows, the plant- and industry-level measures for productivity are constru-cted from the LRD for the census years1977, 1982, and 1987. The details of themeasurement of gross output, inputs, andfactor elasticities (measured via cost shares)are discussed in the Appendix but essent-ially follow that of Baily, Hulten, andCampbell (1992). The standard difficultmeasurement issues in constructing mea-sures of real output and inputs (in partic-

ular, for example, the construction of thereal capital stock) that are always confrontedin measuring TFP are amplified in thistype of microeconomic analysis. In addi-tion, the plant-level data are incomplete onsome important dimensions. For example,other than some limited informationcollected on contract workers, the ASMand thus the LRD do not include informa-tion on purchased services.

The decomposition considered herefor a given industry is as follows:20

(7) D lnTFPt = Scontinuers

uit -k D lnTFPit +

Scontinuers

(lnTFPit-k ] lnTFPt -k)Duit +

Scontinuers

D lnTFPitDuit +

Sentering plants

uit(lnTFPit] lnTFPt -k) ]

Sexiting plants

uit -1(lnTFPit-k] lnTFPt -k) .

The first term in this decompositionrepresents a within plant component basedon plant-level changes, weighted by initialoutput shares in the industry. The secondterm represents a between-plant compon-ent that reflects changing output shares,weighted by the deviation of initial plantproductivity from the initial industryindex. The third term represents a covari-ance term. The last two terms representthe contribution of entering and exitingplants, respectively.

In this decomposition, the between-plant term and the entry and exit termsinvolve deviations of plant-level product-ivity from the initial industry index. For acontinuing plant, this implies that anincrease in its output share contributespositively to the between-plant componentonly if the plant has higher productivitythan average initial productivity for theindustry. Similarly, an exiting plant contri-butes positively only if the plant exhibitsproductivity lower than the initial average,and an entering plant contributes posi-

20 This decomposition differs fromthat in Baily, Hulten, andCampbell (1992) in a few sub-tle, but important, respects.The differences and conse-quences are discussed in detailin the Appendix.

Decomposition of Total Factor Productivity Growth*

Census NetPeriod Total Within Between Covariance Entry1977-87 10.73 5.84 -1.11 4.03 1.971977-82 2.43 -0.30 -1.26 3.52 0.431982-87 8.26 4.76 -1.39 3.92 0.96

*Selected periods, percentage increases during the period.SOURCE: Tabulations from LRD, based on decomposition in Equation 7.

Table 3


65

MAY/JU N E 1997

tively only if the plant has higher produc-tivity than the initial average.

This decomposition is undertaken atthe four-digit industry level for 1977-87,1977-82, and 1982-87, using plant-leveldata from the CM. Weighted averages ofthe industry-level decompositions arereported in Table 3. Following Baily,Hulten, and Campbell (1992), the weightsused to aggregate across industries are theindustry share of nominal gross output,averaged over the beginning and endingyears of the period over which the changeis measured.21 Several interesting patternsemerge. First, the within-plant componentis quite important but is far from tellingthe entire story. For example, for theperiod 1977-87, the within-plant compo-nent accounts for about half of the averageindustry change. The between-plant com-ponent is uniformly negative but relativelysmall, while the covariance term is uni-formly positive and large. For the 1977-87period, the covariance term accounts forabout 40 percent of the average industrychange. It is clear from this result that theshift in output towards plants that are alsoincreasing productivity is a major factor inaccounting for the average industry change.Net entry plays an important supportingrole as well. For the 1977-87 period, netentry accounts for about 18 percent of theaverage industry change. Taken together,these results imply that about half of theincrease in productivity for the averageindustry is accounted for by compositioneffects involving the reallocation of outputacross production sites.

The contribution of the various com-ponents varies over time and apparentlythroughout the cycle. The period 1977-82exhibits very modest average productivitygrowth. Interestingly, both the within-plant and the between-plant componentsare negative for this period. The modestincrease in the overall average during thisfive-year horizon is accounted for by a rel-atively large and positive covariancecomponent, as well as a positive net-entrycomponent that offsets the contribution ofthe within- and between-plant compo-nents. In contrast, the period 1982-87

exhibits robust average productivitygrowth, with large positive contributionsfrom the within-, covariance, and net-entry components.

Table 4 provides information aboutsome of the underlying determinants ofthe decomposition by reporting outputshares of entering and exiting plants andthe weighted average of productivity levelsfor continuing, entering and exitingplants.22 The reported productivityindexes are relative to the weightedaverage for all plants in 1977. Enteringplants tend to be smaller than exitingplants, as reflected in the generally smalleroutput shares of entrants (relative toexiting plants). Entering plants in period ttend to have higher productivity than thelevel of productivity in period t2k forexiting and continuing plants, but entrantsexhibit slightly lower productivity thancontinuing plants in period t. Exitingplants from period t2k tend to have lowerproductivity than continuing plants inperiod t2k. Thus, entering plants tend todisplace less-productive exiting plants, butenter with about the same productivity ascontinuing plants.

These results are very much in thespirit of the findings reported by Baily,Hulten, and Campbell (1992); Olley andPakes (1992); and Bartelsman andDhrymes (1994). The message that emergesis that the reallocation of output acrossplants plays a very important role inaccounting for aggregate measures of pro-ductivity growth (specifically, here,

21 It is important to emphasizethat the decomposition is forindustry productivity growthand the weighted averageacross industries does not cap-ture the reallocation of outputbetween industries. Thus, theweighted industry averages arenot directly comparable to over-all changes in productivity fortotal manufacturing. In spiteof this disclaimer, a comparisonwith overall changes in produc-tivity growth in total manufac-turing, based on theBartelsman and Gray (1995)published ASM data, yieldsquite similar overall patterns.

22 As in Table 3, the industry-levelproductivity indexes are weight-ed by the average of the indus-try share in nominal gross out-put in the beginning and end-ing periods.

Output Shares and Relative Productivity

Output Shares Relative Productivity Indexes

Exiting Entering Exiting Entering Contiuing ContiuingCensus Plants Plants Plants Plants Plants PlantsPeriod (t-k) (t) (t-k) (t) (t-k) (t)

1977-87 20.1 15.9 0.99 1.10 1.00 1.111977-82 7.7 7.7 1.01 1.01 1.02 1.031982-87 12.2 8.3 0.99 1.09 1.01 1.11

SOURCE: Tabulations from LRD. TFP indexes are relative to calculated TFP index in1977 for all plants, based on Equation 5.

Table 4

through the reallocation of output towardsestablishments with rising productivity).Furthermore, the relative contribution ofthe reallocation varies through time.Putting these two results together suggeststhat documenting and understanding theprocess of reallocation is important forunderstanding the determinants and thefluctuations in aggregate productivitygrowth.

Putting the Pieces TogetherRecent evidence from studies using

establishment-level data make a primafacie case that aggregate fluctuations inkey aggregates like employment, invest-ment, and productivity can only be under-stood by building from micro evidence.Large and time-varying rates of within-industry job reallocation indicate that microheterogeneity is pervasive and plays animportant role in characterizing under-lying driving forces and in characterizingfluctuations. The bursts of permanent jobdestruction at the onset of recessions areclosely linked to the observed persistencein unemployment rates during the cycle.Nonlinear micro adjustment of labor andcapital inputs, in combination with thisheterogeneity, imply time-varying elastici-ties with respect to aggregate shocks. Theunderlying reallocation also plays a funda-mental role in characterizing aggregateproductivity dynamics. The reallocation ofoutput towards establishments with risingproductivity and the supporting contribu-tion of more-productive entering plantsdisplacing less-productive exiting plantsaccount for about half of the growth inaverage industry productivity in U.S. man-ufacturing during the 1980s. In addition,the contribution of the process of realloca-tion to productivity growth varies overtime, suggesting that understandingfluctuations in aggregate productivitygrowth requires tracking the contributionof reallocation.

Many of the findings thus far should,of course, be viewed as only suggestivebecause they primarily reflect the behaviorin U.S. manufacturing and are plagued bytheir own measurement and conceptual

problems (this is particularly true for theanalysis of productivity). However,progress on these dimensions can only bemade by developing and improving therequisite longitudinal micro databases—the topic to which we now turn.

DATA COLLECTION,PROCESSING, ANDMEASUREMENT

Building the microeconomic databasesrequired to pursue a longitudinal micro-economic approach to measurement andanalysis of aggregate fluctuations is aformidable challenge. My discussion inthis section is cast in terms of the databasecollection, processing, and measurementissues that must be confronted to pursuethis approach, given the current practicesof the U.S. statistical agencies. In addition,the discussion highlights economists’limited understanding of a variety of keyconceptual and measurement issues thatserve as additional obstacles to this approach.

The ideal, of course, is to build a com-prehensive longitudinal establishmentdatabase that would be based on a repre-sentative longitudinally matched sample ofestablishments, including a representativesample of births and deaths. This data setwould have unique, time-invariant esta-blishment identifiers, enabling linking ofestablishments over time, as well as indica-tors-of-ownership structure so that esta-blishments of multi-unit companies couldbe linked together. Variables in the dataset would include detailed informationabout location, establishment age, industry,output, capital, labor, and intermediateinputs (including energy and purchasedservices), as well as detailed informationabout wages and prices. Measurement ofoutput would include a detailed break-down of the products and/or servicesprovided by this establishment. Such adata set could be used for micro studies ofestablishment and firm behavior, as well asfor characterizing the connection betweenmicro dynamics and aggregate fluctuationsalong the lines discussed earlier in the sec-

23 I have neglected many of theapplications of longitudinalbusiness data to a variety ofother topics and questions.See McGuckin (1995) for anoverview of the type of analy-sis that has and can be donewith longitudinal micro data onthe business population.Bartelsman (1995) provides arelated discussion that also con-siders related findings for theNetherlands.

MAY/JU N E 1997

FE D E R A L R E S E RV E BA N K O F S T. L O U I S

66

tion entitled, “Micro Heterogeneity andAggregate Fluctuations: A Brief Review ofRecent Evidence.”23

Is it possible under current practicesto build anything remotely resembling thiswishful fiction? Building this type of data-base requires a comprehensive, integratedapproach to the collection of statistics onthe U.S. business population.24 As I indi-cated at the beginning of this article,various federal agencies conduct differentsurveys and censuses to collect currentinformation on the U.S. business popula-tion. The various pieces of informationare matched at an aggregate (e.g., indu-stry) level.

It is instructive in this regard toconsider briefly the various data sourcesdrawn on to produce published industryreal output and productivity (i.e., laborand total factor) indices. Although it iswell beyond the scope of this article toprovide a complete characterization of thedata sources and procedures the statisticalagencies use to measure macro aggregates,25

even a crude characterization reveals thenature of the process and the limitations interms of building micro databases.

One key ingredient in buildingpublished aggregate statistics are the eco-nomic censuses taken every five years bythe Bureau of the Census. The economiccensuses collect data from the universe ofall establishments, covering the manufac-turing, wholesale trade, retail trade,service, construction, agriculture, transpor-tation, and mineral industries. The primarydata collected are payroll and nominalgross revenue. For many sectors, informa-tion on employment, intermediate inputs,and asset expenditures are also collected,although the level of detail varies dramati-cally across sectors. This information fromthe economic censuses on gross revenue,intermediate input expenses, and assetexpenditures are vital for the Bureau ofEconomic Analysis (BEA) to build itsindustry input-output tables. These infre-quently revised tables (the latest input-output table now available is from 1987)are used by BEA, along with annualindustry-level tabulations of gross output,

to construct measures of sectorial value-added. The annual industry-level tabu-lations of gross output are based on theannual surveys conducted by the CensusBureau, which collects information ongross revenue, as well as information fromother sources for sectors with inadequateannual surveys. Constructing sectorialreal gross output measures, as well as realvalue-added, requires the generation ofboth output and intermediate input defla-tors. The Bureau of Labor Statistics (BLS)collects the output deflators separately.The intermediate input deflators arederived from the output deflators and theinput-output matrixes. The sectorial realvalue-added measures that emerge are thecore of the product side of the grossdomestic product (GDP) accounts and inturn are used for a variety of other purposes,including generating labor and total factorproductivity statistics (e.g., those producedby the BLS).26

In terms of the measurement of otherinputs for officially published productivitystatistics, the employment, hours, and pay-roll information is based on the BLS esta-blishment survey (the 790 data), which inturn is benchmarked to the ES-202 data,based on state unemployment insuranceadministrative data. Since hours data inthe BLS 790 data are restricted to nonsuper-visory workers, hours data are further sup-plemented by information from theCurrent Population Survey (CPS).

Investment expenditures by industryare generated by using the shipments,imports, and exports of capital goods fromthe annual surveys conducted by theCensus Bureau and other sources, alongwith the input-output tables. Construc-tion of capital stocks are generated usingperpetual inventory methods that in turnrequires investment deflators and deprecia-tion rates. Because deflators and depreci-ation rates vary widely across asset types,the measurement of real expenditures andstocks by industry requires detailed infor-mation on asset expenditures by industry.Historically, no data have been collectedon detailed asset by industry (rather, atbest, information distinguishing between

24 This call for a comprehensive,integrated approach to the col-lection of business populationstatistics is not new. (See, forexample, Griliches, 1994; theBonnen report, 1981; andTriplett, 1991, for related dis-cussion and additional refer-ences).

25 See, for example, Carson(1987), for an excellentoverview of the data andsources used to generate theNational Income and ProductAccounts.

26 BLS uses different interpolationand extrapolation proceduresbetween economic censuses sothat BLS measures of sectorialproductivity differ from thosegenerated from the BEA proce-dures.

MAY/JU N E 1997


67

27 This combination of imputa-tion, interpolation, and extrapo-lation procedures raise ques-tions about the use of thesecapital stocks by industry formeasurement and research pur-poses. For example, theseprocedures raise particularlyserious questions about usingthese capital stocks by industryto investigate hypothesesregarding issues such as capi-tal-skill or technology-skill com-plementarity.

28 One source of incompletenessin this discussion is the neglectof higher frequency sources ofdata for monthly and quarterlyaggregate statistics.

29 Some company-based surveys(e.g., the new Annual CapitalExpenditures Survey) ask com-panies to provide informationabout their activities in sepa-rate two-digit or three-digitindustries.

30 “Well-known” is in quotesbecause, although many ofthese issues are well-known inthe statistical community, theyare not generally well-knownby research economists usingthe data.

MAY/JU N E 1997


68

structures and equipment has been collec-ted). The BEA capital-flows tables used toallocate detailed asset types acrossindustries are based on auxiliaryinformation. For example, in describingthe release of a new Capital Flows Table(CFT) in 1985, Silverstein (1985) notesthat “most distributions were made in pro-portion to some indicator, such asemployment, assumed to be correlated withthe use of the commodity.” Like the rest ofthe input-output tables, the construction ofthe CFT is an arduous, time-consumingprocess. For example, Silverstein (1985)reports that the CFT released in 1985 wasbased on the 1977 input-output accounts(from the 1977 Economic Censuses). Thetypical indicator used to allocate assets toindustries in the CFT released in 1985 isthe occupational mix of employment byindustry, based on the occupation-by-industry report from the 1970 Census ofPopulation, extrapolated to 1977. Thus,real capital expenditures and stocks byindustry after 1985 are based in part oninformation collected many years earlierand on assumptions of some fixed relation-ship between asset use by industry and theoccupational mix of employment byindustry.27

This depiction is incomplete. Never-theless, it causes one to imagine that aggre-gate statistics emerge from some great blackcauldron, mixed together with data from analphabet soup of surveys.28 A host of well-known problems arise with this merging ofdiverse sources of data (collected by avariety of different agencies) havingdifferent sample frames and consequentlydifferent properties. The level of detail anddisaggregation varies substantially acrosssectors so that strong assumptions are nec-essary to match the data across sources. Ina similar manner, the information collectedis in many cases incomplete. Heroicassumptions underlying imputation proce-dures, along with matching data fromdifferent sources, are therefore necessary toconstruct the measure of interest (e.g., themeasurement of real capital stocks andexpenditures, using the CFT). Anotherproblem is that the benchmark detailed

estimates are available only at five-yearintervals and often with a substantial lag sothat the higher frequency (e.g., annual,quarterly, and monthly) data involvesubstantial interpolation and extrapolation.

A related problem in matching the dataacross sources is that some of the under-lying data are from sample frames at theestablishment level while others are at acompany or employer taxpayer-identifi-cation level. This distinction matters interms of industrial classification of the rele-vant outputs and inputs. Many of thelargest companies have multiple establish-ments, operating in a variety of industries(crossing two-digit and one-digit bound-aries). The sectorial classification of activitywill differ in important ways, depending onhow the activities of such large companiesare allocated across industries. Establish-ment-based surveys tend to classify all ofthe activity in a given establishment in asingle detailed (e.g., four-digit) industrywhile company-based surveys classify allactivity for an entire company in a single,broader (e.g., two- or three-digit) industrybased on the major activity of the company.29

This implies that, even under a commonindustrial classification system, company-based and establishment-based surveysyield different pictures of the industrialcomposition of activity. Matching informa-tion across such sources at the industrylevel has obvious problems.

The objective here is not simply to reit-erate the “well-known” problems associ-ated with building existing aggregate statis-tics from myriad sources, but to considerthe formidable challenge of what would berequired to take heterogeneity and aggrega-tion issues seriously in the measurementand analysis of aggregate fluctuations.30

The discussion in the remainder of this sec-tion accordingly proceeds as follows. First,I review some problems with the manufac-turing sector (having arguably the bestlongitudinal micro data) since this discus-sion is instructive for the more generalproblems that must be confronted inbuilding a comprehensive longitudinaldatabase. Second, I discuss the problemswith matching data from myriad sources,

both cross sectionally and longitudinally atthe micro level. Third, I discuss theconceptual challenges in building new lon-gitudinally based aggregate statistics.Finally, I conclude with some briefremarks that contrast the concerns raisedhere about measurement with the manyother vital measurement issues weconfront in analyzing aggregatefluctuations.

Problems With the Data on Businesses

The sector with arguably the bestannual data collected in a manner suitablefor longitudinally based analysis is manu-facturing. Detailed data come from theASM and the CM. Not only do the ASMand CM contain a wealth of informationabout individual establishments, the five-year panel rotation of the ASM and thecomprehensive CM provide a means oflinking the data longitudinally. As high-lighted by the discussion of the studies inthe above section entitled, “Micro Hetero-geneity and Aggregate Fluctuations: ABrief Review of Decent Evidence,” the lon-gitudinally based data that have resulted(i.e., the LRD) offer the opportunity tostudy the dynamics of employment, wages,investment, and productivity.31 However,even here, there are substantial limitationsthat in many respects are becoming moresevere with time. Oddly enough, some ofthe growing limitations of the ASM andCM are being generated by well-intentionedefforts to improve the overall quality ofeconomic statistics, along with concerns ofreducing the reporting burden on compa-nies. Discussing these limitations isillustrative of the problems that must beovercome in building longitudinal establish-ment databases.

A number of key variables have beeneliminated from the ASM through theyears. In 1980, the series on quarterlyhours for production workers was drop-ped. After 1988, much of the detail oncapital stocks and capital expenditures wasdropped. Before 1988, the ASM includedbeginning- and end-of-period book values

for equipment and structures, new expen-ditures on equipment and structures, usedexpenditures, and retirements. Since1988, only new expenditures on equip-ment and structures have been included.

The motivation for these deletions istypically based on the argument that theseries in question duplicates informationthat some other survey (perhaps by someother statistical agency) collects and iseliminated to reduce costs and reportingburdens. For example, the hours dataused in the National Income and ProductAccounts and BLS productivity tabulationsare derived from the BLS 790 and the CPS.Since ASM hours data were not a vital partof creating published aggregates, they weredeemed expendable. Likewise, detailedASM data on capital stocks and expend-itures were deemed expendable, given themanner in which the BEA capital stocksand expenditures are constructed andgiven the recently initiated Annual CapitalExpenditures Survey (ACES).32

These deletions, however, severelylimit the ASM and thus the LRD as asource of longitudinal establishment datafor the 1990s and beyond. In terms of theempirical studies discussed in the abovesection on micro heterogeneity and aggre-gate fluctuations, these changes severelylimit the ability to conduct futureanalogous types of analyses. For example,the analysis on nonlinear adjustment ofemployment is based on a sampleterminated in 1980, given the eliminationof the quarterly hours series. Likewise, thestudies of plant-level investment dynamicsdescribed in the section on microheterogeneity and aggregate fluctuationsused samples terminating in 1988, giventhe elimination of the detail onexpenditures, retirements, and bookvalues. In general, investment andproductivity studies using the LRD wereseverely hampered after 1988, given theelimination of the detail on capital stocksand expenditures.

Although the information contained inthe ASM is deteriorating over time, it isimportant to emphasize in this contextthat the ASM still contains a vast amount

31 Furthermore, the manufactur-ing sector comes relativelyclose at times to the “plug &play” approach advocated inthe section entitled “Can theExisting Microeconomic DataBe Linked Across Sources andTime” in that many of thespecial surveys can be linked atthe micro level. For example,detailed data on the type oftechnology adopted at plantshave been collected in theSurvey of ManufacturingTechnology (SMT). A varietyof recent studies (see, e.g.,Doms, Dunne, and Troske,1997) have matched themicro data on technology useto the CM micro data.

32 ACES has some quite attractivefeatures in that it covers allsectors and periodically willprovide detailed asset expendi-tures information to provideneeded information on the mixof assets being accumulated.


69

MAY/JU N E 1997

of information relative to the datacollected for many nonmanufacturing sec-tors (e.g., services). The annual surveysconducted by the Bureau of the Census fornonmanufacturing sectors often arelimited to collecting information on grossrevenue, employment, and payroll. Somesurveys for individual sectors contain lim-ited information on expenses but eventhen the expenses are often combinedunder one item (e.g., total operatingexpenses). The sampling unit is typicallynot only an establishment but rather amixture of establishments, companies, andbusiness units identified by taxpayeremployer identification numbers (EINs).Note that companies may use multipleEINs. Furthermore, the annual surveysoutside of manufacturing do not have thefive-year panel-rotation feature of the ASMthat permits longitudinal analysis.

Can the Micro Data Be Linked?Given the manner in which business

population data are currently collected inthe United States, one obvious question is,can the various micro data sets at the sta-tistical agencies be appropriately linked atthe micro level? In principle, there is noreason why all the ingredients for longitu-dinal establishment-based statistics andanalyses of employment, investment, orproductivity need to be collected in asingle survey or by a single statisticalagency. However, as emphasized above,the problem is that the collection of infor-mation from the U.S. business populationis not based on a comprehensive,integrated approach. Two related issuesmust be considered in this context: First,can the micro data be matched cross-sectionally across sources? Second, canthe micro data be matched longitudinally?Several examples illustrate the problem oflinking the data on both of thesedimensions at the micro level.

As a first example, consider the possi-bility of using ACES as the source of microdata on investment and using the CensusBureau’s annual surveys of various sectorsas the source of information for shipments

or revenue (e.g., to measure gross output).For concreteness, it is illustrative toconsider specifically the problems ofmatching up the micro data from ACESand the ASM. An immediate problem isthat ACES is a company-based survey,while the ASM is an establishment-basedsurvey. Although the ASM includes com-pany identifiers, it is a sample of establish-ments and thus not all establishments ofmulti-unit establishment companies areincluded. This implies that one could notsuccessfully aggregate the ASM to acompany level and match to ACES.Another problem is the nature of the panelrotation. The ASM is drawn every fiveyears, with a representative sample ofbirths added each year to the ongoing five-year panels. A new ACES sample is drawnevery year. Although large certainty estab-lishments (ASM) and certainty companies(ACES) are included in every survey, smallbusiness units (either companies or estab-lishments) cannot be linked across panelrotations. Because a new ACES sample isdrawn every year, it cannot be used to gen-erate a representative matched panel ofcompanies across years, much less used tomatch longitudinally to the LRD or ASM.

For another example, consider the BLS790 establishment survey. It containsmonthly information on hours, employ-ment, and payroll on an establishmentbasis but huge obstacles arise that wouldhave to be confronted to link it up to thevarious micro data sets that contain infor-mation on shipments, other inputs, andcapital expenditures at the Census Bureau.One serious obstacle is that current confi-dentiality laws and restrictions severelylimit statistical agencies’ data sharing atthe micro level. The confidentialityprotections are essential since it is impera-tive to protect the confidentiality ofrespondents data from households andcompanies. Each of the statistical agenciesis committed to providing such protectionas reflected in the current set of similarrestrictions that each agency has in place.What is needed is for micro data at all ofthe federal statistical agencies to beprotected under a common set of restric-

33 Some recent proposed legisla-tion (H.R. 3924) would put allstatistical agencies under thesame confidentiality restrictionsand thus permit data sharingfor statistical purposes acrossthe statistical agencies.

MAY/JU N E 1997


70

tions so that data can be shared across thestatistical agencies.33 Even after over-coming these monumental legal problems,the linkage would face serious obstacles.The BLS 790 data is an establishmentsurvey but is not based on a representativesample. Furthermore, the BLS 790 surveyis voluntary, so many longitudinal holesexist in the micro data. BLS has a variety ofprocedures to overcome these limitationsin building aggregates from the 790 surveybut these limitations restrict the usefulnessof the 790 as a micro database.34

What needs to be done to develop amore comprehensive, integrated approachto the collection of statistics by the U.S.business population? It is beyond thescope of this article to explore thestatistical agencies’ organizational changesneeded to achieve a comprehensive,integrated approach to the collection ofbusiness population statistics. One avenuefor achieving such objectives is to create acentral statistical agency for the UnitedStates. It is worth noting that both Canadaand France have such central statisticalagencies and have surpassed the UnitedStates in terms of the development of lon-gitudinal business databases. Indeed, inFrance, a longitudinal database withmatched employers and employees hasbeen developed that yields a host of addi-tional possibilities beyond those discussedabove (see, e.g., Kramarz, 1994). Even ifcreating a central statistical agency is notfeasible, the creation of a virtual centralagency through close coordination anddata sharing is essential to collect businesspopulation statistics in a comprehensiveand integrated fashion. This call for coor-dination across statistical agencies is farfrom new, but one argument made byTriplett (1991) and McGuckin (1995) isworth repeating in this context. Boththese authors argue forcefully that thecoordination across the statistical agenciesis only possible through maintaining andbuilding the capability for research andanalysis at the statistical agencies. Theirbasic argument is that research capabilityat the statistical agencies is essential forproviding the necessary links between the

producers of the data and the users of thedata. The voice of new ideas and researchin the program planning process is crucialin the current context. My main argumenthere is that new research with longitudinalbusiness population data points towards aneed for rethinking the manner in whichwe collect and process data for producingaggregate statistics. This rethinking isonly possible if this new research has avoice at the statistical agencies, with influ-ence on the operational collection andprocessing of statistics.

Beyond the organizational changesthat may be required, one key is to use acommon master business establishmentlist and to follow a “plug & play” approachto the collection of business populationstatistics. If all business population statis-tical surveys are establishment based anddrawn from a common frame that main-tains consistent establishment, company,and industry identifiers, then the data canbe matched at the micro level. The mostdesirable approach is to keep all surveys atthe establishment level since mixing datafrom establishment and company surveysgenerates the type of problem discussedabove in matching ACES to ASM data.35

Another essential aspect to a successful“plug & play” approach to building longi-tudinal business data is to ensure that therotation of establishments in an individualsurvey over time is such that it permitscreating a representative sample of longitu-dinally matched plants (along with arepresentative sample of establishmentbirths and deaths). The frequency of panelrotation of the establishments in surveys isalso important to consider in this contextbecause it will affect the frequency atwhich longitudinal analysis can be conduc-ted. To the extent that the data are collec-ted by different surveys for the same sectorthat will be matched cross-sectionally,appropriate coordination of the panel rota-tion is required.

In spite of the somewhat pessimistictone I have taken above, the United Statesis not that far away from achieving someaspects of this comprehensive, integratedapproach for U.S. business population sta-

34 BLS uses a ratio-of-change esti-mation procedure to overcomethese problems in generatingaggregates from the survey.Furthermore, the BLS bench-marks some of the tabulations(e.g., employment) from the790 to its master businessestablishment data file (the ES-202 data discussed later in thearticle).

35 It may be, however, that insome industries it is very diffi-cult to collect information atthe establishment level.Furthermore, some data itemsare inherently company-levelvariables (e.g., financial assetsand liabilities) or are difficult tocollect at the establishmentlevel (e.g., exports). Thesedifficulties do not negate theimportance of measuring cer-tain types of behavior at theestablishment level (e.g.,employment growth, invest-ment, and productivity growth)but rather highlight the need tolink relevant establishment-level behavior with relevantcompany-level variables (e.g.,financial variables andexports).

36 One question is why bothagencies are maintaining busi-ness-establishment lists. Asdiscussed above, legal data-sharing restrictions lurk at theroot of this redundancy.


71

MAY/JU N E 1997

tistics. The Census Bureau and the BLSnow both maintain master business estab-lishment lists.36 The Standard StatisticalEstablishment List (SSEL) maintained bythe Bureau of the Census is a master busi-ness list based on administrative data andsurvey sources that cover all employerswith at least one employee. The businessunit in the SSEL is typically an establish-ment, and the file contains establishment,enterprise, location, and industry identi-fiers, as well as basic information on pay-roll and first quarter employment fromInternal Revenue Service files. The SSELis the frame from which the CensusBureau conducts the economic censuses,as well as its various annual surveys. Fur-thermore, it is through the SSEL, theannual Company Organization Survey, andthe economic censuses that the CensusBureau tries to maintain accurateestablishment and firm identifiers. TheBLS maintains a similar administrativerecord file (the ES-202 data), based onUnemployment Insur- ance files. The ES-202 data are the frame underlying the BLSsurveys of establishments. Furthermore,the BLS conducts a multi-site workplacesurvey to track the ownership linkagesbetween establishments.

The SSEL and the ES-202 canpotentially serve as the core on which tobuild comprehensive, longitudinal estab-lishment databases. Indeed, severalongoing projects use both of these datasets on a limited basis. For example, plansare under way to use the ES-202 data toproduce annual and quarterly gross job-creation and job-destruction statistics byindustry, state, size class, and wage classfor the entire U.S. economy. This ispossible with the ES-202 data withoutlinking it to any other information becausethe ES-202 data already contain employ-ment data.37 Thus, the administrativerecord databases currently used for sampleframes should be viewed as primary datasources for longitudinally based statistics.

Using either the SSEL or ES-202 as alongitudinal frame or database requiressubstantial effort. Both are based on taxrecords and the core identifiers in the files

are thus tax identification numbers (EINs)and/or Unemployment Insurance (UI)account numbers. A variety of ownershipand organizational structure changes yieldchanges in EIN and UI account numbers.Furthermore, multi-unit firms may main-tain one or several tax identificationnumbers and again may change these inresponse to a variety of circumstances.Longitudinal linkage problems in usingthese files are substantial. As an example,in constructing the LRD during the 1972-88 period, staff and researchers at theCenter for Economic Studies at the Bureauof the Census detected more than 5,000longitudinal linkage problems in the ASMeven after the Census Bureau processingprocedures had assigned permanent plantidentifiers (PPNs) that were supposed tobe permanently affixed to a particular loca-tion. Pursuing these longitudinal linkageproblems outside of manufacturing is amonumental task in light of the more than7 million establishments with more thanone employee.38 Nevertheless, recentefforts at both the Census Bureau and BLSindicate that these longitudinal linkageproblems can be overcome.39

To summarize briefly, collecting business population data in a comprehen-sive and integrated manner is essential for building the type of longitudinal businessdatabases necessary for understanding the underlying driving forces for keymacro aggregates like employment,investment, and productivity growth. The path towards a more comprehensive, inte-grated approach involves the following steps.

• Permit data sharing across the statis-tical agencies and thus create acommon master business establish-ment list.

• Exploit the information in the admin-istrative data (underlying the businessestablishment lists) to the fullestextent to build longitudinally basedstatistics.

• Base censuses and draw samples for allsurveys of businesses from the masterbusiness list so that micro data and

37 Similarly, the SSEL has alreadybeen used to generate statisticson the changes in the distribu-tion of employment growth bycompany and establishmentsize in a contract with theSmall Business Administration(see, e.g., Trager and Moore,1995, for more discussion).Furthermore, linking the microrecords in the SSEL, economiccensuses, and annual surveyswould permit analysis of thejoint distribution of employ-ment growth and labor-produc-tivity growth (measured usinggross output measures) at theestablishment level.

38 Furthermore, greater conceptu-al problems arise outside ofmanufacturing in specifyingwhat one means by an ongo-ing establishment. In principle,the Census Bureau assigns aPPN that reflects ongoing activi-ty at some fixed, physical loca-tion. In the retail trade andservice sectors, it may be inap-propriate over the course oftime to link a particular retaillocation that houses a varietyof different retailers selling awide variety of products andservices.

39 See, for example, Spletzer(1995) and Trager and Moore(1995).

MAY/JU N E 1997


72

associated aggregate statistics can belinked.

• Introduce appropriate panel rotationin the samples of businesses in thesurveys to permit construction of lon-gitudinal statistics from survey data.

Although these steps would be extremelyuseful for many statistical and research pur-poses (including those advocated here), evenlarger payoffs await a more fundamentalchange in the manner in which businessstatistics are collected. The myriad surveysand censuses of businesses conducted by thedifferent statistical agencies impose aheavy burden on respondents and yield ahost of problems in linking the data at themicro and the aggregate level for the usersof the statistics. Finding some effectivemeans of streamlining this process so thatall the data from an individual business fora given period is collected at one timewould yield tremendous benefits to bothrespondents and data users.

Parsimonious Ways to Summarizethe Micro Hetergeneity

In addition to the problems of buildingthe requisite longitudinal business databases,this approach to aggregate analysis has avariety of other measurement and conceptualproblems that need to be confronted. Onegeneral problem is that parsimonious waysof summarizing and aggregating the relevantinformation have yet to be developed. Even ifall the logistical obstacles discussed aboveare overcome, it is unlikely that longitudinalbusiness databases created at the statisticalagencies will ever be widely accessible to theresearch and policymaking communities.Confidentiality restrictions inherently implylimited access to the micro data with associ-ated monitoring to prevent inadvertentdisclosure.40 Furthermore, even with theincreasing speed and disk capacity ofcomputers, the underlying micro databasesare enormous.

The question then is whether newaggregate measures of the distribution ofthe micro changes and activity can be

developed that would prove useful foranalysis of aggregate fluctuations. Grossjob-creation and job-destruction rates arean example of new aggregates that can begenerated from longitudinal micro data.Furthermore, one could easily imaginethat other basic, descriptive decomposi-tions like those used in Equation 7 todecompose industry productivity growthcould prove quite useful. Knowing the rel-ative contribution of within-establishment,between-establishment, covariance, andnet-entry components of aggregatemeasures of productivity growth wouldundoubtedly shed considerable light onthe determinants of the aggregates.

Because a basic insight of thisapproach is that higher moments matterfor aggregate fluctuations, it would alsoappear at first glance that the studies onnonlinear micro adjustment and aggregatefluctuations yield promising suggestionsfor aggregates that could be created. Theproblem here is that the higher momentsthat matter involve the difference betweendesired and actual variables. For example,the measurement of the desired capitalstock is model dependent with a numberof reasonable alternative specifications.This model dependence is apt to be ageneric issue in considering the linkbetween micro and macro behavior. How-ever, even without consensus on specifi-cation, the analysis in the section on microheterogeneity and aggregate fluctuationsprovides some suggestions of measuresthat might be useful. For example, theanalysis of the aggregate implications ofnonlinear micro adjustment for employ-ment described in the same sectiondepends critically on measuring the distri-bution of establishment deviationsbetween actual and “normal” hours perworker at the establishment. The relatedanalysis of nonlinear micro adjustmentand aggregate investment dynamicsdepends critically on the distribution ofestablishment output-capital ratios. Bothof these cases yield relatively simple andintuitive suggestions of potentially relevantmeasures of the distributions of activity atthe establishment level. The challenge is,

40 The opening of a CensusResearch Data Center (RDC) inBoston and at Carnegie-MellonUniversity are one importantway these limitations on dataaccess can be overcome. TheCensus RDC’s permitresearchers to work with themicro data without coming toCensus-headquarters inWashington which maintainingstrict monitoring of theresearch output to prevent dis-closure of the confidentialmicro data. A Census Bureauemployee is detailed to eachRDC to provide the necessaryliaison between the researchersand the Census Bureau.

MAY/JU N E 1997


73

in general, to find creative ways to summa-rize the relevant microeconomicdistributions of activity and change in amanner that can be widely used (i.e., notidiosyncratically linked to a narrowly spec-ified model or functional form).

CONCLUSIONLet me close by briefly comparing and

contrasting these concerns about hetero-geneity and aggregation issues with thehost of other measurement problems con-fronting our measurement of real activity.The conceptual issues (e.g., how to measureoutput in certain service industries), aswell as the associated limited data collectedin the service sector, are first-order problems.Dealing with quality change and new goodsand services are perennial problems in themeasurement of output, inputs, and prices.Given limited budgets for statistical agencies,changes in data collection and processingprocedures that address these issues deservehigh priority. However, my view is thatimplementing the data and processing pro-cedures required for addressing theseheterogeneity and aggregation issuesshould be on the list of priorities as well.Addressing these heterogeneity and aggre-gation issues is in many ways complemen-tary with addressing the other measurementdifficulties that we face. After all, trying tomeasure investment, real output, and pro-ductivity growth at the establishment levelforces consideration of the conceptual issuesand the limited data availability problemsat their most basic and primitive level.Furthermore, the longitudinal microeconomicapproach to the production of aggregatestatistics advocated in this article implies acomprehensive, integrated approach to thecollection and processing of statistics thathas many benefits and is long overdue.

REFERENCESAnderson, Patricia M., and Bruce D. Meyer. “The Extent and

Consequences of Job Turnover,” Brookings Papers on EconomicActivity: Microeconomics, 1994, pp. 177-236.

Andolfatto, David, and Glenn M. MacDonald. “Endogenous

Technological Change, Growth, and Aggregate Fluctuations,” unpub-lished manuscript, University of Rochester, 1993.

Baily, Martin N., Eric J. Bartelsman, and John Haltiwanger. “LaborProductivity: Structural Change and Cyclical Dynamics,” NBERWorking Paper no. 5503, March 1996.

_____, Charles Hulten, and David Campbell. “Productivity Dynamicsin Manufacturing Plants,” Brookings Papers on Economic Activity:Microeconomics, 1992, pp. 187-249.

Bartelsman, Eric. J. “Using Microdata for Economic Research and PolicyAnalysis,” unpublished manuscript, Central Planning Bureau,Netherlands, 1995.

_____ and Wayne Gray. “The NBER Manufacturing ProductivityDatabase,” unpublished manuscript, Board of Governors of theFederal Reserve System, 1995.

_____, Eric. J., and Phoebus J. Dhrymes. “Productivity Dynamics:U.S. Manufacturing Plants, 1972-86,” Board of Governors of theFederal Reserve System Finance and Economics Discussion Series, No.93-43, January 1994.

Bils, Mark. “The Cyclical Behavior of Marginal Cost and Price,”American Economic Review (vol. 77, no. 5, 1987), pp. 838-55.

Blanchard, Olivier, and Peter Diamond. “The Beveridge Curve,”Brookings Papers on Economic Activity (1989:1), pp. 1-60.

_____ and Peter Diamond. “The Cyclical Behavior of Gross Flows ofU. S. Workers,” Brookings Papers on Economic Activity (1990:2),pp. 85-143.

Bonnen, James. “Improving the Federal Statistical System: Issues andOptions,” Statistical Reporter, 1981, pp. 133-221.

Caballero, Ricardo. “A Fallacy of Composition,” The American EconomicReview (vol. 82, no. 5, 1992), pp. 1279-92.

_____ and Eduardo M.R.A. Engel. “Beyond the Partial-AdjustmentModel,” The American Economic Review (vol. 82, no. 2, 1992), pp.360-64.

_____ and _____. “Microeconomic Adjustment Hazards andAggregate Dynamics,” Quarterly Journal of Economics (vol. 108, no.2, 1993), pp. 359- 84.

_____ and _____. “Explaining Investment Dynamics in U.S.Manufacturing: A Generalized (S,s) Approach.” NBER working paperno. 4887, October 1994.

_____, _____, and John C. Haltiwanger. “Aggregate EmploymentDynamics: Building from Microeconomic Evidence,” The AmericanEconomic Review (March 1997), pp.115-37.

_____, _____, and _____. “Plant Level Adjustment andAggregate Investment Dynamics,” Brookings Papers on EconomicActivity (1995:2), pp. 1-39.

_____ and Mohamad L. Hammour. “The Cleansing Effects ofRecessions,” The American Economic Review (Vol. 84, 5, 1994), pp.1350-68.

Campbell, Jeffrey R. “Entry, Exit, Technology, and Business Cycles,”

MAY/JU N E 1997


74

Rochester Center for Economic Research Working Paper No. 407,1995.

_____ and Kenneth Kuttner. “Macroeconomic Effects of EmploymentReallocation,” Carnegie-Rochester Series on Public Policy, forthcom-ing.

Carson, Carol S. “GNP: An Overview of Source Data and EstimatingMethods,” Survey of Current Business (July 1987), pp. 103-26.

Cooley, Thomas, Jeremy Greenwood, and Mehmet Yorukoglu. “TheReplacement Problem,” unpublished manuscript, University ofRochester, 1994.

Cooper, Russell, John Haltiwanger, and Laura Power. “MachineReplacement and the Business Cycle: Lumps and Bumps,” NBERWorking Paper no. 5260, September 1995.

Davis, Steven J., and John Haltiwanger. “Gross Job Creation andDestruction: Microeconomic Evidence and MacroeconomicImplications,” NBER Macroeconomics Annual, 1990, pp. 123-68.

_____ and _____. “Driving Forces and Employment Fluctuations:New Evidence and Alternative Fluctuations,” unpublished manuscript,University of Maryland, 1996.

_____, _____, and Scott Schuh. Job Creation and Destruction,MIT Press, 1996.

Doms, Mark, and Timothy Dunne. “Capital Adjustment Patterns inManufacturing Plants,” unpublished manuscript, Center for EconomicStudies, U.S. Bureau of the Census, 1994.

_____, _____, and Kenneth Troske. “Workers, Wages, andTechnology,” Quarterly Journal of Economics (February 1997), pp.253-90.

Ericson, Richard, and Ariel Pakes. “Markov-Perfect Industry Dynamics: AFramework for Empirical Work,” Review of Economic Studies (62, 1,1995), pp. 53-82.

Farber, Henry S. “Are Lifetime Jobs Disappearing? Job Duration in theUnited States: 1973- 1993,” Industrial Relations Section, PrincetonUniversity working paper no. 341, January 1995.

Foote, Christopher. “Trend Employment Growth and the Bunching ofJob Creation and Destruction,” unpublished manuscript, University ofMichigan, 1995.

Griliches, Zvi. “Productivity, R&D, and the Data Constraint,” TheAmerican Economic Review (vol. 84, no.1, 1994), pp. 1-23.

Gordon, Robert J. “Problems in the Measurement and Performance ofService-Sector Productivity in the United States,” NBER WorkingPaper, no. 5519, March 1996.

Hall, Robert E. “Labor Demand, Labor Supply, and EmploymentVolatility,” NBER Macroeconomics Annual 1991, pp. 17-47.

_____. “Lost Jobs,” Brookings Papers on Economic Activity(1995:1), pp. 221-56.

Hamermesh, Daniel S. Labor Demand, Princeton University Press,

1993.

Jovanovic, Boyan. “Selection and the Evolution of Industry,”Econometrica (vol. 50, no. 3, 1982), pp. 649-70.

_____ and Glenn M. MacDonald. “Competitive Diffusion,” Journal ofPolitical Economy (vol. 102, no. 1, 1994), pp. 24-52.

Kramarz, Francis. “The French Experience on Unbalanced Panels ofFirms and on Matched Employer-Employee Panels,” unpublished manuscript, INSEE, 1994.

Lucas, Robert E. Jr. “Understanding Business Cycles,” Journal ofMonetary Economics, Supplementary Series 1977, pp. 7-29

McGuckin, Robert H. “Establishment Microdata for Economic Researchand Policy Analysis: Looking Beyond the Aggregates,” Journal ofBusiness and Economic Statistics (vol. 13, no. 1, 1995), pp. 121-6.

Mortensen, Dale T., and Christopher A. Pissarides. “Job Creation andJob Destruction in the Theory of Unemployment,” Review ofEconomic Studies (61, 3, 1994), pp. 397-415.

Olley, Steven G., and Ariel Pakes. “The Dynamics of Productivity in theTelecommunications Equipment Industry,” NBER working paper, no.3977, January 1992.

Silverstein, Gerald. “New Structures and Equipment by Using Industry,1977,” Survey of Current Business, November 1985, pp. 26-35.

Solow, Robert M. “Investment and Technical Progress,” in Symposiumon Mathematical Methods in the Social Sciences: 1959, KennethArrow, Samuel Karlin, and Paul Suppes, eds., Stanford UniversityPress, 1960.

Spletzer, James. “The Contribution of Establishment Births and Deathsto Employment Growth,” unpublished manuscript, Bureau of LaborStatistics, 1995.

Trager, Mitchell L., and Richard A. Moore. “Development of aLongitudinally-Linked Establishment Based Register, March 1993through April 1995,” U.S. Bureau of the Census Working Paper,1995.

Triplett, Jack E. “The Federal Statistical System’s Response to EmergingData Needs,” Journal of Economic and Social Measurement (vol. 17,no. 3-4, 1991), pp. 155-77.

May/June 1997

Federa l Reser ve Bank o f S t . Lou i s

75

MEASURING PLANT-LEVELTOTAL FACTOR PRODUC-TIVITY AND DECOMPOSINGTHE AGGREGATE

Measurement Issues

The Census of Manufactures (CM)plant-level data used in the analysis in thesection on productivity dynamics containsinformation on shipments, inventories,book values of equipment and structures,employment of production and nonpro-duction workers, total hours of productionworkers, and cost of materials and energyusage. For the most part (exceptionsnoted below), the measurement method-ology used in the section on productivitydynamics follows closely that of Baily,Hulten, and Campbell (1992). Real grossoutput is measured as shipments adjustedfor inventories, deflated by the four-digitoutput deflator for the industry in whichthe plant is classified. All output andmaterials deflators used are from the four-digit Bartelsman and Gray (1995) data.Labor input is measured by total hours forproduction workers, multiplied by theratio of total payroll for all workers pluspayments for contract work to payroll forproduction workers. This latter multipli-cation factor acts as a means for account-ing for both hours of nonproduction andcontract workers. Materials input is mea-sured as the cost of materials deflated bythe Gray-Bartelsman materials deflator.Capital stocks for equipment andstructures are measured from the bookvalues deflated by capital stock deflators(where the latter is measured as the ratioof the current dollar book value to theconstant dollar value for the two-digitindustry). Energy input is measured as thecost of energy usage, deflated by the Gray-Bartelsman energy-price deflator. Thefactor elasticities are measured as theindustry average cost shares, averaged overthe beginning and ending year of theperiod of growth. Industry cost shares aregenerated by combining industry-level

data from the Gray-Bartelsman data withthe Bureau of Labor Statistics (BLS) capitalrental prices.

The CM does not include data on pur-chased services (other than that measuredthrough contract work). Baily, Hulten, andCampbell used a crude estimate ofpurchased services based on the two-digitratio of purchased services-to-materialsusage available from the Bureau of LaborStatistics KLEMS data (where KLEMSrefers to capital, labor, energy, materialsand service inputs). Baily, Hulten andCampbell applied the two-digit ratio fromthe aggregate KLEMS data to the plantlevel data on materials. Because this is atbest a crude adjustment that will not pro-vide much help in decomposing product-ivity growth within four-digit industries,this adjustment was not incorporated inthe analysis of the section on micro hetero-geneity and aggregate fluctuations.Furthermore, a comparison of the resultsin Baily, Hulten, and Campbell with thosegenerated here yields quite similar resultsfor the overlap industries (they considered23 industries) when I used their exactdecomposition methodology.

The data used are from the mailuniverse of the CM for 1977, 1982, and1987. In the CM, very small plants (typi-cally fewer than five employees) are excludedfrom the mail universe and denotedadministrative record cases. Payroll andemployment information on such verysmall establishments are available fromadministrative records (i.e., the StandardStatistical Establishment List) but theremainder of their data are imputed. Suchadministrative record cases are excludedfrom the analysis in the section on produc-tivity dynamics. In addition to the usualproblems in using book-value data, for plantsthat were not in the Annual Survey of Man-ufactures (about 50,000-70,000 plants) butin the mail universe of the CM, book-valuedata are imputed in years other than 1987.Baily, Hulten, and Campbell investigatedthis issue and found little sensitivity onthis dimension. This partly reflects the

MAY/JU N E 1997


76

Appendix

relatively small capital cost shares in totalfactor costs when materials are included.

As a further cross-check on sensitivityto measurement issues, the analysis in thesection on productivity dynamics was alsoconducted for labor productivity, using bothgross output and value-added meas- uresof labor productivity. Besides the indepen-dent interest in labor productivity, themeasurement of labor productivity (particu-larly on a gross output basis) is less fraughtwith measurement problems. Interestingly,quite similar results are generated whenusing output weights as the relevant sharesin the decomposition in Equation 7 butnow with plant-level productivity measuredin terms of labor productivity. An employ-ment-weighted decomposition of laborproductivity yields roughly similar results;however, the within-plant and net-entrycomponents play a more substantial role.Further exploration of these differences isbeyond the scope of this article but inter-esting to note, given currently availabledata collection and processing procedures.That is, analysis for sectors other thanmanufactur-ing along the lines conductedin the section on productivity dynamicswill, given the scant data collected on inputsfor most sectors, need to be based on adecomposition of labor productivity mea-sured on a gross output-per-worker basis.

DECOMPOSITION ISSUESAlthough the measure of plant-level

and aggregate industry-level total factorproductivity follows that of Baily, Hulten,and Campbell very closely, the decomposi-tion in 7 differs from that in Baily, Hulten,and Campbell in two important respects.The Baily, Hulten, and Campbell decompo-sition involved the three following terms:

• A within-plant component (theydenote this term as fixed shares) thatis identical to the within-plant compo-nent in Equation 7;

• A between-plant component (denotedas share effect) measured as the sum ofthe changes in the output shares for

each continuing plant weighted byending level of plant-level productivity;

• A net-entry component measured asthe output-weighted average product-ivity of entrants less the output-weighted average productivity ofexiting plants.

Thus, one difference is that by usingending-level plant-level productivity in their“share effect,” their share effect capturesboth the between-plant and the covarianceterm in the decomposition in Equation 7.Second, the between- and net-entry termsin Baily, Hulten, and Campbell do not involvedeviations of the relevant plant-level pro-ductivity from the initial average level ofproductivity in the industry. A consequenceof this formula- tion is that even if all plantshave the same productivity in period t2kand t, the Baily, Hulten, and Campbelldecomposition yields a nonzero between-plant term and an offsetting nonzero net-entryterm if the share of output due to enteringplants is different than the share of outputdue to exiting plants. Because the size ofexiting plants is typically larger than enteringplants, in practice this yields a bias towardsa positive between-plant term and a nega-tive net-entry term. This partly explainswhy Baily, Hulten, and Campbell find thatthe contribution of net entry is very smalland sometimes negative.

On a more general methodologicalnote, this discussion highlights the subtlebut important differences in between andwithin decompositions in a balanced panelversus an unbalanced panel. For balancedpanels, it is unnecessary to deviate the rel-evant term (e.g., initial plant-level produc-tivity) in the between component fromthe initial overall average because the sumof the changes in shares is zero. One thusobtains identical results with and withoutdeviating from means. This property doesnot hold for unbalanced panels necessita-ting a refinement of the standard betweenand within decomposition.

MAY/JU N E 1997


77

"Measuring and Analyzing Aggregate Fluctuations - St. Louis Fed

Documents