Professional Documents
Culture Documents
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Oxford University Press and The Review of Economic Studies, Ltd. are collaborating with JSTOR to digitize,
preserve and extend access to The Review of Economic Studies.
http://www.jstor.org
and
Uncertainty Investment
Dynamics
NICK BLOOM
StanfordUniversity,Centrefor EconomicPerformance,and NBER
STEPHENBOND
Institutefor Fiscal Studiesand Universityof Oxford
and
JOHN VAN REENEN
LondonSchool of Economics,
Centrefor EconomicPerformance,and CEPR
1. INTRODUCTION
Recent theoretical analyses of investment under uncertaintyhave highlighted the effects of
irreversibilityin generating"realoptions"(e.g. Dixit andPindyck, 1994). In these models, uncer-
taintyincreasesthe separationbetween the marginalproductof capitalwhichjustifies investment
and the marginalproductof capital which justifies disinvestment.This increases the range of
inaction where investmentis zero as the firm prefers to "wait and see" ratherthan undertak-
ing a costly action with uncertainconsequences. In short, investmentbehaviourbecomes more
cautious.
Firm-level data are attractivefor investigatingthis effect of uncertaintyon the degree of
cautionsince empiricalmeasuresof uncertaintycan be constructedbased on shareprice volatility
(e.g. Leahy and Whited, 1996). One importantdifficultyfor directtesting of real options models
of investmentunderuncertaintyusing firm data, however, is the extreme rarityof observations
with zero investmentin annual consolidated accounts. If we believed that these firms make a
single investmentdecision in each year,this lack of zeros would rejectthe canonicalreal options
model of a single investmentdecision with its region of inaction. However,given the extensive
evidence of discrete and lumpy adjustmentsin more disaggregatedplant-level data (e.g. Doms
and Dunne, 1998), this lack of zeros at the firm level is suggestive of aggregationover types of
capital,productionunits, and time.
391
1. See, for example, Bertola and Caballero(1994), Caballeroand Engel (1999), Abel and Eberly (2001), and
Doyle and Whited (2001). Thomas (2002) and Veracierto(2002) find that in general equilibriummodels the impact of
non-convex investmentcosts on the businesscycle may be small. These papersare necessarilybased on relativelysimple
models of firminvestment-including a constantlevel of uncertainty-to enablecomplex generalequilibriummodelling.
Our focus here is on much richer(partialequilibrium)micro-modelsthat include fluctuationsin the level of uncertainty.
These are appropriatefor estimationon firm-leveldata.
2. SIMULATINGINVESTMENTDYNAMICSUNDER UNCERTAINTY
The typical model in the literatureconsiders investmentin a single partiallyirreversiblecapital
good, with a Cobb-Douglas revenue function and demand conditions that follow a Brownian
motion process with constantvariance.Investmentonly occurs when the firm'smarginalrevenue
productof capital (MRPC) hits an upperthreshold,given by the traditionaluser cost of capital
plus an option value for investment. Similarly, disinvestmentonly occurs when the marginal
revenueproducthits a lower threshold,given by the user cost for selling capital less an option
value for disinvestment.The firmchooses to wait and do nothingif its MRPClies between these
two thresholds.
As the MRPC evolves stochasticallyover time this approachpredictsthat the firm will un-
dertakesporadicburstsof investmentor disinvestment,consistentwith the typical evidence from
plant-level data (see, for example, Doms and Dunne, 1998; or Nilson and Schiantarelli,2003).
Abel and Eberly (1996) show by comparativestatistics that the option values are increasingin
the (time-invariant)level of uncertainty.This suggests that firms which face a higher level of
uncertaintyare less likely to respondto a given demandshock.
TABLE 1
Episodes of zero investmentin differenttypesof data
Observationswith zero investment(%)
Buildings Equipment Vehicles Total
Firms 5-9 0.1 n.a. 0.1
Establishments 46.8 3-2 21.2 1.8
Single plants 53.0 4-3 23.6 2.4
Small single plants 57.6 5.6 24.4 3-2
2.2. Thesimulationmodel
We start by parameterizingone model from the general class of supermodularhomogeneous
models that we are considering.Firms are assumed to operate a large collection of individual
productionunits, with the numberchosen to ensure that full aggregationhas occurred.In the
simulationthis is set at 250 units per firm, chosen by increasingthe numberof units until the
resultswere no longer sensitive to this number.5
Each unit faces an iso-elastic demandcurve for its output,which is producedusing labour
and two types of capital. Demand conditions evolve as a geometric random walk with time-
varying uncertaintyand have a unit-specificidiosyncraticcomponentand a common firm-level
component. Demand shocks, uncertaintyshocks, and optimizationoccur in monthly discrete
time. Labouris costless to adjustwhile both types of capitalare costly to adjust.
2.2.1. The production unit model. In the basic model each productionunithas a reduced-
form supermodularrevenuefunction R(X, K1, K2) given by
In the simulationwe set a = 0.4 and fl = 0-4, correspondingto a 25% mark-upand constantre-
turnsto scale in the physical productionfunction,with equal coefficientson each type of capital.
Demand conditions are a composite of a unit-level (PU) and a firm-level(pF) component,
p = pU x pF. The unit-level demand(or productivity)conditions evolve over time as an aug-
mentedgeometricrandomwalk with stochasticvolatility:
PU = pU (1 + u (at) + at VtU), VU
- N
(0, 1) (2.4)
Here p (at) is the mean drift in unit-level demand conditions, a2 is the variance of unit-level
demand conditions,a* is the long-runmean of at, p, is the rate of convergenceto this mean,
and a2 is the varianceof the shocks to this varianceprocess. The terms VTUand Wtare the i.i.d.
shocks to unit-leveldemandand varianceconditions,respectively.
The firm-leveldemandprocess is also an augmentedgeometricrandomwalk with stochastic
volatility,which for tractabilitywe assume has the same mean and variance:
Hence, the overall demandprocess log P has drift 2,u(at) and variance2a2. While this demand
structuremay seem complex, it is formulatedto ensurethatunits withinthe same firmhave linked
investmentbehaviourdue to the common firm-level demand shocks and level of uncertainty,
but also display some independentbehaviourdue to idiosyncraticshocks. The baseline value of
2,u(at) is set to 4% (averagereal sales growth),invariantto the level of uncertainty,althoughwe
also reportbelow some experimentsthat allow for more generaldrifts.
2.2.2. Solving the production unit model. The complexity of the model necessitates
numericalsimulation,but analyticalresults can be used to show that the problemhas a unique-
valued continuoussolution7and an (almost everywhere)uniquepolicy function.This means our
numericalresults will be convergentwith the uniqueanalyticalsolution.
In principle, we have a model with too many state variablesto be solved using numerical
methods given the currentcomputingpower. The unit's optimizationproblem, however,can be
simplifiedby noting thatthe revenuefunction,adjustmentcost function,depreciationschedules,
andexpectationsoperatorsareall jointly homogeneousof degreeone in (P, K1, K2). This allows
us to normalizeby one statevariable-capital type 1-simplifying the model anddramaticallyin-
creasingthe speed of the numericalsolutionroutine.This effectively gives us one state"forfree",
in thatwe estimateon two majorstatespaces (j and 2), butfor threeunderlyingstatevariables.
The optimizationproblem(before normalization)can be statedas
V(Pt, Kit, K2t, Ut) = max R(Pt, Kit + Ilt, K2t + I2t) - C(Pt, Kit, K2t, Ilt, 12t)
Ilt,12t
1
+ E[V(Pt+1, (Kit +11t)(1 -6), (K2t+12t)(1 -6), t+l)],
1+r
where r is the discount rate, 6 is the depreciationrate, E[.] is the expectationsoperator,Ijt is
investmentin type j (j = 1,2) capitalat time t, and Kjt is the stock of type j capital.Using the
homogeneityin (P, K1, K2) this can be rewrittenas
Kit V(Pt*, 1, Kt, at) = max Klt R(Pt*, 1 + I*t, Kt (1 + I2t)) - KltC(Pt*, 1, K*, I*I*t tK2t)
11t'I2t
1
+ K1t+lE[V(Pt', 1 K*t+,
I+r at+l)],
where starredvariablesare K =- K, P* = ,I = , and I = normalizationby
2. Upon
Kit this simplifies to
6. Our choice of adjustmentcost parametersis based on the literaturewhere available,in particularCooper and
Haltiwanger(2006). The qualitativeresultsfrom our analysis of the simulateddataarenot sensitive to moderatechanges
in the adjustmentcost parametervalues, althoughas discussedin Section 3.2, they are sensitive to the type of adjustment
costs considered.
7. An applicationof Stokey and Lucas (1989) for the continuous,concave, and almost surely bounded normal-
ized returnsand cost function for models with partialirreversibilities(this section) and quadraticadjustmentcosts, and
Caballeroand Leahy (1996) for the extension to models with fixed costs in Section 3.2.2.
which is a function of only the state variables ( K, a). We let uncertainty,at, take five
,
equally spaced values from 0.1 to 0.5, with a symmetricmonthly transitionmatrix that is ap-
proximatelycalibratedagainst(the varianceand autocorrelationof) our stock-returnsmeasureof
uncertaintyfor U.K. listed firms,describedin Section 4.1. The simulationis run on a state space
of ' , a) of (100,100,5).8
2.2.3. Aggregation to firm-level data. Simulateddata are generatedby taking the nu-
merical solutions for the optimal investmentfunctions and feeding in demand and uncertainty
shocks at a monthlyfrequency.The simulationis runfor 60 monthsto generatean initial ergodic
distribution.Annual firm-levelinvestmentdata are then generatedby aggregatingacross the two
types of capital, the 250 units, and 12 months within each year. Capital stocks and the level of
the demandconditions are summed across all units at the end of each year, while uncertaintyis
measuredas the averageyearly value.
Using the model and solution method outlined above we generatesimulatedinvestmentand de-
manddatafor a panel of 50,000 firmsand25 years.We confirmthe two implicationsfor short-run
investmentdynamics highlighted in Section 2.1 by consideringthe relationshipbetween firm-
level annual investment rates and demand growth in this simulated data. As the drift in the
demandprocess is common to all firms, and the idiosyncraticshocks are averagedacross 250
productionunits, there is a simple correspondencebetween demand growth and the firm-level
demandshock in this simulation.
Figure 1 presents lowess-smoothed non-parametricplots9 of investment against demand
growthfor observationsaroundthe 10th,25th, 50th, 75th, and 90th percentilesof the distribution
of uncertainty(Ut).10 Investmentrates are measuredas annualinvestmentdivided by the capi-
tal stock at the beginning of the year, and annualdemandgrowthis measuredas the percentage
change comparingthe beginningandthe end of the year.The firstimplication-that the short-run
response of investmentto demandshocks will be lower at higheruncertainty-indicates that the
slope of these response functions is lower at higher levels of uncertainty.It is evident that these
non-parametricregression estimates do indeed become flatteras the level of uncertaintyrises,
consistent with the first implication of the theoreticalmodel. In quantitativeterms, comparing
investmentresponses with -10% and +25% demandgrowth, the gradientof the investmentre-
sponse to demandgrowthapproximatelydoubles when moving from the thirdquartileto the first
quartileof the distributionof uncertaintyand approximatelytriples when moving from the 90th
percentileto the 10th percentile.Hence, differencesin the level of uncertaintygeneratesubstan-
tial variationin the short-runresponse of investmentto demand shocks, and this is clearly seen
8. We also need the optimal control space of (I,*, 1*) of dimension (100,100), so that the full returnsfunction in
the Bellman equationhas dimensionality(100,100,100,100,5). The programmeand a manualexplainingthe underlying
techniquesare availableat http://www.stanford.edu/~nbloom/ or from nbloom@stanford.edu
9. Lowess smoothingestimates a linearregressionat each datapoint, using Cleveland's(1979) tricubeweighting
over a moving window of 5% of the data, to generatea non-parametricallysmoothed data series. Lowess is similar to
Kernelsmoothing,but uses informationon both the mean and the slope of the dataand so is more efficient in estimating
functions with continuousfirst derivatives,which our aggregateddata have asymptotically(in the numberof production
units).
10. As this varianceparameterhas a five-pointprocess in the underlyingmonthly model we obtain considerable
clustering of observationsaroundthese values, even in the average annual firm data. Although we use 1.25 million
generatedobservations,there are no observationsin the sample at the 10th and the 25th percentilesof uncertaintywith
annualdemandgrowthabove 27% and 64%, respectively,so the lines are not estimatedbeyond these points.
80
-50th Percentileof
a) uncertainty
25thPercentile of
60 - uncertainty
•
10thPercentileof
40- uncertainty
75thPercentile of
uncertainty
20 - 90th Percentile of
Suncertainty
0-
-25 0 25 50 75 100
Averageannualfirmdemandgrowth(%)
FIGURE1
Averageinvestmentresponseto demandshocks at differentlevels of uncertainty,simulateddata
3. EVALUATINGOUR EMPIRICALSPECIFICATION
The next step is to investigatethe empiricalimportanceof these propertiesof short-runinvest-
mentdynamicsin actualfirm-leveldata,which requiresan appropriateeconometricspecification.
If we observedthe trueunderlyingdemandshocks and demandvariancethis would be relatively
straightforwardas we could, for example,use the same non-parametricapproachused in the pre-
vious section to analyse short-runinvestmentresponsesto exogenous demandshocks. However,
in real firm-leveldata-setswe only observeproxies for demandgrowthsuch as firm sales growth
and proxies for uncertaintysuch as share price volatility. Among other issues, this requiresus
to deal with the problem that outcomes like sales and share prices arejointly determinedwith
the firm'sinvestmentdecisions. To do this we considerGMM estimatesof dynamiceconometric
investmentequations.
Ourstartingpoint is a reduced-formECM thatprovidesa flexible distinctionbetween short-
run influences on investmentrates and longer term influences on capital stocks. This has been
11. See, for example, Hall, Mairesseand Mulkay(1999) and Bond, Harhoffand VanReenen (2007).
12. The representationtheoremof Engle and Granger(1987) shows thatthe dynamicrelationshipbetween two I(1)
series thatare cointegratedcan be formulatedas an errorcorrectionrelationship.
13. Both this specificationand the results in Bloom (2000) are based on a single productionunit with one type of
capital.To check that this provides an accurateapproximationfor our aggregatedfirm-leveldata, we confirmedthat log
capitalwas cointegratedwith log sales in our simulateddata,with a coefficient of 1-008 on log sales. In Sections 3.2 and
4.4 we also considerrelaxingthe restrictionin (3.2) thatthis coefficient is unity.
TABLE2
Samplecorrelationsin the simulateddata
14. All the simulation data, actual data, and estimation code requiredto produce Tables 2-7 are available on
http://www.stanford.edu/-nbloom/and from nbloom@stanford.edu
TABLE3
Econometricestimationon the simulateddata
of "error-correcting" behaviour,with the actualcapital stock adjustingin the long run towardsa
targetthat is cointegratedwith its frictionless level. We find no evidence here that a permanent
increasein the level of uncertaintywould affect the level of the capital stock in the long run, but
thereis an indicationthatincreasesin uncertaintyreduceinvestmentin the shortrunin ways that
are not fully capturedby our multiplicativeinteractionterm.
Column(2) of Table3 uses insteadthe empiricalcounterpartsto the demandanduncertainty
variables,based on annuallevels of simulatedsales (Yit) and the within-yearstandarddeviation
of simulatedmonthly stock returns(SDit). As these variablesarejointly determinedwith invest-
ment decisions we treatthem as endogenousand reportGMM estimates.To mimic our empirical
analysis of real companydatamore closely, we also allow for the possibility of unobservedfirm-
specific effects here and estimatethis specificationin first differences.The instrumentsused are
the second and thirdlags of our simulatedinvestment,capital, sales, and uncertaintymeasures,
following Arellano and Bond (1991). A Sargan-Hansentest of overidentifyingrestrictionsdoes
not rejectthis specification,and thereis no significantevidence of second-orderserialcorrelation
in the first-differencedresiduals.While the parameterestimates are less precise in this case, we
again detect significantevidence thatuncertaintyinfluencesthe short-runresponseof investment
to demand shocks and that this response is convex. It should be noted, however, that this was
not always the case if we imposed simplerdynamic specificationsthat were rejectedby the test
of overidentifyingrestrictions(e.g. if we omit the errorcorrectionterm). This illustratesthe po-
tentialimportanceof controllingfor longer runinvestmentdynamicswhen testing the properties
of the short-runresponses to demandshocks. For other calibrationsof the simulationmodel we
found that alternativedynamic specificationsor instrumentsets may be required.The negative
coefficient on the interactionterm and the positive coefficient on the squaredterm, however,
3.2. Simulationrobustnesstests
To assess the generality of our predictionson the uncertainty-demandgrowth interactionterm
and on the demandgrowthsquaredterm,we now investigatewhetherthese effects are found for
an alternativerevenuefunctionand for alternativetypes of adjustmentcosts.
15. In a competitivemodel with shocks to outputprices and a flexible factor(such as labour)the MRPC is convex
in demandconditions, so uncertaintyhas a positive impacton the expectedMRPC.For example, with a revenuefunction
R = ZKa Lb (where Z is a demandprocess, K is capital,and L is labour),afteroptimizingout labournet revenueequals
CZ1/(1-b)Ka/(1-b) (where C is a constant)and the MRPC equals Z1/(1-b)K(a+b-1)/(1-b). If Z is a geometric
Brownian process with drift and variance a, then so the expected
E[dZ/(1-b)/Zl/(1-b)] = ( +~ i t
t b,
growth of MRPC +
equals (p 1, which is in
increasing uncertainty.However, as Caballero (1991) notes, the
) and whether the underlying
sign of this effect is sensitive to assumptions
_ 1___ such as the degree of imperfectcompetition,
shocks are to prices or quantities.Underalternativeassumptionsthe MRPC can become concave in demandconditions,
with a negative impact of uncertainty.To qualitativelysimulatethese positive and negative Hartman-Abeleffects in our
linear homogeneous specification,we adjustour demanddrift term by +- -,
noting that the quantitativeeffects would
also dependon the exact convexity/concavityof the underlyingMRPCin demandconditions.
TABLE4
Robustnesstests on the simulateddata
Dependentvariable:(Iit/Ki,tl) (1) (2) (3) (4) (5)
Estimationmethod OLS OLS OLS OLS GMM
Type of data for P and a True True True True Empirical
Adjustmentcosts Partial Fixed Quadratic All All
Revenuefunction CES Cobb- Cobb- Cobb- Cobb-
Douglas Douglas Douglas Douglas
Demandgrowth, Apit 0.400 0-586 0-486 0.378 0.630
(0-041) (0-093) (0.051) (0-039) (0.065)
Demandgrowthsquared,Ap2 0.029 -0-012 -0-004 0.011 0.403
(0-004) (0-031) (0.015) (0-004) (0-114)
Changein uncertainty,Aai t -0-389 - 1031 -0-048 -0-666 -0-064
(0-158) (0-283) (0-220) (0-141) (0-042)
Uncertainty,ait -0-006 -0-505 -1-216 0-033 0-101
(0-071) (0-131) (0-090) (0-066) (0-294)
Uncertaintyx demandgrowth, -1-566 1.766 -0-844 -1-791 -0-701
ait * Apit (0-373) (0-854) (0-427) (0-346) (0-294)
DemandECM term, (p - k)i,t_1 0.186 0.205 0-323 0.180 0-204
(0-008) (0-024) (0-020) (0-006) (0-098)
Second-orderserial correlation(p-value) 0.267
Sargan-Hansentest (p-value) 0.875
The associated linear homogeneous revenue function is then defined as R(P, K1, K2)
P -Y (K + K2)7/f, where P = Xa/(1-). We set f = 0.5 and y = 0.8.
Column (1) of Table4 presentsOLS results for simulatedfirm-leveldata with this alterna-
tive CES specification,using the true demandand uncertaintyvariables.We again find that the
short-runresponseof investmentto demandshocks is convex and thathigheruncertaintyreduces
this impacteffect of demandshocks on investment.First-differencedGMMestimates,using sales
as a measure of demand and stock-returnvolatility as a measure of uncertainty,also yielded a
significantpositive coefficient on the sales growth squaredterm and a significant negative co-
efficient on the uncertaintyinteractionterm.16This suggests that our empiricaltests can detect
these effects on short-runinvestmentdynamicswith this alternativespecificationof the revenue
function.
16. Coefficients(standarderror)of 0.627 (0-132) on the sales growthterm and -1.452 (0-467) on the uncertainty
interactionterm.
3.2.2. General adjustment costs. A number of previous papers, including Abel and
Eberly (1994), Bloom (2006, 2007), and Cooper and Haltiwanger(2006) have noted that dif-
ferent forms of adjustmentcosts can have significantlydifferentimplicationsfor investmentbe-
haviour.Ourcore predictionsarebased on a model with partialirreversibilities,butin this section
we investigatewhetherthey are also found using two additionaltypes of adjustmentcosts: fixed
disruptioncosts and quadraticadjustmentcosts.
Fixed disruptioncosts: When new capitalis addedto the productionprocess some downtime
may result in a fixed loss of output,howeverlarge the investment.For example, the factory may
need to close for a fixed period while a refit is occurring.For the simulationwe assume that the
fixed cost of adjustmentfor either type of capital is 5% of annualsales, which is approximately
calibratedon a monthlybasis from the annualestimatesin Cooperand Haltiwanger(2006).
Quadratic adjustmentcosts: The costs of investmentmay also be related to the rate of
adjustmentwith highercosts for more rapidchanges, which we specify as Cquad,j j ( K-)2
for j = 1,2. Forthe simulationwe assumethat2j = 0-3 for both types of capital,againcalibrated
roughly on a monthlybasis from Cooperand Haltiwanger's(2006) annualestimates.17
Since both these adjustmentcosts arejointly homogeneousof degree one in (P, K1, K2) the
cost function C(P, K1, K2, I, I2) is also homogeneous, permittingthe same normalizationby
capital type 1 and the resultingacceleratednumericalsolution as outlinedin Section 2.2.
Column (2) of Table4 presentsOLS results for simulatedfirm-leveldata with fixed adjust-
ment costs only, againusing the truedemandand uncertaintyvariables.Interestingly,in this case
we find that higher uncertaintyhas the opposite effect on the impact effect of demand shocks
on investment.This suggests that the "cautionaryeffect" of uncertaintyon short-runinvestment
dynamicsis sensitive to the form of adjustmentcosts, even withinthe class of non-convexadjust-
ment costs. The reasonfor this positive effect is that,underfixed costs, investmentis undertaken
as a jump process with the level of investmentdeterminedso as to return - and to targetlev-
els between theirthresholds,ratherthanto hold them continuouslyat theirinvestment 2 thresholds
as underpartialirreversibility.When uncertaintyrises and the thresholdsmove furtherapart,this
targetlevel moves by less, so thatthe gap between the targetand the investmentthresholdgrows
and the amount of investmentrequiredto reach the target also grows. This positive impact of
uncertaintyon the level of investmentundertakenat each investingunit offsets the negativeeffect
of uncertaintyon the numberof investingunits, leading to a positive effect of uncertaintyon the
responseof investmentto demandshocks in our simulationwith fixed costs only. Since the mag-
nitudeof these opposingeffects is likely to be sensitiveto the exactparameterizationof the model,
the sign of this short-runeffect is probablyambiguousunder(pure)fixed costs. We also find no
significantnon-linearityin the responseof investmentto demandshocks in this experiment.
Column (3) of Table 4 reportsthe OLS results for a simulationwith quadraticadjustment
costs only. In this case we find a smallerresponseof investmentto demandshocks at higherlev-
els of uncertainty,but again no significantindicationof non-linearityin the short-runresponses.
Again these results suggest that these propertiesof short-runinvestmentdynamics are sensitive
to the type of adjustmentcosts. In this experimentwe also find a stronglong-runeffect of uncer-
taintyin reducingthe level of the capitalstock,18althoughwe note thatthis effect could be offset
by a positive Hartman-Abeltype effect of the kind consideredin Table3.
17. For simplicity we have assumedthatboth the fixed and the quadraticadjustmentcosts are identical for the two
types of capital, with no cross-effects. In experimentsallowing for differentlevels of these adjustmentcosts for the two
types of capital,we found qualitativelysimilarresults.Ourapproachcould allow for more generalspecificationsof these
adjustmentcosts, with cross-effects,but we leave this for futureresearch.
18. With fixed costs or partialirreversibilities,the value functionis linearoutsidethe centralregionof inaction,with
slope equal to the purchase/resaleprice of capital, so that the value function is concave only in the region of inaction.
With quadraticadjustmentcosts, the value functionis globally curved,therebygeneratinggreaterglobal concavity.
Finally,in columns (4) and (5) we reportthe resultsfor a simulationthat combines all three
types of adjustmentcosts, with the same parametervalues used previously. For both the OLS
resultsusing the "true"explanatoryvariablesin column (4) andthe first-differencedGMMresults
using the observableproxies in column (5), we find evidence that higher uncertaintymakes the
responseof investmentto demandshocks more cautious,and thatthis responseis convex, with a
proportionatelylargerresponseto largershocks. At least for this combinationof adjustmentcost
parameters,based on the evidence presentedin Cooperand Haltiwanger(2006), we find that the
propertiesof the short-runinvestmentdynamics seem to be dominatedby the effects of partial
irreversibilitythatwe highlightedin Section 2.
In other robustnesstests we also experimentedwith, first, changing the discount rate from
10%to 5%, and second, relaxingthe restrictionin oureconometricspecificationthatthe long-run
coefficient on log sales is unity, as in equation(3.2). The signs and statisticalsignificanceof the
coefficients on the additionaluncertaintyinteractionand squareddemand growth terms in our
augmentedECMs were robustin both cases, both in the OLS and in the first-differencedGMM
results.19
4. EMPIRICALRESULTSFOR COMPANYDATA
In this section we want to investigateif our theoreticalpredictionshold in real firm-leveldata.
We use firm-leveldatafor an unbalancedpanel of 672 publicly tradedU.K. manufacturingfirms
between 1972 and 1991. We use pre-1991 U.K. databecause the accountingregulationsrequired
firmsto reportinvestmentexpenditureconsistentlythroughoutthis period, togetherwith the net
book value of capitalfrom acquisitionof subsidiariesandrevenuefrom sales of fixed assets. This
enablesus to constructrelativelyaccuratemeasuresof bothinvestmentanddisinvestment.We use
data on publicly tradedcompanies because these allow us to constructmeasuresof uncertainty
from high-frequencystock marketreturnsdata.
4.1. Uncertaintymeasures
19. For the simulateddata with a 5% discount rate, a consistent finding was that the effect of uncertaintyon the
short-runresponseof investmentto demandshocks was largerthanwe found with a 10%discountrate.This is consistent
with the highervalue of real options at a lower discountrate.These results are availableon requestfrom the authors.
4.3. Estimationresults
Our main econometric results are estimated using the system GMM proceduredeveloped by
Arellano and Bover (1995) and Blundell and Bond (1998). This combines a system of equations
in first differencesusing suitablylagged levels of endogenousvariablesas instruments,as in the
basic first-differencedGMM estimator(see Arellano and Bond, 1991), with equationsin levels
for which lagged differencesof endogenousvariablesare used as instruments.Unobservedfirm-
specific effects areeliminatedfromthe first-differencedequationsby the transformation.The key
requirementis thatthe additionalinstrumentsused in the levels equationsshouldbe uncorrelated
with the unobservedfirm-specificeffects in the investmentequation, which is tested using the
Sargan-Hansentest of overidentifyingrestrictions.The advantageis that, if these additionalin-
strumentsare valid, the system GMM estimatorshouldhave greaterefficiency and smallerfinite
sample bias than the correspondingfirst-differencedGMM estimator.The reportedresults treat
both sales and stock-returnsvolatilityas endogenousvariables,with the precise instrumentsused
noted in the tables. Similarresults were found using a range of alternativeinstrumentsets, and
our main findings concerningthe short-runeffects of sales growthand uncertaintyon company
investmentwere also found using the first-differencedGMM estimator.
Our main specificationis based on equation(3.4), with currentand lagged cash flow vari-
ables as additionalcontrols.Such termsare often found to be informativein micro-
(Cit/Ki,t-1)
econometric investment equations and may reflect either financing constraints (see Fazzari,
TABLE5
Econometricestimatesusing U.K.companydata
20. We reportthe squaredcorrelationcoefficient between actual and predictedlevels of the investmentrate. This
squaredcorrelationmeasureis equivalentto the standardR2 in an OLS regression,and is recommendedas a goodness-
of-fit measurefor instrumentalvariableregressionsby, for example,Windmeijer(1995).
Column (2) of Table5 adds a squaredtermin currentreal sales growthto this basic specifi-
cation. In line with our resultsfor the simulateddataunderpartialirreversibilityin Section 3, we
find significantpositive coefficientson both the level andthe squareof real sales growth.Column
(3) adds a range of uncertaintyterms to this extended errorcorrectionspecification.The main
result of interest here is the significantnegative coefficient on the interactionterm. The linear
uncertaintyterms (the change in uncertaintyand the lagged level of uncertainty),in contrast,are
found to be only weakly significant,with a joint test of their exclusion from the specificationin
column (3) not rejected(p-value= 0-17). We includethese termsherepartlyto ensurethatthe sig-
nificantcoefficient on the interactiontermis not the resultof omittingrelevantlinearuncertainty
termsand partlyto investigatewhetherthereis significantevidence of a long-runeffect of uncer-
tainty on capital accumulation.The insignificanceof the lagged level of uncertaintyin column
(3) formallyrejects the presence of such a long-runeffect, althoughthe imprecisionwith which
we estimatethis coefficientsuggests thatthis test may not be very powerful.21Omittingthis term
in column (4) results in an insignificantcoefficient on the short-runchange in uncertaintyterm,
which we also omit from our preferredparsimoniousspecificationin column (5). Thus the only
effect of uncertaintyon companyinvestmentbehaviourthat we can detect with a high degree of
statisticalconfidenceis the interactionwith the impacteffect of currentreal sales growth.22
Table 6 investigates this interactioneffect further.Here we decompose our stock-returns
measure of uncertainty(SDit) into three components-a macroeconomiccomponent,common
to all firms in a particularyear a time-invariantfirm-specificcomponent (SDi); and an
(SD/);
idiosyncratictime-varyingcomponent(SDit = SDit - SDt - SDi). Columns (1)-(3) include in-
teractionsbetween real sales growth and each of these uncertaintyvariablesindividually,while
column (4) includes all threeinteractiontermsjointly.
The interactionbetween firm-levelreal sales growthand a purely macroeconomicmeasure
of uncertainty,includedin column (1) of Table6, is the least informativeof our threevariables.23
The interactionwith a time-invariantfirm-specificmeasure of uncertainty,reportedin column
(2), is only weakly significant,while columns (3) and (4) show that it is the interactionbetween
sales growth and the idiosyncratictime-varyingcomponent (SDit) of our uncertaintymeasure
thatis most informative.However,because the coefficientson the remaininginteractiontermsin
column (4) areestimatedimprecisely,we can easily acceptthe restrictionof common coefficients
on these threeinteractions,as imposed in our preferredempiricalspecification.
4.4. Robustnesstests
We conducteda numberof robustnesstests, some of which arereportedin Table7. In column (1)
we omit the cash-flow variables,which causes the test of overidentifyingrestrictionsto reject,
but does not affect the sign or the significanceof the coefficientson the interactiontermbetween
uncertaintyand sales growthandon the squaredsales growthterm.In column (2) we adda further
interactionwith firm size, defininga "big firm"dummyvariableBIGit that takes the value 1 for
observationswith real sales above the sample median and 0 otherwise. Ourresult indicates that
the effect of uncertaintyon the impact effect of sales growth on investmentis not significantly
differentbetween the smaller and larger firms within our sample. The point estimate suggests
that this effect may be smallerfor the relativelylarge firms.One possible explanationis that our
21. That is, our results do not rule out the possibility of an economically significantnegative long-run effect of
uncertaintyon capital accumulation,althoughwe cannotconfirmthe presenceof such an effect with any confidence.
22. Bloom and Van Reenen (2002) show a linked finding that higher uncertaintyreduces the responsiveness of
firms' productivityto new patents.
23. The limited informationthat we find in macroeconomicas opposed to microeconomicvariationin our uncer-
tainty measure may help to explain why annualtime-series studies of aggregateinvestmentdata have often not found
significanteffects of uncertaintyvariables.
TABLE6
Separatingtime,firm, and residual variationin uncertainty
Dependentvariable:(lit/Ki,tl) (1) (2) (3) (4)
Sales growth(Ayit) 0.127 0-141 0.474 0.499
(0-052) (0-053) (0-182) (0-184)
Cash flow (Cit/Ki,tl) 0.270 0.263 0.287 0.280
(0-124) (0-127) (0-122) (0-124)
Lagged cash flow (Ci,t-1/Ki,t-2) 0-261 0.270 0.264 0-273
(0-079) (0-080) (0-076) (0-077)
Errorcorrectionterm (y - k)i,t_1 0.054 0.056 0.047 0.049
(0-027) (0-027) (0-026) (0-026)
Sales growthsquared(Ayit)2 0.497 0-507 0.534 0.537
(0-170) (0-157) (0-148) (0-162)
Time uncertaintyx sales growth 0-016 -0-051
(SDt)* (Ayit) (0-150) (0-136)
Firmuncertaintyx sales growth -0-130 -0-136
(SDi) * (Ayit) (0-105) (0-107)
Residualuncertaintyx sales growth -0-225 -0-230
(SDit) * (Ayit) (0-102) (0-103)
Goodness of fit-Corr(I/K, IK)2 0.307 0.298 0.311 0.288
Second-orderserial correlation(p-value) 0.096 0.094 0.132 0.106
Sargan-Hansen(p-value) 0.399 0.490 0-383 0-452
24. Switching from daily to monthly returnsdata is expected to increase the sampling varianceof SDit fivefold.
The monthly (weekly) measure of uncertaintyhas a correlationcoefficient with the daily measure of 0.784 (0-900).
Using the weekly uncertaintymeasure,the coefficient (standarderror)on the uncertainty-salesgrowthinteractionterm
was estimatedto be -0-145 (0-082), aboutmid-way between the estimatesusing the daily and monthlymeasures.
TABLE7
Robustnesstests on U.K.companydata
25. This is done by multiplying SDit by the ratio of equity to (equity + debt), with equity measuredusing the
marketvalue of shares(ordinaryand preference)and debt measuredusing the book value of all long-termdebt.
25thPercentileof uncertainty
4.0-
50thPercentileof uncertainty
2.0-
Ca
90thPercentileof uncertainty
0.0-
0 2 4 6 8 10
Years after the sales shock
FIGURE 2
Investmentresponseto a sales shock at differentlevels of uncertainty,U.K. firm-leveldata
26. See Bloom, Bond and Van Reenen (2003) for more details of these and furthersimulations.The exact size of
the sales shock makesrelativelylittle differenceto the results.
27. For comparison,the increase in averageuncertaintyfor our sample firms after the 1973 oil crisis is similarin
magnitudeto the increasefrom the firstquartileto the thirdquartileof our sampledistribution.
10thPercentileof uncertainty
1.01 -
75th Percentile
ofuncertainty
Percentileof uncertainty
1.00
-905h
0 2 4 6 8 10
Yearsafterthe sales shock
FIGURE 3
that the "short-run"effects persist for a significantperiod. Even after 10 years there is still a
noticeable differencebetween the predictedincreasesin capital stocks, in response to the same
demandshock, at differentlevels of uncertainty.
6. CONCLUSIONS
This paperdevelops two implicationsof partialirreversibilityfor the short-rundynamics of in-
vestment. First, investmentwill respond more cautiously to a given demand shock at higher
levels of uncertainty(due to wider thresholdsfor the zone of inaction), and second, investment
will have a convex response to positive demandshocks (due to aggregationand supermodular-
ity). We confirmthese implicationsusing numericalmethodsto solve a model with two types of
capital, a rich mix of adjustmentcosts (partialirreversibility,quadratic,and fixed), time-varying
uncertainty,alternativefunctionalforms for the revenuefunction,and aggregationover time and
productionunits. We propose and evaluatean econometricspecificationthat is designed to test
for these propertiesof short-runinvestmentdynamicsusing firm-leveldata.We reportevidence
that both the "cautionary"and the "convexity"effects are found using a measureof uncertainty
based on stock-returnsvolatility for a large panel of manufacturingfirms.Throughboth numer-
ical and econometric simulationswe show that these effects are economically important-the
investmentresponse to a demandshock is doubledby moving from the thirdquartileto the first
quartilein the distributionof our measure of uncertaintyand quadrupledby moving from the
90th to the 10thpercentile.
This indicatesthatthe one standarddeviationincreasein measuresof uncertaintyobserved
aroundmajorshocks, like September11, 2001 and the 1970's oil shocks, could seriouslyreduce
the responsivenessof investmentto subsequentmonetaryor fiscal policy. While we do not model
the behaviourof labourdemand,the existence of similarlabourhiringand firingcosts would im-
ply thathigheruncertaintywould also make firmsmore cautiousin theiremploymentresponses.
This is importantas policy-makerstypicallywantto respondto majorshocks, but the behavioural
responses to any given policy stimulusmay be much lower thannormalin these periods of high
uncertainty.
@ 2007 The Review of Economic StudiesLimited
APPENDIX.DATAAPPENDIX
The companydataare takenfrom the consolidatedaccountsof manufacturingfirmslisted on the U.K. stock market.
We deleted firms with less than three consecutive observations,brokethe series for firmswhere accountingperiodsfell
outside the range300-400 days (due to changes in yearends), and excludedobservationsfor firmswherethere arejumps
of greaterthan 150%in any of the basic variables.These dataare obtainedfrom the Datastreamonline service.
Investment(I): Totalnew fixed assets (DS435) less sales of fixed assets (DS423).
Capital stock (K): Constructedby applying a perpetualinventoryprocedurewith a depreciationrate of 8%. The
startingvalue was based on the net book value of tangiblefixed capital assets (DS339) in the firstobservationwithin our
sample period, adjustedfor previousinflation.Subsequentvalues were obtainedusing accounts data on investmentand
asset sales and an aggregateseries for investmentgoods prices.
Sales (Y): Totalsales (DS104), deflatedby the aggregateGDP deflator.
Cashflow (C): Net profits(earnedfor ordinary,DS 182) plus depreciation(DS 136).
Uncertainty(a): The computationof this variableis describedin the text. For each companywe take the daily stock
marketreturn(DatastreamReturnsIndex, RI). This measure includes on a daily returnsbasis the capital gain on the
stock, dividendpayments,the value of rightsissues, special dividends,andstock dilutions.We then computethe standard
deviationof these daily returnson a year-by-yearbasis, matchedprecisely to the accountingperiod.We trimthe variable
so thatvalues above five are set equal to five. The results are robustto droppingthese 10 observations.
REFERENCES
ABEL, A. (1983), "OptimalInvestmentunderUncertainty",AmericanEconomicReview,73, 228-233.
ABEL, A. and EBERLY,J. (1994), "A Unified Model of InvestmentunderUncertainty",American Economic Review,
84, 1369-1384.
ABEL, A. and EBERLY,J. (1996), "OptimalInvestmentwith Costly Reversibility",Review of Economic Studies, 63,
581-593.
ABEL, A. and EBERLY,J. (1999), "TheEffects of Irreversibilityand Uncertaintyon CapitalAccumulation",Journalof
MonetaryEconomics,44, 339-377.
ABEL, A. and EBERLY,J. (2001), "Investmentand q with Fixed Costs: An EmpiricalAnalysis" (Mimeo, Wharton
School, Universityof Pennsylvania).
ARELLANO, M. and BOND, S. (1991), "Some Tests of Specificationfor Panel Data: Monte Carlo Evidence and an
Applicationto EmploymentEquations",Reviewof EconomicStudies,58, 277-297.
ARELLANO,M. and BOVER, O. (1995), "AnotherLook at the Instrumental-Variable Estimationof ErrorComponents
Models",Journalof Econometrics,68, 29-52.
BERTOLA,G. and CABALLERO,R. (1994), "Irreversibilityand AggregateInvestment",Reviewof Economic Studies,
61, 223-246.
BLOOM, N. (2000), "The Dynamic Effects of Real Options and Irreversibilityon Investmentand LabourDemand"
(WorkingPaperNo. W00/15, Institutefor Fiscal Studies).
BLOOM, N. (2006), "The Impact of UncertaintyShocks: A Firm-LevelEstimationand a 9/11 Simulation"(Mimeo,
Stanford).
BLOOM,N. (2007), "Uncertaintyand the Dynamics of R&D",AmericanEconomicReview(forthcoming).
BLOOM, N., BOND, S. and VAN REENEN, J. (2003), "Uncertaintyand CompanyInvestmentDynamics: Empirical
Evidence for UK Firms"(Discussion PaperNo. 4025, CEPR).
BLOOM, N. and VAN REENEN, J. (2002), "Patents,Real Options and Firm Performance", Economic Journal, 112,
C97-C 116.
BLUNDELL,R. and BOND, S. (1998), "InitialConditionsand Moment Restrictionsin Dynamic Panel Data Models",
Journalof Econometrics,87, 115-143.
BOND, S., HARHOFF,D. and VAN REENEN, J. (2007), "Investment,R&D and FinancialConstraintsin Britainand
Germany",Annales d'Economieet de Statistique(forthcoming).
BOND, S., KLEMM,A., NEWTON-SMITH,R., SYED, M. and VLIEGHE,G. (2004), "The Roles of Expected Prof-
itability,Tobin'sq and CashFlow in EconometricModels of CompanyInvestment"(WorkingPaperNo. 222, Bank
of England).
BOND, S., MOESSNER,R., MUMTAZ,H. and SYED, M. (2005), "MicroeconometricEvidence on Uncertaintyand
Investment"(Mimeo, Institutefor Fiscal Studies).
CABALLERO,R. (1991), "On the Sign of the Investment-Uncertainty Relationship",AmericanEconomic Review, 81,
279-288.
CABALLERO,R. and ENGEL, E. (1999), "ExplainingInvestmentDynamics in U.S. Manufacturing:A Generalized
(S,s) Approach",Econometrica,67, 783-826.
CABALLERO,R. and LEAHY, J. (1996), "Fixed Costs: The Demise of Marginal Q" (Working Paper No. 5508,
NBER).
CAMPBELL,J., LETTAU,M., BURTON, G. M. and XU, Y. (2001), "HaveIndividualStocks Become More Volatile?
An EmpiricalExplorationof IdiosyncraticRisk",Journalof Finance, LVI (1), 1-43.
CHETTY,R. (2007), "InterestRates, Irreversibility,and Backward-BendingInvestment",Review of Economic Studies,
74, 67-91.
CLEVELAND,W. (1979), "RobustLocally WeightedRegressionand SmoothingScatterplots",Journalof the American
StatisticalAssociation, 74, 829-836.
COMIN,D. and PHILIPPON,T. (2006), "TheRise in Firm-LevelVolatility:Causes and Consequences",NBERMacro-
economicsAnnual,20, 167-227.
COOPER,R. and EJARQUE,J. (2003), "FinancialFrictions and Investment:A Requiem in Q", Review of Economic
Dynamics, 6, 710-728.
COOPER, R. and HALTIWANGER,J. (2006), "On the Nature of Capital AdjustmentCosts", Review of Economic
Studies,73, 611-633.
DAVIS,S., HALTIWANGER,J., JARMIN,R. and MIRANDA,J. (2006), "Volatilityand Dispersionin Business Growth
Rates: Publicly Tradedversus PrivatelyHeld Firms"(WorkingPaperNo. w12354, NBER).
DIXIT,A. and PINDYCK,R. (1994) InvestmentUnder Uncertainty(Princeton,NJ: PrincetonUniversityPress).
DOMS, M. and DUNNE, T. (1998), "CapitalAdjustmentPatternsin ManufacturingPlants", Review of Economic
Dynamics, 67, 117-145.
DOYLE, J. and WHITED,T. (2001), "Fixed Costs of Adjustment,Coordination,and IndustryInvestment",Review of
Economicsand Statistics,83, 628-637.
EBERLY,J. and VAN MIEGHEM,J. (1997), "Multi-FactorDynamic InvestmentunderUncertainty", Journal of Eco-
nomic Theory,75, 345-387.
ENGLE,R. and GRANGER,C. (1987), "Cointegrationand Error-Correction: Representation,Estimationand Testing",
Econometrica,55, 251-276.
ERICKSON,T. and WHITED,T. (2000), "MeasurementErrorand the Relationshipbetween Investmentand q", Journal
of Political Economy,108, 1027-1057.
FAZZARI,S., HUBBARD,R. andPETERSEN,B. (1988), "FinancingConstraintsandCorporateInvestment",Brookings
Papers on EconomicActivity,1, 141-195.
GUISO, L. and PARIGI, G. (1999), "Investmentand Demand Uncertainty",QuarterlyJournal of Economics, 114,
185-227.
HALL, B., MAIRESSE, J. and MULKAY,B. (1999), "Firm-LevelInvestmentin France and the United States: An
Explorationof WhatWe Have Learnedin TwentyYears",Annales d'Economieet de Statistique,55-56, 27-69.
HARTMAN, R. (1972), "The Effect of Price and Cost Uncertaintyon Investment",Journal of Economic Theory, 5,
258-266.
LEAHY,J. andWHITED,T. (1996), "TheEffects of Uncertaintyon Investment:Some Stylized Facts",Journalof Money
Creditand Banking,28, 64-83.
MERTON, R. (1980), "On Estimating the Expected Return on the Market",Journal of Financial Economics, 8,
323-361.
NILSON, O. and SCHIANTARELLI,E (2003), "Zeroesand Lumpsin Investment:EmpiricalEvidence on Irreversibili-
ties and Nonconvexities",Reviewof Economicsand Statistics,85, 1021-1037.
REDUTO DOS REIS, I. (1999), "IrreversibleInvestmentand Fixed Costs. Evidence from ARD EstablishmentLevel
Data"(Mimeo, UniversityCollege London).