The Sensitivity of Empirical Studies to Alternative …...The Sensitivity of Empirical Studies to...

F EDERAL RESERVE BANK OF ST. LOU IS

51

NOVEMBER/DECEMBER 1996

The Sensitivityof EmpiricalStudies toAlternativeMeasures of theMonetary Baseand Reserves Michael J. Dueker andApostolos Serletis

T

he adjusted monetary base has longbeen an index measure that aggregatesthe effects of Federal Reserve open-

market operations, discount windowlending and reserve-requirement changes.Traditionally the adjustment procedure hasassumed that a change in required reservesleads to a one-to-one change in thedemand for reserves. Recently, however,Anderson and Rasche (1996) have revisedthe St. Louis measure of the adjusted mon-etary base to reflect that many depositoryinstitutions are not bound by reserverequirements, so their demand for reservesdoes not change with further reductions inreserve requirements. Using bank-leveldata, Anderson and Rasche determine foreach financial institution whether it isbound at the time of the change inrequired reserves, and they adjust thereserves only of bound banks.

A second innovation of Anderson andRasche is to include banks’ requiredclearing balances in the measure of bankreserves. In terms of helping to reduce thelikelihood of payments-related overdraftsin banks’ accounts with the Fed, clearingbalances are functionally equivalent to thereserve account deposits that are formallypledged towards satisfying reserve require-

ments. Furthermore, many nonboundbanks freely choose to hold deposits at theFed under clearing balance agreements.

Hence Anderson and Rasche’srevisions affect both components of tradi-tional measures of adjusted reserves: thesource base (through the addition ofrequired clearing balances) and the reserveadjustment magnitude (through moresophisticated analysis of how changes inreserve requirements affect the demand forreserves). Anderson and Rasche’s articlecontains a full description of the revisionsthey make to adjusted reserves in the post-1980 period.

Because the adjusted monetary baseand adjusted reserves have seen wideempirical use as indicators of monetarypolicy, it is important to know whether theAnderson and Rasche revisions alterexisting empirical evidence regarding thestance and potency of monetary policy.The present paper applies both the old andrevised measures in empirical models toexamine whether the revisions cause theconclusions to change. Any altered empir-ical results would be a

consequence of thedata revision, but not a rationale or justifi-cation for the revision. One would expectthat the growth rates of the revised and theold measures of the adjusted monetarybase would differ most greatly in periodssurrounding reserve requirement changesand periods when required clearingbalances change rapidly. Just as it is neces-sary to examine whether the switch fromfixed-weight to chain-type measures ofGDP alters our understanding of businesscycles and recent economic events, similartests of monetary policy indicators shouldaccompany revisions to the adjusted mon-etary base.

The first section of this article presentsMcCallum and Hargraves’ (1995) mone-tary impulse measure for the revised andold base measures. In a similar vein, weuse McCallum’s (1987, 1988) monetarybase rule as a model of U.S. monetary

Michael J. Dueker is a senior economist at the Federal Reserve Bank of St. Louis. Apostolos Serletis is a professor of economics at theUniversity of Calgary. Nick Meggos provided research assistance.

FEDERAL RESERVE BANK OF ST. LOU IS

52

1 Note that the revised St. Louisadjusted base and reservesmeasures used in this articlewere those available in January1996. Both series have under-gone further minor revisionssince then.


policy and estimate the coefficients andnominal GDP targets that best describe the growth in both the revised and oldadjusted base measures. The secondsection compares Vector Autoregressions(VARs) that use either the revised or oldmeasures of the base and reserves, startingwith the model from Sims (1992) andBernanke and Blinder (1992) that viewsinnovations in the federal funds rate asmonetary policy shocks. We also examineVARs proposed by Eichenbaum (1992);Christiano, Eichenbaum and Evans (1995);and Strongin (1995), all of which viewinnovations in non-borrowed reserves as amonetary policy shock. The section con-cludes with a VAR from Bernanke andMihov (1995) that is tailored to take intoaccount changes in the Federal Reserve’soperating procedures when identifyingmonetary policy shocks.

The third section looks at the new andold measures of the monetary base andreserves in the context of a structural VAR,following the “monetarist” specification ofHaslag and Hein (1995). The questionposed here is whether open-market opera-tions and reserve requirement changeshave differential effects on the economy.As another example, we repeat the Loun-gani and Rush (1995) “credit-view”regressions that test for significant effectsfrom reserve-requirement shocks onoutput and investment.

Empirical models that focus on variablescorresponding to reserve-requirementshocks, such as Loungani and Rush, besthighlight the differences between the twobase series. Reserve-requirement shocksdo not have a significant impact on outputwhen the old base and reserves measuresare used, but they do exert a significantinfluence when the revised measures areapplied. In the estimated McCallum rule,we find differences in the velocity forecasterrors between the old and revised bases,especially after 1985. In the post-1985period, the revised base measure leads to alower estimated target nominal GDPgrowth than the old base. In the VARs, theresults are qualitatively the same for boththe revised and old bases, although thereare minor differences in the variancedecompositions. The variance decomposi-tions from a structural VAR of Haslag andHein (1995) also yield the interestingresult that the old definitions imply a rela-tively large role for source-base shocks inexplaining inflation. The revised data, onthe other hand, attribute inflation variabil-ity less to source-base shocks and more toshocks to the inflation rate itself.

The revised St. Louis adjusted mone-tary base and reserves series have beenchained back to 1959.1 The VARs insection II use this full 1960-1995 sample.Several other empirical investigations usedshorter sample periods as noted in thetext. As Anderson and Rasche (1996)explain, the old and revised measuresdiffer most significantly in the post-1980period, because required clearing balancescame into existence in 1980 and grew inpopularity, and the percentage of non-bound banks grew appreciably. To see thedifferences, it helps to focus on reserves,because the currency component of thebase is identical for both measures, and 90percent of the base consists of currency.Figure 1 illustrates year-over-year growthrates of the old and revised measures ofadjusted St. Louis reserves. The revisedseries shows faster growth from 1985 to1988, perhaps presaging the build-up ofinflation that peaked in 1990. The revisedmeasure also experienced slower growth

New Measure

Old Measure

Percent Change over Previous Year

Reserves Growth

25

20

15

10

5

0

-5

-1081 82 83 84 85 86 87 88 89 90 91 92 93 94 951980

Figure 1


53

2 In the chart, the point plottedas the monetary impulse in1990Q1, for example, is thetwo-year moving average cover-ing 1988Q1 to 1990Q1. Atthat same date, the “lead” ofnominal GDP growth is a two-year moving average covering1989Q1 to 1991Q1.


than the old measure in 1991 prior to asluggish recovery.

THE MONETARY BASE INMONETARY POLICYINDICATORS

McCallum and Hargrave’sMonetary Impulse Measure

In a recent study for the InternationalMonetary Fund, McCallum and Hargraves(1995) present a comparison of the “mone-tary impulses” generated across time bycentral banks in the G-7 countries. Theysum monetary base growth, adjusted forchanges in reserve requirements, and amedium-term forecast of velocity growth.With this measure of monetary stimulus,McCallum and Hargraves examine whethercurrent monetary policy actions (summa-rized by growth in the adjusted monetarybase) suggest higher or lower inflationarypotential, relative to current inflation.Hence their monetary impulse measureserves as a natural vehicle for the jointstudy of differences in growth ratesbetween the two adjusted base measuresand the differences in their predicted veloc-ities. A given sustained level of monetaryimpulse ought to translate eventually intonominal spending growth of equal magni-tude, due to the velocity adjustment:

∆yt\t-1 = ∆mt + ∆[y - m]t\t-1,

where y is nominal spending and m is themonetary base (both in logs). Subscript t\t-1indicates a forecast based on informationavailable through the previous period. Thevelocity forecasts in our calculated monetaryimpulse measures are from a conditionallyheteroscedastic time-varying coefficientmodel estimated via the Kalman filter, usingquarterly data from 1960. The first threeyears of data were reserved for an initial esti-mate of the coefficients. The explanatoryvariables in the model are lagged velocity,lagged M1 growth, and the lagged firstdifference of the three-month Treasury billrate. Further details on the velocity fore-casting method are in Dueker (1993).

For a visual test of the predictive powerof the monetary impulse measures, Figure 2plots two-year moving averages of the mon-etary impulse measures with nominal GDPgrowth in the following year.2 Both mone-tary impulse measures overpredicted thesurge in nominal GDP growth between 1983and 1985, because they underpredicted thefall in velocity that followed the disinflationof the early 1980s. The revised adjustedbase yielded a monetary impulse measurethat predicted the magnitude of the increasein nominal GDP growth from 1983 to 1985but overstated the level. The monetaryimpulse measure from the old base, on theother hand, underpredicted nominal GDP’sacceleration but overstated the level by less.Since 1985, both monetary impulsemeasures have generally overpredictedfuture nominal GDP growth, primarilybecause of two slowdowns in nominal GDPgrowth: one in 1986 and another correspon-ding with the 1990-91 recession. For thepast six years, however, the monetaryimpulse measure from the revised base hasremained consistently below that from theold base by about 0.5 percent, on average.

McCallum’s Rule as an Estimated Equation

Another empirical exercise similar inspirit to the monetary impulse measuredescribed above consists of estimating aneconometric model in which the central

Monetary Impulse Measuresfor the United StatesPercent Change from Previous Year

14

12

10

8

6

4

2

069 71 73 75 77 79 81 83 85 87 89 91 93 951967

New Base Impulse

Old Base Impulse

Nominal GDP Growth(one year later)

Figure 2


54


bank’s willingness to let the monetary basegrow at its historical rate elucidates itsimplicit nominal GDP growth targets. Inthis section we present another method ofderiving base-implied nominal GDPgrowth rates, and we compare results forboth base measures. The starting point isMcCallum’s rule, which sets quarterly basegrowth according to a target growth ratefor nominal GDP, a forecast of base velo-city growth, and a feedback parameter onthe gap between the actual and targetlevels of nominal GDP:

(1) ∆mt = λ0t - ∆[y - m]t\t-1 +λ 1t( y-y)t-1.

The parameter λ0t is the (possiblytime-varying) target rate of nominal GDP

growth. The log of the monetary base ism, y is log nominal GDP, and (y-m) is thelog of base velocity. Finally, y is the targetvalue of y. The forecast of base velocityindicates how fast money ought to grow inorder to hit the nominal GDP growthtarget. Simulations of McCallum’s rulerequire a fully-specified stochastic modelof how nominal GDP growth will respondto the rule-implied base growth.

As an alternative to simulations,Dueker and Fischer (1996) use McCallum-type rules as models of policy and estimatethe parameters that best describe observedbase growth as an outcome of nominalGDP targeting. We conduct the estimationexercise twice, once for the revised (new)measure of the adjusted monetary base andonce for the old measure, and look for dif-ferences in the path of the inferred targetrates for nominal GDP growth. If the twomeasures are nearly perfect substitutes,then the inferred target rates of nominalGDP growth ought not differ appreciably.As described earlier, the model used togenerate one-step-ahead forecasts is atime-varying parameter model withheteroscedastic errors. Estimates of theforecast error variances for the growth ratesof the velocities of the old and revisedadjusted base measures differ across time,especially after 1985. Figure 3, whichfocuses on the post-1985 period, showsthat the forecast error variance for thevelocity of the old measure of the adjustedbase is about 15 percent greater than thecorresponding variance for the revisedmeasure of the adjusted base.

Inserting the velocity forecasts intoEquation 1 and adding a mean-zero errorterm that has time-varying variance

s2t, we

arrive at an econometric model of nominalGDP targeting. We also drop fromEquation 1 the feedback from the level ofnominal GDP, because all estimationsfound the feedback parameter λ1 to benear zero:

(2) ∆mt = λ0t - ∆[y - m]t\t-1 +et.

Because the implicit nominal GDPgrowth target of the Federal Reserve has

Forecast Error Variances forVelocities of Adjusted Base Measures

1.75

1.50

1.25

1.00

0.75

0.50951985 86 87 89 90 91 92 93 9488

Old Base

New Base

Figure 3

13121110

9876543

1972

Percent Change over Previous Year

Nominal GDP Growth Targetsfrom Two Base Measures

74 76 78 80 82 84 86 88 90 92 94

Target from Old BaseTarget from New Base

Moving Average of Nominal GDP Growth

Figure 4


55

likely changed over the past 20 years, weallow λ0 to vary over time according to anunobservable, two-state Markov process.Thus, the estimation problem includesuncovering estimates of the probabilitiesthat the GDP will be in high and low nom-inal growth target states in different timeperiods. The inferred value of the implicitgrowth target will be the probability-weighted sum of the high-growth andlow-growth parameters. For notation weuse λ0(S1) to convey that λ0 is a parametertied to state variable S1, where S1 is abinary variable representing the nominalgrowth targets; similarly, notation for thevariance s2(S2) indicates that s2 changeswith another state variable S2. To allowfor the possibility that interest-ratechanges have fatter-tailed distributionsthan the normal distribution, we allow e tohave a student-t distribution. Subject tothese Markov-switching state variables,Equation 2 becomes Equation 3.

(3) ∆mt = λ0(S1) - ∆[y - m]t\t-1+et, whereet

, student-t(mean = 0, n, s2(S2)),Variance (e) = s2 (S2)n/(n - 2), andS1 e h0,1j S2 e h0,1j.

The two Markov processes areassumed to undergo transitions betweentheir states independently from oneanother for reasons of tractability. In thiscase, the transition probabilities can besummarized as

(4) Prob.(S1t = 0 | S1t-1 = 0) = p1,Prob.(S1t = 1 | S1t-1 = 1) = q1,Prob.(S2t = 0 | S2t-1 = 0) = p2, andProb.(S2t = 1 | S2t-1 = 1) = q2.

Note that without the independenceassumption for S1 and S2, we would haveto estimate sixteen transition probabilitiesinstead of four.

Estimation Results for U.S. Data 1973-1995

Using data from the past 23 years(1973-95), we obtain two sets ofparameter estimates, one from each

measure of the adjusted base. Table 1 con-tains the parameter estimates. The mostimportant difference between the two setsof parameters is in the estimated low nom-inal GDP growth target: λ0(S1 = 1) = 7.07when the old adjusted monetary base is thepresumed policy instrument and λ0(S1 = 1)= 6.09 when the new adjusted monetarybase is used. Figure 4 (opposite page)plotsthe model-implied target or baseline rate ofnominal GDP growth, where the inferredtarget is equal to the probability-weightedsum of λ0(S1 = 0) and λ0(S1 = 1). The esti-mated target or baseline growth rates areplotted with a two-year centered movingaverage of actual nominal GDP growth.Figure 4 shows that the 6.0 percentnominal GDP growth target inferred fromthe new adjusted monetary base bettermatches the actual trend in nominal GDPgrowth since 1988, relative to the 7.0 per-cent trend inferred from the old adjustedmonetary base. This finding concurs withthe differences from Figure 2 between the


Indicator Model of Base Growth

Parameter Revised Measure Old Measure

λ0(

S1 = 0) 9.71 (.325) 10.9 (.365)

λ0(S1 = 1) 6.09 (.336) 7.06 (.346)

s2(S2 = 0) 1.60 (.043) .116 (.035)

s2(S2 = 1) .606 (.211) .623 (.235)

p1 .960 (.034) .914 (.068)

q1 .950 (.042) .943 (.046)

p2 .947 (.048) .958 (.033)

q2 .888 (.083) .903 (.073)

1/n 6E-04 9E-05 (1.5E-04)

Log-Likelihood –84.98 –76.29No. of parameters 9 9

Note: Standard errors are in parentheses.

Table 1


56

3 Because the VARs are run inlevels and the coefficients havenon-standard unit-root distribu-tions, the marginal significancelevels reported in the tables arenot exact. Nevertheless, theyare still useful for relative com-parisons between specificationsemploying the new and oldbase measures.

4 Since the St. Louis adjustedreserves series has always beenspliced across a break in 1980,it is not technically correct tosubtract nominal borrowingsfrom adjusted reserves to arriveat a measure of nonborrowedreserves. All previous studiesthat use St. Louis adjustedreserves to calculate nonbor-rowed reserves have used thesimple difference, however, andours is no exception, given thatno Reserve AdjustmentMagnitude (RAM) exists that istailored specifically to nonbor-rowed reserves.


two corresponding monetary impulse mea-sures, in which the implied nominal GDPgrowth rate is lower from approximately1989 on, when the revised measure of themonetary base is used.

IS THE ESTIMATED ROLE OF MONETARY POLICY IN VARs ROBUST TO DIFFERENT MEASURES OF THE ADJUSTED BASEAND RESERVES?

In this section, we consider both non-structural and structural VAR specificationsapplied to data from 1960 to the present.In nonstructural VARs, the ordering of thevariables matters for the inferences. Hence,when analyzing the effects of monetarypolicy shocks, the econometrician mustdecide for the entire sample whether inte-rest rates are set before the money supplyis determined (an interest-rate targetingoperating procedure) or the money supplyis exogenously set before interest rates aredetermined (a monetary targeting operatingprocedure). We begin with a comparison ofnonstructural VARs in which the interestrate is assumed to be determined before themoney supply. The comparison of interestis to run the revised and old monetary mea-sures through otherwise indentical models.We conclude the section with a structuralVAR that is designed to identify monetarypolicy shocks in a framework that allowsfor changes in operating proceduresbetween interest-rate and monetarytargeting within the sample period.

Four Variable VAR Models withFed Funds Rate Policy Shocks

We look at Sims’ classic four-variable(interest rates, money, the price level andoutput) VAR. We use monthly data inlevels.3 The federal funds rate, FF, is theinterest rate and is assumed to be deter-mined before the money supply, assuggested by Sims (1980, 1992) and Ber-nanke and Blinder (1992). The price levelis measured by the log of the consumer

price index, P, money by various variables,and output by the log of the industrial pro-duction index, Y. Among the monetaryvariables, MB stands for the monetarybase, TR for total reserves, and NBRD fornonborrowed reserves. The order of thevariables in the VAR is {FF, MB, P, Y}. Forthe lag length, we follow Sims (1987) andset it equal to one year plus one period;the extra month is added because it cansometimes capture seasonal effects not

VAR RESULTS: FOUR-VARIABLESYSTEMS WITH THE WOLD ORDERING {FF, M, P, Y}

Marginal Significance Levels for Exclusion of Lags

Equation FF M P Y

OLDMBFF .000 .698 .078 .000OLDMB .000 .000 .491 .011P .000 .007 .000 .255Y .091 .007 .094 .000NEWMBFF .000 .839 .130 .003NEWMB .000 .000 .326 .000P .000 .008 .000 .166Y .000 .004 .053 .000OLDTR FF .000 .855 .112 .000OLDTR .000 .000 .024 .005P .000 .026 .000 .675Y .125 .035 .238 .000NEWTRFF .000 .790 .140 .003NEWTR .000 .000 .014 .001P .000 .023 .000 .488Y .020 .044 .180 .000OLDNBRDFF .000 .550 .161 .002OLDNBRD .003 .000 .026 .037P .000 .010 .000 .326Y .510 .176 .427 .000NEWNBRDFF .000 .720 .151 .003NEWNBRD .001 .000 .034 .062P .000 .000 .000 .245Y .147 .408 .449 .000

NOTES: Sample period, monthly data: 1959:1 - 1996:6. Themodels have been estimated with 13 monthly lags using theSims (1992) and Bernanke and Blinder (1992) {FF, M, P, Y}ordering. Low values imply strong marginal predictive power.Variance decompositions are for a five-year horizon.

Table 2


57


removed by seasonal adjustment ofthe data.

Table 2 reports the marginal signifi-cance levels of Granger causality F-testsfor six different measures of money: old andrevised monetary base; old and revised totalreserves; old and revised non-borrowedreserves. In particular, p-values are for thenull hypothesis that lags of a particularright-hand-side variable (indicated in thecolumn heading) can be excluded fromone of the system’s equations (indicated in the row heading). The variance decom-positions show percentages of the five-year forecast-error variance of a variableexplained by its own shocks versus shocksto the other variables.

The marginal significance levels inTable 2 indicate that lags of the fed fundsrate do not significantly influence indus-trial production in the systems wheremoney is measured by the old monetarybase or total reserves, but they do have asignificant impact in the system with thenew monetary base and new total reserves.In the system with nonborrowed reserves,lags of the funds rate do not significantlyaffect output when either measure is used.4

Thus, the evidence regarding whethermonetary policy shocks significantly affectoutput varies according to the measure ofmoney that is used. One possible expla-nation for the contradictory conclusionsreached about the predictive power of fedfunds rate shocks across the different mea-sures of money is that Granger causalitytests are sensitive to the nonorthogonalitybetween the right-hand-side variables (see,for example, Bernanke and Blinder, 1992,for a discussion of this issue). In thisregard, it should be noted that the fedfunds rate correlates more with the baseand total reserves than with nonborrowedreserves:

ρ(FFt, OLDNBRDt) = .282 andρ(FFt, NEWNBRDt) = .248.

The forecast error variance decomposi-tions in Table 3 show that innovations inmoney (independently of how it is mea-sured) explain a very small percentage of

the variance of output. In contrast, inno-vations in the fed funds rate explain a veryhigh percentage of the variance of indus-trial production, especially when money ismeasured by the base, about 56 percent inthe case of NEWMB. Replacing NEWMBwith NEWTR, however, reduces thisstatistic to 37 percent and to 29 percentwhen NEWNBRD is used. Moreover, in thevariance decompositions of Table 3, thenew measures of money perform generallythe same as the corresponding old ones.

Solid lines in Figures 5-7 (followingpages) show the impulse response func-

VAR RESULTS: FOUR-VARIABLE SYSTEMS WITH THE WOLD ORDERING {FF, M, P, Y}

Forecast Error Variance Decompositions (60-Month Horizon)

Equation FF M P YOLDMBFF 49.947 1.366 16.367 32.318OLDMB 62.058 35.216 0.069 2.655P 8.327 1.930 72.761 16.981Y 59.762 0.395 7.167 32.674NEWMBFF 49.518 3.046 19.843 27.592NEWMB 54.139 37.076 2.080 6.703 P 10.629 1.720 72.393 15.255Y 56.539 0.556 6.549 36.355OLDTR FF 45.634 2.134 10.839 41.391OLDTR 60.444 28.465 3.196 7.896P 4.608 3.088 66.331 25.971Y 39.888 6.984 6.075 47.051NEWTRFF 38.009 8.895 17.483 35.611NEWTR 34.235 36.913 26.888 1.962P 4.506 1.729 70.959 22.804Y 36.519 9.137 7.920 46.421OLDNBRDFF 45.186 2.284 8.159 44.369OLDNBRD 59.438 30.656 0.821 9.082P 4.209 0.730 62.210 32.849Y 31.155 2.754 12.827 53.262NEWNBRDFF 36.202 8.672 12.462 42.662NEWNBRD 32.935 43.716 14.916 8.431P 3.456 1.438 65.846 29.258Y 28.662 4.382 13.312 53.643

Table 3

tions of each of the four variables (interestrate, money, prices, and output) to each ofthe six measures of money. Responses(based on orthogonalized innovations withthe ordering as shown in the charts) areplotted over a horizon of five years. Dashedlines denote one standard deviation band,computed according to the Monte Carlomethod described in Doan (1992, Example10.1) with 500 draws from the posteriordistribution of the VAR coefficients andthe covariance matrix of the innovations.5

In general, the qualitative responsespictured in Figures 5-7 do not differ sub-stantially between the old and revisedmoney series. Minor differences are that,in Figure 5, the response of output to ashock in the old base measure is persis-tently positive, whereas the revised basemeasure predicts negative effects of a mon-etary shock on output at long horizons.However, the responses are not signifi-cantly different from zero in either case.In Figure 7, we see a more statistically sig-nificant liquidity effect at longer horizonsusing the revised nonborrowed reservesseries. Moreover, the decrease in the fed-eral funds rate from a positive non-borrowed reserves shock is persistent,without the sign changes seen from the old nonborrowed reserves series.

Permutations of the NonstructuralVAR Approach to IdentifyingMonetary Policy Shocks

We also considered four alternativenonstructural VAR specifications from theliterature, such as Eichenbaum’s (1992) re-ordering of the four-variable model: (P, Y,NBRD, FF). Sims (1992) also suggested thatsensitive commodities prices could captureinformation on future inflationary pressurebeyond that embodied in the consumerprice index. Christiano, Eichenbaum andEvans (1995) estimated orderings of VARswith commodities prices that corre-sponded with monetary targeting andinterest-rate targeting. Strongin (1995)proposed that innovations to the ratiobetween the supply of nonborrowedreserves to the total demand for reserves

(NBRX) could represent monetary policyshocks. In each case, however, we foundthat it did not make a substantive differ-ence whether we used the old series or therevised one in these VARs. One exceptionmight be the Strongin (1995) model,where for the old measures, TR accountsfor about 9 percent of the variance of out-put and NBRX for about 45 percent. Withthe revised series, these statistics become16 percent and 37 percent, respectively, sothere is some sensitivity to changes in themoney measure. On the other hand,NBRX explains about 10 percent of thevariance of FF at the 5-year horizon, irre-spective of whether the old or the newseries is used.

Bernanke and Mihov’s StructuralVAR and Changes in OperatingProcedures

In the nonstructural models discussedabove, the VAR framework forces the ana-lyst to decide for the entire sample periodwhether monetary policy was conductedthrough interest-rate targeting or monetarycontrol. In contrast, the Bernanke andMihov (1995) model of monetary policyshocks allows the monetary policy oper-ating procedure to switch over timebetween interest-rate targeting and mone-tary control. In their structural VAR,Bernanke and Mihov identify a monetarypolicy shock that is a linear combinationof the forecast errors of total reserves, non-borrowed reserves and the fed funds rate.Depending on the operating procedure,the monetary policy shock inferred fromthe model can be a function of one ormore of these forecast errors. Given thatthe Fed changed its operating proceduretemporarily and controlled growth in non-borrowed reserves from 1979 to 1982 (aperiod that falls in the middle of our sam-ple), the Bernanke-Mihov allowance forchanges in operating procedure seemsappropriate. For space considerations, wedo not discuss the quantitative derivationof the monetary policy shock in theBernanke-Mihov model. Full details areavailable in Bernanke-Mihov (1995).


58


5 Ninety percent confidence inter-vals have become the norm foranalyzing VAR output.


59


Shock to OLDMB

1 10 19 28 37 46 55

0.018

0.012

0.006

-0.000

-0.006

Response of P

0.0075

0.0050

0.0025

0.0000

-0.00251 10 19 28 37 46 55

Response of NEWMB0.0075

0.0050

0.0025

0.0000

-0.00251 10 19 28 37 46 55

Response of OLDMB

0.300.200.100.00

-0.10-0.20-0.30

1 10 19 28 37 46 55

Response of FF

Shock to NEWMB

Response of FF0.300.200.100.00

-0.10-0.20-0.30

1 10 19 28 37 46 55

1 10 19 28 37 46 55

0.018

0.012

0.006

-0.000

-0.006

Response of P

1 10 19 28 37 46 55

0.012

0.008

0.004

0.000

-0.004

-0.008

Response of Y

1 10 19 28 37 46 55

0.012

0.008

0.004

0.000

-0.004

-0.008

Response of Y

Month Month

Month Month

Month Month

Month Month

Figure 5

Impulse Responses, {FF, MB, P, Y} ModelMonthly data: 1959:1–1995:6


60


Shock to OLDTR

Month Month

Month Month

Month Month

Month Month

Shock to NEWTR

0.300.200.100.00

-0.10-0.20-0.30

1 10 19 28 37 46 55

Response of FF

1 10 19 28 37 46 55

0.0150

0.0100

0.0050

0.0000

-0.0050

Response of NEWTR

1 10 19 28 37 46 55

0.01000.00750.00500.00250.0000

-0.0025-0.0050

Response of OLDTR

1 10 19 28 37 46 55

0.010

0.005

0.000

-0.005

-0.010

Response of P

1 10 19 28 37 46 55

0.005

0.000

-0.005

-0.010

-0.015

Response of Y

1 10 19 28 37 46 55

0.005

0.000

-0.005

-0.010

-0.015

Response of Y

1 10 19 28 37 46 55

0.005

0.000

-0.005

-0.010

Response of P

1 10 19 28 37 46 55

0.300.200.10

-0.00-0.10-0.20-0.30

Response of FF

Figure 6

Impulse Responses, {FF, TR, P, Y} ModelMonthly data: 1959:1–1995:6


61


Shock to OLDNBRD

0.020

0.015

0.010

0.005

0.000

-0.005

1 10 19 28 37 46 55

0.005

0.000

-0.005

-0.010

0.0050

0.0025

0.0000

-0.0025

-0.0050

-0.0075

Response of P

0.012

0.008

0.004

0.000

-0.004

Response of Y

0.005

0.000

-0.005

-0.010

Response of P

0.200.10

-0.00-0.10-0.20-0.30-0.40

Shock to NEWNBRD

Response of FF0.300.200.100.00

-0.10-0.20-0.30

1 10 19 28 37 46 55

1 10 19 28 37 46 55

Response of OLDNBRD

1 10 19 28 37 46 55

0.005

0.000

-0.005

-0.010

1 10 19 28 37 46 55

1 10 19 28 37 46 55

1 10 19 28 37 46 55

Response of NEWNBRD

1 10 19 28 37 46 55

Response of FF

Month Month

Month Month

Month Month

Month Month

Response of Y

Figure 7

Impulse Responses, {FF, NBRD, P, Y} ModelMonthly data: 1959:1–1995:6

The model contains five variables(including commodities prices denotedPCOM) and is structural, so that the orderof the variables is no longer key to identifi-cation. Following Bernanke and Mihov(1995), we present impulse response func-tions to a “typical” monetary policy shock.Figure 8 contains plots of the impulse

response functions using the old definitionsof total reserves and nonborrowed reserves.Unlike some of the simpler VAR models(see Figure 5, for example), in this modelthe price level rises in response to a positivemonetary policy shock, although the 90percent confidence interval still containszero. Output increases for about three years



62

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

Response of GDP

0

0.0080.0060.0040.0020.000

-0.002-0.004-0.006

Response of PGDP

0

0.008

0.006

0.004

0.002

0.000

-0.002

Response of PCOM

0

0.020

0.015

0.010

0.005

0.000

-0.005

-0.010

Response of TR

0

0.004

0.002

0.000

-0.002

-0.004

-0.006

-0.008

Response of NBR

0

0.0080.0060.0040.0020.000

-0.002-0.004-0.006-0.008

4 8 12 16 20 24 28 32 36 40 44

Response of FFR

0

0.60.40.20.00.2

-0.4-0.6-0.8

Figure 8

Bernanke-Mihov Impulse Responses to a MonetaryPolicy Shock (Old Reserves Measure)

in response to a positive monetary policyshock and commodity prices respondweakly. Figure 9 has the correspondingcharts for the revised measures of adjustedreserves. The most noteworthy differencebetween results for the old and revisedseries is that with the revised series, the 90percent confidence interval for the response

of the price level (denoted PGDP for GDPdeflator) to an expansionary monetarypolicy shock is unambiguously positive atlong horizons (after about 30 months).Otherwise, the response patterns for out-put and commodity prices are similar tothose derived from the old measures ofadjusted reserves.



63

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

4 8 12 16 20 24 28 32 36 40 44

Response of GDP

0

0.0080.0060.0040.0020.000

-0.002-0.004-0.006

Response of PGDP

0

0.008

0.006

0.004

0.002

0.000

-0.002

Response of PCOM

0

0.020

0.015

0.010

0.005

0.000

-0.005

-0.010

Response of TR

0

0.004

0.002

0.000

-0.002

-0.004

-0.006

-0.008

Response of NBR

0

0.010

0.005

0.000

-0.005

-0.010

Response of FFR

0

0.6

0.4

0.2

0.0

0.2

-0.4

-0.6

Month Month

Month Month

Month Month

Figure 9

Bernanke-Mihov Impulse Responses to a MonetaryPolicy Shock (New Reserves Measure)

6 We estimate only their bench-mark “monetarist” model andnot their other “Keynesian” or“supply-shock” models.



64

OPEN MARKET OPERATIONS AND RESERVE REQUIREMENTCHANGES AS MONETARYPOLICY INDICATORS

Haslag and Hein’s Search forDifferential Impacts

An index of monetary policy actionssuch as the adjusted monetary baseassumes that open-market purchases,which increase the supply of bankreserves, are economically equivalent toreserve-requirement reductions that “liber-ate” an equal amount of reserves by reduc-ing the demand for reserves. Haslag andHein (1995) have recently examined thisassumption in their article, “Does it matterhow monetary policy is implemented?”They look for differential impacts ofreserve-requirement shocks and source-base shocks in the dynamic responses of

output and inflation. For this reason, theyuse structural vector autoregressions toidentify the shocks and study the impulseresponses. We re-examine their structuralVAR to see if the revised adjusted mone-tary base provides new and different evi-dence regarding this question.6 One pieceof analysis that we add to Haslag and Heinis some judgment regarding the model-implied variance decompositions.

Given that the adjusted monetary baseequals the source base plus the reserveadjustment magnitude (AMB = H + RAM),Haslag and Hein separate changes in thesource base and changes in the reserveadjustment magnitude (RAM) that com-prise the adjusted base into two station-ary components:

Ht = ∆Ht /[(AMBt + AMBt-1)/2] and

RAM^t = ∆RAMt /[(AMBt + AMBt-1})/2].

They include these two variables in athird-order VAR with the M2 money multi-plier (mm2), inflation and real GDP. Themoney multiplier is included becausehigh-powered money might respond tochanges in the multiplier. The VAR resid-uals (u) are decomposed into orthogonal“structural” residuals (z) through a set ofcontemporaneous restrictions:

(5) uRAM = z1,

uH = b1 uRAM + b2 umm2 + b3 uINF + z2,

umm2 = b4 uINF + b5 uGDP + z3,

uINF = b6 uRAM + b7 uH + z4, and

uGDP = b8 uRAM + b9 uH + b10 umm2 + z5.

Using quarterly data from 1980-95, weobtained parameter estimates (shown inTable 4) for the just identified model, byapplying both the revised and the oldadjusted monetary bases to measure thesource base and RAM. More important thanthe contemporaneous relations captured bythe coefficients, however, are the impulseresponses and variance decompositions,especially the way in which output andinflation respond to innovations in H andRAM. Figure 10 provides the impulseresponses of output and inflation to reserve-

Structural Decompositionof VAR Residuals (Revised and Old Definitions ofAdjusted Monetary Base for 1980-95)

Coefficient Revised data Old data

b1 .743 .705

b2 .501 .406

b3 .675 1.31

b4 –5.55 –4.10

b5 5.49 4.61

b6 –.061 –.295

b7 –.163 –.434

b8 .048 –.206

b9 .933 .145

b10 –1.42 –1.89

sz1 .569 .602

sz2 .413 .463

sz3 3.78 3.00

sz4 .219 .248

sz5 1.12 1.28

Table 4

7 Confidence intervals for theimpulse responses were notderived since they requiredMonte Carlo simulations of re-estimations of the struc-tural decompositions, whichproved to be fragile in their convergence.



65

requirement and source-base shocks.Haslag and Hein (1995) look for evidencethat the two monetary-policy innovationsgenerate similar impulse responses. Figures10a and 10b report results for GDPresponses, using the revised and old datadefinitions. In each case, output growth ishigher for about four quarters before return-ing to zero or turning slightly negative.7

This finding is qualitatively very similar tothat of Haslag and Hein. The estimatedimpulse response to a required reservesshock, however, is quite volatile and appar-ently imprecisely estimated, as one wouldexpect, since reserve requirement changesare infrequent. Nevertheless, the resultsconcur with those found by Haslag andHein, using either the revised or old defini-tions of the monetary base, in that theoutput response paths are qualitatively similar for both source-base and required-

reserve shocks. With respect to inflation,the revised monetary base data suggest dif-ferent impacts from source-base shocks andrequired-reserve shocks in Figure 10c. Apositive shock to source-base growth has apersistently positive effect on inflation,whereas a required-reserve shock has asomewhat negative (and almost certainlyinsignificant) impact on inflation. The oldmonetary base data give a similar picture butfind a much greater magnitude than therevised data for the response of inflation to a source-base shock in Figure 10d.

Haslag and Hein do not report variancedecompositions, but they are importantdiagnostics for a structural model, because ithelps to know how often these impulses arestriking and how large they are. The firsttwo rows of Table 5 contain the forecastvariance decomposition for output for 16-quarter forecasts. This decomposition is

2 4 6 8 10 12 142 4 6 8 10 12 14

2 4 6 8 10 12 14

a. GDP New b. GDP Old

0.16

0.12

0.08

0.04

0.00

-0.04

-0.080

^RAM

d. Inflation Old

^ H

c. Inflation New0.16

0.12

0.08

0.04

0.00

-0.04

-0.080

^RAM

^ H

0.250.200.150.100.050.00

-0.05-0.10

0

^RAM

^ H

2 4 6 8 10 12 14

0.250.200.150.100.050.00

-0.05-0.10

0

^RAM

^ H

Month

Month Month

Month

Figure 10

Responses of GDP and Inflation(post-1980) to RA^M and H

^



66

representative for all horizons. We find noreal distinction between the revised defini-tion and the old definition of the adjustedmonetary base regarding the roles played bysource-base and required-reserve shocks indetermining the variance of output. Thedecomposition of the inflation forecast error,given in the last two rows of Table 5, showsthat the old definition gives a much largerrole to source-base shocks. The revised defi-nition, in contrast, attributes more of thevariation in inflation to persistent, idiosyn-cratic shocks to inflation itself. Hence ourmain finding in using the revised definitionversus the old one is that, according to therevised measure, inflation appears lessdriven by innovations in high-poweredmoney growth than the old adjusted mone-tary base would suggest.

The Impact of Reserve RequirementChanges on Output and Investment

For empirical evidence that theamount of financial intermediation is animportant determinant of output andinvestment, Loungani and Rush (1995)look to reserve requirement changes asexogenous shifts in the effective tax rateon intermediation. They note that othermeasures of intermediation are simulta-neously determined with real output,whereas reserve requirement changes aregenerally made for technical reasons notassociated with the business cycle. Their

hypothesis is that decreases in reserverequirements lower the tax on intermedia-tion and act as a positive shock to banklending and real economic activity.

Loungani and Rush (1995) obtain a sta-tionary measure of the magnitude of thereserve requirement changes by taking logdifferences of the ratio between the adjustedand unadjusted monetary base. Theydenote the log of this ratio as F and its firstdifference as DF. The unadjusted base isthe source base from the Fed’s balancesheet, and the adjusted base adds a reserveadjustment magnitude (RAM) to couch theimpact of reserve requirement changes interms of an equivalent change in high-pow-ered money. Loungani and Rush also usean alternative measure of DF based on theratio of adjusted to unadjusted reserves toabstract from the large role that currencyplays in the monetary base.

The basic empirical proposition ofLoungani and Rush is that reserverequirement changes, summarized by DF,have significant explanatory power inreduced-form OLS regressions of outputgrowth and investment growth. Our pur-pose is to run the Loungani and Rushregressions using both the revised (new)and old definitions of the adjusted mone-tary base and adjusted reserves to see ifthe conclusions regarding the potency ofmonetary policy change.

In their regressions, Loungani andRush include other conditioning variablesin addition to DF. They find little sensi-tivity to four alternative measures ofmonetary policy. We follow their use ofthe change in M1 growth as a conditioningvariable. Twice differencing the log of M1ensures that a change in the money growthrate does not permanently affect the out-put growth rate. They also condition onthe real stock return, the change in theStandard and Poor’s 500 index, as a mea-sure of changes in real wealth. Lags ofthese conditioning variables are includedin the regression, because it may take timefor real economic activity to respond toshocks. We follow Loungani and Rush byusing eight lags of each explanatory vari-able, but we do not use the contempora-

Output and Inflation VarianceDecompositions for 16-QuarterForecast Errors (1980-95)

Percent of Variance by Innovation Source (Rows Sum to 100)

RAM Source base M2 mult. Inflation GDPOutput:

Revised Base 12.6 3.3 62.6 10.7 10.9

Old Base 15.3 6.0 54.5 10.1 14.1

Inflation:

Revised Base 8.3 23.6 1.9 62.3 3.9

Old Base 6.4 47.7 2.2 36.5 7.2

Table 5



67

neous (zero lag) values. One might arguethat the explanatory variables are exoge-nous vis-a-vis today’s output, but we choseto include only lagged values.

Denoting (in logs) output as Y, invest-ment as I, and stock prices as S, and thedifferencing operator as D, the regressionspecification is

(6) DYt = b + pDYt-1 + S8

k=1 dk DFt-k +

S8

k=1Uk DDM1t-k + S

8

k=1λk DSt-k,

where DF is either ∆ln(AMB/MB) or∆ln(ARES/RES). AMB is the adjusted mon-etary base, MB is the unadjusted monetarybase, ARES are adjusted reserves, and RESare unadjusted bank reserves. Anotherregression is run with investment spendingas the dependent variable. As discussedabove, each of the regressions is run twice,once using the revised (new) definition ofthe adjusted base and reserves, and onceusing the old definitions.

Loungani and Rush (1995) measurethe potency of monetary shocks by look-ing at the statistical significance of thecumulative effect of a reserve-requirementchange on output and investment. Thus akey hypothesis test is whether the sum ofdk coefficients from Equation 6 is signi-ficant. Following Loungani and Rush, wereport the sum of the coefficients on DFthrough both four and eight lags and theprobability values for the hypothesis thatthe sums are zero in quarterly data from1962 to 1995.

The first two columns of Table 6 con-tain results for the base measures of DF,with results for the old and revised (new)measures of the base side by side. The pri-mary difference in results between the oldand revised measures of the adjusted baseis that reserve-requirement shocks do nothave a significant effect on output at the 5percent level when the old measure isused, but are significant at 2 percent to 3percent when the revised measure isapplied. The results for investment acti-vity, on the other hand, do not appear

sensitive to which measure of the adjustedmonetary base is used.

The last two columns of Table 6provide the same tests using changes inthe ratio of adjusted to adjusted reserves as the measure of reserve-requirementchanges. The significance of the effect ofreserve-requirement shocks on output dif-fers between the old definition and therevised one even more dramatically forreserves than for the base. For the oldadjusted monetary base, the sum of coeffi-cients on DF was significant at 10 percent,but is not even close for the old adjusted-reserves measure. Meanwhile, the revisedadjusted reserves measure is significant inthe output equation. As with the corre-sponding monetary base measures, the olddefinition shows little difference in signifi-cance levels from the revised definitionwith respect to the investment regression.Nevertheless, the fact that the impact isstatistically significant with either measurereinforces and adds robustness to theLoungani-Rush finding that exogenousreserve-requirement shocks matter.

Effects of Adjusted Base and Adjusted ReservesShocks on Output and Investment Growth

Sum of coefficients and

p-values

old AMB revised AMB old ARES revised ARESOutput: DDM1 –.498 –.443 .082 –.029

(.35) (.39) (.26) (.80)DS 0.20 .019 0.20 .018

(.025) (.036) (.027) (.045)DF(4) .405 .559 .055 .127

(.082) (.021) (.34) (.041)DF(8) .527 .695 .082 .166

(.088) (.032) (.26) (.035)

Investment: DDM1 –1.22 –.715 .550 .548(.52) (.78) (.34) (.34)

DS .118 .107 .114 .103(.007) (.013) (.008) (.018)

DF(4) 2.84 3.15 .600 .749(.017) (.010) (.039) (.018)

DF(8) 4.53 4.63 .954 1.16(.004) (.005) (.010) (.004)

Table 6

Based on their finding that reserve-requirement shocks affect investment morestrongly than output, Loungani and Rushconclude their paper with a regression ofinvestment on the ratio of bank credit tototal credit. Their focus on the investmentchannel stems from their finding differentialsignificance levels for reserve-requirementshocks between the investment and outputregressions. Our findings suggest that theimpact of reserve-requirement shocksmight be significant to both measures ofeconomic activity if one uses the revisedmeasures of adjusted reserves and adjustedmonetary base.

CONCLUSIONSOur investigation has identified

several empirical results and conclusionsthat are sensitive to the choice between theold and revised measures of the adjustedmonetary base and reserves. Our esti-mates suggest that the forecast errorvariances for base velocity are smaller forthe revised monetary base from approx-imately 1985 to the present, a finding thatleads to different inferences regarding thetarget rate of nominal GDP growth. In theVARs, which are designed to measure theeffects of monetary policy shocks, thequalitative results are not sensitive to thechoice of base and reserves measures,although the variance decompositionsdiffer somewhat. The structural VAR fromBernanke and Mihov (1995) has its chiefdiscrepancy in the inflation variancedecomposition, where the old measuregives relatively more weight to source-baseshocks. Finally, because of their focus onreserve-requirement shocks, the “creditview” regressions of Loungani and Rush(1995) offer a good place to distinguishbetween the two base measures. The alter-native base measures give differentanswers regarding the significance ofreserve-requirement shocks on output.While reserve-requirement shocks are animportant determinant of investment forboth measures, only the revised base mea-sure suggests that reserve-requirementshocks significantly influence output.

REFERENCESAnderson, Richard G., and Robert H. Rasche. “Measuring the Adjusted

Monetary Base in an Era of Financial Change,” this issue of the

Review, pp. 3-37

Bernanke, Ben S., and Alan S. Blinder. “The Federal Funds Rate and theChannels of Monetary Transmission,” The American Economic Review(September 1992), pp. 901-21.

_____ and Ilian Mihov. “Measuring Monetary Policy,” NBER WorkingPaper (1995).

Christiano, Lawrence J., and Martin Eichenbaum. “Identification andthe Liquidity Effect of Monetary Shocks,” Business Cycles, Growth andthe Political Economy, A. Cuikerman, L.Z. Hercowitz, and L.Leiderman, eds., MIT Press (1992), pp. 335-70.

_____, _____, and Charles Evans. “The Effects of Monetary Policy Shocks: Some Evidence from the Flow of Funds,” Review ofEconomics and Statistics (February1996), pp. 16-34.

Doan, Thomas A. RATS User’s Manual Version 4, Estima (1995).

Dueker, Michael. “Indicators of Monetary Policy: The View from Implicit Feedback Rules,” this Review (September/October 1993),pp. 23-40.

_____ and Andreas M. Fischer. “Inflation Targeting in a Small OpenEconomy: Empirical Results for Switzerland.” Journal of MonetaryEconomics (May 1996), pp. 89-103.

Eichenbaum, Martin. “Comments” on “Interpreting the MacroeconomicTime Series Facts: The Effects of Monetary Policy” by C. Sims,European Economic Review (June 1992), pp. 1001-11.

_____ and Charles L. Evans. “Some Empirical Evidence on the Effectsof Shocks to Monetary Policy on Exchange Rates,” Quarterly Journalof Economics (November 1995), pp. 975-1009.

Haslag, Joseph H., and Scott E. Hein. “Does It Matter how MonetaryPolicy is Implemented?” Journal of Monetary Economics (April1995), pp. 359-86.

Loungani, Prakash, and Mark Rush. “The Effect of Changes in ReserveRequirements on Investment and GNP,” Journal of Money, Credit andBanking (May 1995), pp. 511-26.

_____ “The Case For Rules in the Conduct of Monetary Policy: AConcrete Example,” Economic Review, Federal Reserve Bank ofRichmond (September/October 1987), pp. 10-18.

McCallum, Bennett T. “Robustness Properties of a Rule for MonetaryPolicy,” Carnegie-Rochester Conference Series on Public Policy(1988), pp. 173-203.

_____ and Monica Hargraves. “A Monetary Impulse Measure forMedium-Term Policy Analysis.” IMF Staff Studies for the WorldEconomic Outlook (1995), pp. 52-69.

Pagan, Adrian R., and John C. Robertson. “Resolving the LiquidityEffect,” this Review (May/June 1995), pp. 33-54.

Sims, Christopher A. “Macroeconomics and Reality,” Econometrica(June 1980), pp. 1-48.



68

_____. “Interpreting the Macroeconomic Time Series Facts: The Effects of Monetary Policy,” European Economic Review (June 1992),pp. 975-1000.

Stock, J.H., and M.W. Watson. “Testing for Common Trends.” Journal of the American Statistical Association (December 1988), pp. 1097-107.

Strongin, Steven. “The Identification of Monetary Policy Disturbances:Explaining the Liquidity Puzzle,” Journal of Monetary Economics(August 1995), pp. 463-97.



69

The Sensitivity of Empirical Studies to Alternative …...The Sensitivity of Empirical Studies to...

Documents

Transcript of The Sensitivity of Empirical Studies to Alternative …...The Sensitivity of Empirical Studies to...