ROBERT J. GORDON NorthwesternUniversity
Can the Inflalon of the 1970s be Explained?
BY MANY STANDARDS inflationhas been a "surprise"during the past six years. Errorsin forecastinginflation have increasedmarkedly compared with earlierperiods. For instance,during the interval 1971:3 to 1975:4 the root mean-squareerror of the Livingstonpanel of economistsin fore- castingthe consumerprice index six months ahead was 3.5 percentage points at an annualrate, comparedwith an errorof 1.6 percentagepoints over the previousseventeen years.' Not only did the panel forecastersfail to predictthe increasedvariance of the inflationrate in the 1970s, but also they fell far shortin predictingthe cumulativetotal price changebetween 1971 and 1976-24.0 percentcompared with the actual change of 34.0 percent.2Most of the error occurredduring the four quartersof 1974, Note: This researchhas been supportedby the National Science Foundation. I am grateful to my research assistant, Joseph Peek, for his superb efficiency in compiling and creating the complex data base on which the paper depends. Helpful suggestions were received from participantsin seminars at Northwestern,the University of Cali- fornia at Berkeley, and the Federal Reserve Banks of San Francisco and Philadelphia. 1. The Livingston forecasts were obtained from John A. Carlson, "A Study of Price Forecasts,"Annals of Economic and Social Measurement,vol. 6 (Winter 1977), table 1, pp. 33-34. I calculated the errorsby comparingthe six-month-aheadforecasts with the change in the consumer price index in the two relevant quarters. For in- stance, Carlson's calculation of the predicted quarterly rate of change between December 1973 and June 1974 is comparedwith the average quarterlyrate of change of the CPI in the first and second quartersof 1974. The "previousseventeen years" runs from 1954:1 to 1971:2. 2. The actual figure refers to the sum of the quarterlyrates of change of the CPI in the interval 1971:3 through 1976:2. The forecast figure is the sum of the six- month predicted changes calculated by Carlson from the Livingston panel data for the ten surveys between June 1971 and December 1975. 253 254 Brookings Papers on Economic Activity, 1:1977 with an actualincrease of 11.6 percent,almost twice the 6.0 percentin- creaseforecast six monthsin advance.3 In searchingfor an explanationfor this inflation,this paper can be likened to an investigativereport following a railroador airline crash. The news of the disaster-in this case, the failure to forecast inflation accurately-was reportedlong ago and by now is well known. But what can we say beyondthe fact that the disasteroccurred? Just as transporta- tion investigationsattempt to determinewhich specificparts of the ma- chinefailed, and to recommendimprovements, so here the relationshipof the inflationrate to other importanteconomic variables is studiedto de- termineas preciselyas possible what was differentabout the experience of the 1970s, and what lessons can be learnedfrom past mistakes.Which theories and structuralrelationships relevant for predicting inflation remainintact, and which require surgeryor euthanasia?What are the implicationsfor policy? Most econometricmodels base their inflationforecasts on structural price and wage equations,either a single pair for the aggregateeconomy, or a largerset of disaggregatedequations. In my own past work on infla- tion, I have specifiedand estimatedaggregate price and wage equations, and have studiedthe sensitivityof the resultsto alternativespecifications, estimationmethods, and sampleperiods. This paperinvestigates the per- formanceof my price-wagemodel in trackingthe inflationof the 1970s, and studies the implicationsof its successes and failuresfor the future conduct of economic policy. The paperis dividedinto threesections, one on the priceequation, one on the wage equation,and one on dynamicsimulations in which the two equationsinteract. 1. Structuralprice equation. An equationthat explains price change with wage changeas a predeterminedvariable is a componentof almost all large-scaleeconometric models of the U.S. economy. In a previous paperI arguedthat the total increasein prices relativeto wages between mid-1971 and late 1975 was almost exactly what would have been pre- dicted by a structuralprice equationfitted to the 1954-71 period, and 3. The errors for the forecasts from five large-scale models compiled by McNees were similar. The four-quarter-aheadforecast made in 1973:4 for the change in the GNP deflator to 1974:4 was 6.04 percent; the actual was 11.04 percent. See the re- vised reprintof Stephen K. McNees, "An Evaluation of Economic Forecasts: Exten- sion and Update," New England Economic Review (September/October 1976), pp. 30-44. Robert J. Gordon 255 that the timing of postsampleerrors was consistentwith the hypothesis that prices had been held down by controls in 1971-72 and then re- boundedwhen controls were terminated in 1974.4This paperextends this test throughthe end of 1976, notes the effectsof recentdata revisionson the originalprice equation,and explores alternativeexplanations of its overpredictionof pricechange in 1975 and 1976. 2. Structuralwage equation. Can a wage equationspecified in 1971 and estimatedfor pre-1971 data explain the behavior of wage change since 1971? Whatwas the impact of 1973-74 "supplyshocks" on wage change, and how should policy respond to future supply shocks?5Has high unemploymentduring 1975 and 1976 held down wage increasesby moreor less thanwould have been expectedon the basisof pre-1971rela- tionships?Finally, can the pre-1971 data or the 1971-76 experience distinguishamong the variousproxies for labor markettightness used by differenteconometric investigators?" 3. Dynamic simulations.How potent are high unemploymentand a slack economyin slowingthe inflationrate? What would have been the consequencesfor inflationof an alternativeexpansionary policy in 1974? Is the Carteradministration's planned economic recovery consistent with its goal of deceleratinginflation? A dynamicsimulation in whichthe price and wage equationsinteract can provideanswers to these questions.
Behaviorof the Main Variables,1969-76
Table 1 displaysthe behaviorover the 1969-76 period of severalim- portantmeasures of changesin prices, wages, money, and nominal de- mand.The figuresare annualrates of change.The firstcolumn covers the ten quartersprior to the impositionof the controlsprogram in 1971, the second column covers the two quartersinfluenced by the 1971 freeze, and the next five columnsshow for the five years 1972-76 the sum of the quarterlyrates of changefor the four quartersof each year. The official price indexes displayedin the first four lines uniformly 4. Robert J. Gordon, "The Impact of Aggregate Demand on Prices," BPEA, 3:1975, pp. 613-62. 5. See Robert J. Gordon, "AlternativeResponses of Policy to External Supply Shocks,"BPEA, 1:1975, pp. 183-204. 6. Robert J. Gordon, "Inflationin Recession and Recovery," BPEA, 1:1971, pp. 105-58. _~~~~~~~~~~~~~~~~~~~~~~~~ ^ 4> - N 9C2. OM
0\ cn enen en cn tn e u AY ~~~~~~~~~~~~~~~~~~~~~~~~~~~~tNXNoW A=3- ** so m W o Oo o mt looS.U e~~~~~~2 o
$:~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~C^ s .5- 1 !~~~~~~~~~~~~~t., eq 0 N o N W- o C tr .4 V
3 > >, t to~caN m 0t r-o W W N tn t_ D EX
> X A. > B~~~~~~~~~~~~~~~~~~~~~~~~~~~~C C; o- C;X n ~~~~~~~~~~~~~~~~~~~~~~~~~~~OX ^XXOWbX. l'
4-- a Xt>% O"W4"0 bb X -- OF x PC C to C3E -
| t > Cmt N ^ q X,I m m n W | aa
W~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~CWI J rB 9. o e X~~~~~~~~~~~~~CIR:l tn C m \ t N OSaN bWootQ cl ,, m$>EadEIc ? > * 1c ?c t?- g 0
SE St Q O r w n @ ar =~~~~~~~~~~~~~~~~~~o8 g aX S
, WE i E*6 :
StructuralPrice Equations
In an earlier paper I estimated structuralprice equations that ex- hibited relatively strong effects of aggregate demand on the price 7. Gordon, "Impactof Aggregate Demand,"pp. 622-29, 656-60. 258 BrookingsPapers on Economic Activity, 1:1977 "markup,"that is, on the relationshipof the aggregateprice level to the aggregatewage level. These equations appeared able to explain the cumulative1971-75 inflationusing coefficients estimated through 1971:2. Althoughthe postsampleprediction errors were large, their timing was consistentwith the interpretationthat the controlshad temporarilyheld down the pricelevel. In table 2, the firstcolumn lists the coefficientsof a versionof the "core"equation as publishedin 1975.8 The specificationof the various price equationspresented in table 2 correspondsto thatderived in my 1975 paper.The pricelevel net of excise and sales taxes is markedup over total cost by a marginthat dependson the level of excess demandfor commodities.Total cost in turnconsists of unit labor cost, materialsprices, and the user cost of capital.After each variableis transformedinto a percentagerate of change,and when tech- nicalchange is assumedto be labor-augmenting,an equationis derivedin which the rate of changeof pricesdepends on each of the variableslisted in table 2: (1) the rate of changeof an excise-taxterm; (2) the rate of changeof the relativeprice of materials;(3) the deviationof the growth rate of actualproductivity from its trend;(4) the rate of changeof wages minusthe trendgrowth rate of productivity-"trendunit laborcost"; (5) the rate of changeof the relativeprice of capitalgoods; and (6) a proxy for the excess demandfor commodities,either the rate of change of the ratioof unfiliedorders to capacity(UFO/C), or the rate of changeof the gapbetween actual and potential output. While in the earlier paper equations including the two alternative proxieswere essentially identical, the samecannot be said of the equations reestimatedwith new data from the 1976 revisionof the nationalincome accounts.The data revisions reduce the statisticalsignificance of most variables when either demand proxy is used, but the version using UFO/C is affectedmost adversely(compare columns 2 and 3). The out- put-gapequation is superioron almostevery count, with a lower standard errorof estimateand highert ratioson everyindependent variable. In contrastto the initial core equation,which trackedthe cumulative postsampleprice changevery closely, both of the new equationsin col- umns 2 and 3 overpredictinflation during 1971-76 very substantially. The problem is not that inflationhas been mysteriouslylow over the five-yearextrapolation interval, but ratherthat the sum of coefficientson 8. See ibid., pp. 634-35, for the equations, and p. 639 for an illustration of the predictionerrors of one equation. RobertJ. Gordon 259 labor cost (line 4) is so far above 1.0 that a significantoverprediction buildsup. The same cumulativepostsample overprediction is exhibitedin column4, whereboth demandvariables are excluded.An interestingfea- tureof the no-demandversion is the highercoefficient on materialsprices, whichin the postsampleextrapolation captures more of the 1974 upsurge in prices and allows the equation to achieve a lower postsampleroot mean-squareerror. But the highercoefficient on materialsprices adds to the overpredictionof the equationin column 4, offsettingthe lower co- efficienton laborcost. The postsampleperformance of the best equation-that in column 3 -is markedlyimproved when the sum of coefficientson labor cost is constrainedto equal precisely 1.0. The constrainedequation in column 5 fits the sampleperiod about as well as the unconstrainedversion. While the root mean-squareextrapolation error is only slightlyimproved in the constrainedversion, the cumulativeoverprediction disappears. The actualchange in the deflatorfor nonfoodproduct net of energyand the predictedvalue from the constrainedequation of column 5 are dis- played in figure 1. A comparisonof the curve marked"actual" (solid line) and that labeled "fittedvalues (1954:2-1971:2 sample period)" (dotted line) reveals that the equationunderpredicts inflation at the end of its sampleperiod in early 1971, but then overpredictsin late 1971 and throughout1972 by a cumulative2.44 percentagepoints. If interpreted as a measureof the effect of the controlsprogram, that figurelies at the low end of the rangeestimated in my previouspapers. Next, the cumulativeunderprediction error in the two years endingin 1975:1 is 6.13 percentagepoints, more than double the 1971-72 over- prediction.That findingis not consistentwith my previousinterpretation that all of the 1973-75 underpredictioncan be attributedto the effect of the unwindingof controls. A more plausible interpretationis that the equation goes astray by exaggeratingthe lag between wage and price changesin an abnormalperiod in which firms recognizedthat controls had endedand reactedto postcontrolwage increasesby passingthem for- wardto customersmuch faster than they normallywould have done. A final puzzleis why the inflationrate in 1976 was consistentlybelow the predictionof the equations-in figure1 the cumulativeoverprediction is 0.92 percent.One way to isolate any recent change is to examinethe predictionsof a similarstructural price equationreestimated through the end of 1976. C'o- Oe 00
2 *~~~~~~~*~~1- ~ " m 0 in0 O C0 0 0i
a~~~~~~~~~~~~~1C1 C) t.-oN o -o
N 00 C1 00 0~~~~0~~-~~ 0 0~ ?~~~~~~~~~~~~~~~~ |1 C) C1 S 0z -COO_ co CY-CPI ~ ~ ~ ~ ~ I
~j~0ON00 0 M
- b4 oa C-C1 - ~e1 0 0oV-)
.8~~~~~~~~~~~.:SM ,
1 ;2I., 0 0 8r0 . -
C C'4 S) C)~~~~~~~~~~~C) 0 si l, > 8 W a r r
C ) C,, 0.. C1 , l , o C 0) C) 0 0:I
C)
C,, ~~~~~~~~~~0 C Cu - ~~~~0~ ~ ~ ~ ~ e- e&4 0 0 .N. o- a E ;
tn C4 ~ C4 ~ * . . m O O * * r 4 OO~~~~~~~~~~~~~~~C C* * bm_o .*A* O N O G O N O t_ t) n N Q* * .: Y.<~~~~~~D =~~~~~~~~~0e Qb I I I I O a *e~~~~~~~~~~~~~5 C
,~~~~~~~~~~~~~~~~~~oi , or
* * No O ^ o o O ~~~~~~~~~~~~~~~~~'=-E; .0 V. crO- c : *~~~~~~~~~~~~ I o _ t 0 c = C<) ?U t _ ^ 4 mur X 3 SgQ0 O~~~~~~~~~~~~~~~~~cO :gc g c O b * * * cot * * 3~~~~~~~tf)C R w3 '-0 . C' Cv5 N- :j 0 ai. ? Cm .t = 3 ' *S S 2 i,yEu1 C~~~~~4 _, Q< *0 .0 0~~~~~~~~~~~~~~~~~~0 go 4-4 PIZ0 U,~ ~~~~~~~~~~~~~~~~~~~O 411,~~~~~~~~~~~~~~~~~~~~4 0~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Po 5-4 4- o 0 ~ ~ ~ ~ ~ ~ ~ ~ ~~~0* t ~ ~ ~ ~ ~ - 0 c~~~~~~~~~- ~~ ~ ~ ~ ~ ~~O)C Robert J. Gordon 263 Column6 in table 2 reportsthe coefficientsof the extendedequation. The effect of price controls is capturedby two dummy variables,one covering the six-quarterinterval beginningin 1971:3, and the second coveringthe four-quarterinterval beginning in 1974:2. The coefficients of the dummyvariables are highly significantand cumulateto a value of -1.98 percentof the controlsperiod and + 2.04 for the postcontrols rebound(there is no constraintimposed to force these cumulativetotals to equaleach other). Column7 amendscolumn 6 by constrainingthe sum of the coefficients on trendunit laborcost to equal 1.0. To highlightthe differingtime paths of the two sets of predictions,based on columns 5 and 7, respectively, fittedvalues for the extendedequation are displayedin figure 1 with the impactof the dummyvariables excluded. The majordifferences occur in the 1973-75 period,when the extendedequation does a muchbetter job of capturingthe timingof the accelerationand subsequentdeceleration of inflation. This performanceis achieved by three shifts in coefficients when the equationis extended.First, the coefficientson labor cost shift sufficientlyto reducethe mean lag by 1.6 quarters.9This allows more of the postcontrols,1974 bulge in wage change to influenceprice change in 1974, ratherthan in 1975. Second, the coefficienton materialsprices is higher,which raises predictedinflation in 1973-74 while reducingit in 1975. Third, the coefficienton currentproductivity change is higher, allowing the negative values of productivitychange in late 1973 and throughout1974 to boost predictedprice change. What is the properinterpretation of the shifts in coefficientswhen the sampleperiod is extended?Any coefficientin a time-seriesregression is sensitiveto conditionsinside the sampleperiod. Thus it is not surprising that an equationestimated for the relativelyplacid 1954-71 periodmisses some aspectsof the timingof pricingdecisions by firmsduring 1971-76, 9. The mean lag of 4.8 quarters in the 1954-71 equation seems unreasonably long. When that sample period is split in half, the mean lag falls to 2.9 quartersfor 1954-62 but rises to 8.1 quartersfor 1963-71. A close examination of the data leads me to suspect that erratic movements of the series on compensation per manhour (CMH) in the latter period forced the computer to "stretchout" the lags. The alter- native wage index, average hourly earnings (AHE), moved more smoothly and actually is more successful as the wage variable for the equation in column 5. It cuts the standarderror from 0.234 to 0.213, and the mean lag from 4.8 to 4.0 quarters. I now believe that, despite its narrower scope, AHE is the preferable wage variable for price (as well as wage) equations, returningto a judgment reflected in my 1971 paper. 264 Brookings Papers on Economic Activity, 1:1977 a period that includedprice and wage controls, a tremendoussurge in materialsprices, and an unprecedentedslump in productivity. StructuralWage Equations Structuralwage and price equationssuitable for estimatingthe surpris- ing aspectsof the 1971-76 inflationare containedin a paperthat I wrote in early 1971.10 While the specificationof the structuralprice equations reportedin table 2 and figure 1 was alteredsomewhat in 1975 and thus incorporatesknowledge of events to that point, no such reevaluationof the 1971 wage equationshas yet been carriedout.1' Thus this section on wage behaviorin the last five years can identifygenuine "surprises" rela- tive to 1971 expectations. The firstcolumn of table 3 presentsthe relevantstatistics of the "final" 1971 wage equation.12The dependentvariable is the two-quarterrate of change in a privatehourly earningsindex, the AHE variablementioned above,which is adjustedby the Bureauof LaborStatistics to excludethe effectsof changesin overtimeand of interindustryemployment shifts, and whichincorporates as well an adjustmentto includethe effectsof changes in fringebenefits (includingemployer contributions for social security). Coefficientsfor two of the independentvariables in the equationsare not listedin table 3, the constantterm and the constrainedeffect of changes in the social security tax rate. The first three listed independentvari- ables are proxies for labor markettightness-unemployment dispersion amongdemographic subgroups, the "disguisedunemployment rate" (the differencebetween the actuallabor force and its trend), and the "unem- ploymentrate of hours" (the differencebetween private hours per week and its trend). The official unemploymentrate does not appearin the equation;the three labor marketvariables are all correlatedwith it and incorporateits influence.Although only currentvalues of the three vari- ablesare included in the wageequation, each of the threereacts to changes in outputwith a differinglag pattern,allowing output changesand thus 10. "Inflationin Recession and Recovery." 11. Detailed comparisons of the performance of the 1971 wage equations with alternative versions proposed by other authors are contained in Robert J. Gordon, "Wage-PriceControls and the Shifting Phillips Curve,"BPEA, 2:1972, pp. 385-421. 12. This information is copied from "Inflationin Recession and Recovery,"table 1, equation 11. Robert J. Gordon 265 changesin labor marketconditions to influencewages with a distributed lag. Two pricevariables are listed (lines 6 and 7). The firstis a distributed lag of past changesin the personalconsumption deflator, with lag weights obtainedfrom a separateregression of the nominalinterest rate on past inflation.The second is the differencebetween changesin the "product price" (nonfarmdeflator) and the consumptiondeflator. The final vari- able (line 9) is the rate of changein the employee-taxvariable, the sum of the effectivetax rate on personal income and the employee's effective socialsecurity tax rate.'3 Data revisionsbetween 1971 and 1976 alter the coefficientsand their statisticalsignificance, as is evidentin comparingcolumn 1, whichis based on the originaldata, and column 2, which is based on the most recently revised data. Ironically,the "naturalrate hypothesis,"in the form of a coefficientof unity on price inflation,is vindicatedby the revisionsin the official data. The unemployment-dispersionvariable becomes insignifi- cant while the coefficienton inflationincreases in lines 6 and 7; as I showedin 1972, the dispersionvariable and high coefficientson inflation are substitute explanations of wage change in the 1954-70 sample period.'4 When the sample period is extended by two quarters,in column 3, coefficientsshift furtherbut the results are reasonablysatisfactory. Al- though the unemployment-dispersionvariable has faded away, the coefficientof the disguised-unemploymentvariable remains significant and that of unemploymentof hoursis considerablyincreased and enhancedin statisticalsignificance as comparedwith column 1. The coefficientson the price variables strongly indicate that wage change fully incorporates changesin price inflationand that it is influencedby changesin product prices,not consumerprices. 13. The 1971 specification,with the social security tax appearingboth as a con- straint on the left-hand side of the equation and as part of the employee-tax variable on the right-handside, allows measurementerror to bias downwardthe coefficienton the employee-tax variable. In columns 2 through 7 this bias is eliminated by defining the employee-tax variable as the two-quarterchange in 1/(1 - rp), where rT is the effective personal income tax rate. This and the replacement of the nonfarm deflator by the deflator for nonfood business product net of energy are the only changes in specificationin moving from column 1 to column 2. Each equation includes a con- stant term and a social security tax constraint,not shown in table 3. 14. See my "Wage-PriceControls," figure 1, p. 402. !%Y~~~~~~~~~~~~~~~~~~~~~~~~~~c:F'@U : t I m SI s s O)W;;s I * o@~~~~~C F;>00C) o bW4 * o 0 > ut ? w X s m oi t Y ~~~~~~~~~~~~~~~~~~~00o em > t 2 o XU z sm S % 'SmtY 2 0O tm ?? gO CX O~~~~~~~C z iz 0 S , t , 3 . q -4 10 0 10 .0 0 a CDM 0 4.. -4 00 tn -0 bD 'S V-4 tn "-4 .14 eq C) en Caco O O C; bo cl c,3w -0 > Q)04) C 00 ON - P4 v) "-q .2.0 0 C's ON > 03 ON CN O C9 C13 a crj ,6 C;cli 6> O' 0 V3 C;3 .03 cl Cfj > cd 0 P4 CrN 1-- C's w en C) C's rn (U 0 tf) 00 -.0 O'- c; 0) 03 44 C; C; C; ci C; lul . U) 0 >, crj 0 R 03 (D O cl o 0 0) 0 03 "-q 00 C-4 ,O , 40 0 ., CZ (7 C,34 0 V) C; C; C; C; to clj cri cq T$ 0 w 0 0C,3 Cc .40 cncn k.0 00 O en 0 VD V) C) U - 0 CO ,5000. la C; C; 00 O U U 0U0 O CZ"O C.)0C's _O-0 1wS., to crj 0 v--4 000, 10 -1 r. v--q 1-0U 10 0 C,4 O 0 C) 03 .0 O .2 E'nclC4 cl 'O crj00 :3O crj 0 (n r) - O 00 0 .0 =C13 0 .0 to Q) O r. C; CD 0 cq0> C,3 C,3 0.) C13 ZQ3 a W.-C) 0 0 OC'3 C,3 t" 0 0 0>0 .1 U0 0r. 00000 C3 'C"i aj A C) > C, to r. 'cn' 10 0 S C130.0 4) r-, - r. crj t,.., to0crj OOC! 10 oo U OC, 0 4 O) O C'3 U.r,.s 0 w 0 OO0 = > C13 'Z Q) M O CZto 2 'O) Q) :3 0 00 crj w O 'IC) CZ tn (1) C13ci O I-0r-4 03 Ei tr) cd C;3 rr,,,. C.)- 00.) .0 4 0 cts (1) > Cd ; 0 4.4 : -, U cd 0 U3 X 00 .0 10,0.)2U O' 0 .- 0. - crl 1 ON 04N ;.. O ?64 0 O 0 Li C) C) C13 O;> (U 040 cn cq 0 0 0 o 0 0 U0 I"Cd E 04 (a to 0p0 11 " 0 4'. Cl O "O C'3 0 0 .00.0 U 0 U , 0 .60 Cs 4b), 'w' 00 C3 q C13C,$ "10 P-4 v" 10 la .4 268 BrookingsPapers on Economic Activity, 1:1977 As in the case of the structuralprice equations,the postsampleextrap- olation errorsof the wage equationare vastly largerthan the in-sample standarderror (lines 12 and 13 of column 3). Two separateextrapola- tions areperformed; the lowerfigures in lines 13 and 14 resultfrom using the nonfarmbusiness deflatoras the "productprice" while the upper figuresresult from using the deflatorof nonfood businessproduct net of energy.'5The cumulativeoverprediction given in line 14 is much higher whenthe nonfarmdeflator is used. This is the firstindication of a conclu- sion that emergesvery stronglyin this section: none of the 1973-74 infla- tion in food and energyprices "got into" wages, and all pre-1971 wage equationsthat allow any influenceof food and energy prices drastically overpredictthe cumulative1971-76 wageincrease. Just as the postsampleextrapolations of the structuralprice equation were superiorwhen the sum of labor-costcoefficients was constrainedto be 1.0, the extrapolationsof the wage equationimprove when the sum of the price coefficientsis constrainedto be 1.0. The constraintis intro- duced by changing the arrangementof the price variables. Since the result in column 3 indicates that only the product price "matters"- since the 1.085 coefficienton the consumptiondeflator in line 6 is virtually cancelledby the 0.974 coefficienton "minus"the consumptiondeflator in line 7-the product price is entered directlyin line 8 with the sum of coefficientsconstrained to equal 1.0. Now the size of the coefficienton line 7 measures(with reversesign) the separateinfluence of the consump- tion deflator;a coefficientof 0.0 would indicatethat only productprices matter,and a coefficientof -1.0 that only consumptionprices matter. The constrainedequation in column 4 fits the sample period slightly better than the unconstrainedversion does, and achieves a markedim- provementin the postsampleroot mean-squareerror. The cumulative postsampleoverprediction is cut to slightlymore than 1 percentagepoint whenthe deflatorfor nonfoodbusiness product net of energyis used as the productprice. Nevertheless,the postsampleperformance is by no means perfect, as is clear in figure2 from a comparisonof the solid, "actual," line with the dotted line representingthe postsamplepredictions of col- umn 4. The equation underpredictsin 1972 and 1973. Although the similarunderprediction in the four quartersprior to controlsin 1970-71 15. The coefficientsin columns 2 through 8 are based on the deflatorfor nonfood business product net of energy. RobertJ. Gordon 269 complicatesthe verdict,the performancesuggests that tlle controlspro- gramdid not reducewage changeat all; beyondthat, wage changeduring the controlsprogram did not even reflectthe decelerationof prices. The other majorerror in the extrapolationis a substantialoverprediction of wage changeduring 1975 and 1976. A possibleinterpretation of the pat- ternof theseerrors is presentedbelow. The strongevidence that productprices and not consumerprices mat- ter suggeststhat the major determinantof wage behavioris the demand for laborby firms,not the needs of workersor union aggressiveness.That, in turn,raises the questionof whetherwage changesdepend basically on demandconditions in the productmarket rather than exclusivelyin the labormarket. Considerableexperimentation with lag structuressuggests that the effect of the commoditymarket on wagescan be representedby a pairof proxies for excess demand: ( 1 ) the gap betweenactual and potentialoutput, and (2) the first differencein the gap (the same variableused in the price equation).16 Whenthe pairof output-gapvariables replaces the threelabor marketvariables of the originalspecification, the standarderrors of esti- mate improveslightly (comparecolumns 3 and 5). The same holds true for a comparisonof the respectiveversions with constrainedprice co- efficientsin columns 4 and 6. The postsampleperformance of the con- strained output-gapversion in column 6 is markedlybetter than that reportedin column 4 by the criteriaof both the root mean-squareerror and the cumulativeerror. When the productprice is representedby the deflatorfor nonfood businessproduct net of energy,the outputequation in column6 can trackcumulative wage changebetween 1971 and 1976 to within0.1 percent. The output-gapequation in column 6 is remarkablein attributingvir- tually all of the impact of the demandfor commoditieson wages to the changein the outputgap. The coefficienton the level of the outputgap is 16. The output gap is equal to potential output minus actual output, with the difference divided by potential output. The level of potential output is a trend that equals actual output when unemploymentequals the natural rate of unemployment. Details of the methodology for estimating the natural unemployment rate are con- tained in Robert J. Gordon, "StructuralUnemployment and the Productivity of Women," in Karl Brunner and Allan H. Meltzer, eds., Stabilizationof the Domestic and InternationalEconomy, Carnegie-RochesterConference Series on Public Policy, vol. 5 (Amsterdam: North-Holland, 1977), pp. 181-229. 270 BrookingsPapers on Economic Activity, 1:1977 so small,and so weakstatistically, that it plays onilya trivialrole, implying that an economywith output gaps of 6 percentand -6 percentwould have almost exactly the same rates of wage inflation,given the rate of price inflation.This implicationof the output-gapversion in column6 conflicts withthe vastbody of previousresearch, including the originalspecification in columns1 and 4, in which the dominantlabor marketvariable is dis- guised unemployment,which tends to be correlatedmore with the level of total unemploymentthan with its rate of change. Finally, in constructingtable 3, I extendedthe sample period of the wageequations to the end of 1976. Resultsfor the unconstrainedversions are shownin columns7 and 8. Dummyvariables for the controlsare in- cludedin the equationfor the sametime intervalsas in table 2, and imply not only that controls in 1971-72 did not hold down wages, but that wages increasedmore than would have been expected in light of the moderatingimpact of the controls on price inflation.The improvement in fit in the extendedversion with the originalspecification is evidentin the contrastbetween the dotted and dashedlines in figure2. At the cost of only a slight deteriorationin the trackingof wage changein 1969-71, the extendedequation is able to cut drasticallythe overpredictionof wage changein 1975. Other than the inclusionof dummyvariables, the main differencein the extendedequation in column 8 is a markedincrease in the absolute value of the coefficienton the level of the output gap. The recessionap- pearsto have been more effectiveduring 1975-76 in holdingdown wage changethan would have been predictedfrom the sampleperiod ending in 1971:2. The output-gapequations estimated for the 1954-71 periodtend to exhibita relativelyflat short-runPhillips curve, because of the influence of the rapidwage changeduring the recessionof 1970-71. Equationsesti- matedto the full 1954-76 perioddisplay a highercoefficient on the level of the outputgap, reflectingthe reducedrates of wage changein 1975-76. The same contrastis evident in a comparisonof the coefficienton the unemploymentof hours in columns4 and 7, the two equationsthat are plottedin figure2. Is it the 1970-71 periodthat shouldbe consideredthe outlier,or 1975-76? Some previousresearch suggests an unusualspread in 1970-71 between union and nonunion wage change which may be associatedwith the timingof unionneg,otiations over the 1967-71 period. Basedon this evidence,I tend to favorthe interpretationthat the 1970-71 000 ko/ 0o 0 V. PW .41~ t~~* .No~ ~~0 r4j-4 - 0*~~~~~~~~. o~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ol .5 0%~~~~~~~~~~0 C** 0 "",' I '..-.00 272 BrookingsPapers on Economic Activity, 1:1977 period was unusual,and hence to preferthe coefficientsin the extended equationsin columns7 and 8. Some authors have developed models of wage-settingbehavior in which wage changedepends not on price change, as in table 3, but only on the past behaviorof wages. While it is plausibleto argue that both firmsand workers base wage changes on wagechanges recently granted to comparableemployees in other firmsor industries,both theory and the data decisivelysupport a role for price change.'7When a distributedlag on past changesin wage ratesis substitutedfor price changein the 1954- 71 period, using the specificationof column 5 in table 3, the sum of squaredresiduals triples. For the longer 1954-76 period, the sum of squaredresiduals rises by 59 percent.Further, the patternof residuals indicatesthat the "wage-wage"version cannot explain any of the acceler- ation of wage changebetween 1973 and 1974. PolicySimulations A dynamicsimulation of the wage and price equations,which allows for the effectsof wages on pricesand priceson wages,provides an assess- ment of the inflationaryimplications of altermativepaths of economicre- coveryand of the requiredduration of a "stableprices at any cost"policy that prevents recovery and maintains today's output gap. Policy simulationswith a two-equationwage-price model have both disadvantagesand advantages as compared to simulations using the large-scaleforecasting models. The maindisadvantage is thatthe specifica- tion mustbe restrictedto rely (largelyif not entirely) on a single exoge- nous variable-for example,the outputgap-which "drives"the simula- tion. Offsettingadvantages are that the simulationresults may be more easily studied,interpreted, and understood,and that the equationsthat underliethe simulationsare similarly"open for inspection." The policy simulations derive alternativepaths of inflation in the nonfood sector net of energyimplied by alternativeexogenous paths of the output gap. Since relative energy prices are likely to rise over the next few years, the correspondingpaths for the GNP deflatorwould all lie above that presentedin figure3. 17. See particularly Robert E. Hall, "The Process of Inflation in the Labor Mar. ket," BPEA, 2:1974, pp. 343-93, and my criticismsof that paper, pp. 394-99. 00 0 * I~~~~~~~~~~~~~~~~~~~~0 * 1 ~~~~~~~~~~~~~~~~~~00 - 0 1 I~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ oPo '~~~~~~~~~~~~~~*0 I-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~0 *p.4~ ~ * oO 0 I en 0 274 Brookings Papers on Economic Activity, 1:1977 Becausethe previousanalysis leads to the conclusionthat the extended- periodprice equationcontains a more plausiblelag patternon trendunit laborcost, andthat the steeperPhillips curve in the extended-periodwage equationis likelyto be more accurate,the simulationspresented here are based on the price equationin table 2, column 7, and the wage equation in table 3, column8. The wage equationthat uses the outputgap rather thanthe unemploymentvariables of the originalspecification is employed to avoid the problemof creatingequations that link those unemployment variablesto the output gap. Tax rates were all assumedto remain unchangedat their values in 1976:4, and the changein the relativeprices of capitaland consumption goods was set equal to zero in all simulations.Simple equations were de- veloped to relate changesin materialsprices and the change in the pro- ductivitydeviation to the changein the output gap. Furtheradjustments were made to ensurethat the inflationrate would neither acceleratenor deceleratewhen the outputgap was zero. To obtainthis resultin dynamic simulations,it is not enoughto constrainthe sum of coefficientson wages in the price equation,and on prices in the wage equation,to be equal to 1.0. Three other importantrestrictions must be imposed:First, the trend rate of productivitygrowth in the price equationmust be set equal to the constantterm in the wageequation. This switch,from 1.96 to 2.13 percent annually,is small enough to be acceptableand within the range of the standarderror in the equationoriginally used to estimatethe productivity trend.Second, the growthrate of the wage variablein the price equation must equal that of the wage variablein the wage equation.Third, there mustbe no changein relativematerials prices. Figure 3 correspondsto these assumptionsand displays three com- binationsof inflationand unemployment.Path A is an implausiblyrapid recoverythat reducesthe outputgap from its 6.2 percentrate at the end of 1976 to zero by 1978:1. At first inflationis predictedto slow down moderately,benefitting from the lagged influenceof low rates of change in wages and prices in 1976, but then an accelerationbegins. The "rate of change"effects of a rapidlyfalling output gap push inflationclose to 7 percentin late 1978, followed by an adjustmentto the long-run"steady state"rate of 6.4 percent. A slower recovery,path B, reaches a zero gap in 1980 (the quarter before the next presidentialelection), ratherthan in early 1978. Slower growthhas both transitoryand permanentbenefits. Inflation is lower by RobertJ. Gordon 275 as much as 1.3 percenitagepoints at an annualrate in late 1978, and the long-run"steady state" rate of inflationis 0.4 pointslower."8 Sincepath B correspondsmost closely to the recoverypath apparently desiredby the Carteradministration, this "optimistic"simulation con- flicts with the administration'savowed aim of reducingunemployment while simultaneouslyachieving a decelerationof inflationto 4 percent. Even on the optimisticassumption of zero changein relativeenergy and food prices,the administration'spolicy goals are inconsistent. The thirdalternative in figure3, path C, shows the rate of deceleration of inflationthat would obtain if the outputgap were held permanentlyat 6.2 percent.The inflationrate would fall rapidlyduring 1977, reflecting the delayedimpact of the lower-than-predictedactual rates of wage and price change during 1976. Subsequently,a furthermodest slowdownof inflationwould occur, beginningwith a 0.24 percentagepoint drop in the inflationrate in 1978, wideningto a decelerationof 0.36 percentage point per year in 1986. This turtle-likedeceleration of inflationreflects the extremelyweak effect of a high output gap on wage behavior, and the absenceof any effect of a maintainedgap on price behavior. In my own judgment,the assumptionsunderlying the simulationsre- flectedin the figurelean towardthe optimisticside. First, as noted above, they ignorethe prospectof risingrelative prices of energyover the years ahead.Second, they assumeno upwardtrend in relativematerials prices, in contrastwith the actually observed trend of 2.0 percent a year for 1963-76 (adjustedto a constantoutput gap). Third, they assume that compensationper manhour and average hourly earningswill grow at equal rates, when in fact the formerhas outpacedthe latter by 0.3 per- centagepoint a year on averagesince mid-1971. If that trend were as- sumed to continue, it would put added upward pressureon the price equationfor any path of averagehourly earningspredicted by the wage equation.Alternative, more pessimistic,assumptions could easily add 1 to 2 points to the inflationrate by 1980 and as much as 3 to 4 points by 1986. 18. As an example of a more optimistic conclusion, a "control solution" recently published by Data Resources, Inc., predicted that the economy could reach 5.5 per- cent unemployment in 1980, a path roughly equivalent to my path B, with only a 5.4 percent change in the GNP deflator in 1980. See Otto Eckstein and others, Eco- nomic Issues and Parameters of the Next 4 Years (Data Resources, Inc., 1977), table 6, p. 30, solution "CONTROL1229." 276 BrookingsPapers on Economic Activity, 1 :1977 Conclusion All approachesfail to explainthe increasedvariance of inflationduring 1971-76 as comparedto the pre-1971 period. But overall, the cumula- tive amountof inflationsince 1971 can be explained-even overexplained -by establishedeconometric procedures. Both the structuralprice and the structuralwage equations can track the cumulativechange in the pricesof nonfoodbusiness product net of energyand in wagesto withina percentagepoint, once they incorporatethe sensibleconstraint that sums of coefficientsof priceson wagesand wages on pricesequal unity. The analysisof thispaper leads to the followinginteresting conclusions. First, the short-runPhillips curve relatingwage changeto unemploy- ment or the output gap may well be steeper than implied by equations estimatedfor sample periods ending in 1971. While this result helps to explainwhy wage changes were so moderatein 1976, it implies that a rapideconomic recovery may bring about a greateracceleration in infla- tion than some commentatorsappear to anticipate. Second,the speed of recoverymatters, in both the price and the wage equations.It is the rate of change of the output gap that influencesthe rateof changeof pricesrelative to wages,and there is also a partialimpact from the speed of the change in output in the output-gapversion of the wage equations. Third,the abilityof productprices and the outputgap alone to explain wage behaviorsuggests that the demandfor labor by firms is the main determinantof wages,and that autonomousactions or reactionsby work- ershave littleimpact. Fourth, as in previouspapers, I conclude that price controls worked temporarily,with a decline in the price level followed by a rebound,but thatwage controls had if anythinga perverseeffect. Why the effectiveness of the controls programshould have been limited to prices is a puzzle thatothers may be betterable to answer.The implicationsfor wage guide- lines or jawboningare not reassuring. Fifth, none of the increasesin food or oil prices in 1973-74 appears to havebeen incorporatedinto wages.In the contextof my previousstudy of supply shocks, this implies that policymakerscould have stimulated nominalincome growthto accommodatesome of the effect of food and oil priceswithout setting off an endlessinflationary spiral. But the strong RobertJ. Gordon 277 demandeffects exhibited in the equationsof this paper suggestthat such a policy of accommodationwould have substantiallylessened the deceler- ationof inflationbetween 1974 and 1976.19 Sixth,perhaps most important,the outlook for inflationis rathergrim. Despitethe continuingoutput gap, the statisticalevidence presented above indicatesthat any furtherdeceleration in inflationis highly unlikely. On the contrary,it points to the probabilityof some accelerationas the economycontinues its recovery.While the extentof that accelerationwill dependon the speedof the recovery,inflation rates of 6 or 7 percentseem likely for the next severalyears, compared with the 5 percentrate during 1976. Any seriouseffort to eliminateinflation through demand restraint wouldbe exceedinglycostly; a strategyof maintainingthe late 1976 output gap mightbring the inflationrate down to 2 percentby the mid-eighties, but only through a loss of output that would substantiallyexceed $1 trillion. Finally, as a corollaryto this unpleasantverdict, the recoveryitself is likely to requirea maintainedgrowth of monetaryaggregates above rates that now seem acceptableto the FederalReserve, in orderto financean annualgrowth of nominal gross national product of 12 or 13 percent duringthe rest of the decade. How the makersof monetarypolicy will react to this dilemmaremains to be seen. Discussion SEVERAL participantscommented on the substantialdifferences between the coefficientsin the price and wage equationsfitted through1971 and thosefor the periodas a whole.They questionedthe stabilityof the under- lying structurein light of these changes.George Perry noted, in particu- lar, the much greaterrole played by the level of the output gap in the full-periodestimate of the wage equation.Franco Modiglianicautioned againstdrawing the inferencethat consumerprices do not matteron the 19. A hypothetical accommodativepolicy that maintainedthe output gap at zero in 1974-76 would have caused substantial extra inflation, reaching a peak in mid- 1975 of 3.8 percentage points over that which actually occurred, and then tapering off to an excess of 2.0 percentage points in late 1976. This conclusion is based on a dynamic simulation of the same equations as are used in figure 3. 278 Brookings Papers on Economic Activity, 1:1977 basis of revised data and updated equationswhen the earlier evidence suggestedotherwise. ArthurOkun remarked that path C in figure3 impliedextremely asym- metricaleffects of excess demandand excess supply. According to the simulation,it takes a 6 percentGNP gap for a whole decadeto eliminate an inflationrate that resulted from much smaller excesses over potentialin the past. Yet thereis no nonlinearityin the equationsused for that simu- lation. Gordonresponded that the apparentasymmetry resulted from the contributionto inflationin the past of variablesother than excess de- mand-particularlytax rates and materialsprices-which are artificially held constantin the simulationsof figure3. Perry commentedthat over the yearsgrowing awareness of inflationmay have causedthe priceeffects on wages to rise, in line with Michael Wachter'sanalysis of a shifting Phillipscurve. Okun suggested that any advocateof extremedemand re- straintwould have a far more optimisticview than path C, relyingon a "hawkish"policy stance to reverseinflationary expectations. He reminded the groupthat WilliamFellner had developed that line of argumentin detail. PenttiKouri was somewhatsurprised to see that such a small role was assignedto consumerprices in U.S. wage behavior during the recent period,unlike other industrial countries. In the UnitedKingdom and Italy, in particular,deteriorations in the terms of trade had inducedwage de- mandsaimed at sustainingprevious real wage levels. Perryrecalled, how- ever, that in his researchon Europeanwage behaviorhe had found that value-addedand wage-wageeffects dominated those of consumerprices. Modiglianidisagreed with Perryover the importanceof consumerprices, especiallyin the case of Italian wage behavior.Wachter suggested that food and fuel inflationmight have ultimatelygotten into wages in the United States if not for the severityof the recent recession.He found it quite plausiblethat price feedbackson wages interactwith demandcon- ditions. EdmundPhelps believedthat the nonfood nonfarmdeflator may have performedbetter than the consumerprice index because the formeris a proxyfor laggedwage changes. He wouldprefer to see a laggedwage term used insteadof prices;he felt that, on theoreticalgrounds, the appropri- ate variableis the expected rate of wage change, which should have a coefficientof unity given labor demand and supply. Gordon reported, however,that in tests he conductedlagged prices performedbetter than RobertJ. Gordon 279 laggedwages. Christopher Sims expressedsome amusementthat the best wage equationhad no labormarket variables in it. This resultconformed with his belief that wage and price equationscannot be distinguishedas applyingto differentcategories of behavior.It was preferableto consider them as interestingstatistical reduced-form summaries of the dynamic relationshipsamong the variables. Wachterfound it implausiblethat disguisedunemployment, which is composedlargely of marginalworkers, could exertthe majorinfluence on wage changes attributedto it by Gordon'swage equation.He also ob- served that the small demand effects in the early price equations meant thattogether the wage and priceequations formed an almostpurely auto- regressivesystem, in whichprices and wagesfed upon each otherwithout beinginfluenced significantly by demand. Gordonsupported Sims' interpretationthat the wage and price equa- tions representedreduced-form summaries of dynamic relationships. Many differentvariables shared the major cyclical movements of the sampleperiod, preventing statistical discrimination among finely differen- tiated hypotlleses.As Gordon saw Wachter'scriticism of the disguised unemploymentvariable, it attemptedto place a structuralinterpretation on a variablethat simplyrepresented a generalizeddemand effect and that performedno better or worse than the output gap, an alternativeproxy. Gordon concludedthat, even when the wage and price equationswere viewed as reducedforms, severalconclusions emerged strongly, particu- larly the importantrole of inertiain the wage-priceprocess, and of the rate of change of output. On the other hand, no confidencecould be placedon the impactof the level of outputwithout an informedjudgment on unusualaspects of the 1970 and 1973-75 recessions.