minimum wages and employment: a case study of the fast

sitreEcNaSecoThdaplrepomAlthpaEcUn(HheRe

exqusu

Minimum Wages and Employment: A Case Study of theFast-Food Industry in New Jersey and Pennsylvania: Reply

By DAVID CARD AND ALAN B. KRUEGER*

nndeir--

themoddschtetaafy

ine-

-mllydem

esm

ee

iveey’sofdntjorof

ea92ionsthet-y-ey

Replication and reanalysis are important edeavors in economics, especially when new fiings run counter to conventional wisdom. In thComment on our 1994American Economic Review article, David Neumark and William Wascher (2000) challenge our conclusion thatApril 1992 increase in the New Jersey minimuwage led to no loss of employment in the fast-foindustry. Using data drawn from payroll recorfor a set of restaurants initially assembled by Riard Berman of the Employment Policies Institu(EPI) and later supplemented by their own dacollection efforts, Neumark and Wascher (hereter, NW) conclude that “... the New Jerseminimum-wage increase led to a relative declin fast-food employment in New Jersey” compared to Pennsylvania.1 They attribute the discrepancies between their findings and ours to problein our fast-food restaurant data set. Specificathey argue that our use of employment datarived from telephone surveys, rather than fro

e-in-anothe

terys-em-r.

Sm-ee-

dS

* Card: Department of Economics, Evans Hall, Univer-y of California, Berkeley, CA 94720, and National Bu-au of Economic Research; Krueger: Department oonomics, Princeton University, Princeton, NJ 08544, andtional Bureau of Economic Research. The analysis inctions I, II, and III, subsection E, of this paper is based onnfidential Bureau of Labor Statistics (BLS) ES-202 datae authors thank the BLS staff for assistance with thesta. Although the BLS data are confidential, persons em

oyed by an eligible organization may apply to BLS forstricted access to ES-202 data for statistical research puses. Data from our 1994 paper are available via anony

ous FTP from the minimum directory of irs.princeton.edu.l opinions and analysis in this paper reflect the views ofe authors and not the U.S. government. We thank seminrticipants at Princeton University, the National Bureau oonomic Research, the University of Pennsylvania, theiversity of California-Berkeley, the Kennedy Schoolarvard University), and Larry Katz and John Kennan forlpful comments, and the Princeton University Industrialations Section for research support.1 In the March 1995 version of their paper, NW reliedclusively on 71 observations collected by EPI. Subseent versions have also included information from theirpplemental data collection.

139

--

-

--

s,-

payroll records, led us to draw faulty inferencabout the effect of the New Jersey minimuwage.

In this paper we attempt to reconcile thcontrasting findings by analyzing administrativemployment data from a new representatsample of fast-food employers in New Jersand Pennsylvania, and by reanalyzing NWdata. Most importantly, we use the BureauLabor Statistics’s (BLS’s) employer-reporteES-202 data file to examine employmegrowth of fast-food restaurants in a set of machains in New Jersey and nearby countiesPennsylvania.2 We draw two samples from thES-202 files: a longitudinal file that tracksfixed sample of establishments between 19and 1993, and a series of repeated cross sectfrom the end of 1991 through 1997. BecauseBLS data are derived from unemploymeninsurance (UI) payroll-tax records, the emploment measures are free of the kinds of surverrors that NW allege affected our earlier rsults. In addition, because the ES-202 dataclude information for all covered employers infixed group of restaurant chains, there isreason to doubt the representativeness ofBLS sample.

A comparison of fast-food employmengrowth in New Jersey and Pennsylvania ovthe period of our original study confirms the kefindings in our 1994 paper, and calls into quetion the representativeness of the sample assbled by Berman, Neumark, and WascheConsistent with our original sample, the BLfast-food data set indicates slightly faster eployment growth in New Jersey than in thPennsylvania border counties over the time priod that we initially examined, although inmost specifications the differential is small anstatistically insignificant. We also use the BL

f

.e-

r--

arf

l

-2 The ES-202 data are also known as the Business Es-

tablishment List.

7

wd

raltd

cav

he-nnu

uy

enetin

e

nygi

ir

n

a

htheolles)lylyhemna-u-ls,nt

tthe

i-

es-

-rein

t-rem-r

-eraep-Sa.ntir-

st--e

nt

3 The first question on the Pennsylvania form requeststhe “Total covered employees in pay period incl. 12th ofmonth.” Employers are asked to report employment for eachmonth of the quarter. A copy of these forms is availablefrom the authors on request. Other points to note about theES-202 data include: they are not restricted to employerswith any minimum number of employees, or to employeeswho have earned any minimum pay in the pay period; thereis no information on hours of work; the pay period may varyacross employers, or within employers for different work-ers; employees on vacation or sick leave should be includedif they are paid while absent from work.

1398 THE AMERICAN ECONOMIC REVIEW DECEMBER 2000

data to examine longer-run effects of the NeJersey minimum-wage increase, and to stuthe effect of the 1996 increase in the fedeminimum wage, which was binding in Pennsyvania but not in New Jersey, where the staminimum wage already exceeded the new feeral standard. Our analysis of this new poliintervention provides further evidence thmodest changes in the minimum wage halittle systematic effect on employment.

In light of these results we go on to reexamine the Berman-Neumark-Wascher (BNWsample and evaluate NW’s contention that trise in the New Jersey minimum wage causemployment to fall in the state’s fast-food industry. Our reanalysis leads to four main coclusions. First, the pattern of employmegrowth in the BNW sample of fast-food restarants across chains and geographic areaswithinNew Jersey is remarkably consistent with ooriginal survey data. In both data sets emploment grew faster in areas of New Jersey whwages were forced up more by the 1992 mimum-wage increase. The differences betwethe BNW sample and ours are attributabledifferences in the BNW sample of Pennsylvanrestaurants, which unlike the more comprehesive BLS sample, and our original samplshows arise in fast-food employment in thestate. Second, the differential employment trein the BNW Pennsylvania sample is driven bdata for restaurants from a single Burger Kinfranchisee who provided all the Pennsylvandata in the original Berman sample.

Third, the employment trends measuredthe BNW sample are significantly different forestaurants that reported their payroll data oweekly, biweekly, or monthly basis. Establishments that reported on a biweekly basis hfaster growth than those that reported onmonthly or weekly basis. We suspect that tdifferent reporting bases matter becauseBNW employment measure is based on payrhours (rather than actual numbers of employeand because weekly, biweekly, and monthaverages of payroll hours were differentialaffected by seasonal factors, including tThanksgiving holiday and a major winter storin December 1992. Regardless of the explation, a higher fraction of Pennsylvania restarants reported their data in biweekly intervaleading to a faster measured employme

yl

-e-

yte

-)ed

-t-

r-

rei-n

oa-

,

d

a

n

a-dae

growth in that state. Once the employmenchanges are adjusted for the reporting bases,BNW sample shows virtually identical growthin New Jersey and eastern Pennsylvania. Fnally, a reanalysis of publicly available BLSdata on employment trends in the two statshows no effect of the minimum wage on employment in the eating and drinking industry.

Based on all the evidence now available, including the BLS ES-202 sample, our earliesample, publicly available BLS data, and thBNW sample, we conclude that the increasethe New Jersey minimum wage in April 1992had little or no systematic effect on total fasfood employment in the state, although themay have been individual restaurants where eployment rose or fell in response to the higheminimum wage.

I. Analysis of Representative BLS Fast-FoodRestaurant Sample

A. Description of BLS ES-202 Data

On April 1, 1992, the New Jersey state minimum wage increased from $4.25 to $5.05 phour, while the minimum wage in Pennsylvaniremained at $4.25. To examine the effect of thNew Jersey minimum-wage increase using reresentative payroll data, we applied to the BLfor permission to analyze their ES-202 datThe ES-202 database consists of employmerecords reported quarterly by employers to thestate employment security agencies for unemployment-insurance tax purposes. The firquestion on the New Jersey UI tax form requests the “Number of covered workers employed during the pay period which includes th12th day of each month.”3 The BLS maintainsthese data as part of the Covered Employme

ole

uodmingerblm

02and

916erlyo

anein

eninn-in-in

tht iove

lgt.

n-nn

ghe-—wweadre

heoreri-ialia,a.nd

rlyomof

ro

—e.s-seenthssebeonab-atofto

led.s”or

edvelrsy-d.nedy

ly,ionande

1399VOL. 90 NO. 5 CARD AND KRUEGER: MINIMUM WAGE AND EMPLOYMENT, REPLY

and Wages Program. We analyze two typessamples from the ES-202 file: a longitudinal fiand a series of repeated cross sections.

The longitudinal sample consists of restarants belonging to a set of the largest fast-fochains.4 Restaurants in the sampled chains eployed 13 percent of all employees in the eatand drinking industry in New Jersey and eastPennsylvania in 1992. There is consideraoverlap between the restaurants in the BLS saple and those in our original sample.5 Our sam-ple of fast-food restaurants from the ES-2data was drawn as follows. We first selectedrecords for all establishments in the eating adrinking industry (SIC 5812) in New Jersey aneastern Pennsylvania in the first quarter of 19first quarter of 1994, and fourth quarter of 199Then restaurants in the sampled chains widentified from this universe by separatesearching for the chains’ names, or variantstheir names, in the legal name, trade name,unit description fields of the ES-202 file. If thname of an included chain was mentionedany of these text fields the record was thvisually examined to ensure that it belongedthe sample of included restaurant. In additiorecords forall eating and drinking establishments from these quarters were visuallyspected to identify any fast-food restaurantsthe relevant chains that were missed bycomputerized name search. If a restauranone of the relevant fast-food chains was discered that was not identified by the initial namsearch, the computerized name-search arithm was amended to include that restauran

The original Card-Krueger (CK) sample cotained data on restaurants in 7 counties of Pesylvania (Bucks, Chester, Lackawanna, LehiLuzerne, Montgomery, and Northampton). Bcause this is a somewhat idiosyncratic groupwith some counties located right on the NeJersey border and others off the border—decided to expand the sample to include 7ditional counties: Berks, Carbon, Delawa

4 For confidentiality reasons, BLS has requested that wnot reveal the identity or number of these chains. We careport, however, that there are fewer than 10 chains in thsample.

5 We reached this conclusion by comparing the distribu-tion of restaurants by three-digit zip code and chain in thetwo data sets.

f

-

-

ne-

lld

,.e

fd

,

en-

o-

-,

-,

Monroe, Philadelphia, Pike, and Wayne. In tresults that follow, we present estimates fboth our original 7 counties and for the largset of 14 counties. The map in Figure 1 indcates the location of the restaurants in our initsurvey, the original 7 counties in Pennsylvanand the additional 7 counties in Pennsylvani

Once restaurants in the relevant chains acounties were identified, we merged quarterecords for these restaurants for the period frthe first quarter of 1992 to the fourth quarter1994 to create a longitudinal file.6 To mirror theCK sample, only establishments with nonzeemployment in February or March of 1992—the months covered by wave 1 of our surveywere included in the longitudinal analysis filThe final longitudinal sample contains 687 etablishments. A total of 16 (2.3 percent) of theestablishments had zero or missing employmin November or December of 1992, the montcovered by wave 2 of our original survey. Theestablishments either closed or could nottracked because their reporting informatichanged. In 1992, less than 1 percent of estlishments had imputed employment data (this, cases where the state filled in an estimateemployment because the establishment failedreport it).

A potential limitation of the BLS longitudinasample for the present paper should be notThe ES-202 data pertain to “reporting unitthat may be either single establishment unitsmultiestablishment units. The BLS encouragemployers to report their data at the county leor below in the early 1990’s. Some employewere in the process of switching to a countlevel reporting basis during our sample perioConsequently, some restaurants that remaiopen were difficult to track because thechanged their reporting identifiers. Fortunatemost of the restaurants that were in this situatcould be tracked by searching addressesother characteristics of the stores. All of th

ene

6 Additionally, to ensure that the sample consisted ex-clusively of restaurants (as opposed to, e.g., headquarters ormonitoring posts), the authors restricted the sample to es-tablishments with an average of five or more employees inFebruary and March 1994, and average monthly payroll peremployee below $3,000 in 1992:Q1 and 1992:Q4. Theserestrictions eliminated 17 observations from the originalsample of 704 observations.


FIGURE 1. AREAS OF NEW JERSEY AND PENNSYLVANIA COVERED BY ORIGINAL SURVEY AND BLS DATA

ne

geross-tl-ely

toeeesse

hevendeiestahehe

-

os-ng

rya-2el

restaurants that were not linked to subsequemonths’ data were assumed closed and assignzero employment for these months, even thousome of these restaurants may not have closThis is probably a more common occurrence foNew Jersey than Pennsylvania: 0.4 percentthe Pennsylvania restaurants had zero or miing employment at the end of 1992, as compared to 3.4 percent of New Jersey restauranIn our original survey, 1.3 percent of Pennsyvania restaurants and 2.7 percent of New Jersrestaurants were temporarily or permanentclosed at the end of 1992.7

Also note that because firms are allowedreport on more than one unit in a county in thBLS data, some of the records reflect an aggrgation of data for multiple establishments. Waddress both of these issues in the analybelow. Importantly, however, these problemdo not affect the repeated cross-sectional filthat we also analyze.

f-

”7 An interviewer visited all of the nonresponding stores

in both states to determine if they were closed in ouroriginal survey.

td

hd.

f-

s.

y

-

is

s

To draw the repeated cross-sectional file, tfinal name-search algorithm described abowas applied each quarter between 1991:Q4 a1997:Q3. Again, data were selected for thsame chains in New Jersey and the 14 countin eastern Pennsylvania. Every month’s dafrom the sampled quarters was selected. Tcross-sectional sample probably provides tcleanest estimates of the effect of the minimumwage increase because it incorporatesbirths aswell as deaths of restaurants, and because psible problems caused by changes in reportiunits over time are minimized.

B. Summary Statistics and Differences-in-Differences

Table 1 reports basic employment summastatistics for New Jersey and for the Pennsylvnia counties, before and after the April 199increase in New Jersey’s minimum wage. PanA is based on the longitudinal BLS sample ofast-food restaurants. In the first row, the “before” period pertains toaverageemployment inFebruary and March of 1992, and the “afterpertains toaverageemployment in November

ties

ge

2)

3)

)

sylvaniafor Newed in Cardservationin the


TABLE 1—DESCRIPTIVE STATISTICS FOR FAST-FOOD RESTAURANTS DRAWN FROM BLS ES-202DATA AND CARD-KRUEGER SURVEY

Means with standard deviations in parentheses:

New Jersey 7 Pennsylvania counties 14 Pennsylvania coun

Before After Change Before After Change Before After Chan

A. BLS ES-202 DataFebruary–March 1992 to

November–December 199237.2 37.6 0.41 42.5 42.4 20.12 44.8 44.3 20.53

(19.9) (21.0) (9.82) (23.2) (23.5) (10.94) (53.7) (59.9) (12.3

February 1992 to November 1992 37.2 37.8 0.57 42.7 42.220.54 44.9 44.4 20.58(19.9) (20.9) (10.12) (23.8) (23.2) (12.82) (53.6) (60.4) (13.8

March 1992 to March 1993 37.2 34.8 22.48 42.3 37.5 24.80 44.7 40.7 24.0(20.1) (20.0) (13.99) (22.8) (18.6) (22.74) (54.0) (54.5) (18.1

B. Card-Krueger Survey DataFebruary 1992 to November 1992 29.8 30.0 0.19 33.1 30.922.23 NA NA NA

(12.5) (13.0) (9.82) (14.7) (10.6) (11.98)

Notes:Sample sizes for the first two rows are 437 for New Jersey, 127 for Pennsylvania 7 counties, and 250 for Penn14 counties; sample sizes for third row are 436, 127, and 250, respectively; sample sizes for the last row are 309Jersey and 75 for Pennsylvania. The 7 Pennsylvania counties used in the middle columns are the same counties usand Krueger (1994); these 7 counties are a subset of the 14 counties in the last three columns (see text). The unit of obfor the BLS data is the “reporting unit,” which in some cases includes multiple establishments. The unit of observationCard-Krueger data is the individual restaurant.

o-eye

alreedy-y,

hea

t-n-lewrs2

e-

c-edge

ntnser-tagel-

-

).sin-

rsin

is

and December of 1992.8 The second row re-ports employment figures for February and Nvember, which were the most common survmonths in our original telephone survey. Ththird row shows data for the 12-month intervfrom March 1992 to March 1993. Finally, focomparison, panel B of Table 1 reports thcorresponding employment statistics calculatfrom the CK survey. Note that for comparabilitwith BLS data, we have calculated total employment for restaurants in our original surveby adding together the number of full-timepart-time, and managerial workers.9

Several conclusions are apparent from tmeans in Table 1. First, the BLS data indicateslight rise in employment in New Jersey’s fasfood restaurants over the period we studied, aa slight decline in employment in Pennsylvania’s restaurants over the same period. Our tephone survey data indicate a net gain in NeJersey relative to Pennsylvania of 2.4 workeper restaurant, whereas the BLS data in rowindicate a smaller net gain of 1.1 workers b

tal,eas,e

8 In one case, employment was zero in March 1992, sothe February figure was used.

9 This approach differs from Card and Krueger (1994),which weights part-time workers by 0.5 to derive full-timeequivalent employment.

d

-

tween February and November of 1992. Seond, between March 1992 and March 1993, thBLS data indicate that both New Jersey anPennsylvania experienced a decline in averaemployment, with the decline being larger inPennsylvania. Third, the average employmelevel in the BLS data is somewhat greater thathe average level in our data, probably becausome of the observations in the BLS data petain to multiple establishments. Fourth, our daand the BLS data both suggest that averarestaurant size was initially larger in Pennsyvania than in New Jersey. By contrast, the BNWdata set indicates that “full-time equivalent employment” was initially greater in New Jerseythan in Pennsylvania (see Section III belowFinally, the BLS data indicate that the resultfor the 7 Pennsylvania counties that we usedour initial sample and the wider set of 14 counties are generally similar.

Neumark and Wascher (2000) and othehave emphasized the fact that the dispersionfull-time employment changes in our data setgreater than the dispersion in changes in tohours worked in the BNW data. Interestinglythe BLS payroll data display roughly the samstandard deviation of employment changeswas found in our original sample. For examplein New Jersey the standard deviation of th

ita

dsurmeis

neeb-i-wa

sen-o

LSnn.8

ey

-ubenrkmururoy

retsthuton

ulllyese-

he

e-

-y

-f

-

’s

s-alre-ineeees.-

gey-ry–2.oru-n-hee-

olenga

e

10 Although Neumark and Wascher (2000 footnote 9)argue that variability in employment growth should besmaller for fast-food restaurants in a small geographic areathan in a sample such as Davis et al.’s set of manufacturingestablishments, it should be noted that gross employmentflows are considerably higher in the retail trade sector thanin the manufacturing sector (see Julia Lane et al., 1996).

11 The proportionate employment change was calculatedas the change in employment divided by the initial level ofemployment. We use total number of full-time and part-time workers in our data for comparability to the BLS data.Neumark and Wascher (2000) show that some other ways ofmeasuring the proportionate change of employment (e.g.,using average employment in the denominator) and somesample restrictions (e.g., eliminating closed stores from thesample) increases the dispersion in our data relative totheirs.

mult ting


12 Observations that are not reported as subunits of-iunit establishments are either part of a multiunit repor

change in employment across reporting unbetween February and November of 1992 w10.12 in the BLS data, which slightly exceethe standard deviation calculated from our svey data (9.82) over approximately the samonths. One problem with this comparisonthat some of the BLS reporting units combitwo or more restaurants that may have bebroken out over time, whereas the unit of oservation in our original survey was the indvidual restaurant. To address this issue,restricted the BLS sample to reporting units thinitially had fewer than 40 employees: thesmaller reporting units are almost certainly idividual restaurants. The standard deviationemployment changes for this truncated Bsample is 9.0 for New Jersey and 6.8 for Pesylvania; these figures compare to 8.0 and 8respectively, if we likewise truncate our survdata.

More generally, the criticism that our telephone survey was flawed because of the sstantial dispersion in measured employmgrowth in our sample strikes us as off the mafor three reasons. First, reporting errors in eployment data collected from a telephone svey are not terribly surprising. Dispersion in odata is not out of line with measures basedother establishment-level employment surve(e.g., Steven J. Davis et al., 1996).10 Second,employment changes are thedependent vari-able in our analysis. As long as the measument error process is the same for restauranNew Jersey and Pennsylvania, estimates ofdifference in employment growth based on odata will be unbiased. We know of no reasonsuspect that the New Jersey and Pennsylvamanagers who responded to our survey womisreport employment data in a systematicadifferent way. Moreover, all of our telephoninterviews were conducted by a single profesional interviewer. Third, any comparison of thstandard deviation of full-time equivalent employment changes is potentially sensitive to t

ss

-

n

et

f

-,

-t

--

ns

-ine

r

iad

-

way data on hours, or combinations of part-timand full-time employees, are scaled. For example, in their analysis NW convert weekly payroll hours data into a measure of employment bdividing by 35, but a smaller divisor wouldobviously lead to larger dispersion of employment in their data. The standard deviation oproportionatechanges in employment is invariant to scaling and is fairly similar in all threedata sets: 0.29 in the BLS data, 0.35 in BNWdata, and 0.39 in our earlier survey data.11

C. Regression-Adjusted Models

Panels A and B of Table 2 present basic regresion estimates using the BLS ES-202 longitudinsample of fast-food restaurants. The models psented in this table essentially parallel the maspecifications in Card and Krueger (1994). Thdependent variable in the first two columns is thchange in the number of employees, while thdependent variable in the last two columns is thproportionate change in the number of employeeFollowing Card and Krueger (1994), the denominator of the proportionate change is the averaof first- and second-period employment. Emploment changes are measured between FebruaMarch 1992 and November–December 199Columns (1) and (3) include as the only regressa dummy variable indicating whether the restarant is located in New Jersey or eastern Pensylvania. These estimates correspond to tdifference-in-differences estimates that can be drived from row 1 of Table 1. The models incolumns (2) and (4) add a set of additional contrvariables: dummy variables for the identity of threstaurant chain, and a dummy variable indicatiwhether the reporting unit was a subunit ofmultiunit employer.12

The regression results in panel A of Tabl

geioll,

yetyysth

teielsin’

-n

ose

e-e.-wof-SareSr-ai-ry

ting

1

4

2

Subunitdata,are in


firm or the only restaurant owned by a particular reporunit.

TABLE 2—BASIC REGRESSIONRESULTS; BLS ES-202 FAST-FOOD DATA AND CARD-KRUEGER SURVEY DATA

Explanatory variables

Dependent variable:

Change in levels Proportionate change

(1) (2) (3) (4)

A. All of New Jersey and 7 Pennsylvania Counties, BLS Data

New Jersey indicator 0.536 0.225 0.007 0.009(1.017) (1.029) (0.029) (0.029)

Chain dummies and subunit dummy variable No Yes No YesStandard error of regression 10.09 9.99 0.286 0.28R2 0.001 0.029 0.000 0.046

B. All of New Jersey and 14 Pennsylvania Counties, BLS Data


Chain dummies and subunit dummy variable No Yes No YesStandard error of regression 10.80 10.63 0.303 0.29R2 0.002 0.042 0.005 0.071

C. Original Card-Krueger Survey Data


Chain and company-ownership dummies No Yes No YesStandard error of regression 10.28 10.25 0.385 0.38R2 0.009 0.025 0.001 0.024

Notes:Each regression also includes a constant. Sample size is 564 for panel A, 687 for panel B, and 384 for panel C.dummy variable equals one if the reporting unit is a subunit of a multiunit employer. For comparability with the BLSemployment in the CK sample is measured by the total number of full- and part-time employees. Standard errorsparentheses.

2, which are based on the employment chanfor restaurants in the same geographic regsurveyed in our earlier work, indicate smapositive coefficients on the New Jersey dummvariable.13 Each of the estimates is individuallstatistically insignificant, however. We interprthese estimates as indicating that New Jerseemployment growth in the fast-food industrover this period was essentially the same awas for the same set of restaurant chains in7 Pennsylvania counties.

In panel B, regression results are presenusing the wider set of 14 Pennsylvania countas the comparison group. These results aindicate somewhat faster employment growthNew Jersey following the increase in the state

2geve

13 Because, in principle, the BLS sample contains thepopulation of fast-food restaurants in the designated chainan argument could be made that the OLS standard errounderstate the precision of the estimates. Nonethelesthroughout the paper we rely on conventional tests of statistical significance.

sn

y

’s

ite

dso

s

minimum wage. Only in the proportionatechange specifications without covariates, however, is the difference in growth rates betweeNew Jersey and Pennsylvania restaurants clto being statistically significant.

For comparison, panel C contains the corrsponding estimates from our original samplThese estimates differ (slightly) from those reported in our original paper because we nomeasure employment as the unweighted sumfull-time workers, part-time workers, and managerial workers to be comparable to the BLdata. The estimates based on our samplequalitatively similar to those based on the BLdata, with positive coefficients on the New Jesey dummy variable. In addition, in both datsets the inclusion of additional explanatory varables does not add very much to the explanatopower of the model.

D. Specification Tests

The BLS data analyzed in Tables 1 andsuggest that the New Jersey minimum-waincrease had either no effect, or a small positi

s,rss,-

rseyate

yof a


TABLE 3—SENSITIVITY OF NEW JERSEY EMPLOYMENT GROWTH DIFFERENTIAL TO

SPECIFICATION CHANGES; BLS ES-202 FAST-FOOD RESTAURANT SAMPLE

Specification and sampleChange in levels

(1)Proportionate change

(2)Sample

size

A. New Jersey and 7 Pennsylvania Counties

1. Basic specification 0.225 0.009 564(1.029) (0.029)

2. Excluding closed stores 0.909 0.031 549(0.950) (0.024)

3. Excluding closed stores unlessimputation code5 9

0.640 0.022 553(0.973) (0.025)

4. Drop large outlier 0.251 0.009 563(0.970) (0.028)

5. Proportionate change with initialemployment in base

— 20.001 564(0.032)

6. Excluding New Jersey shore 0.032 0.008 480(1.092) (0.030)

7. March 1992 to March 1993employment change

2.345 0.007 563(1.678) (0.035)

8. February 1992 to November 1992employment change

1.05 0.013 564(1.10) (0.032)

B. New Jersey and 14 Pennsylvania Counties

1. Basic specification 0.272 0.032 687(1.029) (0.024)

2. Excluding closed stores 0.639 0.055 671(0.776) (0.021)

3. Excluding closed stores unlessimputation code5 9

0.338 0.044 675(0.787) (0.021)

4. Drop large outliers 0.72 0.032 685(0.78) (0.023)

5. Proportionate change with initialemployment in base

— 0.020 687(0.024)

6. Excluding New Jersey shore 0.069 0.030 603(0.924) (0.025)

7. March 1992 to March 1993employment change

1.196 0.009 686(1.258) (0.027)

8. February 1992 to November 1992employment change

0.624 0.027 687(0.927) (0.024)

Notes:The table reports the coefficient (with standard error in parentheses) for the New Jedummy variable from a regression of the change in employment [column (1)] or proportionchange in employment [column (2)] on a New Jersey dummy variable, chain dummvariables, a dummy variable indicating whether the restaurant is reported as a subunitmultiestablishment employer, and a constant.

ino

m-m

sinne

theyem

en

a-3

a-gerare

od,

effect, on fast-food industry employmentNew Jersey vis-a`-vis eastern Pennsylvania. Tprobe this finding further, in Table 3 we exaine a variety of other specifications and saples. Panel A of the table presents results uour original 7 Pennsylvania counties, and paB uses the wider set of 14 counties. In all ofmodels, we include a full set of chain dummvariables and the subunit dummy variable. Rsults are reported for both the change in e

-gl

--

ployment specification [column (1)] and thproportionate employment growth specificatio[column (2)].

For reference, the first row replicates the bsic specifications from Table 2. Rows 2 andexamine the sensitivity of our results to alterntive choices for handling stores with missinemployment data in November–Decemb1992. In the base specification these storesassigned 0 employment in the second peri

laretseasisdemniasor

thinme

w

aetraa

rshs

hg

seuavi

er-nl-

n-the

-w

enttheistore8

chaseyr

elr

terenase

ony–er

-yutn-’stoch

at-n

e-rn


which is equivalent to assuming that they aclosed. Recall that some of these stores mhave actually remained open but changedporting identifiers. In row 2, we delete from thsample all stores with missing employment dain the second period; this is equivalent to asuming that all of these stores remained opbut were randomly missing employment datFinally, in row 3, we use the imputation codein the ES-202 database to attempt to distingubetween closed stores (with an imputation coof 9) and those that had missing data for othreasons. In particular, we add back to the saple any restaurant with missing employmedata (or those with 0 reported employment)they were assigned an imputation code indicing a closure. In our opinion, this is the moappropriate sample for measuring the effectthe minimum wage on the set of stores that wein business just before the rise in the minimumA comparison of the results in rows 2 and 3 withe base specifications indicates that eliminatstores with missing or zero second-period eployment, or including such stores only if thimputation code indicated the store was closetends to increase the coefficient on the NeJersey dummy variable.

Two of the observations in the sample haemployment increases about twice the mewave 1 size; the next largest increase was lthan the mean size.14 These large employmenchanges may have occurred because one fchisee acquired another outlet, or for other resons. To probe the impact of these two outliethey are dropped from the sample in row 4. Testimates are not very much affected by theobservations, however.

In Card and Krueger (1994) we calculated tproportionate change in employment with averaemployment over the two periods in the denomnator. (This procedure is widely used by analyof micro-level establishment data, e.g., Davisal., 1996.) This specification was selected becawe thought it would reduce the impact of mesurement error in the employment data. We haused that specification in Tables 2 and 3 of thpaper. The specification in row 5 of Table 3, however, measures the proportionate change in e

ar-ss

14 Large negative employment changes are more likelbecause of restaurant closings.

ly

e-

a-n.

her-

tft-tfe.

g-

d,

dnss

n--,

ee

eei-tstse-es-m-

ployment with the first-period employment in thdenominator. With this specification, New Jesey’s employment growth is slightly lower thathat in the 7-county Pennsylvania sample, athough employment growth in New Jersey cotinues to be greater than in Pennsylvania when14-county sample is used.

In row 6 we eliminate from the sample restaurants that are located in counties on the NeJersey shore. These counties may have differseasonal demand patterns than the rest ofsample. The results are not very different in thtruncated sample, however. Another wayhold seasonal effects constant is to compayear-over-year employment changes. In rowwe measure employment changes from Mar1992 to March 1993. This 12-month change hthe added advantage of allowing New Jersemployers more time to adjust to the higheminimum wage. The relative change in the levof employment in New Jersey is notably largewhen March-to-March changes are used.

Finally, in row 8 we measure employmenchanges from February 1992 to Novemb1992. As mentioned, these are the months whthe preponderance of data in our survey wcollected. It is probably not surprising that thesresults are quite similar to the base specificatiestimates, which use the average FebruarMarch 1992 and average November–Decemb1992 employment data.

On the whole, we interpret the BLS longitudinal data as indicating that fast-food industremployment growth in New Jersey was abothe same, or slightly stronger, than that in Pensylvania following the increase in New Jerseyminimum wage. It is nonetheless possiblechoose samples and/or specifications in whiemployment growth was slightly weaker inNew Jersey than in Pennsylvania. This is whone would expect if the true difference in employment growth was very close to zero. Inone of our specifications or subsamples do wfind any indication of significantly weaker employment growth in New Jersey than in eastePennsylvania.

II. Repeated Cross Sections from the BLSES-202 Data

As described above, we also used the quterly BLS ES-202 data to draw repeated cro

y

crease.


FIGURE 2. EMPLOYMENT IN NEW JERSEY AND PENNSYLVANIA FAST-FOOD RESTAURANTS, OCTOBER 1991TO SEPTEMBER 1997

Note: Vertical lines indicate dates of original Card-Krueger survey and the October 1996 federal minimum-wage inSource:Authors’ calculations based on BLS ES-202 data.

ods

enyl-etiee

omhihlto

enr,hstwn-14le

hey-

t intleentars

m.75yl-

te’sewom-eworm-

ng-edy

ill

sections of fast-food restaurants for the perifrom 1991 to 1997. We used these crossectional samples to calculate total employmfor New Jersey, for the 7 counties of Pennsvania used in our original study, and for thbroader set of 14 eastern Pennsylvania counin each month. Figure 2 summarizes the timseries patterns of aggregate employment frthese files. For each of the three geograpregions, the figure shows aggregate montemployment in the fast-food industry relativetheir respective February 1992 levels.

The figure reveals a pattern that is consistwith the longitudinal estimates. In particulabetween February and November of 1992—tmain months our survey was conducted—fafood employment grew by 3 percent in NeJersey, while it fell by 1 percent in the 7 Pensylvania counties and fell by 3 percent in thePennsylvania counties. Although it is possibto find some pairs of months surrounding tminimum-wage increase over which emplo

-t

s-

cy

t

e-

ment growth in Pennsylvania exceeded thaNew Jersey, on whole the figure provides litevidence that Pennsylvania’s employmgrowth exceeded New Jersey’s in the few yefollowing the minimum-wage increase.

A. The Effect of the1996Federal Minimum-Wage Increase

On October 1, 1996, the federal minimuwage increased from $4.25 per hour to $4per hour. This increase was binding in Pennsvania, but not in New Jersey, where the sta$5.05 minimum wage already exceeded the nfederal standard. Consequently, the same cparison can be conducted in reverse, with NJersey now serving as a “control group” fPennsylvania’s experience. This reverse coparison is particularly useful because any lorun economic trends that might have biasemployment growth in favor of New Jerseduring the previous minimum-wage hike w

o

tew

n-mbere

niaosia

tiom

ulds-maser-th

nnnin

e-ineenntactha-e

io

ldugeinnihe

taly

r-to

rss

ng-reen-

d-ntedfalf

,hat

-gndu-ide-ns

i-adasetsorWs-s-ed

15 The BNW data that we analyze were downloadedfrom www.econ.msu.edu in November 1997.

16 A referee pointed out to us that Neumark and Wascher(2000 Appendix A) offers a different rationale for takingover the data collection, namely, “to get data on all types ofrestaurants represented in CK’s data.”

ani hisv ple,r sam-


17 The most recent version of NW’s data set includesndicator variable for restaurants collected by EPI. Tariable shows a total of 81 restaurants in the EPI samepresenting the 72 restaurants in the original Berman

now have the opposite effect on our inferencethe effect of the minimum wage.

The results in Figure 2 clearly indicate greaemployment growth in Pennsylvania than in NeJersey following the 1996 minimum-wage icrease. Between September 1996 and Septe1997, for example, employment grew by 10 pcent in the 7 Pennsylvania counties and 2 percin New Jersey. In the 14-county Pennsylvasample employment grew by 6 percent. It is psible that the superior growth in Pennsylvanrelative to New Jersey reflects a delayed reacto the 1992 increase in New Jersey’s minimuwage, although we doubt that employment wotake so long to adjust in this high-turnover indutry. We also doubt that Pennsylvania’s strong eployment growth was caused by the 1996 increin the federal minimum wage, but there is ctainly no evidence in these data to suggest thathike in the federal minimum wage caused Pesylvania restaurants to lower their employmerelative to what it otherwise would have beenthe absence of the minimum-wage increase.

To more formally test the relationship btween relative employment trends and the mimum wage using the data in Figure 2, westimated a regression in which the dependvariable was the difference in log employmebetween New Jersey and Pennsylvania emonth, and the independent variables werelog of the minimum wage in New Jersey reltive to that in Pennsylvania and a linear timtrend. For the 7-county sample, this regressyielded a positive coefficient with at-ratio of5.2 on the minimum wage. Although we wounot necessarily interpret this evidence as sgesting that a higher minimum wage causemployment to rise, we see little evidencethese data that the relative value of the mimum wage reduced relative employment in tfast-food industry during the 1990’s.

III. A Reanalysis of the Berman-Neumark-Wascher (BNW) Data Set

A. Genesis of the BNW Sample

The conclusion we draw from the BLS daand our original survey data is qualitativedifferent from the conclusion NW draw fromthe data they collected in conjunction with Beman and the EPI. This discrepancy led us

f

r

er-nt

-

n

-e

e-t

-

t

he

n

-s

-

reanalyze the BNW data, devoting particulaattention to the possible nonrepresentativeneof the sample.15 Problems in the BNW samplemay have arisen because a scientific samplimethod was not used in the initial EPI datacollection effort, and because the data wecollected three years after the minimum-wagincrease, rather than before and after the icrease, as in our original survey.

A fuller account of the origins of the BNWsample is provided in our earlier paper (Card anKrueger, 1998). In brief, an initial sample of restaurants from two of the four chains included iour original study was assembled by EPI in la1994 and early 1995. According to Neumark anWascher (2000 Appendix A), this initial sample orestaurants was drawn partly by using informindustry contracts, and partly from a survey ofranchisees in theChain Operators Guide.Werefer to this initial sample of 71 observationsaugmented with data for one New Jersey store tclosed during 1992, as the “original Berman sample.” Following the release of early reports usinthese data by Berman (1995) and Neumark aWascher (1995a), data collection continued. Nemark and Wascher (1995b) reported that “to avoconflicts of interest we subsequently took over thdata collection effort from EPI, so that the remaining data came from the franchisees or corporatiodirectly to us.”16 During the period from March toAugust 1995 they added information for 18 addtional restaurants owned by franchisees who halready supplied some data to EPI, as wellinformation from 7 additional franchisees and onchain. We refer to this sample of 154 restauranas the Neumark-Wascher (NW) sample. Data f9 other restaurants were supplied by EPI after Ntook over data collection (see Neumark and Wacher, 1995b footnote 9). We include these 9 retaurants in the pooled BNW sample, but excludthem from the original Berman subsample anfrom the NW subsample.17

ei

etsn

ehednge

ar

6b09e7

ers-

,ests-art,

erta

rel-ndtad.

odinhentoll

to

ey 163;Jersey


ple and the 9 restaurants which were provided directlyEPI after March 1995.

TABLE 4—DESCRIPTIVE STATISTICS FOR LEVELS AND CHANGES IN EMPLOYMENT BY STATE, BNW DATA

Means with standard deviations in parentheses:Difference-in-differencesNew Jersey-Pennsylvania

(standard error)

New Jersey Pennsylvania

Before After Change Before After Change

Total payroll hours/35:1. Pooled BNW sample 17.5 17.5 20.1 15.1 15.9 0.8 20.85

(5.5) (5.9) (3.4) (4.0) (5.9) (3.5) (0.49)2. NW subsample 17.7 16.7 21.0 13.4 12.4 21.0 20.05

(6.1) (6.3) (3.3) (3.8) (4.9) (3.5) (0.61)3. Original Berman

subsample17.1 19.3 2.1 16.9 20.4 3.4 21.28(3.5) (4.3) (2.7) (3.4) (4.3) (2.1) (0.63)

Nonmanagement employment:4. Pooled BNW sample 24.8 28.4 3.6 29.0 31.3 2.2 1.39

(6.0) (6.8) (3.0) (5.5) (6.8) (4.7) (1.20)

Notes:See text for description of employment variables and samples. Sample sizes are as follows. Row 1: New JersPennsylvania 72. Row 2: New Jersey 114; Pennsylvania 40. Row 3: New Jersey 49; Pennsylvania 23. Row 4: New19; Pennsylvania 33.

Although NW attempted to draw a completsample of restaurants not included in the orignal Berman sample, they successfully collectdata for only a fraction of fast-food restauranin New Jersey and eastern Pennsylvania beloing to the four chains in our original study.18

We can obtain a lower bound estimate of thnumber of restaurants in this universe from tnumber of working telephone listings we founin January 1992 in the process of constructiour original sample. In New Jersey, where thgeographic boundaries of the sample frameunambiguous, we found 364 valid phone numbers, whereas the BNW sample contains 1restaurants (see Card and Krueger, 1995 TaA.2.1). In eastern Pennsylvania, we found 1working phone numbers in the 7 counties wsurveyed, whereas the BNW sample contains

rye-

18 Neumark and Wascher’s letter to franchisees statethat they planned to “reexamine the New Jersey-Pennsyvania minimum-wage study” and emphasized that they werworking “in conjunction with ... a restaurant-supported lob-bying” organization. This lead-in may have affected re-sponse patterns for restaurants with different employmentrends in New Jersey and Pennsylvania, accounting for thelow response rate. We asked David Neumark if he couldprovide us with the survey form that EPI used to gather theidata, and he informed us, “To the best of my knowledgethere was no form; this was all solicited by phone” (e-mailcorrespondence, December 8, 1997).

-d

g-

e-3le

2

restaurants in 19 counties.19 These comparisonssuggest that the BNW sample includes fewthan one-half of the potential universe of retaurants. If the BNW sample wererandomthiswould not be a problem. As explained belowhowever, several features of the sample suggotherwise. In particular, the Pennsylvania retaurants in the original Berman sample appeto differ from other restaurants in the data seand also exhibit employment trends that difffrom those in the more comprehensive BLS daset described above. Conclusions about theative employment trends in New Jersey aPennsylvania are very sensitive to how the dafor this small subset of restaurants are treate

B. Basic Results

Table 4 shows the basic patterns of fast-foemployment in the pooled BNW sample andvarious subsamples. The first three rows of ttable report data on NW’s main employmemeasure, which is based on average payrhours reported for each restaurant in Februaand November of 1992. Franchise owners rported their data in different time intervals—

dl-e

tir

r

19 BNW’s sample universe covers a broader region ofeastern Pennsylvania than ours because BNW define theirgeographic area based on our three-digit zip codes. Thesezip codes encompass 19 counties, although our sampleuniverse only included restaurants in 7 counties.

eee

leolfnteswSnnern-

inesese

veanr-eg-mllyl-

ple.r

inotoeenhen-an

-ese

erfp-lylers

-ortflea

inhe

antf-l-ris

y-ne

n-wtesyor

y-rdig-

21 To compare relative changes in hours and employees


weekly, biweekly, or monthly—for up to thre“payroll periods” before and after the rise in thminimum wage. NW converted the data (for thmaximum number of payroll periods availabfor each franchisee) into average weekly payrhours divided by 35. As shown in row 1 oTable 4, this measure of full-time employmefor the pooled BNW sample shows that storwere initially smaller in Pennsylvania than NeJersey (contrary to the pattern in the BLS E202 data), and that during 1992 stores in Pesylvania expanded while stores in New Jerscontracted slightly (also contrary to the pattein the BLS ES-202 data). The “difference-indifferences” of employment growth is shownthe right-most column of the table, and indicatthat relative employment fell by 0.85 full-timequivalents in New Jersey from the period jubefore the rise in the minimum wage to thperiod 6 months later.

In rows 2 and 3 we compare these relatitrends for restaurants in the original Bermsample and in NW’s later sample. The diffeence in relative employment growth in thpooled sample is driven by data from the oriinal Berman sample, which shows positive eployment growth in both states, but especiastrong growth in Pennsylvania. All 23 Pennsyvania restaurants in the original Berman sambelong to a single Burger King franchiseThus, any conclusion about the growth of aveage payroll hours in the fast-food industryNew Jersey relative to Pennsylvania hingesthe experiences of this one restaurant opera

The final row of Table 4 reports relativtrends in an alternative measure of employmavailable for a subset of restaurants in tpooled BNW sample—the total number of nomanagement employees. In contrast to the ptern for total payroll hours, nonmanagemeemployment rosefaster in New Jersey thanPennsylvania.20 Taken at face value, these findings suggest that the rise in the New Jersminimum wage was associated with an increa

20 Among the subset of stores that reported nonmanage-ment employment, the difference-in-differences in averagpayroll hours/35 is20.43, with a standard error of 0.55.Thus, there is no strong difference between relative payrohours trends in the pooled BNW sample and among thsubset of restaurants that reported nonmanagement emploment.

l

--

y

t

-

e

-

nr.

t

t-t

y

in employment and a small decline in hours pworker.21 Unfortunately, although one-half orestaurants in the original Berman sample suplied nonmanagement employment data, on10 percent of restaurants in the NW subsampreported it. Thus, the BNW sample available fostudying relative trends in employment versuhours is very limited.

C. Regression-Adjusted Models

The simple comparisons of relative employment growth in Table 4 make no allowances fother sources of variation in employmengrowth. The effects of controlling for some othese alternative factors are illustrated in Tab5. Each column of the table corresponds todifferent regression model fit to the changesemployment observed for restaurants in tpooled BNW sample.

Column (1) presents a model with onlyNew Jersey dummy: the estimated coefficiecorresponds to the simple difference-odifferences reported in row 1 of Table 4. Coumn (2) reports a model with only an indicatofor observations in the NW subsample. Thvariable is highly significant (t-ratio over 8) andnegative, implying that restaurants in the NWsubsample had systematically slower emploment growth than those in the original Bermasample. The model in column (3) explores theffect of chain and company-ownership cotrols. These are jointly significant and shoconsiderable differences in average growth raacross chains, with slower growth among RoRogers and KFC restaurants than Wendy’sBurger King outlets.

Finally, the model in column (4) includesindicators for whether the restaurant’s emploment data were derived from biweekly omonthly intervals (with weekly data the omittecategory). These variables are also highly s

e

lley-

it is convenient to work with logarithms, so scaling is not anissue. For the sample of 55 observations that reported bothnumbers of employees and hours, the difference-in-differ-ences of log payroll hours is20.018; the difference-in-differences of log nonmanagement employees is 0.066; andthe difference-in-differences of log employees minus loghours is 0.084 (t-ratio 5 2.28). Thus, the apparent oppositemovement in hours and employees is statistically significantfor this small sample.

.62

change in


TABLE 5—ESTIMATED REGRESSIONMODELS FORCHANGE IN AVERAGE PAYROLL HOURS/35, BNW DATA

Specification:

(1) (2) (3) (4) (5) (6) (7)

New Jersey 20.85 — — — 20.36 20.66 20.09(0.49) (0.44) (0.41) (0.42)

NW subsample (15 yes) — 23.49 — — 23.44 — —(0.42) (0.43)

Chain dummies:Roy Rogers — — 23.56 — — 23.14 21.98

(0.81) (0.85) (0.89)Wendy’s — — 20.85 — — 20.71 21.35

(0.67) (0.67) (0.70)KFC — — 26.51 — — 26.30 26.56

(0.90) (0.90) (0.89)Company-owned — — 20.89 — — 21.31 20.72

(0.76) (0.81) (0.95)Payroll data type:

Biweekly — — — 1.73 — — 1.65(0.52) (0.52)

Monthly — — — 22.60 — — 21.06(0.48) (0.89)

R2 0.01 0.23 0.41 0.30 0.23 0.10 0.45Standard error of regression 3.47 3.07 2.70 2.95 3.08 3.32 2

Notes:Standard errors are in parentheses. Sample consists of 235 stores. Dependent variable in all models is theaverage weekly payroll hours divided by 35 between wave 1 and wave 2.

orea-

taha

rs

ft

lyorfr

aseve-oenw

es

thehebyionnt.ntsv-theftandtheby

etedrs,heenr-

y

ltsentent-hed

nificant, and suggest that the reporting basisthe payroll data has a strong effect on measuemployment trends. Relative to restaurants thprovided weekly data (25 percent of the sample), restaurants that provided biweekly daexperienced faster hours growth between ttwo waves of the survey, while restaurants thprovided monthly data had slower hougrowth.

An important lesson from columns (1)–(4) oTable 5 is that a wide variety of factors affecmeasured employment growth in the BNWsample. Many of these factors are also highcorrelated with the New Jersey dummy. Fexample, a disproportionately high fraction oNew Jersey stores in the pooled sample weobtained by NW. Since the NW subsample hslower growth overall, this correlation might bexpected to influence the estimate of relatiemployment trends in New Jersey. Additionally, the Pennsylvania sample contains nonethe slow-growing KFC outlets. Thus, it may bimportant to control for these other factors wheattempting to measure the relative trend in NeJersey employment growth.

The models in columns (5)–(7) include thindicator for New Jersey outlets and variou

fdt

et

e

f

subsets of the other covariates. Notice thataddition of any subset of controls lowers tmagnitude of the New Jersey coefficient20–90 percent, and also improves the precisof the estimated coefficient by 10–15 perceNone of the estimated New Jersey coefficieare statistically significant at conventional leels once the other controls are included inmodel. Simply controlling for an intercept shibetween restaurants in the NW subsamplethe balance of the pooled data set reducessize of the estimated New Jersey coefficientover 50 percent.

The addition of controls indicating the timinterval over which the hours data were reporhas a particularly strong impact on scaled houand on any inference about the effect of tminimum-wage increase in these data. Evcontrolling for chain and ownership characteistics, the biweekly payroll indicator is highlstatistically significant (t 5 3.19). InCard andKrueger (1997 Appendix) we present resusuggesting that the differences in employmgrowth across reporting intervals are not drivby specific functional form assumptions or ouliers. We are unsure of the reasons for thighly significant differences in measure

tes-aconh

toalsu-r-ecaltere-forni

onas13estabv

,le

rmi

tstheeisea

rsleenntwcehe

rs,

hes-d

re-r-

ls-cht

oftole,eyrs-

feinu-


growth rates between restaurants that repordata over different payroll intervals, but we supect this pattern reflects differential seasonal ftors that systematically led to the mis-scalinghours in some pay periods. For example, marestaurants are closed on Thanksgiving. TThanksgiving holiday probably was more likelyhave been covered by monthly payroll intervthan weekly or biweekly ones, which would spriously affect the growth of hours worked. Unfotunately, Neumark and Wascher did not colldata on the number of days stores were actuopen during their pay periods, or on the dawhich were spanned by the pay periods coveby the data they collected.22 Consequently, no adjustment to work hours can be made to allowwhether stores were closed during holidays. Aother factor that may have affected changespayroll hours for restaurants that reportedweekly versus biweekly or monthly intervals wa massive winter storm on December 10–1992, which caused two million power outagand widespread flooding, and forced many eslishments in the Northeast to shut down for seeral days (seeElectric Utility Week,December 211992). Some pay intervals in the BNW sampmay have been more likely to include the stothan others, leading to spurious movementspayroll hours.23

Absent information on whether restauranwere closed because of Thanksgiving orDecember 10–13 storm for some part of thpay period, the best way to control for theextraneous factors is to add controls for the p

heW

wIna

watelr

22 In view of this fact, we disagree with Neumark andWascher’s (2000) assertion that because they collectehours worked for a “well-defined payroll period (which isspecified as either weekly, biweekly or monthly)” the BNWdata set should provide a more reliable measure of employment changes than our survey data. Because Neumark aWascher failed to collect the dates covered by their payrolperiods, or the number of days the store was in operatioduring their payroll periods, there are potential problemssuch as the correlation between employment growth and threporting interval that cannot be explained in their data.

23 These factors are unlikely to be a problem in ouroriginal survey data or in the BLS data because the numbeof workers on the payroll should be unaffected by tempo-rary shutdowns, and because the BLS consistently collecteemployment for the payroll period containing the 12th dayof the month. It is possible that weather and holiday factorsaccount for the contrasting results discussed previously fohours versus number of workers in the BNW data set.

d

-fye

tlysd

-n

,

--

n

r

y

period to the regression model for scaled houchanges. The results in column (7) of Tab5 show that the addition of controls for thpayroll reporting interval has a large effect othe estimated New Jersey relative employmeeffect, because a much lower fraction of NeJersey restaurants supplied biweekly data. Onthese differences are taken into account, temployment growth differential between NewJersey and Pennsylvania all but disappeaeven in the pooled BNW sample.

D. Alternative Specifications and Samples

An important conclusion that emerges fromTables 4 and is that the measured effects of tNew Jersey minimum wage differ between retaurants in the original Berman sample anthose in the subsequent NW sample. Table 6ports the estimated coefficient on the New Jesey dummy from a variety of alternative modefit to the pooled BNW sample, the NW subsample, and the original Berman sample. Earow of the table corresponds to a differenmodel specification or alternative measurethe dependent variable; each column refersone of the three indicated samples. For exampthe first row reports the estimated New Jerseffect from models that include no othecontrols: these correspond to the differencein-differences reported in Table 4.

Row 2 of the table illustrates the influence othe data from the single Burger King franchisethat supplied the Pennsylvania observationsthe original Berman sample. When the restarants owned by this franchisee are excluded, testimated New Jersey effect in the pooled BNsample becomes positive.24 Without this own-er’s data it is impossible to estimate the NeJersey effect in the original Berman sample.the NW sample, however, the exclusion hasnegligible effect.

Row 3 of Table 6 shows the estimated NeJersey coefficients from specifications thcontrol for chain and company ownership. Thresults in row 4 control for the type of payroldata supplied to BNW (biweekly, weekly, o

d

-ndln

e

r

d

r

24 This franchisee supplied data on 23 restaurants (all inPennsylvania) to the original Berman/EPI data-collectioneffort, and on three additional restaurants (all in New Jer-sey) to NW’s later sample.

has 71sample

ge payrollours forin panelfirst and


TABLE 6—ESTIMATED RELATIVE EMPLOYMENT EFFECTS IN NEW JERSEY FORALTERNATIVE SAMPLES AND

SPECIFICATIONS, BNW DATA

Pooled BNWsample

Neumark-Waschersample

Original Bermansample

A. Change in Average Payroll Hours/35

1. No controls 20.85 20.05 21.28(0.49) (0.61) (0.63)

2. Exclude first Pennsylvaniafranchisee, no controls

0.37 20.11 —(0.56) (0.62)

3. Controls for chain and ownership 20.66 20.27 20.95(0.41) (0.53) (0.64)

4. Controls for chain, ownershipand payroll period

20.09 20.22 0.58(0.42) (0.54) (0.72)

B. Change in Payroll Hours/35 Using First Pay Period per Restaurant

5. No controls 20.55 0.18 20.85(0.50) (0.63) (0.67)

6. Controls for chain, ownership,and payroll period

0.07 20.15 0.96(0.43) (0.52) (0.78)

C. Proportional Change in Average Payroll Hours/35

7. No controls 20.06 20.00 20.09(0.05) (0.07) (0.07)

8. Controls for chain and ownership 20.05 20.03 20.04(0.05) (0.07) (0.07)

9. Controls for chain, ownership,and payroll period

20.01 20.03 20.08(0.05) (0.07) (0.08)

Notes: Pooled BNW sample has 235 observations; NW sample has 154 observations; original Berman sampleobservations. In row 2, data for 26 restaurants owned by one franchisee are excluded. In this row only, pooled BNWhas 209 observations; and NW sample has 151 observations. Dependent variable in panel A is the change in averahours between the first and second waves, divided by 35. Dependent variable in panel B is the change in payroll hthe first payroll period reported by each store between the first and second waves, divided by 35. Dependent variableC is the change in average payroll hours between the first and second waves, divided by average payroll hours in thesecond waves. Standard errors are in parentheses.

-cIny

re

-lylei-h-

aisrnB

r-

hat-

hea

fi--

sinhere

tery-s ar-onthe

in

monthly). As noted, once controls for the payroll period are included, the New Jersey effefalls to essentially zero in the pooled sample.the original Berman sample, the New Jerseeffect becomes positive when controls aadded for the payroll period.

In most of their analysis NW utilize an employment measure based on the average scahours data taken over varying numbers of paroll periods across restaurants in their samp(Data are recorded for up to three payroll perods per restaurant in each wave). To check tsensitivity of the results to this choice, we constructed a measure using only thefirst payrollperiod for restaurants that reported more thone period. In principle, one would expect thalternative measure to show the same patteas the averaged data. As illustrated in panel(rows 5 and 6) of Table 6, the use of the alte

t

ed-.

e

n

s

native employment measure leads to results tare uniformly less supportive of NW’s conclusion of a negative employment effect in NewJersey. Even in the original Berman sample tuse of the simpler hours measure leads to33-percent reduction in the New Jersey coefcient, and yields an estimate that is insignificantly different from zero.

Finally, panel C of Table 6 reports estimatefrom models that use the proportional changeaverage payroll hours at each restaurant—ratthan the change in the level of averaghours—as the dependent variable. The latspecifications are more appropriate if emploment responses to external factors (such arise in the minimum wage) are roughly propotional to the scale of each restaurant. Inspectiof these results suggests that the signs ofNew Jersey effects are generally the same as

mos

lesihlletitithothrr

yt-leelel

eyte

n

e

aigu

epeesswges

entfrs,

heentnn-rnur

i-ofoaeic-heu-ndseeh.-antnehes”n

ntme

re.--

e-nde

l-lee

25 The first three digits of the postal zip code do notcorrespond to any conventional geographic entity.

26 The situation is more complex if the BNW data aretreated as a noisy measure of the truth, e.g., because ofsampling or nonsampling errors. In particular, letl j repre-sent the reliability of the observed employment changes (byzip code and chain) in surveyj ( j 5 1, 2). In this case, ifthe measurement errors in the two surveys are uncorrelated,the probability limit of the regression coefficient from alinear regression of the employment change in survey 1 onthe change in survey 2 isl2 (the reliability of the secondsurvey), and the probability limit of theR2 is l1 z l2 (theproduct of the reliability ratios).


the corresponding models for the levels of eployment, although the coefficients in the prportional change models are relatively leprecise.

Our conclusion from the estimates in Tab6 is that most (but not all) of the alternativeshow a negative relative employment trendNew Jersey, although the magnitudes of testimated effects are generally much smathan the naive difference-in-differences esmate from the pooled BNW sample. The esmated New Jersey effect is most negative inoriginal Berman sample. In the NW samplein the pooled sample that excludes data forPennsylvania franchisee who supplied Beman’s data, the relative employment effects asmall in magnitude and uniformly statisticallinsignificant (t-ratios of 0.7 or less). These paterns highlight the crucial role of the originaBerman data in drawing inferences from thBNW sample. Without these data (or more prcisely, without the observations from the singBurger King franchisee who provided the initiaPennsylvania data) the BNW sample providlittle indication that the rise in the New Jerseminimum wage lowered fast-food employmenEven with these data, once controls are includfor the payroll reporting periods, the differencebetween New Jersey and Pennsylvania are uformly small and statistically insignificant.

E. Consistency of the BNW Sample with thCard-Krueger and BLS Samples

Neumark and Wascher (2000) argue ththere is “severe measurement error” in our orinal survey data and argue at length that odependent variable has a higher standard dation than theirs. In Card and Krueger (1995 p71–72) we noted that our survey data containsome measurement errors, and tried to assthe extent of the errors by using reinterviemethods. Since measured employment chanare used as thedependentvariable in our anal-ysis, however, the presence of measuremerror does not in any way affect the validity oour estimates or our calculated standard erroprovided that the mean and variance of tmeasurement errors in observed employmchanges are the same in New Jersey and Pesylvania. Neumark and Wascher’s conceabout bias due to measurement error in o

--s

ner--ere-e

-

s

.d

si-

t-r

vi-.d

dependent variable is only relevant if the varable either contains no signal, or if the meansthe errors are systematically different in the twstates. To check the validity of our original datit is useful to compare employment trends in thtwo data sets at the substate level. The publuse versions of both data sets include only tfirst three digits of the zip code of each restarant, rather than full addresses. This limitationecessitates comparisons of employment trenby restaurant chain and “three-digit zip-codarea.”25 We also compare the BNW data to thBLS data at the chain-by-zip-code level, whicpoints up further problems in the BNW sample

A useful summary of the degree of consistency between the two samples is provided bybivariate regression of the average employmechanges (by chain and zip-code area) from osample on the corresponding changes from tother. In particular, if the employment changein the BNW sample are taken to be the “truechange for the cell, then one would expect aintercept of 0 and a slope coefficient of 1 froma regression of the observed employmechanges in our data on the changes for the sazip-code area and chain in the BNW data.26

This prediction has to be modified slightly if theemployment changes in the BNW sample a“true” but scaled differently than in our surveyIn particular, if the ratio of the mean employment level in our survey to the mean employment level in the BNW sample isk, then theexpected slope coefficient isk.

Table 7 presents estimation results from rgressing employment growth rates by chain azip-code area from our fast-food sample on thcorresponding data from the BNW sample. Athough 98 chain-by-zip-code cells are availabin our data set, only 48 cells are present in th

specificn changeata set.

data aret squaresrrors are


TABLE 7—COMPARISONS OFEMPLOYMENT GROWTH BY CHAIN AND ZIP-CODE AREA, CARD-KRUEGER DATA VERSUS

BERMAN-NEUMARK-WASCHER DATA

New Jersey andPennsylvania

New Jerseyonly

Pennsylvaniaonly

A. Using Combined BNW Sample

Intercept 20.32 0.41 23.91(0.56) (0.50) (1.77)

Change in employment in BNW sample 0.78 0.90 0.65(0.22) (0.19) (0.68)

R2 0.22 0.38 0.09Standard error 8.97 7.35 10.76p-value: intercept5 0, slope5 1 0.47 0.65 0.07Number of observations (chain3 zip-code cells) 48 37 11

B. Using Combined BNW Sample Excluding Data from One Franchisee

Intercept 20.26 0.36 23.52(0.54) (0.51) (1.67)

Change in employment in BNW sample 0.87 0.91 0.93(0.21) (0.20) (0.73)

R2 0.28 0.39 0.17Standard error 8.56 7.40 10.27p-value: intercept5 0, slope5 1 0.71 0.72 0.14Number of observations (chain3 zip-code cells) 46 36 10

Notes:Dependent variable in all models is the mean change in full-time employment for fast-food restaurants of achain in a specific three-digit zip-code area, taken from the Card-Krueger data set. Independent variable is the meain payroll hours divided by 35 for fast-food restaurants (in the same chain and zip-code area) taken from the BNW dIn panel B, restaurants in the BNW sample obtained from the franchisee who provided Berman’s Pennsylvaniadeleted prior to forming average employment changes by chain and zip-code area. All models are fit by weighted leasusing as a weight the number of observations in the chain-by-zip-code cell in the Card-Krueger data set. Standard ein parentheses.

lts)dain

neath7ipuhtlen

atge

en

1

ase

lva-tala-ean

rothectre

entle.eltan-

a.ts

m-myl-

pooled BNW sample. Column (1) shows resufor these cells, while columns (2) and (3present results separately for chain-by-zip-coareas in New Jersey and Pennsylvania. The dunderlying the analysis are also plottedFigure 3.

Inspection of Figure 3 and the regressioresults in Table 7 suggests that there is a rsonably high degree of consistency betweentwo data sets: the correlation coefficient is 0.4The two largest negative outliers are in zcodes containing a high proportion of restarants from EPI’s Pennsylvania sample. In ligof this finding, and the concerns raised in Tab6 about the influence of the data from the frachisee who supplied these data, we showparallel set of models in panel B of Table 7 thexcludes this owner’s data from the averachanges in the BNW sample.

Looking first at the top panel of the table, thregression coefficient relating the employmechanges in the two data sets is 0.78. AnF-testfor the joint hypothesis that this coefficient is

eta

-e.

-

-a

t

and that the intercept of the regression is 0 ha probability value of 0.47. Comparisons of thseparate results for New Jersey and Pennsynia suggest that within New Jersey the two dasets are in closer agreement. Across the retively small number of Pennsylvania cells thsamples are less consistent, although we conly marginally reject the hypothesis of a zeintercept and unit slope. Because we usedsame survey methods and interviewer to colledata from New Jersey and Pennsylvania, theis no reason to suspect different measuremerror properties in the two states in our sampA comparison of the results in the bottom panof the table shows that the exclusion of dafrom the franchisee who provided EPI’s Pensylvania sampleimprovesthe consistency of thetwo data sets, particularly in PennsylvaniWhile not decisive, this comparison suggesthat the key differences between the BNW saple and our sample are driven by the data frothe single franchisee who supplied the Pennsvania data for the Berman sample.

nisioth ted.

in


FIGURE 3. GRAPH OF CK DATA VS. BNW DATA AT THE

THREE-DIGIT ZIP-CODE-BY-CHAIN LEVEL

Notes:Shaded triangles indicate Berman/EPI Pennsylvasample. The line shown on the graph is the WLS regresfit using the subsample collected by NW. Weights arenumber of restaurants in the cell based on CK data.

ofhepesnl-mith

ae

iho

n

tsll-onter

e

tri-rso-

nu-s

-r

etn-e

b-otll.l-ei-ees

he

y-

To further explore the representativenessthe BNW data, the BLS ES-202 data and tBNW data were both aggregated up to the thrdigit zip-code-by-chain level for common zicodes and chains. Specifically, we calculataverage employment changes for establiments in each of these cells in the BLS data ain BNW’s data, and then linked the two cellevel data sets together. Because the BNW saple does not contain all of the restaurantseach cell, the sets of restaurants covered intwo data sets are not identical.27 Nonetheless, ifthe two samples are representative, the cellerages should move together. The resulting clevel data set was used to estimate a setregressions of the employment changeBNW’s data on the employment change in tBLS data. If the BNW sample is unbiased, nother variable should predict employmegrowth in that data set, conditional on true employment growth.

Column (1) of Table 8 reports coefficienfrom a bivariate regression in which the ceaverage employment change calculated frthe BNW data set is the dependent variable athe cell-average employment change calculafrom the BLS ES-202 data set is the explanatovariable. There is a positive relationship b

27 Only the subsample of cells with nonmissing data inboth cell-level data sets was used in the analysis.

an

e

TABLE 8—COMPARISONS OFEMPLOYMENT GROWTH BY

CHAIN AND ZIP-CODE AREA, BERMAN-NEUMARK-WASCHER

DATA VERSUSBUREAU OF LABOR STATISTICS

ES-202 DATA

(DEPENDENT VARIABLE: AVERAGE CHANGE IN

EMPLOYMENT, BNW DATA)

(1) (2) (3)

Average change inemployment, BLS data

0.17 0.10 0.09(0.05) (0.05) (0.05)

Fraction of sample collectedby EPI

— 3.46 —(0.70)

Fraction of sample collectedby EPI 3 Pennsylvania

— — 4.33(1.19)

Fraction of sample collectedby EPI 3 New Jersey

— — 3.22(0.75)

Constant 0.04 20.95 20.96(0.34) (0.34) (0.34)

R2 0.19 0.49 0.50

Notes: Weighted least-squares estimates are presenWeights are equal to the number of BNW observationsthe cell. Cells are composed of the three-digit zip-code-bchain areas. Standard errors are in parentheses.

.t,

ee-

dh-d

-ne

v-ll-ofne

t-

mddy-

tween the two measures of employmenchanges. Because the BNW employment vaable is scaled hours, not the number of workeemployed, one would not expect the slope cefficient from this regression to equal 1. Incolumn (2), we add a variable to the regressiomodel that measures the proportion of restarants in each cell of BNW’s data set that wacollected by EPI. Lastly, in column (3) we in-teract this variable with a dummy variable indicating whether the cell is in New Jersey oPennsylvania.

The results in column (2) indicate that thproportion of observations in BNW’s cells thawere collected by EPI has a positive effect oemployment growth, conditional on actual employment growth for the cell as measured by thBLS data. This finding suggests that the susample of observations collected by EPI are nrepresentative of the experience of the ceMoreover, the larger coefficient on the Pennsyvania interaction in column (3) suggests that thproblem of nonrepresentativeness in the orignal Berman data is particularly acute for thPennsylvania restaurants. Together with thother evidence in Tables 4–6, this finding leadus to question the representativeness of tEPI’s sample, and of the pooled BNW sample

Neumark and Wascher (2000) argue tha“The only legitimate objection to the validity of

oe

rtn,

i

nuob

isnoisamg

Eemaeur

elum.doev

bnhes,ingrer

t

e of

w

of

m-e

ith

emo-ofll-

nt

n,

ntseirethu-

28 At the restaurant level this proportionate gap is definedas the difference between the new minimum wage and therestaurant’s starting wage in wave 1 divided by the startingwage in wave 1. (The gap is set to zero if the starting wageis above the new minimum wage.) TheGAPmeasure in ourregressions is the weighted average of the restaurant-levelproportionate gaps, where the weights are the number ofrestaurants in the cell.


the combined sample is that some observatiadded by the EPI were not drawn from thChain Operators Guide,but rather were forfranchisees identified informally.” This assetion is incorrect, however, if personal contacwere used to collect data from some restauralisted in theGuide and not others. MoreoverNeumark and Wascher’s separate analysisrestaurants listed and not listed in theGuidedoes not address this concern. All of the restarants in their sample could have been listedthe Chain Operators Guide,and the samplewould still be nonrepresentative if personal cotacts were selectively used to encourage a sset of restaurants to respond, or if a nonrandsample of restaurants agreed to participatecause they knew the purpose of the survey.

F. Patterns of Employment Changes WithinNew Jersey

The main inference we draw from Table 7that the employment changes in the BNW aCard-Krueger data sets are reasonably highly crelated, especially within New Jersey. Larger dcrepancies arise between the relatively smsubsamples of Pennsylvania restaurants. A coparison of the BLS and BNW data sets also sugests that the Pennsylvania data collected byand provided to Neumark and Wascher skew thresults. The consistency of the New Jersey saples is worth emphasizing since, in our originpaper, we found that comparisons of employmgrowth within New Jersey (i.e., between restarants that were initially paying higher and lowewages, and were therefore differentially affectby the minimum-wage hike) led to the same concsion about the effect of the minimum wage as coparisons between New Jersey and Pennsylvania

To further check this conclusion we mergethe average starting wage from the first waveour original fast-food survey for each of thchain-by-zip-code areas in New Jersey with aerage employment data for the same chain-zip-code cell from the BNW sample. We thecompared employment growth rates from tBNW sample in low-wage and high-wage celldefined as below and above the median startwage in February–March 1992. The results asummarized graphically in Figure 4. As in ouoriginal paper, employment growth within NewJersey wasfasterin chain-by-zip-code cells tha

ns

-sts

of

u-n

-b-me-

dr--ll-

-PIir-

lnt-

d--

f

-y-

had to increase wages more as a consequencthe rise in the minimum wage.

We also merged the average proportionalgapbetween the wave-1 starting wage and the neminimum wage from our original survey to thecorresponding chain-by-zip-code averagesemployment growth from the BNW sample.28

We then regressed the average changes in eployment (DE) on the average gap measur(GAP) for the 37 overlapping cells in NewJersey. The estimated regression equation, wstandard errors in parentheses, is:

(1) DE 5 22.001 17.98GAP R250.09.(1.11) (9.75)

The coefficient on theGAP variable in BNW’sdata is similar in magnitude to the estimate wobtain if we use the New Jersey micro data froour survey to estimate the corresponding micrlevel regression (13.1 with a standard error6.6). Furthermore, if we estimate another celevel model with BNW’s employment data andadd dummy variables indicating the restaurachain, we obtain:

(2) DE 5 0.771 12.00GAP(0.78) (5.76)

1 Chain Dummies R2 5 0.78.

In this model, the coefficient on theGAPvariableis slightly smaller than in the bivariate regressiobut it has a highert-ratio. The positive estimatedcoefficients on theGAPvariable indicate that em-ployment rose faster at New Jersey restauralocated in areas that were required to raise thentry wage the most when the minimum wagincreased. The pattern of employment growrates within BNW’s sample of New Jersey resta

inmr-

dleg-deouin-ewo

sey

re94adesbler-

ur

ly,hateyeyr-ereew

toultsd-

u-gre-inge202

a

zeroles

vaniability

ip-coden the cell


29 The wage-gap variable was assigned a value offor all Pennsylvania restaurants. Including dummy variab

that indicate whether restaurants are located in Pennsylor New Jersey thus completely absorbs interstate variain the wage-gap variable.

FIGURE 4. AVERAGE FTE EMPLOYMENT IN LOW- AND HIGH-WAGE AREAS OF NEW JERSEY, BEFORE AND AFTER 1992MINIMUM -WAGE INCREASE

Notes:Average FTE employment is calculated from BNW data set. Restaurants were aggregated to the chain-by-zlevel, and divided into low-wage and high-wage areas based on whether the average starting wage for restaurants iin the CK data set was above or below the median starting wage in February–March 1992.

rants supports our original finding that the risethe minimum wage had no adverse effect on eployment growth at lower-wage relative to highewage restaurants in the state.

Neumark and Wascher (2000) also performecell-level analysis using a wage-gap variabHowever, they only selectively replicate our oriinal analysis. Neumark and Wascher (2000)scribe the analysis of the wage-gap variable in1994 paper as follows: “This experiment contues to identify minimum-wage effects off of thdifference in employment growth between NeJersey and Pennsylvania, but adds informationthe extent to which the minimum-wage increawould have raised starting wages in New JerseThis description is incomplete because it ignothe estimates in column (v) of Table 4 of our 19paper, which included dummy variables for broregions in Pennsylvania and New Jersey. In thestimates, identification of the wage-gap variaarises entirely from differences within New Jesey.29 Indeed, this was a major motivation for o

-

a.

-r

n

.”s

e

analysis of the wage-gap variable. UnfortunateNeumark and Wascher only report results texclude the region controls in their Table 5. Thdo report in their text, however, that when threstrict their sample just to cells within New Jesey, they find that restaurants in areas that wrequired to raise their wages the most by the NJersey minimum-wage increase also tendedhave faster employment growth. These resfrom within New Jersey confirm an essential fining of our original paper.

G. Other Evidence for the Eating andDrinking Industry

In the final section of their Comment, Nemark and Wascher present an analysis of aggate-level data for the entire eating and drinkindustry, taken from two publicly availablsources: the BLS-790 program and the ES-program. The former are only available on

drinking)rcentageis theare for 7

(3) use

ent rate.y adjustedember.


TABLE 9—ESTIMATES OF MINIMUM -WAGE EFFECTS ONEMPLOYMENT GROWTH IN THE EATING AND DRINKING INDUSTRY

FROM FEBRUARY/MARCH TO NOVEMBER/DECEMBER

BLS 790 data, New Jersey and PennsylvaniaES-202 data, New Jersey and7 counties of Pennsylvania

NW(1)

Revised data(2)

Revised data andcorrectly datedunemployment

(3)NW(4)

Correctly datedunemployment

(5)

Minimum-wage change 20.15 20.11 20.03 20.11 20.01(0.10) (0.10) (0.08) (0.10) (0.07)

New Jersey 20.05 20.38 20.41 0.25 0.21(0.95) (0.95) (0.80) (0.96) (0.72)

Change in unemploymenta 20.32 20.39 21.47 20.49 21.84(0.47) (0.46) (0.44) (0.46) (0.39)

R2 0.15 0.13 0.38 0.14 0.51

Notes:Standard errors are in parentheses. Models are estimated using changes in employment in SIC 58 (eating andbetween February/March and November/December of 1982–1996. Dependent variable in columns (1)–(3) is the pechange in statewide employment in SIC 58 from the BLS 790 program. Dependent variable in columns (4)–(5)percentage change in employment from the ES202 program; New Jersey data are statewide and Pennsylvania datacounties only. Columns (1) and (4) are taken from Neumark and Wascher (2000 Table 10). Models in columns (2)–revised BLS 790 data for 1995 and 1996. Sample size is 30 observations.

a In columns (1), (2), and (4) the change in unemployment represents the change in the annual average unemploymIn columns (3) and (5) the change in unemployment represents the difference between the average of the seasonallrates in February and March and the average of the seasonally adjusted rates in the following November and Dec

bywindsecent

oichgeterueary

tivb-the

ingthe

ar,mesu-rtor

-

td.s-reare

ch-e

oriodry/oftafit

30 The BLS-790 data are revised after their initial re-lease. We are uncertain of when Neumark and Wascherassembled their data set; however, their data for 1982–1994are identical to the data available as of January 1999. Thedata we used for the estimates in the table are the final BLSemployment estimates, and not subject to revision.

31 Similarly, using the revised data affects their estimatesbased on December-to-December changes. The minimum-wage coefficient from the model in panel A, column (4), oftheir Table 10 falls to20.12 (standard error 0.08) with therevised data.

statewide basis, while the latter are availablecounty, permitting a comparison between NeJersey and the 7 counties of Pennsylvaniacluded in our original survey. Neumark anWascher summarize their findings from thedata as providing “... complementary evidenthat [the] minimum wage reduces employmein the restaurant industry.”

Columns (1) and (4) of Table 9 reproduce twkey regression models from their Table 10, whsuggest a negative impact of the minimum waon employment. The specifications are fit to staspecific changes in employment between Febary/March and November/December of each yfrom 1982 to 1996, and include as explanatovariables the percentage change in the effecminimum wage in the state, an indicator for oservations from New Jersey, and the change inannual unemployment rate from the precedcalendar year to the calendar year in whichdata are observed.

As the other results in the table make clehowever, the estimated impacts of the minimuwage are extremely sensitive to minor changin the data or control variables used by Nemark and Wascher. In column (2) we repoestimates that simply replace NW’s data f

-

--r

e

1995 and 1996 with the revised BLS-790 employment data.30 The effect of the minimumwage is smaller and statistically insignifican(t 5 21.1) when the revised data are useNotice also that the minimum-wage effects etimated from the BLS-790 and ES-202 data asimilar when the revised data are used [compcolumns (2) and (4)].31 This similarity might beexpected since the BLS-790 data are benmarked to the ES-202 data. In column (3) wreport estimates from a model that controls fchanges in unemployment over the same peras the dependent variable (i.e., from FebruaMarch to November/December). The usechronologically aligned unemployment daleads to a very noticeable improvement in the

fiehse0sc

ne

8tn

atecy

e-

hlei-tgr

s

nlinoninyr

e

r

st-

lyy’ss---me

seysp.

etaetu-niae

nttool-inllbyenurn-neseiveonownndtive


of the model, and to a larger and more signicant coefficient on the unemployment ratMore importantly, it also leads to a mucsmaller estimated minimum-wage effect. Ashown in column (5), the effect is about thsame on estimates derived from the ES-2data. Controlling for properly aligned changein the unemployment rate, the estimated effeof the minimum wage is negligible.32

We have investigated a number of other extesions to the findings in Table 9. For exampladding two more years of data from the BLS-79series (albeit based on preliminary data for 199leads to coefficient estimates that are similarthose reported in column (3). Similarly, adding aadditional year of ES-202 data has little effect othe results in column (5). We also consideredalternative estimation method that regressesdifference in employment growth rates betweNew Jersey and Pennsylvania on the differenin the changes in minimum wages and unemploment between the states. This specificationperhaps most comparable to the “differencin-differences” specification used in our original paper. Theseresults are very consistent witthe estimates in columns (3) and (5): for exampthe coefficient of the relative minimum-wage varable in models for the BLS-790 employment dais 0.003 (standard error 0.11) without controllinfor relative unemployment, and 0.02 (standaerror 0.12) controlling for relative changes in unemployment.

Based on the findings in Table 9, and thefurther analyses, we conclude that changesthe minimum wage in New Jersey and Pensylvania over the 1980’s and 1990’s probabhad little systematic effect on employmentthe eating and drinking sectors of the twstates. Thus, contrary to the impression coveyed by Neumark and Wascher’s analysbased on the unrevised employment data aan incorrectly aligned aggregate unemploment variable, the aggregated BLS data aquite consistent with our findings from th

s

--

nt,ofs-h

32 Again, a similar pattern is found using the Decemberto-December changes in the BLS-790 data also analyzedNeumark and Wascher. In particular, the estimated coefficient of the minimum-wage variable falls to20.09 (stan-dard error 0.07) controlling for December-to-Decembechanges in unemployment.

-.

2

t

-,0)

o

nnhenes-

is-

,

a

d-

ein-

y

-sd-e

fast-food sector in Section I, and with ouoriginal survey results.

IV. Conclusion

After analyzing BNW’s data, our original sur-vey data, publicly available BLS data, and moimportantly, the BLS ES-202 fast-food establishment data, we reach the following conclusion:Theincrease in New Jersey’s minimum wage probabhad no effect on total employment in New Jersefast-food industry, and possibly had a small poitive effect.We have previously written that, because of frictions in the labor market, a minimumwage increase can be expected to cause sofirms to reduce employment and others to raiemployment, with these two effects potentiallcancelling out if the rise in the minimum wage imodest (Card and Krueger, 1995 especially p13–14). If this view is correct, then it would not bsurprising to find some specifications and dadefinitions that yield a negative impact of thminimum wage on employment. But we doubthat a representative survey of fast-food restarants in New Jersey and eastern Pennsylvawould show a significant adverse impact of thminimum wage on total employment.

The only data set that indicates a significadecline in employment in New Jersey relativePennsylvania is the small set of restaurants clected by EPI. Results of this data set standcontrast to our survey data, to the BLS’s payrodata, and to the supplemental data collectedNeumark and Wascher. The difference betwethe New Jersey-Pennsylvania comparison in ooriginal survey and BNW’s data cannot be recociled by inherent differences between a telephosurvey and administrative payroll records becauthe BLS ES-202 data are based on administratpayroll records. Instead, we suspect the commdenominator is that representative samples shstatistically insignificant and small differences iemployment growth between New Jersey aeastern Pennsylvania, while the nonrepresentasample informally collected for Berman produceanomalous results.

An alternative interpretation of the full spectrum of results is that the New Jersey minimumwage increase did not reduce total employmebut it did slightly reduce the average numberhours worked per employee. Neumark and Wacher (1995b) reject this interpretation. Althoug

-by-

r

reerien/heinursasisrsm

-,

henn

o

st-ll

-

A”e

nt.”


we are less quick to rule out this possibility, we askeptical about any conclusion concerning avage hours worked per employee that relso heavily on the informally collected BermaEPI sample, and the exclusion of controls for tlength of the reporting interval. Moreover, withNew Jersey the BNW data indicate that hogrewmoreat restaurants in the lowest wage areof the state, where the minimum-wage increawas more likely to be a binding constraint. Thfinding runs counter to the view that total houdeclined in response to the New Jersey minimuwage increase.

REFERENCES

Berman, Richard. “The Crippling Flaws in theNew Jersey Fast-Food Study.” Mimeo, Employment Policies Institute, Washington, DCMarch 1995.

Card, David and Krueger, Alan B. “MinimumWages and Employment: A Case Study of tFast-Food Industry in New Jersey and Pesylvania.”American Economic Review, Sep-tember 1994,84(4), pp. 772–93.

. Myth and measurement: The new ecnomics of the minimum wage. Princeton, NJ:Princeton University Press, 1995.

-s

se

-

-

-

. “A Reanalysis of the Effect of the NewJersey Minimum Wage Increase on the FaFood Industry with Representative PayroData.” Industrial Relations Section WorkingPaper No. 393, Princeton University, December 1997 [available at http://www.irs.princeton.edu/pubs/working_papers.html].

Chain Operators Guide.Various issues.Davis, Steven J.; Haltiwanger, John C. and Schuh,

Scott. Job creation and destruction. Cam-bridge, MA: MIT Press, 1996.

Electric Utility Week. “Two-Day ‘Nor’ Easter’Cuts Power to Nearly 2 Million in North-east.” December 21, 1992, p. 7.

Lane, Julia; Stevens, David and Burgess, Simon.“Worker and Job Flows.”Economics Letters,April 1996, 51(1), pp. 109–13.

Neumark, David and Wascher, William. “TheEffects of New Jersey’s Minimum WageIncrease on Fast-Food Employment:Re-Evaluation Using Payroll Records.Unpublished manuscript, Michigan StatUniversity, March 1995a; August 1995b.

. “Minimum Wages and Employment:A Case Study of the Fast-Food Industry iNew Jersey and Pennsylvania: CommenAmerican Economic Review, December2000,90(5), pp. 1362–96.

minimum wages and employment: a case study of the fast

Documents