labor supply of new york city cabdrivers: one day at a time ·...

LABOR SUPPLY OF NEW YORK CITY CABDRIVERSONE DAY AT A TIME

COLIN CAMERER

LINDA BABCOCK

GEORGE LOEWENSTEIN

RICHARD THALER

Life-cycle models of labor supply predict a positive relationship betweenhours supplied and transitory changes in wages We tested this prediction usingthree samples of wages and hours of New York City cabdrivers whose wages arecorrelated within days but uncorrelated between days Estimated wage elasticit-ies are signicantly negative in two out of three samples Elasticities of inexperi-enced drivers average approximately 2 1 and are less than zero in all threesamples (and signicantly less than for experienced drivers in two of threesamples) Our interpretation of these ndings is that cabdrivers (at least inexperi-enced ones) (i) make labor supply decisions ldquoone day at a timerdquo instead of inter-temporally substituting labor and leisure across multiple days and (ii) set a loosedaily income target and quit working once they reach that target

I INTRODUCTION

Dynamic models of labor supply predict that work hoursshould respond positively to transitory positive wage changes asworkers intertemporally substitute labor and leisure workingmore when wages are high and consuming more leisure whenits pricemdashthe forgone wagemdashis low (eg Lucas and Rapping[1969]) While this prediction is straightforward it has proveddifcult to verify Estimated elasticities of intertemporal substi-tution have generally been low and insignicant or even nega-tive whether they are based on aggregate [Mankiw Rotembergand Summers 1985] cohort [Browning Deaton and Irish 1985]or panel [Altonji 1986] data (see also Laisney Pohlmeier and

Many thanks to Bruce Schaller (NYC Taxi and Limousine Commission) fordata and helpful discussions James Choi Kim Morgan and Dov Rosenberg forresearch assistance Charles Brown Jeffrey Dominitz John Engberg John HamSeth Sanders and Lowell Taylor for helpful discussions Andrei Shleifer for theWeber quote two referees and editor Lawrence Katz for extraordinarily thoroughcomments and colleagues at many workshops The Caltech brown bag lunchseminar NBER Behavioral Labor Economics meeting and Behavioral EconomicsSummer Camp (sponsored by the Russell Sage Foundation) Carnegie Mellon Uni-versityrsquos Heinz School of Public Policy and Management the University of Cali-fornia (Irvine and Berkeley) Departments of Economics the MITHarvardBehavioral Economics seminar the University of Chicago Labor Workshop theJudgmentDecision Making Society and the Econometric Society meetings Ad-dress correspondence to the rst author at Division of Social Sciences 228ndash77California Institute of Technology Pasadena CA 91125 camererhsscaltechedu

q 1997 by the President and Fellows of Harvard College and the Massachusetts Instituteof TechnologyThe Quarterly Journal of Economics May 1997

Staat [1992] Pencavel [1986] and cf Mulligan [1995]) Howeverthese results are difcult to interpret because actual wagechanges are rarely transitory so the hypothesis of intertemporalsubstitution must be tested jointly along with auxiliary assump-tions about persistence of wage shocks formation of wage expec-tations etc As a result the frequently observed negative wageelasticities can plausibly be attributed to specication error

The ideal test of labor supply responses to transitory wageincreases would use a context in which wages are relativelyconstant within a day but uncorrelated across days In such asituation all dynamic optimization models predict a positive rela-tionship between wages and hours worked due to the negligibleimpact on life-cycle wealth of a one-day increase in wage (egMaCurdy [1981 p 1074])

Such data are available for at least one group of workersNew York City cabdrivers Drivers face wages that uctuate ona daily basis due to demand shocks caused by weather subwaybreakdowns day-of-the-week effects holidays conventions etcAlthough rates per mile are set by law on busy days driversspend less time searching for customers and thus earn a higherhourly wage These wages tend to be correlated within days anduncorrelated across days (ie transitory)

Another advantage of studying cabdrivers is that unlikemost workers they choose the number of hours they work eachday because drivers rent their cabs from a eet for a xed fee (orown them) and can drive as long as they like during a continuoustwelve-hour shift Furthermore most analyses of labor supplymeasure hours (and sometimes income) by self-reports For cab-drivers better measures of hours and income are available fromldquotrip sheetsrdquo the drivers ll out and from meters installed in cabswhich automatically record the fares

Using these data we investigate the relationship betweenwages and hours worked and nd little evidence for intertem-poral substitution Most of the elasticities we estimate are nega-tive drivers tend to quit early on high wage days and to drivelonger hours on low wage days Elasticities for inexperienceddrivers are around 2 1 for each of the three samples of cabdriverswe used in our study The results are robust to outliers and differ-ent specications There are several possible explanations forthese negative elasticities Some explanations can be ruled outbut others require more data to evaluate

Assuming that the alternative explanations for negative

QUARTERLY JOURNAL OF ECONOMICS408

elasticities are not correct two major conclusions can be drawnfrom the nding of negative elasticities Both conclusions pointto the importance of psychological factors that are not incorpo-rated in conventional dynamic models of labor supply

First it is difcult to explain negative wage elasticities witha model that has more than a one-day time horizon for decision-making Imagine for example that cabdrivers had an earningstarget (an idea we return to below) beyond which they derivedzero marginal utility of income If applied at the daily level sucha target would produce wage elasticities of 2 1 because as thewage increased on a particular day drivers would cut back theirhours proportionately to earn a daily income that just meets thetarget (since exceeding it adds no utility) However if a targetwere applied at even a two-day levelmdashie if drivers had a two-day earnings targetmdashestimated elasticities would be positive fora wide range of plausible specications Drivers would intertem-porally substitute between the two days working long hours onthe rst day if it turned out to be high wage day and cutting backon hours if it were a low-wage day Thus for plausible incomeutility functions a one-day time horizon for labor supply deci-sions is necessary to explain strongly negative wage elasticities

Second if drivers take a one-day horizon for elasticities tobe substantially negative requires the marginal utility of incometo drop substantially sharply around the level of average dailyincome Analytically this property is familiar as a high degree ofincome risk aversion A psychological account of the source of thishigh risk aversion which came from conversations with manycabdrivers is that drivers drive as if they have an income targetwhen they get near the target the probability of quitting for theday rises sharply (as if the marginal utility of income drops a lot)1

1 For example Weber [1958] wrote ldquo raising the piece-rates has oftenhad the result that not more but less has been accomplished in the same timebecause the worker reacted to the increase not by increasing but by decreasingthe amount of his work A man for instance who at the rate of 1 mark per acremowed 2ndash12 acres per day and earned 2ndash12 marks when the rate was raised to125 marks per acre mowed not 3 acres as he might easily have done thus earn-ing 375 marks but only 2 acres so that he could still earn the 2ndash12 marks towhich he was accustomedrdquo In their widely used microeconomics textbook Pindyckand Rubinfeld [1989 p 503] write about a student who has a one-summer hori-zon ldquoIn real life a backward-bending labor supply curve might apply to a collegestudent working during the summer to earn living expenses for the school yearAs soon as a target level of earnings is reached the student stops working andallocates more time to leisure activities An increase in the wage rate will thenlead to fewer hours worked because it enables the student to reach the targetlevel of earnings fasterrdquo

LABOR SUPPLY OF NYC CABDRIVERS 409

Such a target might be set at a driverrsquos average earnings levelat some round number such as $150 or by a simple formula suchas twice the daily fee for leasing the cab

Both the idea that cabdrivers make labor supply decisionsone day at a time and that they seem to have a target wage oraspiration level are consistent with much other research in psy-chology and economics Indeed this research motivated us tostudy the behavior of cabdrivers in the rst place because theone-day targeting hypothesis predicts negative elasticities andhence directly competes with the standard theory

Taking one day at a time is consistent with considerable re-search which suggests that people ldquobracketrdquo decisions narrowlysimplifying decisions by isolating them from the entire stream ofdecisions they are embedded in [Read and Loewenstein 1996]For example people are risk averse to single plays of smallgambles even though they typically face many uncorrelatedsmall risks over time that diversify away the risk of a single playAnother example closely related to the cabdriversrsquo daily deci-sions is betting on horse races Bettors seem to record the bettingactivity for each day in a separate ldquomental accountrdquo [Thaler1985] Since the track takes a percentage of each bet most bet-tors are behind by the end of the day Studies show that they tendto shift bets toward long shots in the last race in an attempt toldquobreak evenrdquo on the day [McGlothlin 1956] This implies that thebehavior on a given day depends much more on the outcome ofearlier bets that same day than on the outcome of bets on previ-ous days or on expectations of future days (in violation of a life-cycle theory of betting)

Narrow bracketing of decisions can produce other decisionanomalies that are not based on risk taking For example Readand Loewenstein [1995] conducted an experimental study ofvariety-seeking among trick-or-treaters on Halloween Childrenwho were told to take any two pieces of candy at a single housealways chose two different candies Those who chose one candyat each of two adjacent houses (from the same set of options) typi-cally chose the same candy at each house Normatively the chil-dren should diversify the portfolio of candy in their bag but infact they only diversify the candy from a single house Decisionisolation has also been observed in some strategic situationsJohnson et al [1996] found that subjects in a three-stage ldquoshrink-ing-pierdquo bargaining experiment often did not bother to look aheadand nd out how much the ldquopierdquo they bargained over wouldshrink if their rst-stage offers were rejected


The notion that drivers are averse to falling below a targetincome is also consistent with many other ndings There isample evidence from psychological studies that judgments anddecisions depend on a comparison of potential outcomes againstsome aspiration level or reference point [Helson 1964 Kahnemanand Tversky 1979 Tversky and Kahneman 1991] For bettors atthe track for example breaking even is a signicant clear refer-ence point In other situations reference points could be deter-mined by past income or consumption (eg Bowman Minehartand Rabin [1996]) by social comparison (eg Duesenberry[1949]) or by expectations for the future Regardless of what thereference points are the general nding is that people are ldquoloss-averserdquomdashthey dislike achieving outcomes below a reference pointabout twice as much as they like exceeding the reference point bythe same absolute amount2

Benartzi and Thaler [1995] use the same combination of nar-row bracketing and loss aversion that we use to explain the eq-uity premium puzzlemdashthe tendency for stocks to offer muchhigher rates of returns than bonds over almost any moderatelylong time interval In their model the equity premium compen-sates stockholders for the risk of suffering a loss over a short hori-zon They show that if investors evaluate the returns on theirportfolios once a year (taking a narrow horizon) and have apiecewise-linear utility function which is twice as steep for lossesas for gains then investors will be roughly indifferent betweenstocks and bonds which justies the large difference in expectedreturns If investors took a longer horizon or cared less aboutlosses they would demand a smaller equity premium Two papersin this issue [Thaler Tversky Kahneman and Schwartz 1997Gneezy and Potters 1997] demonstrate the same effect inexperiments

There is other eld evidence of narrow bracketing and lossaversion in stock trading and consumer purchases Investors whoown stocks that have lost value hold them longer than they holdldquowinningrdquo stocks before selling [Odean 1996 Weber and Camererforthcoming] Purchases of consumer goods like orange juice falla lot when prices are increased compared with how much pur-chases rise when prices are cut (ie price elasticities are asym-

2 Other applications of loss aversion include Kahneman Knetsch and Tha-ler [1990] on ldquoendowment effectsrdquo in consumer choice and contingent valuationof nonmarket goods Samuelson and Zeckhauser [1988] on ldquostatus quo biasesrdquoand Bowman et al [1996] and Shea [1995] on anomalies in savings-consumptionpatterns


metric eg Hardie Johnson and Fader [1993]) These datasuggest that like the trick-or-treaters mentioned above investorsand consumers isolate single decisionsmdashselling one stock or buy-ing one productmdashfrom the more general decisions about the con-tents of their stock portfolio or shopping cart (contrary to portfoliotheories in nance and the economic theory of consumer choice)Note that losses loom largest when decisions are isolated be-cause otherwise losses on a single stock or product can be com-bined with gains from other decisions in a single mental accountSo the assumptions of narrow bracketing and loss aversion rela-tive to a reference point are both needed to explain thesephenomena

II EMPIRICAL ANALYSES

In this section we use data on trip sheets of New York Citycabdrivers to explore the relationship between hours that driverschoose to work each day and the average daily wage A trip sheetis a sequential list of trips that a driver took on a given day Foreach trip the driver lists the time the fare was picked up anddropped off and the amount of the fare (excluding tip) Fares areset by the Taxi and Limousine Commission (TLC) For the rstperiod we study (1988) the fares were $115 per trip plus $15 foreach 15 of a mile or 60 seconds of waiting time For the secondperiod we study (1990 and 1994) fares were $150 per trip plus$25 each 15 of a mile or 75 seconds of waiting time In bothperiods a $50 per-trip surcharge is added between 8 PM and6 AM

Our data consist of three samples of trip sheets We describeeach data set briey here and include longer descriptions as Ap-pendix 1 The rst data set TRIP came from a set of 192 tripsheets from the spring of 1994 We borrowed and copied thesefrom a eet company Fleet companies are organizations that ownmany cabs (each afxed with a medallion which is required tooperate it legally) They rent these cabs for twelve-hour shifts todrivers who in our sample period typically paid $76 for a dayshift and $86 for a night shift The driver also has to ll the cabup with gas at the end of the shift (costing about $15) Driversget most of their fares by ldquocruisingrdquo and looking for passengers(Unlike many cities trips to the airport are relatively raremdasharound one trip per day on average) Drivers keep all the faresincluding tips The driver is free to keep the cab out as long as hewants up to the twelve-hour limit Drivers who return the cab


late are ned When a driver returns the cab the trip sheet isstamped with the number of trips that have been recorded on thecabrsquos meter This can then be used to determine how carefully thedriver has lled in the trip sheet

The measure of hours worked is obtained directly from thetrip sheet It is the difference between the time that the rst pas-senger is picked up and the time that the last passenger isdropped off Total revenue was calculated by adding up the fareslisted on the trip sheet The average hourly wage is total revenuedivided by hours worked

Many of the trip sheets were incomplete since the numberof trips listed by the cabdriver was much fewer than the numberof trips recorded by the meter Therefore we exclude trip sheetsthat listed a number of trips that deviates by more than two fromthe metered number This screen leaves us with 70 trip sheetsfrom thirteen drivers (eight of whom drive on more than one dayin the sample)

The advantage of the TRIP data set is that we can use thetrip sheets to measure the within-day autocorrelation in hourlyearnings as well as differences in earning across days Eventhough taxi fares are xed by the TLC earnings differ from dayto day because of differences in how ldquobusyrdquo drivers are that iswhether they spend most of the day with passengers in their cabor have to spend a lot of time searching for passengers

The second and third data sets of trip sheets were obtainedfrom the TLC3 The TLC periodically samples trip sheets to sat-isfy various demands for information about drivers and earnings(eg when rate increases are proposed) In these two data setshours and the number of driver-listed trips are obtained from thetrip sheets and the number of recorded trips fares and milesdriven is obtained from the meter

The TLC developed a screen to discard incomplete tripsheets To pass this screen the number of trips on the meter mustexactly match the number of trips listed by the driver and addi-tional criteria must also be met (see Appendix 1 for details) Be-cause the TLC provided us with the summary measures but notthe trip sheets themselves we are unable to create an alternativescreening procedure so we use their screened data for ouranalyses

The rst of the TLC data sets TLC1 is a summary of 1723

3 See NYTLC [1991 1992] for descriptive analyses of the NYC taxi businessbased on these data sets


trip sheets collected mostly during October 29 to November 51990 This data set includes three types of drivers daily eetdrivers lease-drivers who lease their cabs by the week or monthand others who own a medallion-bearing cab and drive it Mostowner-drivers rent their cab out to other drivers for some shiftsimposing constraints on when and how long they can drive Thosewho do not rent out their cabs can drive whenever they want

The screened data contain 1044 trip sheets and 484 drivers(234 of whom drove more than one day in the data) The mainadvantages of this sample are that it includes several observa-tions for each of many drivers and contains a range of differenttypes of drivers

The second TLC data set TLC2 is a summary of 750 tripsheets mostly from November 1ndash3 1988 This data set samplesowner-drivers as well as drivers from mini-eet companies (mini-eets usually lease cabs to drivers weekly or monthly) We dis-card 38 trip sheets using the TLC screen leaving us 712 tripsheets The main differences between TLC2 and TLC1 are thatno drivers appear more than once in the data in TLC2 and thefares set by the TLC in TLC2 are slightly lower

The analyses reported in the body of the paper use only thescreened samples of trip sheets for all three data sets Appendix3 reports sample statistics for the screened and ldquoscreened-outrdquodata for TRIP and TLC1 (TLC2 is not compared because so fewobservations are screened out) It also replicates the basic regres-sions reported in the paper including the screened-out data Nosubstantive conclusions are changed

To learn about important institutional details we conducteda phone survey of fourteen owners and managers at eet compa-nies that rent cabs to drivers The average eet in New York oper-ates 88 cabs so the responses roughly summarize the behavior ofover a thousand drivers The institutional details they reportedhelp make sense of the results derived from analysis of hours andincome data

Sample Characteristics

Table I presents means medians and standard deviations ofthe key variables Cabdrivers work about 95 hours per day takebetween 28 and 30 trips and collect almost $17 per hour in reve-nues (excluding tips) Average hourly wage is slightly lower in theTLC2 sample because of the lower rates imposed by the TLC dur-ing that time period The distributions of hours and hourly wages


TABLE ISUMMARY STATISTICS

Mean Median Std dev

TRIP (n 5 70)Hours worked 916 938 139Average wage 1691 1620 321Total revenue 15270 15400 2499 Trips listed on sheet 3017 3000 548 Trips counted by meter 3070 3000 572High temperature for day 7590 7600 821Correlation log wage and log hours 5 2 503 The standard deviation of log hoursis 159 log wage is 183 and log total revenue is 172 The within-driver standarddeviation of log revenue is 155 and across drivers standard deviation is 017TLC1 (n 5 1044)Hours worked 962 967 288Average wage 1664 1631 436Total revenue 15458 15400 4583 Trips counted by meter 2788 2900 915High temperature for day 6516 6400 859Correlation log wage and log hours 5 2 391 The standard deviation of log hoursis 263 log wage is 351 and log total revenue is 347 The within-driver standarddeviation of log revenue is 189 and across drivers standard deviation is 158TLC2 (n 5 712)Hours worked 938 925 296Average wage 1470 1471 320Total revenue 13338 13723 4074 Trips counted by meter 2862 2900 941High temperature for day 4929 4900 201Correlation log wage and log hours 5 2 269 The standard deviation of log hoursis 382 log wage is 259 and log total revenue is 400

are presented in Appendix 2 In the TRIP data the average tripduration was 95 minutes and the average fare was $513

One feature of the data is that the variation in hours workedand number of trips in the TRIP sample is substantially lowermdashabout half as largemdashas in the TLC1 and TLC2 samples Recallthat a key difference is that TRIP consists of only eet driverswho rent their cabs daily while TLC1 consists of eet lease andowner-drivers and the TLC2 consists of lease and owner-driversFigure II below is a distribution of hours broken up by driver-type for the TLC1 data It is clear from the histograms that thedifferences in variation in the key variables across data sets (seeAppendix 2) are driven by the differences in driver-types acrossthe data sets


Wage Variability within Days and between Days

In the empirical analyses below we estimate labor supplyfunctions using the daily number of hours as the dependent vari-able and the average wage the driver earned during that day asthe independent variable (both in log form) The average wage iscalculated by dividing daily total revenue by daily hours4 How-ever this assumes that the decisions drivers make regardingwhen to stop driving depend on the average wage during the dayrather than uctuations of the wage rate during the day

Within-day uctuations are important to consider becausenegatively autocorrelated intraday hourly wage rates could leaddrivers who are actually driving according to the predictions ofthe standard theory to behave as if they were violating it Ifautocorrelation is negative on a day with a high wage earlyin the day drivers will (rationally) quit early because high hourlywages are likely to be followed by low-wage hours Conversely ona day with low early wages drivers will drive long hours ex-pecting the wage to rise If hourly autocorrelations are zero orpositive however we can rule out this alternative explanation(unless drivers think the autocorrelation is negative when itis not)

To investigate how the hourly rate varied within the day weused the trip-by-trip data available in the TRIP sample Dayswere broken into hours and the median hourly wage for all driv-ers during that day and hour were calculated We then regressedthe median hourly wage (across drivers driving that hour) on theprevious hourrsquos median wage estimating an autocorrelation of493 (se 5 092)5 The second-order autocorrelation is even higher(578) and the third- and fourth-order autocorrelations are alsopositive and signicant When hourly wage is regressed on twoprevious lags both coefcients are greater than 40 and are sig-nicantly different from zero If we divide days into rst and sec-ond halves the correlation between median wages in the twohalves is 406 The patterns imply that when a day starts out as

4 This is similar to the method traditionally used in the labor supply litera-turemdashdividing yearly (or monthly) income by yearly (or monthly) hours to get thewage rate

5 Weighting the median observations by the number of drivers used to con-struct that observation did not change the standard error and changed the esti-mate only slightly to 512

6 The p-value of 15 for this correlation is higher than conventional levelsbut note that the sample size for this correlation is only fourteen (because eachobservation is a day)


a high wage day it will probably continue to be a high wage dayThe eet managers surveyed weakly agreed7 with these patternssaying the within-day autocorrelation is positive or zero (nonesaid it was negative)

Wages are signicantly different across days (p 0001 forTRIP and TLC1 too few days to permit a test for TLC2) Themedian (across drivers) of the average hourly wage for a dayranges from a low of $1393 to a high of $2062 in the TRIP dataand a low of $1556 to a high of $1935 in the TLC1 data Wagesare also virtually uncorrelated across days When we ran regres-sions of the mean or median wage on day t on the mean or medianwage on day t 2 1 the regression coefcient was 2 07 and insig-nicant (p 7)

Since wages are virtually uncorrelated across days andfairly stable within days they are ideal for calculating the laborsupply response to a transitory change in wage

Wage Elasticities

For each of the three data sets we calculate the simple corre-lation between (log) hours and (log) wages These statistics pro-vided in Table I are 2 503 2 391 and 2 269 Figure I showsscatterplots of log hours and log wages in the three sampleswhich corroborate the negative correlations Regressions of (log)hours on (log) wages are provided in Table II for the three datasets TRIP and TLC1 include multiple observations for eachdriver so either the standard errors are corrected to account forthe panel nature of the data or driver xed effects are included8

We also include two weather measures in the regression thehigh temperature for the day and a dummy variable for rain(which does not vary in TLC1 since it did not rain in that timeperiod) These variables control for shifts in labor supply that oc-cur if driving on a rainy day is more difcult and driving on a

7 Fleet managers were asked whether ldquoa driver who made more money thanaverage in the rst half of a shiftrdquo was likely to have a second half which wasbetter than average (three agreed) worse than average (zero) or about the sameas average (six) Expressing the target-income hypothesis two eet managersspontaneously said the second half earning were irrelevant ldquobecause drivers willquit earlyrdquo

8 The xed effects control for the possibility that drivers vary systematicallyin their work hours or their target income (see Section III) independent of thewage There are not enough observations per driver to allow driversrsquo elasticitiesto vary However we estimated some individual-driver regressions using the TRIPsample for those drivers with many daily observations Most of the wage elasticit-ies were signicantly negative


FIGURE IHours-Wage Relationships

warmer day has a higher opportunity cost (perhaps because for-gone leisure is more pleasurable) Also included is a dummy vari-able for the shift driven and a dummy variable for a weekdayversus weekend day (although all shifts are during the week inthe TLC2 data)9

9 Shifts are described in detail in Appendix 1 Briey in the TRIP and TLC2samples the dummy indicates night shift (versus day or afternoon) and in theTLC1 sample there are two shift dummy variables (night and day versus ldquootherrdquo)reecting the greater heterogeneity of driving arrangements in this sample Theestimates are changed very little if no shift designations are used No additional


TABLE IIOLS LOG HOURS WORKED EQUATIONS

Sample TRIP TLC1 TLC2

Log hourly wage 2 411 2 186 2 501 2 618 2 355(169) (129) (063) (051) (051)

High temperature 000 2 000 001 002 2 021(002) (002) (002) (002) (007)

Shift during week 2 057 2 047 2 004 030 mdash(019) (033) (035) (042)

Rain 002 015 mdash mdash 2 150(035) (035) (062)

Night shift dummy 048 2 049 2 127 2 294 2 253(053) (049) (034) (047) (038)

Day shift dummy mdash mdash 000 053 mdash(028) (045)

Fixed effects No Yes No Yes NoAdjusted R2 243 484 175 318 146Sample size 70 65 1044 794 712Number of drivers 13 8 484 234 712

Dependent variable is the log of hours worked Standard errors are in parentheses and are corrected forthe nonxed effects estimates in coulmns 1 and 3 to account for the panel structure of the data Explanatoryvariables are described in Appendix 1

In TRIP the wage elasticities depend substantially onwhether or not driver xed effects are included in the model Inthe rst column (no driver xed effects) the estimated wage elas-ticity is 2 411 and is signicantly different from zero Includingdriver xed effects which are jointly signicant lowers the esti-mated elasticity to 2 186 which is no longer signicantly differ-ent from zero10

improvement in t is obtained if day of the week dummy variables are includedrather than a weekday versus weekend dummy variable

10 One way to make use of the large amount of screened-out data in TRIPis to impute missing hours for the incomplete trip sheets by multiplying thedriver-listed hours by the ratio of meter-recorded trips to the number of driver-listed trips For example if a driver listed only 16 trips in 5 hours of driving butthe meter recorded 24 trips this method would impute 75 total hours of drivingThis method yields OLS estimates of 2 549 (se 5 156 n 5 162) and 2 276 (se 5071 n 5 158) for the TRIP sample without and with xed effects These estimatesare slightly more negative and more precisely estimated than those for thescreened sample reported in Table II Another method of imputation assumesthat drivers stopped lling out their trip sheets when they got busy (so that theaverage wage during the missing hours is higher than during the listed hours)This method scales up the number of hours by a factor that is less than the ratioof meter-recorded trips to driver-recorded trips (since it assumes the hours-per-trip is smaller for the missing trips) and actually makes the estimates evenmore negative


In the TLC1 and TLC2 samples elasticities are stronglynegative more precisely estimated and quite robust to includingxed effects (the estimates range from 2 355 to 2 618) In allthree samples analyses that reduce outlier inuence (such asmedian regression) indicate that the results are not sensitive tooutliers

The difference between the wage elasticities in these samplesand the xed-effects estimate in the TRIP sample is a bit curiousHowever recall that TRIP consists entirely of eet drivers (whopay daily) while the TLC samples also includes weekly andmonthly lease-drivers and owner-drivers Lease-drivers andowner-drivers have more exibility in the number of hours theydrive (since eet drivers are constrained to drive no more thantwelve hours) We report below (in Table V) that elasticities forthe eet drivers are substantially smaller in magnitude (lessnegative) than for lease- and owner-drivers The TRIP samplewhich is all eet drivers reects this compositional difference indriver types

Controlling for Measurement Error

Measurement error is a pervasive concern in studies of laborsupply Although the data on hours come from trip sheets ratherthan from memory they may include recording errors11 If thereis ldquoclassicalrdquo measurement error in hours (the errors are ldquowhitenoiserdquo and are uncorrelated with hours [Maddala 1992]) thisleads to a predictable bias in the wage elasticity Since the aver-age hourly wage is computed by dividing daily revenue by re-ported hours overstated hours will produce high hours-low wageobservations and understated hours produce low hours-highwage observations creating spuriously negative elasticities Thisbias can be eliminated if we can nd an instrument for wage thatis uncorrelated with the measurement error in hours We usesummary statistics of the distribution of hourly wages of otherdrivers that drove on the same day and shift (the 25th 50th and

11 Measurement error in income may also occur due to the omission of tipsSuppose that true income equals income from fares times (1 1 t) where t is theaverage tip percentage If 1 1 t is independent of fares when taking logs themeasurement error will be independent of measured income causing no bias inthe wage elasticity (Sherwin Rosen suggested that on high-demand days frus-trated passengers searching for cabs might add voluntary surcharges eg wavingmoney at cabdrivers This would cause a bias the highest hourly wages would bemost understated and the true elasticity would be even more negative than weestimate it to be)


75th percentiles) as instruments for own wage These instru-ments that summarize the ldquowagerdquo for the day should be uncorre-lated with a particular driverrsquos measurement error

The rst-stage regression of average wage on the 25th 50thand 75th percentiles of the other driver wage distribution is pre-sented in the bottom half of Table III The joint test of the nullhypothesis that all coefcients are zero can be easily rejected Thetop half of Table III reports estimated elasticities using these in-struments and including weather shift and weekday dummiesas explanatory variables The elasticities are less precisely esti-mated using the instrumental variables (as is common) but areeven more negative For TRIP and TLC1 estimates with andwithout xed effects are included The basic ndings from TableII are maintained in the IV estimation elasticities are negativeand signicantly different from zero except in the TRIP samplewhen xed effects are included

The results in Table III are quite robust with respect to vari-ous specications We also estimated specications that used asinstruments 1) the mean wage of other drivers on the same dayand shift 2) the 25th 50th and 75th percentiles of the other driv-ers on that day wage distribution although not broken down byshift and 3) percent of miles driven that are ldquoliverdquo (during whicha passenger is in the cab) The basic results are unchanged whenthese other specications are used12

How Do Elasticities Vary with Experience

Drivers may learn over time that driving more on high wagedays and less on low wage days provides more income and moreleisure If so the labor supply curve of experienced drivers wouldhave a more positive wage elasticity than that of inexperienceddrivers There are good measures of driver experience in thesedata sets In the TLC data sets the TLC separated drivers intoexperience groups for TLC1 those with greater or less than fouryears of experience and in TLC2 those with greater or less thanthree years of experience These group measures are absent inthe TRIP data However cabdriver licenses are issued with six-

12 In unreported regressions we also tried using daily subway ridership asan instrument for wage However this instrument did not predict wages well inthe rst stage We tried to obtain data on hotel occupancy or convention atten-dance but could not Note that conventions are an ideal instrument because theyare most likely to shift demand without also shifting the disutility of effort (andhence the supply curve)


digit numbers (called hack numbers) in chronological order sothat lower numbers correspond to drivers who obtained their li-censes earlier13 Using their license numbers we use a median

13 This is not a perfect measure of actual experience because for examplesome drivers with old licenses may only drive sporadically However licensesmust be renewed each year so that sufciently inactive drivers probably let theirlicenses expire

TABLE IIIIV LOG HOURS WORKED EQUATIONS


Log hourly wage 2 319 005 2 1313 2 926 2 975(298) (273) (236) (259) (478)

High temperature 2 000 2 001 002 002 2 022(002) (002) (002) (002) (007)


Rain 2 007 2 001 mdash mdash 2 130(042) (041) (070)


Day shift dummy mdash mdash 2 030 068 mdash(038) (048)

Fixed effects No Yes No Yes NoSample size 70 65 1044 794 712Number of drivers 13 8 484 234 712

Dependent variable is the log of hours worked Standard errors are inparentheses and are corrected for the nonxed effects estimates in columns 1 and3 to account for the panel structure of the data Instruments for the log hourlywage include the summary statistics of the distribution of hourly (log) wages ofother drivers on the same day and shift (the 25th 50th and 75th percentiles)

First-stage regressions

Median 316 026 2 385 2 276 1292(225) (188) (394) (467) (4281)

25th percentile 323 287 693 469 2 373(160) (126) (241) (332) (3516)

75th percentile 399 289 614 688 479(171) (149) (242) (292) (1699)

Adjusted R2 374 642 056 206 019P-value for F-test of 000 004 000 000 020instruments for wage

Dependent variable is the log of average hourly wage Standard errors are inparentheses Regressions also include weather and shift characteristics (dummyvariable for rain high temperature during the day dummy variable for shift on aweekday and time of shift dummy variables) as explanatory variables


TABLE IVIV LOG HOURS WORKED EQUATIONS BY DRIVER EXPERIENCE LEVEL


Experience level Low High Low High Low HighLog hourly wage 2 841 613 2 559 2 1243 2 1308 2220

(290) (357) (406) (333) (738) (1942)Fixed effects Yes Yes Yes Yes No NoSample size 26 39 319 458 320 375P-value for difference 030 666 058in wage elasticity

Dependent variable is the log of hours worked Standard errors are in parentheses Regressions alsoinclude weather and shift characteristics (dummy variable for rain high temperature during the day dummyvariable for shift on a weekday and time of shift dummy variables) as explanatory variables Instrumentsfor the log hourly wage include the summary statistics of the distribution of hourly (log) wages of otherdrivers on the same day and shift (the 25th 50th and 75th percentiles)

split to divide drivers into low- and high-experience subsamplesfor the TRIP data14

Table IV presents the wage elasticities estimated separatelyfor low- and high-experience drivers All regressions use instru-mental variables and all include xed effects (except of coursefor TLC2) In all three samples the low-experience elasticityis strongly negative generally close to 2 1 The wage elasticityof the high-experience group is signicantly larger in magni-tude for the TRIP and TLC2 samples (p 5 030 and 058respectively)15

How Do Elasticities Vary with Payment Structure

The way drivers pay for their cabs might affect their respon-siveness of hours to wages if for example the payment structureaffects the horizon over which they plan Alternatively it mightaffect the degree to which they can signicantly vary hours acrossdays The TLC1 sample contains data from three types of pay-ment schemes daily rental (eet cabs) weekly or monthly rental(lease cabs) or owned Table V presents elasticity estimates in

14 The number of observations in the low- and high-experienced samples forthe TRIP data are not equal because the median split is done on drivers not tripsheets and there are different sample sizes for each driver

15 An alternative approach is to use the median wage directly as a regressorskipping the rst-stage regression This lowers the adjusted R2 substantially (asis expected) but does not alter the sign or magnitude of the estimates reported inTable III systematically (TRIP and TLC2 estimates become more negative andTLC1 estimates become less negative) The large estimate and standard error onthe high-experience TLC2 elasticity reported in Table IV do become smaller( 2 135 and 968 respectively) but that does not change the conclusion that expe-rience makes elasticities less negative


TABLE VIV LOG HOURS WORKED EQUATIONS BY PAYMENT STRUCTURE TLC1 DATA

Type of cab Fleet Lease OwnedLog hourly wage 2 197 2 978 2 867

(252) (365) (487)Fixed effects Yes Yes YesSample size 150 339 305

Dependent variable is the log of hours worked Standard errors are in parentheses Regressions alsoinclude weather and shift characteristics (dummy variable for rain high temperature during the day dummyvariable for shift on a weekday and time of shift dummy variables) as explanatory variables Instrumentsfor the log hourly wage include the summary statistics of the distribution of hourly (log) wages of otherdrivers on the same day and shift (the 25th 50th and 75th percentiles) Fleet cabs are rented daily leasedcabs are rented by the week or month and owned cabs are owned by the drivers

the three payment categories from the TLC1 sample All regres-sions are estimated using instrumental variables and includedriver-xed effects

All wage elasticities in Table V are negative The elasticitythat is smallest in magnitude for eet drivers is not signicantlydifferent from zero The lease- and owner-driver wage elasticitiesare approximately 2 9 and are signicantly different from zeroPart of the explanation for the lower elasticity for eet drivers isa technical one Since they are constrained to drive no more thantwelve hours the dependent variable is truncated biasing theslope coefcient toward zero

Could Drivers Earn More by Driving Differently

One can simulate how income would change if driverschanged their driving behavior Using the TLC1 data we takethe 234 drivers who had two or more days of data in our sampleFor a specic driver i call the hours and hourly wages on a spe-cic day t hit and Wit respectively and call driver irsquos mean hoursover all the days in the sample hi By construction the driverrsquosactual total wages earned in our sample is S thitWit

One comparison is to ask how much money that driver wouldhave earned if he had driven hi hours every day rather than vary-ing the number of hours (ie if his labor supply curve of hoursagainst wages was at) Call this answer ldquoxed-hours earningsrdquo(FHE) S thiWit

Is FHE greater than actual earnings We know that on aver-age hit and wit are negatively correlated so that the differencebetween FHE and actual earnings will be positive in general Infact drivers would increase their net earnings by 50 percent onaverage (stderror 5 04 percent) if they drove the same number


of hours (hi) every day rather than varying their hours every dayIf we exclude drivers who would earn less by driving xed hours(because their wage elasticity is positive) the improvement inearnings would average 78 percent And note that if leisure util-ity is concave xed-hours driving will improve overall leisureutility too

These increases in income arise from following the simplestpossible advicemdashdrive a constant number of hours each day Sup-pose instead that we hold each driverrsquos average hours xed butreallocate hours across days as if the wage elasticity was 1 1Then the average increase in net income across all drivers is 10percent Across drivers who gain the average increase is 156percent16

III WHY MIGHT LABOR SUPPLY BE DOWNWARD-SLOPING

Our results lend support to the common nding that elastici-ties are not strongly positive for temporary changes in wages In-deed wage elasticities estimated with instrumental variables aresignicantly negative in two out of three samples Two additionaleffects we observe are that wage elasticities are signicantlyhigher for experienced drivers in two of three samples and wageelasticities are signicantly more negative for lease- and owner-drivers than for eet drivers These two additional regularitiesalong with other patterns in the data as well as informationgleaned from our telephone survey of eet managers allow us toevaluate alternative explanations for the observed negative elas-ticities We begin by discussing the explanation we favor thenevaluate three others suggested by colleagues referees and con-tentious friends

Daily Income Targeting

As explained in the Introduction one possible explanationfor the negative hours elasticities is that cabdrivers take a one-day horizon and set a target (or target range) and quit when the

16 Still another gure one can compute is the optimal reallocation of hoursto earn the largest possible wage total This calculation will yield a wage elasticitysubstantially larger than the 1 1 value used above But such a calculation willrequire drivers to work 12-hour shifts (or longer for eet and owner-drivers with24-hour shifts) on all the high-wage days and quit very early on low-wage daysThis pattern will raise variation into leisure hours (which will lower overall utilityif variation in leisure is undesirable) Without some accounting for the utility offorgone leisure simply knowing how much more income the drivers would earnis not of much interest


target is reached This decision rule can be modeled by marginalutility of income declining substantially around the average dailyincome level This explanation was suggested to us by severaldrivers in New York City and also rang true to many of the eetmanagers we surveyed They were asked to choose which one ofthree sentences ldquobest describes how many hours cabdrivers driveeach dayrdquo Six eet managers chose ldquoDrive until they make acertain amount of moneyrdquo Five chose the response ldquoFixed hoursrdquoOnly one chose the intertemporal substitution response ldquodrive alot when doing well quit early on a bad dayrdquo (One manager saidldquoall of the above it depends on the driverrdquo)

While daily income targeting may seem ad hoc to laboreconomists it is as we discussed in the Introduction consistentwith general principles of decision-making that have been ob-served in many other domains In fact the theory we use here isvery similar to that used by Benartzi and Thaler [1995] in theirpaper about the equity premium puzzle and is implicit in evi-dence of disposition effects in stock trading and asymmetric priceelasticities in consumer brand choice

A utility function for daily income with a target referencepoint could result from various underlying psychological pro-cesses For example targeting is a simple decision rule it re-quires drivers to keep track only of the income they have earnedThis is computationally easier than tracking the ongoing balanceof forgone leisure utility and marginal income utilitymdashwhich de-pends on expected future wagesmdashrequired for optimal intertem-poral substitution Working a xed number of hoursmdashldquohours-targetingrdquomdashis equally simple but drivers (especially inexperi-enced ones) may not realize that this alternative rule generatesmore income and more leisure Note that a weekly or monthlyearning target is much more difcult to implement because adriver would need to decide how much to earn on each day (giventhe wage opportunity cost of time etc on that day) A daily earn-ings target produces a much simpler rule simply drive until oneearns the target

Daily targets can also serve a second purpose like manymental accounts they help mitigate self-control problems (seeShefrin and Thaler [1992])17 There are two kinds of self-control

17 The use of a short horizon and income target to avoid temptation sug-gests that these features can be thought of as a self-imposed liquidity constraintbut could also be empirically distinguished from liquidity constraint imposed bylimited wealth and borrowing power


problems drivers might face First driving a cab is tedious andtiring and unlike many jobs work hours are not rigidly set Driv-ers are free to quit any time they want A daily income goal likean author imposing a daily goal of written pages establishes anoutput-based guideline of when to quit A weekly or monthly tar-get would leave open the temptation to make up for todayrsquos short-fall tomorrow or next week and so on in an endless cycle

Drivers could also keep themselves from quitting too earlyby setting daily targets that vary positively with wages early inthe day (ie they plan to work longer hours after a few goodhours and allow themselves to quit early after a few bad hours)Such a wage-dependent targeting rule creates intertemporal sub-stitution but also creates a second self-control problem driversmust save the windfall of cash they earn from driving long hourson a high-wage day so they can afford to quit early on low-wagedays But a drive home through Manhattan with $200ndash$300 incash from a good day could be an obstacle course of temptationsfor many drivers Given these two self-control problems substi-tuting over a weekly or monthly horizon may be too difcult sodaily targeting results Of course like most self-control strate-gies it yields less income and leisure than a person with perfectself-control would earn

A strong form of the target income hypothesis in which thetarget is constant across days and is the same for all driverscan be easily rejected This hypothesis predicts that daily incomeshould not vary much across days but it clearly does (see TableI) And the fact that (log) daily income has more variance within-drivers than across-drivers (see Table I again) implies that tar-gets vary more across days than across drivers The constant-target hypothesis also predicts the log hours-log wage relationwill be linear but adding a quadratic term improves tsignicantly

While the constant-target hypothesis can be rejected incometargeting in some form is useful for explaining two features ofthe data First for drivers with a one-day horizon and additivelyseparable income and leisure utility income utility must be quiteconcave around the average income level to explain elasticities asextremely negative as 2 1 which are evident in the inexperi-enced-driver regressions (Table IV)18 Strong concavity is of

18 Assume a one-day horizon no nonwage income wage 5 w hours 5 hincome y 5 hw and leisure L 5 24 2 h and an additively separable utility func-


course a possible feature of any utility function The idea thatworkers are ldquoloss-averserdquo around an income target (they dislikefalling short much more than they like exceeding it) is one simpleexplanation of extreme concavity that is consistent with muchother evidence

Second there are fewer low-hours days among eet driverswho pay daily (in the TLC1 sample see Figure II) The reluctanceto work short days is consistent with the hypothesis that eetdrivers use their daily lease fee as one reference point and areparticularly reluctant to quit before reaching it (compared withlease-drivers who pay weekly or monthly and owner-drivers)

The daily income-target hypothesis also seems to account forthe effect of experience rather naturally experienced drivers whohave larger elasticities either learn over time to take a longerhorizon (and to resist the temptations of quitting early andsquandering cash from good days) or to adopt the simple rule ofdriving a xed number of hours each day (Similarly we suspectthat experienced gamblers are less likely to allow within-day out-comes to inuence their subsequent choices They learn ldquonot tocount the money while theyrsquore sitting at the tablerdquo) Alternativelysome drivers may just lack these qualities They will have lessleisure and income and will be selected out of the experienced-driver pool Either way experienced drivers will have more posi-tive wage elasticities

Liquidity Constraints

Negative elasticities could occur because cabdrivers facestrongly binding liquidity constraints Liquidity-constraineddrivers who must earn a certain amount of money each day mustdrive long hours when wages are low This explanation seems un-likely for two reasons

First according to our eet manager survey almost all lease-drivers pay their weekly or monthly fees in advance Most of the

tion v(y) 1 u(L) with v() and u() both concave Assuming workers maximize util-ity and differentiating gives the elasticity equation (dhdw)(wh) 5 (1 2 yr(y))(yr(y) 1 hr(L)) where r(y) 5 2 v0 (y)v0 (y) and r(L) 5 2 u0 (L)u9 (L) are risk-aversion coefcients For u(L) concave (r(L) 0) the elasticity becomes negativefor r(y) 1y (eg more concave than log utility) The elasticity becomes increas-ingly negative as r(y) gets larger but does not reach 2 1 unless r(y) become in-nite (corresponding to a kink at the income target reference point cf Bowman etal [1996]) If leisure utility is convex or if leisure and daily income are strongcomplements then it is easier to generate negative elasticities (then a wage in-crease raises income holding hours xed which triggers an increase in leisureutility and causes an optimizing worker to cut hours and consume more leisure)


FIGURE IIHistograms of Hours Worked by Driver Ownership Class

(TLC1 only)

eet drivers pay at the end of the day but most eet managerssaid drivers could sometimes pay later Since lease-drivers pay inadvance and eet drivers can pay late most drivers do not needto drive long hours on low-wage days to scrape together enoughcash to pay the lease fee immediately at the end of their shift

Second the liquidity constraint argument implies thatpeople who are not liquidity constrainedmdashsuch as those withsubstantial wealthmdashshould not display negative elasticities Atthe time the data were generated cab medallions were worth


about $130000 so owner-drivers presumably have substantiallymore wealth or borrowing power than nonowners Therefore me-dallion ownership is a weak proxy for wealth If daily liquidityconstraints are responsible for the negative wage elasticitiesdrivers who own their cabs would have larger elasticities thandrivers who rent from eets or lease The empirical results inTable V show the opposite

Breaks

The trip sheets used to measure work hours do not distin-guish between idle time spent searching for fares and consciousbreaks that might be considered leisure If drivers are taking lotsof breaks on low-wage days for example that could explain whythey appear to work longer hours on those days than on high-wage days Perhaps if we could subtract these leisure breaksfrom hours worked the true wage elasticity would be morepositive

We do not have good data on the amount of break time driv-ers take but there are three reasons to think self-administeredbreaks do not explain all three regularities First in early anal-ysis using the TRIP sample breaks of more than 30 minutes wereremoved when calculating hours The results were similar tothose reported here Second various assumptions about how theamount of break time varies with wages help bound the effectthat excluding breaks would have For a plausible range of as-sumptions the true wage elasticity will not be positive if the mea-sured elasticity is negative19 Third to explain the increasedelasticities of experienced drivers requires the assumption that

19 Call measured hours m true (unobserved) hours t and breaks b Bydenition t 5 m2 b Taking derivatives dtdw 5 dmdw 2 dbdw Noting thatdmdw appears to be negative in our analyses we can ask how dtdw wouldchange for plausible values of a break response dbdw If breaks are xed inlength across days (eg breaks are taken for meals or coffee) then dbdw 5 0and dtdw 5 dmdw 0 Another possibility is that breaks respond to wageslike nonwork leisure does Dening nonwork leisure L 5 242 m if dbdw 5 dLdw then dbdw 5 dmdw so that dtdw 5 2(dmdw) 0 The opposite possibil-ity is that breaks and nonwork leisure are perfect substitutes (drivers do not carewhether they take breaks on the job or after work at home) and unresponsive towage so dbdw 5 2 dLdw Then dbdw 5 2 dmdw so dtdw 5 0 These threesimple assumptions show that for values of dbdw in the interval [ 2 |dLdw||dLdw|] dtdw remains negative or zero For dtdw to be positive re-quires that breaks respond more strongly to wages than leisure L does and in theopposite direction drivers must really like taking breaks on slow (low-wage) daysand dislike them on busy days though they exhibit the opposite pattern of leisurepreferences This is conceivable (and could be tested with better data) but nomore plausible than the other three assumptions which produce nonpositivedt dw


the length of breaks they take responds less strongly to wagesthan for inexperienced drivers which is possible but not sup-ported by any particular intuition or evidence

Increasing Disutility of Effort

Since fares are xed within each sample a high-wage day isa busy day in which a driver picked up many fares or drove themfurther Drivers may get tired faster on these high-wage days andquit earlier due to fatigue

The hypothesis implicit in this explanation is not just thatdriving with a passenger in your cab is hard work but that car-rying a passenger is harder work than searching for one Almostall of the managers in our survey said the opposite The eetmanagers were asked to compare two hypothetical drivers (A) ldquoadriver who worked 10 hours found fares very quickly drove 30trips and spent little time cruising looking for passengersrdquo and(B) ldquoa driver who worked 10 hours drove 20 trips and had a hardtime nding fares so he spent a lot of time cruising looking forpassengersrdquo Ten eet managers said the 20-trip driver would beldquomore tired at the end of the dayrdquo Only one said the busy 30-tripdriver would be more tired (Two managers said the two driverswould be equally tired or did not know) This makes sense giventhe logistics of searching for passengers in Manhattan Drivingto a specic destination probably requires less attention thandriving while searching for a potential passenger who is trying tohail a cab on either side of the street and preparing to swerveacross trafc to reach the passenger

The earning-money-is-tiring hypothesis also does not easilyexplain the effect of experience unless one assumes that inexpe-rienced drivers get relatively more tired carrying passengers andexperienced drivers get relatively more tired searching for pas-sengers The opposite effect could easily be true if experienceddrivers learn the easiest places to nd fares then searching forpassengers becomes relatively less tiring for them

Participation

The hours equation is estimated using only days on whichcabdrivers worked positive hours If unobserved factors affecteddriversrsquo decisions about whether to work at all (or ldquoparticipaterdquo)and those factors also affected their hours decisions the wageelasticity will be biased [Heckman 1979] The sign of the bias willbe opposite of the sign of the correlation between the error terms


in the hours and participation equations If unobserved shocks toparticipation and hours are positively correlated for example thewage elasticity will be downward-biased One way to control forthis ldquoselection biasrdquo is to collect data on participation Unfortu-nately we do not have these data However there are severalreasons to think that selection bias is not severe enough to ex-plain the substantial negative elasticities First including driverxed effects in the hours equation mitigates the omitted variableproblem that leads to the correlation between the hours and par-ticipation error terms Second in the survey eleven of the four-teen eet managers said drivers usually have a regular scheduleof shifts each week When drivers skip days about half the eetmanagers said those drivers had to pay their fees anyway or suf-fered some penalty so they have a large incentive to stick to theirschedule Also a driver cannot always participate on an unsched-uled day even if he decides to Cabs are not always available be-cause medallion owners tightly schedule them to maximize thelease fees they collect While owner-drivers are not strictly sub-ject to a regular schedule most of them rent their cabs to anotherdriver or two the remaining days effectively constitute a regularschedule for themselves Because of the regularity in the driversrsquoschedules there is not that much variation in unobserved factorsthat affect participation and there should be little selection bias

IV DISCUSSION AND CONCLUSIONS

Dynamic theories of labor supply predict a positive laborsupply response to transitory uctuations in wages Previousstudies have not been able to measure this elasticity preciselyand the measured sign is often negative contradicting the theo-retical prediction These analyses however have been plaguedby a wide variety of estimation problems

Many of these estimation problems are avoided by estimat-ing labor supply functions for taxi drivers Drivers have exibleself-determined work hours and face wages that are highly corre-lated within days but only weakly correlated between days (souctuations are transitory) The fact that our analyses yieldnegative wage elasticities suggests that elasticities of intertem-poral substitution around zero (or at least not strongly positive)may represent a real behavioral regularity Further support forthis assertion comes from analyses of labor supply of farmers[Berg 1961 Orde-Brown 1946] and self-employed proprietors


[Wales 1973] who like cabdrivers set their own hours and oftenhave negative measured wage elasticities These data suggestthat it may be worthwhile to search for negative wage elasticitiesin other jobs in which workers pay a xed fee to work earn vari-able wages and set their own work hoursmdashsuch as shing somekinds of sales and panhandling

Of course cabdrivers farmers and small-business proprie-tors are not representative of the working population Besidessome demographic differences all three groups have self-selectedonto occupations with low variable wages long hours and (in thecase of farmers and cabdrivers) relatively high rates of accidentsand fatalities However there is no reason to think their planninghorizons are uniquely short Indeed many cabdrivers are recentimmigrants who by immigrating are effectively making long-term investments in economic and educational opportunity forthemselves and their children

Because evidence of negative labor supply responses to tran-sitory wage changes is so much at odds with conventional eco-nomic wisdom these results should be treated with cautionFurther analyses need to be conducted with other data sets (asin Mulligan [1995]) before reaching the conclusion that negativewage elasticities are more than an artifact of measurement orthe special circumstances of cabdrivers If replicated in furtheranalyses however evidence of negative wage elasticities callsinto question the validity of the life-cycle approach to laborsupply

APPENDIX 1 DESCRIPTION OF DATA SETS

Trip Sheet DataData Set 1 TRIP

We collected 192 trips sheets from a eet company in NewYork City that rents cabs daily to drivers This sample consists of27 cabdrivers who drove during the days April 24 1994 to May14 1994 A trip sheet is a sequential list of trips that a drivertook on a given day For each trip the driver lists the time thefare was picked up and dropped off and the amount of the fare(excluding tip) The company uses these trip sheets for insurancepurposes (they are not used for taxes) When a driver returns thecab the trip sheet is stamped with the number of trips that havebeen recorded by the meter in the cab


Our measure of hours worked is obtained directly from thetrip sheet It is the difference between the time that the rst pas-senger is picked up and the time that the last passenger isdropped off We calculated total revenue by adding up the fareslisted on the trip sheet Average hourly wage is total revenue di-vided by hours worked

Not all trip sheets we obtained were complete because thenumber of trips listed by the cabdriver is sometimes much lessthan the number of trips recorded by the meter Because we arecalculating hours and total daily revenue from the trip sheet weneed to screen out incomplete trip sheets (because hours and reve-nues will be too low) We decided to use only trip sheets wherethe number of trips listed by the driver is within two of the num-ber of trips stamped by the meter Using this to screen the tripsheets we are left with 70 trip sheets and 13 drivers Eight ofthese drivers appear more than once in the screened data

There are important differences between the data retainedin the screened sample and those that are not used in the anal-ysis (the screened-out data) In Appendix 3 we provide summarystatistics for the key variables for both samples As expected thenumber of trips listed by the driver in the screened sample ismuch greater than in the screened-out sample This causes hoursworked in the screened sample to be greater than in the screened-out sample However the average wage (for the trips listed) doesnot differ between the two samples This is some evidence thatwhether or not the cabdriver lls out the trip sheet completely isnot related to how ldquobusyrdquo the cabdriver is In Appendix 4 we alsopresent the basic regressions from the paper when the entiresample of trip sheets is used rather than only the screenedsample

For our screened sample drivers either worked the afternoonor evening shift We dened the afternoon shift to be those driv-ers who picked up their cabs before 430 PM and the evening shiftas those picking up their cabs after 430 PM (all drivers in oursample picked up their cabs between 100 PM and 725 PM) Ap-proximately 30 percent of the trip sheets are for the afternoonshift The results are not sensitive to whether a shift is dened

For the 70 shifts the average trip duration was 95 minutesand the average time searching for the next fare was also 95minutes The average fare per trip was $513 The percent of timethat a driver spent with a passenger in the cab was 517

There is no direct information on the experience of the driv-


ers In the analysis we use hack numbers which are issued inconsecutive order by the TLC We use a median split on the hacknumber to separate the drivers into the ldquolowrdquo or ldquohighrdquo experi-ence group

To try to control for factors other than wages that might af-fect hours we collected measures of the weather on the days inour sample from The New York Times It rained during approxi-mately one-third of the shifts and the high temperature aver-aged almost 76 degrees Seventy-three percent of the shifts wereduring the week

TLC Data

We use two data sets of trip sheets collected by the New YorkCity Taxi and Limousine Commission (TLC1 and TLC2) A tripsheet is a sequential list of trips that a driver took on a givenshift For each trip the driver lists the time the fare was pickedup and dropped off and the amount of the fare On each tripsheet the driver also stamps the following output from the meternumber of trips the meter logged at the start and end of the driv-errsquos shift (the difference is number of trips taken by the driver)number of miles at the start and end of the shift number of milesldquoliverdquo (with a passenger) and total revenue this shift (excludingtips)

Our measure of hours worked is obtained directly from thetrip sheet It is the difference between the time that the rst pas-senger is picked up and the time that the last passenger isdropped off Our measure of total revenue is obtained directlyfrom the meter (we do not have revenues from the trip sheet) Wecalculate the average hourly wage by dividing total revenue fromthe meter by the number of hours worked from the trip sheet

Not all trip sheets were lled out completely because thenumber of trips listed by the cabdriver is sometimes much lessthan the number of trips recorded by the meter Because the TLCcalculates hours from the trip sheet a screen is needed to elimi-nate these incomplete trip sheets Also the TLC has indicatedthat the meters malfunction occasionally recording negativenumbers of trips or negative revenues The TLC developed ascreen to discard trip sheets To pass this screen the number oftrips on the meter must exactly match the number of trips listedby the driver and the percent of ldquolive milesrdquo (percent of milesdriven when driver has a passenger) is between 20 and 91

We were not given the trip sheets themselves but only the


summary measures such as hours driven revenue collectedshift the number of trips and whether the trip sheet passed thescreen Because we do not have the trip sheets themselves wecannot devise an alternative screening procedure Therefore weuse the TLC screen for the analysis in the paper However inAppendix 3 we present sample statistics for the screened and thescreened-out sample and present the basic regression results inAppendix 4 for the unscreened (full) sample

Data Set 2 TLC1

This data set consists of 1723 completed trip sheets collectedfor a study by the New York City Taxi and Limousine Commission(TLC) The shifts occurred mostly during the time period October29 1990 to November 5 1990 The screen developed by the TLCeliminates 658 trips sheets and we eliminate 21 additional tripsheets due to missing hack numbers (we need hack numbers tocorrect the standard errors in the nonxed-effects model and toestimate the xed-effects models) Summary statistics of thescreened-out sample are given in Appendix 3 (note that the num-ber of observations is 646 rather than 658 because 12 observa-tions are omitted because of missing data on hack number ordate driven)

In the screened sample here are 1044 trip sheets logged by484 drivers Of the 1044 trip sheets 34 percent are from eetcompanies 355 percent are leases and 305 percent are fromowner-drivers The NYC TLC estimates that of all shifts driven in1990 22 percent are from eet companies 30 percent are owner-drivers and 40 percent are leases (8 percent are other) There-fore this sample overrepresents eet company shifts

The TLC provided measures of experience for the drivers inthis sample Approximately 45 percent of the shifts in the samplehave drivers with less than four years of experience

We obtained from the TLC variables that indicated whattime the driver began driving and what shift they had designatedfor that driver (ldquodayrdquo ldquonightrdquo or ldquootherrdquo) However we realizedthat the TLCrsquos designations were not consistent across driversFor example there might be many drivers that began driving atAM most of which were labeled as the ldquodayrdquo shift However somedrivers that also began driving at AM were labeled ldquootherrdquo shiftWe decided to make the shift designations consistent so that all


drivers beginning at the same time were labeled as the sameshift (To do this we used the majority designation and assignedit to everyone) Drivers beginning their day between AM and AMare ldquodayrdquo shift (336 percent of trip sheets) between 3PM and10PM are ldquonightrdquo shift (287 percent of trip sheets) ldquootherrdquo is aresidual shift category (377 percent trip sheets) All the analysesin the paper are run using these shift designations However theresults are not sensitive to our particular denition of shift Theresults are qualitatively unchanged if we use the original TLCdenitions (the only difference is in the IV xed-effects modelsfor TLC1mdashthere is no experience effect) Furthermore if no shiftdesignation is used at all the results are identical to those pre-sented in the paper

We obtained measures of the weather from The New YorkTimes for the days in this sample It did not rain on any of thedays in the sample The high temperature averaged about 65 de-grees Thirty-four percent of the shifts were on the weekend

Data Set 3 TLC2

This data set consists of 750 trip sheets taken from mini-eetand owner-drivers Mini-eets are smaller operations than eetsand usually lease cabs to drivers weekly or monthly We cannotidentify which trips sheets come from mini-eets and which areowner-drivers There is only one observation per driver mostlyfrom November 1 2 or 3 1988 The screen used by the TLCeliminates 38 trip sheets (screen is described above) leaving 712trip sheets for our analysis Summary statistics for the screenedand screened-out sample are provided in Appendix 3 The regres-sion results do not change at all (mostly because so few tripsheets are eliminated with the screen)

We obtained shift and experience measures from the TLCAlthough we do not have the time drivers began their shifts theTLC designated the shift either ldquodayrdquo or ldquonightrdquo Fifteen percentof the screened sample are night shift and 85 percent are dayshift All shifts in this sample are during the week Forty-six per-cent of the shifts are with drivers with fewer than three yearsof experience

We obtained measures of the weather from The New YorkTimes It rained on approximately 5 percent of the shifts and thehigh temperature averaged just over 49 degrees


APPENDIX 2 DISTRIBUTION OF HOURS AND WAGES SCREENED SAMPLE

TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712

HoursMinimum 627 078 1005 660 442 42510 706 618 57525 829 809 78850 938 967 92575 1010 1108 110090 1091 1250 125095 1114 1377 1400Maximum 1141 2343 2225

Average hourly wageMinimum 1120 328 2175 1275 1088 96110 1332 1244 113825 1497 1432 129950 1620 1631 147175 1845 1836 164590 2192 2105 183995 2295 2363 1948Maximum 2543 5056 3560

APPENDIX 3 COMPARISON OF SCREENED DATA WITH SCREENED-OUT DATA

Screened Screened outTRIP Mean Median Std dev Mean Median Std dev

Hours worked 916 938 139 694 724 290Average wage 1691 1620 321 1741 1710 467Total revenue 15270 15400 2499 11400 12388 4769 Trips listed on 3017 3000 548 2281 2300 959sheet Trips counted 3070 3000 572 3419 3600 689by meterSample size 70 122Correlation (log 2 502 2 431hours log wages)

Screened Screened outTLC1 Mean Median Std dev Mean Median Std dev

Hours worked 962 967 288 991 966 344Average wage 1664 1631 436 1788 1664 836Total revenue 15458 15400 4583 16213 16100 5839 Trips counted 2788 2900 915 3084 3100 1345by meterSample size 1044 646


Correlation (log 2 391 2 487hours log wages)


Hours worked 938 925 296 1003 1013 140Average wage 1470 1471 320 1019 1036 226Total revenue 13338 13723 4074 10062 10417 2292 Trips counted 2862 2900 941 1976 1950 717by meterSample size 712 38Correlation (log 2 269 2 234hours log wages)

APPENDIX 4 LOG HOURS WORKED EQUATIONS USING FULL SAMPLE

Sample TRIP TLC1

OLS resultsLog hourly wage 2 1402 157 2 410 2 468

(753) (113) (053) (028)Fixed effects No Yes No YesAdjusted R2 198 882 197 232Sample size 192 183 1690 1316IV resultsLog hourly wage 2 609 190 2 1164 2 1305

(439) (244) (387) (273)Fixed effects No Yes No YesSample size 192 183 1690 1316IV by experience results Low High Low HighLog hourly wage 127 281 2 373 2 1194

(406) (242) (319) (412)Fixed effects Yes Yes Yes YesSample size 91 92 564 732

Dependent variable is the log of hours worked Standard errors are in parentheses and are corrected(for the nonxed-effects models) to account for the panel structure of the data All regressions also includeweather and shift characteristics (dummy variable for rain high temperature during the day dummy vari-able for shift on a weekday and time of shift dummy variables) as explanatory variables Instruments forthe log hourly wage include the summary statistics of the distribution of hourly (log) wages of other driverson the same day and shift (the 25th 50th and 75th percentiles)

CALIFORNIA INSTITUTE OF TECHNOLOGY

CARNEGIE MELLON UNIVERSITY


UNIVERSITY OF CHICAGO

APPENDIX 3 CONTINUED



REFERENCES

Altonji Joseph G ldquoIntertemporal Substitution in Labor Supply Evidence fromMicro Datardquo Journal of Political Economy XCIV (1986) s176ndashs215

Benartzi Shlomo and Richard Thaler ldquoMyopic Loss Aversion and the Equity Pre-mium Puzzlerdquo Quarterly Journal of Economics CX (1995) 73ndash92

Berg Elliot J ldquoBackward-Sloping Labor Supply Functions in Dual EconomiesmdashThe Africa Caserdquo Quarterly Journal of Economics LXXV (1961) 468ndash92

Bowman David Debby Minehart and Matthew Rabin ldquoLoss Aversion in a Sav-ings Modelrdquo University of California at Berkeley working paper 1996

Browning Martin Angus Deaton and Margaret Irish ldquoA Protable Approach toLabor Supply and Commodity Demands over the Life-Cyclerdquo EconometricaLIII (1985) 503ndash43

Duesenberry J Income Saving and the Theory of Consumer Behavior (Cam-bridge MA Harvard University Press 1949)

Gneezy Uri and Jan Potters ldquoAn Experiment on Risk Taking and EvaluationPeriodsrdquo Quarterly Journal of Economics CXII (1997) 631ndash645

Hardie Bruce G S Eric J Johnson and Peter S Fader ldquoModeling Loss Aversionand Reference-Dependence Effects on Brand Choicerdquo Marketing Science XII(1993) 378ndash94

Heckman James ldquoSample Selection Bias as a Specication Errorrdquo EconometricaXLVII (1979) 153ndash61

Helson Harry Adaptation-Level Theory (New York NY Harper and Row 1964)Johnson Eric J Colin F Camerer Talia Rymon and Sankar Sen ldquoLimited Com-

putation and Fairness in Sequential Bargaining Experimentsrdquo University ofPennsylvania Department of Marketing Working Paper 1996

Kahneman Daniel Jack Knetsch and Richard Thaler ldquoExperimental Tests of theEndowment Effect and the Coase Theoremrdquo Journal of Political EconomyXCVIII (1990) 1325ndash48

Kahneman Daniel and Amos Tversky ldquoProspect Theory An Analysis of Decisionunder Riskrdquo Econometrica XLVII (1979) 263ndash91

Laisney Francois Winfried Pohlmeier and Matthias Staat ldquoEstimation of LaborSupply Functions Using Panel Data A Surveyrdquo in Matyas and Sevestre edsThe Economics of Panel Data Handbook of Theory and Applications (Dor-drecht The Netherlands Kluwer 1992)

Lucas Robert E Jr and Leonard A Rapping ldquoReal Wages Employment andInationrdquo Journal of Political Economy LXXVII (1969) 721ndash54

MaCurdy Thomas E ldquoAn Empirical Model of Labor Supply in a Life-Cycle Set-tingrdquo Journal of Political Economy LXXXIX (1981) 1059ndash85

Maddala G S Introduction to Econometrics 2nd edition (New York NY Macmil-lan Publishing Company 1992)

Mankiw N Gregory Julio J Rotemberg and Lawrence H Summers ldquoIntertem-poral Substitution in Macroeconomicsrdquo Quarterly Journal of Economics C(1985) 225ndash51

McGlothlin William H ldquoStability of Choices among Uncertain AlternativesrdquoAmerican Journal of Psychology LXIX (1956) 604ndash15

Mulligan Casey lsquoThe Intertemporal Substitution of WorkmdashWhat Does the Evi-dence Sayrsquo University of Chicago Population Research Center working paper95-11 June 1995

NYC Taxi and Limousine Commission ldquoTaxi Trip and Fare Data A Compen-diumrdquo October 29 1991 NYC Taxi and Limousine Commission ldquoThe NewYork City Taxicab Fact Bookrdquo May 1992

Odean Terry lsquoAre Investors Reluctant to Realize Their Lossesrsquo University ofCalifornia-Berkeley Working Paper 1996

Orde-Brown G Labour Conditions in East Africa (London Colonial OfceHMSO 1946)

Pencavel John ldquoLabor Supply of Men A Surveyrdquo in O Ashenfelter and RLayard eds Handbook of Labor Economics Volume I (Amsterdam TheNetherlands North-Holland 1986) pp 3ndash102

Pindyck Robert S and Daniel L Rubinfeld Microeconomics (New York Macmil-lan 1989)

Read D and G Loewenstein ldquoThe Diversication Bias Explaining the Differ-


ence between Prospective and Real-Time Taste for Varietyrdquo Journal of Ex-perimental Psychology Applied I (1995) 34ndash49

Read D and G Loewenstein ldquoTemporal Bracketing of Choice Discrepancies be-tween Simultaneous and Sequential Choicerdquo Carnegie Mellon UniversityDepartment of Social and Decision Sciences Working Paper 1996

Samuelson William and Richard Zeckhauser ldquoStatus Quo Bias in Decision Mak-ingrdquo Journal of Risk and Uncertainty I (1988) 39ndash60

Shea John ldquoUnion Contracts and the Life-CyclePermanent-Income HypothesisrdquoAmerican Economic Review LXXXV (1995) 186ndash200

Shefrin Hersh M and Richard H Thaler ldquoMental Accounting Saving and Self-Controlrdquo in G Loewenstein and J Elster eds Choice Over Time (New YorkRussell Sage Foundation Press 1992)

Thaler Richard ldquoMental Accounting and Consumer Choicerdquo Marketing ScienceIV (1985) 199ndash214

Thaler Richard Amos Tversky Daniel Kahneman and Alan Schwartz ldquoHow My-opic Loss-Averse Investors Learn from Experiencerdquo Quarterly Journal of Eco-nomics CXII (1997) 647ndash661

Tversky Amos and Daniel Kahneman ldquoLoss Aversion in Riskless Choice AReference-Dependent Modelrdquo Quarterly Journal of Economics CVI (1991)1039ndash61

Wales Terence J ldquoEstimation of a Labor Supply Curve for Self-Employed Busi-ness Proprietorsrdquo International Economic Review XIV (1973) 69ndash80

Weber Martin and Colin F Camerer ldquoThe Disposition Effect in Securities Trad-ing An Experimental Analysisrdquo Journal of Economic Behavior and Organi-zation forthcoming

Weber Max The Protestant Ethic and the Spirit of Capitalism (New York NYCharles Scribner amp Sons 1958)


Staat [1992] Pencavel [1986] and cf Mulligan [1995]) Howeverthese results are difcult to interpret because actual wagechanges are rarely transitory so the hypothesis of intertemporalsubstitution must be tested jointly along with auxiliary assump-tions about persistence of wage shocks formation of wage expec-tations etc As a result the frequently observed negative wageelasticities can plausibly be attributed to specication error

The ideal test of labor supply responses to transitory wageincreases would use a context in which wages are relativelyconstant within a day but uncorrelated across days In such asituation all dynamic optimization models predict a positive rela-tionship between wages and hours worked due to the negligibleimpact on life-cycle wealth of a one-day increase in wage (egMaCurdy [1981 p 1074])

Such data are available for at least one group of workersNew York City cabdrivers Drivers face wages that uctuate ona daily basis due to demand shocks caused by weather subwaybreakdowns day-of-the-week effects holidays conventions etcAlthough rates per mile are set by law on busy days driversspend less time searching for customers and thus earn a higherhourly wage These wages tend to be correlated within days anduncorrelated across days (ie transitory)

Another advantage of studying cabdrivers is that unlikemost workers they choose the number of hours they work eachday because drivers rent their cabs from a eet for a xed fee (orown them) and can drive as long as they like during a continuoustwelve-hour shift Furthermore most analyses of labor supplymeasure hours (and sometimes income) by self-reports For cab-drivers better measures of hours and income are available fromldquotrip sheetsrdquo the drivers ll out and from meters installed in cabswhich automatically record the fares

Using these data we investigate the relationship betweenwages and hours worked and nd little evidence for intertem-poral substitution Most of the elasticities we estimate are nega-tive drivers tend to quit early on high wage days and to drivelonger hours on low wage days Elasticities for inexperienceddrivers are around 2 1 for each of the three samples of cabdriverswe used in our study The results are robust to outliers and differ-ent specications There are several possible explanations forthese negative elasticities Some explanations can be ruled outbut others require more data to evaluate

Assuming that the alternative explanations for negative








































Mean Median Std dev
















Wage Elasticities












Log hourly wage 2 411 2 186 2 501 2 618 2 355(169) (129) (063) (051) (051)



Rain 002 015 mdash mdash 2 150(035) (035) (062)


























Log hourly wage 2 319 005 2 1313 2 926 2 975(298) (273) (236) (259) (478)



Rain 2 007 2 001 mdash mdash 2 130(042) (041) (070)






Median 316 026 2 385 2 276 1292(225) (188) (394) (467) (4281)

25th percentile 323 287 693 469 2 373(160) (126) (241) (332) (3516)

75th percentile 399 289 614 688 479(171) (149) (242) (292) (1699)
























































(TLC1 only)





Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES














































































Mean Median Std dev
















Wage Elasticities












Log hourly wage 2 411 2 186 2 501 2 618 2 355(169) (129) (063) (051) (051)



Rain 002 015 mdash mdash 2 150(035) (035) (062)


























Log hourly wage 2 319 005 2 1313 2 926 2 975(298) (273) (236) (259) (478)



Rain 2 007 2 001 mdash mdash 2 130(042) (041) (070)






Median 316 026 2 385 2 276 1292(225) (188) (394) (467) (4281)

25th percentile 323 287 693 469 2 373(160) (126) (241) (332) (3516)

75th percentile 399 289 614 688 479(171) (149) (242) (292) (1699)
























































(TLC1 only)





Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES



















































Wage Elasticities












Log hourly wage 2 411 2 186 2 501 2 618 2 355(169) (129) (063) (051) (051)



Rain 002 015 mdash mdash 2 150(035) (035) (062)


























Log hourly wage 2 319 005 2 1313 2 926 2 975(298) (273) (236) (259) (478)



Rain 2 007 2 001 mdash mdash 2 130(042) (041) (070)






Median 316 026 2 385 2 276 1292(225) (188) (394) (467) (4281)

25th percentile 323 287 693 469 2 373(160) (126) (241) (332) (3516)

75th percentile 399 289 614 688 479(171) (149) (242) (292) (1699)
























































(TLC1 only)





Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES














































Log hourly wage 2 411 2 186 2 501 2 618 2 355(169) (129) (063) (051) (051)



Rain 002 015 mdash mdash 2 150(035) (035) (062)


























Log hourly wage 2 319 005 2 1313 2 926 2 975(298) (273) (236) (259) (478)



Rain 2 007 2 001 mdash mdash 2 130(042) (041) (070)






Median 316 026 2 385 2 276 1292(225) (188) (394) (467) (4281)

25th percentile 323 287 693 469 2 373(160) (126) (241) (332) (3516)

75th percentile 399 289 614 688 479(171) (149) (242) (292) (1699)
























































(TLC1 only)





Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES

























































Log hourly wage 2 319 005 2 1313 2 926 2 975(298) (273) (236) (259) (478)



Rain 2 007 2 001 mdash mdash 2 130(042) (041) (070)






Median 316 026 2 385 2 276 1292(225) (188) (394) (467) (4281)

25th percentile 323 287 693 469 2 373(160) (126) (241) (332) (3516)

75th percentile 399 289 614 688 479(171) (149) (242) (292) (1699)
























































(TLC1 only)





Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES




























































































(TLC1 only)





Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES









































Breaks










Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES













































Participation
























TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES





























































TLC Data







Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES









































Data Set 2 TLC1








Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES










































Data Set 3 TLC2






TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES









































TRIP TLC1 TLC2n 5 70 n 5 1044 n 5 712













Sample TRIP TLC1













REFERENCES












































Sample TRIP TLC1













REFERENCES








































REFERENCES








































labor supply of new york city cabdrivers: one day at a time ·...

Documents