American Economic Review 2008, 98:3, 1069–1082

http://www.aeaweb.org/articles/php?doi=10.1257/aer.98.3.1069

Reference-Dependent Preferences and Labor Supply:

The Case of New York City Taxi Drivers

By Henry S. Farber*

Increased attention has been paid in recent years to deviations from the standard neoclassical
model of consumer behavior. A substantial segment of this work focuses on reference-dependent
preferences where there is a change in the shape of the utility function at some base (reference)
level of income or consumption. These models have strong predictions for how responses to
changes in prices are affected by the actual level of consumption or income relative to the refer-
ence level. A difficulty in bringing this class of models to the data is that the reference level of
income or consumption is seldom observed, and we have only weak information on what deter-
mines the reference point.
There are important reasons to understand how these considerations affect estimation of labor
supply elasticities. Evaluation of much government policy regarding tax and transfer programs
depends on having reliable estimates of the sensitivity of labor supply to wage rates and income
levels. To the extent that individuals’ levels of labor supply are the result of optimization with
reference-dependent preferences, the usual estimates of wage and income elasticities are likely
to be misleading.
In this study, I develop an empirical model of daily labor supply that incorporates reference-
dependent preferences but does not require that the reference level of income be observed or
defined in advance. I apply this model to data on the daily labor supply of New York City taxi
drivers by allowing taxi drivers to have a reference level of daily income. The estimates suggest
that, while there may be a reference level of income on a given day such that there is a discrete
increase in the probability of stopping when that income level is reached, the reference level
varies substantially day to day for a particular driver. Additionally, most shifts end before the
reference income level is reached. Essentially, the data show more smoothness in the relationship
between income and the continuation and stopping probabilities than seems consistent with an
important role for reference-dependent preferences.

I. Reference-Dependent Preferences

The theory of reference-dependent preferences has its roots in the concept of loss aversion.
Amos Tversky and Daniel Kahneman 119912 present an analysis of choice in a riskless frame-
work where the concavity of utility as a function of income changes at some reference or status
quo value of income 1YR2. Marginal utility is higher and increasing in income, at incomes below
YR. Marginal utility is lower and decreasing in income, at incomes above YR. This preference
structure implies that, for equivalent absolute dollar gains and losses around YR, individuals give
up more utility from the dollar loss than they receive from the dollar gain.
This core idea has been credited as the explanation for the “endowment effect” (Richard Thaler
1980). The endowment effect, which has much experimental support, states that ­ individuals

*  Industrial Relations Section, Firestone Library, Princeton University, Princeton, NJ 08544-2098 (e-mail: farber@
princeton.edu). I thank Orley Ashenfelter, David Card, Avinash Dixit, Robert Solow, and two anonymous referees for
helpful comments and discussion. Participants in numerous workshops also contributed useful ideas and comments.
Susan Merino, Gregory Evans, Julia Stahl, and Hannah Pierce provided able research assistance.
1069
1070 THE AMERICAN ECONOMIC REVIEW june 2008

appear to value items they own more than identical items owned by others. For example, an eas-
ily replicable experiment is to distribute commemorative mugs to a random subgroup of some
experimental subject pool, to elicit from those who received the mugs a willingness to accept
(WTA) a given amount in exchange for the mug, and to elicit from those who did not receive
the mugs a willingness to pay (WTP) a given amount to purchase a mug. It is a routine finding
that the average WTA is substantially greater than the average WTP. Indeed, this relationship
between WTA and WTP is considered a test of reference-dependent preferences (Ian Bateman
et al. 1997).
While there has been continued refinement of the theoretical ideas underlying reference-depen-
dent preferences (e.g., Alistair Munro and Robert Sugden 2003; Botond Kőszegi and Matthew
Rabin 2006), there is relatively little direct field evidence on reference-dependent preferences.
One example is David Genesove and Christopher Mayer (2001), who examine seller behavior in
the housing market using the purchase price as the reference point. They find clear evidence that
sellers are more risk averse when faced with losses. This is consistent with a utility function that
changes concavity at the reference point.
There is an emerging literature on labor supply in settings where workers are free to set their
labor supply (e.g., Colin Camerer 1997; Gerald S. Oettinger 1999; Ernst Fehr and Lorenz Goette
2007; and Farber 2005). None of this literature allows directly for the possibility of reference-
dependent preferences and, by and large, this literature finds substantial positive labor supply
elasticities consistent with the standard neoclassical labor supply model.
A notable exception directly relevant to my study is the finding of a negative labor supply elas-
ticity among New York City taxi drivers by Camerer et al. (1997). They interpret their finding as
evidence of target earnings behavior by taxi drivers. A target earner is defined as an individual
who works until earnings reach the target level, and then quits. Target earnings behavior is an
extreme version of reference-dependent preferences, where, at income levels less than the tar-
get/reference level, the marginal utility of income is extremely high and, at income levels greater
than the target/reference level, the marginal utility of income is extremely low. Essentially, there
is such a sharp kink in the utility function at the reference/target level that workers always quit
when income reaches the reference/target level. This is clearly nonoptimal from a neoclassical
perspective, since it implies quitting early on days when it is easy to make money and working
longer on days when it is harder to make money. Utility would be higher by allocating time in
precisely the opposite manner. In an earlier paper (Farber 2005), I critiqued the conceptual and
econometric approach used by Camerer et al., and I analyzed new data (also used in this study)
on the labor supply of New York City taxi drivers using a completely different econometric
framework. In that study, I found no evidence of target earnings behavior.
One can also interpret work that considers the effect of relative income or consumption com-
parisons across individuals on behavior as implicitly using reference-dependent preferences. The
idea is that the interpersonal comparisons set up a reference point that introduces an asymme-
try between gains and losses relative to this reference point. For example, David Neumark and
Andrew Postlewaite (1998) use relative income concerns to understand the increase in labor sup-
ply of married women. As more married women enter the labor force and raise their families’
incomes, the reference level of income that families use to judge their well-being increases. The
result is that more women enter the labor force so that they do not suffer the substantial loss in
utility that comes from family income below the reference level.


See Kahneman, Jack L. Knetsch, and Thaler (1991) for a nice discussion of the endowment effect and its relation-
ship to reference-dependent preferences.

I survey this literature briefly in Farber (2005).
VOL. 98 NO. 3 farber: reference-dependent preferences and labor supply 1071

II. Modeling Labor Supply with Reference-Dependent Preferences

I implement reference-dependent preferences by allowing the marginal utility of income to be

higher at income levels below the reference level than at income levels above the reference level.
In this case, the utility function is continuous everywhere, but marginal utility is discontinuous
at the reference level. This is a natural formulation, given the roots of the reference point concept
in the literature on loss aversion.
I show in my earlier work (Farber 2005) that it is not appropriate to characterize the labor sup-
ply of a taxi driver as a response to a parametric wage, regardless of the specific utility function.
An alternative approach I use is to consider the end of each passenger trip as a decision point
regarding whether to continue driving or to stop. In the remainder of this section, I derive the
implications of a simple model of reference-dependent preferences for a driver’s labor supply
decision, assuming that future earnings are stochastic.
One parametrization of a driver’s utility (as a function of accumulated income and hours) after
trip t during a shift that incorporates a reference point is
c
(1)  U 1Yt, ht 2 5 11 1 aI  3Yt , T]2 1Y 2 T2 2 h11n,
11n t

where I 3·4 is the indicator function and

• ht 5 hours worked on shift at end of trip t,

• Yt 5 income earned on shift at end of trip t,
• T 5 reference income,
• a 5 parameter determining sharpness of “kink” in utility function (a . 0),
• c 5 parameter determining disutility of work, and
• n 5 elasticity parameter (wage elasticity of labor supply 5 1/n, defining Y 5 Wh ).

This utility function is continuous, but kinked at income level T, with the marginal utility of
income discontinuously higher below the kink (a . 0 ) than above the kink.
Next, I express the utility on trip t 1 1 as a function of accumulated income and hours as of
the end of trip t and the fare and time associated with trip t 1 1. Denote the next fare by ft11 and
the time to search for and complete the next fare by tt11. On this basis, the utility after trip t 1 1
can be written as
c
(2)  U 1Yt11, ht112 5 1Yt 1 ft11 2 T2 11 1 aI  3Yt 1 ft11 , T4 2 2 1h 1 tt112 11n.
11n t

The marginal utility derived from the next trip t depends on earnings associated with the next
fare and the time necessary to complete this trip. For analytic convenience in working with this
marginal utility, I approximate the second (hours) term of the utility function using a first-order
Taylor series expansion around accumulated hours at the choice point (ht ). I further simplify the
expression for marginal utility by assuming that marginal utility on the next trip is at the higher
“pre-reference” rate when accumulated income prior to that trip is anywhere below the reference


Given the relatively short length of trips, the first-order approximation is likely to be quite good.
1072 THE AMERICAN ECONOMIC REVIEW june 2008

income level (even if the next fare causes accumulated income to exceed the reference income
level). The resulting expression for marginal utility is

(3)  MUt11 5 ft11 11 1 aI  3Yt , T4 2 2 tt11chnt.

Dividing both sides of equation (3) by the time required for the t 1 1 trip (tt11) yields

(4)  MVt11 5 vt11 11 1 aI  3Yt , T4 2 2 chnt,

where MVt11 denotes the marginal utility per unit time on the t 1 1 trip, which I call “marginal
value.” The quantity vt11 5 ft11 /tt11 is the wage per unit time on the t 1 1 trip.
The marginal value of continuing to drive is a linear function of the uncertain wage on the
t 1 1 trip. Let E 1vt 2 5 mvt so that expected marginal value is

(5)  E 3MVt114 5 mvt11 11 1 aI  3Yt , T4 2 2 chnt.

Conditional on the expected wage, the expected marginal value of continuing to drive declines
monotonically as the shift progresses and income and hours worked accumulate. There is a
smooth increase in the marginal disutility of hours, as hours accumulate at all hours and income
levels. In contrast, the marginal utility of income is constant, aside from the discontinuous drop
when the reference income level is reached. If the expected wage is nonincreasing during a
shift, then expected marginal value is monotonically declining in t. In this case, the driver would
continue to drive until the expected marginal value of the next trip is negative. On this basis, the
function relating the continuation probability to accumulated income will exhibit an incremental
decline at the reference income level. There should be no other relationship between the continu-
ation probability and accumulated income conditional on a set of observables indicating hours
worked and factors affecting expected fares and times (e.g., local, time of day, day of weak,
weather).
If, in fact, the expected wage were nonincreasing, I could implement a structural stopping
model based on the expected marginal value in equation (5). However, it is unreasonable to
assume that the expected wage is nonincreasing. For example, there may be lulls in the mid-
dle of a shift where waiting times are rather high. Drivers may not quit for the day at these
points because they expect fares to be more plentiful later in the shift. Given that the decision to
stop driving for the day is irreversible, consideration of the option value of continuing to drive
becomes important. A complete solution to the optimal stopping problem in this context requires
the solution of a stochastic dynamic optimization problem that incorporates this option value.
Rather than provide this solution (which is not usefully amenable to empirical implementation
with the available data), I implement a reduced-form approach to the problem.


In the language of dynamic optimization models, the model states that the driver has two state variables, Yt and ht.
The decision to stop is a smoothly increasing function of ht but a discontinuous function of Yt, with a discrete jump in
the likelihood of stopping when Yt reaches T.

This is a result of the constant marginal utility of income assumption built into the specification of the utility
function. This assumption is standard in the literature on intertemporal substitution in labor supply, and I discuss its
plausibility in more detail in Farber (2005).

This problem arises in other domains, for example, if there is a substantial literature on the retirement decision in
the presence of defined-benefit pension plans with nonmonotone marginal values of continuing to work. James H. Stock
and David A. Wise (1990) present an analysis of retirement in this context that offers an alternative to a full solution
to the problem. See John Rust (1989) and James Berkovec and Steven Stern (1991) for other approaches to estimating
dynamic retirement models.
VOL. 98 NO. 3 farber: reference-dependent preferences and labor supply 1073

III. An Empirical Model of the Labor Supply Decision

Consider the end of each passenger trip as a decision point regarding whether to continue or
to stop driving. A driver will quit for the day when the expected value of continuing to drive first
becomes negative. A reasonable approximate solution to the full dynamic stopping problem can
be implemented empirically as a discrete choice problem. After any trip t during a shift, a driver
can calculate the forward-looking expected value of continuing to drive (the continuation value).
This will be a function of many factors, including hours worked so far on the shift and variables
that affect expectations about future earnings possibilities. It is also affected by accumulated
shift income in an unusual way: when accumulated income is less than the reference income
level, there is an opportunity for the driver to derive utility from additional fare income at the
higher “pre-reference” rate.
In order the implement the model empirically, I first define a latent variable Cijt for driver i on
shift j after trip t that represents the forward-looking expected value of continuing to drive (the
continuation value). An empirical representation of this latent variable is

(6)  Cijt 5 Xijt b 1 dI  3Yijt , Tij 4 1 eijt ,

where

• Tij represent the reference income level for driver i on shift j,

• Yijt represent the income level for driver i on shift j at trip t,
• Xijt is a vector of variables that determine the difference between current utility and the
continuation value,
• b is a parameter vector to be estimated,
• eijt is a random component with a standard normal distribution.
• I  3Yijt , Tijt 4 is an indicator function that equals one if accumulated income is less than the
reference income level and equals zero otherwise, and
• d is a positive parameter that represents the increment to the continuation value when the
reference income level is greater than income.

The second term in equation (6) embodies the incremental utility value of income below the
reference income level that could be earned by continuing to drive.
Driver i on shift j will continue driving after trip t if Cijt $ 0 and stop if Cijt , 0. The prob-
ability of continuing conditional on Tij is

(7)  c
P ijt ZTij 5 F 3Xijt b 1 dI  3Yijt , Tij 4 4 ,

where F 3·4 is the standard normal cumulative distribution function.

If the reference income 1Tij 2 level were known, this specification would imply the usual probit
model, and estimates could be derived by maximum likelihood in the usual way. But the refer-
ence income level is not known and needs to be estimated. Suppose the reference income level
is

(8)  Tij 5 ui 1 mij ,

where ui is an individual mean reference income level and mij is a random component distributed
normally with mean 0 and variance s2m. The mij represent daily deviations from ui in the reference
income level.
1074 THE AMERICAN ECONOMIC REVIEW june 2008

The next step is to derive a likelihood function based on the probability that a shift ends (the
“shift probability”) after trip t in the case where the reference income level is unknown but
defined by equation (8). I present a brief description of the likelihood function here. Think of an
observation as a shift. If the driver stops after trip t, he did not stop after the first t21 trips. Based
on equation (7), it is straightforward to write the joint probability of this pattern conditional
on a value for the reference income level 1Tij 2. Given the assumed distribution for the reference
income level (equation (8)), I “integrate out” Tij to derive the unconditional probability of driver i
on shift j stopping after trip t. The likelihood function for my sample of 538 shifts for 15 drivers
is constructed from these unconditional shift probabilities. In Section V, I compute maximum
likelihood estimates of the parameters of the model 1b, ui, s2m, and d 2.

A. What Constitutes a Test of Reference Dependence in Labor Supply?

A central goal of the analysis is to determine whether daily reference income levels are an
important factor in the labor supply decisions of taxi drivers. Estimates of the model derived in
this section can be used to achieve this goal. The test I use is subjective but relies on a substantive
understanding of the potential role of reference-dependent preferences in labor supply.
Reference dependence is important to the extent that it implies predictable behavior changes
when income reaches the reference income level. In the context of the model, there is a discrete
change in the continuation probability at the reference income level. Empirically, the model
implies that drivers stop with higher probability, though not with certainty, when the reference
income level is reached. Drivers may continue to work beyond this point on some days if the
option value of future fares is sufficiently high.
A test would be straightforward if the reference income level were observed. In this case, one
could examine the data for a discrete increase in the stopping probability at the known reference
levels. Given that reference income levels are not observed, the test must rely on estimates of the
parameters of the econometric model.
The key parameter controlling the change in the continuation probability at the reference point
is d (the increment to the continuation value when income has not reached the reference level). A
substantial positive value for d would imply a significant increase in the probability of stopping
when the reference income level is reached.
A large d is not sufficient for reference-dependent preferences to be useful in predicting labor
supply. It must also be the case that a given individual’s reference income levels are roughly con-
stant or vary predictably over reasonable periods of time. If this were not the case, then variation
across days in labor supply could be rationalized as an optimal response to random changes in
the reference level of income. The standard deviation 1sm 2 of the daily random component deter-
mining reference income is a measure of how much the reference income level varies from day to
day. If the standard deviation is small relative to the mean 1ui 2 , then I can conclude that reference
incomes are fairly stable. In this case, knowing ui could provide important information regarding
when driver i will stop. Specifically, a driver’s mean reference income level 1ui 2 will be strongly
predictive of his daily income.
Intuitively, if reference dependence is important, it is likely that most shifts will end with driv-
ers reaching their reference income level (though not necessarily stopping at that point). If this
is not the case, then reference dependence is not likely to have a substantial effect on behavior.
Thus, another test of the model is to examine the predicted probability that the reference income


The notation is cumbersome due to the discrete shift in the continuation probability when Yijt reaches
Tij. I present the derivation of the likelihood function in the online Appendix available at http://www.aeaweb.
org/articles/php?doi=10.1257/aer.98.3.1069.
VOL. 98 NO. 3 farber: reference-dependent preferences and labor supply 1075

level has been reached at shift’s end. Given the total income observed on each shift, this is a
straightforward calculation based on the estimated parameters of the model.
To summarize, evidence consistent with reference-dependent preferences playing an impor-
tant role in determining daily labor supply of the taxi drivers in my sample would be:

• A substantial decline in the continuation value when the reference income level is reached
(as indexed inversely by the estimated value of d );

• An estimated value of sm that is small relative to the values estimated for the mean reference
income levels (the u vector or the mean of the distribution of ui );

• A driver’s mean reference income level (ui ) strongly predictive of daily income; and

• Attainment of the reference income level on most shifts in the sample.

I examine each of these pieces of evidence in Section V.

IV. Data

The data necessary to carry out my analysis are available on “trip sheets” that drivers fill
out during each shift. Each trip sheet lists the driver’s name, hack number, and date, along
with details of each trip. The information for each trip includes the start time, start location,
end time, end location, and fare. In order to obtain a sample of trip sheets, in the summer of
2000 my research assistants created a list of taxi leasing companies from the current edition
of the New York City Yellow Pages. After contacting more than 70 leasing companies, one
was found that was willing to provide trip sheets. We were sent 244 trip sheets for 13 drivers
covering various dates over the period from June 1999 through May 2000. We contacted the
leasing company again in the summer of 2001, and we were sent an additional 349 trip sheets
for 10 drivers covering various dates over the period from June 2000 through May 2001. Two
of the drivers appear in both groups, so I have a total of 593 trip sheets for 21 drivers over the
period from June 1999 through May 2001. A few of these trip sheets refer to common dates,
so I have data on 584 shifts. Because I am trying to estimate driver-specific mean reference
income levels, I dropped six drivers with ten or fewer shifts from my analysis. These six driv-
ers accounted for a total of 46 shifts, so the analysis sample has a total of 538 shifts and 12,187
trips for 15 drivers.
The drivers in my sample leased their cabs weekly for a fee of $575. Each driver pays for his
own fuel and keeps all of his fare income and tips. An unfortunate consequence of receiving
the trip sheets in an unsystematic fashion is that I have no information on the number of shifts
worked. If a trip sheet is not available for a specific driver on a given day, I cannot determine if
that driver did not work on that day or if the trip sheet was simply not provided. This prevents me
from examining in any conclusive way inter-day relationships in labor supply.
I performed several regularity checks to insure that the trip sheets were reasonably complete
and internally consistent. Where they were not, I cleaned the data using a set of reasonable rules.
These rules are outlined in detail in Farber (2005).
Almost all trips (92 percent) started and ended in Manhattan. My earlier work using these
data (Farber 2005) suggests that there are important differences in stopping behavior between
trips ending outside Manhattan and trips ending within Manhattan, but no significant differences
within these two broad locations. I proceed using an indicator to distinguish Manhattan from
other locations.
1076 THE AMERICAN ECONOMIC REVIEW june 2008

Table 1—Shift Level Summary Statistics, by Driver

Number of Average Working Driving Waiting Break Total Average

Driver shifts trips hours hours hours hours income wage
Driver 1 39 23.56 6.85 4.32 2.53 0.90 157.58 23.16
Driver 2 14 12.29 3.89 2.78 1.11 2.41 97.10 25.11
Driver 3 40 22.10 6.28 4.52 1.76 0.39 147.51 23.89
Driver 4 23 16.52 6.46 3.98 2.48 2.11 144.96 23.65
Driver 5 24 22.29 6.47 4.42 2.05 0.74 160.71 25.59
Driver 6 37 25.32 7.78 5.13 2.64 0.86 172.44 22.54
Driver 7 19 25.58 7.17 5.47 1.70 0.54 162.02 23.23
Driver 8 45 20.27 6.35 3.90 2.45 1.65 133.19 21.46
Driver 9 13 19.46 6.15 4.03 2.13 0.55 157.95 25.78
Driver 10 17 21.29 7.06 4.49 2.57 0.64 165.84 23.59
Driver 11 70 25.06 6.84 4.56 2.28 0.93 172.01 25.62
Driver 12 72 28.10 8.53 5.84 2.69 0.60 203.05 24.01
Driver 13 33 17.06 6.91 4.63 2.29 0.97 163.51 23.91
Driver 14 46 24.46 7.10 4.80 2.30 0.67 156.23 22.00
Driver 15 46 19.17 5.32 3.66 1.66 0.24 128.97 24.72
All 538 22.65 6.84 4.58 2.26 0.87 160.03 23.80

I additionally collected data from the National Atmospheric and Oceanic Administration
(NOAA) on temperature and precipitation in New York City. I collected daily average, minimum
and maximum temperatures, and total daily rainfall and snowfall in Central Park. I also col-
lected hourly rainfall data at LaGuardia Airport.

A. Summary Statistics

Table 1 contains average statistics by shift for each driver. Hours worked per day is defined as
the sum of driving time (the sum over trips of the time between the trip start time and the trip
end time) and waiting time (the sum over trips of the time between the end of the last trip and the
start of the current trip). Waiting time is substantial, accounting for 33 percent of working time
on average. Break time averages about 52 minutes per shift.
There is substantial variation across drivers in average hours worked per day, with means
ranging from 3.89 to 8.62. Still, the majority of the variation in daily work hours is within-driver
variation across days. The standard deviation of daily work hours is 2.50. The R-squared from
a regression of daily hours on a set of driver fixed effects is 0.152, with a residual root mean
squared error of 2.33. Panel A of Figure 1 contains a histogram of hours worked for the 538
shifts. The distribution is single-peaked, with the mode at eight hours.
There is also substantial variation across drivers in total fare income per day, with means
ranging from $97.10 to $228.26. Not surprisingly, daily income covaries strongly with daily
hours with a simple correlation of 0.91. As with hours, the majority of the variation in daily
income is within-driver variation across days. The standard deviation of daily income is $59.82.
The R-squared from a regression of daily income on a set of driver fixed effects is 0.153 with a
residual root mean squared error of $55.77. Panel B of Figure 1 contains a kernel density estimate
of daily income.


Income per day is the sum of fares. Tip income is not measured or accounted for.

All kernel density estimates in this study use the Epanechnikov kernel. The bandwidths are listed in the figures.
VOL. 98 NO. 3 farber: reference-dependent preferences and labor supply 1077

   

     

,FSOFM
EFOTJUZ
'SBDUJPO


                      
"
)PVST
XPSLFE
JO
TIJGU #
4IJGU
JODPNF

]
CBOEXJEUI  




 
,FSOFM
EFOTJUZ

,FSOFM
EFOTJUZ
 




                    
$
5PUBM
NJOVUFT
PO
USJQ
]
CBOEXJEUI   %
'BSF

] CBOEXJEUI
 

Figure 1. Distributions of Hours and Income, by Shift and Kernel Density Estimates, of Total Trip Times
and Fares

The last column of Table 1 contains the daily average for each driver of their hourly wage rate
(total income divided by working hours). These show less inter-driver variation, ranging from a
low of $21.46 to a high of $25.78. The standard deviation of the daily wage rate is $4.52. Most of
this is within-driver variation, as the R-squared from a regression of the daily wage on a set of
driver fixed effects is 0.082 with a residual root mean squared error of $4.39.
Panel C of Figure 1 contains a kernel density estimate of the distribution of total trip times (the
sum of waiting and travel times) for the 12,187 trips in my sample.10 Median time per trip is 15
minutes and the mean is 18.1 minutes. That trips are this short reflects the fact that 92.5 percent
of the trips in my sample begin and end in Manhattan. The fact that the great majority of trips
are short implies that opportunities to stop driving occur frequently.
Panel D of Figure 1 contains a kernel density estimate of the distribution of fares.11 The
median fare is $5.30, and the mean fare is $7.06. Once again, the small size of the fares (which
exclude tips) is due to the fact that most trips are intra-Manhattan. The average intra-Manhattan
fare is $5.93, while the average fare that starts or ends outside Manhattan is $20.76. The small
blip at $30 represents the flat rate between Kennedy Airport and Manhattan in force during my
sample period. The fact that the great majority of fares are small implies that the first opportu-
nity to stop driving after the reference income is exceeded will generally arise at an income level
not greatly in excess of the reference income level.

10
I have truncated this distribution at 90 minutes for total trip time. There are 38 trips with times greater than 90
minutes.
11
I have truncated this distribution at $50 for fares. There are eight trips with fares greater than $50.
1078 THE AMERICAN ECONOMIC REVIEW june 2008

Table 2—Maximum Likelihood Estimates of Reference-Dependent Labor Supply Model

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

b̂ (contprob) (constant) 20.691 — — —
(0.243)
û (mean ref inc) 159.02 206.71 250.86 —
(4.99) (7.98) (16.47)
d̂ (cont increment) 3.40 5.35 4.85 5.38
(0.279) (0.573) (0.711) (0.545)
ŝ2 (ref inc var) 3,199.4 10,440.0 15,944.3 8,236.2
(294.0) (1,660.7) (3,652.1) (1,222.2)

Driver ûi (15) No No No Yes

Vars in cont prob
Driver FE’s (14) No No Yes No
Accum hours (7) No Yes Yes Yes
Weather (4) No Yes Yes Yes
Day shift and end (2) No Yes Yes Yes
Location (1) No Yes Yes Yes
Day-of-Week (6) No Yes Yes Yes
Hour-of-Day (18) No Yes Yes Yes
Log(L) 21,867.8 21,631.6 21,572.8 21,606.0
Number of parameters 4 43 57 57

Note: The sample includes 12,187 trips in 538 shifts for 15 drivers. Models 2–4 also include a constant in the X vec-
tor for the continuation probability. The hours from 5 am to 10 am have a common fixed effect. Standard errors are
reported in parentheses.

V. Estimation of the Labor Supply Model

In this section I present maximum likelihood estimates of the parameters 1b, ui, s2m, and d 2
of the reference-dependent labor supply model based on the likelihood function described in
Section III and on the data described in the previous section.
The first column of Table 2 contains estimates of a restricted model that constrains all driv-
ers to have the same mean reference income level and the X vector to contain only a constant.
The evidence is mixed with regard to the role of reference income levels in determining labor
supply. The estimated mean reference income level is reasonable at $159.02. The estimate of d,
which indexes (inversely) the change in the probability of continuing to drive once the reference
income level for the day is reached, is substantial at 3.40 and statistically significantly different
from zero. This would seem to be strong evidence of the importance of reference-dependent
preferences in this context.
There is substantial inter-shift variation, however, around the mean reference income level.
The estimated variance of 3,199.4 implies a standard deviation of $56.60. To the extent that this
represents daily variation in the reference income level for a particular driver, the predictive
power of the reference income level for daily labor supply would be quite limited.
The second column of Table 2 contains estimates of the model that include a set of variables
in the continuation function (equation (6)) that are meant to capture earnings opportunities and
other factors that would affect a driver’s continuation probability. These are:

• Indicators for eight categories of hours worked at trip end;

• Six indicators for the day of week;
• Indicators for 18 clock hours at trip end;
• A day-shift indicator;
VOL. 98 NO. 3 farber: reference-dependent preferences and labor supply 1079

• An interaction of day-shift with clock hours 3–4 pm meant to capture the likelihood that a
day-shift driver must turn the car over to a night-shift driver at that time of day;
• Four variables measuring weather, including daily snowfall, hourly rainfall, high heat (maxi-
mum temperature . 5 80 degrees), and cold (minimum temperature , 30 degrees); and
• An indicator for non-Manhattan location.

These 39 variables are clearly important in determining labor supply as the log-likelihood
improves from 21,867.8 to 21,631.6 when they are included in the model. The estimated mean
reference income level increases to $206.71. The estimate of d increases to 5.35, implying once
again that reference income factors may be important, but the inter-shift variance of reference
income around its mean increases dramatically to 10,440.0 implying a standard deviation in the
reference income level of $102.17.
Column 3 of Table 2 contains estimates of the model where the continuation function addi-
tionally includes driver fixed effects. These capture inter-driver differences in the marginal disu-
tility of labor. The driver fixed effects dramatically improve the fit of the model, increasing the
log-likelihood from 21,631.6 to 21,572.8. The estimates of the reference-dependence param-
eters continue to be characterized by (a) a substantial increment 1d 2 to the continuation function
before the reference income level is reached; and (b) substantial variation in the reference income
level.
The first three columns of Table 2 show an important constraint in the models, that is, all driv-
ers are assumed to have the same mean reference income level. Indeed, the large estimate of the
variance in the reference income level could be due to differences across drivers in their mean
reference income level. The obvious solution is to estimate separate mean reference incomes for
each driver. However, estimation of two sets of driver fixed effects in the model is asking too
much of the data.12 As a result, I estimate a model with unique mean reference income levels for
each driver 1ui 2 , but without driver fixed effects in the continuation function. These estimates are
contained in column 4 of Table 2.
Allowing for different mean reference income levels improves the fit of the model substan-
tially, with the log-likelihood improving from 21,631.6 (in column 2 of the table) to 21,606.0.
The hypothesis that all drivers that have the same mean reference income level can be rejected at
any reasonable significance level 1 p-value 5 3.8E-6) using a likelihood-ratio test. Interestingly,
the model with fixed effects in the continuation function (column 3) fits the data better (a higher
log-likelihood value) than does the model with driver-specific mean reference incomes.13
With regard to estimates of parameters related to the role of reference income levels on labor
supply, the evidence remains mixed. The general pattern across the specifications in Table 2
is similar, and I focus the discussion on the estimates in column 4. The estimate of d, which
indexes (inversely) the change in the probability of continuing to drive once the reference income
level for the day is reached, is substantial at 5.38 and significantly different from zero ( p-value
, 1.E-10). Given the standard normal distribution of the latent variable determining the prob-
ability of continuing to drive, as shown in equation (6), this implies that the continuation prob-
ability is close to zero once the reference income level is reached, regardless of when in the shift
is reached. This would seem to be strong evidence of the importance of reference-dependent

12
I did attempt to estimate such a model, and I was not able to find an optimum, despite considerable effort. It is
likely that more observations (both in the form of more drivers and more shifts per driver) are necessary to estimate
such a model.
13
While I was not able find a unique optimum for the full model that nests the two models in columns 3 and 4 of
Table 2, I was able to find log-likelihood values in the range of 21,554.4. This is sufficiently high to be able to reject
that either of the sets of fixed effects is zero. Thus, it appears that mean reference income levels do vary by driver even
after driver differences in the continuation function are accounted for.
1080 THE AMERICAN ECONOMIC REVIEW june 2008

Table 3—Maximum Likelihood Estimates of preferences in this context. While drivers may
Reference-Dependent Labor Supply Model
(Column 4 of Table 2)
or may not stop before their reference income
level is reached, they are almost sure to stop
Variable b̂ Driver û once they reach that level.
Hour # 2 2.242 1 245.66 With the allowance for driver-specific mean
10.6012 128.272 reference levels, the variance of the reference
Hour 3–5 1.420 2 93.06
10.2482 137.872
income level reflects within-driver day-to-day
Hour 6 0.446 3 166.95 variation in reference income. What is perhaps
10.1532 120.972 surprising is that the estimate of the within-
Hour 7 0.210 4 181.83
10.1312 131.722
driver variance remains very large at 8326.2,
Hour 9 0.094 5 185.18 implying a within-driver daily standard devia-
10.1872 121.632 tion in reference income of $91.25. Based on
Hour 10 20.601 6 284.67
10.2102 140.532
the assumption of normality, the inter-quartile
Hour 11 20.763 7 211.44 range for a particular driver with a mean refer-
10.2582 135.282 ence income level of $250 is from $188 to $311,
Hour $ 12 21.088 8 190.40
10.3272 118.762
and the 95 percent confidence interval is from
Min temp , 30 20.033 9 182.52 $99 to $400. The large variances estimated in
10.1302 139.132 the first three columns of Table 2 are not sim-
Max temp $ 80 0.003 10 186.61
10.1532 132.002
ply the result of different drivers having differ-
Hourly rain 20.225 11 217.17 ent mean reference income levels.
10.1782 116.132 Table 3 contains more detailed estimates of
Daily snow 0.035 12 231.93
10.1712 112.802
the full model in column 4 of Table 2. The esti-
Non-Manhattan 20.568 13 181.39 mates of b, in the first column, show a strongly
10.1662 127.142 decreasing continuation probability with hours
14 232.82
123.792
worked on the shift. The standard normal
15 150.82 latent variable determining the probability
121.512 of continuing (equation (6)) decreases by 3.3
from the beginning of the shift to the twelfth
Notes: The sample includes 12,187 trips in 538 shifts for
15 drivers. The model also includes in the continuation hour. The weather does not appear to be sig-
function a constant, 6 indicators for day of the week, a nificantly related to the labor supply decision.
day-shift and end-of-day shift indicator, and 18 indicators Ending a fare outside of Manhattan results in
for hour of the day. The hours from 5 am to 10 am have
a common fixed effect. Standard errors are reported in a significantly lower probability of continuing,
parentheses. See column 4 of Table 2 for more details. probably due to the fact that many drivers live
outside of Manhattan.
The estimates of the individual mean refer-
ence income levels (the u vector) vary significantly across drivers, from a low of $93.06 for driver
2 to a high of $284.67 for driver 6. The mean reference income is $196.16 with a standard devia-
tion across the 15 drivers of $44.70.

VI. Does Reference Dependence Have Predictive Value?

In Section IIIA, I listed four types of evidence consistent with reference dependent preferences
playing an important role in determining daily labor supply.

1. It is clearly the case that the continuation probability decreases dramatically when the refer-
ence income level is reached. The estimated value of d of 5.38 (column 4 of Table 2) implies
a decrease of over five standard deviations in the standard normal distribution of the continu-
ation value when the reference income level is reached. It appears that drivers are almost cer-
tain to stop after they reach their reference income level for the day.
VOL. 98 NO. 3 farber: reference-dependent preferences and labor supply 1081

2. The large estimated within-driver inter-day variance in the reference income level implies that
there is little consistency in a driver’s reference income level day to day.

3. There is a strong positive correlation between both average income and average hours across
shifts for a given driver and that driver’s estimated mean reference income level (ui ). These
correlations are 0.76 and 0.86, respectively. In order to investigate whether ui is a strong pre-
dictor of daily income, I estimate the following simple regression at the shift level of income
on the estimated mean reference income level for the 538 shifts in my sample:

(9) Yij 5 a0 1 a1ûi 1 gij ,

where Yij is income for driver i on shift j. While the estimate of a1 is significantly positive at
0.408 (s.e. 5 0.063), it is substantially less than one and the R2 from this regression is only
0.07. Thus, very little of the overall variation in income across shifts can be accounted for by
variation across drivers in the estimated mean reference income level.14

4. The probability that the reference income level is attained on a given shift is relatively low.
The probability that income on shift j for driver i 1Yij 2 exceeds the reference level for that shift
is the CDF of a normal distribution with mean ui and variance equal to 8,236.2. This is:

(10) P 1Tij # Yij 2 5 F 3 1Yij 2 ui 2/sm 4 .

Figure 2 contains a plot of the fraction of shifts where the predicted probability that the refer-
ence income level was reached during the shift falls in each of ten intervals from zero to one.
It is clear that the reference income level is reached in only a minority of cases. The mean
probability is 0.34, implying that the reference income level was reached in only 34 percent
of shifts in the sample. The seventy-fifth percentile of the distribution of probabilities that the
reference income level is reached is 0.49, and the ninetieth percentile of the same distribu-
tion is 0.63. Thus, it appears that drivers generally stop before the reference income level is
reached.15

Taken together, these findings leave a puzzle. The reference-dependent utility model builds a
particular nonlinearity into the relationship of income with the continuation probability, and the
large estimated value of d is evidence for this nonlinearity: drivers generally will stop if they
reach their reference income level. The reference income level for a given driver varies unpre-
dictably day to day, however, and most shifts end before the driver reaches his reference income
level. Essentially, the data show more smoothness in the relationship between income and the
continuation and stopping probabilities than seems consistent with an important role for refer-
ence-dependent preferences. The relatively large unexplained within-driver variation in income
across shifts also belies an important role for daily reference dependence in this context.

14
In fact, most of the variation in income across shifts is within-driver variation. A regression of daily income on a
set of driver fixed effects yields an R2 of 0.15.
15
In order to investigate the possibility that drivers end their shift when they are near, but not necessarily at, the
reference income level, I repeat this analysis calculating for each shift the probability that, at shift’s end, accumulated
income exceeds 90 percent of the reference income level. While these probabilities are higher, they are still fairly low.
The mean probability that a shift ends at some point after 90 percent of the reference income level has been reached is
0.41, implying that 90 percent of the reference income level was reached in only 41 percent of shifts in the sample.
1082 THE AMERICAN ECONOMIC REVIEW june 2008

¢

¢

¢

¢
1S :  5
BU
TIJGU
FOE

¢

¢

¢

¢

¢

¢

    

'SBDUJPO
XJUI
1S :  5
JO
JOUFSWBM
BU
TIJGU
FOE

Figure 2. Distribution of Predicted Probability of Income Exceeding the Reference Level at Shift End
(Based on parameters in column 4 of Table 2)

references
Bateman, Ian J., Alistair Munro, Bruce Rhodes, Chris Starmer, and Robert Sugden. 1997. “A Test of the
Theory of Reference-Dependent Preferences.” Quarterly Journal of Economics, 112(2): 479–505.
Berkovec, James, and Steven Stern. 1991. “Job Exit Behavior of Older Men.” Econometrica, 59(1): 189–210.
Camerer, Colin, Linda Babcock, George Loewenstein, and Richard Thaler. 1997. “Labor Supply of New
York City Cabdrivers: One Day at a Time.” Quarterly Journal of Economics, 112(2): 407–41.
Farber, Henry S. 2005. “Is Tomorrow Another Day? The Labor Supply of New York City Cabdrivers.”
Journal of Political Economy, 113(1): 46–82.
Fehr, Ernst, and Lorenz Goette. 2007. “Do Workers Work More if Wages Are High? Evidence from a Ran-
domized Field Experiment.” American Economic Review, 97(1): 298–317.
Genesove, David, and Christopher Mayer. 2001. “Loss Aversion and Seller Behavior: Evidence from the
Housing Market.” Quarterly Journal of Economics, 116(4): 1233–60.
Kahneman, Daniel, Jack L. Knetsch, and Richard H. Thaler. 1991. “The Endowment Effect, Loss Aver-
sion, and Status Quo Bias: Anomalies.” Journal of Economic Perspectives, 5(1): 193–206.
Kőszegi, Botond, and Matthew Rabin. 2006. “A Model of Reference-Dependent Preferences.” Quarterly
Journal of Economics, 121(4): 1133–65.
Munro, Alistair, and Robert Sugden. 2003. “On the Theory of Reference-Dependent Preferences.” Jour-
nal of Economic Behavior and Organization, 50(4): 407–28.
Neumark, David, and Andrew Postlewaite. 1998. “Relative Income Concerns and the Rise in Married
Women’s Employment.” Journal of Public Economics, 70(1): 157–83.
Oettinger, Gerald S. 1999. “An Empirical Analysis of the Daily Labor Supply of Stadium Vendors.” Jour-
nal of Political Economy, 107(2): 360–92.
Rust, John. 1989. “A Dynamic Programming Model of Retirement Behavior.” In The Economics of
Aging, ed. David A. Wise, 359–98. Chicago: University of Chicago Press.
Stock, James H., and David A. Wise. 1990. “Pensions, the Option Value of Work, and Retirement.” Econo-
metrica, 58(5): 1151–80.
Thaler, Richard. 1980. “Toward a Positive Theory of Consumer Choice.” Journal of Economic Behav-
ior and Organization, 1(1): 39–60.
Tversky, Amos, and Daniel Kahneman. 1991. “Loss Aversion in Riskless Choice: A Reference-Dependent
Model.” Quarterly Journal of Economics, 106(4): 1039–61.