Topics in Applied Microeconomics

Extended Referee Report

December 19, 2023
2807948A

1 Preface
Being keen on cognitive bias research, I gave a presentation about the article "Does market
experience eliminate market anomalies?" (2003) by J. List. Unfortunately, later I found out
that the author did not make the data set available, therefore it was unfeasible for me to
replicate and extend the selected paper. Thus, I decided to change the paper and chose one
of the approved articles from the reading list, namely "New York City Cab Drivers’ Labor
Supply Revisited: Reference-Dependent Preferences with Rational Expectations Targets for
Hours and Income" (2011) by Crawford and Meng. In this report, I will describe the paper’s
main findings, replicate the main result and present the sub-analysis by the time of day to
check the robustness of the results.

2 New York City Cab Drivers’ Labour Supply Revisited

2.1 Summary of Results
A neoclassical model of labour supply predicts a positive wage elasticity of hours, yet a
study by Camerer et al. (1997) found a strongly negative elasticity of hours. This discrepancy
implies that if in reality, the drivers’ labour supply is the result of optimization with reference-
dependent preferences, the usual estimates of elasticities are misleading. Hence, Crawford
and Meng (2011) depart from the neoclassical theory and propose a new model of cab drivers’
labour supply based on reference-dependent preferences, having targets for both income and
hours.
This paper uses reference-dependent utility theory developed by Kőszegi and Rabin
(2006), which is more general than models used in previous studies on the labour supply.
Other research mainly focused on income targets, while Crawford and Meng (2011) use hours
targets as well. Moreover, a model by Kőszegi and Rabin endogenize targets by setting them
equal to theoretical rational expectations of hours and income. In this paper, authors proxy
drivers’ point-expectation income and hours targets via the sample averages up to but not
including the day in estimating.
The driver’s utility function V (I, H ∣ I r , H r ), being a weighted average of consump-
tion utility (U1 (I) + U2 (H)) and gain-loss utility R (I, H ∣ I r , H r ) used in the analysis is as
follows:

V (I, H ∣ I r , H r ) = (1 − η) (U1 (I) + U2 (H)) + ηR (I, H ∣ I r , H r ) , where

R (I, H ∣ I r , H r ) = 1(I−I r ≤0) λ (U1 (I) − U1 (I r ))
+ 1(I−I r >0) (U1 (I) − U1 (I r )) + 1(H−H r ≥0) λ (U2 (H) − U2 (H r ))
+ 1(H−H r <0) (U2 (H) − U2 (H r )) ,

where I and H denote his income earned and hours worked that day, and I r and H r denote
his income and hours targets for the day. Note that, in the spirit of Kahneman and Tversky,
Topics in Applied Microeconomics 2807948A

losses have a constant weight λ, i.e. the coefficient of loss aversion. Estimates that allow for
different λ for income and hours yield no significant difference, so authors assume λ is the
same for both.
Authors use Farber’s data (2005, 2008), and start with estimating probit models of the
probability of stopping with various control strategies. Also, they split the sample according
to whether a driver’s earnings for the first x hours of the day are higher or lower than his
proxied expectations. When controlled for all available relevant variables, they find that
when x > 1, the number of hours driven is significant, and when x < 1, the income earned
during the day is significant. This reversal of the pattern is inconsistent with the neoclassical
model, while is explained by a reference-dependent model, in which drivers are more likely
to stop when the second target is reached, regardless of whether it is a target for hours or
income. This evidence allows authors to confirm that reference-dependent theory matches
cab drivers’ behaviour well and allows them to carry on with the structural estimation.
The main result of the article is the empirical estimations for the model of the reference-
dependent preferences developed in the paper. Following Farber (2008), authors assume the
following form of the utility function:
θ
U (I, H) = I − H 1+ρ ,
(1 + ρ)

where ρ is the elasticity of the marginal rate of substitution. They derive the likelihood
function and present structural estimates of the parameters for 5 different control techniques
to check robustness. The main result is replicated and discussed in subsection 2.3.

2.2 Discussion
This study extends previous analysis of drivers’ labour supply, building on the theory of
reference-dependent utility. Development on the issue was made by Farber (2005) who
collected and analysed data also used in the discussed paper. He found that the probability
of stopping work on a day is significantly related to realized income that day, although the
effect is not significant after including fixed effects and relevant controls. Taking this analysis
further, Farber (2008) estimated a structural model based only on income targets, and found
that this model fits reality better than a neoclassical model. Crawford and Meng in contrast
realized that both hours and income targets are important and adopted Kőszegi and Rabin’s
model for New York drivers data.
It seems that the method used, namely changing the neoclassical model to the one based
on reference-dependent preferences is a good fit for this case. Of course, there are many
possible models with different utility functions that can describe taxi drivers’ behaviour,
but the evidence and the results of the paper prove that Kőszegi and Rabin’s model is
appropriate for this data. The main finding suggests that given the reference-dependent
utility assumption, there is no contradiction between a negative wage elasticity and a positive
incentive to work, unlike when a neoclassical model is estimated.
In many other jobs, people do not choose the amount of work they do, i.e. many people
have fixed-hour contracts, which is not the case with taxi drivers, who can choose when to
stop working. This allowed the authors to obtain the described results as we can observe
how drivers make decisions based on the hours driven and income achieved during the day.
Therefore, these results are not easily generalisable as such data may not be available for
many other professions. However, the basis of this model, namely the reference-dependent
utility may apply to other job markets, which is an important direction for future research
and further testing of a neoclassical model.

2
Topics in Applied Microeconomics 2807948A

There are many possible further extensions of the paper (many of which are not feasible
to us as they require additional data). A more comprehensive investigation of taxi drivers’
behaviour and people with different jobs but similar choices with larger data sets (the one
used in this study only includes 20 drivers) is necessary. Moreover, Crawford and Meng
(2011) use a simple estimation for targets, and the results may not be robust if wage and
income targets are modelled using a different method. The authors believe that further
extensions of the paper will yield a reference-dependent model of labour supply that will
prove that such a model significantly improves over the neoclassical model. However, it
seems that the use of another data set may prove the choice of the reference-dependent
model to be wrong, and the findings of Crawford and Meng are specific to their data set, as
shown in the following study.
Farber (2015) used a new (2009-2013) data set consisting of information on 8,802 New
York taxi drivers to estimate the model of labour supply. He found that the reference-
dependent model does not describe their behaviour well, while the results are consistent
with the neoclassical model, as drivers tend to respond positively to increases in earning
opportunities (labour supply elasticities are positive and range from 0.4 to 0.8). Hence, it
is clear that the research on the issue is far from over, as there is evidence in favour of both
neoclassical and reference-dependent models. More investigation of the behaviour of taxi
drivers and people with similar jobs is needed.
Moreover, Farber (2015) finds that different drivers may have different elasticities, and
while some drivers have reference-dependent preferences, others may be optimizers according
to the neoclassical model. Farber splits the sample to night shift drivers and day shift drivers
and finds that the majority of drivers have a positive elasticity, while night shift drivers in
general behave differently. In this spirit, we divide the data set into weekend shift drivers
and weekday shift drivers in the extension presented in the next section.

2.3 Replication
In this subsection, we will replicate and discuss the evaluation of the structural model, the
main result of the article. The authors did make a replication package available, but the
code did not work correctly and it was not easy to replicate the results. We use the data set
that the authors provided, and our Stata code is based on the one that the authors provided,
with all the bugs and errors fixed.
The authors estimate the parameters of the model using the maximum likelihood
method. Moreover, they test the robustness of their results for different estimations of
wage and income targets. They show that the result is robust in 2 out of 3 cases. For
brevity, we will only replicate the main result (Crawford and Meng, 2011, table 4, column
1), and omit the robustness check, as it yields the same result and is similar qualitatively.
As we have said before, we used the available replication package, but we had to try many
methods to find a way to fix all the bugs and estimate the parameters. The main result
estimates are reported with higher accuracy than in the article to confirm replication and
are available in Table 1.
We have estimated the structural model and received the same result as the authors
reported in the article. Following the paper, we control for driver fixed effects, day of the
week, hour of the day, location, and weather. The estimates are significant at the 1% level,
including both λH and λI , ruling out that the model is reduced to a neoclassical level (when
η(λ − 1) = 0). Next, we will present a sub-analysis in which we divide the shifts for the
weekend and weekdays and estimate the same structural model to show that the results are
not robust.

3
Topics in Applied Microeconomics 2807948A

Table 1: Structural Estimates: Replication and Sub-analysis

Replication Weekend Weekdays Night hours

η(λH − 1) 1.3091*** -0.640 -0.414*** -0.675
η(λI − 1) 0.5114*** -1.000 -0.626 -1.000
θ 0.0349*** 0.815 0.324** 0.556**
ρ 0.5971*** 0.976*** 0.763** 1.123***
σ+ 0.1264 0.218 0.081 0.237
Observations 10,337 3,894 8,256 7,194
Log-likelihood -1,321 -449 -556 -919
We indicate significance only and not p-values for brevity (p-values es-
timates are the same as in the article). Control variables include driver
fixed effects, day of the week, hour of the day, location, and weather.
***Significant at the 1 per cent level. **Significant at the 5 per cent level.

3 Sub-analysis: Weekend and Workdays

Many possible extensions are not feasible to us, as they require other data, hence we proceed
with a straightforward, yet very insightful, sub-analysis and divide the data into weekdays
(Monday to Thursday) and weekends (Friday to Sunday). We consider Friday a weekend as
during this day of the week, the use of a taxi may be more similar to the way people use it
on Saturday and Sunday. Also, this divides data in a more balanced way, as Sunday is the
day with the smallest number of observations.
The results of the estimation are presented in Table 1, columns 2 and 3. For the
weekend shifts, we can not reject the hypothesis that η(λ − 1) = 0. For weekdays η(λH − 1) is
significant at the 1% level (however, the estimate is very different to the result in the article),
while η(λI − 1) is not, hence the sub-analysis shows that the results of the paper may not be
robust if we split the data set. Similarly, the estimate of θ is close to the one in the paper,
but only significant for weekdays at the 5% level, and not for the weekend. Our estimates
for ρ (significant at the 1% level) and σ (not significant) are close to the coefficients in the
article. We omit the test of whether λH is equal to λI because this question is not relevant
to our extension.
Moreover, for the sake of interest and going beyond our extension, we also evaluate the
model using night hours only (with the end of the ride from 5:00 pm to 5:59 am), following
Farber (2015). The estimates can be seen in Table 1, column 4. The estimate for ρ is again
significant at the 1% level, and the coefficient is not far from other estimations. θ, similarly
for weekdays, is significant at 5% level. However, neither η(λI − 1) not η(λH − 1) is not
significant, as well as the estimate of σ.
These results may suggest that, despite being robust to different modelling of hour
and income targets, the results of the paper are specific to the data set used, which includes
only 20 drivers. This statement is supported by the results of Farber (2015) who found
that different drivers may behave differently, and most of them optimize according to the
neoclassical model. In our analysis, we have shown that the results obtained by Crawford
and Meng (2011) are not robust under a sub-sample estimation. It is clear that there is
evidence in favour of both the reference-dependent and the neoclassical predictions, hence
more research in this area is needed to conclude which model fits the behaviour of taxi drivers
and professionals with similar choice sets best.

4
Topics in Applied Microeconomics 2807948A

References
[1] Camerer, C., Babcock, L., Loewenstein, G., and Thaler, R. (1997). Labor supply of new
york city cabdrivers: One day at a time. The Quarterly Journal of Economics.

[2] Crawford, V. P. and Meng, J. (2011). New York City cab drivers’ labor supply revisited:
Reference-dependent preferences with rational-expectations targets for hours and income.
American Economic Review, 101(5):1912–1932.

[3] Farber, H. S. (2005). Is tomorrow another day? the labor supply of new york city
cabdrivers. Journal of Political Economy, 113(1):46–82.

[4] Farber, H. S. (2008). Reference-dependent preferences and labor supply: The case of
new york city taxi drivers. American Economic Review, 98(3):1069–1082.

[5] Farber, H. S. (2015). Why you can’t find a taxi in the rain and other labor supply lessons
from cab drivers. The Quarterly Journal of Economics, 130(4):1975–2026.

[6] Kahneman, D. and Tversky, A. (1979). Prospect theory: An analysis of decision under
risk. Econometrica, 47(2):263–291.

[7] Kőszegi, B. and Rabin, M. (2006). A model of reference-dependent preferences. The

Quarterly Journal of Economics, 121(4):1133–1165.