HAL Id: hal-03393093
https://sciencespo.hal.science/hal-03393093
Preprint submitted on 21 Oct 2021
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Distributed under a Creative Commons Attribution - ShareAlike 4.0 International License
When in Rome… on local norms and sentencing decisions
David Abrams, Roberto Galbiati, Emeric Henry, Arnaud Philippe
To cite this version:
David Abrams, Roberto Galbiati, Emeric Henry, Arnaud Philippe. When in Rome… on local norms
and sentencing decisions. 2019. �hal-03393093�
LIEPP Working Paper
April 2019, nº88
When in Rome…
on local norms and sentencing
decisions
David ABRAMS
University of Pennsylvania
Roberto GALBIATI
CNRS-Sciences Po and CEPR
roberto.galbiati@sciencespo.fr
Emeric HENRY
Sciences Po and CEPR
Arnaud PHILIPPE
University of Bristol
www.sciencespo.fr/liepp
© 2019 by the author. All rights reserved.
How to cite this publication :
David ABRAMS, Roberto GALBIATI, Emeric HENRY and Arnaud PHILIPPE, When in Rome… on
local norms and sentencing decisions, Sciences Po LIEPP Working Paper n°88, 2019-04-17.
When in Rome...
on local norms and sentencing decisions
∗
David Abrams
1
Roberto Galbiati
2
Emeric Henry
3
Arnaud Philippe
4
April 17, 2019
Abstract
In this paper, we show that sentencing norms vary widely even across geographically
close units. By examining North Carolina’s unique judicial rotation system, we show
that judges arriving in a new court gradually converge to local sentencing norms. We
document factors that facilitate this convergence and show that sentencing norms are
predicted by preferences of the local constituents. We build on these empirical results
to analyze theoretically the delegation trade-off faced by a social planner: the judge
can learn the local norm, but only at the cost of potential capture.
1
University of Pennsylvania
2
CNRS-SciencesPo and CEPR
3
SciencesPo and CEPR
4
University of Bristol
∗
We wish to thank Charles Angelucci, Alberto Bisin, Patrick Le Bihan, Olivier Marie, Aurelie Ouss,
Salvatore Piccolo, Shanker Satyanath, Joel Van Der Weele, and seminar participants at the University
of Amsterdam, University of Bristol, NYU, Oxford University, University of Pennsylvania, University of
Bergamo, Collegio Carlo Alberto, University of Paris Ouest, Erasmus University, CELS and IAST-Toulouse
for useful comments and discussions. Kriti Mahajan, Chloe Nibourel, Alyssa Staats and Anton Strezhnev
provided excellent research assistance. Ryan Fackler is owed tremendous thanks for conveying his extensive
knowledge of the data. Roberto Galbiati acknowleges the support of the LIEPP through a public grant
overseen by the French National Research Agency (ANR) as part of the Investissements dAvenir program
LIEPP (reference: ANR-11-LABX0091, ANR-11-IDEX-0005-02)
1
1 Introduction
The ancient Romans appreciated the merits of a flexible interpretation and application of the
law, as evidenced by the Latin maxim summum ius, summa iniuria (Cicero, de officiis, 30),
meaning “supreme law, supreme injustice.” In modern legal practice, legislators still grapple
with the necessity of granting judicial autonomy to allow for flexibility while setting forth
legal rules. This autonomy means that sentences will vary along at least three dimensions:
case, judge, and location. At the case level, judges take into account specific characteristics,
such as mitigating or aggravating factors, in assigning criminal sentences. Sentences also vary
by judge, due to differences in ideology, or potentially capture by local interests. Finally,
sentences may vary spatially, due to local sentencing norms. While judges’ biases have
been extensively studied (Abrams et al. 2012, Lim et al. 2015, Cohen and Yang, 2018
and Schanzenbach and Tiller, 2007), we know little about the existence, magnitude, and
determinants of spatial variation in sentencing.
This lack of evidence is primarily due to the empirical difficulty of disentangling geo-
graphical differences in sentencing norms from spatial variation in unobservable crime char-
acteristics or spatial clustering in judges’ preferences. Analyzing a unique setting where
judges must rotate across judicial districts, we find that judges arriving at a new court grad-
ually converge to the local sentencing averages – appearing to follow the well known proverb,
“when in Rome, do as the Romans do.” We thus both demonstrate the existence of local
sentencing norms and document how judges dynamically adapt to them.
Later in the paper, we discuss the implications of these results for the question of optimal
delegation of sentencing to judges. Judges, when granted autonomy, can obtain information
about case specificities, but also, as we emphasize in our main contribution, about local
norms, information that is costly for the planner to obtain. Letting judges adapt to case
specific details is clearly socially desirable if the planner wants to establish a precise grada-
tion of sentences. Adaptation to local norms of behavior is a more contentious issue. On
the one hand, the planner has a set of guiding principles that should not be affected by local
preferences. However, the planner may also want to at least partially adapt to local norms
to limit the cost of enforcement, which can increase when sentences conflict with local pref-
erences (Acemoglu and Jackson, 2017 and Hay and Shleifer, 1998). We present a theoretical
model, building on the empirical evidence and including these trade-offs, that compares the
primary instruments used in practice to constrain judges: sentencing guidelines, rotation,
and judicial elections.
Our empirical analysis exploits the unique institutional features of North Carolina’s Su-
perior Court system and a rich dataset including all criminal justice decisions made in these
2
courts between 1998 and 2010. In this system, judges are subject to elections as well as regu-
lar spatial rotation. The State is divided into eight divisions, each of which are subsequently
divided into a variable number of districts. Judges are elected in one particular district, but
they do not stay there permanently. Every six months, in January and July, they have to
change district within their division, according to a schedule determined by the State Chief
Justice.
The starting point of our analysis, and one of our main contributions, is to show that
judges gradually adapt to local sentencing practices when they arrive in a new district.
Specifically, for each judicial district, we compute the average sentence in a crime category
given by “senior judges” (judges elected before 1998), who presumably already have good
knowledge of the local conditions. We refer to this average sentence at the district crime level
as the “local sentencing norm”. Restricting the data to decisions made by “junior judges”
(judges elected after 1998), for whom we observe the full history of sentencing, we show
that the absolute value of the distance between the chosen sentence and the local sentencing
norm decreases with the number of cases examined by the judge in the district. Our results
imply that comparing the hundredth case to the first in any given district, the judge gives a
sentence 24 days closer to the local sentencing norm.
By documenting the evolution of judicial sentencing as a function of time spent in a
district, our methodological approach shows both the existence of a local sentencing normal
and gradual adaptation to it. In particular, our identification strategy overcomes several
challenges, that we discuss in detail in Section 4.1. A first concern is that judges recently
arrived in a district might be assigned cases more distant from the local norm. We provide
evidence that this is not the case by showing that observable characteristics, such as crime
type or the defendant’s sex or race, do not evolve over time spent in a district. Second,
selection might explain our results if judges are able to affect the time they spend in the
different districts. To eliminate this possibility, we focus on a balanced sample of judges,
restricting our working sample to the first 400 decisions in a district for judges making at
least 400 decisions. Finally, by using judicial rotation as our source of identification, we can
separate adaptation to a local norm from learning about the law in general.
Further, we differentiate decisions made by judges working in their home district (which
we call the home district) from those made outside and show that convergence only happens
for those in the second category. This is consistent with the idea that judges are already aware
of local sentencing practices in their home district. Digging into the details of the learning
process we document that early sentences made in a district - before learning takes place
- are partly determined by the sentencing norm that prevails in the judge’s home district,
an indicator of the judge’s intrinsic preferences. When judges from relatively tougher (more
3
lenient) home districts rotate to a new location, their sentencing converges to the new local
sentencing norm from above (below).
Our analysis, exploiting assignment of judges to districts, also allows us to establish that
spatial variation in sentencing is much larger than judge-level variations. Regressing sen-
tences on judge and district fixed effects while controlling for crime characteristics, we find
that the standard deviation in district fixed effects is twice as large as that of judge fixed
effects. This suggests that, while the literature has mostly focused on judicial biases, a sig-
nificant part of the variation attributed to judges could in fact be due to local characteristics.
After having shown the existence of a local norm of sentencing, in Section 4.3 we examine
what factors seem to explain variations in the norm. There are three main candidates. Local
norms could reflect constraints that district judges take into account (such as constraints on
the police force or on prisons), preferences of the population (norms of behavior and customs)
or just a court-specific culture unrelated to fundamentals. Our analysis suggests that the
main driver is local preferences: controlling for district and crime fixed effects, variations in
the prevalence of a particular crime in a district decreases average sentences for that crime
in that district. We interpret this as the judge adapting to the norm of greater tolerance
towards certain types of crimes. In a similar vein, we show a district’s sentencing norms are
correlated with a community’s votes in referenda on criminal justice questions. Site-specific
variables measuring resource constraints, in particular prison overcrowding, and those trying
to capture culture, do not appear to play a role.
Our empirical findings show that the judge, when granted autonomy, adapts both to
case characteristics (such as mitigating factors) as well as local preferences. While adapting
to case characteristics is welfare-enhancing for a planner wanting to have a gradation of
sentences, adaptation to local norms is a more contentious issue. We build a theoretical
model to approach this problem and to characterize optimal delegation of sentencing to
judges.
In the model, a representative citizen has a preferred sentence that varies by district.
The planner has a statewide preference and would ideally not adapt to local preferences, but
faces a cost of enforcement that depends on the distance between the sentence chosen and the
preferred sentence at the district level.
1
Given this constraint, the planner prefers to move
in the direction of the local norm, but only partially. However, the planner observes neither
case-specific characteristics nor the local norm, and thus delegates the decision to a judge,
who, for each case, can observe the specific details, and can also observe local preferences
1
Such a constraint is also present in papers discussing the interaction between law enforcement and social
norms and customs (Acemoglu and Jackson, 2017 and Hay and Shleifer, 1998), noting that when the two
are too distant enforcement costs increase.
4
for a cost. The judge also wants to adapt to the local norm, because, for instance, she risks
having her decision appealed if she does not. In terms of sentencing preferences, the judge
can either be aligned with the planner with some probability, or prefer tougher sentences.
Within this environment, we build on our empirical findings by comparing social welfare
under three main instruments of judicial oversight: judicial rotation, sentencing guidelines,
and judicial elections. In the model, the planner’s preferred policy partially adapts to the
local norm, taking enforcement costs into account. The judge, under rotation, also partially
adapts, as observed in the data. Our main finding is that whenever the uncertainty on
unobservable crime characteristics is large, rotation will be the socially preferred instrument.
Sentencing guidelines in this case naturally perform poorly since they cannot condition on
ex ante unobservable characteristics. Elections also perform poorly, because, at equilibrium,
all types of judges pool on the same sentence to avoid revealing their type, and accordingly
ignore information on case specific conditions. However, when the uncertainty is small,
sentencing guidelines may be preferred, as they avoid the cost of having a biased judge
decide.
The trade-off between information gathering and excessive adaptation to local interests
is relevant in many contexts beyond the judiciary. In the corporate context, such as con-
sulting, local representatives are often used to gather valuable local information and forge
relationships with particular clients. However increased familiarity can also lead to favoring
client interests over those of the company if relationships grow too close. To counteract this
tendency, employees in this context are often rotated.
Similarly, over time, local politicians acquire intricate knowledge of the problems facing
their jurisdiction but are also susceptible to capture
2
by local power brokers. Elections and
term limits are two approaches to address these problems, with known limitations. Our
investigation of this phenomenon in the judiciary is both interesting in its own right, and
also allows us to empirically describe the general trade off. We exploit the fact that we
observe the full set of criminal decisions made by a judge over time. In general, the full work
output of many other professionals is not readily observable nor quantifiable.
The paper is related to different strands of literature. First, it relates to studies on
judicial decision-making. As mentioned above, most of the recent research in this area aims
at explaining inter-judges variation in criminal justice sentencing. Some papers have shown
how judges’ personal characteristics help explain their decisions. For instance, Lim et al.
(2016) study the influence of judges’ political orientation and demographic characteristics
on criminal sentencing decisions in Texas. The authors document substantial heterogeneity
2
Throughout this paper we use capture to mean influence which may cause judicial sentences to deviate
from the social optimum, but are not necessarily illegal.
5
in sentencing harshness across judges. Cohen and Yang (2018) show that judge’s political
orientation contributes to gender and racial disparities in the US federal courts. Yang (2014)
shows that, when they are not constrained by sentencing guidelines, judges discriminate more
against minorities in criminal justice decisions. Abrams, Bertrand and Mullainathan (2012)
find evidence that heterogeneity in judicial decisions is driven by defendant race. Berdejo and
Chen (2017) show how judges are influenced by the political climate modifying their decisions
when presidential elections are closer. There is also a substantial theoretical literature that
studies how behavioral biases and individual motivations shape judicial decisions in both
tort and criminal law (Gennaioli and Shleifer, 2008 and Bordalo et al. 2015).
The literature on spatial variations is much more limited. Yang (2014) documents spatial
variation in sentencing across the US but cannot distinguish between crime characteristics
and different local practices. Lim et al. (2016), using data from Texas state district courts
where judges overlap in different districts, find results in contrast to ours: judge fixed effect
tend to be more variable than district fixed effects, and judges do not seem to adapt to
behavior of other judges in their district. One important distinction with their work is that,
unlike in our institutional setting, judges sort in particular districts through the electoral
process. Ichino et al. (2003) show that Italian labor judges, take into account local con-
straints and adapt their reintegration decisions to labor market circumstances. One of our
main contributions is also to describe how judges learn about local sentencing norms. To the
best of our knowledge, the economic literature empirically documenting learning in judicial
decision making is scarce, but Coviello et al.(2018) are a notable exception documenting how
judges learn by doing how to treat similar cases.
Beyond judicial decision making, our paper contributes to a recent literature explaining
spatial variations in the provision of public services. In particular, our approach is similar to
two recent papers explaining spatial variation in medical doctors’ diagnostic practices in the
US (Molitor, 2018 and Finkelstein et al. 2016), although in our setting, the usual challenges
to identification due to endogenous sorting are overcome by the institutional design.
Finally, our study contributes to the literature studying the effects of policies aimed at
reducing the capture of public officials. The closest examples are recent papers analyzing the
effects of term limits on politicians behavior (Coviello and Gagliarducci, 2017 and Dal Bo
and Rossi, 2011) and how UK tenured judges respond to political pressure (Blanes i Vidal
and Leaver, 2011).
The paper is organized as follows. In Section 2, we describe the institutional setting.
In Section 3, we present the data and show motivating evidence that spatial variations are
larger than judge-level variation. Section 4 discusses our identification strategy, presents the
evidence on adaptation of judges to local practices, and describes factors that correlate with
6
these local sentencing norms. Finally, in Section 5, we use the empirical findings to build a
theoretical model that characterizes optimal delegation of sentencing to the judge.
2 Institutional Setting
In North Carolina, the Superior Courts have general trial jurisdiction over civil and criminal
matters. The Superior Court system uniquely combines judicial elections, judicial rotation,
and sentencing guidelines. While sentencing guidelines and judicial elections are quite com-
mon, since the majority of States use elections for at least some judicial positions, rotation
is a relatively unique feature of modern state courts. It was, however, more common in the
past, when judges would “ride circuit” to provide justice to more rural areas.
3
In practice, the State is divided into eight divisions, each of which are partitioned into
districts (see figure 1 in the Appendix A).
4
Elections take place separately in each district,
and rotation is then organized at the division level, as we describe below.
Elections
The selection of Superior Court Judges takes place via non-partisan elections. Judges
are elected in a home district for an eight-year term. If a vacancy arises in the middle of a
term, the governor fills the vacancy by appointment, which is effective until the next general
election. There are no term limits; after their term ends, judges can run again for elections.
Although they are required to run again in the same district, they usually end up doing so.
5
Rotation
Superior Court judges elected in each district have the obligation to rotate every six
months, in January and July, within the division where their home district is located. This
rotation rule was established in the North Carolina Constitution of 1868. The primary
motivation for the rotation system is to avoid capture by the local community. As expressed
by a Superior Court Judge in the local press,“because judges are elected, building alliances
through campaigns and asking for campaign contributions largely with people in their local
districts, rotation reduces bias and the perception that contributors have better access or
influence.”
6
A similar motivation is presented in Bobbitt (1948) who also emphasizes the
3
South Carolina and Nevada are the only two other states that still use some sort of judicial rotation.
4
Before 1999, the State was divided into 4 divisions, which was then increased to 8 to reduce the distance
judges had to travel.
5
In our sample, only 2 judges choose to run in a different district.
6
See ”“Riding the circuit: Traveling judges program ensures impartiality, comes with cost”, in the Times
News. He adds, “Rotation mitigates the building of inappropriate or overly-familiar relationships and en-
hances the perception that a truly impartial judge will be presiding. In this way, judicial independence is
enhanced, and an independent judiciary is a principle of the highest value.”
7
benefit of avoiding an excessive connection between judges and particular lawyers.
7
The
additional argument provided in Bobbitt (1948) is that, if the judge is too immersed in the
local community, he might act “on the basis of fixed ideas of his own or on the basis of local
reports rather than on the basis of the evidence before him.”
The Chief Justice of the North Carolina Supreme Court establishes the rotation schedule.
There are no formal rules that the Chief Justice needs to follow, even though there is a
common understanding that judges should return to their home district (the district of their
election) at least once every two years. Judges can present motivated objections to their
particular assignment. This official rotation policy is in fact respected in practice. Using
only observations in our data, we plot in Figure 2 the average probability of moving district
by month, where the district of work in a month is defined as the district where a majority
of cases are decided. Consistent with the rotation policy, a large majority of judges move to
a new district in January and July.
8
Sentencing guidelines
Judges’ decisions for criminal cases in North Carolina are structured by the State’s sen-
tencing laws that went into effect in 1994.
9
To choose a sentence for a felony case, the judge
first has to determine the offense class to which the felony belongs. She then establishes the
prior criminal record of the offender. There are ten offense classes ranked by severity and
six different criminal history levels. Once the offense class and the criminal history of the
offender are determined, the judge must establish which aggravating and mitigating factors
apply to the case. Given these choices, the case falls into a cell in the sentencing grid. Each
cell defines the range of possible minimum sentences. The judge has to select an appropriate
sentence within this minimum sentence range (see Figure 10 in Appendix D). For instance,
an offender convicted for an armed robbery (a class D offense) and having prior record of two
offenses, faces a default sentencing range of minimal sentences (called presumptive sentence
range) ranging from 79 to 97 months. If the judge finds aggravating circumstances, the
range shifts to 97-121, whereas mitigating circumstances move it to 58-78 months. Thus,
depending on the circumstances, and the criminal history, the minimal sentence for an armed
7
”If it be true, or if it be thought by the litigants or by the public, that the judge, whether through
personal friendship with lawyer or with litigant, or through previous experience with lawyer or with litigant,
or through political, business, social or other connections, will be swayed, consciously or unconsciously, by
considerations other than the law and the evidence in the particular case, the prestige and usefulness of the
judge is greatly impaired. It may be well to recall that, in every controversy before the court, each decision
a judge makes is a decision against some person or persons.”
8
The frequency of moves in these two months is larger than 80 percent. There are some deviations,
especially in the one month preceding and following the scheduled moves, presumably to smooth transitions.
9
The sentencing procedure is detailed in the North Carolina ’Structure Sentencing Training and Reference
Manual’ available at https://www.nccourts.gov/assets/documents/publications/sstrainingmanual09.pdf
8
robbery can actually vary between 58 and 121 months.
Finally, the law may stipulate a sentence disposition for each combination of offense class
and prior record level: the sentence may be active (i.e. the offender is under custody) or
suspended. If a defendant receives an active sentence, he must serve the entire minimum sen-
tence defined above. For certain class-record combinations, the judge has to decide whether
the sentence should be active or suspended.
10
3 Data and descriptive statistics
3.1 Data
Sentencing data
Our data comes from the North Carolina Administrative Office of the Courts and includes
the universe of felony cases decided in North Carolina superior courts from 1998 to 2010.
For each decision, we know the week of the decision and the identities of the defendant,
defense attorney, district attorney, and judge. The data also includes the main demographic
characteristics of the defendant as well as their criminal charges. The dataset includes
343,776 sentencing decisions with final disposition dates between 1998 and 2010. It is worth
noting that our main unit of analysis (a case) is defined by aggregating all outstanding
charges for a defendant that are disposed of at one time. The construction of the data is
described in more detail in Appendix D.
In the main analysis, our main sentencing variable is the minimum active sentence chosen
by the judge as in the procedure described in Section 2. If multiple cases are disposed of at
the same time for the same defendant, we consider the maximum of the minimum sentences,
i.e the decision made for the most severe offense. In Appendix B, we consider the robustness
when using as dependent variable the total sentence, both active and non.
Master schedule and electoral data
We also obtained the master schedule produced by the Chief Justice that describes the
assignment of each judge across the districts. We use this data to guarantee that we correctly
identify the judge; we restrict the data to observations where the judge is in the division
recorded in the master schedule. The identity of the judge is then used to match with the
result of judicial elections.
10
As in the rest of the U.S., the vast majority of criminal cases in North Carolina are resolved via plea
bargain, where the sentence is agreed to between the prosecution and defense. Even though the judge is
often not directly involved in plea bargains, she is required to approve the sentence, and thus still exerts a
great deal of influence over the outcome (Abrams and Fackler, 2018 and Mnookin and Kornhauser, 1979).
9
Other sources
We obtained voting data from North Carolina’s Board of Elections. First, results on
voting by district for all presidential elections held during our time period. Second, data
on voting in four referenda held before, during and after our time period, that involved
proposals related to the judicial system (for details on these proposals, see Appendix D). To
compute jail overcrowding, we used the Annual Survey of Jails (ASJ) Data Series, which is
collected by the Bureau of Justice Statistics.
3.2 Descriptive statistics
Table 1 reports descriptive statistics. In panel A we present case and judicial characteristics,
separately for 70 junior judges (those elected or appointed after 1998) and 91 senior judges
(those elected or appointed before 1998). This distinction will prove essential in our empirical
analysis.
More than 80 percent of defendants are young males, (the mean age is 31 years), and
more than half of them have a criminal record. Half of defendants are black, and races other
than black and white represent less than 10 percent of cases. Judges spend around half of
their time in their home district, senior judges have average tenure in office of 14 years, and
junior judges have an average tenure of 7 years. The average sentence is about 500 days,
while the active sentence is about one year.
In Panel B, we present average characteristics of the 50 districts in North Carolina. There
are, on average, 11 judges and 23 district attorneys who take at least one decision in the
district. Naturally, the pool of defense attorneys is larger, with an average of 61 per district.
The rest of panel B presents summary statistics of political preferences and population
characteristics that we’ll use in our analysis of the determinants of the sentencing norms.
North Carolina tends to vote more Republican.
3.3 Spatial variation in sentencing decisions
The premise of our study is that spatial variation in sentencing exists. We document in
Figure 3 that this is indeed the case. Sentences vary widely not only by district but also by
crime type. For instance, sentences for violent crimes range from an average of around 400
days in lenient districts to approximately 800 for the harshest.
11
11
Forsyth County, which contains Winston-Salem, is the outlier in all panels of Figure 3. In private
conversation, a local official suggested that local prosecutors have a norm of aggressive prosecution, especially
among most serious crimes.
10
Several factors could cause this spatial variation. First, the distribution of crimes commit-
ted or the unobserved crime characteristics within a crime type could differ across districts.
Second, the pool of judges differs by division and could vary in harshness.
12
We can account
for these variations using the following specification:
S
ijdc
= βX
i
+ δ
j
+ γ
d
+ ν
c
+ (1)
where S
ijdc
is the sentence decided by a judge j , in district d for given crime c and i a case
indicator, δ
j
, γ
d
and ν
c
, respectively judge, district and crime fixed effects and X
i
are case
characteristics, such as age, gender and race of the defendant.
In the left panel of Figure 4, we present the distribution of district fixed effect γ
d
where
equation (1) is estimated using only the decisions made by senior judges, who have acquired
good knowledge of each district.
13
After controlling for case characteristics and judge fixed
effects, large variations across districts remain. Excluding two outliers, fixed effects by
districts still vary within a 200 days band.
These results also allow us to examine how much of the deviation in sentencing from a
crime level average can be explained by judicial versus geographic variation. As highlighted
in the introduction, distinguishing spatial and judicial fixed effects relies on our institutional
setting, where judges rotate across districts. In the right panel of Figure 4, we present
the judicial fixed effects for the senior judges who take more than 500 decisions and visit
at least two districts during our observation period. Variability in district fixed effects is
approximately twice as large as variability in judicial fixed effects. This finding suggests that
spatial variation in sentencing is indeed an important feature per se of the criminal justice
system over and beyond the variation explained by judges’ fixed effects. Understanding
spatial variation is thus crucial for describing the system’s design.
Even though in estimating equation (1), we control for crime fixed effects (with 615
categories) and for observable characteristics of the defendant, the possibility remains that
variations across districts is explained by systematic geographical differences in crime char-
acteristics, unobserved by the econometrician but observable to the judge.
14
The local fixed
effect could therefore include two components: first, differences in average unobservable crime
characteristics, second, different behavior across space in sentencing for the same crime char-
acteristics, which we call the local sentencing norm. Our identification strategy, presented
12
Even though we have rotation and thus constraints on sorting by judges, judges can still sort across
divisions and might be able to affect the time they spend in each district.
13
The separate use of senior judges will be described in more detail when we discuss identification in
Section 4.
14
This explanation seems unlikely, since observable characteristics such as race, gender and sex do not
reduce the geographical variance much.
11
in the next section, allows us to both show the existence of this sentencing norm and the
gradual adaptation of judges to it.
4 Adaptation to local sentencing norms
4.1 Identification
Two features of the institutional setting in North Carolina provide important benefits for
identification. First, we have access to a relatively large dataset with a significant number
of judges arriving in new jurisdictions, at different points in their career, all under the same
State law. Second, we can distinguish decisions made by judges in their home district versus
other decisions made elsewhere, which helps us to distinguish adaptation to a norm from
learning about the law.
We divide our sample of judges in two groups: junior and senior judges. For junior
judges, the entire sentencing history is observed, including the number of decisions made in
each district. We use this set of judges to construct our working sample. We use the senior
judges to compute the local effect LocSent
dc
, defined as the average sentence given by senior
judges in district d for a given crime category c.
15
The idea is that, if the local effect includes
a local sentencing norm component, these judges will have already spent more time in the
district, allowing them to have better knowledge of this norm. We show in Section 4.2 that
our main results are robust to changes in the definition of LocSent
dc
, in particular using
only senior judges elected in the district to define the local effect.
16
In our main specification, we examine how sentences evolve compared to LocSent
dc
:
|S
ijdc
− LocSent
dc
| = α
1
Order
jd
+ α
2
Order
jd
∗ 1
ElecDistr
(2)
+ α
3
1
ElecDistr
+ α
4
Order
j
+ βX
i
+
where S
ijdc
is the sentence decided by a judge j , in district d for given crime c and i a
case indicator. Order
j
counts the number of cases treated by judge j up to date t (judge
experience) and Order
jd
counts the number of cases treated by judge j up to date t in
the district d (judge local experience).
17
Finally 1
ElecDistr
measures whether the judge is
15
An alternative would be to extract this variable as a crime * district fixed effect in a modified version
of equation (1). However, this would imply extracting 615*50 fixed effects and most of them would be very
imprecisely estimated.
16
If the sample of judges used to define LocSent
dc
have not fully learned the norm, this would create a
noisy measure of the local effect. It is reasonable to think that elected judges would have better knowledge
of the local effect, hence the robustness check.
17
We will also run quantile regressions of equation (2) where the dependent variable is taken without
12
sentencing in her home district.
The main parameter of interest is α
1
, which measures the impact of the number of
decisions made in a district on the distance to the local effect, and provides two essential
pieces of information. First, under the assumption that the unobservable characteristics of
the case (to the econometrician) are always observed by the judge, a value of α
1
significantly
different from zero (i.e an evolution in time of the judge’s sentencing behavior) indicates
the existence of a local sentencing norm. Indeed, if variations in sentencing norms were
only driven by variations in unobservable case characteristics, since the judge observes them,
there would be no evolution in time and α
1
would be zero. A negative value of α
1
indicates
gradual adaptation to the local sentencing norm as a function of number of cases seen in a
district.
The first potential identification challenge is related to assignment of cases to judges. If
the judge is assigned cases more distant from the average case when she first arrives in a
district, α
1
would potentially capture this assignment dimension rather than adaptation to
the local sentencing norm. One way to address this concern is to examine how observable
characteristics vary as a function of the order of decisions in a district. We examine this in
Table 2 where we estimate equation (2) using different case characteristics as the dependent
variables. In panel A, we report the coefficients when we use the defendant characteristics
as dependent variables, and in panel B, the crime categories. Reassuringly, we find that the
order of the sentences made by a judge in a given district does not have a significant impact
on the type of case or the characteristics of the defendant.
The second potential challenge is to differentiate between adapting to the local sentencing
norm and merely learning about the law. This concern is alleviated by the fact that different
judges arrive in the same district at different points in their careers, due to the rotation
system. In the specification, we separately include judicial experience in the district and
overall experience of the judge, with the latter effect being measured by parameter α
4
.
We also allow for the fact that judges likely already know the norms in their home district
by interacting the variable Order
jd
with a home district dummy. Under the reasonable
assumption that a judge does not learn more quickly about the law in her home district
compared to outside, if we find α
2
to be significantly different from zero, this would suggest
that we are indeed measuring adaptation. In particular, we would expect to find α
2
positive,
and roughly of the same absolute magnitude as α
1
, consistent with the idea that the judge
already knows the local sentencing norm in her home district.
The third potential challenge is the risk that judges can affect the time they spend in
the absolute value. If judges come closer to the local effect over time, we expect the coefficients for lower
quantiles (resp. higher) to be positive (resp. negative).
13
a district. As explained in Section 2, the master schedule is controlled by the State Chief
Justice. While we have no evidence of favoritism in assignment, judges are allowed to make
scheduling requests. If endogenous sorting does occur, finding α
1
to be significant (i.e. the
fact a judge behaves differently for early decisions versus late) could capture the fact she is
able to spend more time in districts where she is closer to the norm. To address this concern,
we use a fully balanced dataset. We homogenize the dataset by focusing on the first 400
decisions made by a judge in a district where the judge takes at least 400 decisions.
18
This
restriction leaves us with a dataset containing 48 different judges.
4.2 Results: adaptation to the local sentencing norm
Before presenting the results of the estimation of equation (2), we first provide graphical
evidence of adaptation over time to the local sentencing norm. Figure 5 presents the results
of an OLS regression of the following model, separately for different quantiles:
S
ijdc
− LocSent
dc
=
7
P
k=1
α
k
∗ 1
order=k
∗ 1
ElecDistr
+
7
P
k=1
β
k
∗ 1
order=k
∗ (1 − 1
ElecDistr
).
The figure shows how sentences evolve with the number of decisions made in a district.
We separately present the results for decisions made in the judge’s home district (right
panel) and in other districts (left panel). There is no clear evolution of sentence length for
judges hearing cases in their home district (right panel). This contrasts markedly with what
happens when a judge is outside her home district (left panel). In this case, the left panel of
Figure 5 shows that the gap between the 90th and the 10th percentile of the distribution of
distance between the sentence and the local target (S
ijdtc
− LocSent
dc
) reduces over time,
19
while the median of the distribution remains very stable.
20
These initial graphical results not only show the existence of and adaptation to a lo-
cal sentencing norm, but the contrast between the left and right panels also validates our
identification strategy. Those in their home district a priori know the local target better as
they already had to campaign in the district, and presumably live there. The fact that they
do not change their sentencing behavior over time confirms that we observe adaptation to
a local norm rather than learning about the law, which should occur uniformly across all
districts that are subject to the same legal rules.
18
We also perform robustness checks in Section 4.2, where we use thresholds of 500 and 300 decisions.
19
Note that the same judge can take decisions in different parts of the distribution.
20
In Figure 8 in Appendix B we reports the same results described above when using as a dependent
variable the total sentence rather than only the active sentence. The patterns described above remain
substantially unchanged.
14
The graphical results are confirmed in Table 3 where we estimate variations of equation
(2) using both OLS and quantile regressions, with standard errors clustered at the judge
level. Column (1) estimates the equation, removing the absolute value from the dependent
variable. For this specification we find no effect of the number of cases in a district on the
distance between the sentence and the local norm. This can be seen as indirect evidence of
the absence of strategic allocation of cases over time. If more (less) severe cases with respect
to the local sentencing norm were systematically assigned to judges with less experience in a
district we would expect a positive (negative) and significant effect of the order of decisions
in the district. Column (7) estimates equation (2), where the dependent variable is the
absolute value of the distance between the sentence and the norm, and shows that α
1
is
indeed negative and significantly different from zero. The magnitude of the effect is large:
the value of α
1
implies that, comparing a judge’s 100th case in a district to the first, the
sentence will be 24 days closer to the local norm, when she is outside her home district.
Parameter α
2
is positive and slightly smaller in absolute value than α
1
. The sum of the two
parameters is not statistically different from zero, confirming that this adaptation is absent
for decisions made in the judge’s home district.
Columns (2) through (6) present the results of quantile regressions with the difference
between the sentence and local norm as the dependent variable. The overall results point to
convergence to the local norm throughout the judicial sentencing distribution. The coefficient
of interest, α
1
, is positive for the reported percentiles below 50 and negative above. For each
quantile, α
1
and α
2
are of opposite signs. Magnitudes are particularly large for the extreme
quantiles: for the 90th percentile, comparing a judge’s 100th case in a district to the first,
the sentence is 66 days closer to the local norm. Figure 6 presents a telling graphical
representation of these results, where we plot the coefficients for different quantiles of the
distribution. The coefficient is positive and significant for the lower quantiles and negative
and significant for the higher quantiles.
21
We consider a number of robustness checks of our main results in Table 4. First, we
examine whether our results are sensitive to the definition of the local sentencing norm. In
the main specification the variable LocSent
dc
was computed using only decisions made by
senior judges, with the idea that they already had time to adapt to the norm.
22
We show
in column (1) that estimating equation (2) when using an even more conservative definition
of LocSent
dc
, restricting ourselves to senior judges in their home district, yields comparable
estimates of α
1
and α
2
. In column (2), we use a more liberal definition of the norm, that
21
Figure 9 in Appendix B shows equivalent results when including suspended sentences.
22
The sample size drops with this specification because for some district*crime categories we do not have
any decisions made by senior elected judges.
15
incorporates all decisions by junior judges after their 400th in a district (these decisions are
used to compute LocSent
dc
, but not to estimate the model). We find that the results still
hold with this alternative definition of the local norm.
In order to rule out concerns of potential correlation between time spent in a district
and judicial preferences, in our main specification we use a balanced sample of decisions,
including just the first 400 decisions made by a judge in a district where the judge makes
at least 400 decisions total. In column (3) we conduct the same exercise using the first 300
decisions and the first 500 in column (4). In both cases, the magnitude of the effects are
very similar to those reported in Table 3 column (6). In column (5), we estimate (2) adding
defendant races, sex, and age; column (6) includes judge fixed effects. The results are robust
to the addition of these extra controls. This comes as no surprise since by focusing on a
balanced sample, we rule out composition effects linked to the order of decision. Finally,
in column (7) we use an alternative definition of sentence that adds suspended to active
sentences. Again, the results hold substantially unchanged.
What determines the starting point of the learning process when a judge first arrives in a
district? Figure 7 shows that when the sentencing norm in the judge’s home district is higher
(lower) than in the current district, the difference between the judge’s sentences and the local
average tends to be positive (negative) at the beginning of her tenure in the district. This
effect disappears once the judge has examined a sufficient number of cases in the district.
The key role of the home district could just be a proxy for the judge’s preferences, i.e. she
gets elected in a district with similar sentencing preferences.
Our main results show the existence of a local norm and gradual adaptation of judges to
it. Our preferred interpretation of the adaptation process is that the judges want to adapt
to the local norm, are initially unaware of it and gradually learn about it, case after case.
We cannot rule out, however, that this process corresponds more to judges initially resisting
the local norm and gradually giving in, for instance because local actors (e.g. attorneys,
prosecutors) learn how to maneuver the judge. For exposition purposes, we will refer to the
adaptation process as learning about the norm, but return to this point when discussing
welfare in Section 5.
4.3 Local sentencing norm
The previous sections established the existence of a local sentencing practice and gradual
adaptation of judges to this sentencing norm. Although we cannot establish the determinants
of the sentencing norm in a causal manner, in this section we explore correlates of variables
meant to capture the three main possible explanations for this norm. As mentioned in the
16
introduction, the sentencing practices could reflect: local norms of behavior, local district
resource constraints (on police or prisons) or, finally, a court-specific culture.
We measure the first dimension, local norms of behavior, in two ways. The idea of these
local norms is that some illegal behavior might be more or less socially acceptable to the
community.
23
Our first approach is thus to construct the district level prevalence of certain
crimes, measured as the proportion of crimes in that district committed in a particular
crime category.
24
Our second approach to measure local norms of behavior is to measure
local political preferences. We use the results of the three presidential elections run during
our time period (2000, 2004, 2008) as well as local referenda on judicial questions, described
in Appendix D.
To measure the second possible determinant of the sentencing norm, we use one partic-
ular constraint, prison overcrowding. However, overcrowding is not a major issue in North
Carolina, where only 20 of the districts have an issue of overcrowding. Even for these dis-
tricts, the ratio of convicts to beds in close to 1, as described in Table 1. We might therefore
be underestimating the impact of local constraints.
Measuring local culture at the courtroom level is more challenging since it is, by definition,
the aggregation of characteristics of the permanent staff (attorneys and clerks) working
in the district. Rather than attempting to measure it directly, we examine whether the
concentration in the attorney or the district attorney market has an influence on average
sentences. Specifically, we construct an Herfindhal Index of attorney or district attorney
concentration at the district level. The index for attorneys is defined as
HI
a
d
=
X
a∈A
d
(N
ad
/N
d
)
2
where A
d
is the set of attorneys working predominantly in district d, N
ad
is the number of
cases in d where a was the attorney and N
d
is the total number of cases in district d. This
index is calculated by year and averaged over the years in our sample. The Herfindhal Index
for district attorneys is calculated in a similar fashion. If the local sentencing norm is defined
by attorneys, a higher Herfindhal Index for attorneys should lead to lower sentences, since
attorneys should presumably try to obtain lower sentences for their clients.
Results are presented in Table 5 where we regress the local sentencing norm LocSent
dc
,
on the determinants described above, controlling throughout for crime fixed effects. The
first result is that an increase in crime c’s prevalence in district d has a significant negative
23
Research in criminology has presented evidence in line with this hypothesis (Eisenstein et al. 1988, and
Ulmer and Johnson 2004), the idea is also supported by various discussions we had with defense attorneys
and prosecutors.
24
Specifically the variable is constructed as 100 ∗ (cases
dc
/cases
d
) − 100 ∗ (cases
c
/cases).
17
effect on the sentencing norm, both controlling for district level controls (columns (3) to (5))
or including district fixed effects (column (2)). A 1 percent increase in the crime prevalence
decreases the average sentence in the district by 4 days for that crime category. Note that
the alternative interpretation is that the causality goes in the opposite direction: lower
sentences for certain crimes in a district induce rational criminals to adapt to it. This
explanation appears unlikely, since it would require very detailed knowledge of variation in
the way judges behave locally.
25
In contrast, neither local courtroom culture, captured by the concentration measures
(introduced in column (3)), nor local constraints (introduced in column (4)) have an impact.
In column (5) we introduce the political variables. The overall political leaning does not play
a role. There seems to be a systematic link with the results of the 2004 and 2014 referenda.
Note that the link is stronger when we consider in Table 6 in Appendix B, the sentencing norm
calculated using total sentences rather than active sentences. Interestingly, the referendum
of 2004 proposed to make judges accountable by fixing their first term at 2 years, while
the 2014 referendum transferred power to the judge, by allowing criminal defendants to be
sentenced by a judge rather than by a jury. The results in column (5) suggest that in districts
where the judges are tougher, the citizens are more willing to transfer prerogatives to them.
This is consistent with the idea that on average, citizen desire higher sentences than judges,
which will relate to the next section’s welfare discussion.
5 Judicial delegation
We have presented in the previous sections empirical evidence on several consequences of
delegating the sentencing decision to the judge. First, as already shown in the literature,
judges adapt to characteristics of the crime and criminal. Second, as shown in the main
contribution of this paper, delegation allows judges to learn and adapt to the local sentencing
norms. The results presented in Section 4.3 provide suggestive evidence that this sentencing
norm is linked to local preferences.
This raises the question of why a social planner would want to constrain the freedom of
the judge, since setting sentencing guidelines or rotation policies would decrease the speed of
learning and the ability to adapt to local conditions. The answer to this question critically
depends on whether adaptation to local norms is socially desirable. It is natural to think
that the planner wants to stick to legal principles that are not abandoned in the face of
local customs. Moreover, with delegation, the judge might end up favoring certain groups,
25
Papers that find incentive effects of sentences (Drago et al. 2009 for instance), use large and salient legal
changes in expected sentences, while our variations are small and not codified.
18
consciously or unconsciously, when spending too much time in a district, a concern at the
root of the rotation policy.
26
However, adaptation can also be socially desirable if it allows
easier enforcement of the law in an environment of higher social acceptance (as in Acemoglu
and Jackson 2017).
This discussion highlights the trade-offs involved in delegating the decision to the judge
and offers some rationale for planners constraining judicial autonomy with the use of tools
such as sentencing guidelines, rotation, or elections. Building on our empirical findings, we
thus consider the question of the optimal constraint on delegation in a model that captures
the aforementioned main ideas.
5.1 A model of judicial delegation
Each period, a judge examines a case, indexed by time t. All players discount the future
at the same rate δ. Each case has its own characteristics (of the crime and the criminal
as shown in the data), summarized by the case severity denoted s
t
distributed according to
distribution f with mean 0.
The citizens
There is a representative citizen in each district d. The citizen has the following period
utility when the sentence is set at S
t
for case t:
u
d
c
= −(S
t
− (θ + β
d
+ s
t
))
2
where β
d
is the local norm in district d, i.i.d drawn from distribution g with mean 0; θ is the
target sentence in the state for a case of average severity s
t
= 0. Thus, the desired sentence
of the citizen in district d, for a case of severity s
t
, is S
d
c
= θ + β
d
+ s
t
. This representation
fits the “penalty should fit the crime” heuristic. Aggravating conditions (higher s
t
) and
sentencing in tougher districts (higher β
d
) lead citizens to demand higher sentences.
The planner
We consider a planner whose utility from sentencing is given by −(S
t
− (θ + s
t
))
2
, i.e
who, for a given s
t
, wants sentence θ + s
t
. In other words, the planner would like to stick to
the basic principles of the law, embedded in θ, and not bend to local preferences. However,
26
This is expressed in Bobbitt (1948): “If it be true, or if it be thought by the litigants or by the public, that
the judge, whether through personal friendship with lawyer or with litigant, or through previous experience
with lawyer or with litigant, or through political, business, social or other connections, will be swayed,
consciously or unconsciously, by considerations other than the law and the evidence in the particular case,
the prestige and usefulness of the judge is greatly impaired.”
19
as in Acemoglu and Jackson (2017), when sentences are not aligned with local norms, this
creates an additional cost of enforcement that the planner has to take into account. We
introduce this idea of difficulties in enforcement in a reduced form way, assuming that there
is a cost α(S
t
− S
d
c
)
2
where S
d
c
, defined above, is the desired sentence of citizens in district
d. Parameter α measures the strength of this enforcement motive. Thus the utility of the
planner is:
u
d
p
= −(S
t
− (θ + s
t
))
2
− α(S
t
− S
d
c
)
2
Given these preferences, if the planner observed s
t
and β
d
, she would optimally choose
sentence
S
d
p
= θ + s
t
+
α
1 + α
β
d
when α = 0, the planner’s desired sentence is θ + s
t
and naturally, as α becomes large, the
desired sentence converges to the representative citizen’s preferred sentence S
d
c
= θ + β
d
+ s
t
.
The judge
The judge, when the sentence deviates from the local norm, has to pay a cost γ(S
t
−S
d
c
)
2
,
corresponding either to the cost of disagreement with the local lawyers or the increased risk
of having the decision appealed. It could also correspond to the risk of capture, which
initially motivated the rotation policy, which would imply that spending time in a district
and developing friendship and connections renders deviations from desired sentences more
costly.
In terms of absolute preferences for sentencing, we assume the judge could be of two
types. The regular type (with prevalence p) has the same preferences as the planner. The
tough type (prevalence 1 - p) is biased in favor of longer sentences, with optimal sentence
net of argumentation cost θ + s
t
+ ζ.
The per period utility of a judge of type j, who also receives a wage w per case indepen-
dently of the sentence chosen, is thus given by:
u
d
j
= −(S
t
− (θ + s
t
+ ζ
j
))
2
− γ(S
t
− S
d
c
)
2
+ w
with ζ
j
= 0 for the normal judge and ζ
j
= ζ for the tough judge.
We assume that γ > α; in other words, the judge’s cost of deviating from the norm is
higher than the planner’s concern for enforcement costs.
Information
For each case t, the judge observes the severity s
t
at no cost simply by listening to the
case specifics. Consistent with the empirical evidence, we assume that if the judge is elected
20
in a district she observes β
d
at no cost. However, if she is not in her home district, the judge
needs to pay a fixed cost C to learn β
d
.
The planner, however, is unable to observe either β
d
or s
t
, and needs to delegate the
decision to the judge. She has three instruments at her disposal to discipline the judge:
• Sentencing guidelines: fixed sentence S
g
independent of the district and the unob-
servable crime characteristics.
• Rotation: the judge is required to move district every N cases.
• Elections: at the end of each period, there is a probability e that an election is run
between the incumbent judge and a challenger. When an election is run, the citizen
can observe the last decision made by the judge. We also assume that the wage w that
the judge obtains per case (or the probability of an election) is high enough that for
any sentence S and any case specific factors s
t
, the judge prefers choosing sentence S to
guarantee reelection in case of a next period election rather than taking an otherwise
preferable decision that guarantees electoral defeat.
5.2 Results
The goal of this model is to understand how the planner will choose among the different
tools to delegate the decision to the judge. The essence of the comparison is given in the
following Proposition:
Proposition 1 There exists
˜
V and
¯
V such that:
1. if the variance of s
t
, V (s
t
) ≥
˜
V , rotation is the socially preferable instrument,
2. if V (s
t
) = 0 and V (β
d
) ≤
¯
V , sentencing guidelines are socially preferable to rotation.
The key tradeoff can be understood by describing the main instruments. All details can
be found in the Appendix.
If the planner chooses sentencing guidelines, she sets the sentence optimally, without
knowing s
t
or β
d
. Given this uncertainty and the fact these variables have zero mean, she
sets S
g
= θ, and the realized welfare is
W
g
= −(1 + α)V (s
t
) − αV (β
d
) ,
welfare that decreases when the uncertainty on s
t
and β
d
increases. W
g
is particularly
sensitive to uncertainty on s
t
, since better knowledge of local conditions allows the planner
21
to better tailor the sentence to her preferences, but also better to minimize the cost of
enforcement.
If the planner chooses rotation, she will choose the number of cases per rotation N
sufficiently large for the judge to find it profitable to invest in learning the local norm. The
benefit of rotation is that the judge takes into account the severity s
t
and the norm β
d
. The
costs are twofold. First, the judge, will excessively adapt to the norm because she bears
more cost than the planner from deviation. Second, with probability 1−p the judge is tough
and has preferences that are not aligned with the planner.
The last instrument, elections, has very particular features. The equilibrium with elec-
tions is characterized by pooling on a single sentence that ignores the case specifics s
t
(as in
the comparison between elected and appointed civil servants, Maskin and Tirole, 2004). In-
deed, a separating equilibrium cannot exist, since, in such equilibria, higher sentences would
always be more likely to be chosen by a tough judge. Depending on whether the districts
prefer a tough or a normal judge, there would always be a type willing to mimic the other.
Thus, under an election system, sentences do not depend on s
t
.
Proposition 1.1 then follows naturally. Since conditions s
t
are ignored under both elec-
tions and sentencing guidelines, rotation dominates when the variance of s
t
is large, i.e when
there is a large variance in case details. Proposition 1.2 follows from the fact that when both
the variance of s
t
and β
d
is small, the benefits of rotation disappear and there only remains
the cost that the judge might be tough and have preferences not aligned with the planner.
In the model, as expressed in Proposition 1 a single instrument is socially preferred,
and there is no room for complementarities between instruments. In practice, the judicial
system in North Carolina uses a mix of all three instruments. However, the sentencing
guidelines implemented in the state are actually fairly large intervals of acceptable sentences.
In this sense, they could help achieve a better balance between the different dimensions of
the trade-off. Indeed, under rotation, for relatively more socially acceptable crimes (high
β
d
), the judge adapts by giving lower sentences. This may be socially beneficial because
enforcement costs can arise when the sentences are far from the social norms. However,
there is also the risk that the judge adapts excessively to these local conditions. In this sense,
loose sentencing guidelines, as those implemented in North Carolina where a relatively large
interval of sentences is authorized for each crime, might be optimal to generate constrained
adaptation.
22
6 Conclusion
In this paper, we have provided evidence on the existence of local sentencing norms and
shown that judges outside their home district gradually converge to them. The results are
robust to alternative definitions of the norm and different ways of defining the working
sample. We also document factors affecting the speed of the convergence process.
When we discuss welfare implications in the model of Section 5, a key element is whether
this convergence is considered to be socially desirable or not. In the model, some adaptation
to local norms is desirable, because sentences that are in conflict with local customs are
harder to enforce. This is consistent with the empirical finding that districts with more
prevalent crime, which we interpret as being more socially acceptable, have lower sentencing
norms.
23
References
Abrams, D.S., Bertrand, M. and Mullainathan, S.. 2012. “Do Judges Vary in
Their Treatment of Race?”Journal of Legal Studies, 41(2): 347-383.
Abrams, D. and Fackler, R.. 2018. “To Plea or Not to Plea: Evidence from North
Carolina.”Working Paper, University of Pennsylvania.
Acemoglu, D. and Jackson, M.O.. 2017. “Social Norms and the Enforcement of
Laws.”Journal of the European Economic Association, 15(2): 245-295.
Blanes i Vidal, J. and Leaver, C.. 2011. “Are tenured judges insulated from political
pressure?”Journal of Public Economics, 95(7-8): 570-586.
Bobbitt, W.H.. 1948. “Rotation of Superior Court Judges.” North Carolina Law
Review, 26: 335-349.
Bordalo, P., Gennaioli, N. and Shleifer, A.. 2015. “Salience Theory of Judicial
Decisions.” Journal of Legal Studies, 44(1): 7-33.
Besley, T. and Payne, A.A.. 2013. “Implementation of Anti-Discrimination Policy:
Does Judicial Selection Matter.” American Law and Economics Review, 15(1): 212-251.
Berdejo, C. and Chen, D.. 2017. “Electoral Cycles Among U.S. Courts of Appeals
Judges.” Journal of Law and Economics, 60(3): 479-496
Drago, F., Galbiati, R. and Vertova, P.. 2009. “The Deterrent Effect of Prison
Evidence from a Natural Experiment.” Journal of Political Economy, 117(2): 257-280.
Finkelstein, E., Gentzkow, M. and Williams, H.. 2016. “Sources of Geographic
Variation in Health Care: Evidence from Patient Migration.” Quarterly Journal of Eco-
nomics, 131(4): 1681-1726.
Cohen, A. and Yang, C.. 2018. “Judicial Politics and Sentencing Decisions.” Ameri-
can Economic Journal: Economic Policy, Forthcoming.
Coviello, D. and Gagliarducci, S.. 2017. “Tenure in Office and Public Procurement.”
American Economic Journal: Economic Policy, 9(3): 59-105.
Coviello, D., Ichino, A. and Persico, N.. 2018. “Measuring the Gains from Labor
Specialization: Theory and Evidence.” mimeo, Northwestern University.
Dal Bo, E. and Rossi, M.. 2011. “Term Length and the Effort of Politicians.” Review
of Economic Studies, 78(4): 59-105.
Eisenstein, J., Flemming, R.B., Nardulli, P.F.. 1988. “The Contours of Justice:
Communities and Their Courts.” Boston: Scott, Foresman.
Gennaioli, N. and Shleifer, A.. 2008. “Judicial Fact Discretion.” Journal of Legal
Studies, 88(2):1-35.
Hay, J.R. and Shleifer, A.. 1998. “Private Enforcement of Public Laws: A Theory of
Legal Reform.” American Economic Review, 88(2): 398-403.
24
Ichino, A., Polo, M. and Rettore, E.. 2003. “Are Judges Biased by Labor Market
Conditions?” European Economic Review, 47(5): 913-944.
Lim, C.. 2013. “Preferences and Incentives of Appointed and Elected Public Officials.”
American Economic Review, 103 (4):1360-1397.
Lim, C., Silveira, B. and Snyder, J.. 2016. “Do Judges’ Characteristics Matter?
Ethnicity, Gender, and Partisanship in Texas State Trial Courts.” American Law and Eco-
nomics Review, 18 (2): 302-357.
Maskin, E. and Tirole, J.. 2004. “The Politician and the Judge: Accountability in
Government.” The American Economic Review, 94 (4): 1034-1054
Mnookin, R.H. and Kornhauser, L.. 1979 “Bargaining in the Shadow of the Law:
The Case of Divorce” The Yale Law Journal, 88( 5): 950-997
Molitor, D.. 2018. “The Evolution of Physician Practice Styles: Evidence from Cardi-
ologist Migration.” American Economic Journal: Economic Policy, 10(1): 326-356.
Schanzenbach, M. M. and Tiller, E.H.. 2007. “Strategic Judging under the U.S.
Sentencing Guidelines: Positive Political Theory and Evidence.” Journal of Law, Economics
and Organization, 23(1): 24-56.
Ulmer J.T. and Johnson, B.. 2004. “Sentencing in Context: A Multilevel Analysis.”
Criminology, 42 (1):137-177.
Yang, C.S. 2014. “Have Interjudge Sentencing Disparities Increased in an Advisory
Guidelines Regime? Evidence from Booker.” NYU Law Review, 89(4): 1268-1342.
25
Appendix A
26
Figure 1: Judicial Map of North Carolina
27
Figure 2:
0 0.2 0.4 0.6 0.8 1
Probability
1 2 3 4 5 6 7 8 9 10 11 12
Month of Decision
Note: Probability of moving per month, using data from the master schedule. A move is defined as a change in the district
where the majority of decisions are made by a judge in a month.
28
Figure 3: Spatial Variation in Sentencing
100 200 300 400 500 600
Mean sentence (judges starting before 1998)
0 10 20 30 40 50
Distict ID
All crimes
0 200 400 600
Mean sentence (judges starting before 1998)
0 10 20 30 40 50
Distict ID
Drug crimes
300 400 500 600 700 800
Mean sentence (judges starting before 1998)
0 10 20 30 40 50
Distict ID
Violent crimes
0 200 400 600 800
Mean sentence (judges starting before 1998)
0 10 20 30 40 50
Distict ID
Property crimes
Note: Each figure shows the mean sentence by district among senior judges (those appointed before 1998) for: all crimes (top-left panel), drugs (top right), violent crimes
(bottom left), and property crimes (bottom right)
29
Figure 4: Comparison of Spatial and Judicial fixed effects
-100 0 100 200 300
District FE
0 10 20 30 40 50
Distict ID
-100 -50 0 50 100
Judge FE
0 10 20 30 40
Judge ID
Note: The left panel displays the distribution of district fixed effects. The right panel displays judicial fixed effects from equation (1) estimated using only senior judges.
30
Figure 5: District Experience and Sentencing
-200 0 200 400
Coefficients
1-50 51-100 101-150 151-200 201-250 251-300 301-351
Number of Decisions in District
10th percentile Median
90th percentile
Non-home Districts
-200 -100 0 100 200
Coefficients
1-50 51-100 101-150 151-200 201-250 251-300 301-351
Number of Decisions in District
10th percentile Median
90th percentile
Home District
Note: Each panel presents results of three separate quantile regressions of the specification in Equation 2. In each case, the dependent variable is a measure of the relative
severity of sentencing, and the independent variable is the number of cases in a district. The left panel presents results for judges outside their home district; the left for judges
in their home district. Standard errors are clustered at the judge level and error bars indicate the 95% confidence interval.
31
Figure 6: District Experience and Sentencing, by quantile
-3 -2 -1 0 1
Coefficient
0 20 40 60 80 100
Quantile
Non-home Districts Home District
Note: This figure presents the results of 19 different quantile regressions (from q5 on the left to q95 on the right). Each quantile
regression measures the effect of the order of cases in a district on the quantile of the distance between sentences and the local
sentencing norm. The dashed line presents estimates for judges in their home district and the solid line for judges in non-home
districts. Standard errors are clustered at the judge level and error bars indicate the 95% confidence interval.
Figure 7: Impact of the Home District Local Norm
-100 -50 0 50
Distance Between Sentence and Local Norm
1-100 101-200 201-300 301-400
Number of Decisions in District
Home – Local above median Home – Local below median
Note: Each bar presents the average distance between the assigned sentence and the local sentencing norm for judges not
elected in the district. Judges with stricter (i.e. higher) home district local norms are represented by the blue bars on the left,
and those with more lenient home districts are represented by the red bars on the right. Decisions are organized in groups of
100.
32
Table 1: Descriptive Statistics
Panel A: Case Characteristics
Junior Senior All
mean sd mean sd diff t
Defendant characteristics
female 0.16 0.37 0.15 0.36 0.01 (7.73)
black 0.51 0.50 0.57 0.50 -0.06 (-32.24)
white 0.42 0.49 0.36 0.48 0.06 (35.19)
other race 0.08 0.05 0.11 0.08 -0.02 (-36.34)
minor 0.04 0.19 0.04 0.19 -0.00 (-0.52)
age 31.07 10.49 30.90 10.34 0.18 (4.85)
criminal history 1.68 1.15 1.65 1.13 0.03 (7.16)
first offense 0.46 0.50 0.45 0.50 0.01 (3.82)
Judge characteristics
tenure 6.56 5.19 14.23 6.08 -7.67 (-378.90)
Judge in district of election 0.47 0.50 0.55 0.50 -0.08 (-47.79)
Outcomes
Sentence (days) 501.85 1450.18 524.52 1569.44 -22.67 (-4.32)
active sentence (days) 360.45 1469.63 386.22 1586.64 -25.77 (-4.85)
Observations 198109 145667 343776
Panel B: District Characteristics
mean sd min max
Judicial characteristics
n. judge 11.5 3.97 6 25
n. district attorney 21.66 14.96 7 96
n. attorney 61.08 39.45 13 192
prison overcrowding .99 .13 .76 1.48
population characteristics
share Democrat in presidential .46 .07 .36 .59
unemployment rate .04 .006 .03 .06
prop black .22 .07 .07 .32
prop white .69 .12 .39 .88
prop female .52 .01 .47 .55
Observations 50 50 50 50
33
Table 2: Balancing tests
Panel A: Defendant Characteristics
(1) (2) (3) (4) (5) (6)
First offense Female Black Age Minor Senior
Nb case in district -0.0000756 -0.000000596 0.000192
∗
0.000348 0.000000504 0.00000221
(0.0000555) (0.0000344) (0.000105) (0.00110) (0.0000227) (0.00000383)
Nb case in district * Elected 0.0000200 0.0000185 0.000000574 0.000316 -0.00000575 -0.00000338
(0.0000693) (0.0000388) (0.0000821) (0.00133) (0.0000247) (0.00000518)
Elected 0.00294 0.00482 -0.0583
∗
0.474 -0.00922 0.00171
(0.0194) (0.00992) (0.0347) (0.360) (0.00681) (0.00114)
Nb case -0.00000643 -0.00000560 -0.0000653
∗∗∗
0.000416
∗∗∗
-0.00000622
∗∗∗
0.000000415
(0.00000777) (0.00000543) (0.0000211) (0.000113) (0.00000200) (0.000000422)
Observations 33069 33069 33069 32469 32469 32469
Adjusted R
2
0.000 0.000 0.007 0.001 0.001 -0.000
Panel B: Crime Categories
(1) (2) (3) (4) (5) (6)
Assault Burglary Drug Fraud Larceny Robbery
Nb case in district -0.0000284 0.0000705 0.0000702 0.0000275 0.00000211 -0.0000424
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Nb case in district * Elected 0.0000194 -0.0000639 0.00000679 0.00000944 -0.000000228 0.0000328
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Elected -0.00633 0.0110 -0.00975 -0.000931 -0.00483 -0.00738
(0.005) (0.014) (0.020) (0.009) (0.009) (0.010)
Nb case 0.00000605
∗∗
-0.0000179
∗∗
-0.00000548 -0.00000249 -0.00000384 0.00000907
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Observations 33069 33069 33069 33069 33069 33069
Adjusted R
2
0.000 0.001 0.000 0.000 0.000 0.001
Notes. Panel A: First offense is a dummy variable equaling one if it is the offender’s first criminal case. Female and Black are dummies equaling one for defendants who are
female and black, respectively. Minor and Senior are dummies equaling one if defendant is below 18 and above 64, respectively. Panel B: Each regression presents the effect
on a dummy equaling one if the crime committed corresponds to the category mentioned in the header. In both panels: Case number in district is the order of the case in the
district for the judge. Elected is a dummy equaling one if the judge is elected in the district. Case number is the order of the case over the judge’s entire career. Standard
errors are clustered at the judge level. t-Statistics are reported in parentheses. p<0.1,
∗∗
p<0.05,
∗∗∗
p<0.01
34
Table 3: Impact of Case Order on Distance Between Sentence and Local Norm sentence
(1) (2) (3) (4) (5) (6) (7)
OLS q10 q25 q50 q75 q90 |distance|
Case number in district -0.064 0.390
∗∗∗
0.148
∗
-0.048
∗
-0.084 -0.661
∗∗∗
-0.236
∗∗∗
(0.052) (0.112) (0.085) (0.029) (0.069) (0.134) (0.084)
Case number in district 0.049 -0.249
∗
-0.101
∗
0.024 0.090 0.457
∗∗∗
0.164
∗∗∗
* Elected (0.056) (0.146) (0.060) (0.028) (0.056) (0.127) (0.055)
Elected -6.974 17.573 26.304 10.652 -5.487 -63.161
∗
-25.419
(13.527) (34.429) (17.756) (7.666) (5.560) (33.392) (15.149)
Case number 0.012 -0.084
∗∗∗
-0.020 0.010
∗∗
0.030 0.128
∗∗
0.046
∗∗
(0.009) (0.021) (0.018) (0.004) (0.030) (0.050) (0.021)
Observations 33069 33069 33069 33069 33069 33069 33069
Notes. In columns 1 to 6, the outcome variable is the distance between the sentence and the average sentence, determined locally. Column 1 presents the result when using an
OLS regression, while columns 2 to 6 present the results when using quantile regressions. The quantile is indicated in the header. In column 7, the outcome variable is the
absolute value of the distance between the sentence and the average sentence, determined locally. Case number in district is the order of the case in the district for the judge.
Elected is a dummy equal to one if the judge is elected in the district. Case number is the order of the case over the judge’s entire career. Standard errors are clustered at the
judge level and reported in parentheses. * p<0.1,
∗∗
p<0.05,
∗∗∗
p<0.01
35
Table 4: Robustness of Convergence to Local Norm
(1) (2) (3) (4) (5) (6) (7)
Norm S elected Norm S+J>400 300 500 S-D Controls Judge fe Active+inactive part
Case number in district -0.356
∗∗∗
-0.241
∗∗∗
-0.155
∗∗
-0.250
∗∗∗
-0.244
∗∗∗
-0.248
∗∗∗
-0.211
∗∗∗
(0.106) (0.086) (0.068) (0.071) (0.085) (0.086) (0.066)
Case number in district 0.261
∗∗∗
0.152
∗∗∗
0.152
∗∗
0.126
∗∗∗
0.170
∗∗∗
0.171
∗∗∗
0.135
∗∗
* Elected (0.075) (0.055) (0.067) (0.045) (0.054) (0.058) (0.051)
Elected -21.326 -22.234 -34.411
∗∗
-15.317 -24.770 -18.199 -20.432
(23.524) (14.899) (16.970) (17.562) (14.789) (20.593) (13.614)
Case number 0.061
∗∗
0.047
∗∗
0.020
∗
0.067
∗∗∗
0.047
∗∗
0.049
∗∗
0.045
∗∗∗
(0.029) (0.021) (0.012) (0.017) (0.021) (0.021) (0.016)
Observations 19361 33809 29372 33751 32469 33069 33069
Adjusted R
2
0.012 0.007 0.002 0.009 0.024 0.025 0.007
Notes. In every column, the outcome variable is the absolute value of the distance between the sentence and the average sentence, determined locally. In column 1, the average
sentence is calculated among senior judges elected in the district. In column 2, the average sentence is calculated among all senior judges and junior judges after their 400th
decision in the district. In column 3, we keep only the first 300 decisions made by a junior judge in a district. In column 4, we keep only the first 500 decisions made by a
judge in a district. In column 5, we control for the following case characteristics: sex, race, age, and dummies for below 18 and above 64. In column 6, we control for judicial
fixed effects. In column 7, the outcome variable is the distance between total (active plus inactive) sentence and the local sentencing norm, calculated using total sentences.
Case number in district is the order of the case in the district for the judge. Elected is a dummy equaling one if the judge is elected in the district. Case number is the order
of the case over the judge’s entire career. Standard errors are clustered at the judge level. Standard errors are clustered at the judge level and reported in parentheses. *p<0.1,
∗∗
p<0.05,
∗∗∗
p<0.01.
36
Table 5: Correlates of Local Norms
Sentencing Norm
(1) (2) (3) (4) (5)
crime prevalence in d -4.668
∗∗∗
-4.639
∗∗∗
-4.683
∗∗∗
-4.678
∗∗∗
-4.658
∗∗∗
(1.572) (1.618) (1.575) (1.578) (1.583)
HI for attorneys 0.104 0.105 0.066
(0.068) (0.071) (0.079)
HI for district attorneys -0.023
∗
-0.021
∗
-0.007
(0.012) (0.011) (0.011)
prison overcrowding -43.353 -41.639
(45.878) (38.474)
majority democrat 29.949
(25.126)
vote referendum 1996 1.114
(3.593)
vote referendum 2004 -3.727
∗
(2.023)
vote referendum 2010 -2.864
(3.693)
vote referendum 2014 3.086
(1.885)
Observations 948 948 948 948 948
Adjusted R
2
0.867 0.890 0.870 0.870 0.873
Note: All columns include crime fixed effects. We exclude crime categories that have fewer than 100 cases over the entire time
period. Column (2) includes district fixed effects. Standard errors are clustered at the district level and reported in parentheses.
∗
p<0.1,
∗∗
p<0.05,
∗∗∗
p<0.01.
37
Appendix B: robustness using total sentence
Figure 8: Effect of decision order in the district on distance to the average sentence, using
total sentences rather than active sentences
-200 0 200 400
Coefficients
1-50 51-100 101-150 151-200 201-250 251-300 301-351
Number of Decisions in District
10th percentile Median
90th percentile
Non-home Districts
-100 0 100 200
Coefficients
1-50 51-100 101-150 151-200 201-250 251-300 301-351
Number of Decisions in District
10th percentile Median
90th percentile
Home District
Notes. This figure reproduces Figure 5 for total sentences rather than active sentences
Figure 9: Effect of decision order in the district on distance for different quantiles, for total
sentences
-2 -1.5 -1 -.5 0 .5
Coefficient
0 20 40 60 80 100
Quantile
Non-home Districts Home District
Notes. This figure reproduces Figure 6 for total sentences rather than active sentences
38
Table 6: Correlates of the sentencing norm, measured using total sentences
Sentencing Norm
(1) (2) (3) (4) (5)
crime prevalence in d -4.010
∗∗∗
-3.974
∗∗∗
-4.023
∗∗∗
-4.020
∗∗∗
-3.993
∗∗∗
(1.416) (1.454) (1.418) (1.419) (1.425)
HI for attorneys 0.083 0.083 0.043
(0.070) (0.071) (0.071)
HI for district attorneys -0.023
∗
-0.022
∗
-0.003
(0.012) (0.012) (0.011)
prison overcrowding -19.576 -17.662
(49.336) (38.349)
majority democrat 37.367
(24.640)
vote referendum 1996 1.769
(3.663)
vote referendum 2004 -6.066
∗∗∗
(2.053)
vote referendum 2010 -4.861
(3.664)
vote referendum 2014 3.437
∗
(1.804)
Observations 948 948 948 948 948
Adjusted R
2
0.870 0.899 0.873 0.872 0.878
Notes. This table reproduces Table 5 for total sentences rather than active sentences
39
Appendix C
We analyze the instruments in turn before proving the results of Proposition 1. We
provide a lemma characterizing each instrument, with proofs provided later in the section.
Sentencing guidelines
The sentencing guideline is the sentence that maximizes the expected utility of the plan-
ner, uninformed about s
t
and β
d
, as expressed in the following lemma:
Lemma 1 The sentencing guideline is S
g
= θ and yields expected welfare of:
W
g
= −(1 + α)V [s
t
] − αV [β
d
] (3)
The expected welfare obtained under sentencing guidelines decreases when the uncertainty
on s
t
and β
d
increase. Welfare is particularly sensitive to uncertainty on s
t
, since better
knowledge of local conditions allows the planner to better tailor the sentence to her prefer-
ences, but also to minimize the cost of enforcement.
Rotation policy
The judge has to decide whether to invest C to obtain information about the local norm
β
d
. This choice is determined by the difference in expected utility when informed about β
d
and the expected utility without that information. The result is summarized below:
Lemma 2 The planner rotates the judge every N
∗
periods, where N
∗
is defined as the lowest
value of N such that:
1 − δ
N
1 − δ
γ
2
1 + γ
V (β
d
) ≥ C,
The planer under this rotation policy obtains a per period expected welfare of:
W
r
= −
γ
2
+ α
(1 + γ)
2
V (β
d
) − (1 − p)(1 + α)
1
(1 + γ)
2
ζ
2
(4)
The planner leaves the judge long enough in the district to ensure she acquires the informa-
tion. If V (β
d
) increases, the judge rotates more frequently because information has a high
value for the judge who no longer needs to be incentivized as much. The expected welfare of
the planner under the rotation policy decreases with V (β
d
), as this increases the expected
cost of enforcement, and decreases with the bias of the judge ζ.
40
Elections
We examine the final tool at the disposal of the planner, elections. Consistent with the
empirical evidence we presented, showing that for judges in their home district there is no
convergence to the sentencing norm, we assume that the judge, in her home district, knows
β
d
without having to invest C.
The citizen, in the case of an election in a soft district β
d
< 0, wants to reelect a regular
candidate and vote against a tough candidate. When focusing on the welfare-maximizing
equilibrium, we obtain that:
Lemma 3 If judicial election is the only tool at the planner’s disposal, then all judges,
regardless of their type or information on case specifics s
t
, pool on a single sentence. If the
judge pools on the socially preferred sentence, the expected welfare is given by:
W
e
= −
α
(1 + α)
V (β
d
) − (1 + α)V (s
t
) (5)
Lemma 3 shows that all judges, normal or tough, pool on a single sentence and ignore infor-
mation on case specifics. Indeed, if there were at least two sentences chosen in equilibrium,
the normal judge would always be more likely to choose the lowest one and would thus
guarantee her own electoral defeat. In the welfare-maximizing equilibrium, the sentence on
which they pool is the planner’s preferred sentence S
e
= θ +
α
1+α
β
d
.
Comparing instruments
Proposition 1 immediately follows from Lemmas 1-3. We restate the result:
Proposition 1 There exists
˜
V and
¯
V such that:
1. if V (s
t
) ≥
˜
V , rotation is the socially preferred instrument,
2. if V (s
t
) = 0 and V (β
d
) ≤
¯
V , sentencing guidelines are socially preferred to rotation.
Lemma 2 shows that welfare under rotation optimal policy is independent of V (s
t
), while
it is strictly decreasing both under sentencing guidelines and elections according to lemmas
1 and 3, so result 1 immediately follows, with a benchmark value
˜
V of the variance of s
t
,
such that, if the variance is larger than that, rotation dominates. The key insight is that,
in the pooling equilibrium, the judge ignores the local conditions s
t
. The second result is
based on how, according to Lemma 1, when V (s
t
) = 0 and V (β
d
) converge to zero, welfare
under sentencing guidelines goes to zero, while it is strictly negative for rotation due to the
41
potential bias of the judge. Thus, there exists a critical value of the variance of β
d
such that
sentencing guidelines dominate.
Proof Lemma 1
For a given sentencing guideline S, the utility, given that s
t
and β
d
have zero means, is
given by:
E
u
d
p
= −E
(S − (θ + s
t
))
2
− αE
(S − (θ + β
d
+ s
t
))
2
= −S
2
+ 2θS − θ
2
− E
s
2
t
− αS
2
+ α2θS − αE
s
2
t
− αE
β
2
d
− αθ
2
= −(1 + α)S
2
+ 2(1 + α)θS − (1 + α)θ
2
− (1 + α)V [s
t
] − αV [β
d
]
The first order condition thus yields that the sentencing guideline is chosen as S
g
= θ. Thus,
expected social welfare under sentencing guidelines is given by:
W
g
= −V [s
t
] − α (V [s
t
] + V [β
d
])
= −(1 + α)V [s
t
] − αV [β
d
]
Proof Lemma 2
We first determine the utility of the judge who invests in information and faces a case
with severity s
t
in a district with norm β
d
. The sentence that solves the first order condition
in this case is given by:
S
r,i
= θ + s
t
+
1
1 + γ
ζ
j
+
γ
1 + γ
β
d
So the per period utility for the judge when she has information on β
d
and s
t
is:
u
r,i
= −
θ + s
t
+
1
1 + γ
ζ
j
+
γ
1 + γ
β
d
− (θ + s
t
+ ζ
j
)
2
− γ
θ + s
t
+
1
1 + γ
ζ
j
+
γ
1 + γ
β
d
− (θ + s
t
+ β
d
)
2
+ w
= −
γ
1 + γ
2
(β
d
− ζ
j
)
2
− γ
1
1 + γ
2
(ζ
j
− β
d
)
2
= −
γ
1 + γ
(β
d
− ζ
j
)
2
So, the expected utility before obtaining the information is:
W
r,i
= −
γ
1 + γ
V (β
d
) + ζ
2
j
Now, we determine the expected welfare when no information on β
d
is obtained. In this
42
case, the sentence will be fixed at:
S
r,u
= θ + s
t
+
1
1 + γ
ζ
j
For an expected welfare of:
W
r,u
= −
γ
1 + γ
ζ
2
j
− γV (β
d
)
The per period welfare is naturally smaller when the judge is uninformed. There is thus a
minimum number of periods N
r
such that the judge will acquire information if and only if
N ≥ N
r
, where N
r
is defined is defined as the lowest value of N such that:
1 − δ
N
1 − δ
(W
r,i
− W
r,u
) ≥ C,
i.e
1 − δ
N
1 − δ
γ
2
1 + γ
V (β
d
) ≥ C,
The planner will choose rotation after N periods and the welfare of the planner is given
by:
W
r
= −
γ
2
+ α
(1 + γ)
2
V (β
d
) − (1 − p)(1 + α)
1
(1 + γ)
2
ζ
2
j
Proof Lemma 3
Consider an equilibrium such that two level of sentences S
1
and S
2
< S
1
are awarded in
equilibrium. There are benchmark value ˜s for the regular judge and ¯s for the tough judge,
with ¯s < ˜s such that the regular judge chooses S
1
if and only if s
t
≥ ˜s and the tough judge
chooses S
1
if and only if s
t
≥ ¯s. Thus, since the regular type prefers lower sentences on
average, when the voter sees sentence S
2
chosen, she increases her posterior belief that the
incumbent is not tough and does not reelect her. Under our assumption that the wage is
high enough that the judge wants to win the election at all costs, sentence S
2
would never
be chosen in equilibrium.
Therefore, the unique equilibrium is a pooling equilibrium. In the welfare-maximizing
equilibrium, the judges pool on the preferred sentence of the planner S
e
= θ +
α
1+α
β
d
. The
welfare is then given by:
W
e
= −
α
(1 + α)
V (β
d
) − (1 + α)V (s
t
)
Proof Proposition 1
43
1. According to lemmas 1 and 3, expected welfare under rotation and elections (even
in the case of the pooling equilibrium that maximizes welfare) is decreasing in V (s
t
), while
lemma 2 shows that welfare under rotation is independent of V (s
t
). Result (1) directly
follows.
2. According to expression (3), when V (s
t
) = 0 and V (β
d
) converges to 0, expected
welfare under sentencing guidelines goes to zero. On the contrary, according to expression
(4), expected welfare under rotation converges to W
r
= −(1 − p)(1 + α)
1
(1+γ)
2
ζ
2
< 0. Thus
there is a range of values of V (β
d
) such that sentencing guidelines dominate rotation.
44
Appendix D- Web Appendix
7 Variable Description and Data Set Construction
7.1 Case Definition: Charge and Sentence
The first step of our analysis is a case definition. Since a criminal case is often comprised
of multiple charges for a single defendant, and our focus is on overall sentencing for a case,
we build on the procedure described in Abrams and Fackler (2018) in identifying cases with
the same criminal and disposition date, and we define a case as a unique person-date of
disposition. Treating a multiple charges case as a single unit implies that we must decide
which charge to keep. We proceed as follows: we first define the lead charge of an incident
as the charge with the highest associated sentence length. Our main sentencing variable
is then defined as the minimum sentence determined by the judge for the lead charge. As
we have specified in the main text, when a defendant is found guilty to a felony, North
Carolina imposes a sentencing range. If the judge determines the sentence should be active,
the defendant is required to serve the full minimum of the range, and may serve less than
the maximum with good behavior. The final active sentence is the main variable used in
our analysis. It is worth noting that, in order to deal with the outliers in sentences, in our
analysis we winsorize this variable at the 5 percent level.
7.2 Judge identity
The second step to conduct our analysis is to identify the judge dealing with each case. North
Carolina sentencing data reports a judge acronym (with two or three letters) for each case.
In order to identify a specific judge based on these acronyms, we use the Master Schedules
recording in which district and division a judge is in a given week. Using this information,
we construct one, two, and three letter acronyms for each judge in the schedule and match
this with our case data. Using disposition date data and the acronyms, we are able to match
the 84 percent of judges in the master schedule to cases in the sentencing data. We only
keep these observations in our working sample. For judges elected or nominated after the
1998, we observe the whole history of decisions.
7.3 District level demographic variables
We collect various demographic and other district level variables that we use in different steps
of the analysis. District level demographic characteristics are constructed starting from the
45
US Bureau of Census data. We use variables for the year 2010 (the most recent census fully
available). These variables (listed in descriptive statistics table of the paper) are collected
at the county level and are then aggregated at the district level since each district usually
includes more than one county.
7.4 District level administrative and political variables
Prison population data are collected from the National Census of Jails and the Annual Sur-
vey of Jails and are used to construct a crowding metric as the ratio between the value of the
total population of inmates at that county’s jail facilities at the survey date and the rated
capacity of the jail, which measures the maximum number of beds (and therefore overnight
inmates) that could fit into the facility on the date the survey was taken.
Finally, we collect data on referenda about justice. We identified four referenda that took
place in 1996, 2004, 2010, and 2014. The 1996 referendum asked voters about the expansion
of alternative punishments to be used on convicted criminals, such as probation and commu-
nity service. The 2004 referendum aims at clarifying and defining several areas of jurisdiction
of the courts, and changed the term of office of magistrates to provide for an initial term of
2 years and subsequent terms of 4 years. The 2010 is intended to prohibit convicted felons
for running as sheriffs in the state and finally the 2014 introduces the possibility for felons to
waive a trial by jury. We collect data about county level votes in these referenda from North
Carolina’s Board of Elections and then aggregate them at the district level to compute the
percentage of votes in favor or against the main question asked in the referendum.
46
Figure 10: Sentencing guidelines
4
Figure A
*** Effective for Offenses Committed on or after 10/1/13 ***
FELONY PUNISHMENT CHART
PRIOR RECORD LEVEL
OFFENSE CLASS
I
0-1 Pt
II
2-5 Pts
III
6-9 Pts
IV
10-13 Pts
V
14-17 Pts
VI
18+ Pts
A
Death or Life Without Parole
Defendant Under 18 at Time of Offense: Life With or Without Parole
B1
A
A
A
A
A
A
DISPOSITION
240 - 300
276 - 345
317 -397
365 - 456
Life Without
Parole
Life Without
Parole
Aggravated Range
192 - 240
221 - 276
254 - 317
292 - 365
336 - 420
386 - 483
PRESUMPTIVE RANGE
144 - 192
166 - 221
190 - 254
219 - 292
252 - 336
290 - 386
Mitigated Range
B2
A
A
A
A
A
A
157 - 196
180 - 225
207 - 258
238 - 297
273 - 342
314 - 393
125 - 157
144 - 180
165 - 207
190 - 238
219 - 273
251 - 314
94 - 125
108 - 144
124 - 165
143 - 190
164 - 219
189 - 251
C
A
A
A
A
A
A
73 – 92
83 - 104
96 - 120
110 - 138
127 - 159
146 - 182
58 - 73
67 - 83
77 - 96
88 - 110
101 - 127
117 - 146
44 - 58
50 - 67
58 - 77
66 - 88
76 - 101
87 - 117
D
A
A
A
A
A
A
64 - 80
73 - 92
84 - 105
97 - 121
111 - 139
128 - 160
51 - 64
59 - 73
67 - 84
78 - 97
89 - 111
103 - 128
38 - 51
44 - 59
51 - 67
58 - 78
67 - 89
77 - 103
E
I/A
I/A
A
A
A
A
25 - 31
29 - 36
33 - 41
38 - 48
44 - 55
50 - 63
20 - 25
23 - 29
26 - 33
30 - 38
35 - 44
40 - 50
15 - 20
17 - 23
20 - 26
23 - 30
26 - 35
30 - 40
F
I/A
I/A
I/A
A
A
A
16 - 20
19 - 23
21 - 27
25 - 31
28 - 36
33 - 41
13 - 16
15 - 19
17 - 21
20 - 25
23 - 28
26 - 33
10 - 13
11 - 15
13 - 17
15 - 20
17 - 23
20 - 26
G
I/A
I/A
I/A
I/A
A
A
13 - 16
14 - 18
17 - 21
19 - 24
22 - 27
25 - 31
10 - 13
12 - 14
13 - 17
15 - 19
17 - 22
20 - 25
8 - 10
9 - 12
10 - 13
11 - 15
13 - 17
15 - 20
H
C/I/A
I/A
I/A
I/A
I/A
A
6 - 8
8 - 10
10 - 12
11 - 14
15 - 19
20 - 25
5 - 6
6 - 8
8 - 10
9 - 11
12 - 15
16 - 20
4 - 5
4 - 6
6 - 8
7 - 9
9 - 12
12 - 16
I
C
C/I
I
I/A
I/A
I/A
6 - 8
6 - 8
6 - 8
8 - 10
9 - 11
10 - 12
4 - 6
4 - 6
5 - 6
6 - 8
7 - 9
8 - 10
3 - 4
3 - 4
4 - 5
4 - 6
5 - 7
6 - 8
A – Active Punishment I – Intermediate Punishment C – Community Punishment
Numbers shown are in months and represent the range of minimum sentences
Revised: 09-09-13
47
Le LIEPP (Laboratoire interdisciplinaire d'évaluation des
politiques publiques) est un laboratoire d'excellence (Labex).
Ce projet est distingué par le jury scientifique international
désigné par l'Agence nationale de la recherche (ANR).
Il est financé dans le cadre des investissements d'avenir.
(ANR-11-LABX-0091, ANR-11-IDEX-0005-02)
www.sciencespo.fr/liepp
A propos de la publication
Procédure de soumission :
Rédigé par un ou plusieurs chercheurs sur un projet en cours, le Working paper vise à susciter
la discussion scientifique et à faire progresser la connaissance sur le sujet étudié. Il est destiné
à être publié dans des revues à comité de lecture (peer review) et à ce titre répond aux
exigences académiques. Les textes proposés peuvent être en français ou en anglais. En début
de texte doivent figurer : les auteurs et leur affiliation institutionnelle, un résumé et des mots
clefs.
Le manuscrit sera adressé à : [email protected]
Les opinions exprimées dans les articles ou reproduites dans les analyses n’engagent que leurs
auteurs.
Directeur de publication :
Bruno Palier
Comité de rédaction :
Andreana Khristova, Carolina Alban Paredes
Sciences Po - LIEPP
27 rue Saint Guillaume
75007 Paris - France
+33(0)1.45.49.83.61
