econstor

A Service of

zbw

Make Your Publications Visible.

Leibniz-Informationszentrum
Wirtschaft
Leibniz Information Centre
for Economics

Mastrobuoni, Giovanni; Rivers, David A.

Working Paper

Criminal Discount Factors and Deterrence

IZA Discussion Papers, No. 9769
Provided in Cooperation with:
IZA – Institute of Labor Economics

Suggested Citation: Mastrobuoni, Giovanni; Rivers, David A. (2016) : Criminal Discount Factors
and Deterrence, IZA Discussion Papers, No. 9769, Institute for the Study of Labor (IZA), Bonn

This Version is available at:
http://hdl.handle.net/10419/141528

Standard-Nutzungsbedingungen:

Terms of use:

Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen
Zwecken und zum Privatgebrauch gespeichert und kopiert werden.

Documents in EconStor may be saved and copied for your
personal and scholarly purposes.

Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich
machen, vertreiben oder anderweitig nutzen.

You are not to copy documents for public or commercial
purposes, to exhibit the documents publicly, to make them
publicly available on the internet, or to distribute or otherwise
use the documents in public.

Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten,
gelten abweichend von diesen Nutzungsbedingungen die in der dort
genannten Lizenz gewährten Nutzungsrechte.

www.econstor.eu

If the documents have been made available under an Open
Content Licence (especially Creative Commons Licences), you
may exercise further usage rights as specified in the indicated
licence.

 SERIES
PAPER
DISCUSSION

IZA DP No. 9769

Criminal Discount Factors and Deterrence
Giovanni Mastrobuoni
David A. Rivers

February 2016

Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor

 Criminal Discount Factors and
Deterrence
Giovanni Mastrobuoni
University of Essex
and IZA

David A. Rivers
University of Western Ontario

Discussion Paper No. 9769
February 2016

IZA
P.O. Box 7240
53072 Bonn
Germany
Phone: +49-228-3894-0
Fax: +49-228-3894-180
E-mail: iza@iza.org

Any opinions expressed here are those of the author(s) and not those of IZA. Research published in
this series may include views on policy, but the institute itself takes no institutional policy positions.
The IZA research network is committed to the IZA Guiding Principles of Research Integrity.
The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
organization supported by Deutsche Post Foundation. The center is associated with the University of
Bonn and offers a stimulating research environment through its international network, workshops and
conferences, data service, project support, research visits and doctoral program. IZA engages in (i)
original and internationally competitive research in all fields of labor economics, (ii) development of
policy concepts, and (iii) dissemination of research results and concepts to the interested public.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.

 IZA Discussion Paper No. 9769
February 2016

ABSTRACT
Criminal Discount Factors and Deterrence*
The trade-off between the immediate returns from committing a crime and the future costs of
punishment depends on an offender’s time discounting. We exploit quasi-experimental
variation in sentence length generated by a large collective pardon in Italy and provide nonparametric evidence on the extent of discounting from the raw data on recidivism and
sentence length. Using a discrete-choice model of recidivism, we estimate an average annual
discount factor of 0.74, although there is heterogeneity based on age, education, crime type,
and nationality. Our estimates imply that the majority of deterrence is derived from the first
few years in prison.

JEL Classification:
Keywords:

D9, K4

discounting, deterrence, collective pardon, recidivism

Corresponding author:
Giovanni Mastrobuoni
University of Essex
Department of Economics
Wivenhoe Park
C04 3SQ Colchester
United Kingdom
E-mail: gmastrob@essex.ac.uk

*

We would like to thank Philip Cook, Igor Livshits, Lance Lochner, Salvador Navarro, Daniel Nagin, as
well as seminar participants at the University of Essex, the University of Western Ontario, and the
University of Trier.

 1

Introduction

Criminal activity involves an inherent trade-off between the immediate returns associated with committing the crime and the possibility of future punishments if caught. Since punishments are received in the
future, the value that a potential criminal places on them depends on how that individual values future
events. One implication is that individuals who more heavily discount the future will be more likely
to engage in crime. Knowing the degree to which potential criminals discount the future is therefore a
key component to developing a better understanding of what drives criminal behavior and for designing
crime policy. This information is of substantial interest to policymakers and those involved in law enforcement, as it relates directly to the deterrent power of criminal punishments. This is a pressing issue,
as policy makers are currently deciding whether to reform some of the mandatory minimum sentences
enacted in the 1970s and 1980s, which contributed to prison population increases of 400 percent (Hunt
(2015)).1
Discounting has two important implications for deterrence. First, the more an individual discounts
future events, the less deterrent power future punishments have. Second, discounting shapes the relative
deterrent power of punishments received at different points in the future. For example, consider an offender who is choosing whether to use a modus operandi that might aggravate his/her crime (e.g., using
a firearm, being violent, damaging property, etc.). As long as criminals discount the future, individuals
considering crimes with longer baseline sentences will be less deterred by the same sentence enhancements compared to those considering crimes with shorter sentences, as these enhancements would be
served further into the future. The magnitude of this differential effect is increasing in the degree of
discounting.
Discounting also has implications for the debate over the effect of severity of punishment on deterrence. In the original economic model of crime developed by Becker (1968), certainty and severity of
punishment combine to form the expected cost of committing a crime. Since severity is typically associated with sentence length, it is commonly assumed that doubling sentence length leads to a doubling of
the “costs” paid by offenders, and is therefore equivalent to a doubling of the probability of punishment
(certainty). But this is only true if criminals do not discount the future, since increased sentence lengths
are added on at the end. Discounting of future consequences fundamentally changes this relationship, as
1 The US House of Representatives has recently unveiled a sentencing reform legislation as a companion to a bipartisan bill
introduced in the Senate on October 1, 2015, the Sentencing Reform and Corrections Act of 2015 (S. 2123).

2

 it implies decreasing marginal deterrence of sentence length (severity). As a consequence, the marginal
cost to society of a given increase in deterrence is increasing, rather than constant, as sentences are
extended. A related consequence is that as discounting increases, the deterrent power of severity of punishment decreases relative to that of the certainty of punishment. The notion that swiftness and certainty
of the penalty outmatch severity has been theorized at least as early as 250 years ago by Beccaria (1764).
In a survey of the recent evidence for deterrent effects of imprisonment, Durlauf and Nagin (2011)
find that there is little evidence for strong marginal deterrent effects of increasing the severity of punishment, but considerable support in the literature for marginal deterrent effects of increasing the certainty
of punishment (see e.g., Hawken and Kleiman (2009)). One explanation that they offer is low discount
factors among criminals.
The idea that discounting is an important component of criminal behavior has been recognized in the
literature at least since Ehrlich (1973). Cook (1980) highlights that with discounting and a constant perperiod disutility of prison, increasing the severity of punishments will have a greater deterrent effect when
the initial punishment is mild. Davis (1988) and Polinsky and Shavell (1999) develop formal models of
intertemporal crime decisions that explicitly take into account the effect of discounting. Wilson and
Herrnstein (1985) and Katz, Levitt, and Shustorovich (2003) suggest that one explanation for criminal
behavior itself is low discount factors among those individuals that decide to commit crimes. However,
despite prior recognition in the literature of the importance of discounting for criminal behavior (see
also McCrary (2010)), there is very limited empirical evidence about the extent of discounting among
criminals.
In this paper, we provide direct quantitative estimates of discount factors for criminals by taking
advantage of a unique dataset related to a large-scale collective pardon passed in Italy at the end of July,
2006. In an attempt to reduce prison overcrowding, more than 20,000 inmates, corresponding to over
one-third of the entire prison population, were released over a period of a few weeks. Inmates who had
a residual sentence below three years at the time of the collective pardon were immediately released. A
key condition of the pardon was that, if a released inmate is found guilty of a crime in the future, the
pardoned sentence (or residual sentence) is automatically added to the new sentence.
Our dataset consists of individual-level data on each released inmate. In addition to detailed information on the characteristics of each inmate, we observe the length of their original sentence that led to
their pre-pardon incarceration, their residual sentence length, and whether or not they recidivate over a

3

 period of 17 months following the pardon.
Our identification strategy for estimating the discount factor is based on measuring how recidivism
rates decrease with the length of the sentence. Specifically, it is the shape of this relationship that identifies the discount factor. The presence of discounting implies that the marginal effect on offending is
decreasing in sentence length, as individuals discount the later years of a sentence more heavily. This
leads to a convex decreasing relationship between offending (recidivism) and total sentence length. The
more convex the relationship is, the lower the discount factor.
A key challenge for identification arises due to the fact that prison sentences are not randomly assigned. Judges impose sentences based on characteristics of the offender and the offender’s criminal history, many of which will be unobserved to the researcher. This correlation between unobserved drivers of
crime and observed sentence lengths generates an endogeneity problem, which has long been recognized
in the deterrence literature (see Ehrlich (1973) for an early discussion and Levitt and Miles (2007) and
Durlauf and Nagin (2011) for more recent summaries).
The natural experiment generated by the Italian pardon, however, provides a solution. Under the
conditions of the pardon, inmates whose original incarceration happened at different times face different
expected total sentences for the same future crime. Conditional on the length of the original sentence, the
residual sentence only depends on the original date of entry into prison, which is plausibly exogenous
and can be exploited to estimate the discount factor.2 This source of exogenous variation in sentence
length has been used previously by Drago, Galbiati, and Vertova (2009) to measure the average deterrent
effect of imprisonment.
An additional advantage of our analysis is that the policy we study is likely to be particularly salient.
Our dataset consists entirely of individuals who have been previously incarcerated, and therefore are
particularly likely to be aware of sanctioning rules (Kaplow (1990)). The California Assembly Study
(1968) shows, not surprisingly, that knowledge of the maximum penalty for various FBI index type
crimes is far better among incarcerated individuals (62 percent) compared to the rest of the population
(25 percent).3 Furthermore, the same policy was applied simultaneously to a large number of inmates,
which should have served to circulate knowledge and understanding of the policy. Finally, one key aspect
2 Conditioning

on the original sentence is important, because individuals with longer original sentences are more likely to
have longer residual sentences, as well as higher recidivism rates. This is because individuals that are deemed more likely to
recidivate are more likely to be given longer original sentences, and mechanically, individuals with longer original sentences
are more likely to have longer residual sentences at the time of the pardon.
3 See also Nagin (2013) and Lochner (2007) on the importance of criminals’ perceptions in shaping crime.

4

 of the policy (early release) was almost certainly noticed by all affected inmates.
Our estimates imply annual discount factors of 0.74 among previously convicted criminals. This
degree of discounting suggests that while these individuals place a significantly lower value on the future,
it is still the case that punishments, even those received several years in the future, entail non-negligible
costs to a potential offender.
In order to highlight the importance of knowing the magnitude of discounting among criminals, it is
useful to compare the effect of increasing sentence length for different values of the discount factor. Our
estimated discount factor of 0.74 implies that doubling a 5-year sentence increases the disutility of prison
by 22%, whereas doubling a 10-year sentence increases it by only 5%. If, instead, we employ a more
traditional value of the discount factor of 0.95, these increases would be 77% and 60% respectively, suggesting a much larger potential role of sentence increases for deterrence, even at relatively long sentence
levels. Our estimates, therefore, imply that increases in sentence length, either through mandatory minimums, sentence enhancements, or more severe sentencing, are unlikely to have much deterrent effect
when the baseline sentence is already long.
Since our dataset contains detailed information on the characteristics of each inmate, we are able to
examine differences in discount factors across many different dimensions of individual heterogeneity.
We find evidence of substantial heterogeneity in discount factors, suggesting that across different groups
of individuals, the marginal deterrent effect of imprisonment diverges as sentences become longer. The
largest differences in discount factors are associated with differences in educational attainment, age, nationality, and crime types. While the average discount factor is estimated to be 0.74 in our baseline
specification, for some groups (high education and those who commit crimes related to organizing prostitution) the estimated discount factors are much closer to 1 (0.99 and 0.92, respectively). The lowest
discount factors are found for immigrants (0.66) and drug-offenders (0.70). This heterogeneity in discount factors implies important differences in the deterrent effect of imprisonment across the population.
For example, given our estimates, longer prison sentences are likely to generate more effective deterrence
for offenders committing prostitution-related crimes compared to drug offenders.
There are a few studies in the literature that have taken advantage of survey questions designed to
elicit time preferences or measures of impulsivity, and relate these to criminal behavior (see e.g., Nagin
and Pogarsky (2004); Jolliffe and Farrington (2009); Akerlund et al. (2014); and Mancino, Navarro, and
Rivers (2015)). These papers generally find that individuals who are more present-oriented, or impulsive,

5

 are more likely to commit a crime. One drawback of these approaches is that it can be difficult to quantify
the magnitude (as opposed to just the sign) of these relationships given the qualitative measures of time
preference that are often used. There are also concerns regarding the measurement of time preference
from laboratory/experimental studies (Coller and Williams (1999)). The paper most closely related to
ours is Lee and McCrary (2009), who analyze recidivism rates for a group of released juvenile inmates
using a regression discontinuity design around the age of majority (age 18) in the severity of sentencing.
They find only a small decrease in offending at the age of 18, when expected sentence length increases,
which is consistent with low discount factors.
We also contribute to the non-experimental literature that has estimated discount factors for consumers (see the early studies by Friedman (1957) and Heckman (1976)) as well as for various other subpopulations (US military personnel, Warner and Pleeter (2001); US purchasers of durable and storable
goods, Hausman (1979); Ching and Osborne (2015); UK homeowners, Giglio, Maggiori, and Stroebel
(2015); and US workers, Viscusi and Moore (1989)).4 Our dataset contains only individuals who have
been convicted of at least one crime, and therefore our results correspond to the subset of the population
with a criminal record. Relative to the literature that estimates discount factors for the broader population, our results are on the low end of the range of estimates, particularly among those papers that use
observational or non-experimental data. Thus our estimates are consistent with high discounting serving
as a determinant of crime.
The rest of the paper is organized as follows. In Section 2 we describe our dataset covering the
collective pardon in Italy. Section 3 contains the description of our model and Section 4 details our
identification strategy. In Section 5 we present our results and perform several robustness checks. In
Section 6 we discuss the implications of our results for measuring deterrence. Section 7 concludes.

2

Data

Collective pardons are deeply rooted in Italian history. Since World War II there has been on average one
pardon or amnesty every five years (Barbarino and Mastrobuoni (2014)), although in more recent years
there have been only two, in 1990 and in 2006. Such pardons eliminate part of the inmates’ sentences,
typically two or three years, and inmates whose new net sentence drops below 0 are immediately released.
4A

series of papers also estimate discount factors using laboratory experiments (see e.g., Harrison, Lau, and Williams
(2002); Harrison, Lau, and Rutström (2010); Meier and Sprenger (2010); Coller and Williams (1999)).

6

 Given their wide reach, these pardons generate sudden releases of large numbers of inmates.5
On July 31, 2006 the Italian Parliament passed the last pardon, which was enacted shortly thereafter.
We were given access to the incarceration spell of all prison inmates released on this occasion, including
the exact dates of release and re-incarceration through December 2007. A key condition of the release
is that if pardoned inmates are rearrested and convicted, the pardoned sentenced is added to the new
sentence. Conditional on the initial sentence, this generates a plausibly exogenous source of variation in
the expected severity of punishment.6 This variation in expected future sentences allows us to measure
how recidivism varies with sentencing and thus identify the discount factor. We also have information
on inmates’ nationality, age, education, and some other individual characteristics.
These are the same data used in Mastrobuoni and Pinotti (2015), and, compared to Drago, Galbiati,
and Vertova (2009), allow us to follow the inmates an additional 10 months after release, for a total of
17 months. A large number of inmates were released (over 20,000 individuals representing more than
one-third of the prison population).7 In Table 1 we present summary statistics for the variables in our
dataset. The average recidivism rate (within 17 months of release) is about 22%. Because our sample
consists of individuals who have previously been convicted of a crime and been released from prison,
the mean age in our sample (38) is higher than in many datasets on criminal offenders. One consequence
of this is that these inmates are more likely to be familiar with the Italian criminal law and the rules
governing their release. In other words, the deterrent effect of the law is likely to be salient.
Overall the sample is quite uneducated. As shown in Table 1, about 27% of the individuals in our data
have an education at or below primary school, and only 3% completed secondary school. About 62%
of inmates are of Italian origin, 26% are married, 5% are female, and 15% were permanently employed
prior to being imprisoned. 74% have a definitive sentence (i.e., exhausted all appeals).
Table 2 provides a breakdown by crime type. Crime types are not mutually exclusive, as a crime can
be both a property crime and a violent crime (e.g., robbery), and therefore the means of the indicator
variables for each crime type do not sum to 1. Property crime is the most common type (58%), and
prostitution and mafia-related crimes are the least common (2% each).
5 In the 2006 pardon a few crimes were excluded, including those related to the mafia, terrorism, kidnapping, and felony sex
offenses, while other types of violent offenses, such as murder and armed robbery, were eligible.
6 Technically, only individuals who receive a future sentence of at least two years are subject to serving their residual
(pardoned) sentence. We discuss this in more detail in Section 5.3.
7 As in Drago, Galbiati, and Vertova (2009), we drop individuals who were released while still awaiting trial, as we have no
information about their sentence. In addition, we lose a small number of observations when matching the sample of released
inmates used in Drago, Galbiati, and Vertova (2009) with a data extraction made by the Italian Prison Administration 17 months
after the collective pardon.

7

 We supplement our main dataset with auxiliary data on clearance rates and transition probabilities
across crime types. Province-level data on the fraction of cleared crimes come from ISTAT (2005)’s
criminal law statistics and are merged by province (there are 103 provinces in Italy) and crime type, with
the individual-level recidivism data. The definition of a clearance in these data is that at least one suspect
has been identified. As shown in Table 3, the province-level clearance rate is 23% on average, but with
considerable variation both across location and crime type.
While we have information on the crime committed before the pardon, for inmates who are reincarcerated all we observe for the new offense is the date of re-incarceration. Our main dataset also
does not contain information on previous incarcerations that could be used to construct transition probabilities across crime types. Instead, we take advantage of a dataset containing information on all inmates
who served time in prison in one of two prisons located in Milan (Bollate and Opera) between 2001 and
2012. Using the entire history of offenses for these individuals (spanning from 1972-2012), we construct
transition probabilities between crime types and merge these into our dataset. For inmates who commit
a crime that falls into more than one category, we base the transition probabilities on the most serious
crime, based on either the average or median sentence for each crime type. Table 4 shows that both measures generate similar transition rates. The large probabilities on the main diagonal show that criminals
tend to recommit the same types of crimes.

3
3.1

Model
Baseline Model

Consider the decision problem of an individual i who has been pardoned and released from prison after
committing a crime of type j. Given the conditions of the collective pardon described in Section 2, if
this person commits another crime of type k and is caught, their total prison sentence, denoted tsik , will
be equal to the new sentence, nsik , plus a residual sentence rsi , where tsik = nsik + rsi . Each individual
decides whether or not to recidivate, and if so, which crime to commit. For now let us assume that j = k;
the new crime is the same as the original crime. We will relax this assumption in Section 3.2.
Let ci j = 1 denote that person i commits a crime of type j and ci j = 0 otherwise. We normalize the
utility associated with ci j = 0 to be equal to 0. The net expected utility associated with committing a

8

 crime is given by:
 
    
 
V (ci j = 1) = u Xiuj + Pjlc D Xidj ,tsi j , δ Xiδj + εi j .

(1)

The functions u and D capture the utility of committing a crime and the (dis)utility of being caught,
sentenced to prison, and serving that time in prison.8 Xiuj denotes observed characteristics of the utility
received from committing a crime that are related to the individual and the crime. Xidj denotes observed
characteristics of the disutility of going to prison. Xiδj are observable characteristics that are related
to differences in discount factors. Pjlc is the probability of being caught and sentenced to prison for
committing crime j in location l, and εi j captures unobserved drivers of the crime decision. δ is the
discount factor, and the main object of interest here.
   
We assume that each year in prison leads to a constant flow disutility9 of d Xidj and discounting is
T −1 t
exponential.10 Given that ∑t=0
δ =

1−δ T
1−δ

, we have that11


   tsi j 
    1 − δ Xiδj
 
    

    
= d Xidj 
D Xidj ,tsi j , δ Xiδj
.
1 − δ Xiδj
   
   
Letting u Xiuj = α0 + α1 Xiuj and d Xidj = β0 + β1 Xidj , the expected (net) utility of crime can be written
as:


   tsi j 
h
i
1
−
δ
Xiδj
u
d
c
    
V (ci j = 1) = α0 + α1 Xi j + β0 + β1 Xi j Pjl 
 + εi j .
1 − δ Xiδj

(2)

The probability that a crime of type j is committed by individual i is then given by

Pr (ci j = 1) = Pr [V (ci j = 1) > 0] ,

which is equal to
   tsi j 
h
i
1 − δ Xiδj



u
d
c
    
Pr εi j > − α0 + α1 Xi j + β0 + β1 Xi j Pjl 
 .
δ
1 − δ Xi j






8 In

Appendix A, we show that the model for utility in equation (1) can be obtained from a more general dynamic utility
model through a restriction on the option value of future crimes.
9 This disutility includes the pain and suffering due to being imprisoned, the lack of freedom, as well as the opportunity cost
of spending time in prison. We discuss this in more detail in Section 4.
10 In Section 5.3 we consider alternative forms of discounting such as hyperbolic discounting.
11 The equation below implicitly assumes that the first month in prison is served immediately (t = 0). Our results are not
sensitive to this.

9

 This equation is the basis for our estimates in Section 5.

3.2

Allowing for the New Crime to be Different from the Old Crime

We now consider the case in which the new crime is not of the same type as the original crime (i.e.,
j 6= k). Let ci jk = 1 denote that individual i, who previously committed a crime of type j, commits a
crime of type k after being released from prison. The net utility associated with this choice is

 

V ci jk = 1 =

α0 + α1 Xiku +

"
  #
h
i
δ tsik
d
c 1 − δ Xik
  + εik ,
β0 + β1 Xik Pkl
1 − δ Xikδ

where tsik = nsik + rsi . The probability of ci jk = 1 is then given by
 
 
 
 
Pr ci jk = 1 = Pr V ci jk > 0 .

Recall that we only observe whether an individual is caught committing another crime and the day
on which that occurred. We do not observe the type of the new crime type (e.g., violent, property, etc.).
However, even though such information might be of use, it would still be unobserved for inmates who
do not recidivate.
With a slight abuse of notation, we can denote the probability that individual i, who previously
committed crime j, commits another crime of any type as an integral over all possible future crime
choices k
 
Pr (ci j = 1) = ∑ Pr ci jk = 1 Pr (k   j) ,
k

where Pr (k   j) are the probabilities of transitioning from crime type j to crime type k.
In the data, we directly observe the residual sentence rsi and the original sentence osi j . However,
we do not observe the new sentence that an individual would receive for committing a new crime: nsik .
The new sentence is likely to be a function of both the individual i and the crime type k. Therefore, we
assume that the new sentences can be decomposed as nsik = θi ask , where ask is the average sentence for
a crime of type k, and θi is an individual-specific component of severity, reflecting violent attitudes of the
offender, as well as any possible systematic sentencing differences based on characteristics of offenders
that are observable to the judges. We can compute ask =

1
Ik

I

k
osik from our data on original sentences,
∑i=1

where Ik is the total number of people who committed crime k. We can then compute the individual

10

 component as θi =

osi j
as j .

This implies that the new sentence for each possible crime choice k for a given

individual i who previously committed crime j can be constructed as

nsik =

osi j
ask .
as j

(3)

In the baseline model for which j = k, this implies that nsi j = osi j .

3.3

Relationship to Deterrence Models Without Discounting

The primary focus of our paper is estimating criminal discount factors. However, the model described
above also has implications for the literature that attempts to measure the deterrent effect of imprisonment. In order to make this comparison clear, we will compare to the model used by Drago, Galbiati,
and Vertova (2009), which uses a dataset very similar to ours to estimate the magnitude of deterrence,
but the points we make apply more generally.
h

For simplicity, if we drop the characteristics X, and express

1−δ tsi j
1−δ

i

as

h

1−δ osi j
1−δ

+δ

osi j
(1−δ rsi )

1−δ

i
, we

can re-write equation (2) as:

V (ci j = 1) =

α0 + β0 Pjlc

 

1 − δ osi j
1−δ

 

+ β0 Pjlc

 

 
δ osi j (1 − δ rsi )
+ εi j .
1−δ

(4)

This equation resembles the baseline equation that Drago, Galbiati, and Vertova (2009) estimate, which
is given by the following:

yi = α + β0 osi +β1 rsi +εi ,

(5)

where yi is a binary variable for whether individual i commits another crime after being released and εi
is assumed to follow a logistic distribution.12
There are two key differences between equation (4) and equation (5). First, the original sentence
os and the residual sentence rs enter (4) non-linearly. Second, rs does not enter equation (4) separately
from os. The importance of this is that, in Drago, Galbiati, and Vertova (2009), the effect of the residual
sentence does not depend (directly) on the length of the original sentence.13 Put differently, the first year
12 Drago, Galbiati, and Vertova (2009) also estimate specifications which include characteristics of the individual and crime
fixed effects.
13 Because they estimate the model using a logit specification, this is not exactly true. Since the logit model is non-linear, the

11

 of the sentence is assumed to have the same effect as the last year. This is a general feature of most
models of deterrence in which the effect of additional years of imprisonment is not allowed to vary based
on the length of the sentence onto which they are added.
In equation (4), however, the effect of rs depends directly on os (recall that ns = os in our baseline
setup). The intuition for this is straightforward. If an individual is caught committing another crime, the
offender will receive a new sentence ns. In addition to this, the sentence will be extended by rs. In other
words, after the ns months are served, the individual will need to serve an additional rs months. As long
as δ < 1, this distinction will be important. When δ < 1, individuals discount the future, and the effect
of an additional rs months in prison will depend on when that extra time is served, which depends on ns.
The larger ns is, the further into the future the rs months are served, and the lower the effect on current
behavior.
By L’Hopital’s Rule, limδ →1 δ

osi (1−δ rsi )

1−δ

os

i
= rsi and limδ →1 1−δ
1−δ = osi . Therefore, when δ = 1, equa-

tion (4) becomes

V (ci j = 1) = α0 + β0 Pjlc (osi ) + β0 Pjlc (rsi ) + εi j ,

which matches the form of the regression equation in Drago, Galbiati, and Vertova (2009). Thus their
model can be seen as implicitly assuming that the discount factor is equal to 1. Since discount factors less
than 1 generate marginal deterrent effects of imprisonment that are decreasing in sentence length, this
implies that the estimate of deterrence in Drago, Galbiati, and Vertova (2009) is some average measure
of deterrence across sentence lengths. This may help to explain one of their results. They find that for
individuals with the longest original sentences, the effect of the residual sentence is negligible (pgs. 275276). They attribute this to the fact that long sentences are given for serious crimes, and they suggest that
more dangerous inmates (those that commit more serious crimes) are not deterred by prison. However,
given the derivations above, an alternative explanation is that these individuals, if caught again, are likely
to face a lengthy new sentence. The residual sentence will be served after that, and therefore far into the
future. If individuals discount the future, then the residual sentence will have a low effect on utility, and
thus it will not have much of an effect on decisions to recidivate.14
original sentence length will impact the effect of the residual sentence, but in the same way as any other covariate included in
the regression. The original sentence length also has this effect in equation (4), but has an additional effect via its interaction
with the residual sentence.
14 More generally, our model predicts that when the discount factor is less than 1, the effect of the residual sentence on
recidivism should be decreasing in the length of the original sentence, where recall that the original sentence is our proxy for

12

 This analysis has important implications for studies of deterrence. To the extent that punishments are
applied in the future (as is the case with imprisonment), failing to account for it might mask some important heterogeneity in deterrent effects. For example, sentence enhancements or other policies designed to
lengthen sentences for very serious offenses (where baseline sentences tend to be long) will potentially
have very small deterrent effects on crime.

4

Identification and Estimation

 
 
Under the assumption that εi j is independent of Xiuj , Xidj , Pjlc ,tsi j , one could estimate the model described by equation (2) via maximum likelihood. However, as stated earlier, the total sentence tsi j =
nsi j + rsi is potentially correlated with the error εi j . There may be characteristics of the individual, which
make that person more likely to commit crimes, that are unobserved to the econometrician, but observed
by judges. As a result, individuals with higher εi j ’s may also receive longer sentences, which generates a
correlation between εi j and nsi j (and osi j ). This also generates a correlation between εi j and rsi , since individuals with longer original sentences are likely to have longer residual sentences as well. Fortunately,
it is still possible to consistently estimate the parameters of the model. The key identifying assumption is
that conditional on the original sentence received osi j , the total sentence tsi j is independent of the error.
This conditional independence assumption is based on the fact that the timing of the release is dictated
by an (unanticipated) mass prison release. Therefore, the only source of endogeneity in the residual sentence rsi is via the original sentence length. Once we condition on the length of the original sentence, the
residual sentence is independent of the error.15 Since nsik =

osi j
as j ask

(or nsi j = osi j in the baseline model),

controlling for the original sentence also controls for the endogeneity of the new sentence, since as j and
ask do not vary by individual.
To summarize, the error in equation (2) is assumed to be correlated with the original sentence, but
conditional on the original sentence, independent of the residual sentence. Therefore, as we discuss in
more detail in Appendix B, we can control for the endogeneity by expressing ε as a function of the
original sentence:
εi j = h (osi j ; γ) + ui j ,
the new sentence. Although they do not highlight this, this is exactly what Drago, Galbiati, and Vertova (2009) find (see their
Table 4).
15 See Drago, Galbiati, and Vertova (2009), which employs a similar assumption in order to estimate the deterrent effect of
imprisonment, for a detailed justification of this assumption.

13

 where h is a flexible function of os, and u is is assumed to be independent of os. By replacing this
expression in to equation (2), we have the following equation:

   tsi j 
h
i
1
−
δ
Xiδj
u
d
c
    
V (ci j = 1) = α0 + α1 Xi j + β0 + β1 Xi j Pjl 
 + h (osi j ; γ) + ui j ,
1 − δ Xiδj

(6)

 
 
where u is independent of Xiuj , Xidj , Pjlc ,tsi j , osi j . The additional term h (osi j ; γ) controls for the dependence between the error and the total sentence tsi j that would otherwise bias our estimates.16
Intuitively, identification of the discount factor is obtained by measuring how each additional month
of a prison sentence deters future crimes, conditional on the total length of the sentence. If the discount
factor was equal to 0, then variation in the residual sentence would have no impact on the probability
of recidivism, as individuals would place no value on future prison sentences. As the discount factor
increases towards 1, each additional month of residual sentence should have an increasing (negative)
effect on the probability of recidivism. The rate at which the marginal deterrent effect of imprisonment
changes is what identifies the discount factor.
A key identifying assumption is that the flow disutility from prison does not depend on the sentence
length.17 If this were not the case, and the per-period flow disutility decreased by a constant fraction
each period (mimicking the discount factor), then we would not be able to separately identify the discount factor from the rate of decay of the disutility of prison. We are not aware of any estimates in the
economics literature on the relationship between prison time and the per-period disutility of imprisonment. However, studies in criminology and psychiatry have found no evidence of a correlation between
prison time and various measures of prison well-being, e.g., subjective quality of life, depression, anxiety, post-traumatic stress symptoms (see Bukstel and Kilmann (1980); Hochstetler, Murphy, and Simons
(2004); Gullone, Jones, and Cummins (2000)). Furthermore, given our baseline estimates of the discount factor in Section 5 below, the value of the 15th year in prison is only about 1% of the value of the
first year in prison. If instead of being due to discounting, this was purely a result of a decline in the
flow disutility of imprisonment over time, then this would suggest that criminals are close to indifferent
between spending that 15th year in prison versus being released, which we believe is unlikely to be the
case.
While we believe that the assumption of a constant flow disutility of prison is a reasonable one it may
16 Essentially
17 This

 
h osi j ; γ acts as a control function (Heckman and Robb (1985)).
assumption is also employed in Lee and McCrary (2009).

14

 not even be necessary from a policy perspective. A more general interpretation of our results would be
that of decreasing marginal returns to additional years of imprisonment. The policy-maker cares about
the deterrent effect of various sentence lengths when designing policies to reduce crime, and may not
care if the source of diminishing marginal returns to imprisonment come from discounting or from a
decreasing negative value placed on subsequent periods of punishment.
Drago, Galbiati, and Vertova (2009) find that each additional month of residual sentence decreases
the probability of recidivism, which they interpret as evidence of a deterrent effect of imprisonment.
An alternative explanation that has been offered by Philip Cook (see Durlauf and Nagin (2011)) is that
each additional month of imprisonment (and therefore one less month of residual sentence) increases
a prisoner’s criminal capital, which leads to a higher recidivism rate. Since both theories generate a
negative relationship between residual sentence and the propensity to recommit crime, it is difficult to
disentangle the two mechanisms. However, by considering the relationship between discounting and
deterrence, we show that it is possible to distinguish between them.
Under the theory of deterrence, as long as discount factors are strictly below 1, the first months of
the residual sentence should affect behavior the most, since they are discounted the least. Alternatively,
under the theory that being imprisoned is a treatment that increases the propensity to commit crime, one
would expect that the first months of imprisonment have the strongest effects, as inmates acquire the
most knowledge capital initially (i.e., decreasing marginal returns).
We illustrate this idea in Figure 1, for a hypothetical original sentence of 3 years. Since the amount
of time served in prison and the residual sentence are negatively correlated, this implies that last few
months of the residual sentence, which correspond to the earliest months of imprisonment, should have
the largest impact on behavior under the treatment hypothesis, and the opposite under the deterrence
hypothesis. In the left panel, we have plotted the hypothetical marginal effects of an additional month of
residual sentence under both the deterrence and treatment mechanisms. For deterrence, we have assumed
a discount factor of 74%, and for treatment we have assumed that the value of treatment decreases
exponentially at a rate of 74%, (i.e., the criminal capital acquired in prison next year is 74% of the current
year). In the right panel, we have plotted the cumulative effects. The first thing to note about the right
panel is that both mechanisms generate a negative relationship between residual sentence and recidivism,
which is what makes it challenging to distinguish between the two theories. The key difference is in the
shape of the relationship, which allows us to disentangle them. For deterrence it is convex, while for

15

 treatment it is concave.
Intuitively, the shape of the relationship between the (dis)utility of imprisonment (which drives recidivism in the data) and sentence length is what identifies the discount factor in our model. We have
illustrated this in Figure 2, by graphing the cumulative utility from prison as a function of sentence
length for a discount factor of 0.74. The more convex the relationship between residual sentence length
and the probability of recidivism, the larger the degree of discounting that is occurring. In the case of
no discounting (discount factor of 1), this relationship would be linear. Since we find estimates of the
discount factor of around 0.74, this indicates that this relationship in the data is in fact convex. Under
the assumption of decreasing marginal returns to criminal capital accumulation, the relationship between
residual sentence and recidivism is concave. In order for the treatment hypothesis to rationalize the data,
it would need to be the case that the criminal capital returns to being in prison are convex over time. This
would imply that individuals learn very little initially while in prison, but that this learning accelerates
over time, with the most criminal capital being gained in the very last part of incarceration, a scenario
which we believe is less likely. Overall, this insight connecting deterrence and discounting helps us to
separately identify the two hypotheses, and lends support to the deterrence hypothesis.

5

Empirical Results

Before we present the estimates from our model, we first provide some evidence from the raw data that
illustrates how the discount factor is identified. In Table 5, we present linear probability model estimates
of recidivism on residual sentences, conditional on the inmates’ original sentence, for various bins of the
original sentence. If people did not care at all about the future, in other words had discount factors of 0,
then the length of the residual sentence would not matter for recidivism, and we would expect all of the
coefficients on residual sentence to be zero. If they had discount factors of 1, meaning that they were not
discounting the future at all, the coefficient would be negative, but it would be the same no matter the
original sentence. If the discount factor was between 0 and 1, then the coefficient should be decreasing
in absolute value towards 0 with the original sentence, where recall that the original sentence proxies for
the new sentence.
We do indeed find that the group with the shortest original sentences (below one year) has the steepest
marginal reduction in recidivism: as one increases the residual sentence by one month, recidivism drops
by 1.46 percentage points (pp). The second steepest slope coefficient is found for inmates with the next
16

 longest sentences (between one and two years) with a slope indicating a drop in recidivism by 0.50
pp, and so on. For the last group of inmates, those with sentences that are longer than six years, the
regression line is almost flat. The slope for the overall sample of -0.20 pp represents a weighted average
of the various slopes and is, when considering the longer recidivism window, similar in size compared to
the estimated marginal effect of imprisonment of -0.16 pp found in Drago, Galbiati, and Vertova (2009).
The coefficient on residual sentence for the shortest sentences is almost an order of magnitude larger
than the overall value. This indicates that the relationship between the probability of recidivism and
residual sentence is substantially weakening as the original sentence length increases and is consistent
with the presence of non-trivial amounts of discounting behavior by criminals.
As a more direct way to see how the discount factor is identified in the data, we provide some nonparametric plots of the relationship between recidivism and total expected sentence length. Specifically,
we compute the average recidivism rate for each of 40 quantiles of sentence length, and plot them in the
left panel of Figure 3.18 The figure illustrates that as the total sentence length increases, the recidivism
rate decreases, which is consistent with deterrence. Furthermore, the relationship is convex, suggesting
that the marginal deterrent effect of imprisonment is decreasing, consistent with discounting. In the right
panel, we add original sentence as a control in order to deal with the endogeneity of the total expected
sentence.19 To do this, we regress the recidivism indicator on a flexible function of the original sentence
(a cubic polynomial in logs) and dummies for each of the quantiles of sentence length.20 We then plot the
relative coefficients on each of the dummies. There are two key changes in the figure. First, the overall
magnitude of deterrence increases by a factor of almost four. Since sentence length is likely correlated
with unobservable drivers of recidivism, we would expect the effect of increased sentence length to be
biased towards zero. Once we add controls for original sentence to correct for this bias, we see much
larger estimated deterrent effects. The second difference is that the predicted recidivism rates for each
bin of sentence length are much more tightly concentrated along a curve that closely resembles Figure 2.
The curvature of this line is what pins down the discount factor.
Overall, this descriptive evidence suggests non-trivial amounts of discounting that are apparent even
in the raw data on recidivism and sentence length. The tight relationship between recidivism and sentence
length in Figure 3 foreshadows the high degree of statistical precision in our estimates, as well as the
18 We

trim the top and bottom 1% of total sentence length to make the figure easier to see.
on the original sentence, the remaining variation in total sentence is due to the exogenous variation in residual
sentence.
20 Using splines and dummy variables based on the months of original sentence length led to very similar results.
19 Conditional

17

 robustness of our results to alternative model specifications. In order to estimate the specific value of the
discount factor, we now turn to the estimates from our model.

5.1

Baseline Results

We estimate our baseline specification described in equation (6) by maximum likelihood, using a logistic
distribution for the error u, and we model h (os) using a cubic polynomial in the logarithm of the original
sentence.21
In Table 6, we present estimates of the baseline model in equation (6). In our main dataset we do
   
not observe the probability of being caught Pjlc , and in our baseline estimates we assume that this is a
constant. In Section 5.3 below, we bring in some additional external data on clearance rates to capture
this probability, and find that our results are qualitatively unaffected.
We start in column 1 with the simplest specification that includes no covariates. In other words, we
investigate the effect of the residual sentence on recidivism, conditioning on only the original sentence
length to control for the endogeneity of the residual sentence. Our estimates imply an annual discount
factor of 0.71.22 In columns 2-5 we add controls to the estimating equation to allow for the utility from
crime commission and the disutility of imprisonment to vary across individuals. In column 2, we add
controls for age, permanent employment status, gender, marital status, whether the inmate is Italian,
whether the original sentence is definitive (all appeals have been exhausted), and binary indicators for
low education and high education. We define low education as having at most a primary education
(corresponding to less than 5 years of education), and we define high education as an education at or
beyond secondary school (corresponding to at least 13 years of education).
The first thing to note is that the estimated discount factor is relatively unchanged, having increased
slightly to 0.74. Age is associated with lower utility of crime commission, and a higher disutility of
going to prison (although it is not precisely estimated in the first term). One explanation for this is that
as people age, the expected benefit from committing crime is lower, perhaps due to a lower productivity
of the individual in criminal activity, and serving time in prison is more unpleasant the older someone is.
Having a definitive sentence is associated with a higher disutility of prison. This is intuitive given that
those without firm sentences face some chance of having their residual sentence commuted or shortened
21 Model

selection criteria show that this is the preferred specification for the original sentence. We also tried using splines
and dummy variables based on the months of original sentence length, which led to very similar results.
22 If δ is the monthly discount factor, then δ = δ 12 is the annual one.
A

18

 upon appeal. Being Italian is not associated with the utility from committing a crime, but is associated
with a lower disutility of prison. This suggests that going to prison in Italy is worse for foreigners than for
Italians. Since Italians have a higher chance of receiving alternative sentences such as home arrest and
parole, this makes sense (e.g., according to ISTAT (2015), in 2013, foreigners represented 35 percent of
the prison population but only 13 percent of the population receiving alternative sentences). Furthermore,
in Italy, inmates are required by law to be imprisoned in the prison closest to their home, so that family
can most easily visit. However, since many of the inmates that are foreign are illegal immigrants23 ,
by definition they have no official home, which means that they are likely to be imprisoned further from
nearby friends and family. We also find that being married is associated with a lower utility of committing
a crime and a lower disutility of going to prison. We find that gender, permanent employment prior to
the original incarceration, and the level of education of the individual are not associated with either
component of utility.
In columns 3 and 4 of Table 6, we estimate the model using alternative definitions of low education
and high education. We first adjust low education to include only individuals with no completed education (column 3). In column 4, we re-define high education to be a high school degree or higher.24 These
changes have very little effect on the other estimates, including the discount factor. The point estimates
on the indicators for low and high education do change a bit, but none of the estimates are statistically
different from zero. The differences between these education measures do matter below, however, when
we interact them with the discount factors in Section 5.2.
In column 5 we add dummies for the various crime categories, with the idea that the returns to
committing crimes and/or the opportunity cost of going to prison might vary by crime type. The other
coefficients are largely unchanged, and as was the case in the previous results, the estimated discount
factor is relatively unaffected, with a point estimate of 0.70.
Overall, the results in Table 6 indicate that the (annual) discount factor among our sample of previous
criminal offenders is between 0.70 and 0.74. In line with the non-parametric evidence shown in the
second panel of Figure 3, the discount factor is precisely estimated. The 95% confidence interval for
our baseline result in column 2 of is [0.65,0.83], and thus we can easily reject the hypothesis that the
discount factor is either 0 or 1.25
23 In

Italy, 70% of foreign inmates are illegal immigrants (see Italian Ministry of Internal Affairs (2007)).
to our original measure of high education we exclude individuals who went to vocational school.
25 Standard errors, and therefore confidence intervals, for the annual discount factors can be computed via the delta method.
24 Compared

19

 In order to illustrate how discounting affects the relationship between utility and imprisonment, in
Figure 4, we have graphed the present discounted flow disutility of prison as a function of sentence
length, for various values of the discount factor. We normalize the disutility to 100 in period 0. Under
no discounting (discount factor of 1), the relationship would be a straight horizontal line at 100. As the
figure illustrates, even small changes in the discount factor lead to large relative changes in utility, and
therefore behavior, which is what allows us to obtain precise estimates of the discount factor.
Our estimates imply that the disutility of an additional year of imprisonment is 74% of the previous
year. This indicates a fairly large degree of discounting by criminal offenders, but not so much that they
act myopically and ignore any future consequences of their actions. Relative to the first year in prison,
the marginal discounted annual disutility of prison drops to 30%, 7%, and 1% after 5, 10, and 15 years,
respectively. The effect diminishes over time at a reasonably fast rate, but not so fast that the difference
between short- and medium-term sentences becomes meaningless from a deterrence perspective. In other
words, there is scope for a deterrent effect of imprisonment.
For a more traditional discount factor of 0.95, the discounted marginal disutility of imprisonment
falls to 50% of its original value after 15 years. If we compare to smaller discount factors—at values of
0.10 and 0.30—the effect drops to below 1% after only 3 and 5 years, respectively, reflecting negligible
marginal deterrent effects of imprisonment for sentences of more than a few years. This figure helps to
illustrate that our estimates of the criminal discount factor are quite reasonable. At a discount factor of
0.74, there is some drop-off in the deterrent effect, but it does not vanish immediately. If discount factors
were much smaller, then this would imply almost no deterrent effect of imprisonment beyond the first
year or two.
Estimates of the individual discount factor for the general population, based both on lab experiments
(e.g., Coller and Williams (1999); Harrison, Lau, and Williams (2002); Andersen et al. (2014)) and
observational studies (e.g., Viscusi and Moore (1989); Warner and Pleeter (2001); Cagetti (2003)), vary
considerably in the literature, ranging from values close to 0 to close to 1. However, the majority of the
estimates range from discount factors of 0.7 to close to 1. Our estimates are on the low end of those
found in the literature, particularly compared to those from other observational or non-experimental
studies. This is consistent with the idea that those individuals who decide to commit crimes have lower
discount factors than people in the population at large, and suggests that low future time preference could
be an important driver of criminal behavior.

20

 The only other paper that we are aware of that attempts to elicit individual discount factors for
criminals is Lee and McCrary (2009). Using detailed panel data on arrests in Florida, they exploit the
discontinuity in the severity of punishment at the age of majority (18). Despite large increases in the
average severity of punishment at the age of 18, they find very small drops in the probability of arrest.
Using a dynamic extension to Becker’s (1968) model, they map these estimates into plausible values of
the discount factor, and find that the youth they study are essentially myopic. Specifically, they find that
their point estimates are inconsistent with annual discount factors larger than 0.022, although they note
that they cannot statistically rule out larger ones.
There are several factors that likely contribute to the difference between our estimates and those in
Lee and McCrary (2009). First, it is likely that the effect of the policy in our sample is much more
salient. Individuals in our dataset are older, with an average age of 38, compared to 18 in Lee and
McCrary (2009), and therefore likely to be more familiar with the criminal justice system, particularly
the adult system.26 Individuals in our sample are also quite well-informed about the future consequences
of their actions, both in terms of the legal implications (a longer sentence) and the costs associated with
being imprisoned. Tied to their early prison release, which should be apparent to them, is the caveat of
having to serve the residual sentence if they are convicted again. Furthermore, the massive scale of the
pardon and the homogeneity of the policy across individuals (with the important exception of the length
of the residual sentence) should have served to increase the flow of information regarding the pardon
across prisoners.
Second, given the age differences in our samples, it is possible that younger people act more myopically (particularly those who decide to commit criminal acts), perhaps due to problems with self-control
or impulsivity (Wilson and Herrnstein (1985); Jolliffe and Farrington (2009)). It is difficult for us to
compare directly, as no one in our sample is less than 19 years of age, and fewer than 1% are below the
age of 22.

5.2

Heterogeneity in the Discount Factor

In order to investigate whether there are any important differences in discounting behavior across individuals we also allow for discount factors to vary across individuals based on observables. We first
allow the discount factor to vary based on characteristics of the inmates, such as gender, age, education,
26 Using NLSY data for the US, Hjalmarsson (2009) finds that youth perceptions of the change in the sanctions at the age of
majority (age 18) are smaller than the true changes observed in the data.

21

 nationality, etc. Second, we allow the discount factor to vary based on the category of the original crime
committed (e.g., violent, property, drugs). This analysis not only provides a description of how discount
factors vary (or do not vary) across individuals, but also provides a check on our results, as it facilitates
comparison to other results in the literature regarding individual discount factors.
We first estimate the model allowing for the discount factors to depend on individual characteristics,
by interacting the discount factor δ with indicator variables for marital status, gender, low education, high
education, employment status, as well as age and measures of nationality. These results are presented
in Tables 7 and 8. In columns 1 and 2 of Table 7, the point estimates suggest higher annual discount
factors for inmates who are married and for women (6 percentage points (pp) and 9 pp, respectively).
Although the differences are not precisely estimated, our results with respect to gender are consistent
with the literature (see e.g., Cagetti (2003); Coller and Williams (1999)). We are not aware of any results
in the literature regarding a systematic relationship between discount factors and marital status. We find
little evidence of differences based on whether or not an individual was permanently employed at the
time imprisonment prior to the pardon.
We find evidence that discount factors are decreasing with age. In columns 4 and 5 we report results
in which we allow the discount factor to depend on age in a log-linear27 and quadratic fashion, respectively.28 In both cases we find evidence that discount factors decrease in age at a decreasing rate. For the
log-linear specification the discount factor decreases from 0.78 at age 26 (the 10th-percentile of the age
distribution) to 0.68 at age 51 (the 90th percentile). With the quadratic model, the discount factor drops
from 0.76 at age 26 to 0.68 at age 51. Overall there seems to be evidence of a somewhat steep decline in
discount factors early on, followed by a flattening out. There is no consensus in the literature on the relationship between time preference and age, with some papers finding that discount factors decrease with
age (e.g., Meier and Sprenger (2010)) and others finding the opposite (e.g., Warner and Pleeter (2001)).
One potential explanation for our finding of a negative relationship in our dataset is selection. To the
extent that low discount factors are an important driver of crime, individuals who are still committing
crimes later in life are likely those who have the lowest discount factors.
For education, we consider two definitions of both low and high education, following the definitions
used previously. For low education (columns 6 and 7), the differences in discount factors for both definitions are small (less than 2%) and statistically insignificant, although the signs are what we would expect
27 Specifically
28 We

we interact the discount factor with the natural log of the age of the offender minus 18 years.
also tried interacting the discount factor with dummies based on different age bins, and found similar results.

22

 (lower educational attainment is associated with lower discount factors). One likely explanation is that
the average education level of individuals in our dataset is quite low. Therefore distinguishing between
low and very low education levels has negligible effects on the estimates of the discount factor.
In contrast to the measures of low education, we do find sizable differences for high education. The
results in columns 8 and 9 indicate that having a higher education is associated with significantly larger
discount factors. For individuals with an education of at least 13 years, the estimated discount factor is
0.89 compared to the rest of the released inmates at 0.73. When we exclude from the high education
measure those that attended vocational school, the difference is even larger, with an estimated discount
factor of 0.99 for those with high education, a difference of 26 percentage points. While less than 2% of
the sample have this high of a level of education, the difference is very precisely estimated, suggesting
that particularly high levels of education are strongly related to discount factors. Given the nature of
our data, we are unable to determine whether education actually increases discount factors (Becker and
Mulligan (1997); Lochner and Moretti (2004)) or whether people with high discount factors are more
likely to invest in education. Regardless of the explanation, consistent with the literature, we find strong
evidence of a positive relationship between (high) education and discount rates (Viscusi and Moore
(1989); Warner and Pleeter (2001); Cagetti (2003)).
The last set of individual characteristics that we analyze are based on nationality. Recall that we
observe the country of origin for each individual. We first estimate a specification in which we allow
the discount factor to depend on whether the person is Italian or not. The results in column 1 of Table
8 indicate a discount factor of 0.78 for Italians compared to 0.66 for non-natives. This result could be
driven by a number of factors. First, it could be the case that the discount rates of the populations from
which the immigrants originate are lower than in Italy. It could also be the case that those individuals
who both immigrate and commit crimes have particularly low discount factors.
Salience might be important here as well. Immigrants may be less familiar with the Italian justice
system and they may not speak Italian, both of which could lead to a failure to fully recognize the
consequences of their future criminal activities, and therefore lower estimated discount factors in the
data. Regardless of the explanation, one implication of our estimates is that foreigners are less likely to
be deterred by imprisonment, and particularly so for longer sentences.
In addition to examining differences between Italian natives and immigrants, we examine how differences in criminal discount factors vary among immigrants across country of origin. A recent literature

23

 has emerged documenting an empirical cross-country relationship between measures of time preference
and many aggregate behaviors and outcomes such as national income, investment in human and physical
capital, and savings rates (e.g., Galor and Özak (2014) and Dohmen et al. (2015)). In a recent attempt
to determine the cause of these underlying differences in time preference, Chen (2013) examines differences across countries based on the extent to which their languages distinguish between current and
future events, which linguists refer to as future time reference (FTR). His hypothesis is that speaking in
a way that places a larger distinction between current and future events causes people to value the future
less. In support of this, Chen finds a negative relationship between FTR and future-oriented behaviors
such as savings and various health-related activities. In other words, in countries in which the language
places a stronger distinction between the current and future, individuals exhibit less forward-looking
behavior on average.29
In order to investigate the presence of differences in discounting across immigrants based on country
of origin, we merged in the data on FTR from Chen’s paper. We matched the primary languages spoken
in each country30 to the country of origin for each individual. Each language is coded by linguists
as having either strong FTR or weak FTR. If all major languages spoken had either strong or weak
references, we classified the country accordingly. For example, the primary language in Italy is Italian,
which is classified as a strong FTR language. If some languages spoken had strong FTR and others
weak FTR, then we classified the country as “both”. For example, in Belgium, Flemish (weak FTR) and
French (strong FTR) are the primary languages.
We estimate two specifications corresponding to different ways to model countries having FTR of
“both”. The results from these models are reported in columns 2 and 3 in Table 8. In each specification
we also include an indicator for whether an individual is an Italian native or not, to allow for there to be
differences in discount factors between native Italians and immigrants from other countries with strong
FTR. In column 2, for countries with a FTR of “both”, we assign a value of 0.5 to the strong-FTR
indicator, and in column 3, we drop these observations. In line with the results in column 1, we find that
native Italians have the largest discount factors. Our results are also consistent with those in Chen (2013),
as we find that discount factors are strongly correlated with FTR. Immigrants from countries with weak
FTR have significantly higher discount factors than those from countries with strong FTR (12 pp - 14 pp
depending on the specification).
29 See

also Falk et al. (2015) who find a systematic relationship between patience and FTR.
information was collected from the CIA World Factbook.

30 This

24

 An alternative interpretation of these results is that linguistic differences across countries could be
proxying for other cultural factors that lead to differences in time preference. We do not take a stand
on the particular mechanism at work. Instead, we emphasize that our results indicate that the patterns
across country of origin in discounting behavior among criminals in our data are consistent with other
forward-looking behaviors for the population at large. This gives additional credibility to our results and
suggests that they reflect patterns in the general population as well.
Lastly, in addition to examining differences in time preference based on individual characteristics,
we also examine whether there are differences across crime types. In Table 9, we report results from
10 different specifications (one for each crime category) in which we interact the discount factor with
an indicator for whether the person was originally imprisoned for a crime of that type. For most crime
categories, the estimated differences in discount factors are small, both statistically and economically.
However, for two crime categories we find large statistically significant differences: prostitution and
drug-related.
For prostitution crimes, we find a discount factor of 0.92 compared to 0.72 for all other crimes. In
Italy, prostitution is legal, but organized prostitution is not. Therefore prostitution crimes in our data are
those relating to the supply of prostitutes. It is interesting that those individuals participating in this type
of crime have discount factors closer to what we typically assign to economic agents, as these crimes
involve running a business (albeit an illegal one), as opposed to the theft of property or acts of violence.31
We find discount factors that are 10 percentage points lower for violations of the law regarding the
use and selling of drugs. Selection into the use of drugs by a pool of more myopic individuals might
explain this result. It may also be the case that drug usage directly affects time preferences and/or
decision making.

5.3

Robustness Checks

In this section we discuss a number of robustness checks to our baseline specification. In particular,
we control for differences in clearance rates and allow for individuals to transition to different crimes
following the model described in Section 3.2. We also examine the impact of the fact that only individuals
who receive future sentences of at least two years are subject to serving their residual sentence, and we
explore the evidence for hyperbolic discounting behavior.
31 Of

course this is not to say that such crimes do not involve violence, but rather that violence is not the primary activity.

25

 5.3.1

Clearance Rates

Our main dataset does not contain any information on clearance rates. In order to determine whether
variation in clearance rates, across either location or crime type, could affect our estimates of discounting,
we perform two sets of robustness checks. In the first set we attempt to control for regional differences
in clearance rates using location-specific and crime-specific dummy variables. The results are reported
in columns 1-4 of Table 10. In the first column we model the clearance rates using a set of three dummy
variables, corresponding to each of the three areas of Italy (North, Center, and South). In column 2 we
add crime-type dummies. In columns 3 and 4 we repeat the exercise using region dummies (there are
20 regions in Italy). The coefficients on the dummy variables are consistent with clearance rates being
the highest in the North, and the estimated discount factors, which range from 0.70-0.74, are extremely
similar to our baseline estimate.
We also collected separate data by province on both the total number of crimes and the number of
crimes in which the perpetrator is known, using data from ISTAT (2005), to compute clearance rates. It
is important to note that identifying a suspect and clearing the crime via arrest are not necessarily the
same, and therefore these data serve only as proxies for the true clearance rates. Quite surprisingly, the
richest regions, e.g. Lombardy, Piedmont, Veneto, Lazio and Liguria appear to have, by far, the lowest
average clearance rates. One possible explanation is that reporting rates of crimes are different across
regions. The criminal justice system is widely recognized to function less efficiently in the South of
Italy (see Jappelli, Pagano, and Bianco (2005)), which could lead to people being less willing to report
crime32 , and thus inflating the clearance rates. We are therefore cautious not to rely too heavily on these
estimates.33 We first estimate our baseline model using the province-specific average clearance rate, and
then estimate a version using the province and crime-specific clearance rate. In columns 5 and 6 we
report estimates from the models using the clearance rate data. The point estimates of the discount factor
increase slightly (to 0.81) in these specifications, but so does the standard error, and the resulting confidence intervals almost completely cover that of the baseline model. Overall, we conclude that accounting
for heterogeneity in clearance rates has at most a small impact on our estimates of discounting.
32 For most property crimes, the most recent victimization survey shows that reporting rates are approximately 10 percentage
points lower in the South than in the North (see Muratore et al. (2004)).
33 Cook (1979) highlights the endogenous nature of clearance rates (they are mediated by the choices made by criminals) and
advises against using these as a measure of criminal justice system effectiveness.

26

 5.3.2

Transitioning to Different Crimes

In this section we discuss estimates of the model described in Section 3.2, in which we take into account
that the next crime a pardoned individual commits may differ from the one for which he was previously
incarcerated, and thus carry a different sentence length. Using data on the prison population from two
prisons in Milan, we constructed transition probabilities between crimes. For cases in which multiple
crime types are committed, we categorized crimes based on the most serious offense (defined based on
either the mean or median sentence). We constructed predicted sentences for all inmates over all possible
future crime choices based on equation (3). We then integrated the likelihood over these different possible
future sentences, using the transition probabilities as weights. The estimates for this model are reported
in Table 11. The results are similar, both to each other, and to our baseline estimates. The discount
factor is 0.76 using the mean sentence to define the most serious crime, and 0.75 using the median.
Not surprisingly the standard errors increase slightly, but the results are still fairly precisely estimated.
Overall, we conclude that using the original crime type as a proxy for the future crime does not bias our
estimates of time preference.

5.3.3

Sentences Less Than Two Years

Under the Collective Clemency Bill passed in 2006, only individuals who receive a future sentence of
at least two years are subject to serving their residual sentence. About one-third of the individuals in
our sample had original sentences of less than two years. This suggests that many individuals in our
data may not face the residual sentence enhancement. One natural way to deal with this would be to set
the residual sentence equal to zero for these individuals. One potential problem with this approach is
that a common reason for shorter sentences is sentence reductions, and repeat offenders are less likely
to obtain sentence reductions.34 This suggests that not all individuals with original sentences below two
years should expect to receive similarly low sentences for their next offense.
In order to investigate whether this two-year cutoff is driving any of our results, we estimated three
additional specifications that deal with this issue in different ways. First, we simply set the residual
sentence equal to zero for individuals with an original sentence below two years. Second, we set the
expected new sentence equal to two years (plus the residual sentence). Third, we dropped all observations
34 There

are several alternative sentencing arrangements that can reduce an individual’s incarceration spell. The two most
common are probation and home arrest.

27

 with an original sentence below two years. The results are presented in Table 12. For the first two
specifications, the results are very similar to our baseline estimates, with estimated discount factors of
0.77 and 0.74. In the third specification, the point estimate drops to 0.66, but it is not nearly as precisely
estimated, due to the smaller number of observations. The resulting confidence interval completely
covers that of the baseline model. These results indicate that our estimates are not being driven by the
two-year threshold for the law.

5.3.4

Hyperbolic Discounting

Since criminal acts generate immediate rewards but delayed costs that are spread over time, it has been
argued that present bias, or a tendency to focus on the immediate gratification, might explain the engagement in criminal activities (Jolls, Sunstein, and Thaler (1998)). This kind of impatience, which is
very strong for near rewards but later declines over time, is often referred as “hyperbolic discounting".35
While there is growing evidence that at least some individuals exhibit inconsistent time preferences36 ,
for criminal behavior such evidence is still mainly speculative, with the possible exception of very young
offenders (Lee and McCrary (2009)). In order to investigate the presence of time inconsistency in preferences among criminals, we estimate a version of our model under hyperbolic discounting. For mathematical convenience we model hyperbolic preferences in continuous time, where such preferences are
commonly expressed as
1
.
DFHyperbolic-Continuous (t) =
1 + kt

(7)

The parameter k governs the rate of discounting. In the continuous-time formulation, hyperbolic discounting differs from the exponential version in two ways. First, under the hyperbolic model, individuals discount the future to a much larger extent early on. Second, the rate of decay of discount factors
is slower, leading to larger discount factors (relative to exponential discounting) in later periods. Although we have little data with short sentences, we do have data with longer sentences, which allow us
to evaluate the hyperbolic model relative to the exponential model.
For convenience, researchers have often used a quasi-hyperbolic discounting specification, which
is a discrete-time approximation to the continuous-time version. These functions have been adopted
for analytic tractability by Laibson (1997); O’Donoghue and Rabin (1999); DellaVigna and Paserman
35 An overview of hyperbolic discounting can be found in Rabin (1998). In the criminology literature, a related notion is
referred to as impulsivity (Wilson and Herrnstein (1985); Jolliffe and Farrington (2009)).
36 For a recent overview of the empirical literature see DellaVigna (2009).

28

 (2005); Mastrobuoni and Weinberg (2009), among others. With quasi-hyperbolic discounting, discount
factors fall quickly in the first period (or periods), and then afterward decay at a geometric rate. Under
the specification due to Laibson (1997) the discount factors are given by:

DFQuasi-Hyperbolic-Discrete (t) = β × δ t ,
where β < 1 denotes the initial drop in the discount factor. After the first period, discount factors continue
to decline in the same way as under exponential discounting. Due to the nature of our data, we do
not observe decisions for individuals facing extremely short sentences of only a month (or even a few
months). As a result, we are not able to separately identify β from the mean disutility of prison, and the
quasi-hyperbolic discounting model is observationally equivalent to the exponential discounting model.37
Instead, we focus on the continuous-time formulation.
In Table 13, we provide estimates from the hyperbolic form of discounting in equation (7). For
comparison we also estimate a continuous time version of exponential discounting, which is given by

DFExponential-Continuous (t) = e−rt ,
where r is the monthly discount rate. (The estimates for the discrete and continuous time versions of
exponential discounting yield results that are essentially identical to each other.)
Comparing the likelihood values in Table 13, we can see that the model with exponential discounting
outperforms the model with hyperbolic discounting. Furthermore, the hyperbolic discounting parameter
k is not precisely estimated. When k equals zero in equation (7), this implies no discounting, and the
95% confidence interval easily covers zero. The top end of the confidence interval implies a discount
factor in the first year (t = 12 months) of 0.24. Together the results reflect a very large range of potential
discount factors under hyperbolic discounting that are consistent with the data.
Because the parameters of the two different forms of discounting (k and r) affect discount rates in
different ways, it is not straightforward to compare the implied discount factors. In order to facilitate
a comparison, we have computed the implied discount factors under both estimated models and plotted
them in Figure 5. One feature of hyperbolic discounting is that discount factors fall more quickly in
37 From

a policy perspective, separately identifying the quasi-hyperbolic parameter β from the mean disutility of prison may
not even matter, as long as one is not interested in the effects of very short incarceration spells, as both models would generate
the same implications for utility, and therefore behavior, over such horizons.

29

 early periods, but then flatten out, compared to exponential discounting. The implied 1-year discount
factors are 0.74 and 0.44 for the exponential and hyperbolic models, respectively. At about 8 years, they
intersect, with hyperbolic discount factors being higher afterward. Similar to several recent papers in the
literature on estimating time preference (e.g., Warner and Pleeter (2001); Harrison, Lau, and Williams
(2002); Sutter et al. (2013)), we find no evidence of hyperbolic discounting behavior, though we cannot
rule out that the discount function might be hyperbolic over very short periods, giving rise to “quasihyperbolic” discount functions.

6

Implications for Estimates of Deterrence

Our finding of significant discounting behavior among criminals implies that the marginal deterrent effect
of imprisonment is decreasing in sentence length. Estimates that fail to take this into account capture
an “average” effect that is biased downward for early periods and upward for later periods. This is
particularly important to consider when comparing the effectiveness of deterrence across countries and
policies. For example, the elasticity of recidivism with respect to sentence length has been found to
vary considerably in the literature, suggestive of large differences in deterrent power. To illustrate the
importance of discounting in driving estimates of deterrence, we computed this elasticity both for the
Italian pardon in our data and for the estimates in Helland and Tabarrok (2007), a study of the deterrent
effect of three strikes legislation in the US. We find elasticities of -0.45 and -0.055, respectively.38,39 The
fact that the elasticity is almost an order of magnitude larger for the Italian pardon is, at face value,
suggestive of large differences in the deterrent power of increases in prison sentences between Italy and
the US. However, when we simulate the effect of the three strikes policy in our data using our estimated
model40 , the predicted drop in recidivism implies an elasticity of only -0.10, much closer the Helland
and Tabarrok (2007) estimates.
While both policies led to percentage decreases in recidivism of about 17%, the large apparent difference in deterrence is caused by the fact that the sentence enhancements under three strikes are substantially longer. The percentage increase in sentence length was 37% in the Italian pardon, and 313%
total change in recidivism rate
baseline sentence
=
baseline recidivism
months of additional sentence .
Galbiati, and Vertova (2009) report an elasticity of -0.74. The difference is due to the fact that, in order to facilitate
comparison with other results in the literature, we compute the elasticity with respect to the baseline sentence (39 months) and
the baseline recidivism (0.14), as opposed to values that include the changes in sentence and recidivism induced by the policy.
If we follow their formula, we obtain an almost identical estimate to theirs of -0.75.
40 That is, we set all future sentences to be equal to 20 years (the minimum sentence under three strikes).
38 We compute all elasticities using the following formula: %∆recidivism
%∆sentence

39 Drago,

30

 under three strikes. However, the later months of the sentence are heavily discounted, leading to small
incremental effects on recidivism. Using our baseline estimates of the discount factor, the percentage
changes in present discounted (dis)utility from imprisonment are 18% and only 24% for the pardon and
three strikes, respectively. In other words, the true force of the three strikes policy is only marginally
larger than that of the Italian pardon.
For a baseline sentence s and an additional sentence ∆s, we can use equation (2) to compute the per(1−δ s+∆s )−(1−δ s )
. Compared
centage change in present discounted disutility of prison, %∆disutility(δ ) =
1−δ s
to the elasticity with respect to sentence length, the only additional information that is needed to compute
the elasticity with respect to the total disutility of prison,

%∆recidivism
,
%∆disutility(δ )

is the discount factor δ . Using

our discount factor, increasing the disutility of prison by 10% reduces recidivism by 9.3% in our data
and by 7.2% in Helland and Tabarrok (2007).
This analysis may also help to explain why the literature has failed to find systematic convincing
evidence as to the deterrent effect of the severity of punishment (compared to certainty of punishment).
The magnitude of the effect depends on the length of punishments studied, and, for long sentences, the
effect is likely to be quite small after taking into account discounting.

7

Conclusion

A large number of studies have cited excessive discounting as a potential explanation for both engagement in crime as well as the (lack of) responsiveness to increased severity of punishment. Empirical
evidence on the discounting behavior of criminals, on the contrary, is quite scarce. We provide new
evidence on the extent of discounting among criminals by exploiting a quasi-experiment that is driven
by a mass release of Italian prison inmates that took place in 2006. Conditional on the original sentence,
the collective pardon generates a distribution of exogenous sentence enhancements. The raw data show
a monotonic reduction in marginal deterrence as the sentence increases (thereby reducing the additional
disutility of incarceration), suggestive of the presence of discounting. In order to identify discount factors
for the large number of pardoned inmates, we estimate a basic intertemporal model of criminal behavior,
and our findings are very robust to a series of different specifications.
Overall, our baseline estimate of criminal discount factors of 0.74 is on the low end of the range
of individual-level discount factors found in the literature for the general population. This supports the
hypothesis that low future time preference is a driver of criminal behavior. However, the estimates are
31

 still far from zero, suggesting that imprisonment does have the potential to deter crime, as criminals do
not fully discount future punishments.
While it has been recognized that low future time preference could be important in shaping the
behavioral responses to future punishments for crimes, this is rarely taken into account both in empirical
studies on deterrence and in policy debates. Given the delayed nature of imprisonment, discounting and
deterrence go hand-in-hand. As we highlight above, this can have importance implications for estimates
of the deterrent effect of imprisonment, and can help explain the lack of consensus on the magnitude of
deterrence.
Our paper provides one of the first empirical investigations and finds significant, but not complete,
discounting and illustrates that accounting for discounting is critical for designing optimal deterrence
policy. Our estimates suggest that while increased sentence length can have quite strong deterrent effects
for low initial sentence lengths, the incremental effects are much smaller for longer sentence lengths.
From a deterrence perspective it might be beneficial to trade off the costs of incarceration associated
with long prison sentences with policies aimed at increasing the certainty of punishment either through
a higher probability of apprehension or through improved efficiency of the criminal justice system.

Appendix A: Dynamic Utility Model
In this appendix we should how our main estimating equation can be obtained from a more general
dynamic utility model via restrictions on the option value of future crimes. The value of being free and
facing sentence s if caught committing another crime is given by

 
 
 
 
 
 
V f (s) = max u (nc) + δ E V f (s) , u (c) + P (d + δ E [V p (s − 1)]) + δ (1 − P) E V f (s) + ε ,

where u (nc) is the utility from not committing a crime, u (c) is the utility from committing a crime, P is
the probability of being apprehended, V p (s) is the utility of being in prison with a sentence of s years, d
is the per-period disutility of prison, and ε is a random utility term associated with committing a crime.
Without loss of generality, we normalize u (nc) to be equal to zero.
If an individual commits a crime and is apprehended, they are placed in prison for s total periods
periods (the first of which is served immediately), and then released. Therefore the value of being in

32

 prison with a remaining sentence of s − 1 periods can be written as:

p

 

V (s − 1) = d

1 − δ s−1
1−δ

 

 
 
+ δ s−1 E V f (s) .

One way to generate our main estimating equation is to assume that the utility shocks ε vary across
individual, but are constant over time. Under this assumption, there will be a cutoff value of ε (which
depends on observables) that determines whether or not an individual commits a crime when not incarcerated. Individuals with values of ε below the cutoff will not commit crime, which given our normalization
 
 
 
of u (nc) = 0 implies that the value of their option to commit future crime is zero E V f (s) = 0 , and
we obtain the model in equation (2). For individuals with an ε above the cutoff, we can write their value
 
 
functions as follows. Since ε is fixed over time, V f (s) = E V f (s) , and since ε is above the cutoff, these
individuals will choose to commit crime, which implies the following:

V f (s) = u (c) + P (d + δV p (s − 1)) + δ (1 − P)V f (s) + ε.

Plugging in for the value of being incarcerated V p (s), we have
 
 
 
 
1 − δ s−1
s f
+ δ V (s) + δ (1 − P)V f (s) + ε
V (s) = u (c) + P d + δ d
1−δ
f

which we can solve for:
V f (s) =

u (c) + Pd

 

1−δ s
1−δ

 

+ ε,

1 − Pδ s − δ (1 − P)

.

Since these individuals choose to commit crime it must be the case that they prefer to do it earlier rather
than later (due to discounting), meaning that

0+δ 

u (c) + Pd

 

1−δ s
1−δ

 

+ε



 <
1 − Pδ s − δ (1 − P)
  s 


1−δ
 
 
s
u
(c)
+
Pd
+
ε,
1−δ
1−δ
 + ε.
u (c) + Pd
+ (Pδ s + δ (1 − P)) 
s
1−δ
1 − Pδ − δ (1 − P)

33

  
Subtracting δ

s

u(c)+Pd ( 1−δ
1−δ )+ε,
1−Pδ s −δ (1−P)

 
0 < u (c) + Pd

 
from both sides we have

1−δs
1−δ

 


+ ε + (Pδ s + δ (1 − P) − δ ) 

u (c) + Pd

 

1−δ s
1−δ

 

+ ε,

1 − Pδ s − δ (1 − P)


,

which can be rewritten as
 
0<

For δ < 1,

 

1−δ
1−Pδ s −δ (1−P)

 

1−δ
s
1 − Pδ − δ (1 − P)

  
 
 
 
1−δs
u (c) + Pd
+ε .
1−δ

is always positive, and this inequality holds if and only if u (c)+Pd

 

1−δ s
1−δ

 

+

ε > 0. Therefore, we have shown that for individuals with random utility draws ε both above and below
the cutoff, their decisions are consistent with the model we employ in Section 3.1. If we plug in for u (c)
h
i   1−δ X δ tsi j  
( )
u
d
+ εi j >
and d, and add back in subscripts, this is equivalent to α0 + α1 Xi j + β0 + β1 Xi j Pjlc 1−δ iXj δ
( ij)
0, which is the model in equation (2).
 
 
An alternative way to motivate our model is to set E V f (s) = 0, which essentially sets the value
of having the option to commit future crime to be zero. For example, suppose that individuals make a
once-and-for-all decision of whether to commit a crime, and then receive zero payoff afterward. In this
case the model simplifies to:
 
1−δs
V (s) = max 0, u (c) + Pd
+ε .
1−δ
f

 

s

We then have that an individual commits a crime if and only if u (c) + Pd 1−δ
1−δ + ε > 0, which again is
consistent with our model in equation (2).

Appendix B: Conditional Independence Assumption
Recall that ci j is an indicator for whether individual i commits a crime of type j. For notational simplicity,
we drop the subscripts and let c be an indicator for committing a crime. The probability that a crime is
committed conditional on the original sentence os, the residual sentence rs, the other observables X, and

34

 a vector of parameters to be estimated θ , can be written as:

Pr (c = 1   os, rs, X; θ ) = Pr (g (os, rs, X; θ ) + ε > 0)
= Pr (ε > −g (os, rs, X; θ ))
= 1 − Fε os,rs (−g (os, rs, X; θ ))
1 − Fε os (−g (os, rs, X; θ )) ,

where g is the utility model as a function of observables and F is the distribution function for ε. The
last equality follows from the key identifying assumption that conditional on the original sentence os, the
error ε is independent of the residual sentence rs.
Recognizing the dependence between ε and os, we can write

ε = h (os; γ) + u,

where h is an unknown (non-parametric) function. Assuming u is independent of os, and using a simple
change of variables from ε to u, we can then write the probability of crime as

Pr (c = 1   os, rs, X; θ ) = 1 − Fε os (−g (os, rs, X; θ ))
= 1 − Fu os (−g (os, rs, X; θ ) − h (os; γ))
= 1 − Fu (−g (os, rs, X; θ ) − h (os; γ)) ,

where the second equality is due to the change of variables, and the third equality follows from the
independence of u and os. This implies that in addition to our model of the crime choice g, an additional
term needs to be added to the model: h (os; γ). This function controls for the dependence between the
error ε and the original sentence os. This additional term is necessary because we do not know the exact
form of the dependence h between ε and os. It may be the case that os enters h differently from how it
enters g. The model can then be estimated using maximum likelihood, with the caveat that any part of g
that is separable in os is not separately identified from h. Since our only interest is in identifying the part
of g related to the residual sentence, this is not a problem for identification of the discount factor.

35

 References
Akerlund, D., B. Golsteyn, H. Gronqvist, and L. Lindahl. 2014. “Time Preferences and Criminal Behaviour.” Tech. Rep. 8168, IZA.
Andersen, Steffen, Glenn W Harrison, Morten I Lau, and E Elisabet Rutström. 2014. “Discounting
behavior: A reconsideration.” European Economic Review 71:15–33.
Barbarino, Alessandro and Giovanni Mastrobuoni. 2014. “The Incapacitation Effect of Incarceration:
Evidence from Several Italian Collective Pardons.” American Economic Journal: Economic Policy
6 (1):1–37.
Beccaria, Cesare. 1764. On Crimes and Punishments. Philadelphia: Philip H. Nicklin, 1819, 2nd
american edition. translated from the italian by anonymous ed.
Becker, G. S. and C. B. Mulligan. 1997. “The Endogenous Determination of Time Preference.” The
Quarterly Journal of Economics 112(3):729–758.
Becker, Gary S. 1968. “Crime and Punishment: An Economic Approach.” The Journal of Political
Economy 76 (2):169–217.
Bukstel, Lee H and Peter R Kilmann. 1980. “Psychological effects of imprisonment on confined individuals.” Psychological Bulletin 88 (2):469.
Cagetti, Marco. 2003. “Wealth accumulation over the life cycle and precautionary savings.” Journal of
Business & Economic Statistics 21 (3):339–353.
California Assembly. 1968. “Deterrent Effects of Criminal Sanctions-Progress Report of the Assembly
Committee on Criminal Procedure.” .
Chen, M.K. 2013. “The Effect of Language on Economic Behavior: Evidence from Saving Rates, Health
Behaviors and Retirement Access.” American Economic Review 103(2):690–731.
Ching, Andrew and Matthew Osborne. 2015. “Identification and Estimation of Forward-Looking Behavior: The Case of Consumer Stockpiling.” Unpublished working paper, University of Toronto.
Coller, Maribeth and Melonie B Williams. 1999. “Eliciting individual discount rates.” Experimental
Economics 2 (2):107–127.
Cook, Philip J. 1979. “The clearance rate as a measure of criminal justice system effectiveness.” Journal
of Public Economics 11 (1):135–142.
———. 1980. “Research in criminal deterrence: Laying the groundwork for the second decade.” Crime
and justice :211–268.
Davis, M.L. 1988. “Time and Punishment: an Intertemporal Model of Crime.” Journal of Political
Economy 96(2):383–390.
DellaVigna, Stefano. 2009. “Psychology and Economics: Evidence from the Field.” Journal of Economic
Literature 47 (2):315–72.
DellaVigna, Stefano and Daniele M. Paserman. 2005. “Job Search and Impatience.” Journal of Labor
Economics 23 (3):527–587.

36

 Dohmen, Thomas, Benjamin Enke, Armin Falk, David Huffman, and Uwe Sunde. 2015. “Patience and
The Wealth of Nations.” Working Paper.
Drago, Francesco, Roberto Galbiati, and Pietro Vertova. 2009. “The deterrent effects of prison: Evidence
from a natural experiment.” Journal of Political Economy 117 (2):257–280.
Durlauf, Steven N. and Daniel S. Nagin. 2011. “The Deterrent Effect of Imprisonment.” Controlling
Crime: Strategies and Tradeoffs .
Ehrlich, Isaac. 1973. “Participation in illegitimate activities: A theoretical and empirical investigation.”
The Journal of Political Economy :521–565.
Falk, Armin, Anke Becker, Thomas Dohmen, Benjamin Enke, David Huffman, and Uwe Sunde. 2015.
“The Nature and Predictive Power of Preferences: Global Evidence.” Unpublished manuscript, Available at SSRN 2691910.
Friedman, Milton. 1957. A Theory of the Consumption Function. Princeton, NJ: Princeton University
Press.
Galor, Oded and Ömer Özak. 2014. “The Agricultural Origins of Time Preference.” Tech. rep., National
Bureau of Economic Research.
Giglio, Stefano, Matteo Maggiori, and Johannes Stroebel. 2015. “Very Long-Run Discount Rates.” The
Quarterly Journal of Economics 130 (1):1–53.
Gullone, Eleonora, Tessa Jones, and Robert Cummins. 2000. “Coping styles and prison experience as
predictors of psychological well-being in male prisoners.” Psychiatry, psychology and law 7 (2):170–
181.
Harrison, Glenn W., Morten I. Lau, and E. Elisabet Rutström. 2010. “Individual discount rates and
smoking: Evidence from a field experiment in Denmark.” Journal of Health Economics 29 (5):708–
717.
Harrison, Glenn W., Morten I. Lau, and Melonie B. Williams. 2002. “Estimating Individual Discount
Rates in Denmark: A Field Experiment.” American Economic Review 92 (5):1606–1617.
Hausman, Jerry A. 1979. “Individual Discount Rates and the Purchase and Utilization of Energy-Using
Durables.” Bell Journal of Economics 10 (1):33–54.
Hawken, Angela and Mark Kleiman. 2009. “Managing Drug Involved Probationers with Swift and
Certain Sanctions: Evaluating Hawaii’s HOPE: Executive Summary.” Washington, DC: National
Criminal Justice Reference Services .
Heckman, James J. 1976. “A Life-Cycle Model of Earnings, Learning, and Consumption.” Journal of
Political Economy 84 (4):S11–44.
Heckman, James J. and Richard Robb. 1985. “Alternative Methods for Evaluating the Impact of Interventions: An Overview.” Journal of Econometrics 30 (1-2):239–267.
Helland, Eric and Alexander Tabarrok. 2007. “Does Three Strikes Deter?: A Nonparametric Estimation.”
Journal of Human Resources 42 (2).
Hjalmarsson, Randi. 2009. “Crime and Expected Punishment: Changes in Perceptions at the Age of
Criminal Majority.” American Law and Economics Review 11 (1):209–248.
37

 Hochstetler, Andy, Daniel S Murphy, and Ronald L Simons. 2004. “Damaged goods: Exploring predictors of distress in prison inmates.” Crime & Delinquency 50 (3):436–457.
Hunt, Albert R. 2015. “Reforming the Criminal Justice System Is Not Assured.” New York Times .
ISTAT. 2005. “Statistiche Giudiziarie Penali.” Tech. rep., Istituto Italiano di Statistica.
———. 2015. “I detenuti nelle carceri italiane.” Report, Istituto Nazionale di Statistica. URL http:
//www.istat.it/it/archivio/153369.
Italian Ministry of Internal Affairs. 2007. Rapporto sulla criminalità in Italia. Analisi, Prevenzione,
Contrasto.
Jappelli, Tullio, Marco Pagano, and Magda Bianco. 2005. “Courts and Banks: Effects of Judicial Enforcement on Credit Markets.” Journal of Money, Credit and Banking 37 (2):pp. 223–244.
Jolliffe, Darrick and David P Farrington. 2009. “A systematic review of the relationship between childhood impulsiveness and later violence.” Personality, Personality disorder, and violence :41–61.
Jolls, Christine, Cass R Sunstein, and Richard Thaler. 1998. “A behavioral approach to law and economics.” Stanford Law Review :1471–1550.
Kaplow, Louis. 1990. “Optimal deterrence, uninformed individuals, and acquiring information about
whether acts are subject to sanctions.” Journal of Law, Economics, & Organization :93–128.
Katz, Lawrence, Steven D. Levitt, and Ellen Shustorovich. 2003. “Prison Conditions, Capital Punishment, and Deterrence.” American Law and Economics Review 5 (2):318–343.
Laibson, David I. 1997. “Golden Eggs and Hyperbolic Discounting.” The Quarterly Journal of Economics 112 (2):443–477.
Lee, David S. and Justin McCrary. 2009. “The Deterrence Effect of Prison: Dynamic Theory and Evidence.” Nber working papers, National Bureau of Economic Research, Inc.
Levitt, Steven D and Thomas J Miles. 2007. “Empirical study of criminal punishment.” Handbook of
law and economics 1:455–495.
Lochner, Lance. 2007. “Individual Perceptions of the Criminal Justice System.” American Economic
Review 97 (1):444–460.
Lochner, Lance J. and Enrico Moretti. 2004. “The Effect of Education on Crime: Evidence from Prison
Inmates, Arrests, and Self-Reports.” American Economic Review 94 (1):155–189.
Mancino, Maria Antonella, Salvador Navarro, and David A. Rivers. 2015. “Separating State Dependence, Experience, and Heterogeneity in a Model of Youth Crime and Education.” Centre for Human
Capital and Productivity Working Papers 20151, University of Western Ontario.
Mastrobuoni, Giovanni and Paolo Pinotti. 2015. “Legal status and the criminal activity of immigrants.”
American Economic Journal: Applied Economics 7 (2):175–206.
Mastrobuoni, Giovanni and Matthew Weinberg. 2009. “Heterogeneity in intra-monthly consumption
patterns, self-control, and savings at retirement.” American Economic Journal: Economic Policy
1 (2):163–189.
McCrary, Justin. 2010. “Dynamic Perspectives on Crime.” Handbook on the Economics of Crime :82.
38

 Meier, Stephan and Charles Sprenger. 2010. “Present-biased preferences and credit card borrowing.”
American Economic Journal: Applied Economics :193–210.
Muratore, Maria Giuseppina, Isabella Corazziari, Giovanna Tagliacozzo, Alessandro Martini, Alessandra
Federici, Manuela Michelini, Agostina Loconte, Roberta Barletta, and Anna Costanza Baldry. 2004.
La Sicurezza dei Cittadini. 2002 Reati Vittime, Percezione della Sicurezza e Sistemi di Protezione.
Indagine Multiscopo sulle Famiglie “Sicurezza dei cittadini?- Anno 2002. Rome: Istituto Nazionale
di Statistica.
Nagin, Daniel S. 2013. “Deterrence: A Review of the Evidence by a Criminologist for Economists.”
Annu. Rev. Econ. 5 (1):83–105.
Nagin, Daniel S and Greg Pogarsky. 2004. “Time and punishment: Delayed consequences and criminal
behavior.” Journal of Quantitative Criminology 20 (4):295–317.
O’Donoghue, Ted and Matthew Rabin. 1999. “Doing It Now or Later.” American Economic Review
89 (1):103–124.
Polinsky, A Mitchell and Steven M Shavell. 1999. “On the Disutility and Discounting of Imprisonment
and the Theory of Deterrence.” Journal of Legal Studies 28 (1):1.
Rabin, Matthew. 1998. “Psychology and economics.” Journal of economic literature :11–46.
Sutter, Matthias, Martin G. Kocher, Daniela Glätzle-Rüetzler, and Stefan T. Trautmann. 2013. “Impatience and Uncertainty: Experimental Decisions Predict Adolescents’ Field Behavior.” American
Economic Review 103 (1):510–31.
Viscusi, W Kip and Michael J Moore. 1989. “Rates of time preference and valuations of the duration of
life.” Journal of Public Economics 38 (3):297–317.
Warner, John T. and Saul Pleeter. 2001. “The Personal Discount Rate: Evidence from Military Downsizing Programs.” American Economic Review 91 (1):33–53.
Wilson, James Q. and Richard J. Herrnstein. 1985. Crime and Human Nature: The Definitive Study of
the Causes of Crime. New York: Simon and Schuster.

39

 24

0

6

12

Time Served
18

12

6

0

18
24
Residual Sentence

30

36

36

30

24

0

6

12

Time Served
18

12

6

0

18
24
Residual Sentence

30

36

Marginal Effect on Recidivism

Cumulative Effect on Recidivism

0

30

0

36

Deterrence, 74%

Treatment, 74%

Deterrence, 74%

Treatment, 74%

Figure 1: Discounting against Prison Treatment
Notes: The left (right) panel shows the marginal (cumulative) effect of deterrence (solid line) as well as prison treatment
(dashed line) on recidivism. Note that for both deterrence and treatment, the effect of increasing prison sentence is to lower
recidivism. The degree of discounting is set to 74 percent per year, and prison treatment effects exhibit decreasing returns to
scale, meaning that they are larger at the beginning of the time served than at the end. An extra month of residual sentence
corresponds to one less month spent in prison.

40

 0
Cumulative Utility of Prison
0

2.5

5

7.5
10
Years of Sentence

12.5

15

Figure 2: Cumulative Utility of Prison as a Function of Sentence Length
Notes: The figure shows the cumulative utility of prison time as a function
of sentence length assuming an annual discount factor of 0.74. Note that all
values of cumulative utility are negative.

-.4

-.1

Change in Recidivism Rate
-.05

Change in Recidivism Rate
-.3
-.2
-.1

0

Conditional on Original Sentence

0

Unconditional

0

50

100
150
Total Sentence Length (Months)

200

0

50

100
150
Total Sentence Length (Months)

200

Figure 3: Change in Average Recidivism by Total Sentence Length
Notes: The left panel plots the average recidivism rate for each of 40 quantiles of the total expected sentence length. The right
panel adds controls for original sentence length. In order to control for original sentence length we regress the recidivism
indicator on a cubic polynomial in the log of original sentence and dummies for each of the quantiles of total sentence length.
The right panel plots the coefficients on each of the dummies. For both panels the differences in average recidivism are
normalized relative to the maximum recidivism rate across the quantiles.

41

 100
Discounted Flow Disutility
20
40
60
80
0
0

2.5

5

7.5
10
Years of Sentence

12.5

15

Discount Factors:
10%
50%
95%

30%
74%

Figure 4: Flow Disutility of Prison as a Function of Sentence
Length–Different Discount Factors

0

Discounted Flow Disutility
20
40
60
80

100

Notes: The figure shows the discounted flow of disutility of prison time
(normalized to be 100 at the beginning of incarceration) for different annual
discount factors.

0

2.5

5

7.5
10
Years of Sentence

12.5

15

Discount Factors:
Exponential: Continuous

Hyperbolic: Continuous

Figure 5: Flow Disutility of Prison as a Function of Sentence
Length–Different Forms of Discounting
Notes: The figure shows the discounted flow of disutility of prison time
(normalized to be 100 at the beginning of incarceration) for exponential and
hyperbolic discounting. These curves are based on the estimates in Table
13.

42

 Table 1: Summary Statistics
Variable
Recidivism (re-incarceration)
Original sentence
Residual sentence
Definitive sentence
Age
Female
Married
Permanently employed
Primary education or less
No education
Secondary school or above
High school or above
Italian

Mean
0.22
39.90
15.01
0.74
37.72
0.05
0.26
0.15
0.27
0.06
0.03
0.02
0.62

Std. Dev.
0.41
32.67
10.52
0.44
9.79
0.21
0.44
0.36
0.44
0.24
0.18
0.13
0.49

Min.
0
1
0
0
19
0
0
0
0
0
0
0
0

Max.
1
360
36
1
70
1
1
1
1
1
1
1
1

Notes: This table presents summary statistics for the individual characteristics of the pardoned inmates. Sentence durations are
expressed in months. The total number of observations is 19,616.

Table 2: List of Crime Types
Variable
Property
Violent
Illegal detention of weapons
Drug-related
Prostitution
Illegal migration
Organized crime (mafia)
Against the economy and state
Fraud
Other crimes

Mean
0.58
0.26
0.09
0.39
0.02
0.07
0.02
0.26
0.10
0.19

Std. Dev.
0.49
0.44
0.29
0.49
0.13
0.26
0.13
0.44
0.29
0.39

Min.
0
0
0
0
0
0
0
0
0
0

Max.
1
1
1
1
1
1
1
1
1
1

Notes: This table presents summary statistics for the crime types committed by the pardoned inmates. The total number of
observations is 19,616.

43

 Table 3: Clearance Rates
Crime Type
Property
Violent
Drug-related
Organized crime (mafia)
Against the economy and state
Fraud
Other crimes
All

Mean
0.11
0.58
0.73
0.88
0.50
0.33
0.77
0.23

Std. Dev.
0.08
0.17
0.17
0.16
0.11
0.13
0.31
0.11

Min.
0.03
0.14
0.19
0.00
0.21
0.08
0.00
0.08

Max.
0.39
0.94
1.00
1.00
0.83
0.82
1.00
0.63

Notes: The province-level clearance rates are based on ISTAT’s criminal statistics (ISTAT, 2005).

Table 4: Transition Probabilities
Panel A: Most Serious Crime Based on Median Sentence
Property

Violent

Drugrelated

Prostitution

Organized
crime

Against
economy/state

Fraud

Other
crimes

Sum

0.47
0.24
0.13
0.25
0.10
0.28
0.21
0.22

0.14
0.33
0.10
0.00
0.14
0.15
0.13
0.15

0.09
0.12
0.51
0.17
0.10
0.10
0.15
0.25

0.00
0.00
0.00
0.25
0.00
0.00
0.00
0.00

0.01
0.02
0.01
0.00
0.37
0.02
0.04
0.03

0.17
0.15
0.08
0.17
0.13
0.29
0.18
0.15

0.03
0.03
0.03
0.00
0.04
0.05
0.22
0.04

0.09
0.11
0.14
0.17
0.12
0.10
0.08
0.17

1
1
1
1
1
1
1
1

Property
Violent
Drug-related
Prostitution
Organized crime (mafia)
Against the economy and state
Fraud
Other crimes

Panel B: Most Serious Crime Based on Mean Sentence
Property

Violent

Drugrelated

Prostitution

Organized
crime

Against
economy/state

Fraud

Other
crimes

Sum

0.47
0.24
0.13
0.22
0.10
0.28
0.23
0.22

0.13
0.32
0.10
0.09
0.13
0.15
0.14
0.14

0.10
0.13
0.52
0.09
0.11
0.11
0.11
0.27

0.00
0.01
0.00
0.13
0.01
0.00
0.00
0.00

0.01
0.02
0.01
0.00
0.37
0.02
0.03
0.03

0.17
0.15
0.08
0.17
0.13
0.29
0.20
0.15

0.02
0.02
0.02
0.04
0.03
0.04
0.21
0.03

0.09
0.11
0.14
0.26
0.12
0.10
0.07
0.17

1
1
1
1
1
1
1
1

Property
Violent
Drug-related
Prostitution
Organized crime (mafia)
Against the economy and state
Fraud
Other crimes

Notes: The transition probabilities are based on the entire criminal history of a sample of inmates who served at least some
time in one of two different Milan prisons between 2001 and 2012. Since criminals may be imprisoned for an offense related
to multiple crime types, the transitions are based on the crimes with either the highest median sentence (Panel A) or the
highest mean sentence (Panel B).

Table 5: Recidivism and Sentence Length
Original Sentence Length
Coefficient on Residual Sentence

Less than 1 Year

1 - 2 Years

2 - 4 Years

4 - 6 Years

More than 6 Years

All

-1.459
(0.072)

-0.502
(0.042)

-0.257
(0.018)

-0.139
(0.041)

-0.005
(0.075)

-0.196
(0.016)

Notes: Each column contains the estimated coefficient (and standard error) on residual sentence from a linear regression of
recidivism on original sentence and residual sentence, for various bins of original sentence length.

44

 Table 6: Baseline Results
Specification
Monthly Discount Factor (δ )
Annual Discount Factor

(1)
No Covariates

(2)
Baseline Covariates

(3)
Alt. Low Education

(4)
Alt. High Education

(5)
Crime Dummies

0.97142
(0.00537)
0.706

0.97488
(0.00519)
0.737

0.97546
(0.00511)
0.742

0.97469
(0.00526)
0.735

0.97104
(0.00549)
0.703

0.366
(0.409)
0.199
(0.145)
-0.032
(0.016)

0.252
(0.398)
0.174
(0.146)
-0.026
(0.016)

0.174
(0.393)
0.203
(0.146)
-0.028
(0.016)

0.279
(0.399)
0.163
(0.147)
-0.025
(0.016)

0.372
(0.425)
0.133
(0.152)
-0.025
(0.017)

-1.841
(0.379)

-1.169
(0.439)
-0.010
(0.006)
-0.331
(0.195)
-0.024
(0.123)

-1.144
(0.435)
-0.010
(0.006)
-0.336
(0.193)

-1.190
(0.440)
-0.009
(0.006)
-0.333
(0.196)
-0.028
(0.124)

-1.153
(0.459)
-0.010
(0.006)
-0.337
(0.209)
-0.013
(0.131)

Controls for Original Sentence:
Log-sentence
Log-sentence sq.
Log-sentence cu.

Utility of Crime:
Constant
Age
Permanently employed
Primary education or less
No education
Secondary education and above

-0.511
(0.361)

0.333
(0.216)
-0.467
(0.356)

0.026
(0.114)
-0.405
(0.288)
-0.597
(0.137)

0.055
(0.115)
-0.429
(0.286)
-0.609
(0.136)

-0.035
(0.014)
-0.001
(0.000)
0.001
(0.007)
0.007
(0.005)

-0.034
(0.014)
-0.001
(0.000)
0.001
(0.006)

High school and above
Italian
Female
Married

-0.562
(0.387)
-0.568
(0.441)
0.025
(0.115)
-0.412
(0.289)
-0.600
(0.138)

Crime dummies

Disutility of Prison:
Constant

-0.068
(0.018)

Age
Permanently employed
Primary education
No education
Secondary education and above

0.007
(0.013)

-0.006
(0.008)
0.004
(0.012)

-0.014
(0.002)
0.020
(0.005)
-0.020
(0.012)
0.012
(0.005)

-0.014
(0.002)
0.019
(0.005)
-0.019
(0.011)
0.012
(0.005)

0.010
(0.018)
-0.014
(0.002)
0.020
(0.005)
-0.020
(0.012)
0.012
(0.005)

19616
-9846.72

19616
-9854.02

19616
-9848.48

High school and above
Definitive sentence
Italian
Female
Married

-0.035
(0.014)
-0.001
(0.000)
0.000
(0.007)
0.008
(0.005)

19616
-10197.64

-0.051
(0.019)
-0.001
(0.000)
0.002
(0.008)
0.007
(0.005)

0.011
(0.015)

Crime dummies

Obs.
log-likelihood

0.065
(0.136)
-0.443
(0.306)
-0.613
(0.147)
√

-0.012
(0.002)
0.014
(0.006)
-0.018
(0.013)
0.014
(0.006)
√

19616
-9746.41

Notes: This table contains estimates from our baseline model in equation (6). In column
1 we control only for original sentence using a polynomial in log of original sentence. In
column 2 we add our baseline controls. Columns 3 and 4 use alternative measures of
education, and column 5 allows the utility of committing a crime, as well as the
disutility of prison, to depend on a full set of crime-type dummies. Standard errors are
reported in parentheses below the point estimates.

45

 46

19616
-9845.52

0.718
0.780
0.062

0.97280
(0.00547)
0.00674
(0.00417)

(1)
Married

19616
-9846.04

0.728
0.821
0.093

0.97390
(0.00523)
0.00979
(0.00753)

(2)
Female

19616
-9846.72

0.739
0.734
-0.005

0.97506
(0.00547)
-0.00051
(0.00506)

(3)
Permanently
employed

19616
-9842.45

0.776
0.677

0.99526
(0.00837)
-0.00780
(0.00271)

(4)
Log(age-18)

19616
-9842.26

0.760
0.676

1.00842
(0.01459)
-0.00161
(0.00063)
0.000016
(0.000006)

(5)
Age, age sq.

19616
-9846.70

0.739
0.731
-0.008

0.97508
(0.00527)
-0.00086
(0.00389)

(6)
Primary education
or less

19616
-9853.98

0.744
0.728
-0.016

0.97567
(0.00519)
-0.00177
(0.00639)

(7)
No education

19616
-9845.14

0.734
0.885
0.151

0.97456
(0.00535)
0.01529
(0.00694)

(8)
Secondary education
and above

19616
-9846.62

0.733
0.995
0.263

0.97440
(0.00514)
0.02520
(0.00892)

(9)
High school
and above

Notes: In these specifications we allow the discount factor to differ by the variables shown in the first row. Standard errors are reported in parentheses below the point estimates.

Obs.
log-likelihood

At age 26
At age 51

Baseline
With interaction
Difference

Annual Discount Factor

Age squared

Interaction coefficient

Constant

Monthly Discount Factor (δ )

Specification:

Table 7: Heterogeneity in Discount Factors

 Table 8: Heterogeneity in Discount Factors
Specification:

(1)
Nationality

(2)
Nationality-Language 1

(3)
Nationality-Language 2

0.96547
(0.00872)
0.01442
(0.00613)

0.97551
(0.00977)
0.01886
(0.00915)
-0.01674
(0.00751)

0.97956
(0.00851)
0.01418
(0.00602)
-0.01408
(0.00601)

0.784
0.656

0.762

0.780

0.603
0.743

0.656
0.780

19571
-9821.16

19220
-9700.85

Monthly Discount Factor (δ )
Constant
Italian
Strong FTR

Annual Discount Factor
Italian
Immigrant (All)
Immigrant (Strong FTR)
Immigrant (Weak FTR)

Obs.
log-likelihood

19616
-9842.21

Notes: In these specifications we allow the discount factor to differ by nationality and
language. FTR refers to future time reference. There are 45 observations for which
either the country of origin was unknown or we were unable to classify the FTR for the
country. Standard errors are reported in parentheses below the point estimates.

47

 48

19616
-9841.44

0.710
0.033

0.97191
(0.00513)
0.00364
(0.00385)

(1)
Violent

19616
-9805.33

0.758
-0.037

0.97715
(0.00525)
-0.00407
(0.00595)

(2)
Property

19616
-9842.98

0.751
-0.042

0.97637
(0.00497)
-0.00463
(0.00793)

(3)
Arms

19616
-9837.12

0.803
-0.107

0.98190
(0.00511)
-0.01160
(0.00387)

(4)
Drug-related

19616
-9837.18

0.719
0.205

0.97284
(0.00530)
0.02052
(0.00649)

(5)
Prostitution

19616
-9841.82

0.697
-0.279

0.97041
(0.00566)
-0.04047
(0.03629)

(6)
Migrant Law

19616
-9845.78

0.744
-0.072

0.97571
(0.00543)
-0.00825
(0.01619)

(7)
Organized crime (mafia)

19616
-9800.75

0.741
-0.073

0.97529
(0.00510)
-0.00841
(0.00564)

(8)
Against the economy and state

19616
-9843.39

0.727
-0.095

0.97382
(0.00507)
-0.01133
(0.00971)

(9)
Fraud

19616
-9836.25

0.715
0.078

0.97243
(0.00533)
0.00843
(0.00409)

(10)
Other crimes

Notes: In these specifications we allow the discount factor to differ by the types of crimes committed by the inmates before the pardon. Standard errors are reported in parentheses
below the point estimates.

Obs.
log-likelihood

Overall
Difference

Annual Discount Factor

Interaction coefficient

Constant

Monthly Discount Factor (δ )

Interaction variable:

Table 9: Heterogeneity in Discount Factors by Crime Type

 49

19616
-9838.81

√

√

19616
-9752.81

√

0.97524
(0.00551)
0.740

(2)

0.97291
(0.00515)
0.719

(1)

19616
-9821.16

√

0.97111
(0.00504)
0.703

(3)

19616
-9734.41

√
√

0.97392
(0.00554)
0.728

(4)

19616
-9881.57

√

0.98263
(0.00798)
0.810

(5)

19616
-9882.37

√

0.98269
(0.00729)
0.811

(6)

Notes: In column 1 we proxy for clearance rates using dummy variables for each of the three areas of Italy (North, Center, and South). In column 2 we include area dummies as well as
dummies for the crime type of the original offense. In column 3 we use dummies for each region (20 regions), instead of areas. Column 4 adds crime-type dummies. In columns 5 and 6
we use province-level clearance rates (taken from ISTAT, 2005), overall and crime-specific, respectively. Standard errors are reported in parentheses below the point estimates.

Obs.
log-likelihood

Area dummies
Region dummies
Crime-type dummies
Clearance rates by province
Clearance rates by province and crime type

Annual Discount Factor

Monthly Discount Factor (δ )

Specification

Table 10: Robustness Results Using Various Controls for Clearance Rates

 Table 11: Incorporating Transition Probabilities Across Crime
Types

Specification
Monthly Discount Factor (δ )
Annual Discount Factor
Obs.
log-likelihood

(1)

(2)

Mean
0.97750
(0.00616)
0.761
19616
-9854.88

Median
0.97613
(0.00641)
0.748
19616
-9854.98

Notes: These results are based on the model in Section 3.2 in which we
intergrate against the distribution of future crime choices, conditional on the
original crime type. We classify offenses into the most serious crime type,
based on both the mean (column 1) and median (column 2) sentence for
each type. Standard errors are reported in parentheses below the point
estimates.

Table 12: Robustness Regarding the Expected New Sentence
Specification
Monthly Discount Factor (δ )
Annual Discount Factor
Obs.
log-likelihood

(1)
0.97860
(0.00604)
0.771
19616
-9854.23

(2)
0.97477
(0.00771)
0.736
19616
-9852.64

(3)
0.96577
(0.01321)
0.658
13241
-6449.57

Notes: Each column presents a different approach to dealing with original
sentences below two years. Each is based on our baseline specification. In
column 1, we set the residual sentence equal to zero for observations with
original sentences below two years. In column 2 we set the expected new
sentence equal to two years (plus the residual sentence). In column 3, we
drop these observations. Standard errors are reported in parentheses below
the point estimates.

Table 13: Alternative Forms of Discounting

Specification
r

(1)
Exponential–Continuous
0.02543
(0.00532)

k
Obs.
log-likelihood

19616
-9846.72

(2)
Hyperbolic–Continuous

0.10150
(0.07441)
19616
-9848.08

Notes: Standard errors are reported in parentheses below the point
estimates.

50