ArticlePDF Available

Fines for illicit drug use do not prevent future crime: evidence from randomly assigned judges

Authors:

Abstract and Figures

This study uses judicial leniency as an instrumental variable to estimate the effect on recidivism of a monetary penalty for using or possessing a prohibited drug. Using data from the state of New South Wales, Australia, we find that fines have no measurable effect on recidivism. In contrast, ordinary least square estimates mistakenly suggest that fines increase recidivism risk. The results add to existing evidence that sanctions are ineffective in changing drug use behaviour. This should encourage policymakers to seek other ways of stopping or reducing illicit drug consumption among active users.
Content may be subject to copyright.
Fines for illicit drug use do not prevent future crime:
evidence from randomly assigned judges
This document is a post-print; please cite the published version:
Alexeev, S., Weatherburn D. 2022 “Fines for illicit drug use do not prevent future crime: evidence from
randomly assigned judges.” Journal of Economic Behavior & Organization 200:555–575
https://doi.org/10.1016/j.jebo.2022.06.015
Sergey Alexeev1, Don Weatherburn2
Abstract
This study uses judicial leniency as an instrumental variable to estimate the effect on
recidivism of a monetary penalty for using or possessing a prohibited drug. Using data from
the state of New South Wales, Australia, we find that fines have no measurable effect on
recidivism. In contrast, ordinary least square estimates mistakenly suggest that fines increase
recidivism risk. The results add to existing evidence that sanctions are ineffective in changing
drug use behaviour. This should encourage policymakers to seek other ways of stopping or
reducing illicit drug consumption among active users.
JEL codes: I12, I18, K14, K41, K42
Key words: Health Behavior, Public Health, Criminal Law, Litigation Process, Illegal
Behavior and the Enforcement of Law
Highlights:
Comprehensive administrative records on crime from Australia are used to study the
effect of a monetary penalty issued for using or possessing a prohibited drug on future
crime and drug use
The effect is identified with the two-stage least squares estimator taking advantage of
the random assignment of cases to judges
Fines neither deter nor increase the likelihood that drug users will (a) commit another
crime (of any type); (b) commit another use/possession offence involving the same
drug, or (c) commit another use/possession offence involving any drug
These results add to the existing body of evidence that sanctioning active users is
unproductive
Author to whom correspondence should be addressed.
1National Drug and Alcohol Research Centre, UNSW, 22-32 King Street Randwick, NSW 2052, Australia
Email: s.alexeev@unsw.edu.au
2National Drug and Alcohol Research Centre, UNSW, 22-32 King Street Randwick, NSW 2052, Australia
Email: d.weatherburn@unsw.edu.au
2 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
1. Introduction
There is an ongoing debate within Australia and many other countries about how to
combat addiction. Health groups argue for harm reduction policies (e.g., syringe exchange
programs or injection rooms). In contrast, law enforcement groups argue for harsher penalties
and incarceration. The expectation is that sanctions will deter both the person sanctioned
(specific deterrence) and others (general deterrence) from further offending. Although many
jurisdictions have schemes that allow police to issue a small fine or divert prohibited drug
users into treatment, large numbers of those found in possession of a prohibited drug are still
prosecuted in the courts. In Australia, the majority of people who appear in court charged with
possessing or using a prohibited drug receive a fine. In the UK, the most common response
to those found in possession of a prohibited drug is caution but around a quarter of those
convicted receive a fine (Cuthbertson 2017). Fines are also a common penalty for prohibited
drug possession in New Zealand (Wilkins & Sweetsur 2012) and Canada (Maxwell 2017).
Fines may or may not deter offenders, but the convictions that result from criminal
prosecution have adverse consequences. The criminal conviction that results in a fine may
reduce an individual’s subsequent employment and earnings prospects (Waldfogel 1994;
Grogger 1995; Dobbie, Goldin & Yang 2018). It can also lead to deeper entanglement in
the criminal justice system. For example, in New South Wales (NSW), offenders who do not
or cannot pay the fine risk having their driving licence suspended. Driving with a suspended
licence often results in imprisonment (Quilter & Hogg 2018). Given these effects, it is of
considerable interest to know whether fining those convicted of drug use and/or possession
exerts any specific deterrent effect.
Existing research indicates that harsher sanctions exert little if any specific deterrent
effect. In their review of research on the effect of prison on reoffending, for example, Nagin,
Cullen & Jonson (2009, p. 115) concluded that: ‘compared with non-custodial sanctions,
incarceration appears to have a null or mildly criminogenic effect on future criminal
behaviour. That conclusion received strong support in the case of drug offences with the
publication of a study by Green & Winik (2010), which exploited the random assignment of
drug cases to judges in the US District of Columbia to test the effectiveness of imprisonment
as a specific deterrent to further drug offending. They found no evidence that imprisoning
drug offenders had any deterrent effect.
In contrast to the vast literature on the specific deterrent effect of custodial penalties,
studies of the impact of fines on offending are comparatively rare. The studies that have
been conducted have been exclusively focused on drink-driving. The results of this research
have been mixed, with some studies finding a deterrent effect (Yu 1994; Briscoe 2004) but
most finding little if any effect (Briscoe 2004; Moffatt & Poynton 2007; Taxman & Piquero
S ALEXEEV AND D WEATHERBURN 3
1998). Wagenaar et al. (2007) reviewed 19 studies of the deterrent effect of DUI fines, but
only six provided evidence that fines are a deterrent. We have been unable to find any study
on the specific deterrent effect of fines for using or possessing a prohibited drug. It would be
imprudent, nonetheless, to assume they are a deterrent to the use and/possession of prohibited
drugs. On the one hand, fines might work because of the ‘full wallets’ effect a robust
observation that a short-term increase in income leads to a spike in drug use by the drug users
(Dobkin & Puller 2007; Riddell & Riddell 2006). This suggests that fines (i.e., a short-term
drop in income) might discourage drug use or criminal activity. On the other hand, the low
risk of detection associated with illicit drug use may deprive fines (and other sanctions for
drug use) of any deterrent effect.2
The fundamental challenge in any study of deterrence is that penalty severity does
not vary randomly across offenders. Failure to address this non-randomness could lead to
a mistaken understanding of the influence of fines on crime. In this article, we use an
instrumental variable (IV) strategy to understand the role of fines in future crimes. We exploit
the plausibly exogenous variation in the types of sanctions imposed on similar offenders
dealt with by different magistrates in the NSW Local Court. We focus on fines because
they are the most common penalty imposed by the courts on persons convicted of drug
use and/or possession in NSW (and many other jurisdictions). The data on sentencing for
drug possession offences is sourced from the NSW Bureau of Crime Statistics and Research
(BOCSAR). To foreshadow our results; we find no evidence that fines reduce the likelihood of
(a) a new offence of any type (drug or non-drug); (b) a new drug possession offence involving
any kind of prohibited drug; (c) a new drug possession offence involving the same drug type.
In the next section of this article, we provide pertinent background information on the
sentencing process in Local Courts.
2. Background
2.1. The law on drug possession
In NSW, persons found by police to be in possession or use of a prohibited drug can be
dealt with in several ways. If the drug is cannabis, the amount of drug involved is less than
15 grams, and the person admits to the offence, they can be cautioned and released. Cannabis
cautions, however, can only be given twice to the same individual and not to a person who has
prior convictions for a drug offence or prior convictions involving violence or sexual assault.
2About 3.4 million Australians reported using an illicit drug in the last 12 months (Australian Bureau of
Statistics 2020). Only about 1% of this group appeared in court charged with possessing or using a prohibited
drug (Australian Bureau of Statistics 2020).
4 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
After two cautions, or if either of the other just mentioned conditions apply, the person must
be referred to Local Court.3
For concreteness, Table 1 shows the different methods of proceeding against those
apprehended for drug use and/or possession offences in 2018. In approximately 31% of
instances, illicit drug use or possession is dealt with by the imposition of a caution. However,
if a case involving a charge of drug use and/or possession reaches court, it is almost always
dealt with by the Local Court.
Table 1. Distribution of jurisdiction
N Share
Local Court 11,533 69.47%
Cannabis caution 4,700 28.31%
Police caution 354 2.13%
District Court 13 0.08%
Drug Court 1 0.01%
Total 16,601 100%
Notes: Included cases: possession of pro-
hibited drug and self-administer/attempt self-
administer prohibited drug (lawcodes 3145
and 3151).
Source: NSW BOCSAR ROD 2018.
The law concerning the use, possession, sale, manufacture and cultivation of a prohibited
drug in NSW is set out in the NSW Drug Misuse and Trafficking Act (1985). Section 10 of
that Act makes it an offence for a person to possess a prohibited drug. Section 12 makes
it an offence for a person to self-administer or attempt to self-administer a prohibited drug.
The offences of possessing a prohibited drug and self-administering or attempting to self-
administer a prohibited drug both carry a maximum penalty of two years imprisonment or
a $2,200 fine. These are the least serious offences under the Act. They are both summary
offences, which is to say that they must be dealt with in the Local Court.
More serious drug offences, such as supplying, manufacturing, cultivating, or importing
a prohibited drug, must be dealt with in the District Court. To be charged with using or
possessing a prohibited drug, a defendant must be found in possession of no more than the
legislatively prescribed threshold quantity of the drug in question. If a person is found in
possession of any amount higher than the threshold, they are deemed by law to be involved
in supplying a prohibited drug. The District Court can accept a plea to drug possession as
an alternative to ‘deemed supply’ if the defendant can show that the drug they possessed
was solely for personal use. Such instances are rare; however where they occur, it is the
District Court that accepts the plea to a lesser charge, not the Local Court. Since drug use and
3Since 2019, persons found in possession of small quantities of other prohibited drugs can be given an
infringement notice rather than being referred to court. In the period covered by the current study, however,
all other cases involving charges of drug possession were dealt with by the NSW Local Court.
S ALEXEEV AND D WEATHERBURN 5
possession are the least serious offences under the NSW Drug Misuse and Trafficking Act
(1985), and all more serious drug offences must be dealt with by the District Court, there is
no scope within the Local Court for plea bargaining in relation to drug offences.
Among the 11,533 persons convicted of using or possessing or using a prohibited drug in
2018 in Local Courts, more than 95 per cent were convicted of possessing rather than using a
prohibited drug. Almost all (99%) of those charged with using or possessing a prohibited drug
in NSW in 2018 pleaded guilty or chose not to enter a plea (i.e., did not contest the charge).
The high guilty plea rate is not surprising given that offenders are generally only charged
when found using or in possession of a prohibited drug. In practice, as shown in Table 2,
imprisonment is rare (1.32% of cases). Slightly more than half (50.46%) of the offenders
were fined in 2018. Another third (26.95%) received a good behaviour bond.4
Table 2. Distribution of penalties in Local Courts
N Share
Fine 5,819 50.46%
Bond without conviction without supervision 2,394 20.76%
Conditional Release Order without conviction, without supervision 687 5.96%
No conviction recorded 669 5.80%
Conviction only 493 4.27%
Bond without supervision 418 3.62%
Bond with supervision 272 2.36%
No penalty 236 2.05%
Imprisonment 152 1.32%
Community Correction Order with supervision 97 0.84%
Conditional Release Order with conviction, without supervision 86 0.75%
Conditional Release Order with conviction, with supervision 56 0.49%
Community Correction Order without supervision 43 0.37%
Bond without conviction with supervision 24 0.21%
Pre-reform or Children’s Community Service Order 22 0.19%
Suspended sentence with supervision 21 0.18%
Intensive Correction Order 14 0.12%
Suspended sentence without supervision 14 0.12%
Conditional Release Order without conviction, with supervision 12 0.10%
Pre-reform Intensive Correction Order 4 0.03%
Total 11,533 100%
Notes: Included cases: possession of prohibited drug and self-administer/attempt self-administer
prohibited drug (lawcodes 3145 and 3151).
Source: NSW BOCSAR ROD 2018.
4A good behaviour bond is a form of conditional release. The offender must agree to abide by certain
conditions prescribed by the court and undertake not to commit any further offence over the period of the
bond. Failure to comply with this requirement typically results in re-arrest and imposition of more severe
punishment. A supervised bond is one where the offender must regularly report to a probation officer and abide
by his or her directions. An unsupervised bond is one with no supervision requirement. A ‘no-conviction’ bond
involves an undertaking to be of good behaviour over a specified period, but no conviction is recorded against
the offender. ‘No-conviction bonds’ are normally reserved for those who have no prior criminal convictions.
6 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
2.2. The allocation of cases to magistrates
There were 144 Local Courts and 273 magistrates in NSW in the period covered by this
study. Local Court cases in NSW are generally heard in the court nearest to where the offence
was detected or took place. Magistrates are allocated to courts by the Chief Magistrate. In
major urban locations, such as Sydney, Newcastle and Wollongong, courts are typically multi-
court complexes; with each court comprising several courtrooms and a different magistrate
presiding in each courtroom. In non-urban settings, magistrates typically preside over matters
in a single court complex. Magistrates rotate across court complexes in urban and non-urban
areas roughly every three years.
The allocation of cases to magistrates is performed by a listing clerk, who publishes
the daily allocation of cases to courts on the day before a case is set down for hearing.
Once a magistrate has received their list, he/she works their way through the list in the order
determined by the clerk. Cases that are not reached are simply adjourned for the next sitting
day or at a date deemed convenient to the parties. The salient point is that the defendant cannot
influence which magistrate deals with their case, and magistrates cannot pick and choose the
cases they hear. This means that the assignment of cases to magistrates is random, conditional
on the court. The arrangement accords with the ‘equality before the law’ principle, with the
goal of treating all cases ex-ante equally (Judicial Commission of New South Wales 2006).
It also means that the variation in the penalty imposed by different magistrates on persons
convicted of possessing or using a prohibited drug is exogenous, conditional on the controls
included in the study.
2.3. The selection of penalty
In Australia, as in other common law countries, courts enjoy considerable discretion
in the way they deal with convicted offenders. In almost all cases (including cases of drug
use and/or possession), the law specifies the maximum penalty and the types of penalties
a magistrate can choose from. Sentencing courts are then free to choose both the type and
quantum of penalty they consider appropriate given the circumstances of the offence and the
characteristics of the offender. Magistrates can obtain statistical data on the types and range of
penalties imposed for similar types of cases, but they have no legally binding force. The only
material constraint on a sentencing court in cases involving drug use or drug possession is
that the prosecutor can appeal to the District Criminal Court if he or she considers the penalty
too lenient, and defence counsel can appeal to District Court if they consider the penalty
too severe. In practice, defence and prosecution sentencing appeals in drug possession cases
are extremely rare. This is partly because the fines are generally only a few hundred dollars
(appealing the sentence would be much more expensive) and partly because defendants can
S ALEXEEV AND D WEATHERBURN 7
elect to pay their fine in instalments if they are impecunious. The wide discretion enjoyed by
magistrates in determining the penalty for an offence creates considerable scope for variation
between them in the penalty they impose on otherwise similar offenders. As we show later
on in this paper, there is substantial variation between magistrates in the penalty they choose
as well as the severity of the fines they impose.
3. Data
As noted earlier, the data for this study were drawn from the Reoffending Database
(ROD); a reoffending database constructed and maintained by the NSW BOCSAR (Hua &
Fitzgerald 2006). The ROD data used in the current study consists of (a) all Local Court
appearances where the principal offence involved possession or use of a prohibited drug; (b)
the offender was not in custody at the time of the offence; (c) the court appearance occurred
between January 1994 and January 2017.5
The selection criteria resulted in a sample of 76,366 cases. Because we aim to generate
an IV that captures magistrate leniency as accurately as possible (and, thus, have an IV that
is highly predictive of the fines), we use this large sample (referred to as Full Sample) to
construct it. However, when estimating the effect of fines on reoffending, we restrict the
sample to cases where the offender was born during or after 1984 (referred to as Estimation
Sample). This restriction ensures that we have the full criminal history of each person
involved in the study (the age of criminal responsibility is ten and our data runs from 1994).
The same approach is followed by Williams & Weatherburn (2022). They use the same data
source and a similar identification strategy.
We examine the impact of fines on three outcomes: (a) reconviction within two years
for an offence of any type; (b) reconviction within two years for a new drug use/possession
offence involving the same drug type as the index appearance; (c) reconviction within two
years for a drug use/possession offence involving any drug type.
The explanatory variable of interest is an indicator function that assigns one if a
defendant is fined and zero otherwise (see Table 2 for examples of alternative sanctions
used as the comparison group). We consulted past research using ROD to construct control
variables that correlate with recidivism (Stavrou & Poynton 2016). These are age, age
squared, gender, Aboriginal status, number of concurrent offences, an indicator for being
issued a fine in court before, an indicator for prior non-custodial orders, an indicator for
first time offender, an indicator for pleading guilty, the socioeconomic advantage of the area,
5Our data spans January 1994 and January 2019, but we can only use data till January 2017 to ensure each
defendant can be followed for two years. An exception is the reoffence of any type within two years. This
outcome is already constructed in ROD, and we can use data from January 1994 and January 2019.
8 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table 3. Descriptive statistics
Full sample Estimation sample
Variable Mean Std. Dev. Min Max Mean Std. Dev. Min Max
Fine 0.671 0.414 0 1 0.669 0.435 0 1
Reoffence of any type 0.354 0.435 0 1 0.328 0.447 0 1
Plea guilty 0.865 0.423 0 1 0.813 0.401 0 1
Number of concurrent offences 1.245 0.550 0 3 1.231 0.533 0 3
Fined before 0.415 0.493 0 1 0.300 0.458 0 1
Prior non-custodial order 0.342 0.474 0 1 0.284 0.451 0 1
Full time prison before 0.190 0.392 0 1 0.088 0.283 0 1
First time offender 0.352 0.478 0 1 0.534 0.499 0 1
Age 31.13 10.34 17 83 23.11 3.743 17 35
Female 0.159 0.366 0 1 0.170 0.376 0 1
Aboriginal 0.061 0.239 0 1 0.046 0.210 0 1
Socioeconomic disadvantage
Highly advantaged 0.242 0.429 0 1 0.302 0.459 0 1
Advantaged 0.246 0.431 0 1 0.246 0.431 0 1
Disadvantaged 0.253 0.435 0 1 0.221 0.415 0 1
Highly disadvantaged 0.259 0.438 0 1 0.231 0.421 0 1
Type of drug
Amphetamine 0.193 0.395 0 1 0.159 0.366 0 1
Cannabis 0.470 0.499 0 1 0.354 0.478 0 1
Cocaine 0.076 0.266 0 1 0.106 0.308 0 1
Ecstasy 0.185 0.389 0 1 0.322 0.467 0 1
Hallucinogens 0.009 0.094 0 1 0.014 0.118 0 1
Opiates 0.037 0.189 0 1 0.013 0.112 0 1
Other drug 0.020 0.139 0 1 0.023 0.148 0 1
Unknown drug 0.009 0.097 0 1 0.010 0.100 0 1
Areas Remoteness
Major cities 0.750 0.433 0 1 0.804 0.397 0 1
Inner region 0.193 0.395 0 1 0.155 0.362 0 1
Outer region 0.050 0.218 0 1 0.036 0.185 0 1
Remote region 0.007 0.081 0 1 0.005 0.070 0 1
Observations 76,366 33,060
Notes: Summary statistics for the variables used in the study. Full Sample is from 1994 to 2019 and is used for
generating magistrate IV. Estimation Sample, that is used to identify the parameters of interest, is limited to 1994-
2017 (to construct outcome variable), and birth cohorts 1984 onward (for the sample representativeness).
Source: NSW BOCSAR ROD.
area remoteness and drug type. Table 3 provides comprehensive descriptive statistics for the
variables used in data modelling.
Some further detail on fines and reoffence types is reported in Appendix. In particular,
Table A1 describes the distribution of fines by drug type (see Figure A1 for kernel density
estimations). The largest average fines (212$) are associated with amphetamines, the smallest
(127$) with ecstasy. Table A2 categorises the reoffence of any type variable by the crime type
using the Australian and New Zealand standard offence classification (ANZSOC) (Australian
Bureau of Statistics 2011).
S ALEXEEV AND D WEATHERBURN 9
4. Methodology
The empirical framework we use to evaluate the effect of a fine on the risk of
reoffending is based on a linear probability model (LPM). As noted earlier, we circumvent
the endogeneity of fines with an IV approach that exploits the quasi-random assignment of
magistrates to cases. Under the additional assumptions of exclusion and weak monotonicity,
magistrate assignment is a valid IV. Our main modelling approach is two-stage least squares
(2SLS) estimator that takes the form:
Icrime
ict =βI fine
ict +φXict +νict (1)
Ifine
ictj =αzctj +λXict +εict,(2)
where Icrime
ict is the indicator for reoffending in 0-24 months for individual iand case cin
year t;Xict is a vector of case- and defendant-level control variables (see Table 3 for the
list of controls) and court-year fixed effects; Ifine
ict is the endogenous variable of interest an
indicator function if fine is imposed on a drug user; zctj is a measure of magistrate tendency
to choose fine (interchangeably referred as magistrate IV or leniency). The parameter βis of
primary interest and it measures the effect of the fine on the probability of reoffending.
As is now standard practice in studies that use the judicial leniency design, we construct
zctj using the choice of penalty each magistrate imposed in all other cases, breaking the small-
sample correlation between the magistrate’s decision on a particular case and magistrate’s IV
(e.g., Dobbie, Goldin & Yang 2018).
Let the residual fine after removing the interacted court-year fixed effects and defendant-
level control be denoted by:
Ifine
ict =Ifine
ict
θXict =zctj +ρict,(3)
The residual, Ifine
ict , includes the required IV, zctj, as well as idiosyncratic defendant level
variation ρict. These residuals are then used to construct the leave-out measure of the assigned
magistrate for each case that varies across years:
zctj =1
ntj nitj ntj
X
k=0
Ifine
ikt
nitj
X
c=0
Ifine
ict !,(4)
where ntj is the total number of cases seen by magistrate jin year tand nitj is the number
of cases of defendant iseen by magistrate jin year t. Effectively, Equation (4) removes the
residual ρict from Equation (3) by aggregating the residualised fine to the tj cell level and
excluding the defendants whose fine is being instrumented.
10 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
We cluster standard errors by defendant and magistrate. Defendant-level clustering
allows for autocorrelation in outcomes between cases, which is unambiguously necessary.
Magistrate clustering additionally allows for a correlation in outcomes for defendants charged
in a similar location. This accounts for correlation in outcomes that may arise because of
common shocks (e.g., changes to policing). Note, however, that ignoring clustering on the
magistrate level (or clustering on the drug type instead) does not affect the results.
Even though LPM has well-studied limitations (e.g., Horrace & Oaxaca 2006; Lewbel,
Dong & Yang 2012), it is our preferred choice. We motivate this choice following Angrist
& Pischke (2008, p.107), who argue that extra complexities of nonlinear models matter little
when it comes to marginal effects. This holds true in our data as well. When we estimate our
specification with IV probit, the results stay largely unchanged.
Because Ifine
ict is an indicator function, the βparameters show the effect of the average
fine, ignoring a possibility that larger or smaller fines might have a different effect. To assess
the effect of fine severity, we also verify that the main results (no effect of fine on future crime)
hold when the indicator function is multiplied by the amount of the log of inflation-adjusted
fine (with a separate IV constructed for this interaction). In this case, the parameter βprovides
a characterisation of the effect of the fine and the fine amount on the probability of reoffending
(Angrist & Imbens 1995). An additional benefit of this continuous parameterisation is that
the standard errors are almost six times smaller. We also verify that the effects do not change
when we parameterise the effect to be quadratic instead of linear. This tests whether the
effects change, for example, for particularly high fines.
Our main results focus on reoffence within two years, but the results are unchanged if the
time window is increased or reduced. We also verify that the effect is the same for subgroups
of cases defined by gender, drug, and reoffence types (although the estimates for subgroups
are too imprecise to draw strong conclusions).
4.1. Instrument relevance and monotonicity
Figure 1 presents the magistrate IV’s distribution, after partialling out the court-year
fixed effects and the defendant and case observables. Superimposed over the histogram is
the non-parametric regression of an indicator for fine on magistrate IV (a flexible analog
to Equation (2)). The relationship is highly linear, suggesting that linear modelling is
appropriate. The first stage F-statistic is at least 232.09 depending on the outcome of interest,
well above rule-of-thumb cutoff of 10 (Stock & Yogo 2005) and well above a recently
advocated cutoff of 104.7 (Lee et al. 2020). Olea & Pflueger (2013) showed that F-statistic,
even adjusted for heteroskedasticity, autocorrelation, and clustering can be high even when
IVs are weak. They developed the effective F-statistic (EF). When data are conditionally
S ALEXEEV AND D WEATHERBURN 11
Figure 1. Distribution of magistrate leniency measure and first stage
Notes: This figure reports the distribution of the magistrate leniency measure that is estimated using data from
other cases assigned to a magistrate.Solid line is generated by a nonparametric regression of fine on magistrate
IV, residualizing out court-year fixed effects and controls. Dashed lines represent 95% confidence intervals
clustered at the magistrate and defendant level.
Source: NSW BOCSAR ROD.
homoskedastic and serially uncorrelated, the EF is identical to the Cragg & Donald (1993)
statistic recommended by Stock & Yogo (2002). For all outcomes, EF is higher than the
required critical value cutoff of 12.28 (this cutoff corresponds to the conventional rule-of-
thumb cutoff of 10). Table A3 in Appendix report further first stage results for all three
outcomes of interest.
Interpretability of the 2SLS estimand as LATE relies on a monotonicity assumption
(Imbens & Angrist 1994). This assumption in the current context means that if magistrate
A is more likely to choose fines than magistrate B, every defendant fined by magistrate B
would also have been fined by magistrate A, had magistrate A handled the case (for any pair
of A and B). A violation of this assumption would occur, for example, if some magistrates
were more likely to choose fine on average but less likely if a defendant is female. This
assumption may be strong (Bhuller & Sigstad 2022). Instead, we invoke the assumption of
average monotonicity, which only requires unidirectional variation between the IV and the
decision to fine for all subgroups (Frandsen, Lefgren & Leslie 2019).
Column (1) of Table 4 is dedicated to this purpose. In addition, in column (2) we
report first-stage regression on subsample using a measure of magistrate IV constructed
excluding this subsample. For example, we exclude cannabis crimes and generate the IV
on Full Sample, and then estimate the first stage on cannabis crimes of Estimation Sample.
This is known as reverse sample IV, and the idea is that magistrates should be stricter for
a specific case type (e.g., cannabis crimes) if they are stricter in other case types (e.g., all
crimes except for cannabis crimes). For all these subsamples, the first stage estimates are
large, positive, and statistically different from zero. This increases our confidence in the
12 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table 4. Examining monotonicity
(1) (2)
Dependent variable: fine
Subsample Baseline IV Reverse Sample IV
Opiates 0.844*** 0.660***
(0.0444) (0.0811)
Amphetamines 0.724*** 0.961***
(0.0888) (0.1187)
Cannabis 1.136*** 0.296***
(0.0534) (0.0985)
Cocaine 0.684*** 0.950***
(0.0959) (0.109)
Ecstasy 0.744*** 1.073***
(0.0578) (0.0717)
IV above median 0.370*** 0.242***
(0.101) (0.1225)
IV below median 0.776*** 0.920***
(0.109) (0.1420)
First offence 0.924*** 0.723***
(0.0648) (0.159)
Subsequent offence 0.565*** 0.847***
(0.0464) (0.0583)
Highly advantaged 0.567*** 0.985***
(0.0658) (0.0927)
Advantaged 0.654*** 1.049***
(0.0514) (0.0773)
Disadvantaged 0.742*** 1.041***
(0.0770) (0.0974)
Highly disadvantaged 0.998*** 0.235*
(0.0674) (0.111)
Male only 0.739*** 0.291
(0.0463) (0.241)
Female only 0.521*** 1.000***
(0.0649) (0.1940)
Notes: Column (1) reports the first stage estimate of the coefficient
on the IV, where the IV is constructed using Full Sample and
where estimation of the first stage is over the sub-sample defined
by the leftmost column. Column (2) reports the first stage estimate
of the coefficient on the IV, where the IV is constructed using the
complement of the sub-sample used for estimation and the sub-
sample used for estimation is defined by the leftmost column. Robust
standard errors clustered at the individual and magistrate level are
reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
instrument’s monotonicity assumption and provides grounds for believing our estimates can
be given a LATE interpretation. In addition to what is shown in Table 4, we also tested the
correlation on various other subgroups. The first stage estimates never turn negative.
We also explored what happens if control variables are excluded (while fixed effects are
preserved) from the regression reported in Table 4. If magistrate IV is randomly assigned,
control variables should not significantly change the estimates in Table 4, as they should
be uncorrelated with the instrument. As expected, the coefficients in Table 4 do not change
appreciably when controls are included or excluded.
It is important to emphasise the local nature of our results. Our IV estimates represent a
LATE for offenders who could have received a different sanction had their case been assigned
to a different magistrate. Intuitively, this research design compares groups of otherwise
similar individuals (a mix of IV always takers and compliers and a mix of IV never takers and
compliers) who have been fined because their cases were randomly assigned to magistrates
S ALEXEEV AND D WEATHERBURN 13
who showed differing tendencies of imposing fines. The difference in the strength of the
first stage for different subsample reported in column (1) in Table 4 already suggest that the
number of compliers varies in the subsamples. The coefficient at the first stage is higher for
the subgroups with more compliers.
To better understand our LATE, we follow Dobbie, Goldin & Yang (2018, App C) and
compare the observable characteristics in Estimation Sample and the IV compliers. Shares
reported in Table A8 (located in Appendix) show that compliers are largely indistinguishable
from the overall population. The next subsection deals in more detail with the assumed
randomness of magistrate allocation to cases.
4.2. Instrument conditional independence and exclusion restriction
For our magistrate IV to be valid, it must be uncorrelated with defendant and case
characteristics, both observed and unobserved. Table 5 verifies that magistrates in our
sample are randomly assigned to cases using the observed variables. Column (1) regresses
the decision to fine on a variety of variables measured before the court decision. The
demographic and past criminal history variables are highly predictive of the fine, with most
being individually significant. The regression explains 22% of the variation in the dependent
variable.
Column (2) examines whether the same set of characteristics can predict magistrate
IV. We see no statistically significant relationship between magistrate IV and various
demographic and past crime history variables. The estimates are all close to zero, with only
one exception. Given the number of variables we consider, having one weakly significant
coefficient is not unusual since the probability of observing one p-value at this level by
chance alone is large. Importantly the variables are not jointly significant, and the percentage
of variation explained by the model is 0.3%.
This provides evidence that cases are randomly assigned to magistrates, and therefore
that the IV is also independent of the unobserved characteristics. While many factors may
affect magistrate leniency, it is important to keep in mind that the underlying causes should
not matter for our analysis as long as magistrates are randomly assigned.
For our IV estimates to have a causal interpretation, the IV must only affect reoffending
through the channel of receiving a fine. A common concern in the magistrate leniency
design is that magistrates’ other decisions on a case influence reconviction. For example,
in the studies of the impact of incarceration, the length of incarceration (Bhuller et al.
2020; Williams & Weatherburn 2022) or a criminal record that restricts employment (Norris,
Pecenco & Weaver 2021) might be influencing reconviction. Fortunately for our research
14 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table 5. Test of randomization
(1) (2)
Dependent variables
Fine IV
Age 0.0203* -0.0159
(0.00852) (0.0121)
Age squared -0.000238 0.000310
(0.000171) (0.000250)
Female -0.0394*** -0.00833
(0.00627) (0.00760)
Plea guilty -0.000760 0.00484
(0.0119) (0.00644)
Aboriginal 0.00716 -0.0186
(0.00715) (0.0126)
Number of concurrent offences -0.130*** -0.00337
(0.0344) (0.00729)
Fined before 0.456*** 0.00518
(0.0390) (0.00795)
Prior non-custodial order 0.517*** 0.00951
(0.0420) (0.0108)
Full time prison before -0.133* 0.00791
(0.0565) (0.0135)
First time offender -0.175** -0.00374
(0.0645) (0.0164)
Socio-Economic Indexes (Highly advantaged omitted)
Advantaged 0.200*** 0.00793
(0.0336) (0.0114)
Disadvantaged 0.212*** -0.0134
(0.0401) (0.0128)
Highly disadvantaged 0.253*** 0.00102
(0.0374) (0.0136)
Areas Remoteness (Major cities omitted)
Inner region -0.116 0.0290*
(0.0615) (0.0127)
Outer region -0.0174 0.00996
(0.111) (0.0208)
Remote region 0.0853 -0.00392
(0.206) (0.0356)
Drug types (Amphetamines omitted)
Cannabis 0.209*** 0.0136
(0.0401) (0.00912)
Cocaine -0.201*** -0.0227
(0.0562) (0.0143)
Ecstasy -0.216*** -0.0144
(0.0436) (0.0110)
Hallucinogens -0.119 -0.0111
(0.110) (0.0280)
Opiates -0.248* 0.0208
(0.125) -0.00156
Other drugs -0.250** (0.00602)
(0.0900) 0.0108
Unknown drug -0.0455 (0.00613)
(0.182) -0.00142
Constant 1.489* (0.00571)
(0.608) 0.00484
Adjusted R20.221 0.031
Observations 33,060 33,060
Joint p-value <0.000 0.553
Notes: Both regressions have interacted court-year fixed effects. In
column (1), the dependent variable is an indicator for fine. In column
(2), the dependent variable is magistrate tendency to fine estimated
using data from other cases. Robust standard errors clustered at the
individual and magistrate level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
S ALEXEEV AND D WEATHERBURN 15
question, this concern is reduced, as assigning a fine implies that other punishment modalities
have been excluded (Drug Misuse and Trafficking Act 1985).
One potential concern is that the size of the fine may matter more than the fact that an
offender has been fined. Table A12 Panel F reports the results when the dummy for fine is
interacted with the fine amount (with the IV generated specifically for this interaction). We
see the same result. Another potential concern is that the tendency of magistrates to choose
a fine may also correlate with a tendency to select bonds, the second most frequent penalty.
To address this issue, we add a dummy for bonds and generate the IV for it as we do for the
fines. Table A12 Panel E shows that the key results are the same.
To rule out other sources of the exclusion restriction violation, we examine the
hypotheses that the exclusion restriction and monotonicity hold using the test recently
developed by Frandsen, Lefgren & Leslie (2019). This test can be viewed as a generalisation
of the Sargan (1958)’s over-identification test to a heterogeneous treatment effects setting.
Applied to our data, the test fails to reject the joint null that both the exclusion restriction and
monotonicity hold (p-value is 0.23).
To increase the confidence that the violation of the exclusion restriction is not affecting
the results, we also performed supplementary estimates using an imperfect instrument
estimator developed by Nevo & Rosen (2012) and a plausibly exogenous estimator developed
by Conley, Hansen & Rossi (2012). Nevo & Rosen (2012)’s estimates are appropriate when
there is knowledge of the direction of the correlation of an IV with an unobserved error
term, but not necessarily its magnitude. Their estimator relaxes the condition of exogeneity
and only requires that the instrument exhibits a weaker correlation with the equation error
term than the endogenous variable for which it is used. Using this approach gives the bounds
that are consistent with our baseline estimates. Conley, Hansen & Rossi (2012)’s estimates
are well-suited for situations in which the direction of the correlation is not known, but the
magnitude is known. Again, the bounds that we recover give no reason to think that the
exclusion restriction assumption is violated to the extent of influencing the estimates for any
outcome of interest.
5. Results
Table 6 shows the estimated effect of a fine on the probability of reoffending. Each
column corresponds to three types of reoffending within 24 months after the fine is issued.
Column (1) corresponds to reoffence of any type, not necessarily drug-related. Column (2)
corresponds to reoffence when charged with use or possession, and the drug was the same
as that for which the fine was issued. Column (3) corresponds to reoffence when charged
with the use or possession of any drug. All regressions include interacted court-year fixed
16 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
effects. Robust standard errors clustered at the individual and magistrate level are reported
throughout. The upper part of Table 6 reports ordinary least square (OLS) results with and
without the control variables. The bottom part of Table 6 reports 2SLS results with and
without the control variables (controls are also excluded from eq. (3)). Since the sample
changes slightly depending on each outcome, we report the number of observations, the
number of magistrates, courts and defendants, and the dependent variable’s mean value.
Table 6. Fine and reoffending in 0-2 years: main estimates
(1) (2) (3)
Dependent variables
Reoffence Use/possession Use/possession
of any type reoffence reoffence
(any drug) (same drug)
OLS 0.185*** 0.0936*** 0.0651***
(no controls) (0.00705) (0.00577) (0.00438)
OLS 0.0898*** 0.0536*** 0.0318***
(with controls) (0.00516) (0.00564) (0.00419)
2SLS 0.0300 0.0364 -0.00286
(no controls) (0.0435) (0.0359) (0.0266)
2SLS -0.0268 0.0108 -0.0178
(with controls) (0.0395) (0.0390) (0.0275)
Observations 33,060 23,219 23,185
N of magistrates 273 259 259
N of courts 144 143 143
N of defendants 27,931 19,759 19,728
Dependent mean 0.328 0.164 0.102
Notes: The dependent variable is in the header. The independent variable is
a indicator if fine is chosen as a sanction. 2SLS instruments for fine using a
magistrate leniency measure estimated using data from other cases assigned to
a magistrate. All regressions include interacted court-year fixed effects. Robust
standard errors clustered at the individual and magistrate level are reported in
parentheses. See Table A4 for unabridged reporting table.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
The first model reports the OLS results with only court-year fixed effects. The OLS
estimates from the parsimonious model suggest that the imposition of fine increases the
risk of reconviction. The control variables’ inclusion reduces the estimates substantially. For
example, without controls, being fined raises the probability of reconviction for any crime
by 0.185 percentage points. The addition of control variables reduces this estimate by 51%.
This uniform reduction suggests that controls are important for addressing potential omitted
variable bias. We also confirm endogeneity issues applying the Durbin-Wu-Hausman test.
For all OLS models test shows endogeneity with p-value less than 0.000.
The third and the fourth models report 2SLS results where we instrument for fine
using the leave-out measure. These estimates improve upon our OLS estimates by exploiting
plausibly exogenous variation in the decision to fine from the quasi-random assignment
of cases to magistrates. Our estimates show a noisy zero for all three dependent variables
regardless of whether we include or exclude controls. The fine does not affect the probability
of reoffending.
S ALEXEEV AND D WEATHERBURN 17
5.1. Robustness checks
To establish the robustness of the main estimates reported in Table 6, tables contained in
the Appendix report additional results. Table A5 reports the average marginal effects of a fine
on future crime when our specification is estimated with probit instead of LPM. For IV probit
models, we follow the procedure recommended in Wooldridge (2002, Ch. 15). The results
are remarkably similar to those reported by LPM. This is partly because our LPM predicted
probabilities do not fall outside the unit interval.
Table A6 reports LPM result when instead of 0-2 years, the follow-up window for
reoffence is extended to 0-4, 0-6 and 0-8 years. We see no effect for use/possession reoffence
(any drug) only. Use/possession reoffence (same drug) outcome similarly shows that there is
no fine effect on reoffending. We offer estimates for other follow-up periods only for drug-
related crimes, as we only have data for drug crimes and can construct reoffending measures
with different time windows only for these crimes. Recidivism for any type of crime for 0-2
years is already calculated in the data. Table A7, Table A9 and Table A10 report the result
of the interaction of the fine by, respectively, gender, drug, and reoffence type. Results for all
subpopulations point to no effect of fines on future crimes.
Table A12 Panels A and B confirm that generating the IV separately for different drugs
or separately for cannabis and other drugs does not change the results. Panels C and D show
that alternative standard error clustering levels does not influence the result. Panel E extends
our 2SLS model by adding a dummy for bonds (all types grouped together) and generating
the IV for this dummy following the step described in (3) and (4). This once again tests the
exclusion restriction. Panel F interacts the dummy for fine with the fine amount. This test
whether the amount rather than the average fine have a different effect. Table A11 extends
the results reported in Panel F and shows that there is no change in results when we assume
the relationship between the fine amount and reoffending is quadratic instead of linear. This
is important because it allows for the possibility that very high fines have a deterrent effect.
6. Discussion
Many countries rely heavily on fines to deter people from using or possessing prohibited
drugs. The reliance on fines may be a valuable source of revenue for the Government, but
criminal prosecution and high fines can have effects that are arguably out of proportion to
the seriousness of the offence for which they are used. Recognising this, countries such as
Australia, the United Kingdom and New Zealand have opted for diversion or cautioning
schemes when dealing with minor drug offences. Even in these countries, however, large
numbers of individuals found using or possessing prohibited drugs are still prosecuted and
18 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
fined. The effectiveness of these fines in deterring users of prohibited drugs from further drug
crimes has never been examined. Our aim in this study is to see whether those who receive
fines for possessing a prohibited drug are, as a result, less likely to reoffend than those who
did not.
To investigate this issue, we exploit plausibly exogenous variation in the tendency to
choose a penalty. In our case, the penalty is the fine imposed by different magistrates on
offenders convicted of possessing or using a prohibited drug. Our IV approach is combined
with an extensive set of controls and a large sample. We also examine a range of different
outcomes. Despite this, we find no evidence that fines, whether high or low, reduce the risk
of (a) reconviction for an offence of any type (b) reconviction for a new drug use/possession
offence (involving the same drug type as the index appearance) or (c) reconviction for a drug
use/possession offence involving any drug type.
The LATE interpretation of the IV estimates highlights that our conclusions concerning
lack of evidence only apply to offenders who would have received a different penalty if a
different magistrate had dealt with them. Therefore, other identification strategies are needed
to fully understand the effects of fines. Our sample consists of individuals who had exhausted
their caution limits, had more than 15 grams of substance or have prior convictions; therefore,
the findings may not generalise for less ‘professionalised’ criminals.
Studies that use the national survey and cross-regional variation in fines are an important
complement to our work (Farrelly et al. 2001; Saffer & Chaloupka 1999; Chaloupka et al.
2009). They show that higher fines for drug possession lower the probability of a young adult
using drugs, but there is no effect on the frequency of use. This suggests that fines may have
a general deterrent effect on non-users but underscore that neither a fine nor its magnitude
influences the likelihood of continued illicit drug use.
Our results are consistent with the literature on the specific deterrent effect of
sanctioning drug users (Green & Winik 2010; Weatherburn & Yeong 2021). They do not,
however, shed light on why fining individuals convicted of prohibited drug possession has no
measurable effect on their behaviour. The work of Bernheim & Rangel (2004) that initiated
the field of behavioural public economics may be applicable. Their model is grounded in
neuroscience research and constitutes a radical departure from the rational economic models
of addiction and deterrence (Becker & Murphy 1988; Becker 1968). It is based on the idea that
repeated drug use sensitises individuals to environmental cues that trigger irrational usage.
This implies that sanctioning, which is predicated on rationality, cannot work on drug users.
There is no doubt about the harms associated with illicit drug use, even if the scale
of these harms varies from drug to drug (Darke, Lappin & Farrell 2019). Some means
must be found for reducing those harms, but the results presented here give little reason
to believe that prosecuting and fining drug users will exert the desired effect. One approach
S ALEXEEV AND D WEATHERBURN 19
would be to decriminalise the less harmful drugs (e.g., cannabis) and expand the treatment
opportunities (Arora & Bencsik 2021). This approach, however, is not likely to attract much
public or political support in the case of drugs such as heroin, methamphetamine, and cocaine
(Weatherburn, Alexeev & Livingston 2021).
If sanctions for drug use are to be preserved, Galenianos, Pacula & Persico (2012)
have suggested a way of making them more effective. Illicit drug markets are subject to
moral hazard. The moral hazard results from a seller’s capacity to secretly adulterate the
product, while the illegality of trade prevents the emergence of institutions that solve similar
informational problems in legal markets (e.g. third-party certification, product guarantees,
customer reviews). Because these inefficiencies undermine drug markets, economic theory
suggests that policy should create an incentive for sellers to dilute the drugs they sell. If
fines (or sentences) were set proportionally to drug purity, drug sellers would have a direct
incentive to ‘cheat. This would, at least in theory, reduce the potency, increase the price and
thereby help suppress the illicit drug market.
20 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
References
ANGRIST,J.&PISCHKE, J. (2008). Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton
University Press. URL https://books.google.com.au/books?id=ztXL21Xd8v8C.
ANGRIST, J.D. & I MBENS, G.W. (1995). Two-stage least squares estimation of average causal effects in
models with variable treatment intensity. Journal of the American Statistical Association 90, 431–442.
doi:10.1080/01621459.1995.10476535. URL https://www.tandfonline.com/doi/abs/10.1080/01621459.
1995.10476535. https://www.tandfonline.com/doi/pdf/10.1080/01621459.1995.10476535.
AROR A, A. & BEN CS IK , P. (2021). Policing substance use: Chicago’s treatment
program for narcotics arrests. Working paper URL https://drive.google.com/file/d/
1QSxrv7pWWWCffL4O2hPoHBKQQHdU1UCk/view.
AUSTRALIAN BUR EAU OF STATISTICS (2011). Australian and New Zealand standard offence classification
(ANZSOC). URL https://www.abs.gov.au/ausstats/abs@.nsf/mf/1234.0.
AUSTRALIAN BUR EAU OF STATISTICS (2020). Criminal courts, australia, 2018-2019. cat.no. 4513.0. URL
https://www.abs.gov.au/ausstats/abs@.nsf/mf/4513.0.
BECKER, G.S. (1968). Crime and punishment: An economic approach. Journal of Political Economy 76,
169–217. URL http://www.jstor.org/stable/1830482.
BECKER, G.S. & MURPHY, K.M. (1988). A theory of rational addiction. Journal of Political Economy 96,
675–700. doi:10.1086/261558. URL https://doi.org/10.1086/261558. https://doi.org/10.1086/261558.
BERNHEIM, B.D. & RANGEL, A. (2004). Addiction and cue-triggered decision processes. American
Economic Review 94, 1558–1590. doi:10.1257/0002828043052222. URL https://www.aeaweb.org/
articles?id=10.1257/0002828043052222.
BHU LLE R, M., DAHL, G.B., LØ KEN, K. V. & MOGS TAD, M. (2020). Incarceration, recidivism, and
employment. Journal of Political Economy 128, 1269–1324. doi:10.1086/705330. URL https:
//doi.org/10.1086/705330. https://doi.org/10.1086/705330.
BHU LLE R, M. & SI GSTA D, H. (2022). Errors and monotonicity in judicial decision-making.
Economics Letters 215, 110486. doi:https://doi.org/10.1016/j.econlet.2022.110486. URL https://www.
sciencedirect.com/science/article/pii/S0165176522001203.
BRISCOE, S. (2004). Raising the bar: can increased statutory penalties deter drink-drivers? Accident
Analysis & Prevention 36, 919 929. doi:https://doi.org/10.1016/j.aap.2003.10.005. URL http:
//www.sciencedirect.com/science/article/pii/S0001457503001556.
CHALOUPKA, F., GRO SS MA N, M. , BICKEL,W.&SAFF ER , H. (2009). The Economic Analysis of
Substance Use and Abuse: An Integration of Econometric and Behavioral Economic Research.
Chapter 5: The Demand for Cocaine and Marijuana by Youth. National Bureau of Economic
Research Conference Report, University of Chicago Press. URL https://books.google.com.au/books?
id=NLe3LLiszmYC.
CON LEY, T., HANS EN , C. & RO SSI , P. (2012). Plausibly exogenous. The Review of Economics and
Statistics 94, 260–272. doi:10.1162/REST\a\00139. URL https://doi.org/10.1162/REST a 00139.
https://doi.org/10.1162/REST a 00139.
CRAG G, J.G. & D ONA LD , S.G. (1993). Testing identifiability and specification in instrumental variable
models. Econometric Theory 9, 222–240. URL http://www.jstor.org/stable/3532477.
CUT HBE RTS ON , P. (2017). A guide to sentencing data. URL https://www.civitas.org.uk/content/files/
aguidetothesentencingdata.pdf. Civitas: Institute for the Study of Civil Society.
DAR KE, S ., LAPPIN, J. & FAR RE LL, M. (2019). The Clinician’s Guide to Illicit Drugs and Health.
Silverback Publishing. URL https://books.google.com.au/books?id=3eiuxQEACAAJ.
DOBBIE, W., GOLDIN, J. & YANG, C.S . (2018). The effects of pretrial detention on conviction, future
crime, and employment: Evidence from randomly assigned judges. American Economic Review 108,
201–40. doi:10.1257/aer.20161503. URL https://www.aeaweb.org/articles?id=10.1257/aer.20161503.
S ALEXEEV AND D WEATHERBURN 21
DOBKIN, C. & PULL ER , S.L. (2007). The effects of government transfers on monthly cycles in drug abuse,
hospitalization and mortality. Journal of Public Economics 91, 2137–2157. doi:https://doi.org/10.1016/
j.jpubeco.2007.04.007. URL https://www.sciencedirect.com/science/article/pii/S0047272707000631.
DRUG MI SU SE AN D TRAFFICKING ACT (1985). New south wales legislation. URL http://classic.austlii.edu.
au/au/legis/nsw/consol act/dmata1985256/.
FARR EL LY, M.C., BRAY, J.W., ZARKIN, G.A. & WENDLING, B.W. (2001). The joint demand for
cigarettes and marijuana: evidence from the national household surveys on drug abuse. Journal
of Health Economics 20, 51–68. doi:https://doi.org/10.1016/S0167-6296(00)00067-9. URL https:
//www.sciencedirect.com/science/article/pii/S0167629600000679.
FRANDSEN, B.R., LEF GRE N, L.J. & LES LI E, E.C. (2019). Judging judge fixed effects. Working Paper
25528, National Bureau of Economic Research. doi:10.3386/w25528. URL http://www.nber.org/
papers/w25528.
GALENIANOS, M., PACUL A, R.L. & P ER SI CO, N. (2012). A search-theoretic model of the retail market
for illicit drugs. The Review of Economic Studies 79, 1239–1269. URL http://www.jstor.org/stable/
23261385.
GRE EN, D . & WINIK, D. (2010). Using random judge assignment to eestimate the effects of incarceration
and probation on recidivism among drug offenders. Criminology 48, 357–387. doi:10.1111/
j.1745-9125.2010.00189.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1745-9125.2010.
00189.x. https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1745-9125.2010.00189.x.
GROG GE R, J . (1995). The Effect of Arrests on the Employment and Earnings of Young Men. The Quarterly
Journal of Economics 110, 51–71. doi:10.2307/2118510. URL https://doi.org/10.2307/2118510.
https://academic.oup.com/qje/article-pdf/110/1/51/5350414/110-1-51.pdf.
HOR RAC E, W.C. & OAXACA, R.L. (2006). Results on the bias and inconsistency of ordinary least squares
for the linear probability model. Economics Letters 90, 321 327. doi:https://doi.org/10.1016/j.econlet.
2005.08.024. URL http://www.sciencedirect.com/science/article/pii/S0165176505003150.
HUA, J. & FITZGERALD, J. (2006). Matching court records to measure reoffending. BOCSAR NSW Crime
and Justice Bulletins , 12URL https://www.bocsar.nsw.gov.au/Documents/CJB/cjb95.pdf.
IMBENS, G.W. & ANGRIST, J.D . (1994). Identification and estimation of local average treatment effects.
Econometrica 62, 467–475. URL http://www.jstor.org/stable/2951620.
JUDICIAL COMMISSION OF NEW SO UTH WAL ES (2006). Equality before the law: Bench book. URL
http://www.ncjrs.gov/App/publications/abstract.aspx?ID=237101.
LEE , D.S., MCCRARY, J., MOREIRA, M.J. & P ORT ER , J. (2020). Valid t-ratio inference for iv. URL
https://arxiv.org/abs/2010.05058. 2010.05058.
LEWBEL, A., DON G, Y. & YANG, T.T. (2012). Comparing features of convenient estimators for binary
choice models with endogenous regressors. Canadian Journal of Economics/Revue canadienne
d’´
economique 45, 809–829. doi:https://doi.org/10.1111/j.1540- 5982.2012.01733.x. URL https:
//onlinelibrary.wiley.com/doi/abs/10.1111/j.1540-5982.2012.01733.x. https://onlinelibrary.wiley.com/
doi/pdf/10.1111/j.1540-5982.2012.01733.x.
MAXWELL, A. (2017). Adult criminal court statistics in canada, 2014/2015. Juristat: Canadian
Centre for Justice Statistics , 1–21URL https://www150.statcan.gc.ca/n1/pub/85- 002-x/2017001/
article/14699-eng.htm.
MOFF ATT,S.&POYN TON , S. (2007). The deterrent effect of higher fines on recidivism: Driving offences.
NSW Bureau of Crime Statistics and Research URL https://www.bocsar.nsw.gov.au/Publications/CJB/
cjb106.pdf.
NAGI N, D.S., CU LLE N, F.T. & JONSON, C.L. (2009). Imprisonment and reoffending. Crime and Justice 38,
115–200. doi:10.1086/599202. URL https://doi.org/10.1086/599202. https://doi.org/10.1086/599202.
NEVO ,A.&ROS EN , A. M. (2012). Identification with imperfect instruments. The Review of Economics and
Statistics 94, 659–671. doi:10.1162/REST\a\00171. URL https://doi.org/10.1162/REST a 00171.
https://doi.org/10.1162/REST a 00171.
22 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
NORRIS, S., PEC EN CO , M. & WE AVER , J. (2021). The effects of parental and sibling incarceration:
Evidence from ohio. American Economic Review 111, 2926–63. doi:10.1257/aer.20190415. URL
https://www.aeaweb.org/articles?id=10.1257/aer.20190415.
OLE A, J.L.M. & PFLU EG ER , C. (2013). A robust test for weak instruments. Journal of Business & Economic
Statistics 31, 358–369. doi:10.1080/00401706.2013.806694. URL https://doi.org/10.1080/00401706.
2013.806694. https://doi.org/10.1080/00401706.2013.806694.
QUI LTER , J. & HOGG, R. G. (2018). The hidden punitiveness of fines. International Journal for Crime,
Justice and Social Democracy URL https://www.crimejusticejournal.com/article/view/914.
RID DEL L, C. & RI DDE LL , R. (2006). Welfare checks, drug consumption, and health: Evidence from
vancouver injection drug users. The Journal of Human Resources 41, 138–161. URL http://www.
jstor.org/stable/40057260.
SAFF ER, H . & CHALOUPKA, F. (1999). The demand for illicit drugs. Economic Inquiry 37, 401–411. doi:
https://doi.org/10.1111/j.1465-7295.1999.tb01439.x. URL https://onlinelibrary.wiley.com/doi/abs/10.
1111/j.1465-7295.1999.tb01439.x. https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1465-7295.1999.
tb01439.x.
SARGAN, J.D. (1958). The estimation of economic relationships using instrumental variables. Econometrica
26, 393–415. URL http://www.jstor.org/stable/1907619.
STAVROU, E . & POY NTON, S. (2016). The revised group risk assessment model (gram
2): Assessing risk of reoffending among adults given non-custodial sanctions. NSW
Bureau of Crime Statistics and Research URL https://www.bocsar.nsw.gov.au/Publications/CJB/
Report-2016-GRAM2-Group-Risk-Assessment- Model-CJB197.pdf.
STOC K, J. & YOG O, M . (2005). Asymptotic Distributions of Instrumental Variables Statistics with Many
Instruments. New York: Cambridge University Press, pp. 109–120. URL http://www.economics.
harvard.edu/faculty/stock/files/AsymptoticDistrib Stock%2BYogo.pdf.
STOC K, J.H. & YOG O, M. (2002). Testing for weak instruments in linear iv regression. Working Paper 284,
National Bureau of Economic Research. doi:10.3386/t0284. URL http://www.nber.org/papers/t0284.
TAXM AN , F.S. & PIQU ERO , A. (1998). On preventing drunk driving recidivism: an examination of
rehabilitation and punishment approaches. Journal of Criminal Justice 26, 129 143. doi:
https://doi.org/10.1016/S0047-2352(97)00075-5. URL http://www.sciencedirect.com/science/article/
pii/S0047235297000755.
WAGENA AR , A.C., MAL DON ADO -MOL INA , M.M., MA, L., TOB LER , A. L. & KO MRO , K. A. (2007).
Effects of legal bac limits on fatal crash involvement: Analyses of 28 states from 1976 through
2002. Journal of Safety Research 38, 493 499. doi:https://doi.org/10.1016/j.jsr.2007.06.001. URL
http://www.sciencedirect.com/science/article/pii/S0022437507000886.
WALDFOG EL , J. (1994). Does conviction have a persistent effect on income and employment? International
Review of Law and Economics 14, 103 119. doi:https://doi.org/10.1016/0144-8188(94)90039-6.
URL http://www.sciencedirect.com/science/article/pii/0144818894900396.
WEATH ERBUR N, D., AL EX EE V,S.&LIVINGSTON, M. (2021). Changes in and correlates of australian
public attitudes toward illicit drug use. Drug and Alcohol Review doi:https://doi.org/10.1111/dar.13426.
URL https://onlinelibrary.wiley.com/doi/abs/10.1111/dar.13426. https://onlinelibrary.wiley.com/doi/
pdf/10.1111/dar.13426.
WEATH ERBUR N, D. & YEO NG , S. (2021). Does correctional supervision of amphetamine users reduce
the risk of re-offending? Addiction 116, 1472–1481. doi:https://doi.org/10.1111/add.15298. URL
https://onlinelibrary.wiley.com/doi/abs/10.1111/add.15298. https://onlinelibrary.wiley.com/doi/pdf/10.
1111/add.15298.
WILKINS, C. & SWEETSUR, P. (2012). Criminal justice outcomes for cannabis use offences in new zealand,
1991–2008. International Journal of Drug Policy 23, 505 511. doi:https://doi.org/10.1016/j.drugpo.
2012.03.002. URL http://www.sciencedirect.com/science/article/pii/S0955395912000485. Special
Focus Issue: Drug Production and Drug Markets.
S ALEXEEV AND D WEATHERBURN A1
WILLIAMS, J. & WEATHERB UR N, D . (2022). Can Electronic Monitoring Reduce Reoffending? The Review
of Economics and Statistics 104, 232–245. doi:10.1162/rest a 00954. URL https://doi.org/10.1162/
rest a 00954. https://direct.mit.edu/rest/article-pdf/104/2/232/1996118/rest a 00954.pdf.
WOOLDRIDGE, J. (2002). Econometric Analysis of Cross Section and Panel Data. MIT Press. URL
https://books.google.com.au/books?id=cdBPOJUP4VsC.
YU, J. (1994). Punishment celerity and severity: Testing a specific deterrence model on drunk driving
recidivism. Journal of Criminal Justice 22, 355 366. doi:https://doi.org/10.1016/0047-2352(94)
90082-5. URL http://www.sciencedirect.com/science/article/pii/0047235294900825.
Appendix
Figure A1. Distribution of fines by drug type
Notes: This figure reports the distribution of the fine by drug type. Distribution is estimated with kernel density
estimation (epanechnikov).
Source: NSW BOCSAR ROD.
Table A1. Descriptive statistics for inflation adjusted fine by drug
N Mean Min 1st 5th 10th 25th 50th 75th 90th 95th 99th Max Skew
Amphetamines 4,956 212.97 0.04 2.04 6.24 7.18 12.72 25.44 369.90 561 765.0 1,109.60 2,080.00 1.65
Cannabis 9,510 236.25 0.07 3.18 6.47 9.45 13.34 208.00 369.90 545 684.0 1,079.00 2,320.00 1.49
Cocaine 4,233 149.18 0.18 4.64 6.24 6.84 12.24 14.37 222.40 520 686.4 1,079.00 2,080.00 2.33
Ecstasy 9,776 127.81 0.28 4.24 6.24 6.54 12.24 13.92 107.90 464 616.5 1,048.05 3,137.50 3.12
Hallucinogens 1,918 139.87 0.24 3.48 6.24 6.96 12.48 18.72 235.80 436 570.0 1,020.00 1,272.00 2.08
Opiates 1,782 202.99 1.02 1.08 3.12 6.47 12.72 25.20 342.60 510 647.4 1,085.04 1,872.00 2.15
Other drug 2,123 170.83 0.24 1.02 6.12 6.44 12.24 18.36 285.00 520 684.0 1,560.00 1,836.00 2.75
Unknown 1,797 177.62 0.25 0.47 3.12 6.24 12.24 19.08 318.00 510 612.0 1,020.00 1,500.00 1.68
Notes: Summary statistics for the fine variable on Estimation Sample.
Source: NSW BOCSAR ROD.
A2 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table A2. ANZSOC classification of the reoffence of any type variable
N Share
Illicit drug offences 13,067 29.5%
Traffic and vehicle regulatory offences 11,545 26.1%
Acts intended to cause injury 5,579 12.6%
Theft and related 4,113 9.3%
Public order 2,041 4.6%
Property damage and related 1,543 3.5%
Justice procedure offences 1,367 3.1%
Dangerous and negligent acts 1,131 2.6%
Weapons 1,018 2.3%
Fraud and related 821 1.9%
Unlawful entry and related 723 1.6%
Miscellaneous 657 1.5%
Abduction/Harassment 269 0.6%
Robbery and related 202 0.5%
Sexual assault and related 139 0.3%
Homicide and related 30 0.1%
Notes: Summary statistics for the reoffence of any type variable on
Estimation Sample. See Australian Bureau of Statistics (2011) for
the definitions of categories.
Source: NSW BOCSAR ROD.
Table A3. First stage
(1) (2) (3)
Dependent variable: fine
Reoffence Use/possession Use/possession
of any type reoffence reoffence
(any drug) (same drug)
IV 0.744*** 0.724*** 0.736***
(0.0444) (0.0888) (0.0534)
N 33,060 23,219 23,185
F 274.11 232.09 237.99
EF 123.3 122.1 101.1
Notes: The table reports the first stage estimates. Robust
standard errors clustered at the individual and magistrate level
are reported in parentheses. F refers to the F-statistic of
excluded instruments, and EF refers to the effective F-statistic.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
S ALEXEEV AND D WEATHERBURN A3
Table A4. Fine and reoffending in 0-2 years: unabridged reporting Table 6
(1) (2) (3) (4) (5) (6)
Dependent variables Dependent variables
Reoffence Use/possession Use/possession Reoffence Use/possession Use/possession
of any type reoffence reoffence of any type reoffence reoffence
(any drug) (same drug) (any drug) (same drug)
OLS 2SLS
Fine 0.0898*** 0.0536*** 0.0318*** -0.0268 0.0108 -0.0178
(0.00516) (0.00564) (0.00419) (0.0395) (0.0390) (0.0275)
Age -0.0839*** -0.0711*** -0.0368** 0.00169*** 0.00149*** 0.000748**
(0.0121) (0.0124) (0.0113) (0.000253) (0.000275) (0.000254)
Age squared 0.00165*** 0.00150*** 0.000763** -0.0201*** -0.0113 -0.00477
(0.000257) (0.000275) (0.000253) (0.00585) (0.00654) (0.00535)
Female -0.0245*** -0.0126 -0.00634 0.109*** 0.00782 -0.00329
(0.00619) (0.00669) (0.00541) (0.0136) (0.0144) (0.0116)
Plea guilty -0.0607*** -0.0601*** -0.0617*** -0.0507*** -0.0502*** -0.0517***
(0.0104) (0.0103) (0.0116) (0.0115) (0.0114) (0.0126)
Aboriginal 0.106*** 0.00634 -0.00503 0.0104* 0.000569 -0.0111**
(0.0138) (0.0147) (0.0118) (0.00472) (0.00500) (0.00384)
Number of concurrent offences 0.0107* 0.000756 -0.0108** 0.131*** 0.0562*** 0.0313***
(0.00466) (0.00492) (0.00387) (0.00630) (0.00667) (0.00609)
Fined before 0.154*** 0.0642*** 0.0406*** 0.0971*** 0.0238*** 0.0159**
(0.00943) (0.00972) (0.00852) (0.00694) (0.00664) (0.00517)
Prior non-custodial order 0.114*** 0.0303*** 0.0235*** 0.141*** -0.000812 -0.0180
(0.00926) (0.00899) (0.00681) (0.0110) (0.0117) (0.0102)
Full time prison before 0.159*** 0.00695 -0.0176 0.157*** 0.00660 -0.0179
(0.0160) (0.0148) (0.0123) (0.0159) (0.0147) (0.0122)
First time offender -0.175** -0.133* -0.130*** -0.173** -0.113* -0.125***
(0.0645) (0.0565) (0.0344) -0.0644 -0.0567 -0.0331
Socio-Economic Indexes (Highly advantaged omitted)
Advantaged 0.0189** 0.00705 0.00654 0.0132* 0.00467 0.00379
(0.00671) (0.00615) (0.00564) (0.00617) (0.00589) (0.00540)
Disadvantaged 0.0297*** 0.00598 0.00644 0.0221** 0.00321 0.00323
(0.00738) (0.00680) (0.00586) (0.00687) (0.00634) (0.00546)
Highly disadvantaged 0.0607*** 0.0394*** 0.0286*** 0.0518*** 0.0358*** 0.0245***
(0.00926) (0.00780) (0.00679) (0.00857) (0.00699) (0.00634)
Areas Remoteness (Major cities omitted)
Inner region -0.0200 -0.00811 -0.00869 -0.0188 -0.00785 -0.00838
(0.0106) (0.0103) (0.00855) (0.0106) (0.0104) (0.00868)
Outer region -0.0450* -0.0425* -0.0370* -0.0413 -0.0414* -0.0358*
(0.0223) (0.0179) (0.0143) (0.0221) (0.0181) (0.0143)
Remote region 0.0103 -0.125** -0.0627 0.0137 -0.127** -0.0641
(0.0538) (0.0450) (0.0388) (0.0539) (0.0448) (0.0382)
Drug types (Amphetamines omitted)
Cannabis -0.0247** 0.0171 0.0743*** -0.0316*** 0.0141 0.0709***
(0.00936) (0.00928) (0.00766) (0.00894) (0.00886) (0.00726)
Cocaine -0.173*** -0.107*** -0.0679*** -0.160*** -0.103*** -0.0624***
(0.0112) (0.0114) (0.00744) (0.0106) (0.0102) (0.00653)
Ecstasy -0.194*** -0.0902*** -0.0558*** -0.178*** -0.0850*** -0.0497***
(0.0101) (0.00915) (0.00646) (0.00865) (0.00812) (0.00555)
Hallucinogens -0.166*** -0.0840*** -0.0573*** -0.152*** -0.0791*** -0.0516***
(0.0207) (0.0184) (0.0110) (0.0203) (0.0185) (0.0105)
Opiates 0.0667** -0.0185 -0.0604** 0.0627* -0.0192 -0.0613***
(0.0254) (0.0244) (0.0182) (0.0249) (0.0245) (0.0183)
Other drugs -0.0812*** -0.0498* -0.0803*** -0.0718*** -0.0462* -0.0761***
(0.0207) (0.0213) (0.0109) (0.0201) (0.0210) (0.0103)
Unknown drug -0.0101 -0.0356 -0.112*** -0.00608 -0.0351 -0.111***
(0.0255) (0.0310) (0.0212) (0.0252) (0.0309) (0.0209)
Observations 33,060 23,219 23,185 33,060 23,219 23,185
Notes: The dependent variable is in the header. The independent variable is an indicator if fine is chosen as a penalty. 2SLS instrument for fine using a
magistrate leniency measure estimated using data from other cases assigned to a magistrate. All regressions include interacted court-year fixed effects.
Robust standard errors clustered at the individual and magistrate level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
A4 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table A5. Fine and reoffending in 0-2 years: robustness to functional form
(1) (2) (3)
Dependent variables
Reoffence Use/possession Use/possession
of any type reoffence reoffence
(any drug) (same drug)
Probit 0.194*** 0.0999*** 0.0671***
(no controls) (0.00505) (0.00477) (0.00448)
Probit 0.0897*** 0.0537*** 0.0316***
(with controls) (0.00416) (0.00514) (0.00411)
IV probit 0.0194 0.0134 0.00081
(no controls) (0.0436) (0.0349) (0.0266)
IV probit -0.0069 0.0117 -0.0068
(with controls) (0.0314) (0.0581) (0.0204)
Observations 33,060 23,219 23,185
Notes: The dependent variable is in the header. The independent variable
is an indicator if fine is chosen as a penalty. 2SLS instruments for fine
using a magistrate leniency measure estimated using data from other cases
assigned to a magistrate. Marginal effects are reported. All regressions
include interacted court-year fixed effects. Robust standard errors clustered
at the individual and magistrate level are reported in parentheses.
Source: NSW BOCSAR ROD.
Table A6. Fine and use/possession reoffence (any drug): robustness to time of reoffence
(1) (2) (3)
Dependent variable: use/possession reoffence (any drug)
in 0-4 years in 0-6 years in 0-8 years
OLS 0.121*** 0.104*** 0.0640***
(no controls) (0.00849) (0.00840) (0.00750)
OLS 0.0599*** 0.0537*** 0.0282***
(with controls) (0.00827) (0.00805) (0.00761)
2SLS 0.0925 0.0666 -0.0373
(no controls) (0.0618) (0.0552) (0.0567)
2SLS 0.0720 0.0605 -0.0221
(with controls) (0.0594) (0.0545) (0.0515)
Observations 15,887 11,072 8,142
N of magistrates 251 249 248
N of courts 144 143 143
N of defendants 8,835 5,908 6,556
Dependent mean 0.332 0.473 0.638
Notes: The dependent variable is in the header. The independent
variable is an indicator if fine is chosen as a penalty. 2SLS instruments
for fine using a magistrate leniency measure estimated using data
from other cases assigned to a magistrate. All regressions include
interacted court-year fixed effects. Robust standard errors clustered at
the individual and magistrate level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
S ALEXEEV AND D WEATHERBURN A5
Table A7. Fine and reoffending in 0-2 years: gender heterogeneity
(1) (2) (3)
Dependent variables
Reoffence Use/possession Use/possession
of any type reoffence reoffence
(any drug) (same drug)
Male 0.184*** 0.0924*** 0.0669***
OLS (0.00712) (0.00578) (0.00464)
(no controls) Female 0.188*** 0.101*** 0.0547***
(0.0124) (0.0127) (0.00959)
Male 0.0807*** 0.0471*** 0.0305***
OLS (0.00568) (0.00575) (0.00451)
(with controls) Female 0.133*** 0.0873*** 0.0386***
(0.00993) (0.0132) (0.0102)
Male 0.0138 0.0400 0.00786
2SLS (0.0460) (0.0370) (0.0292)
(no controls) Female 0.112 0.0177 -0.0589
(0.0925) (0.0668) (0.0526)
Male -0.0318 0.0200 -0.00636
2SLS (0.0398) (0.0393) (0.0292)
(with controls) Female -0.00176 -0.0389 -0.0800
(0.0738) (0.0711) (0.0506)
Observations 33,060 23,219 23,185
Notes: The dependent variable is in the header. The independent variable is a dummy
for fine interacted by the gender dummy. 2SLS instruments for fine using a magistrate
leniency measure estimated using data from other cases assigned to a magistrate. All
regressions include interacted court-year fixed effects. Robust standard errors clustered
at the individual and magistrate level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
Table A8. Characterization of compliers
(1) (2)
Sample Complier
share share
Opiates 0.013 0.063
Amphetamines 0.159 0.100
Cannabis 0.354 0.869
Cocaine 0.106 0.092
Ecstasy 0.322 0.307
First offence 0.534 0.548
Highly advantaged 0.302 0.102
Advantaged 0.246 0.132
Disadvantaged 0.221 0.235
Highly disadvantaged 0.231 0.659
Male 0.830 0.920
Female 0.170 0.045
Notes: For each subgroup, we report the sample
share, the complier share (first stage coefficient on
the IV for the subsample multiplied by the difference
between the 95th percentile and 5th percentile of the
IV for the subsample).
Source: NSW BOCSAR ROD.
A6 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table A9. Fine and reoffending in 0-2 years: drug type
(1) (2) (3) (4) (5) (6)
Dependent variables Dependent variables
Reoffence Use/possession Use/possession Reoffence Use/possession Use/possession
of any type reoffence reoffence of any type reoffence reoffence
(any drug) (same drug) (any drug) (same drug)
OLS (no controls) 2SLS (no controls)
Amphetamines 0.271*** 0.119*** 0.0478*** 0.0108 -0.00437 -0.103
(0.0122) (0.0112) (0.00839) (0.142) (0.143) (0.0903)
Cannabis 0.245*** 0.135*** 0.129*** -0.113 -0.0315 -0.0176
(0.0104) (0.00820) (0.00677) (0.172) (0.204) (0.130)
Cocaine 0.0505*** -0.00998 -0.0347*** 0.0812 0.0813 0.0868
(0.0143) (0.0139) (0.00939) (0.161) (0.154) (0.126)
Ecstasy 0.00739 0.00683 -0.0198*** 0.111 0.0604 0.0455
(0.00933) (0.00698) (0.00364) (0.110) (0.0582) (0.0431)
Hallucinogens 0.00552 -0.00318 -0.0159 -0.620 -2.906 0.786
(0.0285) (0.0203) (0.00989) (1.534) (21.24) (6.560)
Opiates 0.380*** 0.113*** -0.0116 -0.0805 0.0346 -0.176
(0.0328) (0.0317) (0.0230) (0.344) (0.185) (0.144)
Other drug 0.181*** 0.0461 -0.0548*** -0.189 -0.116 -0.148
(0.0332) (0.0281) (0.0127) (0.297) (0.308) (0.133)
Unknown 0.290*** 0.104* -0.0685** 0.0568 0.0629 0.0158
(0.0342) (0.0475) (0.0246) (0.226) (0.274) (0.188)
OLS (with controls) 2SLS (with controls)
Amphetamines 0.155*** 0.0817*** 0.0421*** -0.0325 0.000670 -0.0951
(0.0129) (0.0119) (0.00970) (0.119) (0.105) (0.0762)
Cannabis 0.0896*** 0.0603*** 0.0576*** -0.112 -0.0181 -0.0315
(0.00798) (0.00967) (0.00823) (0.0915) (0.0926) (0.0801)
Cocaine 0.0509*** 0.0187 -0.0110 0.00883 0.0417 0.0447
(0.0144) (0.0141) (0.00982) (0.0918) (0.0975) (0.0671)
Ecstasy 0.0546*** 0.0333*** 0.00207 0.0119 0.0126 0.00864
(0.00848) (0.00707) (0.00392) (0.0419) (0.0291) (0.0183)
Hallucinogens 0.0569 0.00246 0.0140 -0.625 -2.492 0.597
(0.0388) (0.0368) (0.0171) (2.392) (8.402) (2.142)
Opiates 0.137** 0.0741 0.00610 0.00176 0.0205 -0.227
(0.0442) (0.0491) (0.0280) (0.289) (0.241) (0.192)
Other drug 0.111** 0.0307 -0.0123 0.00303 0.0299 -0.0806
(0.0339) (0.0384) (0.0148) (0.268) (0.286) (0.121)
Unknown 0.119** 0.124* 0.0165 -0.0857 -0.00589 0.0743
(0.0421) (0.0608) (0.0356) (0.304) (0.363) (0.177)
Observations 33,060 23,219 23,185 33,060 23,219 23,185
Notes: The dependent variable is in the header. The independent variable is a dummy for fine interacted by the array of dummies
for different drugs. 2SLS instrument for fine using a magistrate leniency measure estimated using data from other cases assigned to a
magistrate. All regressions include interacted court-year fixed effects. Robust standard errors clustered at the individualand magistrate
level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
S ALEXEEV AND D WEATHERBURN A7
Table A10. Fine and reoffending of any type in 0-2 years: by reoffence type
(1) (2) (3) (4)
Dependent variable: reoffence of any type
OLS OLS 2SLS 2SLS
(no controls) (with controls) (no controls) (with controls)
Acts intended to cause injury -0.582*** -0.622*** -0.873 -0.539
(0.0528) (0.0605) (4.419) (1.047)
Sexual assault and related 0.0719*** 0.0315* -0.0454 -0.0865
(0.0135) (0.0132) (0.243) (0.222)
Dangerous and negligent acts 0.0680 0.0642 -0.329 0.133
(0.0911) (0.0885) (1.014) (0.690)
Abduction/Harassment 0.0541 0.0281 -0.328 -0.323
(0.0363) (0.0359) (0.240) (0.237)
Robbery and related 0.0878 0.0545 -0.756 -1.232
(0.0564) (0.0518) (1.560) (1.882)
Unlawful entry and related 0.165** 0.102 -0.327 -0.702
(0.0509) (0.0552) (0.927) (1.327)
Theft and related 0.122*** 0.0587 -0.525 -0.596
(0.0307) (0.0301) (2.232) (0.984)
Fraud and related 0.139*** 0.0877*** 0.0581 0.0307
(0.0132) (0.0134) (0.700) (0.220)
Illicit drug offences 0.113*** 0.0705** -0.131 -0.297
(0.0231) (0.0235) (0.469) (0.358)
Weapons 0.0710*** 0.0534*** 0.0168 0.0243
(0.0104) (0.00990) (0.129) (0.0913)
Property damage and related 0.103*** 0.0683** 0.0830 0.154
(0.0244) (0.0246) (0.654) (0.259)
Public order 0.0798** 0.0431 -0.360 -0.880
(0.0241) (0.0233) (1.223) (0.958)
Traffic and vehicle regulatory offences 0.115*** 0.0768*** -3.417 -1.692
(0.0213) (0.0207) (24.04) (5.868)
Justice procedure offences 0.0328*** 0.00967 0.0429 0.00323
(0.00956) (0.0101) (0.236) (0.173)
Miscellaneous 0.118*** 0.0816*** -0.181 -0.239
(0.0159) (0.0161) (0.254) (0.252)
Homicide and related 0.0830* 0.0587 0.0862 0.0655
(0.0395) (0.0399) (0.738) (0.265)
Observations 33,060 33,060 33,060 33,060
Notes: The dependent variable is in the header. The independent variable is a dummy for fine interacted by the array of
dummies for different crime. 2SLS instrument for fine using a magistrate leniency measure estimated using data from
other cases assigned to a magistrate. All regressions include interacted court-year fixed effects. Robust standard errors
clustered at the individual and magistrate level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
A8 FINES FOR ILLICIT DRUG USE DO NOT PREVENT FUTURE CRIME
Table A11. Fine and reoffending in 0-2 years: quadratic profile
(1) (2) (3) (4) (5) (6)
Dependent variables Dependent variables
Reoffence Use/possession Use/possession Reoffence Use/possession Use/possession
of any type reoffence reoffence of any type reoffence reoffence
(any drug) (same drug) (any drug) (same drug)
OLS (no controls) 2SLS (no controls)
Fine 0.0518** 0.0101 0.0158 -0.311 -0.233 -0.190
(0.0157) (0.0129) (0.00962) (0.231) (0.279) (0.229)
Fine2-0.00198 0.00110 -0.000208 0.0376 0.0285 0.0230
(0.00188) (0.00150) (0.00112) (0.0272) (0.0329) (0.0268)
OLS (with controls) 2SLS (with controls)
Fine 0.0317* -0.00138 -0.000140 -0.109 -0.170 -0.141
(0.0135) (0.0124) (0.00958) (0.210) (0.265) (0.219)
Fine2-0.00219 0.00142 0.000801 0.0129 0.0204 0.0166
(0.00163) (0.00147) (0.00112) (0.0248) (0.0313) (0.0256)
Observations 33,060 23,219 23,185 33,060 23,219 23,185
Notes: The dependent variable is in the header. The independent variable is a dummy for fine interacted with log-transformed and
inflation-adjusted fine amount and dummy for fine interacted with square of log-transformed and inflation-adjusted fine amount.
2SLS instruments for fine interacted with log-transformed and inflation-adjusted fine amount using a magistrate leniency measure
estimated using data from other cases assigned to a magistrate. All regressions include interacted court-year fixed effects. Robust
standard errors clustered at the individual and magistrate level are reported in parentheses.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
S ALEXEEV AND D WEATHERBURN A9
Table A12. Fine and reoffending in 0-2 years: alternative clustering, drug type grouping, and additional
controls
(1) (2) (3)
Dependent variables
Reoffence Use/possession Use/possession
of any type reoffence reoffence
(any drug) (same drug)
A: Generating IV for each drug separately
2SLS 0.0163 0.0831 0.0379
(no controls) (0.0473) (0.0777) (0.0601)
2SLS 0.051 0.0119 0.01133
(with controls) (0.0308) (0.0791) (0.0651)
B: Generating IV for cannabis and other drugs separately
2SLS 0.0256 0.0739 0.0115
(no controls) (0.0313) (0.0638) (0.0485)
2SLS 0.0145 0.0134 0.01133
(with controls) (0.0801) (0.0715) (0.0411)
C: Clustering on defendant only
2SLS 0.0300 0.0364 -0.00286
(no controls) (0.0671) (0.0517) (0.0813)
2SLS -0.0268 0.0108 -0.0178
(with controls) (0.0608) (0.0598) (0.0411)
D: Clustering on drug type instead of magistrate
2SLS 0.00300 0.00364 -0.00286
(no controls) (0.0888) (0.0711) (0.0941)
2SLS -0.0268 0.0108 -0.0178
(with controls) (0.0501) (0.0468) (0.0311)
E: Controlling and instrumenting for bond
2SLS 0.0211 -0.0214 0.0185
(no controls) (0.0321) (0.0249) (0.0136)
2SLS 0.0111 0.0218 0.0169
(with controls) (0.0311) (0.0191) (0.0154)
F: Interacting with fine amount
2SLS 0.00481 0.00754 0.00379
(no controls) (0.00773) (0.00777) (0.00600)
2SLS -0.000611 0.00289 -0.000933
(with controls) (0.00712) (0.00789) (0.00610)
Observations 33,060 23,219 23,185
Notes: The dependent variable is in the header of the table. Each panel
reports the estimates from different modifications of our 2SLS framework.
*p < 0.05, ** p < 0.01, *** p < 0.001.
Source: NSW BOCSAR ROD.
... In most cases, law enforcement groups argue for harsher fines and penalties to deter crimes. A study in Australia (New South Wales) suggested that fines and harsh penalties increase recidivism (Alexeev & Weatherburn, 2022). Another study in USA (Texas), revealed that offenders who receive fines were more likely to commit additional crimes than those who were not punished. ...
Article
Full-text available
Illegal artisanal gold mining threatens biodiversity conservation within protected areas. The objectives of this study were to: (i) assess the environmental impacts of illegal artisanal gold mining in Chewore Safari Area (CSA) and (ii) assess trends in illegal artisanal gold mining within CSA, northern Zimbabwe. A mixed methods approach was used to collect data (i) from law enforcement records kept at CSA focusing on the period 2014 to 2021, and (ii) from field visits to areas affected by illegal artisanal gold mining. The study found that there was a gold rush into CSA with many people from the adjacent communitie being arrested and prosecuted in the year 2014. A gradual decline trend in illegal gold mining activities was observed. Equipment or tools were confiscated from illegal miners including gold metal detectors. Site observations indicated destructive gold mining practices associated with visible land degradation. The study recommended increased law enforcement and awareness programmes within communities adjacent to CSA and reclamation of mined sites.
... Prior to reporting the tests to support the IV validity, we stress that the ROD dataset has been used multiple times before; and it is well-established that the judge leniency design is feasible in the NSW setting (Williams andWeatherburn, 2022, Alexeev and. We begin by establishing IV randomness, which confirms that the 'equality before the law' principle -the central element of our design -is present in the data. ...
Preprint
Full-text available
A number of recent studies have found evidence that non-prosecution of minor offences reduces the likelihood of further offending. One of the mechanisms which might explain this result is that offenders are prosecuted and who end up with a criminal conviction suffer damage to their future employment and earnings prospects. In this study we test this explanation using a facility under which magistrates in New South Wales, Australia can find an offence proved but dismiss the charge and record no conviction. We exploit the plausibly exogenous variation between magistrates in their willingness to make use of this facility and find no difference in risk of reoffending among DUI offenders whose charges are dismissed and those who are prosecuted and convicted. Our findings suggest that some mechanism other than the avoidance of a criminal conviction may be responsible for the finding that non-prosecution of minor offences lowers the risk of reoffending.
... For the actual IV construction, we perform the residualization on the full sample (separately for males or females) without restricting the population to employed workers only. A bigger 4. One could also note that if, instead of interview, we only had area fixed effects, our steps are similar to the ones taken in the examiner design with an interviewer treated is an examiner (e.g., Alexeev and Weatherburn 2022). If we do follows these steps for our setting, the constructed IV fails the relevance assumption because there is no economic reason for the interviewer's identity to influence the treatment. ...
Preprint
Full-text available
We leverage interviewer assignment in an interviewer-mediated survey as a randomization device to measure the wage penalty of cannabis use. Using the Australian household survey, we show that the female wage penalty for cannabis use is 8.3%. For males, we estimate a noisy zero. This confirms findings from the medical literature that women are biologically more sensitive to cannabis and its health‐related harm. As the legalization of cannabis is typically found to increase its usage, it may also exacerbate the gender wage gap.
... Instead of using the dummies for interviewers directly as in previous medical papers (McGovern et al. 2015), I construct a residualized leave-out IV that is highly predictive of drinking pattern but independent from a parent whose outcome is instrumented. This can be compared to the contemporary practice of using judge sentencing harshness in papers that utilize randomly assigned judges to measure the specific deterrence effect in criminology (Alexeev et al. 2021;Frandsen et al. 2019). ...
Preprint
Full-text available
The work constructs a data-rich, nationally representative dataset with 40516 dyads to quantify the intergenerational transmission of alcohol consumption. Drinking is transmitted only from the parent who shares the offspring's gender. For females, the estimated elasticity is higher than for males. For males, the influence is fully established by the late 20s, while for females continues to the late 30s. The established influence persists throughout life. The discovered drivers are the acquisition of gender-specific drinking norms during the teenage years and the adoption of parental coping mechanisms during parenthood. The genes matter for male transmission but not for females. The endogeneity is found to be empirically negligible.
Article
We estimate the impact of financial sanctions in the US criminal justice system, leveraging nine natural experiments in a regression discontinuity design framework across a diverse range of enforcement levels (17–6,000) and institutional environments. We leverage survey and administrative data to consider a variety of short- and long-term outcomes, including employment, recidivism, household expenditures, and other self-reported measures of well-being. We find robust evidence of precise null effects, including ruling out long-run impacts larger than −391–142 in annual earnings and −0.001–0.01 in annual convictions, with no corresponding payment increases despite salient and heterogeneous enforcement mechanisms. (JEL G51, H76, J31, K14, K41, K42)
Preprint
Estimates in judge designs run the risk of being biased due to the many judge identities that are implicitly or explicitly used as instrumental variables. The usual method to analyse judge designs, via a leave-out mean instrument, eliminates this many instrument bias only in case the data are clustered in at most one dimension. What is left out in the mean defines this clustering dimension. How most judge designs cluster their standard errors, however, implies that there are additional clustering dimensions, which makes that a many instrument bias remains. We propose two estimators that are many instrument bias free, also in multidimensional clustered judge designs. The first generalises the one dimensional cluster jackknife instrumental variable estimator, by removing from this estimator the additional bias terms due to the extra dependence in the data. The second models all but one clustering dimensions by fixed effects and we show how these numerous fixed effects can be removed without introducing extra bias. A Monte-Carlo experiment and the revisitation of two judge designs show the empirical relevance of properly accounting for multidimensional clustering in estimation.
Preprint
Full-text available
A medically supervised injecting room (MSIR) — a facility for the safe injection of illicit drugs — is an example of a harm reduction approach against opioid epidemics. An economic theory of moral hazard suggests that such a facility can intensify drug use by reducing the risk of lethal overdose, mediating an increase in crime, death, and nuisance behavior, thus, depressing local urban development. Beyond that, negative media coverage can generate similar adverse outcomes even without any factual drug intensification. Accordingly, using non-parametric methods and spatial double and triple difference design on the data from Victoria, Australia, we show that the MSIR's opening causes around a 5%–7% reduction in housing value. We conclude that the harm reduction approach can have nontrivial indirect costs outside health domains.
Article
Full-text available
Introduction The present study explores Australian public support for more lenient treatment of persons found in possession of small amounts of illegal drugs for personal use. Methods Data for the study are drawn from the 2013, 2016 and 2019 National Drug Strategy Household Surveys. Bivariate associations were tested using χ ² tests corrected for survey design. Multivariate analyses of public support for legalisation and imprisonment involved logistic regression. Multivariate analyses of public preferences for various responses to those found in possession of illicit drugs involved the use of multinomial regression. Results Support for legalising personal use of three drugs (cannabis, ecstasy and cocaine) has risen significantly since 2013 but remained stable (and low) for two (heroin and methamphetamine). Support for no action, a caution or a warning has grown for those found in possession of cannabis, ecstasy and heroin for personal use but not for those found in possession of methamphetamine. There is, however, more support for responding to this group with treatment, education or a small fine than there is for imprisonment. Discussion and Conclusion There is strong support for legalising use of cannabis. There is little support for legalising use of ecstasy and cocaine but growing public support for a less punitive approach to those who use these drugs. There is little public support for a change in the current approach to heroin and methamphetamine.
Article
We propose a nonparametric test for the exclusion and monotonicity assumptions invoked in instrumental variable (IV) designs based on the random assignment of cases to judges. We show its asymptotic validity and demonstrate its finite-sample performance in simulations. We apply our test in an empirical setting from the literature examining the effects of pretrial detention on defendant outcomes in New York. When the assumptions are not satisfied, we propose weaker versions of the usual exclusion and monotonicity restrictions under which the IV estimator still converges to a proper weighted average of treatment effects. (JEL H76, K41)
Article
Recent literature raises concerns about monotonicity conditions required to interpret IV estimates under heterogeneous effects. A prominent example involves random decision-maker IV designs where decision-makers exhibit systematic differences in both preferences and skills. We develop tests of monotonicity in the context of judicial decision-making using proxies of judicial errors based on appeals and reversals of trial decisions from Norwegian court records. Our tests fail to reject average monotonicity. This suggests that differences in stringencies across judges are not sufficiently driven by skills to raise concerns about the validity of the random judge IV literature.
Article
Every year, millions of Americans experience the incarceration of a family member. Using 30 years of administrative data from Ohio and exploiting differing incarceration propensities of randomly assigned judges, this paper provides the first quasi-experimental estimates of the effects of parental and sibling incarceration in the United States. Parental incarceration has beneficial effects on some important outcomes for children, reducing their likelihood of incarceration by 4.9 percentage points and improving their adult neighborhood quality. While estimates on academic performance and teen parenthood are imprecise, we reject large positive or negative effects. Sibling incarceration leads to similar reductions in criminal activity. (JEL H76, J13, K42)
Article
Aim To examine the specific deterrent effect of offender supervision on the risk of a further offence by persons convicted of amphetamine use and/or possession. Design Observational study of two groups matched on a wide range of factors using augmented inverse probability of treatment weighting. Setting New South Wales (NSW), Australia. Participants A total of 2,099 persons convicted of amphetamine use and/or possession and placed on either a supervised good behaviour bond (1,004) or a bond without supervision (1,096). Measurements Conviction for another amphetamine use/possession offence, any drug offence, assault, or theft within 48 months free time after the index court appearance. Findings Only two of the treatment estimates were consistent with a deterrent effect. Where the outcome was an assault offence, the estimated reduction in risk of re‐offending was 1 per cent. Where the outcome was a theft offence, the estimated reduction in the risk of reoffending was 0.07 per cent. Neither result was statistically significant. The coefficients measuring the effect of supervision on the remaining two outcomes (use/possess amphetamine and use/possess any drug) were both positive and not statistically significant. Conclusion Correctional supervision does not appear to have been effective in New South Wales, Australia, in reducing the risk of reoffending amongst people convicted of amphetamine use/possession.
Article
We evaluate electronic monitoring as an alternative to prison for non-violent offenses. Leveraging plausibly exogenous variation in sentencing outcomes generated by quasirandom assignment of judges, we find electronic monitoring reduces reoffending at both extensive and intensive margins. Compared with prison, electronic monitoring is estimated to reduce the probability of reoffending by 22 percentage points 5 years after sentencing and by 11 percentage points 10 years after sentencing, with the cumulative number of offenses reduced by 45 percent 10 years after sentencing. These results demonstrate that electronic monitoring can have sustained crime-reducing effects.