I’d rather be Hanged for a Sheep than a Lamb
The Unintended Consequences of ‘Three-Strikes’ Laws

Radha Iyengar
Harvard University
October 2007

The opinions and conclusions are solely those of the author. I grateful to Orley Ashenfelter,
David Autor, Hank Farber, Lawrence Katz, Lisa Kahn, Jeffery Kling, Steve Levitt, Alex Mas,
Cecilia Rouse, Philipp Schnabl and participants at the Industrial Relation’s labor lunch for
numerous insightful suggestions. I would also like to thank Tony Crittenden and the California
Bureau of Criminal Information for generous assistance with data sources. I am grateful to the
Princeton Industrial Relation Section, the Woodrow Wilson Society of Fellows, and the Robert
Wood Johnson Foundation for financial support. Any remaining errors are entirely my own.

ABSTRACT
Strong sentences are common “tough on crime” tool used to reduce the incentives for individuals
to participate in criminal activity. However, the design of such policies often ignores other
margins along which individuals interested in participating in crime may adjust. I use
California’s Three Strikes law to identify several effects of a large increase in the penalty for a
broad set of crimes. Using criminal records data, I estimate that Three Strikes reduced
participation in criminal activity by 20 percent for second-strike eligible offenders and a 28
percent decline for third-strike eligible offenders. However, I find two unintended consequences
of the law. First, because Three Strikes flattened the penalty gradient with respect to severity,
criminals were more likely to commit more violent crimes. Among third-strike eligible
offenders, the probability of committing violent crimes increased by 9 percentage points.
Second, because California’s law was more harsh than the laws of other nearby states, Three
Strikes had a “beggar-thy-neighbor” effect increasing the migration of criminals with second and
third-strike eligibility to commit crimes in neighboring states. The high cost of incarceration
combined with the high cost of violent crime relative to non-violent crime implies that Three
Strikes may not be a cost-effective means of reducing crime.

Radha Iyengar
Robert Wood Johnson Health Policy Scholar
Harvard University and NBER
1730 Cambridge St
Cambridge, MA 02138
riyengar@rwj.harvard.edu

2

1. INTRODUCTION

The high crime rates of the 1980s coupled with the belief that prison served as a “revolving
door” for criminal activity, prompted new sentencing laws aimed at increasing sentences for repeat
offenders. One of the most publicized new policies was habitual offender law, commonly called
“Three-Strikes You’re Out”. This paper uses California’s version of this law to estimate how
criminals respond to changes in sentencing policy. California’s Three Strikes law changed the
penalty structure in two ways: it increased the expected penalty for all crimes (intercept shift), and
flattened the penalty gradient with respect to severity of crime (slope shift). I develop a model in
which the increase in the intercept has the expected effect of decreasing crime levels, while the
gradient shift has the unanticipated consequence of encouraging a shift toward more serious crime. I
empirically examine the relative magnitude of these two effects using California’s Three
Strikes law. The results suggest that Three Strikes reduced the overall level of crime but increased
the propensity to commit violent crime. Depending on the societal preferences regarding the cost of
non-violent and violent crimes, such offsetting distributional effects may substantially reduce the
benefits of broad enhancements in sentencing. In addition I find that some of California’s reduction
comes at the expense of other states. This further suggests that single-state enhancements may be
more costly from a national perspective than previously believed.
In part because of the high publicity surrounding the law and in part because it remains
among the most striking examples of across-the-board sentence enhancements, there has been an
extensive literature aimed at estimating the overall effect of Three Strikes law. Early work by
Greenwood et al (1994) estimated huge costs and limited deterrence from the law change based on
projections of current offenders among Three Strikes states. Macallair and Males (1999) compare
counties with strict versus lax Three Strikes enforcement. They find counties that strictly enforced
the sentence enhancements saw negligible effects on crime rates. Marvell and Moody’s (2000)
cross-state analysis also found Three Strikes has little effect on overall crime rates but found a
significant increase in the number of murders. Comparing counties and age groups, Jaimeson
(1999) finds little effect of Three Strikes on criminal participation. Shepherd (2001) compared the
rates of triggering and non-triggering offenses before and after Three-Strikes and found significant
declines in triggering offenses supporting a deterrence effect from expected increased punishment.
Most recently, evidence from Helland and Tabarrok (2007) shows a significant deterrence effect of
3

Three Strikes law on second strike offenders concentrated among violent offenders.
Previous attempts to estimate the effect of Three-Strikes have been limited by the ability to
establish a valid control group. This paper uses the unique structure of Three-Strikes law in which
offenders with the same criminal history but different ordering of crime commission face different
sentencing eligibility to identify the effect of Three-Strikes sentencing eligibility on criminal
activity. In particular, the law required that an individual commit a “record aggravating” or
“triggering” offense in order to activate eligibility for Three-Strikes law sentencing. This meant that
individuals who committed a “triggering” offense followed by a felony faced different potential
sentences than those who committed a felony and then a “triggering” offense. Using individuals
who committed the same crimes but in different orders, I estimate a baseline difference in the
likelihood of re-offending and of committing a violent crime conditional on re-offending prior to the
law change. I estimate the post-Three-Strikes difference in their likelihood of re-offending and of
committing a violent crime conditional on re-offending. Differencing out the baseline likelihood, I
estimate a 9 percentage point decrease in the propensity to re-offend. In part, this appears due to a
lengthened duration of non-participation in criminal activities by repeat offenders.
While Three Strikes had the intended effect of reducing participation in crime, there appear
to be two sizeable unintended consequences of this law. First, there is an 8 percentage point
increase in the propensity to commit violent crime conditional on committing a new crime. This
effect, while smaller than the effect on the participation margin is a non-negligible and socially
costly consequence of broad sentencing policies which apply equal penalties to a crime of varying
severity. Second, some of the reduction in criminal participation in California appears due to the
migratory response of repeat offenders who opt to move to lower-sanctioning states. Thus there
appears to be a “beggar thy neighbor” spillover effect from state level sentencing laws.
This paper adds to the literature attempting to estimate the effectiveness of harsh sentencing
regimes on crime levels. Consistent with more recent literature, I find an overall effect of a decline
in the criminal participation rate among second and third strike eligible offenders and a reduced
propensity to commit record-aggravating offenses among first-strike eligible offenders. In addition,
this paper attempt to systematically disentangle the competing effects of broad sentence
enhancements on both the decision to participate in criminal activity and the selection of type of
crime committed conditional on participation criminal activity as well as the mechanisms by which
reduction in criminal activity is accomplished (e.g. deterrence versus migration). Separating and
4

identifying the margins along which criminals adapt to sentence enhancements can reveal not only
the overall effect of long sentences but also the general responsiveness of criminals to cost-based
incentives and the relative magnitudes of their responsiveness across different margins of
adjustment.
This rest of this paper proceeds as follows: Section 2 presents a framework for considering
sentencing regimes and the specific case of the effects of Three Strikes law on criminal activity.
Section 3 presents the data and strategy used to identify the causal effect of change in the penalty
structure on criminal activity. Section 4 presents the results from an empirical analysis of the effect
of Three Strikes law on the propensity to commit crimes as well as its effects on migration and
crime selection. Section 5 uses the empirical results to estimate the social benefit of changes in
sentencing structure and then offers some concluding remarks.

2. Theoretical Framework for Comparing Sentencing Regimes

Broad sentence enhancements have been a common tool for increasing the potential costs of
criminal activity. Most policies focus on maximizing the effect of these laws on participation in
criminal activity but there are several margins of along which criminals may adjust their behavior in
response to the enhanced sentencing. For example, as illustrated in Stigler (1978) while the most
obvious margin of adjustment is participation, another means of adjustment is the severity of crime
(which assuming returns to crime are increasing in severity may raise the profit of crime despite the
increased cost of enhanced sentencing). In this section I develop a simple framework for
considering the potential effect of different types of sentence enhancements and then consider the
specific case of California’s Three Strikes law.

2.1 Basic Framework
To begin understanding the effect of changes in sentencing policies on a criminal’s
decision, consider a simple version of the rational criminal’s decision-making process (based on
Becker, 1968). An individual will choose to commit a crime only if the utility from this crime, as
defined by the difference between the revenue and the expected cost of committing this crime
(Ucrime), is greater than some reservation utility ( U ):
U crime

≥ U

(1)
5

In such a model, the high cost of crime, typically generated by expected cost of imprisonment, will
cause many individuals not to commit crime at all. To illustrate this relationship more formally,
define the utility from crime:
U T = p ⋅ R T − (1 − p ) ⋅ g ( P T ) − θ ⋅ T

(2)

An individual’s utility from crime expressed above is separated into three terms. The first term is
the returns from crime of type T, RT, which occurs with probability p. The second term is the utility
of a failed attempted at crime type T, for which the criminal must pay PT and receives utility
corresponding to g(PT). This second term occurs with probability 1 – p. The third term is the fixed
cost to the criminal of committing crime T, θ ·T, and occurs regardless of success. In this
framework, θT represents an individual cost specific term that incorporates psychic costs of crime as
well as other individual-specific factors which may generate utility or disutility from crime (i.e. the
additional private return from crime). The distribution θ is described by F(θ) which admits a density
f(θ). Normalize the utility of the outside option to zero. For illustrative purposes, let utility gain

from crime be linear in the type of crime and let the utility from punishment be represented by the
function g which is strictly monotonic in PT and twice differentiable. 1
For simplicity, suppose there are only 2 types of crime. Violent crime has a payoff RH = H
and a penalty of PH and non-violent crime has a payoff of RL = L and a penalty of PL where H > L
and PH ≥ PL. Using these simplifications, it is possible to divide the decisions of individuals in the
distribution into three categories: non-criminal activity, non-violent crime, and violent crime. The
criminal participation margin is defined as the value of θ which sets the utility from non-criminal
activity (T = 0) and non-violent criminal activity (T=L) equal.2 Defining θParticipate as the value that
makes the participation constraint hold at equality, the probability that an individual will participate
in crime is:
Pr(θ ≤ θ Participate ) = F

(

1
L

( pL − (1 − p ) g ( P L ))

)

(3)

The cutoff value from equation 3 is illustrated in figure 1, where F is assumed to be a normal
distribution with mean 0 and variance 1. As illustrated, the distribution of θ generates a range of
1

Note that if g(.) were simply linear then the results in this section would obtain. However, because we wish to allow
for an ambiguous effect of an intercept shift in the level of punishment (rather than the ratio of the punishments), the
added complexity of a more general g(.) function is included.
2
Note that the subsequent analysis assumes, θParticipate< θSeverity. This will be the case if g(PH) is sufficiently small
H − L g(P H )
relative to g(PL). Specifically, this condition will hold if
≥
.
L
g(P L )
6

individuals who participate in crime some of whom would shift their behavior into non-criminal
activity with a small increase in costs and some of whom would not deterred even after a large
increases in the cost of crime. Since g (.) is monotonic in PL, an increase in the penalty for nonviolent crime will shift θParticipate to the left. Thus, some individuals are deterred from engaging in
criminal activity.
The crime severity margin is the boundary value of θ at which criminals decide to
participate in non-violent or violent crime. Individuals who commit violent crime are intuitively the
set of individuals whose value of θ sets the payoffs from non-violent (T = L) as less than those from
violent (T = H). Defining θseverity as the value of θ which sets the payoffs from the two types of
crime equal, the probability that an individual will participate in violent crime is:

 1

Pr(θ ≤ θ Severity ) = F 
( p ( H − L) − (1 − p )( g ( P H ) − g ( P L )) 
H −L


(4)

Thus for individuals with a sufficiently low θ, participation in violent crime will be optimal.
However, equation 4 illustrates that that this decision may be affected by either the penalty for
violent crime (PH), the penalty for non-violent crime (PL), and the relationship between the two
penalties.3

2.2 Changing the Penalty-Severity Gradient
In the basic framework, the existence of the criminal participation margin and crime severity
margin is generated by an assumed difference in the penalty structure which penalizes more serious
crimes (e.g. violent crimes) more severely.4 Suppose the relationship between the penalty for
violent and non-violent crime is described as follows: PL = β·PH where 0 < β ≤ 1. In this setting,
1/β represents the penalty-severity gradient. A small β corresponds to a steep gradient meaning a
very large penalty for violent crime relative to non-violent crime. If β=1 then the crime-severity
gradient is flat meaning that there is no additional penalty for more severe crimes. Using this
3

This follows in the vein of Stigler (1970) and Becker (1968). The debate between the two articles concerns what
role that a penalty gradient with respect to crime severity might play. In his classical model, the efficient criminal
punishment system applies maximal (ideally infinite) punishment to all crimes with low probability of enforcement.
This system is efficient in the sense that it has the highest ratio of crimes deterred relative to cost. Stigler countered
that this effect was concentrated on the participation margin, which he labeled “average deterrence.” In Becker’s
model, the additional penalties for more severe crime, which Stigler labeled “marginal deterrence” introduced
inefficiency in the sense that it potentially lowers this ratio. Stigler’s response suggested that the increased marginal
cost of crimes was necessary to transfer the increased social cost of these crimes onto the individual imposing the
costs on society.
7

representation of the penalty structure, the cutoff values of θ from equations 3 and 4 can be
rewritten as functions of β and PH.
Pr(θ ≤ θ Participate ) = F

(

1
L

( pL − (1 − p ) g ( β P H ))

)

(5)

 1

Pr(θ ≤ θ Severity ) = F 
( p ( H − L) − (1 − p )( g ( P H ) − g ( β P H )) 
H −L


(6)

Taking the derivative of the F(.) in equation (5) with respect to β yields the expression:

−

1
∂g ( β P H )
f (.)(1 − p )
. Thus an increase in β or a flattening of the penalty gradient will decrease
L
∂β

the number of individuals willing to participate. Intuitively, this is because the crime participation
margin is created by the cost of non-violent crime. As illustrated in figure 1, an increase in this cost
will shift θparticipate to the left, decreasing the number of people willing to engage in criminal activity.
Taking the derivative of F(.) in equation (6) with respect to β yields the expression

1
∂g ( β P H )
f (.)(1 − p )
. This suggests that an increase in β or a flattening of the penalty
∂β
H −L
gradient will increase the number of people willing to participate in violent crime. This occurs
because the returns from successful criminal activity are fixed and thus an increase in the cost of
non-violent crime relative to non-violent crime changes the relative profitability of violent crime
relative to non-violent crime. This can be illustrated in figure 1 as a shift to the right of θseverity,
which encompasses a larger fraction of the total distribution as well as a larger fraction of
individuals committing crime.
To summarize, an increase in the penalty of non-violent crime relative to violent crime has
two effects: first, it reduces the number of individuals willing to participate in criminal activity.
Second, it increases the fraction of individuals participating in criminal activity who engage in
violent crime.5

2.3 Changing the Penalty Level
While a change in the penalty-severity gradient has clear predictions for the change in the
crime participation and severity margins, a simple scaled increase in penalties has a more ambiguous
4

This was generated in the above case by assuming PH > PL.
In the language of Becker and Stiglitz, for non-violent crime the average and marginal deterrence effect move in the
same direction and the model predicts an unambiguous decline in non-violent crime. For violent crime, the average

5

8

effect. Suppose rather than changing the penalty gradient, there is an increase in the absolute level
of penalties of crime such that the new penalty, PT is defined as PT =

1

α

PT for T = H, L. 1/α

represents the inflation factor of the new sentencing regime relative to the old and as such 0 < α <1.
In this case the penalty for violent relative to non-violent is the same before and after the change in
penalty structure, that is

PL
P
= L .
PH PH

Again substituting this into equations 3 and 4 yields the following expression:
Pr(θ ≤ θ Participate ) = F

(

1
L

( pL − (1 − p ) g (αP L ))

)

(5)

 1

Pr(θ ≤ θ Severity ) = F 
( p ( H − L) − (1 − p )( g (αP H ) − g (αP L )) 
H −L


(6)

Differentiating equation (5) with respect to α yields the following expression:
−

1
∂g (αP L )
f (.)(1 − p )
. By the same logic as the flattened penalty gradient, the increased penalty
L
∂α

for non-violent crime (i.e. a smaller α) corresponds to a reduction in the number of individuals
willing to engage in criminal activity.
The crime severity margin is more ambiguous. The derivative of equation (6) with respect to

α is:
∂F (θ Severity )
(1 − p )
= − f (θ Severity ) ⋅
[g ' (αPH ) PH − g ' (αPL ) PL ]
∂α
∆

(7)

And thus the sign of this depends on the term g ' (αPH ) PH − g ' (αPL ) PL . The inflated sentencing will
have the same effect as the flattened sentencing if g ' (αPH ) PH − g ' (αPL ) PL < 0 or if

g ' (αPH ) PL
<
.
g ' (αPL ) PH

Intuitively, this requires that the utility from an unsuccessful crime attempt (i.e. g(.)) be sufficiently
concave that the enhanced cost from violent crime has a low cost in utility terms relative to the
enhanced cost from non-violent crime. If g(.) is convex, we might expect that individuals will shift
from violent to non-violent crime because of the escalating disutility from more severe crime. That
is the concavity or convexity of the g(.) function serves the de facto role of flattening or steepening
the penalty gradient for sufficiently high penalties.
To summarize a level increase in the penalties of both violent and non-violent crime has two
and marginal effects move in different directions and the overall effect of the policy is ambiguous.

9

effects: first, it reduces the number of individuals willing to participate in criminal activity. Second,
it has an ambiguous effect on the fraction of individuals participating in criminal activity who
engage in violent crime.

2.4 California’s Three-Strikes Law as an Instrument for Changes in Sentencing Structure
In 1993, Washington and Wisconsin were the first states to adopt Three-Strikes sentencing
laws. By 1997, twenty-two other states and the Federal Government instituted similar statutes.
The common underlying theme among these statutes was severe punishment for recidivist
offenders. Although many states ignored their statute, two important components of California’s
law led it be strictly enforced. First, the broad coverage of the law offered highly enhanced
sentencing for all felonies allowing wide application. Second, lack of judicial discretion
prevented judges from circumventing the law in cases in which its application seemed
unreasonable.6 In California as of 2000, over 40,000 offenders have been sentenced under
Three-Strikes while no other state has even reached 1000 (Zimring, Hawkins, Kamin, 2001).7
Three-Strikes changed the entire sentencing structure for felonies in two distinct ways. In
order to activate Three Strikes sentencing, individuals needed to be convicted of a “record
aggravating” offense. As Table 1 shows, the aggravating offenses are very broad under California
law, ranging from murder and rape to burglary.8 The important aspect of the legal structure was that
California law invokes a second or third strike for any felony, so long as the individual was
previously convicted of an aggravating offense.9
Specifically the structure of the law introduced two distinct changes to the penalty structure.
On the third strike, California’s Three Strikes law required individuals to serve the maximum of
three times the sentence of the current felony or 25 years to life. Eligible individuals did not face
any additional punishment for violent offenses relative to non-violent ones. This corresponds to a
flat penalty-severity gradient (i.e. β =1) and the anticipated effect illustrated above is a decrease in
criminal participation but an increase violent crime conditional upon participation. On the second
6

In California, only prosecutors had discretion as to whether to charge individuals with qualifying offenses until
1997, when the California Supreme Court reinstated judicial discretion.
7
Several studies (National Institute of Justice, 1996; Dickey, 1996; Kessler and Levitt, 1998), as well as anecdotal
observations by the media indicate that Three-Strikes statutes have rarely been invoked anywhere else.
8
Definitions of offenses are presented in Appendix Table 1.
9
In fact, a prior prison sentence is not even required to trigger additional penalties, a unique feature of California law
(Clark, Austin, and Henry, 1997).
10

strike, California’s Three Strikes law criminals faced a doubling of the sentence for the second
felony. Thus, on the second strike, eligible individuals faced inflated sentences for all crimes. This
corresponds to the case where α =0.5 such that PT = 2 PT for both violent and non-violent crimes.
There is an ambiguous effect on the severity of crime conditional on participation that depends on
the nature of the disutility from unsuccessful criminal attempts.
To summarize the predicted overall effect of the law: (1) there is an unambiguous decline in
participation in criminal activity among second and third-strike eligible offenders, (2) there is a shift
to more severe crime among third-strike eligible individuals who participate in crime, and (3) there
is a potential change in the severity level of crime committed by second strike.

3. DATA AND IDENTIFICATION

The analysis in this paper uses a sample of offender records for individuals arrested from
1990-1999 sampled from three California cities: Los Angeles, San Francisco, and San Diego.
The sample thus includes individuals who have been arrested at least one time for a felony,
though many of these individuals will not have been convicted. This data is linked to the
Criminal Offenders Record Information (CORI) which provides information on previous and
future offenses. The retrospective information includes prior convictions, prior sentences served,
and the total number of prior arrests. The information on future criminal activity details all
felony convictions after the year of arrest until 1999. These arrest records also document the
final disposition of the crime for which the individual is under arrest which includes conviction
and sentence length. Finally, the arrest records include some information on personal
characteristics such as age, gender, and race. I partition individuals into three groups: first strike
eligible, second strike eligible and third strike eligible based on their criminal history and current
offense and disposition. Within each group, offenders have between zero and six prior felonies.
In addition, I append information on police spending, prosecution, and other criminal justice
spending from California Criminal Justice Profiles. I also use information on unemployment and
poverty information from the Current Population Survey.
It is worth noting that the data used in this paper, while imperfect, represents a substantial
improvement on previous data used to study the effect of sentencing enhancements on criminal
activity. Much of the previous work uses aggregate crime rates relying on regional and/or
11

temporal variation to identify the effects of sentence enhancements. An exception to this is recent
work by Helland and Tabborak (HT). HT identify the effect of Three Strikes using whether an
individual is convicted of two versus one strikeable offense as an exogenous source of variation.
There are several reasons why this identification may not be the best way to identify the effects of
the law change.
First, to the extent that there are systematic differences in offenders with two convictions
versus one, estimates which only difference post-law change will tend to be biased. While HT
tests this assumption in states without Three-Strikes law and find little difference, if the
willingness or ability of juries to convict individuals of strikeable offenses changes as a function
of the law, then HT’s tests will not be able to ascertain the validity of their identifying
assumption. Indeed, there appears to be a relationship between Three Strikes law on the rate of
negotiated sentences (plea bargains). The fraction of cases decided by jury trial increased almost
10 percent after the enactment of Three Strikes.10 While I cannot causally relate this to Three
Strikes, discussions with district attorneys, defense attorneys and judges suggests that Three
Strikes law has been one of the primary causes for this increase in the rate of cases going to trial.
Because they are likely to face a lengthy sentence regardless of a plea bargain, many defendants
decide not to negotiate a plea bargain in second and third strike cases. Thus many more offenders
choose to go to court in the hopes of avoiding a conviction altogether altering the probability of
conviction after the law change. Moreover, even if Three Strikes did not cause the increase in
trial rates, the concurrent change in sentencing law and trial rates makes it difficult to separately
identify the Three Strikes effect from other changes.
Second, because of the nature of discretion in the criminal justice system, the further
along the process data is collected the more affected by discretion is the data. Ideally, we would
observe all of the criminals who commit crime regardless of detection. Arrest is only one-step
removed from that as it requires only detection by police. Cases brought to trial are several steps
removed, requiring the decision to prosecute, determination of sufficient evidence for trial by a
grand jury, and decision to go to trial. Such discretion can be directly influence by the law
change. For instance, in some areas prosecutors sought Three Strikes enhancements only in
10

This is based on the estimates of the change in the probability of a jury trial conditional on being prosecuted for a
strikeable offense based on California Department of Justice statistics. From the data used in this paper it is not
possible to observe whether criminals were convicted due to a plea bargain or by trial. However, the probability of
conviction does change significantly after the law and this change varies by city.
12

certain cases, such as for certain types of crimes that are particular problems in their county or
where the current offense is serious or violent. While in other counties, prosecutors seek Three
Strikes enhancements in most eligible cases. Similarly, after 1997, judges varied in how often
they dismiss prior strikes, based on discretion afforded to them under the Romero decision.11
Third, there was a great deal of variation in the rate at which offenders who were arrested
faced penalties from Three-Strikes law. A legislative analysis by Brown and Jolivette (2005)
noted considerable variation among counties in the likelihood that an offender who is arrested
would be prosecuted and convicted under the Three Strikes law. For example, Kern County was
over 13 times more likely to send an arrestee to state prison with a strike enhancement than San
Francisco County. This variation makes it difficult to identify the effect of Three Strikes
penalties on offenders independent of prosecutorial conduct.
The approach used in this analysis is to compare similar individuals who faced different
strike eligibility before and after Three Strikes law was introduced. If we could observe the true
underlying propensity of individuals to commit a crime in the pre-Three-Strikes era, and then
their propensity to commit a crime in the post-Three-Strikes era, we could attribute the difference
in the propensity to commit crime to the effect of harsher sentencing (either through
incapacitation or deterrence). In practice it is not possible to observe an individuals true
probability of committing crime. However, we can observe among individuals who had
previously committed a crime, whether their probability of committing another crime changes
after the law change. Specifically, suppose that we believed the underlying distribution regarding
the probability of recidivism was fixed over time except with respect to Three Strikes sentencing.
Then if we observed in a change in the propensity to commit a crime among individuals who
had previously committed crimes—that is a change in the propensity to commit a crime—then
we can attribute that to the average deterrence effect from Three Strikes.
In order to match plausibly similar individuals, I use an individual’s prior criminal history
(PCH) as the source of identification.12 Under Three strikes, individuals with the same criminal
11

On June 20, 1996, the state Supreme Court ruled in People v. Superior Court (Romero)that the court has the
discretion to dismiss prior serious or violent felony convictions under the Three Strikes law. For a discussion of the
evolution of Three Strikes law see Brown and Jolivette (2005)
12
The prior criminal history (PCH) variable is a vector of indicator variables for the types of crimes committed prior
to the current offense, where prior crime categories are murder, rape, assault, robbery, burglary, theft, drugs, and
other miscellaneous felonies.. For example, an individual with two priors in burglary and theft would have non-zero
values for burglary and theft and zero values for all other crime types.
13

history, but different ordering of crimes have different sentencing eligibility. This mismatch
between strikes and felonies arises because while all felony convictions count as strikes after the
first strike, only certain felonies are covered as record aggravating or “triggering” offenses (to
give an individual a record-enhancing strike and evoke the harsher penalties). The list of record
aggravating offenses is presented in Table 2. Using this fact, I assume that individuals with the
same PCH variable have a fixed difference across time in all respects except sentencing
eligibility. Comparing individuals with similar histories but different Three Strikes eligibility
before and after Three-Strikes provides a means to measure the change in propensity to commit
crime as well as the change in propensity to commit a violent crime associated with the law
change.
To illustrate the identification strategy, consider the following example with two criminals
both of whom have previously committed a theft and a burglary. Criminal A first committed a theft
and then committed burglary. Criminal B first committed a burglary and then committed a theft.
Under sentencing guideline prior to Three-Strikes, both these individuals would face similar
sentencing eligibility if they committed a third offense. However, after the Three-Strikes law
change, the ordering of the crimes committed matters. Because burglary is a triggering offense, it
activates Three-Strikes sentencing. All felonies committed after the activation of Three-Strikes then
count as strikes. Thus, if individual A commits a new offense, that offense will count as a second
strike since he has committed no offenses after the burglary. In contrast, a new offense committed
by individual B will count as a third strike because he committed a theft after committing a burglary.
Thus in the post-period, individuals A and B are exposed to different penalties based on the
ordering of their previously committed crimes.
Because there may be differences in the probability of committing a crime and the type of
crime committed by an individual who first commits a less serious crime and then more serious
crime relative to an individual who commits a more serious and then a less serious crime, it is
important to control for the baseline difference in propensity to commit crimes. Thus I compare a
pair of individuals A and B, before and after the law change. I assume that a pair of individuals with
the same criminal history but different orderings of those crimes have a fixed difference in their
probability of committing a new offense. I will attribute the change in the difference between these
two individual’s propensity to commit a crime to Three Strikes sentencing eligibility.
There are two important exclusions in these data that may result in a mis-measured PCH
14

measure. First, juvenile records were not included despite the fact that under Three-Strikes
juvenile offenses may count as a strike if they meet the statutory criteria. Second, out-of-state
felonies count as a strike but are not documented in California arrest records. Thus, while I
might observe individuals who exit the California criminal market, I cannot observe whether
individuals committing offenses in California are first time offenders or migrants from other
states. Barring these exclusions, this data provides a comprehensive set of information regarding
individuals allowing relatively detailed comparison of offenders.
In order to construct the PCH variable, I classify previous convictions into one of seven
categories: murder, rape, assault, robbery, burglary, theft (which includes larceny and motor
vehicle theft), drug crimes, and other crimes. The definitions of these categories are presented in
Table 3. I then construct the PCH variable. PCH is a vector-valued variable which counts the
number of prior convictions in any of the seven offense categories. Returning to the example
above, both criminal’s A and B would have the same prior criminal histories
PCH = [0 0 0 0 1 1 0 0].
In general, it would be troublesome to use prior criminal history as a control variable for
an individual’s innate propensity to commit crime, as the prior history itself may be affected by
the law change. That is, individuals may be deciding whether to commit crimes now based, in
part, on their effect on sentencing for future crimes. In order to avoid including this, I restrict the
sample to individuals who committed their prior offenses before the law change (in 1994). Thus
the retroactive nature of Three Strikes makes the variation in PCH independent of enhanced
sentence eligibility in both the pre- and post-Three Strikes periods.
In addition, because of the censoring that occurs for individuals who commit crimes prior
to 1990, I restrict the analysis to offenders who committed at least one prior criminal activity
between 1990 and 1994. This eliminates the problem of observing individuals who commit
crime pre-1990 and then never commit crime again.
The above restrictions may generate the concern that individuals who commit crime in
the pre-period may be less crime prone than those in the post-period because the sample is in part
selected on the timeframe of an individuals criminal history. To address this all specifications
include the felony rate per criminal year (FRCY). The FRCY provides a measure of the
combination of effect from youth and being a “crime-prone” individual. Specifically, it is
defined as:
15

FRCY =

Number of Felonies Commited
Age of Offender − Time in Pr ison − 18

(6)

I also include offender age as a control variable, which allows both an age effect as well as a rate
effect, conditional on age. 13
The final sample restriction I impose is that I remove all offenders who are serving prison
sentences for the entire analysis period since by construction they cannot recidivate. Because the
time frame for recidivism used in the subsequent analysis is relatively short, this should not
systematically bias the propensity for recidivism before and after the law change.
Summary statistics for the sample used in the analysis are reported in Table 2. When
compared to the statewide criminal population (not reported), the sample differs from the general
population on key demographics. The sample includes a high fraction of minorities especially
blacks than in the population on average. The higher proportion of minorities is due to the
sampling of cities and the higher proportion of blacks is due to the concentration of blacks in Los
Angeles. Comparing outcomes between the cities also yields some notable differences. The
fraction of individuals charged with record aggravating offenses appears significantly higher in
Los Angeles than in San Francisco and marginally higher than in San Diego. This is consistent
with previous literature which suggests Los Angeles was more zealous in its enforcement of
Three Strikes. Similarly, conviction rates of second-strike and third-strike eligible defendants
were significantly higher in both Los Angeles and San Diego, than in San Francisco. This
difference declines after 1997 most likely due to the introduction of judicial discretion.

4. RESULTS
Before looking at the estimated effect of Three Strikes on criminal activity, I verify that ThreeStrikes resulted in sentencing differences by strike eligibility. Table 3 reports the sentencing statistics
before and after Three-Strikes for offenders convicted after arrest.14 It appears that Three-Strikes did in
13

This restriction may raise the concern that individuals in the pre-period were required to commit their current
offenses in more rapid succession than those in the post-period. If these are “worse” criminals in the sense that they
are more likely to recidivate than the reported estimates would tend to overstate the recidivism effects of Three
Strikes Law. Thus in addition to the FRCY inclusion, I test the sensitivity of the reported results to these sampling
restrictions in two ways. First, I include all individuals regardless of the year in which their prior was committed.
Second, I restrict the post sample to 1995-1996. This creates a symmetric timing requirement for pre- and post-law
change samples. Results are consistent across regression and are reported in Appendix Table 3.
14
Although in general, there are not significant differences between individuals with updated information versus
those without, there does appear to be a marginally significant difference between individuals with current offenses
16

fact double sentences on the second strike and dramatically increased sentences on the third strike, as is
required by law. It is worth noting that there is no effect on sentence length for first-strike eligible
offenders but these offenders do face different penalty profiles for future offenses (as noted in Shepherd
2001).

Moreover, while there is variation in the probability of conviction by cities, conditional on

conviction there is no significant difference in the sentences faced by criminals.

4.1 Estimating the Participation Effect
I define the participation effect of Three Strikes as a change in the probability of
committing a crime conditional on strike eligibility. To motivate this interpretation of the
participation effect, consider a latent variable model where we define a variable Y* such that
Y * = U Crime − U . Then, assume that Y*, the difference in utility from criminal and non-criminal
activity, is a function of strike eligibility, prior criminal history, and individual characteristics.
Therefore, we can write Y* as:

Yict* = β 0 + β1 (2 strikesict ) + β 2 (3strikesict ) + β 3 (after * 2 strikes )ict + β 4 (after * 3strikes )ict +

β5 ( PCH ict ) + β 6 (individual controlsi ) + γ t + δ c + ε ict

(7)

In equation (7), 2strikes is an indicator variable for second strike eligibility, 3strikes is an
indicator variable for third strike eligibility, PCH is a vector valued variable detailing an
individual’s prior criminal history, and individual controls include age race, sex, and felony rate
per criminal year. Although the latent variable, Y * is not observable, I can observe whether an
individual chooses to commit a crime (call this variable Y). The observed binary variable Y is 1
if Y* > 0 and 0 otherwise. I can then estimate a linear model of the probability that an individual
chooses to commit a new crime before and after the law passage and use the difference as a
measure of the laws effect on criminal participation.
Because the data is drawn using individuals who are currently under arrest, in order to
estimate how Three Strikes affected the probability of recidivism I examine how their strike
eligibility affects the probability that they commit a crime at some point in the future. For future
crimes, an individual's true strike eligibility includes both the total number of previous
convictions and the current offense if convicted. However, using the true strike eligibility as a
measure of the cost of a future crime is problematic for two main reasons. First, individuals

of assault or drugs who have updated sentencing information.
17

arrested for a felony after the law change may have chosen the type of crime for which they are
currently under arrest in response to the law change. The theoretical prediction that penalty
structure may affect the severity of the crime chosen after the law change makes including the
current offense as part of strike eligibility undesirable. Second, because conviction after Three
Strikes appears to be affected by the law change, a measure of strike eligibility after the law
change includes the endogenously changing conviction rates.
Thus in order to predict the effect of strike eligibility on the probability of recidivating, I
use an individual’s strike eligibility based on his/her prior criminal history committed before the
law change as an instrument for an individual’s true strike eligibility. In order to do this, I
construct an individual’s true strike eligibility as determined by their strike eligibility from their
prior criminal history plus an additional strike if they were convicted of a strikeable offense
(either a felony if Three Strikes was already triggered or a triggering offense). I then construct
four indicators: second strike eligible (strikes2), third strike eligible (strikes3), second strike
eligible after 1994 (after*strikes2) and third strike eligible after 1994 (after*strikes3). I also
construct a PCH based strike eligibility by counting the number of strikes acquired in the period
pre-1994. Then I define the PCH based indicators: strikes2_pch, strikes3_pch,
after*strikes2_pch, after*strikes3_pch which count the number of strikes based on the
individuals prior criminal activity excluding the crime for which they are currently under arrest.
Using these PCH based strike counts, I estimate a first stage of this regression and instrument for
strikes2, strikes3, after*strikes2, after*strikes3 in equation 7. The t-statistics for all first stages
are significant at the 1 percent level.
The data used spans 1990-1999 and thus individuals toward the end for the time series
will be censored. To limit the fraction of the sample that is censored, I estimate the probability
that an individual recidivates within 2-years of his/her release.15 The results of this analysis are
presented in table 4. Columns (1) and (2) compare the OLS and the instrumental variables (IV)
regressions. The OLS appears to be upward biased consistent with a change in the composition
of offenders convicted after Three Strikes changes. If the offenders convicted after the Three
Strikes law change were less likely to recidivate then the post-Three Strikes cohorts would
include some individuals with a lower propensity to recidivate. Thus, some of the reduced
15

The choice of 2 years was based on criminology literature which suggests that most offenders who recidivate will
do so within 2 years of their release from prison. The results presented are not very sensitive to the length of this
18

recidivism from the compositional change in Three-Strikes conviction rates is attributed to the
behavioral response of criminals. The instrumental variable estimate implies a 9 percentage
point (18 percent) reduction in the probability of recidivating among second strike eligible
offenders. The effect for third strike eligible offenders is higher, corresponding to a 14
percentage points or 28 percent reduction. Column (3) includes additional controls for economic
factors and criminal justice expenditures. The estimates appear robust to the inclusion of these
additional variables.
To test how the stringency of enforcement affects the deterrent effect, columns (4)
through (8) report the OLS and IV estimates by city. Because a proportionally high fraction of
the total number of offenders come from Los Angeles, the results from Los Angeles appear
consistent with those in the Three City sample. There does appear a slightly larger difference
between the IV and OLS estimates in Los Angeles relative to either San Diego or San Francisco.
In San Diego, the IV is smaller than the OLS estimate, consistent with the pooled results. In San
Francisco, however, the OLS estimates are smaller than the IV estimates. If they are different,
this would suggest that discretionary use of Three Strikes is resulting in worse criminal being
sentenced under Three Strikes’ harsher sentencing. However, because of the large standard
errors I cannot reject that the IV and the OLS are the same size.

4.2 Estimating the Migration Effect
While the lower probability of recidivating may be due to reduced participation, another
less-desirable way in which crime in California might decline is the migration of repeat offenders
into other states in order to commit crimes. Indeed the probability that a criminal will commit
crime outside of California increased significantly after the Three Strikes law was introduced.
However, because migration might be increasing generally during this time period, I estimate two
specifications attempting to identify the impact of Three Strikes law on inter-state migration of
criminals.
First, I estimate the propensity for criminals to commit crime in California, as a function
of strike eligibility, before and after Three-Strikes law. Specifically, I estimate:
Pr(Crime in CA) = β 0 + β 1 (2 strikes ict ) + β 2 (3strikesict ) + β 3 (after * 2 strikes ) ict
+ β 4 (after * 3strikes ) ict + β 5 ( PCH ict ) + β 6 (individual controls i ) + γ t + δ c

(8)

window. Sensitivity checks using 1-year and 3-years are presented in Appendix Table 2.
19

Again to account for the endogeneity of conviction rates in the post-Three Strikes time period, I
instrument for strike eligibility using the criminal history. The results are presented in columns
(1) through (3) of table 5. The results show a larger participation effect of individuals
committing crime within California. Thus there appears to be a participation effect separate from
the migration effect and the migration effect moves in the opposite direction as the within-state
participation effect.
I next estimate the relationship between strike eligibility and the probability of
committing crime outside of California conditional on recidivating as specified in equation (8).
Consistent with the notion that criminals facing second and especially third strike eligibility
migrate to other states, I find a 6 percentage point increase in migration among second strike
offenders and an 8.5 percentage point change among third strike offenders. Thus it appears that
among offenders committing new crimes, a growing fraction commit those crimes in other states.
The overall migration effect is smaller in magnitude than within state participation but large
relative to the fraction of individuals who migrated to conduct criminal activity prior to the law
change.
In the sample, the two most frequent states to which criminals migrate are Nevada and
Arizona. Nevada’s equivalent of Three-Strikes law applies only for violent offenses and appears
to be rarely invoked.16 Arizona does not have habitual offender legislation. It is worth noting
that these results are likely a lower bound on the estimated probability of migrating. An
individual shows up in the data as having committed a crime in another state if that state requests
criminal records. Because many states may not request criminal records for low-level felonies,
some individuals who migrate out and commit crimes will not appear in this data.

4.3 Estimating the Crime Severity Effect
While Three-Strikes appears to have had the anticipated effect of reducing recidivism
among strike-eligible offenders, it may also have an effect on the distribution of crimes
committed by recidivating criminals conditional on strike eligibility. Following the procedure
used to estimate the participation effect, define V * = U (violent ) − U (non − violent ) . Next,
16

Numerous articles and anecdotal evidence suggest that except in California, Three Strikes statutes are rarely
invoked. See for example New York Times (1996)
20

suppose that V * is a function of an individual's strike eligibility, age-crime rate, prior criminal
history, county characteristics, and individual characteristics, such as age, race/ethnicity, and sex.
Then the

Vict* = β 0 + β1 (2 strikesict ) + β 2 (3strikesict ) + β 3 (after * 2strikes )ict + β 4 (after * 3strikes )ict +

β5 ( PCH ict ) + β 6 (individual controlsi ) + γ t + δ c + ε ict

(9)

Again defining a binary variable V that is 1 if V* is greater than zero and 0 otherwise, I estimate
a linear probability model of the probability that an individual chooses violent crime before and after
the law passage and use the difference as a measure of the laws effect on crime choice. Taking a set of
observations on crime choices, I can estimate the change in the distribution of crime types, i.e. the
marginal deterrence effect. Note that the identification in equation (9) comes solely from individuals
with the same prior criminal history facing differing strike eligibility. Because the sampling requires
that all individuals have current offenses, I can use all individuals with a pre-period determined PCH
and estimate the change in the crime severity differential probability before and after the law change.
Thus unlike in the general participation estimates it is not necessary to estimate an instrumental
variables specification.
Table 6 reports the results of these regressions. Column (1) reports coefficients for a linear
model with the outcome as whether an individual committed a violent crime or not. The estimate of
propensity to commit violent crime indicates that second strike eligible individuals who choose to
commit a felony after Three-Strikes was passed are about 4 percentage points more likely to choose a
violent crime over a nonviolent crime than their counterparts were prior to Three-Strikes. Similarly,
third-strike eligible individuals are about 10 percentage points more likely to commit violent crime.
The similar effect on second strike eligible offenders suggests that the doubling of penalties does not
simply create a higher marginal cost of severity across the board but rather appears to flatten the cost of
more serious crime, possibly due to concavity in the cost function of criminals.

Columns 2 through 8

in Table 6 provide estimates of the probability of committing a given type of crime (conditional on
committing crime). Third strike eligible offenders are more likely to commit rape and robbery and less
to commit burglary and theft. Among second strike eligible offenders the pattern was very similar but
with no significant decline in theft rates. Because burglary is record aggravating offense despite being
nonviolent, offenders who commit crime may be seeking a greater “bang for their buck” by committing
higher payoff, and therefore more violent, crimes. There also appears to be fewer substitutions from
21

burglary but more substitutions from theft among third strike eligible offenders. 17
Overall these results seem consistent with the theory that by eliminating marginal deterrence,
Three-Strikes resulted in a crime distribution that is skewed towards more violent crimes. For
example, the decrease in murder given Three-Strikes seems reasonable since premeditated murder
activates the death penalty, thus preserving marginal deterrence. Therefore, conditional on committing
a violent crime, criminals should substitute away from murder to assault or robbery. The shift away
from non-violent crime towards assault or robbery also seems consistent with the theory the marginal
deterrence is relevant. The most compelling evidence appears in the increase in robbery and the
decrease in burglary. Robbery and burglary are similar crimes in terms of goal, but differ in the
element of force. Moreover, both offenses are record aggravating, which means they generate similar
sentence eligibility.
One alternative explanation for these results is that police officers began charging individuals
with more serious crimes after the passage of Three-Strikes law. If this is correct, then the type of the
crimes committed before and after Three-Strikes are the same and instead police discretion about the
crime with which an offender is charged resulted in more serious charges for Three-Strikes eligible
arrestees. While the use of discretion for an arrest is plausible, it is checked in part by the need for a
judicial arrest warrant. Because the charges for violent felonies, like murder, rape and robbery, are
difficult to compare to any nonviolent or misdemeanor crime it is difficult to imagine that judges would
sanction the substitution of felony charges for lesser degree crimes. Discretion could apply in cases
where individuals are arrested during the commission of a crime or during other exigent circumstances.
However, in these cases it is unlikely that officers would know the strike eligibility of a particular
individual.18

Moreover, it is not necessarily clear that officers would have an incentive to charge more

serious crimes. They might charge less serious crimes after Three Strikes to offenders who they
perceive as less dangerous, which would bias against the results presented in this paper.
Another alternative explanation consistent with the results presented in this paper is that
offenders for non-violent crime are disproportionately deterred from committing crime. Thus,
rather than a substitution effect, the results simply indicate the relative composition effect.
17

These results appear to be linked to Three Strikes law change. I performed falsification checks to test the timing of
the shift in severity by limiting the data to 1990-1993, pre-law change. I artificially assign 1992 as the placebo year
of law change and find no significant changes when estimating the regression equation (9). Results are presented in
Appendix Table 4.
18
This theory would be of greater concern with indictment or conviction level data, where charges often reflect both
the nature of the crime and a bargaining position for plea negotiations. Kesseler and Piehl (1998)
22

Consistent with this view, the unconditional probability of any offense declines, albeit less for
violent than for non-violent crime. Specifically three-quarters of the decline in crime rates
appear to be due to the reduction in criminal participation by non-violent offenders.19 Such
results are consistent with a story where violent criminals are less able to be deterred and thus
enhanced sentencing reduces non-violent crime while keeping violent crime more-or-less
constant. This would produce results similar to those reported in Tables 5 and 6. While there
does not appear to be a systematic way to disentangle the substitution story from the composition
story, I marshal some evidence that suggests that at least some of the effects are due to
substitution. First, there also appears to be a change in the types of crimes committed by firststrike eligible offenders. The probability that a first-strike eligible offender committed a record
aggravating offense declined significantly by 8 percentage points (12 percent).20 Under Three
Strikes the penalty associated with a triggering offense could be higher for individuals expecting
to engage in criminal activity over their lifetime. For these criminals, substituting from record
aggravating to non-record aggravating offenses is consistent with the crime severity substitution
effect. Second, the change in nature of the violent crimes being committed in the post-period
appears more consistent with a substitution story than a compositional story. The increase in the
conditional likelihood to commit violent crime appears entirely driven by robbery and rape.
Thus, the compositional story would require a deterrent effect largely from burglary that results
in the post-Three Strikes distribution being higher only in these two crimes. Finally, an analysis
of lesser-included charges, when looking at the fraction of rapes and assaults which occur during
non-violent crimes in the post-period, there appears to be a significant increase in these rapes
relative to other forms of rapes or assault. Table 7 reports the results from this analysis. For
rapes, lesser-included chargers of other sex-offenses remain relatively constant while rapes with
lesser-included charges of burglary and theft crimes increase. Similarly, assaults with theft
related lesser-included-charges increase while assaults with no lesser-included charges decline.
This combined with the increase in robberies is suggestive of the fact that at least some
individuals may be switching from committing burglary to robbery or may be more willing to
commit a rape or assault during the course of a burglary.
Additionally, the compositional explanation for behavior does not diminish the need for
19
20

This claim is based on estimating equation (7) separately for violent and non-violent offenses.
This evidence is consistent with results in Shepherd (2001) which finds reduced levels of triggering offenses in a
23

marginal deterrence. If offenders who commit violent crimes receive higher payoffs for these crimes,
then harsher penalties are still required to deter these criminals. Thus coincident with this alternative
theory is an alternative justification for maintaining a penalty gradient: proportional sentencing is
necessary to ensure that violent crimes are deterred. This bears directly on the cost-effectiveness of a
an enhanced sentencing policy, but requires a relative valuation of violent, non-violent, and
incarceration costs which are beyond the scope of this paper. However, if this explanation of behavior
is true, then Three Strikes did not encourage any crime that would not have occurred in the absence of
Three Strikes, it simply failed to deter violent crime.

5. CONCLUDING REMARKS
This study presents evidence that an increase in the severity of penalties for all crimes can
generate competing effects. On the one hand, as intended, such a policy appears to reduce participation
in criminal activity. In the case of habitual offender legislation such as Three-Strikes, this effect
appears to be especially concentrated among repeat offenders. Such an effect produces unambiguous
social gains. On the other hand, the broad enhancement of penalty severity reduces the cost of more
severe crimes (such as violent crimes) relative to less severe crimes (such as non-violent crimes). This
produces a social cost for societies who have a distaste for more severe crimes. Thus, while the overall
effect of a sentence enhancement may be a reduction in crime levels, the cost in terms of a higher
fraction of violent crimes may be unpalatable. This study provides additional evidence that criminals,
when faced with harsh penalties in one area, may migrate to other less costly locations. This result is
especially important when considering the efficacy of crime laws which are passed at the state level. If
these laws do littler more than beggar thy neighbor by shifting the worst criminals across the border,
then harsh sentencing regimes may not produce the anticipated reduction in criminal participation but
rather will only serve to shift criminal activity across borders.
In order to better compare the participation, severity, and migration effects, I attempt to quantify
and monetize the estimates of crimes reduced. Based on the estimates presented in this paper, it
appears that on average 148,000 non-violent crimes and 74,000 violent crimes were not committed
each year due to the participation effect of the law change. However, the escalating severity due to the
removal of proportional sentencing resulted in 21,000 additional violent crimes annually. Using
monetized estimates of the cost of crime by the Bureau of Justice Statistics, I estimate that this amounts
structural model using aggregate data.
24

to $193 million dollars.21 A legislative analysis of Three Strikes estimated that the operating costs
resulting from Three Strikes law is nearly one-half billion dollars annually.22 Thus while Three-Strikes
does appear to be effective at deterring crime, the substitution from non-violent to violent crime and the
high cost of incarceration make the law a somewhat costly strategy to reduce crime levels.
Separate from the within state effectiveness of Three-Strikes is the effect of such harsh penalties
on other states. Three Strikes appears to have imposed 50,000 crimes on other states due to the
migration of criminals out of California. Such an affect appears to have been largely unanticipated and
may be extremely costly for other states, especially if the destination states are ill-equipped to handle an
influx of criminals. This effect is particularly important when considering the types of criminal justice
policies advocated because most of these policies occur at the state level. If these laws are successful
in part because they transfer criminal activity across borders then while politically successfully they
maybe socially costly.
The evidence provided in this study highlights the responsiveness of criminals, especially repeat
offenders, to incentive-based penalty schemes. Individuals appear to choose both whether to participate
as well as the form of the participation as a function of the penalty structure. While it may be
surprising that criminals respond so sharply to incentives, sociological evidence (e.g. Shafer) suggests
that criminals are aware of the sentencing structure and their own eligibility for punishment. The
nuanced responsiveness of criminals to smaller enhancements and the effects of strong enhancements
that preserve proportionality with respect to severity are not estimated in this paper and left as an area
of future research.

21

Cost estimates are weighted average of estimated costs for types of crimes from Miller, Cohen, and Wiersema
updated to 2001 dollars.
22
These estimates are substantially smaller than the estimates presented in previous work (such as Greenwood et al.)
The primary reason for the difference is the effect of discretion (i.e. the use of judicial discretion to dismiss prior
strikes and variation among counties in willingness to prosecute offenders under the Three Strikes law.)
25

REFERENCES

Becker, Gary (1968), “Crime and Punishment: And Economic Approach” Journal of Political
Economy 76:169-217

Beres, Linda S., and Thomas D. Griffith (1998), “Did ‘Three-Strikes’ Cause the Recent Drop in
California Crime” An Analysis of the California Attorney General’s Repot”. Loyola of Los Angeles
Law Review 32:101

Brown, Brian and Greg Jolivette (2005) “A Primer: Three Strikes - The Impact After More Than a
Decade” Legislative Analyst's Office Report

California Department of Corrections (2002), Weekly Report Of Population, Data Analysis Unit,
State of California, March 24

Census Population Reports, State Estimates, 1986-1998

Clarke, John, James Austin, and D. Alan Henry (1997), Three-Strikes and You’re Out”: A Review of
State Legislation, National Institute of Justice, Washington D.C.

Cushman, Robert. (1996), “The Effect of ‘Three-Strikes and You’re Out’ on Corrections,” in ThreeStrikes and You’re Out: Vengeance as Public Policy, David Schichor and Dale Sechrest eds., p.
155-174, Thousand Oaks, CA: SAGE Publications.

Dickey, Walter (1996),The Impact of Three-Strikes You’re Out” Laws: What Have We Learned
Washington D.C.: Campaign for An Effective Crime Policy

DiIulio, John and Anne Morrison Piehl (1991), “Does Prison Pay? The Stormy National Debate
over the Cost-Effectiveness of Imprisonment”, The Brooking Review, Fall

Ewing v. California (2003). Certiorari to the Court Of Appeal Of California, Second Appellate
26

District, No.01 –6978. Argued November 5,2002 —Decided March 5,2003

Flynn, Edith E. et al. (1997), Three-Strikes Legislation: Prevalence and Definitions,” in Critical
Criminal Justice Issues: Task Force Reports from the American Society of Criminology to Attorney
General Janet Reno (Washington, DC: US Department of Justice/Office of Justice Programs, NCJ
158837)

Greenwood, et al. (1994) Three-Strikes and You’re Out: Estimated Benefits and Costs of
California’s New Mandatory-Sentencing Law” (Santa Monica, CA: RAND)

Hawken, Angela and Peter Greenwood (2001), “Three-Strikes and You’re Out: A Review of the
Research Evidence On Impacts” Background Report for Three-Strikes Roundtable at RAND

Helland, Eric and Tabarrok (2007) Does Three Strikes Deter? A Non-Parametric Estimation.
Journal of Human Resources forthcoming

Jaimeson, Ross (1999), “Striking Out: The Failure of California's ‘Three-Strikes and You're Out’
Law”, Stanford Law and Policy Review, Fall

Kessler, Daniel and Anne Morrison Piehl (1998), “The Role of Discretion in the Criminal Justice
System” Journal of Law, Economics and Organization, 14(2):256-276

Macallair, Daniel and Michael Males (1999), “Striking Out: The Failure of California of
California’s ‘Three-Strikes You’re Out’ Law.” Sand Francisco, CA: The Justice Policy Institute

Males, Michael and Dan Macallair, and Khaled Taqi-Eddin. (1999). “California Three-Strikes
Ineffective.” Overcrowded Times 10:1 14-16

Marvell, Thomas B., and Carlisle E. Moody (2001), “The Lethal Effects of Three-Strikes Laws,”
Journal of Legal Studies, v. 30, p. 89-106.

27

Miller, Ted, Cohen, Mark A., Wiersema, Brian. (1996) “Victim Costs and Consequences: A New
Look”, National Institute of Justice Research Report

New York Times (1996), “’Three-Strikes’ Rarely Invoked in Courtrooms” A1:1, September 10

Schafer, John (1999), “The Deterrent Effect of Three-Strikes Law”, FBI Law Enforcement Bulletin,
April

Shepherd, Joanna M. (2002), "Fear of the First Strike: The Full Deterrent Effect of California's
Two-and Three-Strikes Legislation." Journal of Legal Studies 31: 159-201.

Stigler, George J. (1970), “The Optimum Enforcement of Laws.” Journal of Political Economy
78:526-36

Zimring, Franklin E., Hawkins, Gordon, and Sam Kamin (2001) Punishment and Democracy:
Three-Strikes and You’re Out in California” Oxford ; New York : Oxford University Press

28

θParticiate

0

.1

.2

.3

.4

θseverity

-4

-2

0
theta

2

4

Figure 1. Example of Distribution of Individual Cost Parameter, θ, and Cut-Off Values

29

Table 1: California Three Strikes Record Aggravating Offenses
Murder
Murder
voluntary manslaughter
Sex Offenses

Rape
Sodomy by force, violence, duress, menace, or threat of injury
Oral copulation by force, violence, duress, menace, or threat of injury
Lewd acts on a child under 14
Continuous sexual abuse of a child

Assault

Attempted murder
Assault with the intent to commit mayhem, rape, sodomy, or oral copulation

Robbery

Any Robbery

Other Violent
Crimes

Mayhem.
Any felony in which the defendant inflicts great bodily injury on any person
Kidnapping
Carjacking
Arson which results in Bodily Harm
Exploding device with intent to injure or kill

Property
Crimes

Arson
Burglary of a Home or Dwelling
Grand Theft

Drug Offenses

Drug Sales to Minors
Drug Trafficking

Other Felonies

Any felony in which the defendant uses a firearm
Threats to victims or witnesses
Extortion
Any felony punishable by death or imprisonment for life.

Violent Felonies

Serious Felonies
(Non-Violent)

Source: California Penal Code, Part 1. Title 16. General Provisions 667

Table 2: Summary statistics of in sample and total population
1-Strike Eligible
12,685

2-Strike Eligible
2,788

3-Strike Eligible
1,659

67%

71%

87%

Black
Hispanic
White

31%
41%
22%

35%
32%
29%

45%
36%
17%

Current Crime
Violent
Property
Drugs
Other

31%
32%
25%
12%

37%
29%
24%
10%

43%
31%
17%
9%

Prior Criminal History
Number of Prior Arrests
Number of Prior Felony Convictions
Number of Violent Convictions

1.1
0.7
0

2.6
2.1
0.8

4.2
3.7
1.2

Current Offense
Convicted on current offense
Receive Life Sentence
Average Sentence Length (in months)

24%
0.3
22

27%
0.7
45

44%
5.3
67

N
Sex
Male
Race

Future Criminal Activity
Probability Recidivate within 2 years
53%
43%
41%
Number of Future Convictions
1.3
2.4
2.2
Number of Future Violent Convictions
0.6
1.0
1.1
Note: Life sentences are entered as 25 years for average sentence length computation. Violent Crimes include murder, rape, assault and robbery. Property crimes
include

Table 3: Median Sentences in Pre and Post Three Strikes Period, by Crime Type and Offender Strikes
1990-1993
1994-1999
(Pre-Three Strikes)
(Post-Three Strikes)
Panel A: First Strike Eligible
Murder
20 years
20 years
Rape
4.9 years
5 years
Assault
4.3 years
3 years, 6 months
Robbery
3 years
3 years
Burglary
9 months
1 year
Theft
6 months
9 months
Drugs
9 months
1 years 2 months
Panel B: Second Strike Eligible
Murder
23 years
27 years
Rape
5 years
9 years
Assault
1 year
2 years, 5 months
Robbery
3years 5 months
6years 5 months
Burglary
1year 3 months
3 years
Theft
1 year
2 years, 8 months
Drugs
1 year
4 years
Panel C: Third Strike Eligible
Murder
20 years
Life
Rape
9 years
30 years
Assault
6.5 years
23 years
Robbery
4 years
21 years
Burglary
2 years
22 years
Theft
1.2
26 years
Drugs
2
25 years
Source: Three County Survey of Arrest Record in Los Angeles, San Diego, and San Francisco, 1990-1999. Offenders who
committed “other” offenses are excluded from the sample. All sentences are truncated at 60 years. Offenders with missing
sentencing data are omitted. Sample size is 17,264.

Table 4. Linear Estimates of Probability of Recidivism by Strike Eligibility
(1)

(2)
(3)
All Three Cities
0.49
-0.1046** - 0.0931**
-0.0814*
(0.0415)
(0.0461)
(0.0418)

(4)
(5)
Los Angeles
0.51
-0.1387**
-0.1143**
(0.0621)
(0.0678)

San Diego
0.46
-0.0742**
-0.0651*
(0.0411)
(0.0440)

-0.1822**
(0.0713)

-0.1434*
(0.0793)

-0.1411**
(0.0721)

-0.2214***
(0.0443)

-0.1685***
(0.0471)

-0.1254**
(0.0493)

-0.1011**
(0.0471)

-0.1119***
(0.0409)

2 strikes
(=1 if second strike eligible)

0.0462
(0.0317)

0.0251
(0.0165)

0.0264
(0.0171)

0.0532
(0.0417)

0.0546
(0.0376)

0.0324
(0.0456)

0.0438
(0.0379)

0.0178
(0.0365)

0.0176
(0.0326)

3 strikes
(=1 if third strike eligible)

0.0643
(0.0436)

0.0471
(0.0795)

0.0238
(0.0624)

0.0526
(0.0757)

0.0564
(0.0795)

0.0471
(0.0795)

0.0496
(0.0775)

0.0784
(0.0736)

0.0464
(0.0714)

Pr(Commit Crime within 2 years)
Dependent Variable Mean
after*2strikes

after*3strikes

(6)

(7)

(8)
(9)
San Francisco
0.47
-0.0675*
-0.0712*
(0.0405)
(0.0410)
-0.1412*
(0.0771)

Y
Y
Y
N
N
N
N
N
N
County Fixed Effects
Y
Y
Y
Y
Y
Y
Y
Y
Y
Year Fixed Effects
Controls for Economic
N
N
Y
Y
Y
Y
Y
Y
Y
Characteristics a
Controls for Police and Judicial
N
N
Y
Y
Y
Y
Y
Y
Y
Spending b
OLS
IV
OLS
OLS
IV
IV
OLS
IV
IV
Estimation Strategy c
1,921
1,921
17,264
17,264
17,264
12,514
12,514
2,829
2,829
Observations
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Coefficients reported are an indicator variable for individuals who are second strike
eligible, and an interaction term between the year indicator variables and strikes indicator variables. Also included in all specifications but not reported are variables for age,
race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history variable is a vector of variables counting the number of times an
individual was convicted of a felony by crime category. Crime categories include murder, rape, assault, robbery, burglary, theft, drug crimes, and other felonies. Standard
errors, reported in parentheses, are clustered by county of arrest
a. Economic Characteristics include county-year measures of unemployment rate and percent of population below poverty.
b. Police and Judicial spending controls include county-year expenditures on police, prosecution, public defense, and judiciary.
c. Instrumental variables estimates instrument for prior criminal history using arrest for offenses.

Table 5. Linear estimates of the Probability of Committing Crime outside of California by Strike Eligibility

Dependent Variable Mean
after*2strikes

(1)
(2)
(3)
Pr(Commit Crime in CA within 2 years)
0.43
-0.1272**
- 0.0882**
-0.0866*
(0.0578)
(0.0465)
(0.0464)
-0.2165**
(0.0779)

-0.1621**
(0.0822)

2 strikes
(=1 if second strike eligible)

0.0652
(0.0465)

3 strikes
(=1 if third strike eligible)

0.0864*
(0.0521)

after*3strikes

(4)
(5)
(6)
Pr(Commit Crime Outside CA within 2 yrs | Commit Crime)
0.13
0.0692**
0.0614**
0.0611*
(0.0299)
(0.0312)
(0.0316)

-0.1599**
(0.0791)

0.0918**
(0.0414)

0.0874*
(0.0466)

0.0867*
(0.043)

0.0541
(0.0454)

0.0610
(0.0465)

0.0031
(0.0315)

0.0061
(0.0376)

0.0059
(0.0371)

0.0412
(0.0678)

0.0431
(0.0645)

0.0093
(0.0516)

0.0167
(0.0613)

0.0163
(0.0615)

Y
Y
Y
Y
Y
Y
County Fixed Effects
Y
Y
Y
Y
Y
Y
Year Fixed Effects
N
N
Y
N
N
Y
Controls for Economic Characteristics a
Controls for Police and Judicial
N
N
Y
N
N
Y
Spending b
IV
OLS
IV
IV
OLS
IV
Estimation Strategy c
17,264
Observations
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Coefficients reported are an indicator variable for individuals who are second strike
eligible, and an interaction term between the year indicator variables and strikes indicator variables. Also included in all specifications but not reported are variables for age,
race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history variable is a vector of variables counting the number of times an
individual was convicted of a felony by crime category. Crime categories include murder, rape, assault, robbery, burglary, theft, drug crimes, and other felonies. Standard
errors, reported in parentheses, are clustered by county of arrest
a. Economic Characteristics include county-year measures of unemployment rate and percent of population below poverty.
b. Police and Judicial spending controls include county-year expenditures on police, prosecution, public defense, and judiciary.
c. Instrumental variables estimates instrument for prior criminal history using arrest for offenses.

Table 6. Linear Estimates of the Change in Crime Severity by Strike Eligibility
(1)
Violent crime

(2)
murder

(3)
rape

(4)
assault

(5)
robbery

(6)
burglary

(7)
theft

(8)
drugs

0.34
0.0412**
(0.0205)

0.01
-0.0048
(0.0031)

0.04
0.0432**
(0.0192)

0.14
0.0289*
(0.0125)

0.15
0.0593**
(0.0237)

0.14
-0.0679***
(0.0260)

0.09
-0.0208
(0.0296)

0.27
-0.0147
(0.0211)

after*3strikes

0.0956***
(0.0295)

-0.0037
(0.0021)

0.0526
(0.0221)

0.0541*
(0.0236)

0.1214*
(0.0525)

-0.0680*
(0.0168)

-0.794*
(0.0628)

0.0713
(0.0448)

2 strikes
(=1 if second strike eligible)

-0.0334
(0.0290)

-0.0014
(0.0017)

0.0267
(0.0342)

0.0947
(0.1016)

0.0372
(0.0244)

-0.0107
(0.0158)

0.1147
(0.0949)

-0.0461
(0.0714)

3 strikes
(=1 if third strike eligible)

-0.0679
(0.0469)

-0.001
(0.0037)

0.0129
(0.0645)

-0.0207
(0.0581)

0.0210*
(0.0436)

-0.0012
(0.0361)

0.1749
(0.0688)

0.0160
(0.0202)

(Probability of committing Y |
Committing a Crime)
Dependent Variable Mean
after*2strikes

County Fixed Effects
Y
Y
Y
Y
Y
Y
Y
Y
Year Fixed Effects
Y
Y
Y
Y
Y
Y
Y
Y
Controls for Economic
Characteristicsa
Y
Y
Y
Y
Y
Y
Y
Y
Controls for Police and Judicial
Y
Y
Y
Y
Y
Y
Y
Y
Spendingb
Observations
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Coefficients reported are an indicator variable for individuals who are second strike
eligible, and an interaction term between the year indicator variables and strikes indicator variables. Also included in all specifications but not reported are variables for age,
race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history variable is a vector of variables counting the number of times an
individual was convicted of a felony by crime category. Crime categories include murder, rape, assault, robbery, burglary, theft, drug crimes, and other felonies. Standard
errors, reported in parentheses, are clustered by county of arrest
a. Economic Characteristics include county-year measures of unemployment rate and percent of population below poverty.
b. Police and Judicial spending controls include county-year expenditures on police, prosecution, public defense, and judiciary.

Table 7. Linear Estimates of Lesser Included Charges (LIC) for Rape and Assault, by Strike Eligibility
Panel A: Rape Sub-Categories
Overall Effect
(all LIC)
Dependent Variable Mean
after*2strikes

after*3strikes

other LIC

burglary or theft LIC

Drug LIC†

no LIC

0.042

0.0223

0.0101

--

0.003

0.0432**

0.0182

0.0331*

--

-0.007

(0.0192)

(0.0132)

(0.0196)

0.0526**

0.0112

0.0475*

(0.0221)

(0.0175)

(0.0283)

0.144

0.017

0.029

0.036

0.062

0.0289*

-0.003

0.0317*

0.0253

-0.0251*

(0.0125)

(0.0163)

(0.0171)

(0.0183)

(0.0142)

0.0541*

-0.012

0.0553*

0.0413

-0.0305

(0.0236)

(0.0261)

(0.0331)

(0.0292)

(0.0274)

17,264

17,264

17,264

17,264

17,264

Y

Y

Y

Y

Y

(0.024)
--

0.0119
(0.0332)

Panel B: Assault Sub-Categories
Dependent Variable Mean
after*2strikes

after*3strikes

Observations
County Fixed Effects

Year Fixed Effects
Y
Y
Y
Y
Y
Controls for Economic
Characteristics a
Y
Y
Y
Y
Y
Controls for Police and Judicial
Spending b
Y
Y
Y
Y
Y
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Coefficients reported are an indicator variable for individuals who are second strike
eligible, and an interaction term between the year indicator variables and strikes indicator variables. Also included in all specifications but not reported are variables for age,
race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history variable is a vector of variables counting the number of times an
individual was convicted of a felony by crime category. Crime categories include murder, rape, assault, robbery, burglary, theft, drug crimes, and other felonies. Standard
errors, reported in parentheses, are clustered by county of arrest.
†There were insufficient numbers of rape offenses with lesser-included-charges related to drugs for specification.
a. Economic Characteristics include county-year measures of unemployment rate and percent of population below poverty.
b. Police and Judicial spending controls include county-year expenditures on police, prosecution, public defense, and judiciary.

Appendix Table 1: Crime Categories and Definitions
Crime
Murder

Definition
All willful (non-negligent) killing of one human being by another

Rape

Forcible sexual contact

Assault

Unlawful attack by one person upon another for the purpose of
inflicting severe or aggravated bodily injury, usually
accompanied by the use of a weapon or by means likely to
produce death or great bodily harm.

Robbery

The taking or attempting to take anything of value from the care,
custody or control of a person or persons by force or threat of
force or violence and/or by putting the victim in fear.

Burglary

Included Offenses (California Penal Code Sections)
Murder (§187)
Voluntary Manslaughter (§192a)
Involuntary Manslaughter (§192b)
Gross Vehicular Manslaughter while intoxicated (§193.5)
Forcible rape, spousal rape (§261, §262)
Forcible Sodomy or Oral Copulation (§286, 288a)
Sexual assault with an object (§289)
Lewd or Lascivious acts of continuous sex abuse of a child (§288, 288.5)
Sexual battery (§243.4)
Mayhem, Aggravated Mayhem (§203, 205)
Torture (§206)
Assault with intent to commit Mayhem or sex offenses (§220)
Assault with Caustic Chemicals or Taser gun (§244, 244.5)
Assault with deadly weapon or by force (§245)
Infliction of injury on spouse, cohabitee or parent of child (§273.5)
Robbery (§211)
First and Second Degree Robbery (§212.5)
Train Robbery, Car Jacking (§214, 215)

The unlawful entry of a structure to commit a felony or theft.
Burglary (§459)
The use of force to secure entry is often a part of burglary but is
Looting (§463)
not required for a burglary charge.
Theft
The unlawful taking, carrying, leading or riding away of property Larceny (§484-502.9)
from the possession or constructive possession of another in
Motor vehicle theft (§10851)
which no use of force, violence or fraud occurs.
Drugs
The unlawful possession, sale, use, growing, manufacturing, and
Any individual subject to California Major Narcotic Vendors Prosecution Law
making of narcotic drugs. The relevant substances include:
(§13883) who is under arrest for violation of the Health and Safety Code
opium or cocaine and their derivatives (morphine, heroin,
Narcotics (§11350-11356.5)
codeine); marijuana; synthetic narcotics (Demerol, methadone);
Controlled Substances formerly classified as restricted dangerous drugs
and dangerous non-narcotic drugs (barbiturates)
(§11377-11382.5)
Note: Definitions from Uniform Crime Reporting Handbook. Not all potentially included offenses are included in the sample

Appendix Table 2. Sensitivity of Linear Probability Estimates of Length of time for Probability of Recidivate
(1)
(2)
(3)
Pr(Commit Crime in CA within 1 years)
0.43
-0.1361**
- 0.1010**
-0.1002*
(0.0498)
(0.0522)
(0.0521)

(4)
(5)
(6)
Pr(Commit Crime in CA within 3 years)
0.13
-0.1125**
- 0.0914*
-0.0913*
(0.0263)
(0.0511)
(0.0499)

-0.2614**
(0.0361)

-0.2211**
(0.0672)

-0.1832**
(0.0323)

-0.1632**
(0.0512)

2 strikes
(=1 if second strike eligible)

0.0553*
(0.0311)

0.0532
(0.0433)

0.0513
(0.0421)

0.0352
(0.0263)

0.0332
(0.0613)

0.0388
(0.0625)

3 strikes
(=1 if third strike eligible)

0.0815
(0.0492)

0.0741
(0.0568)

0.0713
(0.0561)

0.0963**
(0.0273)

0.0716
(0.0512)

0.0713
(0.0555)

Dependent Variable Mean
after*2strikes

after*3strikes

-0.2159**
(0.0671)

-0.1613**
(0.0528)

Y
Y
Y
Y
Y
Y
County Fixed Effects
Y
Y
Y
Y
Y
Y
Year Fixed Effects
N
N
Y
N
N
Y
Controls for Economic Characteristics a
N
N
Y
N
N
Y
Controls for Police and Judicial Spending b
c
IV
OLS
IV
IV
OLS
IV
Estimation Strategy
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Reported values are marginal effects evaluated at the mean. Column (1) dependent
variable is an indicator for whether the current offense is violent. Violent offenses are murder, sex offenses, assault and robbery. The dependent variables for columns (2)(8) are indicator variables for whether an individual committed a given crime type (types are murder, sex offenses, assault, robbery, burglary, theft, drugs). Coefficients
reported are an indicator variable for individuals who are second strike eligible, and an interaction term between the year indicator variables and strikes indicator variables.
Also included in all specifications but not reported are variables for age, race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history
variable is a vector of variables counting the number of times an individual was convicted of a felony by crime category. Crime categories include murder, rape, assault,
robbery, burglary, theft, drug crimes, and other felonies. Standard errors, reported in parentheses, are clustered by county of arrest
a. Economic Characteristics include county-year measures of unemployment rate and percent of population below poverty.
b. Police and Judicial spending controls include county-year expenditures on police, prosecution, public defense, and judiciary.
c. Instrumental variables estimates instrument for prior criminal history using arrest for offenses.

Appendix Table 3. Sensitivity of Linear Probability Estimates for Sample Time Frame
(1)
Dependent Variable Mean
after*2strikes

after*3strikes

(2)
0.43

-0.1046**
(0.0415)

-0.0814*
(0.0418)

-0.1822**
(0.0713)

-0.1411**
(0.0721)

(3)
(4)
Pr(Commit Crime in CA within 2 years)
0.31
-0.1183**
-0.0913
(0.0515)
(0.0564)

(5)

(6)
0.45

-0.1099**
(0.0451)

-0.0899*
(0.0513)
-0.1611**
(0.0783)

-0.1316**
(0.0588)

-0.1292*
(0.0769)

-0.1786**
(0.0727)

2 strikes
(=1 if second strike eligible)

0.0462
(0.0317)

0.0264
(0.0171)

0.0325
(0.0235)

0.0329
(0.0372)

0.0511
(0.0361)

0.0416
(0.0400)

3 strikes
(=1 if third strike eligible)

0.0643
(0.0436)

0.0238
(0.0624)

0.0734
(0.0481)

0.0226
(0.0143)

0.0701
(0.0511)

0.0611
(0.0649)

Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
N
Y
N
Y
N
Y
N
Y
N
Y
N
Y
OLS
IV
OLS
IV
OLS
IV
1990-1999
1990-1999
1990-1999
1990-1999
1990-1996
1990-1996
At least 1 prior
At least 1 prior No restriction on No restriction on At least 1 prior
At least 1 prior
>90,
>90,
date of prior
date of prior
>90,
>90,
All priors <94
All priors <94
All priors <94
All priors <94
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Reported values are marginal effects evaluated at the mean. Column (1) dependent
variable is an indicator for whether the current offense is violent. Violent offenses are murder, sex offenses, assault and robbery. The dependent variables for columns (2)(8) are indicator variables for whether an individual committed a given crime type (types are murder, sex offenses, assault, robbery, burglary, theft, drugs). Coefficients
reported are an indicator variable for individuals who are second strike eligible, and an interaction term between the year indicator variables and strikes indicator variables.
Also included in all specifications but not reported are variables for age, race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history
variable is a vector of variables counting the number of times an individual was convicted of a felony by crime category. Crime categories include murder, rape, assault,
robbery, burglary, theft, drug crimes, and other felonies. Standard errors, reported in parentheses, are clustered by county of arrest
a. Economic Characteristics include county-year measures of unemployment rate and percent of population below poverty.
b. Police and Judicial spending controls include county-year expenditures on police, prosecution, public defense, and judiciary.
c. Instrumental variables estimates instrument for prior criminal history using arrest for offenses.
d. Sample used in the analysis of the paper spans 1990-1999 and in order to be included requires offenders

County Fixed Effects
Year Fixed Effects
Controls for Economic Characteristics a
Controls for Police and Judicial Spending b
Estimation Strategy c
Sample Used for Analysisd

Appendix Table 4. Falsification Checks of Probability of Current Crime Type for Second and Third Strike Eligible Arrestees, 1990-1993
(1)
Violent crime
0.0011
(0.0256)

(2)
murder
-0.0048
(0.0073)

(3)
rape
0.0015
(0.0162)

(4)
assault
0.0164
(0.0123)

(5)
robbery
0.0176
(0.0313)

(6)
burglary
-0.0226
(0.0322)

(7)
theft
-0.0046
(0.0267)

(8)
drugs
-0.0029
(0.0201)

(after 1992)*3strikes

0.0064
(0.0361)

-0.0037
(0.0092)

0.0032
(0.0120)

0.0251
(0.0232)

0.0356
(0.0785)

-0.0327*
(0.0331)

-0.0098
(0.0538)

0.0052
(0.0451)

2 strikes
(=1 if second strike eligible)

0.0042
(0.0132)

-0.0014
(0.0061)

0.0267
(0.0342)

0.0123
(0.0165)

0.0354
(0.0451)

-0.0153
(0.0298)

0.0794
(0.0949)

-0.0171
(0.0645)

3 strikes
(=1 if third strike eligible)

0.0263
(0.0236)

-0.001
(0.0086)

0.0178
(0.0412)

-0.0236
(0.0592)

0.0326
(0.0476)

-0.0312
(0.0301)

0.0189
(0.0465)

0.0245
(0.0465)

(after 1992)*2strikes

County Fixed Effects
Y
Y
Y
Y
Y
Y
Y
Y
Year Fixed Effects
Y
Y
Y
Y
Y
Y
Y
Y
Controls for Economic
Characteristics
N
N
N
Y
N
N
N
Y
Controls for Police and Judicial
Spending
N
N
N
Y
N
N
N
Y
Offender with criminal history pre1990
N
Y
N
N
N
Y
N
N
All Offenders
N
N
Y
N
N
N
Y
N
Observations
Note: Results that are significant at .05 (0.1, 0.01) are reported with **, (*, ***). Reported values are marginal effects evaluated at the mean. Column (1) dependent
variable is an indicator for whether the current offense is violent. Violent offenses are murder, sex offenses, assault and robbery. The dependent variables for columns (2)(8) are indicator variables for whether an individual committed a given crime type (types are murder, sex offenses, assault, robbery, burglary, theft, drugs). Coefficients
reported are an indicator variable for individuals who are second strike eligible, and an interaction term between the year indicator variables and strikes indicator variables.
Also included in all specifications but not reported are variables for age, race, ethnicity, sex, felony rate per criminal year, and prior criminal history. Prior criminal history
variable is a vector of variables counting the number of times an individual was convicted of a felony by crime category. Crime categories include murder, rape, assault,
robbery, burglary, theft, drug crimes, and other felonies. Standard errors, reported in parentheses, are clustered by county of arrest