Social Science Research 45 (2014) 3345

Contents lists available at ScienceDirect

Social Science Research

journal homepage: www.elsevier.com/locate/ssresearch

Link between unemployment and crime in the US:

A Markov-Switching approach q
Firouz Fallahi a,, Gabriel Rodrguez b
a
Department of Economics, University of Tabriz, Tabriz, Iran
b
Department of Economics, Ponticia Universidad Catlica del Per, Peru

a r t i c l e i n f o a b s t r a c t

Article history: This study has two goals. The rst is to use Markov Switching models to identify and ana-
Received 5 February 2013 lyze the cycles in the unemployment rate and four different types of property-related crim-
Revised 14 December 2013 inal activities in the US. The second is to apply the nonparametric concordance index of
Accepted 22 December 2013
Harding and Pagan (2006) to determine the correlation between the cycles of unemploy-
Available online 2 January 2014
ment rate and property crimes. Findings show that there is a positive but insignicant rela-
tionship between the unemployment rate, burglary, larceny, and robbery. However, the
Keywords:
unemployment rate has a signicant and negative (i.e., a counter-cyclical) relationship
Markov-Switching models
Business cycles
with motor-vehicle theft. Therefore, more motor-vehicle thefts occur during economic
Unemployment expansions relative to contractions. Next, we divide the sample into three different subs-
Crime amples to examine the consistency of the ndings. The results show that the co-move-
Concordance ments between the unemployment rate and property crimes during recession periods
are much weaker, when compared with that of the normal periods of the US economy.
2013 Elsevier Inc. All rights reserved.

1. Introduction

Although economic theory anticipates the existence of a positive relationship between unemployment and crime, empir-
ical works have found mixed results. Chiricos (1987) reviewed 68 studies about the relationship between crime and unem-
ployment rate, and found that only less than half of these studies have found positive signicant effects of unemployment on
crime rates. That is, most of these studies have shown a negative or no relationship between crime and unemployment.
Cantor and Land (1985) and Cook and Zarkin (1985) are among those who have tried to explain these mixed ndings of the
empirical studies. Cook and Zarkin (1985) stated that during the recession, the opportunity cost of attempting to crime and
also the legitimate opportunities of earning income deteriorate. Therefore, the number of property and violent crimes esca-
late; they called this the motivation effect.1 In addition, recession might cause a reduction in the policing due to lower tax rev-
enues of governments. The results of this change might be a lower chance of crime detection and punishment; hence, it might
increase the crime rate. On the other hand, during recession, the number and quality of criminal opportunities decline (Cook,
2010), because more people stay at home, which increases the guardianship of properties, and hence, the number of burglary
victims would decrease. The same could be said for the robbery, because people carry less cash, and as Cook (2010) stated, they
have more tendencies to defend what they have. Cantor and Land (1985) called this the opportunity effect. Motivational and
opportunity effects would produce upward and downward uctuations in crime rates; consequently, the net effect of unemploy-
ment, as an indicator of bad economic conditions, on crime is ambiguous. In other words, when the opportunity effect is larger

q
We thank the editor, two anonymous referees, Lynda Khalaf, Marcel Voia, Gamal Atallah, and David Greenberg for their extremely thoughtful and
thorough comments, which substantially improved the quality and presentation of the paper.
Corresponding author. Fax: +98 411 335 6013.
E-mail address: ffallahi@tabrizu.ac.ir (F. Fallahi).
1
In fact, motivational effect, originally, has been introduced by Becker (1968).

0049-089X/$ - see front matter 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.ssresearch.2013.12.007
34 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345

(smaller) than the motivation effect, there would be a negative (positive) link between unemployment rate and crime. Therefore,
the effect of a business cycle, proxied by unemployment, on crime might go in either direction or even it could be null.2
Besides, as Cantor and Land (1985) argued, the occurrence timing of these two effects could be different, i.e., opportunity
effect would appear instantaneously, whereas the motivation would respond with a lag (Arvanites and Dena, 2006). In fact,
the opportunity effect is instantaneous, but the motivation effect needs time; hence, it is lagged. Moreover, these two effects
have opposite signs and, therefore, they might cancel each another.3
As mentioned earlier, the association between the unemployment rate and crime has been investigated extensively in
different studies. However, no unequivocal support for the link between unemployment and crime has been found. Some
of these studies have employed national aggregate data (Kollias and Paleologou, 2012; Cook and Zarkin, 1985; Cantor and
Land, 1987; Hale and Sabbagh, 1991; Allen, 1996; Greenberg, 2001; Fallahi et al., 2012); while others; such as those by Entorf
and Spengler (2000), Levitt (2001), and Edmark (2005), based their investigation on subnational; i.e. state- or county-level
data. Phillips and Land (2012) are among those who analyzed this subject at the county, state, and national levels to shed
more light on the impact of data aggregation level on ndings. Andersen (2012) used neighborhood-level panel data to
examine the association between crime and unemployment. Britt (1997) and Lee and Holoviak (2006) examined age-specic
data on crime; whereas Witt et al. (1999) highlighted the importance of gender on crime.
These studies not only differ in terms of aggregation level of the data used for analysis (e.g., state-level vs. national-levels),
they also differ in their ndings. Several studies have shown a positive relationship between unemployment and property
crime, while others have found a negative relationship. For example, Edmark (2005) and Fallahi et al. (2012) showed that
unemployment and auto-theft co-vary signicantly, which means that the rate of auto-theft is higher during economic con-
traction. On the other hand, studies by Cook and Zarkin (1985), Allen (1996), and Phillips and Land (2012) suggested that the
unemployment rate and rate of auto-theft move in opposite directions, that is, more auto-theft occurs during economic
expansions. Yet, the study by Hale and Sabbagh (1991) is among those that concluded that unemployment has no signicant
inuence on auto-theft.
The same is true for burglary; Cook and Zarkin (1985), Cantor and Land (1987), Allen (1996), and Edmark (2005) identi-
ed a positive co-movement between the unemployment rate and rate of burglary. However, Fallahi et al. (2012) and Phillips
and Land (2012) came to a conclusion that the rate of burglary is signicantly lower during recessions, i.e., a negative rela-
tionship between business cycles and burglary.
Larceny is another property crime that has gained attention from research community. The results about the link between
larceny and unemployment are mixed and contradictory, with Cantor and Land (1987) and Hale and Sabbagh (1991) docu-
menting a positive link, but Phillips and Land (2012) nding a negative relationship. The majority of studies on robbery
found a positive signicant effect of unemployment on robbery (Cook and Zarkin, 1985; Hale and Sabbagh, 1991; Allen,
1996), but Edmark (2005) claimed that there is no signicant link between unemployment and robbery.4
An important step forward in understanding the relationship between crime and unemployment was introducing the ef-
fect of unemployment lag to the literature. Some scholars argue that it is the lagged unemployment, rather than current
unemployment, that has a potential inuence on crime; in fact, they try to distinguish the motivational and opportunity ef-
fects. Taking this point into consideration, Allen (1996) showed that lagged unemployment has a negative effect on larceny,
burglary, and auto-theft. Greenberg (2001) and Allen (1996) revealed the same result for robbery. Yet, in another study, Can-
tor and Land (1987) showed that lagged unemployment positively affects larceny and burglary.
Most of these studies were carried out using multiple regression models, vector autoregression or error correction mod-
els. All of these methods assume a stable behavior of the variable under examination. However, the unemployment rate is
directly related to business cycles and has a cyclical pattern; hence, linear models may provide a weak t. The same is true
for crime as well (see Freeman et al., 1999 for a survey). Consequently, one needs to use nonlinear models, such as the Mar-
kov Switching (MS) models. These models are able to capture changes in the behavior of time-series by allowing the switch-
ing between regimes or states.5 MS models have been used widely in the literature and have proven their ability to explain the
changes in the pattern of time series.
This study has two goals. The rst goal is to examine the behavior of different types of property-related felonies and the
second goal deals with the potential relationship between the unemployment rate and property crime rates in the US econ-
omy. Our sampling period covers 1977:2 until 2004:4 and four types of property crimes6 are analyzed: burglary, larceny, mo-
tor-vehicle theft, and robbery.7

2
The effect of business cycles on crime could be very complex, which makes the prediction of these effects really difcult (Cook, 2010).
3
It is worth mentioning that, as stated by Freeman et al. (1999), the estimated signicant relationship between unemployment and crime does not
necessarily show the causality between these variables.
4
As one of the referees of the present paper points out, one reason for these contradictory results might be the difference in the data type used to study the
link. Some of these papers have used cross-section data; hence, the results show the relationship between unemployment and crime at one point of time.
However, others have used time-series data; hence, the results show the link over time. Consequently, the former can be regarded as a static approach, while
the latter is more similar to dynamic approaches. Therefore, it is better to be cautious in comparing the results of cross-sectional papers with that of the time-
series ones.
5
In other words, MS models allow having multiple endogenously determined structural breaks in the data.
6
These four types of crime cover 92% of the total crime in the US in 2004.
7
As stated by Raphael and Winter-Ebmer (2001) and Grant and Martinez (1997), although robbery is considered a violent offence, in fact, it is motivated
primarily by economic need; therefore, it would be better to study that along with other property crimes.
 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345 35

The Markov-Switching method is used to examine the behavior of the unemployment rate and property crime rates. The
ndings show that there is a positive but insignicant association between the unemployment rate, burglary, larceny, and
robbery. However, the unemployment rate has a signicant relationship with motor-vehicle theft. The contemporaneous
relationship between the rate of unemployment and motor-vehicle theft is negative (which supports the existence of oppor-
tunity effect), which implies that less motor-vehicle theft has been reported during recessions when the unemployment rate
is higher. The results from the correlations between the lagged cycles of unemployment and crime indicate that as time goes
by, the relative size of motivation effect and opportunity effect changes.
To shed more light on the stability of the link between unemployment and property crimes, we divide the sample into
three subsamples. The rst subsample covers data from the rst quarter in 1990 through the last quarter of 2004, when
the crimes in the US are known to have declined. The second subsample locates recession phases in the US and the third
subsample consists of the observations coinciding with the period of normal state in the US economy. The ndings of sub-
sample studies show that the links between unemployment and property crimes are higher for the rst subsample (i.e., the
period from 1990 to 2004) and the third subsample when the US economy was at a normal state. On the other hand, during
the recession periods (the second subsample), the link is much weaker than the links found for the whole sample and the
other two subsamples. In addition, we found that during the period from 1990 to 2004, there was a positive and signicant
relationship between the unemployment rate and robbery.
The remaining portion of this paper is organized as follows. The second section describes the methodology. The third sec-
tion shows the results obtained, and the fourth section deals with synchronization of cycles using the nonparametric Con-
cordance Index of Harding and Pagan (2006), to check the relationship between the cycles of the unemployment rate and
crime variables. Lastly, the fth section gives the conclusion.

2. Methodology

Let yt denote the variable of interest in period t. Different methods can be used to study the behavior of a series; however,
as stated by Hamilton (1994), [m]any variables undergo episodes in which the behavior of the series seems to change quite
dramatically. These changes (breaks) could happen because of various reasons, such as nancial panics, wars, economic cri-
ses, and regulation changes. They can create different regimes or states for the variable. To take these changes into account,
we have different approaches to follow. It is possible to estimate a model for the period before break and another model for
the periods after the break. However, the exact date of break is not known most of the time. Markov switching approach
provides us with a solution. Using this approach, we can estimate a different model for each regime without a priori knowl-
edge about the date of breaks or regime changes.
Following Hamilton (1989), we may propose the following model for the variable yt :
yt  lst /1 yt1  lst1      /k ytk  lst4 t ; 1

where l shows the mean of the variable yt ; st is an unobserved variable indicating the state of the variable yt , and
t  i:i:d. N0; r2 . A full description of the dynamics of yt could be obtained if we have a probabilistic description of how
the variable changes from one regime to another. Such simplest model is a Markov chain,8 which is a model where
Prst jjst1 i; st2 j; . . . ; yt1 ; yt2 ; . . . Prst jjst1 i pij ;
PM
with j1 pij 1; 8i; j 2 f1; 2; . . . ; Mg. In this model, we have M different possible states or regimes for the variable, and the
variable yt moves between these states. Therefore, we have a matrix of transition probabilities
2 3
p11 p12 ... p1M
6p p22 ... p2M 7
6 21 7
P6 7;
4 ... ... ... ... 5
pM1 pM2 . . . PMM
where pij shows the probability of moving from regime i to the regime j.
In the model (1), only the value of the mean is regime-dependent. It is denoted as MSM(2)-AR(k), which indicates a Mar-
kov switching autoregressive model (MS-AR) with regime-dependent mean, two regimes, and an autoregression of order k.
Other specications are available; for example, a model where the mean, variance, and autoregressive coefcients are regime
dependent is denoted by MSMAH(m)-AR(k), where m indicates the number of states (regimes) and k reects the order of
the autoregression. In some cases, instead of the mean, the intercept is considered as regime-dependent to obtain an

8
In the rst-order Markov switching approach, the state of the variable at time t 1 depends on the state of the variable at time t. Therefore, having
information about the present period, we can discuss about the state of the variable at the next period. Sometimes the variable is deterministic, such as the
trafc light that has three lights or states: Green, yellow, and red. If the light is green now, we know that it will turn yellow, and if it is yellow now, we know
that it will turn red. Therefore, knowing the present light or state reveals the next state with certainty.The situation gets complicated when the variable is a
stochastic process. Take weather conditions as an example. Suppose it could be rainy, cloudy, or sunny. Then, knowing that it is sunny today, we cannot be
certain about tomorrows weather condition. We only can assign probabilities on the potential conditions for tomorrows weather. For example, we may say
that the probability that it is sunny the following day is 80%, the probability that it rains the following day is 3%, and the probability of having a cloudy day is
17%. One way to calculate these probabilities is the Markov switching (chains) approach that is a branch of probability theory.
36 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345

MSIAH(m)-AR(k). The models where the mean and intercept are regime-dependent are not equivalent. They imply different
dynamics of adjustment of the variables after a change in regime.
Once the coefcients of the model are estimated and a transition matrix is calculated, one can calculate the probability of
being in state j at each period, based on the information of the whole sample. This series of probabilities is known as smoothed
probabilities. In addition, one can calculate the probability of being at state j at each time only based on the information up to
that date (not the whole sample), and the corresponding series of probabilities is known as ltered probabilities.

3. Empirical results of the MS-AR models

Quarterly data of four property crime series (Burglary, Larceny, Motor-Vehicle Theft, and Robbery) for the US were used.9
The data were obtained from the Uniform Crime Reports of the Federal Bureau of Investigation (FBI) and they are national aggre-
gates. These data cover the period from 1977:2 to 2004:4, and show the number of each crime per 100,000 of population.10 The
data for the quarterly unemployment rates were taken from the Bureau of Labor Statistics.
The data were seasonally nonadjusted. We used the TRAMO/SEATS procedure of Gmez and Maravall (1992) to remove
the seasonality and detect the presence of outliers in the series.
The presence of unit roots was formally tested using the GLS-based Augmented Dickey-Fuller (ADF GLS ) statistic of Elliott
et al. (1996), and the feasible point optimal statistic suggested by Elliott et al. (1996). The null hypothesis of a unit root was
not rejected for any of the crime series, indicating that all crime time series are I(1) processes. On the other hand, the unem-
ployment rate appeared to be stationary, that is, I(0).
Given the previous results of the unit root tests, the property crime series were modeled in rst differences11 and the
unemployment rate entered in levels. A generalized version of the model suggested by Hamilton (1989) was estimated. All esti-
mations were carried out with the MSVAR class for Ox (see Krolzig, 2005).
Before estimating the MS models, a linear model was specied. The lag length was determined using information criteria
such as the Akaike information criterion (AIC), the Schwarz information criterion (SIC), and the HannanQuinn information
criterion (HQ). Using the selected lag structure, MS models with two states were estimated, allowing for changes in the inter-
cept (I), variance (H), and autoregressive parameters (A). Based on the information criteria and the LR (likelihood ratio) sta-
tistic, we selected the best tting MS models. The null hypothesis of linearity was rejected for all selected models.12

3.1. Burglary

For this time series, the MSIAH(2)-AR(4) was selected as the best model, and Table 1 shows the results for this model. The
p-value of the Daviess statistic for the null hypothesis of linearity was 0.004, which showed that for this time series, the null
of linearity should be rejected in favor of the MS model. Overall, the results showed that in the rst regime, a decrease in the
rate of burglary was present, and in the second regime, this rate increased. Therefore, the rst regime stands for lower and
second regime stands for higher burglary rates in the US.
Based on the transition probabilities, it is clear that the rst regime is persistent with a probability of 84%. In addition, the
unconditional probabilities of being at the lower and higher burglary rate regimes are 79% and 21%, respectively. The transition
probabilities revealed that the duration of the rst regime is 6.09 quarters, whereas the second regime has the duration of 1.59
quarters. This indicates the presence of asymmetries in the duration of the regimes of lower and higher burglary rates, intend-
ing that the periods of reduction in the rate of burglary are much longer than those with positive change at this rate.
The results also showed that 79% of the observations for burglary were categorized in the rst regime, indicating that
between the second quarter of 1977 and the end of 2004, in 88 quarters, the rate of burglary was declining. Concerning
the standard deviations, the model allowed for different variances for each regime. In fact, the results showed that the sec-
ond regime was more volatile than the rst one (standard deviation of 7.77, when compared with 5.50).
The smoothed probabilities and the ltered probabilities of being at each regime at each point of time are presented in
Fig. 1. This shows that the rst regime (lower burglary rates) was formed by the periods 1977:21979:3, 1980:31980:4,
1981:31984:3, 1985:11985:4, 1986:21986:4, 1987:21988:3, 1989:21991:3, 1992:21992:3, and 1993:22004:4.
The same gure also shows the observations since 1993:2 until the end of the sample form the rst regime (decreasing bur-
glary rates).13

3.2. Larceny

All information criteria indicated that the preferred model is the MSIA(2)-AR(1). Table 1 presents the results for this mod-
el. The results showed that the rst regime has a negative intercept, while the intercept of the second regime is positive and

9
It should be mentioned that this study used quarterly aggregate data, while several studies have used annual aggregate data.
10
Notice that the rates are reported property crime rates based on crimes reported to the police, not those reported in crime victimization surveys. We would
like to thank an anonymous referee for pointing this out.
11
It is worth noting that as we are using the rst difference of the property crime rates and as differenced data show the short-term uctuations only, the
results should not be used to judge about long-run relationships.
12
Using LR test and Davies (1987) bounded LR test.
13
This nding is in line with the fact that the index crimes in the US have been declining since the 1990s.
 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345 37

Table 1
Results of univariate MS-AR models for crimes.

Burglary Robbery Larceny Motor-Vehicle theft

MSIAH(2)-AR(4) MSI(2)-AR(3) MSIA(2)-AR(1) MSI(2)-AR(1)
Intercepts
l1 2.112 (2.607) 0.922 (4.484) 4.221 (1.818) 1.212 (2.281)
l2 5.352 (2.251) 1.714 (4.857) 34.687 (5.920) 2.534 (3.253)
Autoregressive parameters
/11 0.231 (2.738) 0.248 (2.737) 0.056 (0.558) 0.197 (1.944)
/21 0.050 (0.644) 0.055 (0.619)
/31 0.083 (0.859) 0.143 (1.758)
/41 0.201 (2.207)
/12 0.853 (3.349) 1.092 (5.589)
/22 1.409 (3.765)
/32 0.016 (0.065)
/42 0.046 (0.224)
Standard errors
r1 5.504 1.418 14.909 3.279
r2 7.774 1.418 14.909 3.279
Transition probabilities
p11 0.836 0.955 0.917 0.974
p22 0.370 0.882 0.360 0.923
Durations
Regime 1 6.09 22.11 12.03 38.31
Regime 2 1.59 8.49 1.56 13.02
Likelihood ratio linearity tests probability
Davies 0.004 0.016 0.002 0.028

t-Statistics are in parentheses.

MSIAH(2)-AR(4), 1977 (2) - 2004 (4)

DBurglary
25

1980 1985 1990 1995 2000 2005

Probabilities of Regime 1
1.0

0.5 filtered smoothed

1980 1985 1990 1995 2000 2005

Probabilities of Regime 2
1.0
filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Fig. 1. The series, smoothed, and ltered probabilities of Burglary; MSIAH(2)-AR(4) model.

both of them are signicant. According to these results, the rst regime stands for the periods with decreasing rates of lar-
ceny and the second regime shows the periods with increasing rates of larceny.
The durations of the regimes were 12 and 11.56 quarters, respectively. From 111 observations included in the estimation,
the rst regime accounts for 88% of the observations (98 quarters), while the second regime accounts for 13 quarters. The
38 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345

transition probabilities indicated that the rst regime is more persistent than the second regime, with a probability of 92%,
when compared with 36%, respectively. This means that if there is a decreasing rate of larceny at any period, it will be dif-
cult to return to a regime with an increasing rate.
Fig. 2 shows the smoothed and ltered probabilities of larceny in the two regimes. According to these results, the second
regime (increasing rates of larceny) consists of 1979:31979:4, 1980:21980:2, 1986:11986:1, 1987:11987:1, 1989:1
1989:1, 1989:41989:4, 1991:41992:1, and 1994:11994:1. Notice that most of these periods last only for two quarters.
It can also be seen from Fig. 2 that from 1994 onwards, the larceny in the US was decreasing.

3.3. Motor-vehicle theft

For this variable, the AIC and the LR test suggested that the best model is an MSIAH(2)-AR(1). However, a close investi-
gation of this model revealed that it does not t the data well and it cannot capture the movements of this time series. At the
same time, the SIC suggested an MSI(2)-AR(1) as the best model, which matches the data and its movement well. Therefore,
this model was selected as the best model.
Table 1 shows the results for this model. The intercepts are negative and positive, respectively, for the rst and second
regimes. The rst regime shows the periods with negative changes at the rate of motor-vehicle theft, and the second regime
stands for the periods with positive changes of this rate. Both the regimes are very persistent, and there is a 3% probability of
moving from the rst (lower rates of motor-vehicle theft) to the second regime (higher rates of motor-vehicle theft). The
opposite direction has an 8% probability.
The results also indicated that 32% of the observations are classied in the higher rate regime. The expected duration of
this regime was 13 quarters, when compared with 38 for the rst regime. Fig. 3 shows the smoothed and ltered probabil-
ities of being at each regime. It shows that during 1980:11984:2 and from 1990:2 to the end of the sample, the variable was
at the rst regime, i.e., the rate of motor-vehicle theft in the US had been decreasing during these periods. For the rest of the
sample, the changes in motor-vehicle theft had been positive and the rate of motor-vehicle theft had been increasing.

3.4. Robbery

The SIC suggests the MSI(2)-AR(3) as the best model that captures the movements of robbery in the US The results from
the estimation of this model are shown in Table 1. The rst regime stands for all the periods with negative changes in the rate
of robbery (i.e., lower robbery rates), and the second regime shows all the periods with positive changes in this rate (i.e.,
higher rates of robbery).

MSIA(2)-AR(1), 1977 (2) - 2004 (4)

100
DLarceny
50

-50

1980 1985 1990 1995 2000 2005

Probabilities of Regime 1
1.0
filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Probabilities of Regime 2
1.0
filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Fig. 2. The series, smoothed, and ltered probabilities of Larceny; MSI(2)-AR(1) model.
 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345 39

MSI(2)-AR(1), 1977 (2) - 2004 (4)

20
DMotor Mean(DMotor)

10

-10

1980 1985 1990 1995 2000 2005

Probabilities of Regime 1
1.0 filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Probabilities of Regime 2
1.0 filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Fig. 3. The mean, smoothed, and ltered probabilities of Motor-Vehicle Theft; MSI(2)-AR(1) model.

The expected durations were 22 and 8.5 quarters, respectively, for the rst and second regimes. The lower robbery rates
regime included 70% of the observations (77.3 quarters). Transition probabilities indicated that both the regimes were per-
sistent, with probabilities equal to 95% and 88%, respectively, indicating that if there is a negative change (positive change) in
this rate in any period, then there is a high probability for a negative change (positive change) in the next period.
Fig. 4 shows the smoothed and ltered probabilities for both the regimes. The longest period with a decreasing rate of
robbery is 1992:12004:4, i.e., almost 13 years. The rest of the study period oscillates between the regimes.

3.5. Unemployment rate

Although the information criteria selected the MSI(2)-AR(3) as the best model, it cannot capture the complete character-
istics and movements of the series. The model that can t the data well and describe the changes in the unemployment rate
in the US is the MSIAH(2)-AR(3). The LR linearity test strongly supported the nonlinearity in this variable.
Table 2 shows the results of estimation of this model. This model allows for regime-dependent autoregressive parameters
and variance; hence, there are different estimations for the coefcients of lagged variables in regimes. The estimated stan-
dard deviations are 0.119 and 0.285, respectively, for regime 1 and 2. Therefore, the regime with higher unemployment rates
is more volatile than the rst regime. Regime 1 shows all the periods with low unemployment rate in the US, and the second
regime stands for all the periods with higher unemployment rate. Therefore, the rst and second regimes show the normal
and recession periods in the US economy.
The expected duration of low unemployment is almost three times longer than that of higher rates of unemployment.
During the period of study, 79 quarters form the regime of low unemployment rate and the rest (32 quarters) are classied
as the second regime (higher unemployment). The rst regime is more persistent than the second regime based on the tran-
sition probabilities, with probabilities of 94% and 84%, respectively, to remain at the same state in the next period.
Fig. 5 shows the smoothed and ltered probabilities for both the regimes. The only periods with high unemployment
rates are 1979:41984:3, 1990:31992:1, and 2001:12001:4. These periods almost perfectly match the recession periods
in US. During the rest of the study period, the unemployment rate in the US was at a lower regime.

4. Synchronization of cycles

So far, the MS models for each single time series were estimated and their properties, such as expected durations, tran-
sition probabilities, and asymmetries, have been investigated individually. One of the questions of interest in this study is the
existence of any relationship between the unemployment rate (as an indicator of the position of the economy) and the prop-
erty crime variables. To answer this question, one can test whether the timing of the cycles of the unemployment rate and
40 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345

MSI(2)-AR(3), 1977 (2) - 2004 (4)

5 DRobbery Mean(DRobbery)

1980 1985 1990 1995 2000 2005

Probabilities of Regime 1
1.0
filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Probabilities of Regime 2
1.0
filtered smoothed

0.5

1980 1985 1990 1995 2000 2005

Fig. 4. The mean, smoothed, and ltered probabilities of Robbery; MSI(2)-AR(3) model.

Table 2
Results of MSIAH(2)AR(3) for the unemployment rate.

Regime 1 Regime 2
l 0.139 (1.085) 0.607 (2.271)
/1 1.189 (9.757) 1.613 (8.630)
/2 0.056 (0.302) 0.730 (2.221)
/3 0.166 (1.500) 0.036 (0.192)
Standard errors 0.119 0.285
Duration 16.48 6.14
Likelihood ratio linearity tests probability
Davies 0.000
Transition probabilities
Regime 1 Regime 2
Regime 1 0.939 0.061
Regime 2 0.163 0.837

t-Statistics are in parentheses.

property crimes under study in this research is similar or not. In other words, one can check for synchronization of cycles;
that is, to see whether higher unemployment rate periods are correlated with higher property crime rates regimes or not.
The unemployment rate will be considered as the reference variable and its cycles as the reference cycle. These cycles will
be compared with those of the property crime series to test for common cycles. For this purpose, the Concordance Index (CI),
proposed originally by Harding and Pagan (2006), has been used.
The CI is a nonparametric statistic that shows the proportion of time at which two series are in the same regime. For two
series xt and yt for t 1; 2; . . . ; T, let Sxt Syt be a dummy variable that takes the value of unity when xt yt is in the rst re-
gime and the value of zero when it is not in the rst regime. Then, the CI for xt and yt is given by the following expression:
( )
1
XT X
T
CI T Sxt Syt 1  Sxt 1  Syt : 3
t1 t1
 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345 41

10.0
Unemployment

7.5

5.0

1980 1985 1990 1995 2000 2005

Probabilities of Regime 1
1.0

0.5
filtered smoothed

1980 1985 1990 1995 2000 2005

Probabilities of Regime 2
1.0

0.5

1980 1985 1990 1995 2000 2005

Fig. 5. The series, smoothed, and ltered probabilities of the Unemployment Rate; MSIAH(2)-AR(3).

For example, a value of 0.8 for this index means that two series (xt and yt ) are in the same regime 80% of the time. As CI is
dened as the proportion of time that two series are in the same state, this index is bounded between zero and unity. The CI
index has a value of unity when Sxt Syt and a value of zero when Syt 1  Sxt . These two series are called pro-cyclical if
CI 1. They are counter cyclical if CI 0.
It is natural to state that having a high concordance index means high common cycle. However, the question is how high
should it be to interpret that as pro-cyclical? Even for two unrelated series, the expected value of the concordance index may
be 0.5 or higher.14
The above-mentioned formula for concordance index can be written in a different way as follows:
X
T
CI 1 2T 1 Sxt Syt  lSx  lSy 1 2qS rSx rSy 2lSx lSy  lSx  lSy ; 4
t1

where qS is the estimated correlation coefcient between Sxt and Syt . If Sxt Syt or Syt 1  Sxt , then rSx rSy r2Sx ; hence,
the value of unity for this index corresponds to qS 1 and value of zero to qS 0. Therefore, qS 1 (qS 1) shows that
two cycles are perfectly positively (negatively) synchronized, and they are unsynchronized when qS 0.
Assuming that the two series are statistically uncorrelated (qS 0, the expected value of this index will be:
ECI 1 2lSx lSy  lSx  lSy : 5

The expected value of being at each regime can be measured by dividing the number of periods at that regime by T. Now, this
expected value can be compared with the calculated value from the series. If the former is smaller than the latter, one can
state that there is a link between the cycles. This indicates that the number of periods where the series are in the same state
is higher than that if they were uncorrelated. If the former is larger than the latter, one can conclude that these series are
counter-cyclical. However, the signicance of this result has to be checked to see whether the ratio of these two is statisti-
cally different from one or not.
Another problem that exists in using this index is that it depends on the expected values of Sxt and Syt , that is, their mean.
Suppose that the mean of Sxt and Syt is 0.5 and these two series are unsynchronized; then, the expected value of the concor-
dance index will be 0.5, which conrms the assumption that they are unrelated. However, if the regime that takes the value
of one has higher duration than the other, the mean values of the series will be higher than 0.5. Now, assume that the means
are 0.8; therefore, the expected concordance index will be 0.68, which is higher than 0.5, and one may think that these two
series have common cycle although they are not related. Therefore, the mean value of the series has to be taken into account.
For this purpose, the mean corrected concordance index (Artis et al., 2004) is considered.

14
For example, consider tossing two fair coins; the probability that both coins are in the same state (either heads or tails) is 0.5.
42 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345

Table 3
Concordance Index (CI) between the unemployment rate and property crimes.

Unemployment Unemployment Unemployment rate & Unemployment rate &

rate & Burglary rate & Larceny Motor-Vehicle theft Robbery
CI 0.712 0.685 0.459 0.595
E(CI) 0.658 0.670 0.582 0.578
q
^ 0.178 0.055 0.293 0.039
t-ratio 1.218 0.474 2.545 0.231
CIcorr 0.053 0.015 0.123 0.016

PT
Let Sx T 1 t1 Sxt indicate the estimated probability of being at the rst regime. Then, the mean corrected concordance
index will be

X
T
CIcorr 2T 1 Sxt  Sx Syt  Sy : 6
t1

As mentioned earlier, one of the shortcomings of the concordance index is that it does not allow for a statistical testing of the
result. Harding and Pagan (2006) suggested that one could use a regression model to deal with this problem. To do so, the
following regression can be used:

r1 1
sy Syt a1 qS rSx Sxt ut : 7

Now, the hypothesis that qS 0 can be tested using the t-ratio of the coefcient of r1 Sx Sxt . In this regression, when the null
hypothesis is true, the error term inherits the serial correlation properties of Syt . In addition, Syt is strongly serially correlated;
hence, robust estimated standard errors have to be used (such as HAC Newey-West method). In addition, as Harding and
Pagan (2006), stated CI is monotonic in qS ; therefore, we can shift our attention from CI to the qS obtained from the estima-
tion of (7).
Table 3 shows the calculated CI, E(CI), and q for each pair of variables using the regime classications based on the MS
models in Section 3. The estimated concordance of 0.712 for the burglary and unemployment rate indicates that burglary
and unemployment are in the same state or regime (i.e., rising or declining together) in 71.2% of time. However, if they were
independent, then they had to be in the same regime only at 65.8% of time. Thus, the results show that unemployment and
burglary rates are linked to each other. In addition, as the ratio of CI to E(CI) is higher than 1, the unemployment rate in the
US and burglary are pro-cyclical, which means that burglary increases when the economy is in recession. The same is true for
CI
larceny and robbery. However, for motor-vehicle theft, ECI is lower than unity, which shows that the unemployment rate
and motor-vehicle theft are counter-cyclical.
Also, the estimated correlation between the unemployment rate and burglary, larceny, and robbery is positive; however,
it is negative between the unemployment rate and motor-vehicle theft. This shows that the relationship between the unem-
ployment rate and burglary, larceny, and robbery is pro-cyclical. However, this relationship is counter-cyclical between the
unemployment rate and motor-vehicle theft. According to the robust t-ratios reported in Table 3, only the results for motor-
vehicle theft is statistically signicant at 5%; hence, only one of the property crimes is signicantly contemporaneously con-
cordant with the unemployment rate cycle.
In summary, we found no contemporaneous signicant relationship between the unemployment rate and the following
crimes: burglary, larceny, and robbery. However, there is a negative contemporaneous relationship between the unemploy-
ment rate and motor-vehicle theft in the US that implies that less motor-vehicle theft has been reported during recessions,
when the unemployment rate is higher.15
The previous ndings were based on the assumption that the relationship between crimes and unemployment rate, if
there is any, is contemporaneous. We next calculated the concordance index of the lagged cycles of the unemployment rate
and current cycle of the property crimes to examine the presence of the motivation effect. In doing so, we assumed that
changes in the regime of unemployment rate at t  i may be related to changes in the regime of crime series at time t, where
i shows the number of lagged periods.
The ndings were the same as those earlier for the relationship between the rst and second lags of unemployment rate
cycles, burglary, larceny, and motor-vehicle theft (see Table 4). However, the results changed for robbery, and the correlation
between the cycles of unemployment rate and those of robbery showed a negative relationship. Nevertheless, as these cor-
relations were insignicant even at 10%, we may conclude that there is no signicant link between unemployment and rob-
bery. When more lags were introduced to the model, the sign of the correlation between the lagged cycles of unemployment
and each one of the property crime types changed. Although these correlations were insignicant, they showed that, as time
goes by, the relative size of motivation effect and opportunity effect changes.16 For example, while motivation effect was high-

15
This could be interpreted as evidence that supports the existence of the opportunity effect, as introduced by Cantor and Land (1985) and Phillips and Land
(2012).
16
As we have used quarterly data, what we are talking about here is the relative contemporaneity of a crime opportunity mechanism (Cantor and Land, 1987).
 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345 43

Table 4
The estimated correlations between crime and the lagged unemployment rate.

Unemployment Unemployment Unemployment rate & Unemployment

rate & Burglary rate & Larceny Motor-Vehicle theft rate & Robbery
1 lag 0.176 0.053 0.286 0.038
t-ratio 1.468 0.513 2.347 0.254
2 lag 0.114 0.052 0.235 0.042
t-ratio 1.067 0.520 1.814 0.311
3 lag 0.112 0.017 0.183 0.077
t-ratio 0.994 0.163 1.316 0.585
4 lag 0.171 0.020 0.130 0.112
t-ratio 1.297 0.181 0.878 0.805
5 lag 0.169 0.022 0.077 0.103
t-ratio 1.246 0.193 0.494 0.705
6 lag 0.106 0.024 0.024 0.094
t-ratio 1.004 0.274 0.145 0.592
7 lag 0.019 0.095 0.031 0.041
t-ratio 0.178 1.030 0.183 0.245
8 lag 0.027 0.029 0.086 0.013
t-ratio 0.205 0.293 0.497 0.076

Notice that, the results with no lag (contemporaneous results) are reported in Table 3.

Signicant correlations at 10% or lower.

Table 5
Synchronizations between the unemployment rate and property crimes in different subsamples.

Unemployment Unemployment Unemployment rate & Unemployment

rate & Burglary rate & Larceny Motor-Vehicle theft rate & Robbery
CI 0.817 0.833 0.800 0.883
Subsample from 1990:1 to 2004:4 E(CI) 0.774 0.785 0.806 0.732
q 0.219 (0.944) 0.287 (1.086) 0.062 (0.981) 0.574 (2.152)
The 1st Regime of Unemployment CI 0.911 0.911 0.608 0.696
The 2nd Regime of Unemployment CI 0.219 0.125 0.094 0.344

t-Statistics are shown in parenthesis. The 1st and 2nd Regimes of Unemployment corresponds to the normal conditions and recession in the US economy,
respectively.

er for larceny up to three lags of unemployment cycle, the opportunity effect was higher than the motivation effect for lags high-
er than three. The same was true for motor-vehicle theft; the ndings showed that the motivation effect existed for motor-vehi-
cle theft; however, it was lower than the opportunity effect. Consequently, the contemporaneous effect, which is the sum of the
motivation and opportunity effects, was negative.
To shed more light on the pattern and stability of the relationship between unemployment rate and property crimes, we
divided the sample into three subsamples and checked the consistency of the results obtained across the subsamples. The
rst subsample covered data from the rst quarter in 1990 through the last quarter of 2004, when the crimes in the US
are known to be declining. The second subsample located recession phases in the US, as identied in Section 3.5 and referred
to as the second regime of the unemployment rate. The third subsample consisted of the observations coinciding with the
period of normal state in the US economy, as denoted by regime 1 in Section 3.5. Based on the results from these three subs-
amples (reported in Table 5), the link between the unemployment and property crimes were higher for the period 1990:1
2004:4. During this period, the results showed a signicant and positive association between unemployment and robbery. In
addition, the calculated CIs in the period corresponding to the normal state of the US economy, when the unemployment rate
was low, showed a much higher concordance between unemployment and property crimes, when compared with the whole
sample. However, during recessions, the second regime of the unemployment rate, this link was weaker than the relation-
ships found for the whole sample and the other two subsamples. Therefore, the state of the economy does affect the rela-
tionship between the variables, and the researchers who want to study the economics of crime must take the state of the
economy into consideration.

5. Conclusions

Although economic theory anticipates the existence of a positive relationship between unemployment rate and crime
rates, empirical works have found mixed results. For example, Chiricos (1987) reviewed 68 studies about the relationship
between crime and the unemployment rate, and found that only less than half of these studies have found positive signif-
icant effects of unemployment on crime rates. That is, most of these studies have shown a negative or no relationship be-
tween crime and unemployment.
44 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345

In this study, we used quarterly data from the US economy over the period of 1977:22004:4 to examine the link between
the unemployment rate, as an indicator of economic condition, and four types of property crime. We applied Markov-Switch-
ing models to the unemployment rate and to the rst difference of the four types of property crime to identify downturns
and upturns for each of these variables. The estimated models for the data captured the movements in these variables. For
example, it is known that since the early 1990s, the crime rates dropped substantially in the US, and the estimated models
for the property crimes have captured this fact correctly. Next, the results of these MS-AR models, jointly with concordance
index, were used to test the existence of any relationships between the unemployment rate and property crimes in the US
economy. The ndings showed that there is a positive but insignicant relationship between the unemployment rate, bur-
glary, larceny, and robbery. However, the unemployment rate was found to have a signicant relationship with motor-vehi-
cle theft, and the contemporaneous relationship was negative for motor-vehicle theft. These ndings are in agreement with
those of Phillips and Land (2012) who used annual crime data for the US.
The counter-cyclical link between unemployment and motor-vehicle theft means that during recession, when the unem-
ployment rate is high, the motor-vehicle theft rate is low. One reason for this nding might be the fact that during recession,
more people stay at home and also they guard their properties. These ndings could be interpreted as a support for the
opportunity effect, as argued by Cantor and Land (1985) and Phillips and Land (2012). In other words, the ndings show that
the opportunity effect is larger than the motivation effect for motor-vehicle theft. However, for burglary, larceny, and rob-
bery, the size of motivation effect was found to be larger than the size of the opportunity effect; therefore, the results showed
a procyclical but insignicant link between the unemployment conditions and these three property crimes. Moreover, the
sign of the total effect changed over the time for all the property crime types, and it showed that the size of opportunity
and motivation effects do change by time.
To check the stability of the relationship between the unemployment rate and property crime, we divided the sample into
three subsamples: (1) the rst subsample covering a period from 1990:1 to 2004:4, (2) the second subsample covering the
recession periods, and (3) the last subsample covering the observations from the normal state of the economy, when the
unemployment rate is low. The results showed that the link between unemployment and property crimes increased when
we studied the rst and third subsamples. Furthermore, during the recession period, the co-movements between the unem-
ployment rate and property crime deteriorated.
Finally, it is worth noting that as we used the rst difference of the property crime rates, with the differenced data show-
ing the short-term uctuations only, the results cannot be used to judge the long-term relationship between unemployment
and property crimes. In other words, our results only explain the short-term relationship between the uctuation or changes
in the property crime rates and movements in the unemployment as an indicator of economic condition. Besides, as larceny
has a fairly low rate of reporting, the results for the larceny must be used cautiously.17 Lastly, the relationships found between
unemployment and property crimes only showed that these property crimes and unemployment rate are following similar co-
movements and that it does not necessarily show causality relationships.

References

Allen, R., 1996. Socioeconomic conditions and property crime: a comprehensive review and test of the professional literature. Am. J. Econ. Sociol. 551, 293
308.
Andersen, M.A., 2012. Unemployment and crime: a neighborhood level panel data approach. Soc. Sci. Res. 41 (6), 16151628.
Artis, M., Krolzig, H.M., Toror, J., 2004. The European business cycle. Oxford Econ. Pap. 56 (1), 144.
Arvanites, T.M., Dena, R.H., 2006. Business cycles and street crime. Criminology 44, 139164.
Becker, G.S., 1968. Crime and punishment: an economic approach. J. Polit. Econ. 76, 169217.
Britt, C.L., 1997. Reconsidering the unemployment and crime relationship: variation by age group and historical period. J. Quant. Criminol. 13 (4), 405428.
Cantor, D., Land, K.C., 1985. Unemployment and crime rates in the post-World War II United States: a theoretical and empirical analysis. Am. Sociol. Rev. 50
(3), 317332.
Cantor, D., and Land, K. C., Unemployment and crime rates: A bivariate seasonal ARIMA analysis of monthly data, 19691985. Unpublished paper, 1987.
Chiricos, T., 1987. Rates of crime and unemployment: an analysis of aggregate research evidence. Soc. Probl. 34, 187212.
Cook, P.J., 2010. Property crime-yes; violence-no. Comment on Lauritsen and Heimer. Criminol. Public Policy 9 (4), 693697.
Cook, P.J., Zarkin, G.A., 1985. Crime and the business cycle. J. Legal Stud. 14 (1), 115128.
Davies, R.B., 1987. Hypothesis testing when a nuisance parameter is only identied under the alternative. Biometrika 47, 3343.
Edmark, K., 2005. Unemployment and crime: is there a connection? Scand. J. Econ. 107 (2), 353373.
Elliott, G., Rothenberg, T.J., Stock, J.H., 1996. Efcient tests for an autoregressive unit root. Econometrica 64 (4), 813836.
Entorf, H., Spengler, H., 2000. Socioeconomic and demographic factors of crime in Germany evidence from panel data of the German states. Int. Rev. Law
Econ. 20 (1), 75106.
Fallahi, F., Pourtaghi, H., Rodrguez, G., 2012. The unemployment rate, unemployment volatility, and crime. Int. J. Soc. Econ. 39 (6), 440448.
Freeman, R.B., 1999. The economics of crime. In: Ashenfelter, O., Card, D. (Eds.), Handbook of Labor Economics, vol. 3. pp. 35293571.
Gmez, V., Maravall, A., 1992. Time Series Regression with Arima Noise and Missing Observations Program Tram, Working Paper 92/81. European
University Institute, Florence.
Grant II, Don S., Martinez Jr, Ramiro, 1997. Crime and the restructuring of the US economy: a reconsideration of the class linkages. Soc. Forces 75, 769799.
Greenberg, D., 2001. Time series analysis of crime rate. J. Quant. Criminol. 17 (4), 291327.
Hale, C., Sabbagh, D., 1991. Testing the relationship between unemployment and crime: a methodological comment and empirical analysis using time series
data from England and Wales. J. Res. Delinq. 28 (4), 400417.
Hamilton, J.D., 1989. A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica 57 (2), 357384.
Hamilton, J.D., 1994. Time Series Analysis. Princeton University Press, Princeton.
Harding, D., Pagan, A., 2006. Synchronization of cycles. J. Econom. 132 (1), 5979.

17
We are grateful to David Greenberg for pointing this out to us.
 F. Fallahi, G. Rodrguez / Social Science Research 45 (2014) 3345 45

Kollias, C., Paleologou, S.M., 2012. Labour market conditions and crime in Greece. Int. J. Soc. Econ. 39 (12), 9831000.
Krolzig, H.M., 2005. MSVAR 1.31k for Ox 3.4. <http://www.kent.ac.uk/economics/staff/hmk/index.html>.
Levitt, S.D., 2001. Alternative strategies for identifying the link between unemployment and crime. J. Quant. Criminol. 17 (4), 377390.
Lee, D., Holoviak, S.J., 2006. Unemployment and crime: an empirical investigation. AppliedEconomics Letters 13, 805810.
Phillips, J., Land, K.C., 2012. The link between unemployment and crime rate uctuations: an analysis at the county, state, and national levels. Soc. Sci. Res.
41 (3), 681694.
Raphael, S., Winter-Ebmer, R., 2001. Identifying the effect of unemployment on crime. J. Law Econ. 44 (1), 259283.
Witt, R., Clarke, A., Nigel, F., 1999. Crime and economic activity: a panel data approach. Br. J. Criminol. 39 (3), 391400.