Labor market opportunities and crime in Sweden

(1)

Master thesis 1, 15 ECTS

Labor market opportunities and crime in Sweden

The importance of individuals on the margin

Max Stenman Braarup

(2)

Acknowledgement

I would like to take the opportunity to thank my supervisor Katharina Jenderny at the department of Economics, Ume˚a University, for continuous support and guidance during the process of writing this thesis. Despite a shaky internet connection and a questionable microphone on my part, you managed to give valuable constructive criticism during this process.

(3)

Abstract

This thesis aims to examine the relationship between legal opportunities and crime rates in Swedish counties between 2000-2013. Even though this relationship has been estimated consistently in the last 10-15 years, this thesis contributes by studying the effects of legal opportunities for men with lower levels of education. This group is often over-represented in crime rates but has seldom been studied specifically in relation to crime. According to economic theory, less educated have a smaller opportunity cost of committing a crime compared to those with higher levels of education. This is confirmed by the results in this thesis, as increases in the unemployment level for non-college educated males have a larger effect on property crime rates than the population average unemployment level. As high levels of crime lead to large costs for a society, labor market- and educational programs may have indirect benefits from lower levels of crime. These benefits may be underestimated if not examining individuals on the margin between the legal- and illegal sectors.

(4)

1 Introduction

Since the beginning of the 21st century until mid 2010s the number of people with low level of education in Sweden reduced but the unemployment level increased (SCB, 2016)¹. Uneducated, men in particular, is a group often over-represented in crime participation (Freeman, 1996; Gould et al., 2002). The recent developments on the labor market for uneducated provides an excellent opportunity to analyze the relationship between labor market opportunities for uneducated men and the number of crimes in Swe- den. The interest in jointly analyzing crime and the labor market can be traced back over 60 years and the literature is continuing to grow (Mustard, 2010). The early influ- ential work within the field such as Becker (1968) assumed crime was mainly driven by potential benefits and costs in the illegal sector. Ehrlich (1973) developed this theory to include opportunities in the legal sector. Lately, improvements in the literature have allowed researchers to consistently measure a relationship between the labor market and crime. Part of the progress originate from inclusion of labor market variables that matter for individuals on the margin between the legal and illegal sectors (Mustard, 2010).

Sweden along with its Scandinavian neighbours has, regarding crime, been described as different from the rest of the countries in the EU. The reason being that Scandinavian countries are characterized by smaller prison populations and expertise present in the public debate. This is the case even though in Sweden criminology is one of the smallest social science disciplines (Estrada et al., 2012). Although Sweden is considered as pos- itively different, inhabitants are worried about crime as numbers show 40% are worried about the development of crime in society and 80% believe the crime rate has increased in the last three years (NCCP, 2019). Apart from a source of individual concern, crime is also costly. Individuals lose property and may have to spend money to retrieve it. Vic- tims of violence may be deprived of their income and need transfers to cover for it and

(6)

the public sector spends money to prevent and reduce crime (McCollister et al., 2010).

If bad labor market opportunities increase crime rates, there may be indirect costs and benefits associated with labor market policies that should be considered. An increasing unemployment level generates expenses through public transfers but may also lead to indirect expenses through an increase in the crime rates (Edmark, 2005). Recognizing this, labor market policies aimed to reduce the unemployment level or improve the labor market opportunities in general, perhaps through higher wages or lower unemployment level, may come with additional benefits not directly associated with the labor market.

The aim of this thesis is to examine how labor market opportunities for men with lower levels of education affect the number of crimes in Sweden, during a period categorized by a large variation in unemployment due to the economic crisis. It was early estab- lished in the literature that population averages such as the mean wage and the population unemployment level may fail to locate individuals on the margin between the legal and illegal sectors. However, other measures have seldom been applied in the literature (Mustard, 2010). One exception is Gould et. al (2002) who find that increasing the wage for non-college educated men and decreasing the unemployment level for the same group also reduces crime. Machin and Maghir (2004) find similar evidence using wages at the bottom end of the wage distribution. Also in Sweden, the use of population averages rather than averages for groups or individuals on the margin is common practice. Gr¨onqvist (2011) however, use individual data on youth unemployment and find a robust and positive effect of youth unemployment on crime. Controlling for the timing of crimes (weekend/weekdays), he finds that unemployment increases the time to engage in crime. Nordin and Alm´en (2017) find that long-term unemployment has better explanatory power than population unemployment.

This thesis contributes to the existing literature by examining labor market opportuni-

(7)

ties for men with different levels of education rather than opportunities for the average population. Studying groups potentially on the margin between the legal and illegal sector can increase the evidence of the relationship between the labor market and crime.

Further, learning to what extent labor market outlooks for uneducated men affect crime rates can produce insightful knowledge on how to design labor market policies. Not accounting for indirect costs and benefits stemming from crime when designing labor market programs may underestimate the total effects. Using data on Swedish counties over fourteen years this thesis show the effect of unemployment on property crime is larger when using the unemployment level by educational level compared to population averages. Further, income as a measure of labor market opportunities instead of wages is proven to be insufficient as the effect of income on different crimes are insignificant.

The thesis is outlined as follows. Section 2 review some of the existing literature about the relationship between the labor market and crime. Section 3 establishes a simple theoretical framework and section 4 outlines the econometric specification and some potential issues regarding the estimation. Section 5 describes the various data sources and section 6 presents the results which are discussed in section 7.

(8)

2 The connection between crime and the labor market

Early literature reviews concluded that there was a gap between theory and empirical evidence. Academics and policymakers agreed on the intuition that labor market opportunities affect criminal behaviour, but it was unsupported by the empirical work (Piehl, 1998; Mustard, 2010). Lately, a new generation of research has been able to consistently measure relationships between labor markets and crime due several developments. Con- sidering crime varies across relatively small areas, the possibility to use less aggregated data has been of vast importance to the development. The possibility to collect longitudinal data has provided the ability to control for unobserved differences across regions to reduce omitted variable bias. Further, work dedicated to theoretically and empirically identify groups on the margin between the legal and illegal sector and the opportunities this particular group is facing has led to consistent evidence on the relationship between labor markets and criminal behaviour. Lastly, realizing limits in the data and potential problems such as reverse causality has increased the ability to control for the potentially biased estimates earlier studies produced (Mustard, 2010).

Regarding empirical findings, Grogger (1998) discovers that young men respond to wage incentives and that the racial wage differential explains part of the difference in crime participation rates. Raphael and Winter-Ebmer (2001), using US state-level panel data over the period 1971-1997, find evidence that unemployment increases property crime rates. They cannot find the same evidence for violent crimes². Focusing on non-college educated men in the US, Gould et al. (2002)³ find wages and unemployment affect crime and that wages explain more of the changes in crime rates than unemployment.

Regarding estimates over their study period 1979-1997, they find the non-college wage

2This is a common result in the literature. Violent crime is unlike property crime, seldom motivated by economic reasons (Levit, 2004).

3This study was the first at its time to specifically focus on those, according to the authors, most likely to commit crimes.

(9)

explains 43% of the increase in property crime and the unemployment of non-college men explains 24% of the increase in property crime. A vast majority of the literature fo- cuses on US data, but Foug`ere et al. (2009) find similar effects compared to the studies above, using panel data collected on regions in France during the late ’90s. They observe that increased youth unemployment increases crime and understanding this, the authors argue, policies focusing on reducing youth unemployment can help combat crime. Us- ing UK-data, Machin and Maghir (2004) discovers that crime rates are higher in areas where there are poorer labor market opportunities with depressed wages at the bottom of the wage distribution.

Similar results hold for Swedish data. Edmark (2005) uses panel data on Swedish counties over the years 1988-1999 and estimates that a one-percent increase in unemployment increase aggregate property crime by 0.11 percent. Using municipality data over the period 1996-2000 Agell and Öster (2007) find unemployment affects burglary, auto theft and drug possession. Contradictory to Fougère et al. (2009) and Grönqvist (2011) they cannot find any evidence for the connection between youth unemployment and crime.

Nordin and Alm´en (2017) focus on violent crime and argue long-term unemployment can explain violent crimes better than the general unemployment level. Being left with- out a job for a long time potentially creates a tension leading to violent behaviour. They find evidence to support their theory and confirm the weak effect of general unemployment on violent crimes found in earlier studies.

(10)

2.1 Labor market opportunities for individuals on the margin

A common approach to connect labor markets with criminality is to examine the effect of the unemployment level. A measure like wages potentially better explain the scenery of potential criminals. Unemployment tends to be cyclical and might decrease while other labor market conditions worsen. For example, if individuals drop out of the labor force because of lower wages. Gould et al. (2002) find wages to posses more explanatory power than unemployment to explain crime rates. The same conclusion is drawn in Machin and Meghir (2004) as the association between wages and crime is more robust than that between unemployment and crime. According to the model they adopt where individuals may allocate time to both the legal and illegal sectors, this is not a surprising result. An increase in crime can stem from either less time spent at work or less leisure and in the latter case, unemployment plays no role. Ihlanfeldt (2006) study job opportunities for male youth, based on the jobs youth are qualified to hold within an area and the competition among workers for these jobs. He finds poor job opportunities in poor inner-city neighborhoods explain a large portion of the high crime rates in these areas. Aside from unemployment and wages directly, Lochner (2004) argues education is an important tool to prevent crime as higher education raises wages. This leads to better opportunities in the legal labor market and raises the opportunity cost of incarceration. Moreover, subsidies targeting schooling or job training is likely to reduce crime in the long-run as the skill level increases. Wage subsidies can affect crime in the short run but may discourage investment in human capital in the future and hence increase crime. As outlined above, there are many variables that in theory may affect crime but one has to carefully evaluate which of these variables are best suited to explain crime rates (Mustard, 2010).

(11)

2.2 The issue with population averages

Despite that Ehrlich (1973) discussed problems with using population averages such as the average wage, average income and general unemployment level, research has long failed to implement other measures. The issue arises because the population average likely will fail to identify those deciding whether to engage in the legal or illegal sector.

If individuals choosing between these sectors mainly are less educated and the potential wage they may receive on the labor market is at the bottom end of the distribution, using averages would often bias the estimates. If the wage increases for college educated, and nothing has happened with the wage for uneducated, the average wage will increase but this increase will not influence decisions for individuals that according to theory may commit crimes to a larger extent. The same issue arises when not examining men specifically. Men are over-represented in crime rates which mean averages driven by changes applying to women will fail to provide correct estimates of the effect on crime (Mustard, 2010).

3 Theoretical framework

To put the theory briefly discussed previously in words to a clear context, a simple model of the supply of crime is described below. The model presented follows the model developed in Edmark (2005) which is based on the models developed in Ehrlich (1973) and Freeman (1999). Before the model is derived, it should be noted that in the rest of the thesis, uneducated will sometimes be used and sometimes non-secondary or non-college educated. Uneducated will be used when a distinction between the specific levels of education is unnecessary and the specific levels of education will be mentioned when a distinction should be made.

(12)

3.1 Deriving the model

The individual is assumed to choose between work and illegal activity as a source of income. For simplicity, we assume time can only be spent in one sector. Let W be the potential wage for men in low-skilled occupations in the legal sector and Wc the potential benefit from crime. B is the unemployment benefit and U is the unemployment rate among uneducated men and can be thought of as the probability of an individual being unemployed in the period. Further, P denotes the probability of being caught and S the cost of punishment. This cost is subtracted from the benefit Wc, meaning the benefit is assumed to be kept if getting caught. The variables described above are assumed to be equal for all individuals. Lastly, let F_cbe the positive/negative psychological effect of committing a crime. This cost is assumed to be independently and continuo- https://www.overleaf.com/project/5e833335eedb06000109a6f5usly distributed over the population (Edmark, 2005).

We expect the individual to engage in the illegal sector if the expected potential benefit from crime minus the costs exceeds the potential wage received in the legal sector, i.e if inequality (1) is satisfied.

E (Wc) − Fc> E(W ) (1)

It follows from inequality (1) that an increase in the benefit from crime, a lower psychological cost of committing a crime, or a lower wage in the legal sector will increase the individuals’ tendency to commit crimes. The expected benefit from crime, E (W_c), depends on the probability of getting caught while committing a crime as well as the cost of punishment and the potential benefit from crime. It is defined as follows:

E (W_c) = (1 − P )W_c+ P (W_c− S) (2)

(13)

As stated earlier, letting the unemployment rate reflect the probability of being unemployed we can define the expected return from work as follows:

E(W ) = (1 − U )W + U B (3)

It is now straight forward to rewrite inequality (1) as:

F_c< ((1 − P ) (W_c) + P (W_c− S)) − ((1 − U )W + U B) (4)

Inequality (4) implies an individual will choose to commit a crime when the difference between expected income from crime and the expected income from honest work, exceeds the psychological cost of committing a crime (Edmark, 2005). The market sup- ply of crime depends on the average wage if the effect of an increase/decrease in the wage is the same for low/high wage individuals. If this is not the case the market supply curve will depend on characteristics of the wage distribution rather than the average wage (Tauchen, 2010). Suppose, for example, that the unemployment level for uneducated exceeds the average unemployment level and the wage level for uneducated is below the average wage, all else equal. In this case we may in theory underestimate the incentives to commit a crime as the RHS of (4) increases with a lower W and higher U (assuming W > B and U < 1 ).

3.2 Aggregating the model

The aggregate supply of crime can be derived from (4). At the aggregate level, a larger RHS of (4) ceteris paribus means more individuals satisfy the condition to choose the illegal sector which in turn increases the aggregate supply of crime. As (4) is increasing in U and decreasing in W and B, a higher unemployment level is expected to increase the supply of crime and a higher wage and higher unemployment benefit are expected

(14)

to reduce the supply. To establish an equilibrium the demand side of crime also has to be considered. However strange it might sound, the demand for crime represent a downward sloping relationship between the number of crimes and criminal earnings, W_c. If the number of crimes increases, more time and money is spent on self-protection. All else equal, this increases the costs of committing a crime and hence decreases the net benefit⁴. The demand can also be thought of as downward sloping because of a finite number of criminal targets. Escalating crime levels implies more competition among criminals which leads to a decreasing marginal benefit of committing a crime. The most valuable goods and individuals are targeted first (Ehrlich, 1996). For the latter interpre- tation in particular, income may have an unclear effect as an increase in income may both reduce the supply of crime and increase the demand as there is a larger supply of valuable goods. This will also have implications for the effect of unemployment. A higher level of unemployment should decrease the average income in a region and hence an indirect negative effect on the demand for crime. Controlling for the average income can thus allow us to estimate an effect of unemployment not influenced by the indirect effect on demand (Edmark, 2005).

4The same argument can be applied to violent crimes. As these levels increase, individuals and the public sector increase crime prevention and perhaps sanctions, which decreases the benefit from crime through higher costs (Freeman, 1999).

(15)

4 Econometric specification

From the theoretical model of the supply of crime, the following time- and fixed effects model can be estimated. Using fixed effects allows control for omitted variables vary- ing across the counties but not over time. The yearly time dummies allows control for aggregate time trends in reporting of crime (Gould et al., 2002). Log-linear or log-log specifications are the most common in the literature (Edmark, 2005). The model below is in log-log form and is given as follows:

ln C_it= α_i+ δ_t+ β₁ln u_it+ β₂ln w_it+ β₃ln p_it+ θ ln X_it+ ε_it (5)

i and t indicate county and time respectively and α_iand δtdenote county and time fixed effects. C_it is the number of reported crimes per 100.000 inhabitants, u_it is the unemployment level for uneducated men, witis the average disposable income by educational level for men, p_itis the clear-up rate for the type of crime displayed in C_it(property and violenece) The clear-up rate is typically endogenous which leads to a downward bias on the coefficient⁵. This, because the denominator of the clear-up rate also appears as the numerator of the dependent variable (Gould et al., 2002). In addition, the clear-up rate can possibly be endogenous for another reason. If the technology⁶available to the police to clear up crimes is a function of the number of reported crimes endogeneity is present.

Including the variable does not consistently indicate evidence of this phenomenon but it is, in general, difficult to control for the potential issue (Levitt, 1998). Handling this is not possible within the scope of this thesis and for this reason, results will be presented with the clear-up rate both included and excluded. Xitis a vector of controls. The mean unemployment benefit B is not included in the specification since it is included in the

5When the number of reported crimes are high (LHS) the clear-up rate is low (RHS), ceteris paribus (Levitt, 1998).

6The technology can be defined as the police force ability to convert reported crimes to arrests or equivalent (Levitt, 1998).

(16)

income variable. An apparent issue regarding the connection between the theoretical model and the empirical counterpart is the omission, due to data limitations, of the benefit from crime W_c, the cost of punishment S and psychological cost F_c. As long as they are fixed locally over time or constant across regions but vary over time, they are controlled for through the fixed and time effects.

Social and demographic variables are frequently included to control for factors affecting crime regardless of the labor market situation (Edmark, 2005). Alcohol consumption varies procyclically and affects crime rates independently (Raphael and Winter-Ebmer, 2001) and is therefore included. The share of young men is included as young men are over-represented in crime rates (Gr¨onqvist, 2011) and should for this reason be controlled for. Population density is added as more densely populated areas often experience higher crime rates (Glaser and Sacerdote, 1999). The share of immigrants is included as immigrants tend to be over-represented in crime rates (Gr¨onqvist, 2011). The average income in the regions will be included in an attempt to proxy for the potential illegal benefit and hence the demand side for crime. The goal is to secure that the effect of unemployment is driven by the supply side. Unemployment by educational level should not be affected by the demand side as much as the population average but average income is still included to be certain. Further, average income can control for areas with higher levels of self-protection as wealthier households should spend more to protect their property (Gould et al., 2002). More self-protection should decrease crime rates as the net-benefit from crime decreases. In summary, average income will have ambiguous effects on crime rates why it is difficult to predict what to expect in terms of estimates.

However, it is merely included as a control variable why the specific direction is not of particular interest.

(17)

4.1 Simultaneity bias and instrumental variable approach

If labor markets are affected by crime, ordinary least squares (OLS) estimates will be biased (Mustard, 2010). There are two primary channels where crime may affect the unemployment rates in a region. First, previously sentenced individuals may be unemploy- able which will impact the regional unemployment levels. Second, areas where crime rates are higher may discourage employers who choose to allocate elsewhere (Raphael and Winter-Ebmer, 2001). Higher crime rates may equally affect wages as employers need to pay higher wages to compensate workers for the risk (Gould et al., 2002). In the literature, this is often handled by instrumental variables (IV). Raphael and Winter- Ebmer (2001) find that the OLS estimates underestimate the effect of unemployment on crime rates as compared to the estimates obtained using instruments . Agell and ¨Oster (2007), who also estimates larger effects using instruments, conclude that OLS-estimates seems to set a lower bound for the effect of unemployment on crime. Gould et al. (2002) however, finds the OLS- and the IV-estimates to be similar which indicates that the OLS estimates on unemployment and wages are not driven by endogeneity. Finding a quality instrument can be a tough task. A weak instrument, for example, may lead to biased estimates and standard errors that are too small. Large estimates can thus be a sign of a weak instrument (Murray, 2006). Nordin and Alm´en (2017) conclude it is probable the IV-estimates are biased and IV instead can be viewed as a test of sensitivity. A way to handle potential endogeneity is to use somewhat larger areas as the problem is likely to be larger when smaller areas are being studied. At the county level, employers moving due to crime rates may still happen but is certainly less likely (Edmark, 2005). Gould et al. (2002) conclude state-level rather than county-level measures may solve the potential endogeneity issue. Based on the discussion above, IV is not used in this thesis. As the thesis is based on county-level data, it is less likely that the unemployment level and income is determined endogenously.

(18)

4.2 A possible sensitivity analysis

The use of county data comes with some advantages as previously discussed. There are, however, some issues regarding the choice of data. As discussed by Raphael and Winter- Ebmer (2001), county- and time dummies are not enough if there are county-specific time trends in the unobservables. Such variations could be county-specific availabil- ity of drugs and weapons. They suggest a sensitivity test where a linear and quadratic county-specific time trends are included in the baseline specification. The goal is to analyze whether the inclusion of time trends alters the results of the model and to conclude whether the baseline specification may be contaminated by omitted variable bias. The issue with this approach when having few observations, as Edmark (2005) points out, is the loss of degrees of freedom. If the estimates are altered by the inclusion of time trends, it is hard to conclude if it caused by omitted variables or a large loss in degrees of freedom. This particular sensitivity analysis will not be preformed in this thesis as the reasons behind potentially altered estimates are hard to evaluate. Further, the choice of county data to limit the possibility of endogeneity is believed to be more important than the above mentioned sensitivity analysis.

5 Data

The panel data set consists of annual data for all 21 Swedish counties over the years 2000-2013⁷. As mentioned in section 2, data on smaller regions was a crucial step to consistently estimate relationships between labor market variables and crime. The argument was crime can vary in relatively small regions which means national aggregates fail to capture variations in crime rates. However, areas such as the municipality- or city level may instead be inadequate as criminals may commit crimes or work in other areas

7NCCP changed its definition of regions from counties and municipalities to seven police regions after 2014

(19)

than their area of residence which will generate biased estimates. This is nevertheless possible at the county level but less likely. Further, the clear-up rates which can be used as a measure of the probability of being caught can only be collected at the county level.

Not including this could yield omitted variable bias according to the theoretical model (Edmark, 2005) .

Data on crime is collected from the Swedish National Council for Crime Prevention (NCCP) and consists of reported crimes per 100.000 inhabitants. Property crime includes robbery, burglary, theft, fraud and bike theft. Vehicle thefts are excluded due to a large decrease since the 1990s stemming from developments in vehicle security systems and hence not driven by factors on the labor market (Nordin & Alm´en, 2017).

Using data on reported crimes rather than data on convictions is a common approach for studies using aggregate data⁸. Not all crimes are reported to the police and the under- reporting is often correlated with the type of offense which leads to measurement errors depending on what crime is studied (Gould et al., 2002). This is indeed true for Sweden, as the number of reported crimes is often less than the number of actual crimes and the propensity to report crimes may vary over time due to changes in attitudes or in insurance policies. Nonetheless, reported crimes are assumed to provide a good approximation of the number of committed crimes (NCCP, 2019). For property crimes in particular, it is reasonable to assume the number of reported crimes to be close to the actual number considering reporting a crime is a condition to receive insurance compensation (Edmark, 2005). If the propensity to report crimes varies across regions but not over time, or if it is constant over regions but varies over time the fixed effects estimates are unbiased (Nordin & Alm´en, 2017).

8Data on actual convictions can be retrieved from NCCP, however, it is not clear if a crime committed in year t will lead to a conviction in year t, t+1or t+2, et cetera. If a crime committed one year leads to a conviction in future years, the data on crime committed year t will be incorrect. There may be a way to

(20)

The population average unemployment rate and the unemployment rate by gender and educational level is collected from the STATIV database at Statistics Sweden (SCB).

The educational levels are the previously mentioned non-secondary education and non- college education. Unemployment is defined as the share of the population (20-64)⁹that sometime during the year have been registered as ”openly unemployed” at the National Labor Market Board (AMS). The difference between this measure and other measures of unemployment is graphically displayed and discussed in Appendix B. Long-term unemployment is also collected by gender and educational level and included to compare the results to Nordin and Alm´en (2017). Long-term unemployment is defined as individuals under 25 years of age registered as unemployed at AMS for a period of at least 100 days and for those above 25 years of age that have been registered for at least 6 months.

The income variable is defined as the average disposable income by gender and educational level and is collected from SCB¹⁰. Disposable income includes, apart from labor income, capital income, social allowances and pensions. The hourly/monthly wage is likely a more adequate measure of labor market opportunities. Gould et al. (2002) use an estimate of the average weekly wage and Machin and Meghir (2004) use the hourly wage for example. This was unfortunately not possible in this thesis due to data limitations. When deciding whether to engage in the legal or illegal sector it is reasonable to assume one takes the hourly/monthly wage into consideration rather than the annual disposable income. This means the variable does not completely represent the opportunities on the labor market for an individual on the margin. However, the disposable income should nonetheless differ between uneducated and the average population in part

9Dividing by the labor force can cause variations in the unemployment level due to changes in the labor force (Foug`ere et al., 2009; Nordin & Alm´en, 2017). It was also necessary to retrieve the unemployment rate by educational level.

10SCB presents the data as the number of price base amounts, which is an annually adjusted variable in the Swedish tax- and transfer system. The value will be multiplied by the average price base amount over the period of study. Not multiplying by each years price base amount is an attempt to deflate the variable as the price base amount is calculated with inflation.

(21)

because of different wages. The effect on crime of an increase in the disposable income is hence assumed to be negative, as suggested by the theoretical framework, but due to the inability to differentiate the wage from other sources of income the effect should be weaker.

The clear-up rate is collected from NCCP and is defined as the share of reported crimes classified as as cleared. A crime classified as cleared does not mean a perpetrator has been connected to a specific crime. Instead, it means the police, prosecutor or custom border control has defined the crime as cleared up (NCCP, 2019). This does not directly translate to the risk of being caught and should instead be viewed as a measure of the resources of the police and justice system (Edmark, 2005). Using different measures of deterrence and crime prevention has been discussed thoroughly in the previous litera- ture. The most used is the arrest rate as a measure of the probability of sanctions and the incarceration rate as a measure of the probability of criminal justice actions (Tauchen, 2010).

Population density defined as population per km², the share of young men (15-24) and the share of immigrants is collected from SCB. The share of educated individuals, defined as those with a college degree of 3 years or more, is also collected from SCB.

The shares mentioned above are defined by dividing the number by the population (16- 64). Alcohol consumption is defined as litres of 100% alcohol purchased at the National Liquor Monopoly per capita and is collected from the Swedish Public Health Agency (PHA). For descriptive statistics, see Appendix C.

(22)

6 Results

The main results are displayed in Table 1 and Table 2 where column 1 includes general unemployment level, column 2 non-college unemployment level and column 3 non- secondary schooling unemployment level. In Table 2 the clear-up rate is excluded to examine how sensitive the results are to a potentially endogenous regressor. Table 3 displays the results of regressing a version of equation (5) against violent crime rates. Each table only includes the variables of interest. In Appendix A the full regression results are displayed and the controls briefly discussed.

6.1 Property crime

Table 1: Results of the baseline specification including clear-up rate.

Variables Property (1) Property (2) Property (3)

Income (non-college) -0.12

(0.35)

Income (non-secondary) 0.18

(0.24) Unemployment (pop. average) 0.15 ***

(0.06)

Unemployment (non-college) 0.27 **

(0.09)

Unemployment (non-secondary) 0.19 **

(0.06)

Clear-up rate -0.09 ***

(0.03)

0.09 ***

(0.03)

-0.09 **

(0.04)

Mean income -0.06

(0.13)

-0.04 (0.12)

-0.13 (0.12)

Control variables x x x

Observations 294 294 294

R2 0.21 0.23 0.18

Note: Cluster-robust standard errors in parentheses. *significant at 5%, **significant at 1% , *** signif- icant at 0.1%. All regressions include time and fixed effects. Controls described above are included in each regression but not displayed. For coefficients on all included variables, see Table A.1 in Appendix A. All regressions have been tested for zero fixed effects and this is rejected for (1)-(3).

(23)

The results in Table 1 shows the effect of the unemployment level on crime is higher when studied by educational level compared to the effect of the average population unemployment level. The effect is largest for the group with non-college education (Prop- erty (2)) and the point estimate imply a 1-percent increase in the unemployment level leads to an increase in property crime by 0.27 percent on average. For individuals with pre-secondary education, a 1-percent increase in the unemployment level is estimated to increase property crime by 0.19 percent on average. The point estimates on unemployment by educational level are not statistically significantly different, which can be noticed by studying their respective standard errors. Compared to the estimate of 0.15 percent on the general unemployment level however, the estimates of unemployment level by lower educational level are higher. The results are thus consistent with theory suggesting averages for individuals on the margin possess more explanatory power than population averages.

The income level coefficients are insignificant for all regressions implying the disposable income for non-secondary and non-college educated males has no effect on property crime. The same is true for the mean income level. The probability of being caught, as measured by the clear-up rate, is negative as expected and robust throughout each specification. Point estimates suggest a 1-percent increase in the clear-up rate decreases property crime by 0.09 percent but this coefficient is likely biased downwards due to issues discussed in section 4.

(24)

Table 2: Results of the baseline specification excluding clear-up rate.

Variables Property (1) Property (2) Property (3)

Income (non-college) -0.01

(0.36)

Income (non-secondary) 0.16

(0.24) Unemployment (pop. average) 0.16 ***

(0.06)

Unemployment (non-college) 0.28 **

(0.09)

Unemployment (non-secondary) 0.2 ***

(0.06)

Mean income -0.04

(0.12)

-0.04 (0.11)

-0.12 (0.11)

Control variables x x x

Observations 294 294 294

R2 0.17 0.18 0.14

Table 2 displays the results from a sensitivity analysis where the potential endogenous regressor clear-up rate is excluded. In general, the estimates look similar to the corre- sponding estimates in Table 1. The notable difference is the loss of R², that is, how much of the variation in the dependent variable the models can explain. The similarity between the coefficients in Table 1 and in the sensitivity analysis above is evidence against any apparent issue with the model specification when the clear-up rate is included.

(25)

6.2 Violent crime

Table 3: Results of regressing the baseline specification on violent crime rates.

Variables Violence (1) Violence (2) Violence (3) Violence (4)

Unemployment (pop. average) 0.11 (0.09)

Long-term unemployment (pop.average) 0.05 (0.05)

Long-term unemployment (non-college) 0.05

(0.04)

Long-term unemployment (non-secondary) 0.03

(0.03)

Clear-up rate -0.02

(0.06)

-0.02 (0.05)

-0.01 (0.05)

-0.02 (0.06)

Mean income -0.05

(0.18)

-0.07 (0.16)

-0.5 (0.15)

-0.9 (0.16)

Control variables x x x x

Observations 294 294 294 294

R2 0.13 0.12 0.13 0.11

In Table 3, the baseline model specification is regressed against violent crime. Income by educational level is excluded as this variable should in theory not affect violent crimes¹¹. Column 1 indicates violent crime is unaffected by the population average unemployment level which is supported by results in similar studies. Population average long- term unemployment had no effect as can be seen in columns 2 and columns 3-4 show that long-term unemployment by educational level did not yield significant results. The weak significance is in general an expected results as it is the common conclusion in similar research. However long-term unemployment, as found in Nordin and Alm´en (2017), should have more explanatory power.

11Test were preformed where this variable was included, but it never showed a significant effect. As the

(26)

7 Discussion

Individuals with a lower level of education are assumed to have a smaller opportunity cost of engaging in the illegal sector which is confirmed by the results presented in the thesis. The estimate for individuals with no secondary education was not significantly different than the estimate for non-college education. The expectation was that the effect of unemployment for non-secondary educated would be statistically significantly larger.

A possible explanation is if individuals experience their potential legal labor market situation poor, they may drop out of the labor force. This means they are neglected in the data, and the effect of the unemployment level for this group is downward-biased. It is clear, however, the effect of unemployment on crime is larger when examining unemployment for lower levels of education in comparison to the population average. This is evidence for the theory suggesting the effect of labor market opportunities on crime are larger for groups on the margin between the legal and illegal sectors. Comparing the results to previous studies, the estimate of 0.27 on unemployment for secondary education is higher than estimates in Edmark (2005). She estimates a 1-percent increase in the population average unemployment level to increase property crime rates by 0.11 percent. Notably, the estimate of the population average unemployment is also higher in this thesis. This could indicate that the role of unemployment when studying crime rates has increased overall which means the estimate of 0.27 is not directly comparable.

Further, some of the difference is likely to stem from slightly different specifications of the model or different definitions of unemployment, and this should be kept in mind when comparing the results. Gould et al. (2002) apply a log-linear model which mean the estimated coefficient should be interpreted as the percentage effect of a 1 percentage point increase in the unemployment level. They report estimates between 1.32-2.26 de- pending on specification and if they use IV or not. Interestingly, they use unemployment for non-college educated males which were the highest point estimate for unemployment by educational level obtained in this thesis.

(27)

To present the estimates in a comprehensible context, it is practical to apply some re- cently reported numbers. In 2019 the average number of reported crimes per 100.000 inhabitants in Sweden was 8903. Suppose the unemployment rate for those with at most secondary education increases by 10 percent. This would according to the model lead to an increase in property crime per 100.000 by 240, or in total 24 thousand assuming a population of 10 million. As crime is costly, it is important to account for increases in property crime following increases in unemployment since the benefits of labor market programs may otherwise be underestimated. Even if the point estimate should not be used as an absolute fact, this thesis along with similar research confirms the importance of analyzing margins instead of averages.

Income never displayed a significant relationship with property crime rates throughout the different specifications. This is likely due to income being a bad representation of legal labor market opportunities. The disposable income includes more than the wage and this creates noise leading to a weak estimated effect. Suppose, for example, the unemployment level for non-college or non-secondary educated individuals increases.

Using wages, this would mean a lower average wage among the groups, but the use of disposable income means the decrease can be somewhat offset by an increase in unemployment benefits or similar transfers. The weak effect is hence not evidence that wages are unimportant in the decision whether to engage in the legal or illegal sector. Rather, the weak effect is evidence that income poorly reflects wages in this context. Notably, results not reported showed that the income by educational level appeared negative and significant when unemployment was excluded from the model. With unemployment level included the standard error increased substantially indicating a high correlation between the covariates. This substantiates the view that the disposable income is influenced by variations in unemployment due to transfers.

(28)

Most previous research struggle to estimate similar effects of legal labor market opportunities on violent crimes as the effects found for property crime. The argument is typically that property crime to a larger extent are economically motivated and therefore are based more rationally on the opportunities at hand. On the contrary, it is not far fetched to believe poor opportunities on the labor market create distress and tension, as argued in Nordin and Alm´en (2017), triggering violent behaviour. With this line of thinking, the results reported in this thesis regarding violent crime is somewhat unex- pected. Two possible explanations for this regards to the data. First, as mentioned in Nordin and Alm´en (2017), long-term unemployment varies little from year to year and may in connection with the relatively few number of observations in this thesis cause the insignificant relationship. Second, the fact that this thesis has fewer observations may alone cause insignificant estimates due to large standard errors. More observations, by using municipality data for example, could lead to smaller standard errors and subse- quently significant estimates.

Lastly, a potential issue with the analysis is the omitted factors that were left out due to data restrictions. If the cost of punishment, the illegal benefit and the psychological costs are not controlled for by the entity- and time fixed effects they may according to the theoretical framework yield omitted variable bias. It is not clear in which direction this would bias the results as these variables typically are hard to obtain. However, this is necessary to keep in mind when interpreting the results. Suppose, for example, when the unemployment level increased during the economic crisis, the illegal benefit increased due to lower levels of self-protection or the psychological cost of crime decreased due to distress. In this case, the pure effect of unemployment could be different. However, as previous research also fail to incorporate these variables, the results obtained in this thesis can be compared and the larger effect of unemployment obtained in this study is

(29)

not driven by the omitted factors related to the theoretical framework.

7.1 Future research

Using income as a measure for the wage is showed to be insufficient why future research with better data can learn more about the effect of labor market opportunities. As discussed earlier in the thesis, wages are frequently proven to have more explanatory power than unemployment. Also, further investigating the potential endogeneity of unemployment and wages regarding when it is a problem and what problems it causes, can create a more thorough baseline for similar papers.

7.2 Concluding remarks

The aim of this thesis was to study how labor market opportunities for men with lower levels of education, a group potentially on the margin between the legal and illegal market, affect crime rates in Sweden. Results substantiate the importance to identify groups on the margin when estimating a relationship between crime and the labor market and non-college educated men seem to be such a group in Sweden. Recognizing that the effect of the labor market situation on crime rates is larger when a group on the margin is studied, serves as insightful knowledge when designing labor market policies or educational programs targeting this group.

(30)

References

Becker, G. S. (1968). Crime and punishment: An economic approach. In The economic dimensions of crime (pp. 13-68). Palgrave Macmillan, London.

Edmark, K. (2005). Unemployment and crime: Is there a connection?. Scandina- vian Journal of Economics, 107(2), 353-373.

Ehrlich, I. (1973). Participation in illegitimate activities: A theoretical and empirical investigation. Journal of political Economy, 81(3), 521-565.

Ehrlich, I. (1996). Crime, punishment, and the market for offenses. Journal of Economic Perspectives, 10(1), 43-67.

Estrada, F., Pettersson, T., & Shannon, D. (2012). Crime and criminology in Sweden. European Journal of Criminology, 9(6), 668-688.

Foug`ere, D., Kramarz, F., & Pouget, J. (2009). Youth unemployment and crime in France. Journal of the European Economic Association, 7(5), 909-938.

Freeman, R. B. (1996). Why do so many young American men commit crimes and what might we do about it?. Journal of Economic perspectives, 10(1), 25-42.

Freeman, R. B. (1999). The economics of crime. Handbook of labor economics, 3, 3529-3571.

(31)

Glaeser, E. L., & Sacerdote, B. (1999). Why is there more crime in cities?.

Journal of political economy, 107(S6), S225-S258.

Gould, E. D., Weinberg, B. A., & Mustard, D. B. (2002). Crime rates and local labor market opportunities in the United States: 1979–1997. Review of Economics and statistics, 84(1), 45-61.

Grogger, J. (1998). Market wages and youth crime. Journal of labor Eco- nomics, 16(4), 756-791.

Gr¨onqvist H (2011). Youth unemployment and crime: new lessons explor- ing longitudinal register data. Working Paper 7/2011, (SOFI) Stockholm University

Ihlanfeldt, K. R. (2006). Neighborhood crime and young males’ job oppor- tunity. The Journal of Law and Economics, 49(1), 249-283.

Levitt, S. D. (1998). Why do increased arrest rates appear to reduce crime:

deterrence, incapacitation, or measurement error?. Economic inquiry, 36(3), 353-372.

Lin, M. J. (2008). Does unemployment increase crime? Evidence from US data 1974–2000. Journal of Human resources, 43(2), 413-436.

Lochner, L. (2004). Education, work, and crime: A human capital approach.

International Economic Review, 45(3), 811-843.

(32)

Machin, S., & Meghir, C. (2004). Crime and economic incentives. Journal of Human resources, 39(4), 958-979.

McCollister, K. E., French, M. T., & Fang, H. (2010). The cost of crime to society: New crime-specific estimates for policy and program evaluation. Drug and alcohol dependence, 108(1-2), 98-109.

Murray, M. P. (2006). Avoiding invalid instruments and coping with weak instruments. Journal of economic Perspectives, 20(4), 111-132.

Mustard, D. B. (2010). How do labor markets affect crime? New evidence on an old puzzle.

Nordin, M., & Alm´en, D. (2017). Long-term unemployment and violent crime. Empirical Economics, 52(1), 1-29.

Piehl, A. M. (1998). Economic Conditions. The Handbook of Crime & Punish- ment, 302.

Raphael, S., & Winter-Ebmer, R. (2001). Identifying the effect of unem- ployment on crime. The Journal of Law and Economics, 44(1), 259-283.

Statistics Sweden. (2017). The labour market for persons with a lower level of education 2005-2016 (Statistiska meddelanden Serie Arbetsmarknaden AM 110 SM 1704). Retrieved from https://www.scb.se/contentassets/

bfca27069b504896af1244d3c837f544/am0401_2017k03_sm_

am110sm1704.pdf

(33)

Swedish National Council for Crime Prevention. (2019). Swedish crime sur- vey 2019. Stockholm: Swedish National Council for Crime Prevention

Swedish National Council for Crime Prevention. (2019). Summary of re- ported crimes 2019. Stockholm: Swedish National Council for Crime Prevention

Tauchen, Helen. 2010. Estimating the Supply of Crime: Recent Advances.

In Handbook on the Economics of Crime, edited by Bruce L. Benson and Paul R Zimmerman, 24-52. Northampton, MA: Edward Elgar.

Oster, A.,¨ Agell, J. (2007). Crime and unemployment in turbulent times.

Journal of the European Economic Association, 5(4), 752-775.

(34)

Appendix

A Full regression results with control variables

Table A.1: Results from the regressions displayed Table 1 including coefficients for control variables.

Variables Property (1) Property (2) Property (3)

Income (non-college) -0.12

(0.35)

Income (non-secondary) 0.18

(0.24) Unemployment (pop. average) 0.15 ***

(0.06)

Unemployment (non-college) 0.27 **

(0.09)

Unemployment (non-secondary) 0.19 **

(0.06)

Clear-up rate -0.09 ***

(0.03)

0.09 ***

(0.03)

-0.09 **

(0.04)

Mean income -0.06

(0.13)

-0.04 (0.12)

-0.13 (0.12)

Population density 0.3

(0.39)

0.29 (0.37)

0.34 (0.39) Alcohol consumption 0.04

(0.13)

0.03 (0.13)

0.001 (0.14)

Share educated -0.76 ***

(0.26)

-0.65 **

(0.27)

-0.62 * (0.29) Share young men (15-24) -0.05

(0.49)

0.03 (0.46)

-0.09 (0.48)

Share immigrants -0.03

(0.1)

-0.01 (0.1)

Observations 294 294 294

R2 0.21 0.23 0.18

(35)

Table A.2: Results from the regresions displayed Table 2 including coefficients for control variables.

Variables Property (1) Property (2) Property (3)

Income (non-college) -0.01

(0.36)

Income (non-secondary) 0.16

(0.24) Unemployment (pop. average) 0.16 **

(0.06)

Unemployment (non-college) 0.28 **

(0.09)

Unemployment (non-secondary) 0.2 **

(0.06)

Mean income -0.04

(0.12)

-0.04 (0.11)

-0.12 (0.11)

Population density 0.3

(0.39)

0.24 (0.33)

0.27 (0.35) Alcohol consumption 0.06

(0.15)

0.06 (0.15)

0.02 (0.16)

Share educated -0.83 ***

(0.24)

-0.73 **

(0.25)

-0.67 * (0.29) Share young men (15-24) -0.02

(0.55)

0.07 (0.5)

-0.05 (0.55)

Share immigrants -0.05

(0.09)

-0.03 (0.1)

Observations 294 294 294

R2 0.17 0.18 0.14

(36)

Table A.3: Results from the regressions displayed Table 3 including coefficients for control variables.

Variables Violence (1) Violence (2) Violence (3) Violence (4)

Unemployment (pop. average) 0.11 (0.09)

Long-term unemployment (pop.average) 0.05 (0.05)

Long-term unemployment (non-college) 0.05

(0.04)

Long-term unemployment (non-secondary) 0.03

(0.03)

Clear-up rate -0.02

(0.06)

-0.02 (0.05)

-0.01 (0.05)

-0.02 (0.06)

Mean income -0.05

(0.18)

-0.07 (0.16)

-0.5 (0.15)

-0.9 (0.16)

Population density -0.32

(0.18)

-0.29 (0.21)

-0.30 (0.2)

-0.27 (0.19)

Alcohol consumption 0.39**

(0.12)

0.38**

(0.13)

0.39**

(0.13)

0.37**

(0.12)

Share educated -0.33

(0.52)

-0.25 (0.6)

-0.25 (0.59)

-0.25 (0.58)

Share young men (15-24) 0.34

(0.56)

0.31 (0.57)

0.33 (0.57)

0.34 (0.5)

Share immigrants 0.17

(0.14)

0.19 (0.14)

0.21 (0.14)

0.2 (0.14)

Observations 294 294 294 294

R2 0.13 0.12 0.13 0.11

(37)

B Unemployment comparison by data source

Figure B.1: Unemployment by data source.

Figure B.1 displays the difference in using different sources of data and different definitions of unemployment. The red line shows the numbers of the Labor Force Survey (LFS) at SCB. The blue line displays ”openly unemployed” from AMS which is used in this study. The reason it is below the green line is that it excludes individuals partici- pating in labor market programs. Noticeably there is a difference in the unemployment numbers when comparing the different sources of data. This is due to different calcu- lations. The ”openly unemployed” measure used in this thesis is simply the number of individuals that has been registered as unemployed at the AMS divided by the population 16-24. The measure used in the LFS by SCB is sample based not defined the same way. Even though the numbers differ the trend and variation over time is almost identical when comparing the unemployment level in LFS to that of AMS.

(38)

C Descriptive statistics

Table C.1: Descriptive statistics of all included variables.

Variable Average Standard

deviation Min. Max.

Property crime 7516 1584 4713 12716

Violent crime 2076 505 1022 3546

Income (non-secondary) 185181 24625 128100 244750 Income (secondary) 208009 33255 142740 280350

Income (mean) 300700 57620 205000 553000

Unemployment (non-secondary) 16.98 3.73 9.3 26.5 Unemployment (secondary) 16.31 3.75 8.10 26.9 Unemployment (pop. average) 14.49 3.04 7.8 22.6 Clear-up rate (property crime) 0.13 0.03 0.07 0.29 Clear-up rate (violent crime) 0.36 0.06 0.26 0.54

Labor market opportunities and crime in Sweden