The costs of demolishing a vacant building are often justified on the grounds of crime reduction. I explore this claim by estimating the spatial and temporal effects of demolitions on reported crime in the city of Saginaw, Michigan. To do so, I estimate a model that uses within‐block group variation to compare crime after a demolition occurs to before the permit for that demolition was issued. Results indicate that demolitions reduce crime by about 8 percent on the block group in question and 5 percent on nearby block groups, with the largest impact concentrated one to two months after the demolition occurs.

1 INTRODUCTION The durability of housing has been found to be the primary reason why urban decline is more persistent than growth (Glaeser & Gyorko, 2005). Because houses are not removed as quickly as they are built, negative shocks decrease housing prices more than they decrease population which leads to slower changes than those caused by a positive shock. Because of this, decline is much longer and slower than growth. Demolitions of vacant buildings may be one way to counteract this durability and persistence. However, little research exists on the effects of demolitions or on vacancies in general (Schilling & Logan, 2008). Policy makers and academics suggest that crime is related to both building vacancies and depopulation.1 Not only does crime cause depopulation and vacant structures (Cullen & Levitt, 1999), they argue, but vacant structures cause crime through increased incidences of arson, the sheltering of criminals, and the creation of general disorder (Whitman et al., 2001; Spelman, 1993; Garber et al., 2008; Winthrop & Herr, 2009). With this as one justification, the United States Government spends millions of dollars a year demolishing vacant buildings.2 Between 2008 and 2011, the U.S. Department of Housing and Urban Development (HUD) spent almost $200 million on vacant building demolitions under the Neighborhood Stabilization Program (NSP) which is only one of several funding sources for demolitions (United States Department of Housing and Urban Development, 2011). The city of Flint, Michigan, alone was awarded over $3 million in 2010 through this program—the same year in which the number of murders in the city reached an all‐time high. The city budget was so constrained that year that the jail was shut down and, as a result, police officers had to issue tickets rather than arrest warrants for many offenses.3 Are demolitions the best use of funding in such economically and fiscally stressed cities? In this paper, I seek to answer part of this question by examining the relationship between demolitions and crime. Although previous research has shown that high‐rise public housing demolitions reduce crime (Aliprantis & Hartley, 2015), no research has examined the link between single‐family home demolitions and crime. Vacant building demolitions of single‐family homes differ from public housing demolitions in that they do not redistribute concentrations of low‐income residents to less dense areas. Therefore, it is unclear whether crime is similarly affected by these types of demolitions. I fill this gap in the literature by estimating the effect of vacant building demolitions of single‐family homes on crime. To do so, I construct a unique data set of demolitions and reported crime4 in Saginaw, Michigan, at the block group level. I then estimate a Poisson fixed effects model that uses within‐block group level variation. In this model, each cross‐sectional unit is allowed to have its own linear time trend in addition to a separate level effect, also referred to as a random trend model (Wooldridge, 2002). Because the permitting process for a demolition may itself induce a change in crime, I do not compare crime after the demolition to crime before the demolition, but rather crime after the demolition to crime before the permit is issued. I also analyze the displacement and spillover effects of demolitions and permits through a spatially lagged independent variable model, and the temporal effects through a flexible dynamic model. Results indicate that demolitions do cause a reduction in crime. Specifically, demolitions are associated with a reduction of 7.5 percent of total crimes per block group per month, 6.7 percent of violent crimes, and 9.7 percent of property crimes. This is equivalent to a reduction of about nine total crimes per year for the average block group, two violent crimes, and five property crimes.5 Demolitions also cause positive spillover effects into nearby neighborhoods, reducing crimes there by 5.1 percent, 3.2 percent, and 7.1 percent for total, violent, and property crimes, respectively. Perhaps surprisingly, permits are also associated with a reduction in crime, although the point estimates are not significant in the main model. Anecdotal evidence suggests that permits may cause an increase in crime due to the public nature of the permitting list, which signals to criminals that the structure is empty and that crime can occur there unnoticed. However, I do not find this to be true, perhaps because the buildings were boarded up and secured upon being permitted for a demolition. Overall, these results show that demolitions do work to reduce crime. However, it is unclear whether demolitions fully revert crime to its prevacancy level. If they do not, vacancy‐prevention strategies may be even more effective at minimizing crime. More research is needed to compare the cost–benefit tradeoffs of demolitions and vacancy prevention strategies to determine the most efficient use of public funds.

2 CONCEPTUAL FRAMEWORK There are three theories that lend insight into how demolitions may impact criminal behavior.6 First, the rational choice theory of crime hypothesizes that criminals maximize their economic well‐being by comparing the benefits and costs of crime such as fines, imprisonment, and social stigmatism (Becker, 1968, 1994; Ehrlich, 1973, 1975; Fleisher, 1963, 1966a, 1966b). If the potential gain from committing a crime is sufficiently greater than the combined risk of being caught and the size of the punishment, then the criminal chooses to commit the crime. A demolition may induce a change in the actual or perceived probability of being caught committing a crime due to the removal of the structure. The actual and perceived probability of being caught committing a crime may increase after a demolition due to the lack of a shelter to hide the crime. This may increase reported crime on or near the vacant parcel, but it may also disincentivize individuals from committing crimes in the first place. Or, the perceived probability of being caught committing a crime may decrease because it may appear that the number of eyes on the street has been reduced. This may lead to increased levels of crime on a block group that has undergone a demolition. However, once a house has been demolished, there is less property to steal, reducing the benefits of crime, so property crime might also decline. These changes in the perceived costs and benefits may also cause two types of crime displacement. First, a criminal may choose to commit a different type of crime because he can no longer undergo the activity that he had previously undertaken (Repetto, 1974). For example, if a drug dealer can no longer produce a drug in an abandoned building, he may switch to robbery as a source of income. I examine this displacement by categorizing crimes into different groups and examining their changes before and after a demolition. Crime displacement may also occur after a demolition if a criminal moves to another location in order to commit the same crime. For instance, an illegal drug manufacturer may simply move to another vacant building to use as a laboratory once his original one is demolished. I examine these displacement effects through a spatially lagged independent variable model. The second theory that relates demolitions to criminal behavior is what has been termed the “broken windows” theory (Kelling & Wilson, 1982). This theory states that if a window in a building is broken and left unrepaired, the rest of the windows in the building will soon be broken as well. Window breaking does not occur on a large scale because some areas are inhabited by determined window‐breakers, whereas others are populated by window‐lovers. Rather, one unrepaired window is a signal that breaking more windows costs nothing. In the case of vacant buildings, the theory implies that one vacant building lying decrepit leads to further crime solely based on the signal that the probability of being punished is low. This suggests that demolitions cause positive spillover effects in which crime is reduced not only in the immediate area but in surrounding areas as well. However, this theory is highly disputed. Some authors have found that property disorder does increase crime and that targeting this disorder is a viable crime reduction strategy (Braga & Bond, 2008; Braga et al., 1999; Corman & Mocan, 2005; Kelling & Sousa, 2001), while others find no support for the broken windows theory nor for the proposition that broken windows policing is the optimal use of scares law enforcement resources (Fagan & Davies, 2000; Harcourt & Ludwig, 2006). The third theory that lends insight into the relationship between demolitions and crime is the social disorganization theory of crime. This theory suggests that social capital and cohesion are disrupted when a neighborhood loses population and the social controls that put limits on criminal activity deteriorate (Park & Burgess, 1925; Shaw & McCay, 1932, 1942, 1969). As houses become vacant, crime increases. Previous research has found this to be true with estimates ranging from $1,472 of public safety money spent per vacant property (Winthrop & Herr, 2009) to a doubling of crime rates on block groups with open abandoned buildings (Spelman, 1993). A number of studies look at foreclosures rather than at vacant buildings with most finding a positive relationship between foreclosures and crime (Acevedo, 2009; Arnio, Baumer, & Wolf, 2012; Clark & Teasdale 2005; Ellen, Lacoe, & Sharygin, 2011; Goodstein & Lee, 2010; Harris, 2011; Immergluck & Smith, 2006; Pandit, 2011; Stucky, Ottensmann, & Payton, 2012; Teasdale, Clark, & Hinkle, 2012) and some finding no relationship between the two (Cui, 2010; Jones & Pridemore, 2012; Kirk & Hyra, 2011; Madensen, Hart, & Miethe, 2011). However, demolitions are not likely to reverse this effect since they do not replace these social controls. In addition to the demolition, the permitting process for a demolition may affect crime. Anecdotal evidence suggests that permits increase property crime because individuals patrol permitted structures based on public records and permit stickers on the houses, then strip vacant houses of copper pipes, fixtures, and aluminum siding (Garber et al., 2008). This change in crime during the permit process may cause persistence in crime that may extend into the postdemolition period; once a criminal targets a neighborhood, he may be more likely to return later. However, it is also possible that the permitting process decreases violent and other types of crime because of the frequent entrance and exit of city officials into the house and neighborhood. Different types of crimes may be differentially affected by demolitions. The rational choice theory of crime suggests that both violent and property crimes may be affected—violent since the loss of the structure alters the perceived and/or actual likelihood of being caught, and property because of this and because of the reduction in property to steal. The broken windows theory, alternatively, suggests that property crime is more likely to be affected than violent crime, both on the parcel in question and in the neighborhood as a whole.

3 DATA The data for this study were collected in collaboration with the city of Saginaw's geographical information systems (GIS) and inspections departments. The GIS department provided block level demographics and parcel level crime data and the inspections department provided detailed demolition data, including permit dates, demolition costs, and sources of funding. I aggregate these data into block group level observations to better capture neighborhood level effects and to minimize measurement error, since the likelihood of a crime being assigned to the wrong geographic unit increases as the geographic unit becomes smaller. The data set consist of 73 block groups spanning 24 months: January 2008 through December 2009. Table 1 displays basic summary statistics for crimes, demolitions, and permits on these block groups. Table 1. Summary statistics for crimes, permits, and demolitions per block group/month Variable No. of Obs. Total Mean Std. Dev. Min Max All crimes 1,752 18,197 10.39 8.09 0 64 Violent crimes 1,752 4,812 2.75 2.75 0 24 Property crimes 1,752 6,973 3.98 3.55 0 30 Demolitions 1,752 254 0.14 0.57 0 8 Permits 1,752 251 0.16 0.72 0 12 I break incidents of crime into three categories: all crime, violent crime, and property crime. Violent crime includes murder and nonnegligent manslaughter, forcible rape, robbery, simple assault, and aggravated assault. Property crime includes burglary and breaking and entering, purse snatching, theft from a building, theft from a machine, theft from a vehicle, other larcenies, motor vehicle theft, stolen property, destruction of property, shoplifting, pocket picking, and arson. Violent crime and property crime do not sum to all crime; there are some crimes that are not counted as either type such as fraud, intimidation, weapon law violations, and drug/narcotic violations. Figure 1 displays the number of demolitions and permits that occur each month in my data set. The length of time that a house is permitted for a demolition before it is demolished varies from 0 to 23 months with a mean of 11.33 months. The frequency of permit times can be seen in Figure 2. Figure 1 Open in figure viewerPowerPoint Demolitions and permits by month Figure 2 Open in figure viewerPowerPoint Length of time of permits 3.1 Brief background on Saginaw Saginaw has undergone rapid depopulation since the decline of manufacturing in the latter half of the twentieth century. Foreclosures and land vacancies that had already been increasing were exacerbated by the financial and housing crises of 2007 which is illustrated in Figure 3. Crime is also a serious problem. As mentioned previously, Saginaw ranked as the most violent city in America from 2003 through September of 2008 (Burns, 2010). Figure 3 Open in figure viewerPowerPoint Total residential vacancies in Saginaw, Michigan, 2002–2009 Property values have decreased substantially from an already depreciated level since the financial crisis of 2007. Saginaw home prices devalued by 10.3 percent in 2009 and the unemployment rate reached 19.7 percent (Sperling's Best Places, 2009). Because of the lag between the time when a home loses value and when its official assessed value decreases, assessed property values will continue to feel the effect of both the recession and depopulation long after they have both subsided.7 Because of these issues, the municipal consulting firm of Plante & Moran projects that the city of Saginaw could face a $19.9 million deficit by 2014 if leaders do not adjust to declining revenues and a shrinking population (Engel, 2010). 3.2 Demolition process Demolitions are a key part of Saginaw's policy relating to vacant and abandoned buildings. To implement demolitions, Saginaw uses both Community Development Block Grant and NSP funds provided by the HUD. The buildings that Saginaw demolishes need not be vacant; they only need to be qualified as structurally dangerous. They could be owned by private individuals, corporations, limited liability companies, mortgage companies, the county treasurer, or the county land bank. The city does not take ownership of the property; it only enforces its right and obligation to keep the city safe. Once demolished, the property value drops to the current local value for a vacant lot of equal size. Once a building is put on the dangerous building list, it is demolished in numerical order. There are three ways in which a property can be placed on the dangerous building list. First, a building can be added to the list through a resident complaint filed with the city's complaint department or clerk's office. Second, a building can be added due to an internal complaint from a city worker who observed the property firsthand while in the field or received complaints at a neighborhood meeting. Third, a building can be added through citywide sweeps undertaken by the inspections department each spring and sometimes in the fall (Personal correspondence with Scott Crofoot, Dangerous Buildings Inspector, City of Saginaw). If a building undergoes arson, it is immediately demolished. I use this information to examine the endogeneity of demolitions and arson within Saginaw. I also exclude the 25 emergency demolitions and the arsons that caused them from my dataset because they provide a direct source of endogeneity.8 Although this is a concern for the analysis in that it may be causing me to miss an important crime reduction caused by demolitions, there is additional concern that the structure of the demolition policy is incentivizing residents to commit arson against neighboring vacant homes because if a vacant building undergoes arson, it is moved to the top of the demolition list. If I include these emergency demolitions and arsons in my analysis I am not able to determine whether the demolition policy causes the arsons, whether the arsons cause the demolitions, or what I would truly like to know: whether demolitions reduce the incidence of arson. I do include all other arson that does not immediately lead to a demolition which allows me to determine whether demolitions reduce the incidence of arson in the neighborhood of the demolition. In addition, I perform robustness checks that include and exclude all arson to determine whether they cause bias in my results. The demolition process follows a set procedure. When the house becomes permitted for a demolition, a notice to the owner is placed at the front entry of the building as part of the notification requirements. This occurs once at the beginning of the process and again 75–80 days later when the notice of findings is posted. Fifteen to 30 days after the last posting, the house is measured to provide cubic volume and determine demolition costs as well as inspected for asbestos.9 This activity near the end of the permit period is believed to cause an increase in theft of anything that has a metal content (Personal correspondence with Scott Crofoot, Dangerous Buildings Inspector, City of Saginaw). Table 2 lists crimes that occurred before permit issuance on parcels that underwent a demolition and Table 3 lists crimes that occurred during permit issuance on parcels that underwent a demolition. The most common crime in both lists is arson,10 followed by burglary, larceny, and damage to property. Surprisingly, there are no incidences of drug crimes on these vacant properties. Table 4 lists crimes that occurred on parcels after a demolition took place. Surprisingly, there were still four arsons even after the vacant building was removed. These are likely arsons in which brush or garbage was lit on fire. Notably missing from these tables are drug‐related crimes, perhaps because these are less visible from the outside than a crime like arson, or because they do not generally have a victim to report the crime like in the case of assault. Drug crimes could also be missing due to locational inaccuracies in the crime data at such a small geography. To minimize this potential measurement error, I run the analyses at a higher level of aggregation. Table 2. Crimes that occurred on parcels before a permit was issued Offense Description Frequency Arson – Residence 26 Burglary – Forced Entry – Residence 14 Larceny 6 Damage to Private Property 6 Assault and Battery/Simple Assault 3 Disorderly Conduct 3 Aggravated Assault – Nonfamily – Gun 2 Burglary – Forced Entry – Nonresidence 2 Larceny – Personal Property from Vehicle 2 Robbery – Business Strong Arm 1 Robbery – Street Gun 1 Arson (other) 1 Arson – Burning of Real Property 1 Burglary – No Forced Entry – Residence 1 Larceny – From Yards/Grounds 1 Fraud (Other) 1 Embezzlement – Business Property 1 Retail Fraud, Theft 3rd Degree 1 Disorderly Conduct (Other) 1 Traffic – Furnish False Info to Officer 1 Health and Safety Violations 1 Skipped Number 1 Inspections/Investigations – Lost and Found Property 1 Miscellaneous – General Assistance 1 Table 3. Crimes that occurred on parcels during the permit period Offense Description Frequency Arson – Residence 23 Larceny (Other) 5 Burglary – Forced Entry – Residence 3 Aggravated Assault\Nonfamily – Gun 2 Robbery ‐ Business Strong Arm 1 Robbery – Street Gun 1 Burglary – Forced Entry – Nonresidence 1 Burglary ‐ No Forced Entry – Residence 1 Damage to Property – Private Property 1 Table 4. Crimes that occurred on parcels after a demolition Offense Description Frequency Arson – Residence 4 Dog Law Violations 1 Figure 4 plots the average monthly number of crimes before and after a demolition on block groups that underwent a demolition. Descriptively, overall crime levels decline in the first few months following a demolition, then appear to begin to increase a bit in the subsequent months. Additionally, there appears to be a slight spike in property crime the month of the demolition, likely due to the existence of an active construction site. However, it is important to keep in mind that these numbers do not control for seasonality or other variables that may be correlated with crime. Figure 4 Open in figure viewerPowerPoint Average number of crimes on a block group before and after a demolition

4 EMPIRICAL METHODOLOGY To estimate the effect of a single‐family vacant building demolition on crime, I would ideally compare a block group that underwent a demolition to itself had it not undergone the demolition. This, unfortunately, is not possible. Alternatively, comparing a block group that underwent a demolition to one that did not will be misguided since there may be unobserved factors that make block groups with demolitions different from block groups without demolitions. This leads to inconsistent estimates if using a cross‐section for analysis. Additionally, estimates that do not control for month fixed effects also lead to misleading results since a large portion of the demolitions occurred in the winter when crime levels are the lowest. I therefore use a within‐estimator to compare block groups that underwent a demolition to themselves over time which removes the time‐invariant unobserved heterogeneity of each block group. However, I do not compare crime after a demolition to crime immediately before a demolition because the permit process for a demolition may induce a change in crime that biases results. If the permit process increases property crime, using the permit period as the base period will bias coefficients downwards. Alternatively, violent and other crime may actually decrease during the permit period due to city officials entering and leaving the premises or due to the boarding up of the property. This will cause these types of crime to appear to increase after a demolition when in fact they are reverting to the mean. In order to examine and account for these trends, I estimate both a contemporaneous model with the variables specified as stock variables and a dynamic model with the variables specified as flow variables. The stock of demolitions is the cumulative number of demolitions that have occurred on the block group and the stock of permits is the cumulative number of houses with permits that have not yet undergone a demolition. The flow variable for each is defined as the first difference of the stock variable, i.e.: Because my dependent variable, the number of crimes on a block group in a given month, takes on nonnegative integer values (is a count variable), a linear model for E(y│x) is not ideal because it can lead to negative predicted values (Wooldridge, 2002, p. 388).11 Also, because y can take on the value zero with positive probability, a log transformation is inappropriate. Therefore, I assume that given takes on a Poisson distribution. The Poisson model is as follows: (Winkelmann, 2008 , estimates of β are consistent even if the mean does not equal the variance—when there is overdispersion.12 where(Winkelmann,). Provided that, estimates of β are consistent even if the mean does not equal the variance—when there is overdispersion. Since there may be concern that demolitions occur on block groups that have positive crime trends, I estimate a random trend model that includes block group‐specific time trends using multinomial quasi‐conditional maximum likelihood estimation as described in Wooldridge (1999). I specify both permits and demolitions as stock variables and estimate the following equation: Each variable is measured at the block group and month level. equals the number of permits issued for block group i in month t. equals the number of demolitions on block group i in month t. The coefficient δ 1 on gives us the mean impact of permits on crime from the month of issuance to the month of the demolition. The coefficient β 1 on tells us the mean impact of demolitions on crime from the month of the demolition through the end of the sample period. Block group fixed effects are represented by and represents block group time trends. Time‐invariant variables are captured by the block group level fixed effects and are therefore not included in this equation. I use standard errors clustered at the block group level that are robust to heteroskedasticity and arbitrary forms of error correlation within each block group. Since many crimes go unreported, the crimes in my data set are likely underrepresentative of overall crime in the city. Reported crimes may increase or decrease due to fewer or more eyes on the streets reporting the crimes. Moreover, certain types of neighborhoods are likely to have higher rates of underreporting than others—neighborhoods with lower income, younger, and male victims are more likely to have underreporting of crimes while neighborhoods with a large number of homeowners are less likely to have underreporting (Skogan, 1999). If demolitions are located in areas that are prone to underreporting, my estimates may be attenuated and represent a lower bound on the effect of demolitions on crime. Additionally, time‐invariant neighborhood level differences in demographics are captured by the block group level fixed effect in my model. The only thing that I cannot directly control for are changes in neighborhood population and demographics since they may themselves be endogenous to crime and demolitions. 4.1 Displacements and spillovers Knowing whether crime changes on one block group due to a demolition does not tell us whether overall crime changes or is merely displaced into surrounding neighborhoods. I therefore estimate a model in which I include a spatial lag of demolitions. In other words, I add into the models above the sum of demolitions and permits that occurred in block groups that touch block group i, or block group i’s neighbors. To do so, I estimate the mean effects model with a spatial lag as follows: is an 1 × I spatial weights vector that assigns a weight of 1 to block groups that are contiguous to block group i and a weight of 0 to block groups that are noncontiguous to block group i (Drukker et al., 2011 is an I × 1 vector of the permits on every block group in the city. is an I × 1 vector of the demolitions on every block group in the city. whereis an 1 ×spatial weights vector that assigns a weight of 1 to block groups that are contiguous to block groupand a weight of 0 to block groups that are noncontiguous to block group(Drukker et al.,).is an× 1 vector of the permits on every block group in the city.is an× 1 vector of the demolitions on every block group in the city.

5 RESULTS Results show that demolitions do reduce crime, with or without controlling for the permitting period (Table 5). In fact, when the permitting period is included as a regressor, the effect of demolitions on all crime and violent crime increases in magnitude. This implies that permits actually decrease some types of crime so that when the demolition period is compared to the prepermit period, the effect on those crimes is even larger than when it is compared to the permitting period.13 Table 5. Effect of demolitions and permits for demolitions on crime in Saginaw, Michigan (1) (2) (3) (4) (5) (6) All Crime Violent Crime Property Crime All Crime Violent Crime Property Crime Demolitions −0.049*** −0.049*** −0.063*** −0.075*** −0.067* −0.097** (0.013) (0.018) (0.024) (0.027) (0.034) (0.039) Permits −0.016 −0.011 −0.022 (0.013) (0.016) (0.014) Obs. 1,728 1,728 1,728 1,728 1,728 1,728 Block groups 72 72 72 72 72 72 When permits are included (as well as block group fixed effects and individual block group time trends14), one demolition is associated with a reduction of 7.5 percent of crimes per month per block group, or about 9 crimes per year. For violent crime, one demolition reduces violent crime by 6.7 percent, or 2 crimes per year on average. And for property crime, demolitions cause a reduction of 9.7 percent of crimes, or an average of 5 property crimes per year.15 Figures 5 through 7 show the effect of demolitions on crime before and after a demolition, controlling for the permit period. In this model, I include two leads and five lags of demolitions.16 The effect of a demolition on overall crime is negative and significant two months after the demolition, as is the effect of a demolition on property crime. The effect of a demolition on violent crime is negative and significant one month after the demolition, but significant and slightly positive four months after the demolition, suggesting that violent crime may increase again after a few months although mostly likely not back up to predemolition levels.17 Figure 5 Open in figure viewerPowerPoint Dynamic effect of demolitions on all crime Note: The dark line plots the semielasticities of demolitions on crime from a dynamic random trend model that includes block group fixed effects and time trends. Sample is a monthly panel of all block groups in Saginaw, MI from 2008 to 2009. Crime offenses refer to the number of incidents on each block group in each month. The light gray lines are the 95 percent confidence intervals calculated using robust standard errors clustered at the block group level. Figure 6 Open in figure viewerPowerPoint Dynamic effect of demolitions on violent crime Note: The dark line plots the semielasticities of demolitions on violent crime from a dynamic random trend model that includes block group fixed effects and time trends. Sample is a monthly panel of all block groups in Saginaw, MI from 2008 to 2009. Crime offenses refer to the number of incidents on each block group in each month. The light gray lines are the 95 percent confidence intervals calculated using robust standard errors clustered at the block group level. Figure 7 Open in figure viewerPowerPoint Dynamic effect of demolitions on property crime Note: The dark line plots the semielasticities of demolitions on property crime from a dynamic random trend model that includes block group fixed effects and time trends. Sample is a monthly panel of all block groups in Saginaw, MI from 2008 to 2009. Crime offenses refer to the number of incidents on each block group in each month. The light gray lines are the 95 percent confidence intervals calculated using robust standard errors clustered at the block group level. The permitting period itself has no statistically significant relationship with crime, although the point estimates are negative. Anecdotal evidence from discussions with Saginaw inspectors suggests that crime might increase on permitted properties due to the public nature of the permitting list, which signals to criminals that the structure is empty and that crime can occur there unnoticed. However, this appears not to have occurred. The permitting period had no effect on crime, or potentially a small negative effect. This may be because the buildings were boarded up and secured upon being permitted for a demolition. Demolitions also cause a positive spillover effect into nearby neighborhoods, reducing crime there as well (Table 6). A demolition on a contiguous block group is associated with a reduction of 5.1 percent of total crimes per block group per month, or 6 crimes per year for the average block group.18 For violent crimes, nearby demolitions cause a reduction of 3.2 percent per block group per month, or 1 per year. And for property crime, nearby demolitions cause a reduction of 7.1 percent per block group per month, or about 3 per year. When spatial lags are included, the main effects become insignificant. This is likely due to power—when additional lags are included overall power declines, but the spillover effects remain significant simply because there are more contiguous block groups than the main block group, and therefore more demolitions as well. Table 6. Displacement and spillover effects of demolitions and permits for demolitions on crime in Saginaw, Michigan (4) (5) (6) All Crime Violent Crime Property Crime Demolitions −0.023 −0.043 −0.019 (0.024) (0.032) (0.036) Permits −0.003 −0.017 0.004 (0.014) (0.020) (0.019) Demolitions on contiguous block groups −0.051*** −0.032*** −0.071*** (0.007) (0.011) (0.013) Permits on contiguous block groups −0.014*** −0.004 −0.021*** (0.004) (0.006) (0.006) Observations 1,728 1,728 1,728 Number of block groups 72 72 72 Nearby permits are also associated with a reduction in crime on the block group in question: a permit on a contiguous block group is associated with a reduction of 1.4 percent of crimes per block group per month, or 2 per year on average. This effect is predominantly driven by the effect of permits on property crime, which shows a 2.1 percent reduction on nearby block groups or about 1 per year for the average block group. Spillover effects may be stronger than direct effects for permits due to power: there are more permits on contiguous block groups because there are multiple contiguous block groups for each block group in question. These results support the hypothesis that a demolition may induce a change in the actual or perceived probability of being caught committing a crime due to the removal of the structure, which in turn causes individuals to be less likely to commit a crime and/or be caught doing so. Additionally, there may be less property to steal, reducing the benefits of crime and therefore the likelihood of doing so. The positive spillover effects from the demolitions seen here provide support for the broken windows hypothesis: reducing disorder caused by vacancies may have positive effects not just in the neighborhood of the demolitions but also in nearby neighborhoods. Since the theory suggests that property crime is more likely to be affected by disorder than violent crime, the larger effects on property crimes found here further support this hypothesis. These results may also provide support for the idea that once the unit is removed and individuals are less likely to go to that parcel to commit a crime, they are less likely to commit crimes nearby as well just because they are less likely to visit that neighborhood all together.

6 CONCLUSION This paper provides empirical evidence for the effect of vacant building demolitions on crime in a depopulating city. Demolitions cause a reduction of nine total crimes per year on average; two of which are violent crimes and five property crimes (the remainder are neither violent nor property crimes). This is equivalent to a 7.5 percent reduction in total crime, a 6.7 percent reduction in violent crime, and 9.7 percent reduction in property crime. Additionally, demolitions cause positive spillover effects into nearby neighborhoods, reducing crimes there by 5.1 percent, 3.2 percent, and 7.1 percent for total, violent, and property crimes, respectively. The average deconstruction cost of demolishing a house in Saginaw is $5,021, not including the overhead costs to the city government and federal agencies distributing the funds. This means that at a minimum, demolitions cost $537 per crime reduced, or $2,511 per violent crime and $1,004 per property crime. Crime reduction is only one potential benefit of demolishing a vacant property, though, so these dollars are likely also creating other benefits. Permits for demolitions also appear to cause a slight reduction in crime, contrary to anecdotal evidence that suggest that permits may signal to individuals that crime can go undetected on that property. This reduction in crime caused by permits may be due to the securing of the property during the permitting period. These results coincide with hypotheses garnered from the rational choice theory of crime, in which a demolition increases the actual or perceived costs of committing a crime (by increasing the likelihood of being caught) and reduces the benefits of committing a crime (by removing property to steal) which therefore disincentivizes crime in the area of the demolition. The results may also support the broken windows theory which suggests that demolitions cause positive spillover effects onto nearby properties by removing disorder that breeds further disorder. However, which of these theories is directly driving the results is unknown. The positive spillover effects onto nearby block groups and the lack of evidence of individuals substituting one type of crime for another indicate that demolitions do have a positive net effect, at least in the larger neighborhood. Although Saginaw is a bit of an outlier in terms of its extreme depopulation and vacancy rates, it is not alone. Detroit, Michigan, Youngstown, Ohio, and many other cities in the rust belt and beyond are dealing with similar decline and depopulation. Vacant building demolitions may be one way to at least stem this population decline by reducing crime in the neighborhoods of the demolitions. However, while this paper provides evidence that demolitions do reduce crime, more research is needed to determine whether they are able to fully revert crime back to prevacancy levels. The social disorganization theory of crime suggests that a demolition will not reverse the increase in crime created by a vacancy's disruption to neighborhood cohesion. Therefore, policies and programs aimed at preventing vacancies in the first place may be even better at minimizing crime, depending upon the effectiveness of these programs and the costs of administering them. For example, the money used to demolish a vacant building could be given upstream to a family facing foreclosure and struggling to make their mortgage payments, perhaps preventing the vacancy in the first place. More formal cost–benefit analyses are needed to determine what combination of vacancy prevention and demolition is the best strategy.

ACKNOWLEDGMENTS Thank you to Dionissi Aliprantis, Quentin Brummet, Timothy Dunne, Ronald Fisher, Daniel Hartley, Timothy Hodge, Sung Hoon Kang, Robert Myers, Leslie Papke, Todd Pugatch, Francisca Richter, Eric Scorsone, Mark Skidmore, Gary Solon, Brian Stacy, and seminar participants at Michigan State University, the University of Michigan, and the Federal Reserve Bank of Cleveland. Special thanks to Scott Crofoot, Jeff Klopcic, Dan Sherman, and others from the City of Saginaw for providing the data. This work was supported in part by the Elton R. Smith Endowment in Food and Agricultural Policy at Michigan State University, The Center for Community Progress, and the Committee on the Status of Women in the Economics Profession.

1 In this paper, I use the term “vacant” to refer to blighted or dangerous buildings. Not all of the buildings that are demolished are technically vacant. Some still have residents living in them but are dangerous enough for the city to have the power to demolish.

In this paper, I use the term “vacant” to refer to blighted or dangerous buildings. Not all of the buildings that are demolished are technically vacant. Some still have residents living in them but are dangerous enough for the city to have the power to demolish. 2 Demolitions are also justified as a means of increasing surrounding property values, removing safety hazards, and attracting residents and businesses to the neighborhood.

Demolitions are also justified as a means of increasing surrounding property values, removing safety hazards, and attracting residents and businesses to the neighborhood. 3 For more serious offenses, Flint police officers use the county jail if there is space.

For more serious offenses, Flint police officers use the county jail if there is space. 4 For the remainder of the paper, I refer to reported crime as crime.

For the remainder of the paper, I refer to reported crime as crime. 5 The effect on property crime and violent crime do not add up to the effect on total crime since there are crimes that are neither violent or property included in total.

The effect on property crime and violent crime do not add up to the effect on total crime since there are crimes that are neither violent or property included in total. 6 For an excellent review of the crime literature, see Deller, Amiel, & Deller ( 2011

For an excellent review of the crime literature, see Deller, Amiel, & Deller ( 7 Since the approval of the General Property Tax Act in 1893, property taxes have been the main source of revenue for local governments in Michigan. However, the property tax structure was altered in 1994 by Proposal A which placed a constitutional cap on the growth of taxable value (TV). Since Proposal A was instated, the TV of a property has been allowed to increase only by the lesser of the rate of inflation or 5 percent until the property is transferred (not including additions or new construction). Historically, this value has been below the state equalized value (SEV) which has led to a general decline in property tax revenues as a proportion of property values. It was not anticipated that the SEV would begin to decrease and eventually fall below TV as has already occurred for some properties throughout the state. When this happens, the property tax paid by the owner follows the fall in SEV. In the short run, Proposal A may help to insulate local revenues from the declining home values. However, when house prices do stabilize and begin to increase, TV will be ratcheted downward and local unit fiscal capacity may not recover for years.

Since the approval of the General Property Tax Act in 1893, property taxes have been the main source of revenue for local governments in Michigan. However, the property tax structure was altered in 1994 by Proposal A which placed a constitutional cap on the growth of taxable value (TV). Since Proposal A was instated, the TV of a property has been allowed to increase only by the lesser of the rate of inflation or 5 percent until the property is transferred (not including additions or new construction). Historically, this value has been below the state equalized value (SEV) which has led to a general decline in property tax revenues as a proportion of property values. It was not anticipated that the SEV would begin to decrease and eventually fall below TV as has already occurred for some properties throughout the state. When this happens, the property tax paid by the owner follows the fall in SEV. In the short run, Proposal A may help to insulate local revenues from the declining home values. However, when house prices do stabilize and begin to increase, TV will be ratcheted downward and local unit fiscal capacity may not recover for years. 8 Once a house undergoes any type of fire it is immediately demolished. Because of the direct endogeneity caused by this policy, I cannot include demolitions due to arson in my analysis. There is anecdotal evidence to suggest that this demolition policy is at least in part causing arson in vacant buildings. Homeowners near vacant buildings have learned that if a vacant building undergoes arson it will be demolished more quickly, which neighbors perceive as a benefit. Therefore, even without the estimation issues caused by the endogeneity between demolitions and arson, it would be difficult to determine which incidences of arson were caused by vacant buildings themselves and which were caused by the structure of the demolition policy. In order to deal with this issue, I first remove all demolitions that were caused by arson, and all arsons that led to demolitions from my data set. Because there may still be a concern that some arsons were undertaken in order to induce a demolition or that some demolitions were caused by arsons that were not coded properly, I do a robustness check for endogeneity bias of all regressions by removing all arsons. The results are very similar and of the same sign and significance, suggesting that this concern is not a major issue.

Once a house undergoes any type of fire it is immediately demolished. Because of the direct endogeneity caused by this policy, I cannot include demolitions due to arson in my analysis. There is anecdotal evidence to suggest that this demolition policy is at least in part causing arson in vacant buildings. Homeowners near vacant buildings have learned that if a vacant building undergoes arson it will be demolished more quickly, which neighbors perceive as a benefit. Therefore, even without the estimation issues caused by the endogeneity between demolitions and arson, it would be difficult to determine which incidences of arson were caused by vacant buildings themselves and which were caused by the structure of the demolition policy. In order to deal with this issue, I first remove all demolitions that were caused by arson, and all arsons that led to demolitions from my data set. Because there may still be a concern that some arsons were undertaken in order to induce a demolition or that some demolitions were caused by arsons that were not coded properly, I do a robustness check for endogeneity bias of all regressions by removing all arsons. The results are very similar and of the same sign and significance, suggesting that this concern is not a major issue. 9 The City of Saginaw does not board up houses. Any boarding that occurs is either completed by the owner or by volunteers during an event that takes place each year before Halloween. If the city did systematically board up vacant houses it would likely affect the incidence of crime on that parcel.

The City of Saginaw does not board up houses. Any boarding that occurs is either completed by the owner or by volunteers during an event that takes place each year before Halloween. If the city did systematically board up vacant houses it would likely affect the incidence of crime on that parcel. 10 This excludes the 25 emergency demolitions that were caused by arson.

This excludes the 25 emergency demolitions that were caused by arson. 11 I use levels of crime as my dependent variable rather than the crime rates because no good time‐varying estimate of population per block exists in my data set. The only number that I could use is the number of occupied parcels on each block. However, this number is defined as the number of parcels with a building on it, not the number of parcels with a person living on it. Therefore, this number changes when a demolition occurs and would cause simultaneity in my estimation. In addition, an occupied parcel could have any number of residents on it from a single person home to a multifamily rental. Therefore, a crime rate based on this number would not only be inaccurate but would also cause bias in my results.

I use levels of crime as my dependent variable rather than the crime rates because no good time‐varying estimate of population per block exists in my data set. The only number that I could use is the number of occupied parcels on each block. However, this number is defined as the number of parcels with a building on it, not the number of parcels with a person living on it. Therefore, this number changes when a demolition occurs and would cause simultaneity in my estimation. In addition, an occupied parcel could have any number of residents on it from a single person home to a multifamily rental. Therefore, a crime rate based on this number would not only be inaccurate but would also cause bias in my results. 12 Fixed effects estimations in nonlinear models such as this one generally lead to inconsistent estimates. However, the Poisson distribution can be arbitrarily misspecified and any kind of serial correlation can be present and the fixed effects Poisson estimator is consistent under mild regularity conditions (Wooldridge, 2002 , estimates of β are consistent even if the mean does not equal the variance. Therefore, a fixed effects model is appropriate and over dispersion can be ignored.

Fixed effects estimations in nonlinear models such as this one generally lead to inconsistent estimates. However, the Poisson distribution can be arbitrarily misspecified and any kind of serial correlation can be present and the fixed effects Poisson estimator is consistent under mild regularity conditions (Wooldridge, , estimates of β are consistent even if the mean does not equal the variance. Therefore, a fixed effects model is appropriate and over dispersion can be ignored. 13 I also ran these models at the block level. Although this level allows me to control for a more detailed level of unobserved heterogeneity, results were less robust likely due to measurement error in the crime data at such a small geography. Police officers may report crime at a nearby location, or on the way to a call.

I also ran these models at the block level. Although this level allows me to control for a more detailed level of unobserved heterogeneity, results were less robust likely due to measurement error in the crime data at such a small geography. Police officers may report crime at a nearby location, or on the way to a call. 14 When the models are run without time trends included, demolitions are associated with an increase in crime. This implies that demolitions were undertaken on block groups with increasing time trends, which were biasing the results that did not include these time trends. I therefore include time trends in all of the subsequent models.

When the models are run without time trends included, demolitions are associated with an increase in crime. This implies that demolitions were undertaken on block groups with increasing time trends, which were biasing the results that did not include these time trends. I therefore include time trends in all of the subsequent models. 15 Note that the effect of demolitions on property and violent crime does not add up to the effect on total crime, since there are many crimes included in total crime that do not fall under property or violent crime.

Note that the effect of demolitions on property and violent crime does not add up to the effect on total crime, since there are many crimes included in total crime that do not fall under property or violent crime. 16 Model is robust to fewer lags and leads but begins to lose power and precision with additional ones.

Model is robust to fewer lags and leads but begins to lose power and precision with additional ones. 17 Although there appears to be a slight spike in violent crime immediately prior to a demolition, this effect is not statistically significant indicating that an Ashenfelter dip (or spike in this case) is not a concern (Ashenfelter & Card, 1985

Although there appears to be a slight spike in violent crime immediately prior to a demolition, this effect is not statistically significant indicating that an Ashenfelter dip (or spike in this case) is not a concern (Ashenfelter & Card, 18 To calculate these yearly level averages, I do a back of the envelope calculation by taking the semielasticity derived from the fixed effects Poisson coefficients, multiply that by the average number of crimes on a block group in a month, and multiply that times the number of months in a year.