NIGHT CRIMES: THE EFFECTS OF EVENING DAYLIGHT ON CRIMINAL ACTIVITY An Applied Project Presented to The Faculty of the Department of Economics Western Kentucky University Bowling Green, Kentucky By Lucas Taulbee August 2017
Abstract This study estimates the effect of an extra hour of evening daylight on criminal activity. Using data from the FBI, U.S. Census, BLS, and BEA on U.S. counties that are located close to a time zone line, between-effects estimators are used to determine the relationship between an extra hour of evening daylight and several crime variables. The findings suggest that an extra hour of evening daylight decreases overall violent crime, aggravated assaults, robberies, and arson.
1. Introduction In this paper, I estimate the relationship between evening daylight and criminal activity. Using data from the FBI, U.S. Census, BLS, and BEA, I find (1) An extra hour of evening daylight decreases total violent crime, aggravated assaults, robberies, and arson and (2) that year-round daylight saving time (DST) could result in substantial social cost savings. My findings are largely consistent with the literature in finding that an extra hour of evening daylight decreases certain types of criminal activity. My dataset is comprised of counties that border a different time zone, neighbors of those counties, and neighbors of neighbors. The choice of counties is used to exploit the hour difference in time between counties on either side of a time zone line. This allows for a natural experiment in which the only major difference in counties is the amount of daylight in the evening. The data on my dependent variables (total crime, total violent crime, total property crime, murder, rape, aggravated assault, robbery, burglary, larceny, motor vehicle theft, and arson) come from the FBI s Uniform Crime Reporting database. Data on controls come from the U.S. Census, BLS, and BEA. My findings suggest that an extra hour of evening daylight is associated with a decrease in violent crime. Property crime, however, is largely unaffected. The effect on violent crime seems to be largely driven by a decrease in aggravated assaults and robberies, with little effect on rape and murder. Despite property crime being largely unaffected by an extra hour of evening daylight, one specific category (arson) does seem to be reduced. Even small decreases in crime can result in substantial social cost savings. McCollister et. al. (2010) estimate that a single aggravated assault costs society over $100,000 dollars in tangible and intangible costs. Similarly, they estimate that robberies and arsons cost society $42,310 and $21,103 respectively. Small reductions in these crimes could potentially result in billions of dollars in annual social cost savings. This study is heavily influenced in spirit by Doleac and Sanders (2015) who used regression discontinuity and difference-in-difference approaches to estimate the effects of ambient light on criminal activity. They found that the shift to DST is associated with a 7% decrease in robberies and that the DST extension in 2007 resulted in $59 million in annual social cost savings solely from avoided robberies.
2. Literature review As is the case with much economic research, the foundations of the economics of crime can largely be attributed to the work of Gary S. Becker. Becker (1968) offers an economic analysis to develop optimal policies to combat criminal behavior. Among the factors affecting criminal activity, the probability that an offender will be caught performing an illegal act is a key part of the equation. Becker s work has relevance to my study for two reasons. The first is that his early work inspired much of the research that has been done on the economics of crime. The second, and more direct, reason is the contribution of this concept that the chances of an offender being caught may impact a potential criminal s decision making. It seems reasonable to think that a deterrent such as extra evening daylight may reduce the attractiveness of some criminal acts. To determine the effects of evening daylight on criminal activity, it is important to understand some of the other factors that affect crime. As proposed by Becker (1968), the probability of being caught, apprehended, and punished likely affects an individual s decision whether to commit a crime. In other words, individuals must compare the costs and benefits of criminal activities with the costs and benefits of legal activities to determine if criminal activity is optimal. Buonanno (2003) reviewed much of the literature on the economics of crime and determined that unemployment, education, inequality, social networks, and age are consistently important in explaining crime. In more recent work, Levitt (2004) attempted to explain why crime decreased in the 1990 s. Levitt attributes decreased crime in the 1990 s to four factors; increased incarceration, more police, the decline of crack, and legalized abortion. Levitt s empirical results do seem to suggest that a strong economy may decrease property crime and that changing demographics may have small effects on violent crime and property crime. Levitt s work is an important demonstration that policy decisions can affect crime. In terms of understanding the effects of daylight on criminal outcomes, there is little work that focuses specifically on this topic. Doleac and Sanders (2015) attempt to estimate the effects of ambient
light on criminal activity by taking advantage of shifts to and from DST and a 2007 extension of DST. Doleac and Sanders use both a regression discontinuity design and a difference-in-difference approach to estimate these effects. They find that the shift to DST decreases robberies by 7% and that the 2007 DST extension resulted in $59 million in annual social cost savings from avoided robberies. One other interesting result from the Doleac and Sanders study is that there seems to be a steep decline in robberies immediately following the Spring shift to DST, but that robbery rates begin to level back out within a few weeks. This could be suggestive of an effect caused by changing sleep patterns or an adjustment by criminals, figuring out how to continue committing robberies without being compromised by the extra daylight. Smith (2016) uses regression discontinuity to estimate the effect of DST on fatal vehicle crashes. Smith did find that the Spring shift to DST is associated with a 5 to 6.5% increase in fatal crash risk. However, after employing tests to determine if the increase is associated with shifting of ambient light or sleep deprivation, Smith concludes that it is indeed a change in sleep patterns that is responsible for the increase in fatal crash risk. Calandrillo and Buehler (2008) claim that a change to year-round DST would decrease criminal activity while also eliminating negative effects of Spring and Fall time changes, saving energy by reducing evening peak electricity loads, and saving lives by reducing pedestrian and motor vehicle fatalities. Kamstra et. al. (2000) elaborate on some of the negative effects of Spring and Fall time changes, specifically, the change in sleep patterns that arises as a result. Their findings suggest that the shift to daylight saving time results in a one-day loss in the United States of $31 billion on the NYSE, AMEX, and NASDAQ exchanges. There is certainly no consensus on many of the benefits of year-round DST however. Kotchen and Grant (2008) find that the rationale of saving energy may be questionable as results from their natural experiment in Indiana suggest that DST increases residential energy demand. Becker laid much of the groundwork for further study in the economics of crime. The works by Levitt and other contemporaries help to further form and explain the economic approach to studying crime. Many recent works seek to explain the effects of DST on economic outcomes as well as the general importance of the DST issue. There is a need for further study on the effects of DST and the effects of evening daylight on criminal activity.
3. Methodology This study seeks to estimate the effects of an extra hour of evening daylight on crime. By using only counties that are near a time zone line, we hope to observe counties that are similar in all but one aspect. Counties directly east of a time zone line should effectively experience one extra hour of evening daylight relative to their western neighbors. By taking advantage of this natural experiment, it should be possible to estimate the long run effects of evening daylight on crime. This approach also serves to remove any influence that changing sleep patterns may have on criminal activity. To estimate the relationship between evening daylight and crime, I use panel data betweeneffects. Between-effects estimation is a regression on group means. For each variable, the mean value for each county is used as an observation rather than using each year s value. The advantage of doing this is that it allows us to use panel data to estimate the effect of a time-invariant variable of interest on a time-varying dependent variable. The disadvantage to using between-effects rather than fixed-effects or random-effects is that it does not allow us to fully exploit the time variation in the data. Generically, the equation to be estimated appears as follows: Crime i = α + β 1 East + β n Controls i + ε i (1) where Crime i is the average crime of i-th county in the sample, East is the group dummy, and Controls i are the average of controls of i-th county. Controls will include income, unemployment, population density, and county distance from a time zone line. To fully determine the effects of crime, a similar equation will be run for multiple dependent variables: total crime, total violent crime, total property crime, rape, murder, aggravated assault, robbery, burglary, larceny, motor vehicle theft, and arson. This will determine if evening daylight has an effect only on certain types of crime. The approach employed here contrasts the approach of Doleac and Sanders (2015) who used both a regression discontinuity and a difference-in-difference approach to estimate the effects of ambient light on crime. Ideally, this study hopes to find results that are consistent with the findings of Doleac and Sanders (2015) despite using a vastly different methodology.
4. Data To estimate the effects of evening daylight on crime, this study utilizes county-level data on crimes reported, population density, per capita income, and unemployment rates. The dataset is compiled for U.S. counties that border a different time zone, neighbors of those counties, and neighbors of neighbors. Counties from the state of Arizona were not used because Arizona does not observe DST. The few counties that fall within two time zones were also excluded from the study. Counties are divided into two groups: those that are directly east of a time zone line and those that are directly west of a time zone line. Roughly half of the counties fall within each group. The dataset used for the study contains data for 656 counties over a period of 14 years (1999-2012) for a total of 8,859 observations. The data for crimes reported come from the Federal Bureau of Investigation s Uniform Crime Reporting database. The crime data include numbers on total crime reported as well as detailed breakdown of crimes reported by type. The two major types of crime are violent crime, which consists of murder, rape, robbery, and aggravated assault; and property crime, which consists of burglary, larceny, motor vehicle theft, and arson. These data also include a variable for coverage indication which informs us how much of the data was reported by local agencies and how much was imputed. The mean of this variable is 96.67 percent with standard deviation of 10.98 percent. In raw form, most of the crime variables have Poisson-like distributions, with a large concentration of 0 and 1 observations. To suit the method used in this study, these variables are transformed by scaling them to crimes reported per 1,000 people. This transformation allows the variables to take on a distribution closely resembling a normal distribution. The population density numbers are estimates derived from the 2010 U.S. census. Per capita income and unemployment rates were obtained from the Bureau of Economic Analysis and Bureau of Labor Statistics respectively. For this study, population density and per capita income were transformed using natural logs. These transformations allow us to more easily estimate the effects of percentage changes in the variables on our crime variables while also allowing the distributions of these variables to more closely resemble a normal distribution.
Descriptive statistics can be found in Table 1 below. Table 1 Descriptive Statistics VARIABLES N Mean SD Min Max East of Timeline 8,859 0.500 0.500 0 1 Density per Square Mile of Land Area - Population 8,831 126.6 409.8 0.100 5,495 Per Capita Income 8,848 29,099 8,687 9,242 121,242 Unemployment Rate 8,848 6.153 2.967 1 30.30 Crimes Reported per Thousand People 8,859 23.43 16.99 0 350 Violent Crimes Reported per Thousand People 8,859 2.418 2.524 0 77.26 Property Crimes Reported per Thousand People 8,859 21.01 15.40 0 333.3 Murders Reported per Thousand People 8,859 0.0294 0.0693 0 1.404 Rapes Reported per Thousand People 8,859 0.241 0.271 0 3.875 Robberies Reported per Thousand People 8,859 0.306 0.608 0 13.84 Aggravated Assaults Reported per Thousand People 8,859 1.842 2.124 0 77.26 Burglaries Reported per Thousand People 8,859 5.059 3.829 0 83.33 Larceny Reported per Thousand People 8,859 14.40 11.39 0 250 Motor Vehicle Thefts Reported per Thousand People 8,859 1.412 1.542 0 19.22 Arson Reported per Thousand People 8,859 0.136 0.290 0 16.67 Number of Counties 656 656 656 656 656 The descriptive statistics shown above tell us that half of the observations fall within each group (east and west). In addition, we see that some of the variables have a standard deviation that is larger than the mean, even after being scaled. This is a side-effect of many of the crime variables being substantially skewed to the right. The skewness statistics can be found in Table A-9. The kurtosis statistics (also found in Table A-9) show that the variables have heavy tails. For estimation purposes, the density and income variable are transformed using natural logarithms. Transforming these two variables with logs serves to make distributions closer to normal and allows for easier interpretation of coefficients.
5. Results Table 2 Regression Results (Benchmark Models) VARIABLES (1) (2) (3) Violent Crimes Reported per Thousand People Crimes Reported per Thousand People Property Crimes Reported per Thousand People East of Timeline -0.526-0.443*** -0.0828 (1.147) (0.150) (1.039) Constant 23.30*** 2.600*** 20.70*** (0.816) (0.107) (0.739) Observations 8,859 8,859 8,859 R-squared 0.000 0.013 0.000 Number of Counties 656 656 656 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Results for benchmark models on the 3 primary dependent variables are presented above in table 2. These models only estimate the effect of being east of a time zone line on crime and contain no controls. Column one estimates the effect of being east of a time zone line on total crime, column two estimates the effect on violent crime, and column 3 estimates the effect on property crime. While these models have very little explanatory power, column 2 does demonstrate that we may want to further investigate the effects of ambient light on violent crime. These results may suggest that being located directly east of a time zone line (and having an extra hour of evening daylight) is correlated with.44 fewer violent crimes being reported per thousand people in each county. In other words, counties with an extra hour of evening daylight have, on average, 17% fewer violent crimes reported annually.
Table 3 Regression Results (Significant Findings) VARIABLES (1) (2) (3) (4) Aggravated Assaults Robberies Reported per Reported per Thousand People Thousand People Violent Crimes Reported per Thousand People Arson Reported per Thousand People East of Timeline -0.457*** -0.395*** -0.0595* -0.0206* (0.141) (0.115) (0.0344) (0.0114) Log of Population per Square Mile 0.310*** 0.145*** 0.142*** 0.00364 (0.0395) (0.0323) (0.00969) (0.00321) Log of Per Capita Income 0.0160-0.445 0.371*** 0.148*** (0.411) (0.335) (0.101) (0.0334) Unemployment Rate 0.120*** 0.0967*** 0.0152 0.0214*** (0.0396) (0.0323) (0.00969) (0.00321) County Distance from Timeline = 2 0.147 0.105 0.0597 0.0303** (0.176) (0.144) (0.0431) (0.0143) County Distance from Timeline = 3-0.171-0.0878-0.0634 0.00249 (0.176) (0.143) (0.0430) (0.0143) Constant 0.773 5.534-4.000*** -1.523*** (4.300) (3.507) (1.053) (0.349) Observations 8,820 8,820 8,820 8,820 R-squared 0.155 0.095 0.326 0.105 Number of Counties 653 653 653 653 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The results in Table 3 above are the significant findings derived from the models that include all controls. The models estimate with some consistency that population density and unemployment are associated with more crime. These results also suggest that income may increase certain types of crime while having no effect on others. One of the interesting results here is that an extra hour of daylight is associated with an estimated 18% fewer violent crimes reported annually, on average. Since there was no significant effect on rape or murder, this result is likely driven by the effects of daylight on aggravated assaults and robberies. An extra hour of daylight is associated with approximately 20% fewer aggravated assaults reported on average and possibly around 18% fewer robberies reported on average. Another interesting result is that an extra hour of daylight may also be associated with an estimated 14% fewer annual reports of arson on average. Full results for all dependent variables can
be found in Tables A-1 through A-3. Percentage calculations can be found in tables A-7 and A-8. In tables A-10 and A-11, percentage calculations can also be found for difference from the median of the control group. The results from the models including controls are consistent with the baseline models in terms of the sign of the relationship. It is demonstrated across the board that the counties in the treatment group experience lower rates of violent crime, aggravated assault, robbery, and arson. The statistical significance of the relationships is fairly consistent as well. Violent crime and aggravated assault maintain strong significance when the controls are added while robbery gains some significance and arson loses some. For sensitivity analysis, these same models were run using only counties with populations between 10,000 and 100,000. The results for this analysis can be found in Table A-4 (total crime, total violent crime, and total property crime), Table A-5 (rape, murder, robbery, and aggravated assault), and Table A-6 (arson, larceny, motor vehicle theft, and burglary). This did not largely affect the significant results on the variable of interest, however, it did flip the sign on some of the controls. In many cases where unemployment would increase crime, it seems to be associated with a decrease in crime for this sample. In many cases where income would have no effect or be associated with an increase in crime, it seems to be associated with a decrease in crime. The inconsistence of the control relationships can possibly be explained by inequality. Many of the counties with high unemployment are rural counties that have consistently had high unemployment. In these cases, it seems reasonable that unemployment would have a lesser impact on most types of criminal activity. A reasonable explanation for income increasing crime in larger counties could be that the incomes of the highest earners are increasing while the incomes of others are staying the same. Another explanation is that in the case of robbery and burglary, more crimes may be committed against those with higher incomes because they have more to take.
6. Concusion This study finds that an extra hour of evening daylight is associated with an estimated 18% fewer total violent crimes reported annually, on average. In addition, these results suggest that an extra hour of evening daylight is associated with an approximate 20% fewer aggravated assaults, 18% fewer robberies, and 14% fewer arsons reported on average. At the very least, the study does not find that an extra hour of evening daylight is associated with an increase in any type of crime. The findings of this study contribute to existing literature that suggests ambient light decreases criminal activity. The results of this study can also contribute to the case for year-round daylight saving time. This research could be extended by finding a way to include more counties. This could potentially be done by collecting average sunset times for each county which would allow for the variable of interest to be a continuous variable that will allow for more variation across counties. Finding a way to estimate the distance of each U.S. county from time zone lines may also be useful in selecting counties to include in the estimating dataset. This study is limited by the somewhat arbitrary nature of choosing counties. It could be interesting to see how continuous distance from a time zone line may affect the results. As with any economic study, there are a lot of data limitations that restricted this study. For example, it would have been helpful to have data on age, education, sex, ethnicity, and inequality to improve the reliability of estimates and ensure that there isn t an omitted variable bias problem. However, such data is not reliably available for many these counties or for many of these years. It is worth noting that the distributions of the dependent variables could have some effect on the results. The major issue is the large number of zero observations for the dependent variables. The zeros rule out the possibility of using logarithms to transform the variables, which would allow for easy percentage change interpretations. It also inflates the likelihood of a zero occurring when the variables are scaled per thousand people. This is a problem that could be better addressed using data on more counties and by enabling the possibility of fixed-effects and/or random-effects estimation. An analysis of cost could be very useful in determining how helpful these results are in settling the debate over DST. It is difficult to determine the exact cost imposed by a particular crime as there are many factors that contribute to the cost, such as; victim costs, criminal justice system costs, crime career costs, and intangible costs (McCollister et. al. 2010). It seems to be a consensus, however, that crime is
quite costly. For example, McCollister et. al. (2010) estimate that an aggravated assault carries a social cost of $107,020 on average, a robbery carries a social cost of $42,310, and an arson carries a social cost of $21,103. Doleac and Sanders (2015) estimated that the 2007 extension of DST resulted in $59 billion in annual social cost savings. Given that my estimates of the effect of ambient light on criminal activity are quite large, it s not unreasonable to conclude that a change to year-round DST would result in substantial social cost savings.
Literature Cited Becker, G. S. (1968). Crime and Punishment: An Economic Approach. Journal of Political Economy, 76(2), 169-217. Buonanno, P. (2003). The Socioeconomic Determinants of Crime. A Review of the Literature. University of Milan - Bicocca Working Paper Series, 63. Calandrillo, S. P., & Buehler, D. E. (2008). Time Well Spent: An Economic Analysis of Daylight Saving Time Legislation. Wake Forest Law Review, 43, 45-92. Doleac, J. L., & Sanders, N. J. (2015). Under the Cover of Darkness: How Ambient Light Influences Criminal Activity. Review of Economics and Statistics, 97(5), 1093-1103. Kamstra, M. J., Kramer, L. A., & Levi, M. D. (2000). Losing Sleep at the Market: The Daylight Saving Anomoly. The American Economic Review, 1005-1011. Kotchen, M., & Grant, L. (2008). Does Daylight Saving Time Save Energy? Evidence from a Natural Experiment in Indiana. NBER Working Paper Series. Levitt, S. D. (2004). Understanding Why Crime Fell in the 1990s: Four Factors that Explain the Decline and Six that Do Not. Journal of Economic Perspectives, 18(1), 163-190. 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. Smith, A. C. (2016). Spring Forward at Your Own Risk: Daylight Saving Time and Fatal Vehicle Crashes. American Economic Journal: Applied Economics, 8(2), 65-91.
Appendix Table A-1 Results on Primary Dependent Variables VARIABLES (1) (2) (3) Violent Crimes Reported per Thousand People Crimes Reported per Thousand People Property Crimes Reported per Thousand People East of Timeline -0.779-0.457*** -0.322 (0.968) (0.141) (0.874) Log of Population per Square Mile 3.747*** 0.310*** 3.436*** (0.272) (0.0395) (0.246) Log of Per Capita Income 7.316*** 0.0160 7.300*** (2.828) (0.411) (2.552) Unemployment Rate 0.834*** 0.120*** 0.714*** (0.272) (0.0396) (0.246) County Distance from Timeline = 2 0.654 0.147 0.508 (1.212) (0.176) (1.094) County Distance from Timeline = 3-0.863-0.171-0.692 (1.209) (0.176) (1.091) Constant -67.96** 0.773-68.74** (29.59) (4.300) (26.70) Observations 8,820 8,820 8,820 R-squared 0.305 0.155 0.311 Number of Counties 653 653 653 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The above table shows results for the 3 primary dependent variables; total crime, total violent crime, and total property crime. These models use all observations for which data is available and all controls.
Table A-2 Results on Violent Crime VARIABLES (1) (2) (3) (4) Murders Reported Rapes Reported per Thousand per Thousand People People Aggravated Assaults Reported per Thousand People Robberies Reported per Thousand People East of Timeline -0.395*** -0.00220-0.000593-0.0595* (0.115) (0.00256) (0.0148) (0.0344) Log of Population per Square Mile 0.145*** 0.00387*** 0.0190*** 0.142*** (0.0323) (0.000721) (0.00416) (0.00969) Log of Per Capita Income -0.445-0.00655 0.0966** 0.371*** (0.335) (0.00749) (0.0432) (0.101) Unemployment Rate 0.0967*** 0.000659 0.00716* 0.0152 (0.0323) (0.000721) (0.00416) (0.00969) County Distance from Timeline = 2 0.105 0.000412-0.0181 0.0597 (0.144) (0.00321) (0.0185) (0.0431) County Distance from Timeline = 3-0.0878-0.00682** -0.0132-0.0634 (0.143) (0.00320) (0.0185) (0.0430) Constant 5.534 0.0840-0.845* -4.000*** (3.507) (0.0784) (0.452) (1.053) Observations 8,820 8,820 8,820 8,820 R-squared 0.095 0.066 0.059 0.326 Number of Counties 653 653 653 653 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The above table shows results for the 4 types of violent crime; aggravated assault, murder, rape, and robbery. These models use all observations for which data is available and all controls.
Table A-3 Results on Property Crime VARIABLES (1) (2) (3) (4) Burglaries Larceny Reported Reported per per Thousand Thousand People People Arson Reported per Thousand People Motor Vehicle Thefts Reported per Thousand People East of Timeline -0.0206* -0.338 0.140-0.102 (0.0114) (0.221) (0.642) (0.0890) Log of Population per Square Mile 0.00364 0.658*** 2.453*** 0.322*** (0.00321) (0.0622) (0.181) (0.0250) Log of Per Capita Income 0.148*** 0.0699 6.170*** 0.911*** (0.0334) (0.646) (1.876) (0.260) Unemployment Rate 0.0214*** 0.271*** 0.325* 0.0966*** (0.00321) (0.0622) (0.181) (0.0250) County Distance from Timeline = 2 0.0303** 0.113 0.358 0.00677 (0.0143) (0.277) (0.804) (0.111) County Distance from Timeline = 3 0.00249-0.0984-0.423-0.173 (0.0143) (0.276) (0.802) (0.111) Constant -1.523*** 0.773-58.57*** -9.412*** (0.349) (6.760) (19.63) (2.720) Observations 8,820 8,820 8,820 8,820 R-squared 0.105 0.237 0.291 0.296 Number of Counties 653 653 653 653 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The above table shows results for the 4 types of property crime; arson, burglary, larceny, and motor vehicle theft. These models use all observations for which data is available and all controls.
Table A-4 Sensitivity Analysis on Primary Dependent Variables VARIABLES (1) (2) (3) Violent Crimes Reported per Thousand People Crimes Reported per Thousand People Property Crimes Reported per Thousand People East of Timeline -0.0115-0.477** 0.466 (1.415) (0.206) (1.275) Log of Population per Square Mile 2.268*** -0.0142 2.282*** (0.586) (0.0852) (0.528) Log of Per Capita Income -2.519-1.747*** -0.772 (4.409) (0.641) (3.973) Unemployment Rate -1.101** -0.131** -0.970** (0.433) (0.0630) (0.391) County Distance from Timeline = 2 0.411 0.0274 0.384 (1.776) (0.258) (1.600) County Distance from Timeline = 3-0.597-0.0248-0.572 (1.693) (0.246) (1.525) Constant 50.40 21.51*** 28.89 (46.14) (6.707) (41.57) Observations 4,431 4,431 4,431 R-squared 0.063 0.038 0.075 Number of Counties 362 362 362 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The above table shows results from a sensitivity analysis for the 3 primary dependent variables; total crime, total violent crime, and total property crime. These models are estimated using only the counties with populations between 10,000 and 100,000 in order to diminish the influence of very small counties and very large counties.
Table A-5 Sensitivity Analysis on Violent Crime VARIABLES (1) (2) (3) (4) Murders Rapes Reported Reported per per Thousand Thousand People People Aggravated Assaults Reported per Thousand People Robberies Reported per Thousand People East of Timeline -0.433** -0.00236 0.0143-0.0565 (0.175) (0.00315) (0.0189) (0.0375) Log of Population per Square Mile -0.104 0.00109-0.0168** 0.106*** (0.0725) (0.00130) (0.00783) (0.0155) Log of Per Capita Income -1.617*** -0.0118 0.122** -0.240** (0.546) (0.00980) (0.0589) (0.117) Unemployment Rate -0.0973* -0.000922-0.0103* -0.0226** (0.0536) (0.000964) (0.00579) (0.0115) County Distance from Timeline = 2 0.00356-0.00316-0.0134 0.0404 (0.220) (0.00395) (0.0237) (0.0470) County Distance from Timeline = 3 0.0211-0.00622* -0.0226-0.0171 (0.209) (0.00376) (0.0226) (0.0448) Constant 19.70*** 0.158-0.847 2.498** (5.709) (0.103) (0.617) (1.221) Observations 4,431 4,431 4,431 4,431 R-squared 0.051 0.016 0.045 0.133 Number of Counties 362 362 362 362 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The above table shows results from the sensitivity analysis for the 4 types of violent crime; aggravated assault, murder, rape, and robbery. These models use only the counties with populations between 10,000 and 100,000.
Table A-6 Sensitivity Analysis on Property Crime VARIABLES (1) (2) (3) (4) Burglaries Larceny Reported Reported per per Thousand Thousand People People Arson Reported per Thousand People Motor Vehicle Thefts Reported per Thousand People East of Timeline -0.0200-0.491 0.892 0.0843 (0.0163) (0.320) (0.945) (0.0934) Log of Population per Square Mile -0.00799 0.515*** 1.578*** 0.197*** (0.00674) (0.133) (0.391) (0.0387) Log of Per Capita Income 0.0881* -2.457** 2.345-0.747** (0.0507) (0.998) (2.945) (0.291) Unemployment Rate 0.00747-0.0309-0.875*** -0.0714** (0.00498) (0.0981) (0.290) (0.0286) County Distance from Timeline = 2 0.0499** 0.234 0.111-0.0111 (0.0204) (0.402) (1.186) (0.117) County Distance from Timeline = 3-0.00370-0.0461-0.372-0.150 (0.0195) (0.383) (1.131) (0.112) Constant -0.788 29.15*** -8.235 8.765*** (0.531) (10.44) (30.82) (3.046) Observations 4,431 4,431 4,431 4,431 R-squared 0.042 0.058 0.091 0.103 Number of Counties 362 362 362 362 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The above table shows results from a sensitivity analysis for the 4 types of property crime; arson, burglary, larceny, and motor vehicle theft. These models use only the counties with populations between 10,000 and 100,000.
Table A-7 Percentage Calculations for Difference from Mean (Baseline) Variable Mean of Control Group Treatment Coefficient Percent Difference Significant? Violent Crime 2.6-0.443-17.04% Yes Aggravated Assault 2.013-0.393-19.52% Yes Robbery 0.322-0.048-14.91% No Arson 0.147-0.0279-18.98% Yes Table A-8 Percentage Calculations for Difference from Mean (Control Models) Variable Mean of Control Group Treatment Coefficient Percent Difference Significant? Violent Crime 2.6-0.457-17.58% Yes Aggravated Assault 2.013-0.395-19.62% Yes Robbery 0.322-0.0595-18.48% Yes Arson 0.147-0.0206-14.01% Yes
Table A-9 Skewness, Kurtosis, and Median VARIABLES N Skewness Kurtosis Median East of Timeline 8,859-0.000226 1.000 1 Density per Square Mile of Land Area - Population 8,831 7.776 78.06 29.80 Per Capita Income 8,848 2.100 12.53 27,574 Unemployment Rate 8,848 1.287 5.749 5.400 Crimes Reported per Thousand People 8,859 2.964 36.63 20.54 Violent Crimes Reported per Thousand People 8,859 7.012 149.2 1.832 Property Crimes Reported per Thousand People 8,859 3.061 40.56 18.46 Murders Reported per Thousand People 8,859 6.625 78.44 0 Rapes Reported per Thousand People 8,859 2.565 17.34 0.180 Robberies Reported per Thousand People 8,859 5.948 74.31 0.103 Aggravated Assaults Reported per Thousand People 8,859 9.186 239.0 1.335 Burglaries Reported per Thousand People 8,859 2.677 30.43 4.341 Larceny Reported per Thousand People 8,859 3.663 53.80 12.66 Motor Vehicle Thefts Reported per Thousand People 8,859 3.168 19.57 1.026 Arson Reported per Thousand People 8,859 23.82 1,224 0.0617 Number of Counties 656 656 656 656
Table A-10 - Percentage Calculations for Difference from Median (Baseline) Variable Median of Control Group Coefficient Percent Difference Significant? Violent Crime 2.09-0.443-21% Yes Aggravated Assault 1.61-0.393-24% Yes Robbery 0.11-0.048-44% No Arson 0.07-0.0279-40% Yes Table A-11 Percentage Calculations for Difference from Median (Control Models) Variable Median of Control Group Coefficient Percent Difference Significant? Violent Crime 2.09 Aggravated Assault 1.61 Robbery 0.11 Arson 0.07-0.457-0.395-0.0595-0.0206-22% Yes -25% Yes -54% Yes -29% Yes