key: cord-0869428-c3xcmkuc
authors: McLay, M. M.
title: When "Shelter-in-Place" Isn't Shelter That's Safe: A Rapid Analysis of Domestic Violence Case Differences During the COVID-19 Pandemic and Stay-at-Home Orders
date: 2020-06-03
journal: nan
DOI: 10.1101/2020.05.29.20117366
sha: 3fb0fbea418a1be16d4d6eeac9cfd3791f824f7e
doc_id: 869428
cord_uid: c3xcmkuc

Purpose: This study explored the COVID-19 pandemic's impacts on domestic violence (DV) with the following research questions: 1) Did DV occurring during the pandemic differ on certain variables from cases occurring on a typical day the previous year? 2) Did DV occurring after the implementation of shelter-in-place orders differ (on these same variables) from cases occurring prior to shelter-in-place orders? Methods: Two logistic regression models were developed to predict DV case differences before and during the pandemic. DV reports (N=4618) were collected from the Chicago Police Department. Cases from March 2019 and March 2020 were analyzed based on multiple variables. One model was set to predict case differences since the pandemic began, and another model was set to predict case differences during the shelter-in-place period later that month. Results: Both models were significant with multiple significant predictors. During the pandemic period, cases with arrests were 3% less likely to have occurred, and cases at residential locations were 22% more likely to have occurred. During the shelter-in-place period, cases at residential locations were 64% more likely to have occurred, and cases with child victims were 67% less likely to have occurred. Conclusions: This study offers a rapid analysis of DV case differences since the pandemic and shelter-in-place began. Additional variables and data sources could improve model explanatory power. Research, policy, and practice in this area must pivot to focus on protecting children whose access to mandated reporters has decreased and moving victims out of dangerous living situations into safe spaces.

different time points, utilizing data from a large U.S. city before and during the height of key policy changes in March 2020. This study analyzed relationships between the timeframe of domestic violence occurrence and a number of key variables. This study poses the following research questions:

1. Will domestic violence that occurred during the COVID-19 pandemic in March 2020

have key differences in the presence of a sex offense, weapon use, resulting arrest, presence of child victims, and residential locationality than domestic violence occurring on a typical day in March 2019? 2. Will domestic violence that occurred after the implementation of COVID-19 shelter-inplace orders on March 21, 2020 have key differences in these same variables than domestic violence occurring during the pandemic but prior to shelter-in-place orders?

The study's sample is comprised of police reports from the Chicago Police Department (CPD). Reports on cases that occurred during the months of March 2019 and March 2020 were utilized. Data was downloaded from the Chicago Data Portal (2020). March 2019 was used as a "usual circumstances" comparison to the March 2020 calls, which occurred toward the beginning of the COVID-19 pandemic hitting this city. It is also the month in which a shelter-in-place order was put into effect (Petrella, St. Clair, Johnson, & Pratt, 2020) , which was also a point of analysis.

There were 19,747 police reports documented by CPD on March 2019 cases, and there were 15,852 reports on March 2020 cases. The data was then narrowed to any reports considered domestic incidents, for a total of 6,560 reports. Data was then imported from a CSV file into a . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https: //doi.org/10.1101 //doi.org/10. /2020 Stata data file. The sample was then reduced only to reports involving some kind of physical or sexual violence. These offenses included any kind of assault, any kind of battery, homicide, criminal sexual assault, any sex offense, and physical or sexual offenses against children. The final sample consisted of 4,618 police reports. There was no missing data in the entire sample.

However, this dataset did not include any demographic information about victims or perpetrators; information was limited to basic descriptors of the case, its crime code and characteristics, location data, and basic date/time information.

Two dependent variables were analyzed: occurrence of domestic violence during the pandemic or not, and occurrence of domestic violence during shelter-in-place order or not. These two dependent variables, as well as several independent variables, are described as follows:

Occurrence of domestic violence during COVID-19 pandemic. One dependent variable for this study is whether or not a domestic violence case occurred during the COVID-19 pandemic. For this study's purposes, March 2019 and March 2020 police calls were the two groups analyzed and represented the dummy coding for this variable, with March 2020 cases being the group of interest and March 2019 being the reference group. This coding reflects the study's purpose of determining whether or not there are significant differences in the composition of domestic violence cases before and during the COVID-19 pandemic. To create this variable, any report for domestic violence happening between March 1-31, 2020, was coded as "pandemic" = "1." Any report for domestic violence happening between March 1-31, 2019, was coded as "pandemic" = "0." Occurrence of domestic violence during shelter-in-place order. The other dependent variable for this study is whether or not a domestic violence case occurring during the pandemic . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 3, 2020. . https: //doi.org/10.1101 //doi.org/10. /2020 happened before or after the shelter-in-place order for the city of Chicago took effect, as this order placed additional constraints on individual's interactions and whereabouts. The shelter-inplace, or "stay at home" order went into effect on March 21, 2020 , at 5:00pm (Petrella, St. Clair, Johnson, & Pratt, 2020 . Cases occurring at or after this time were the group of interest, and cases before this time (but still during March 2020) were the reference group. This coding reflects the study's purpose of determining whether or not there are significant differences in the composition of domestic violence cases before and during shelter-in-place. To create this variable, any report for domestic violence happening between March 21 at 5:00pm and March 31 at 11:59pm, was coded as "stayhome" = "1." Any report for domestic violence happening between March 1 at 12:00am and March 21 at 4:59pm was coded as "stayhome" = "0."

Presence of sex crime. The presence of a sex crime during domestic violence cases was one independent variable analyzed. For the purposes of this research, sex crime was operationalized as any of the following crimes: criminal sexual assault, sex offense, and any sexual offenses against children. Data was also manually searched in Excel to ensure that no additional sex crimes were missed. A variable called "sexcrime" was created, and any case containing one of these elements was coded as "1," and any case that did not contain a sex crime was coded as "0." Use of weapon. The use of a weapon during domestic violence cases was another independent variable contributing to these analyses. In CPD records, all cases involving weapons were designated as "aggravated" per the criminal code. Data was also manually searched in Excel to ensure that no additional crimes with weapons were missed. A variable called "weapon" was created, and any case involving the use of a weapon was coded as "1," and any case that did not was coded as "0." . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10.1101/2020.05.29.20117366 doi: medRxiv preprint Whether or not arrest was made. The issuance of an arrest during domestic violence cases was another independent variable contributing to these analyses. In CPD records, all cases in which an arrest was made were coded as such. A variable called "arrest" was created, and any case in which an arrest was made was coded as "1," and any case without was coded as "0." Presence of child victim(s). The presence of one or more child victims during domestic violence cases was another independent variable contributing to these analyses. In CPD records, cases involving children were categorized as "offenses against children." Data was also manually searched in Excel to ensure that no crimes with child victims were missed, and a few crimes in which child victims were referenced in descriptions, but not categorized under "offenses against children," were manually recoded under this category. A variable called "child" was created, and any case involving a child victim was coded as "1." All cases without reported child victims were coded as "0." An interaction term was also created between presence of child victims and presence of sex crimes.

A final independent variable was location, or more specifically, whether or not the violence took place at someone's residence. For the purposes of this research, residence was operationalized as any of the following locations: apartment, residence, nursing home facility, or the grounds, parking lot, or yard of any of these locations. Hotels were not included as residences. Data was manually searched to ensure that all residential locations were coded as such. A variable called "residence" was located, and any case taking place at one of these residences was coded as "1." All cases taking place elsewhere were coded as "0."

Logistic regression analysis was chosen as the primary data analysis tool for this study, as a number of independent variables were of interest, thus justifying the need for a multivariate . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 analysis. Utilizing a maximum likelihood estimator, logistic regression selects coefficients that maximize the probability of reproducing the sample data. Since two "whether or not" situations are being examined, it makes sense to analyze them in terms of probabilities or likelihood of each event occurring. The study offers two models with dichotomous dependent variables: one predicts whether or not a domestic violence case took place during the COVID-19 pandemic, and the other predicts whether or not a domestic violence case took place after the shelter-in-place order went into effect. While only five independent variables were available for the models due to the unique circumstances of the COVID-19 pandemic and the limited data available as a result, there are still enough variables to justify the use of logistic regression. Each model utilized in the study and their formal specifications are outlined as follows.

variables. The first logistic regression model features occurrence of domestic violence before or during the COVID-19 pandemic regressed on the following independent variables: presence of a sex crime, weapon use, resulting arrest, presence of child victim(s), and residential locationality.

Under this model, it is expected that, all other things being equal: when a sex crime is present, probability of domestic violence occurring during the COVID-19 pandemic changes; when a weapon is used, probability of domestic violence occurring during this pandemic changes; when an arrest is made, probability of domestic violence occurring during this pandemic changes; when a child victim is present, probability of domestic violence occurring during this pandemic changes; and when the location of the violence is a residence, probability of domestic violence occurring during this pandemic changes. The intercept in this model is the predicted log odds of domestic violence's occurrence before or during the COVID-19 pandemic when there is no sex crime present, no weapon use, no resulting arrest, no child victim(s), and the location of the . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10.1101/2020.05.29.20117366 doi: medRxiv preprint violence was not a residence. The formal model specification is as follows:

Two-tailed significance testing was performed on each regression coefficient, and odds ratios and their confidence intervals were reported. Two-tailed testing was chosen due to the unprecedented nature of the COVID-19 pandemic and its yet unknown effects on domestic violence. Given the exploratory nature of this study, this test is appropriate. Multicollinearity and influential data checks were performed to ensure these assumptions of logistic regression were met. Joint hypothesis testing and interaction effects were explored based on various bivariate analyses.

Model 2: Occurrence of domestic violence during shelter-in-place order, regressed on multiple variables. The second logistic regression model features occurrence of domestic violence before or during the shelter-in-place order regressed on the same independent variables as the first model: presence of a sex crime, weapon use, resulting arrest, presence of child victim(s), and residential locationality. Under this model, it is expected that, all other things being equal: when a sex crime is present, probability of domestic violence occurring during shelter-in-place changes; when a weapon is used, probability of domestic violence occurring during shelter-in-place changes; when an arrest is made, probability of domestic violence occurring during shelter-in-place changes; when a child victim is present, probability of domestic violence occurring during shelter-in-place changes; and when the location of the violence is a residence, probability of domestic violence occurring during shelter-in-place changes. The intercept in this model is the predicted log odds of domestic violence's occurrence before or during the shelter-in-place order when there is no sex crime present, no weapon use, no resulting arrest, no child victim(s), and the location of the violence was not a residence. The formal model . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10.1101/2020.05.29.20117366 doi: medRxiv preprint specification is below:

Two-tailed significance testing was performed on each regression coefficient, and odds ratios and their confidence intervals were reported. Two-tailed testing was chosen because of the unprecedented nature of the shelter-in-place orders and their yet unknown effects on domestic violence. Given the exploratory nature of this study, two-tailed significant tests are appropriate.

Multicollinearity and influential data checks were performed to ensure these assumptions of logistic regression were met. Similar to the first model, joint hypothesis testing using the likelihood ratio test examined whether or not weapon use and resulting arrest jointly contribute to the model fit. Also similarly, joint hypothesis testing and interaction effects were explored based on various bivariate analyses.

The study's dataset of March 2019 and March 2020 domestic violence cases contains 4,618 records, with complete information on sex offenses, weapon use, arrests made, child victims, and crime locations. The sample is large enough to meet the assumptions of logistic regression. All variables are dichotomous. Descriptive statistics are reported for March 2019 and March 2020 separately and in aggregate (Table 1 ).

In terms of total domestic violence cases, 51.26% (n=2367) of the sample's cases occurred in March 2019 and 48.74% (n=2251) took place in March 2020. The number of total cases was slightly lower during the pandemic month than they were during the previous year's "typical" month. During March 2020, 70.32% (n=1583) of cases took place prior to the shelterin-place, and 29.68% (n=668) of cases took place after. This finding is consistent with the fact . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10.1101/2020.05.29.20117366 doi: medRxiv preprint that the shelter-in-place began on March 21-just over two-thirds of the way through the month.

In terms of overall trends, the overall sample contained cases with sex crimes 2.84% of the time, weapon use 14.83% of the time, arrests made 20.23% of the time, child victims involved 2.97% of the time, and residential locations involved 77.85% of the time. The amount that these varied by occurrence before or during the COVID-19 pandemic are described in Table   1 and was explored in the multiple regression analyses.

Bivariate analyses were conducted on a few independent variables of interest to determine potential need for interaction terms and joint hypothesis testing.

Several interaction effects were tested preliminarily. The most promising was the potential interaction between the presence of a sex crime and the presence of child victims, for both theoretical and data-driven reasons. Victims of sex crimes and child victims are two particularly vulnerable groups who may lack in-person access to resources during a pandemic, and the experience of child sexual victimization is particularly alarming (Finkelhor, Shattuck, Turner, & Hamby, 2014) . Chi-square analysis revealed a strong association between presence of a sex crime and presence of child victims (Table 2) , leading to later testing of interaction effects.

Additionally, weapon use and the involvement of an arrest in domestic violence posed an interesting relationship. Logically, it makes sense that there may be a relationship between a weapon used in domestic violence and subsequent arrest in that situation. Chi-square analysis revealed a significant association between these two variables ( Table 2) . Joint hypothesis testing using likelihood ratio was utilized on these two significantly associated variables, because the size and significance of their association was smaller than the association between sex crimes . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10.1101/2020.05.29.20117366 doi: medRxiv preprint value of 9.07, p=.97. To check for multicollinearity problems, a variance inflation factor (VIF) test was run, resulting in a low mean VIF of 1.09, indicating no problems. Influential data was assessed with Cook's distance and was not a problem for this model. The pseudo R 2 was .004, meaning that the model only accounted for .4% of the variation in the dependent variable.

A joint hypothesis test was also conducted to determine whether or not weapon use and resulting arrest jointly contributed to the model fit. The model was run with all variables, and then run again with weapon use and resulting arrest dropped from the model. In comparing (minus two times) the log-likelihood between the two models and obtaining a test chi-square, a significant chi-square of 8.95, p=.01 was found, indicating that at least one of the variables tested-either weapon use, resulting arrest, or both-was significant.

Model 2: Occurrence of domestic violence in shelter-in-place, regressed on multiple variables. The results of the second logistic regression model, predicting occurrence of domestic violence during the shelter-in-place period regressed on several key variables, provided additional significant findings, but again with low explanatory value (Table 4) . A chi-square test was performed on the log-likelihood ratio, and because its value was greater than the critical chisquare value for a two-tailed test, df=5, the null hypothesis that all coefficients are equal to zero was rejected. In addition to the significant intercept, in this model, residential location was once again significant. This time, however, resulting arrest was not at all significant, whereas the presence of a child victim was. All other being equal, a domestic violence case occurring at a residential location was 64% more likely (than cases at other locations) to have occurred during the shelter-in-place. Similarly, a domestic violence case with a child victim was 67% less likely (than cases without child victims) to have occurred during the shelter-in-place order.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 In terms of overall model appropriateness, Hosmer-Lemeshow goodness-of-fit test indicated that the logistic response function was appropriate with a non-significant chi-square value of 13.66, p=.62. To check for multicollinearity problems, a variance inflation factor (VIF) test was run, resulting in a low mean VIF of 1.05, indicating no problems. Influential data was assessed with Cook's distance and was not a problem for this model. The pseudo R 2 was .01, meaning that the model only accounted for 1% of the variation in the dependent variable. This was a slight improvement over the previous model.

A joint hypothesis test was conducted once more to determine whether or not weapon use and resulting arrest jointly contributed to the model fit. This model was run with all variables, and then run again with weapon use and resulting arrest dropped from the model. This time, in comparing (minus two times) the log-likelihood between the two models and obtaining a test chisquare, the chi-square value was not significant at 1.95, p=.38 was found, so the null hypothesis cannot be rejected, and the conclusion is that that neither variable was significant. Weapon use and resulting arrest did not jointly contribute to this model, unlike the previous one.

[ INSERT TABLE 4] 

This study offers a rapid comparison of domestic violence cases occurring in a large midwestern city before and during the COVID-19 pandemic, contributing to the lack of literature in this critical area. The two models explored-one predicting occurrence of domestic violence during the pandemic, and the other predicting occurrence of domestic violence during the shelter-in-place period-offer multidimensionality in analyzing the pandemic's impact, taking its rapidly-changing nature in terms of impacts and regulations into account. While both models had . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https: //doi.org/10.1101 //doi.org/10. /2020 little explanatory power (.4% and 1%, respectively), the models were highly significant. More important, multiple key characteristics were identified as important predictors of how domestic violence may differ during the pandemic and under shelter-in-place orders.

Residential location was a significant predictor in both models, with greater impact during the shelter-in-place period (22% versus 64%). This makes sense-if more people are staying at home, violence is more likely going to occur in the home than not. This finding may suggest that individuals are taking precautions and staying at home, which suggests compliance with shelter-in-place orders. However, this finding also suggests that domestic violence first responders and service providers may need to focus their efforts on residential areas and methods of reaching individuals who are staying at home. Usual surveillance practices of observing public spaces will not be as useful during a time when violence is happening in private residences.

Two other predictor variables were significant. Arrests were less likely during the pandemic, but only by 3%; this impact did not hold shift when shelter-in-place was ordered.

What did change during shelter-in-place was the presence of child victims-by quite a margin.

With cases with child victims being 67% less likely to have occurred during the shelter-in-place order, the question may be less about occurrence and more about reporting. What happens to reporting violence with child victims when everyone is sheltering in place-especially in terms of access to mandated reporters-and what can be done about it?

While a lesser focus of this analysis, one more notable finding is on the total number of domestic violence cases during the pandemic, as compared to the previous year. With 51.26% (n=2367) of these cases occurring before the pandemic, in this sample, domestic violence cases actually went down. This contradicts previously reported statistics of domestic violence increases during the COVID-19 pandemic (Bosman, 2020). While variables and case types may differ . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 across reports, greater attention may be needed on the source of the report-police versus hotline data-than is currently being given in the public sphere.

This study is limited in a few key ways. A major limitation is the lack of available data.

Several data sources in both Chicago and St. Louis were searched and contacted throughout the research process, without producing anything other than administrative police data. Additionally, this data was limited in scope, and extensive recoding was done to fit it within the confines of quantitative analysis. Particularly concerning was the lack of demographic information about perpetrators and victims, which limited the richness of the analysis and the ability to understand the pandemic's impact on domestic violence for certain communities. It is understandable that data would be sparse during this time, with reduced staffing and competing priorities. Some data, especially confidential domestic violence organization data, is simply not available as quickly as a pandemic would necessitate.

Additionally, there were time and methodological constraints that impacted the depth of analysis with the data available. With more time to explore different crime codes, types of abuse such as telephone harassment, stalking, and order of protection violations could have been included. Finally, low pseudo R 2 results and the limitation to one metropolitan area reduce generalizability and explanatory power of the findings, significant though they may be.

The study's strengths are the large sample, the innovation in working with available data, and the rapidness of response. Utilizing a large city's data allowed for nearly 5,000 domestic violence cases to be analyzed, offering great statistical power for this exploratory study. The study was also able to adapt to available data, quickly recoding as much useful qualitative data as was available into dichotomous variables for testing. The use of both pandemic and shelter-in-. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020 . . https://doi.org/10.1101 /2020 place timeframes for analysis, with a direct comparison group to the year prior, made for a more robust analysis and provides a glimpse at the rapid changes data can take on during this unprecedented time. Lastly, the study was able to make quick use of data that was available only a month prior to analysis. Its rapid approach to analysis and writing during such limitations and strain can serve as a model for other social work research during the COVID-19 pandemic.

This study provides a glimpse at how domestic violence may differ during the COVID-19 pandemic and its shelter-in-place orders, as well as a few key ways in which those differences manifest. Domestic violence first responders and service providers, as well as researchers, must pivot their approaches to understanding, preventing, and treating this problem given the rapidlyevolving nature of COVID-19. Given the significant findings of cases more likely to occur in residences and less likely to be reported on child victims during shelter-in-place, additional resources are needed in this critical areas, to make residential areas safer and protect children who are vulnerable at this time. The latter point is perhaps the most interesting finding of the study and one in which urgent responses may be needed. With in-person access to mandated reporters like teachers and social workers being restricted due to stay-at-home orders, innovative ways of understanding and responding to children's experiences in the home are needed to adapt to these new circumstances.

The low explanatory power of the two regression models could be improved with the identification of additional important predictors. For future research to improve, more sources of data are needed, as are greater availability of variables that might contribute to model explanatory power. Information such as the length of wait time for police response, location of first contact with police and service providers, co-occurring protective order violations, and . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020 . . https://doi.org/10.1101 /2020 elements of virtual/cyber abuse would be useful to further understanding the problem. Future research could compare domestic violence reported to police, to hotlines, and both during the pandemic, compared to usual reporting rates, to determine which services individuals are accessing during this time that looks so different from the norm. Open data sharing between research and practice realms would assist in this regard.

Lastly, funding must be directed to service providers, to move victims out of dangerous living situations, where more violence is being reported, and into safe spaces. Sheltering in place may create safety from COVID-19, which is of great public benefit. However, if physical and emotional safety are put at greater risk as a result, research and resources are needed to make sure individuals are simultaneously safe from both kinds of threats.

The author thanks Brett Drake, Merriah Croston, Maxine Davis, Shenyang Guo, Melissa Jonson-Reid, and Nancy Jacquelyn Perez-Flores for their encouragement and methodological guidance. The author also thanks Amirrah Abou-Youssef and Andrew Houghton for their on-theground wisdom that motivated the continuation of this work, and for their service to the Chicago area community.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10.1101/2020.05.29.20117366 doi: medRxiv preprint . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 Note. x 2 (1, N=4618)=8.0094, p=.005

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 Model Likelihood Ratio X 2 (df) 24.12(5)***

Note: * p<.05, ** p<.01, *** p<.001, from two-tailed test.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 Model Likelihood Ratio X 2 (df) 27.08(5)**

Note: * p<.05, ** p<.01, *** p<.001, from two-tailed test.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 3, 2020. . https://doi.org/10. 1101 /2020 

Crimes -2001 to present