Hello everyone it is noon so we'll get started. First I'd like to thank you all for being here for our final session of the NDACAN summer training series. This is the second summer that we've done this training webinar and we're really happy that you guys could join us. I know many of you have been here throughout the summer on and off. And it is Erin McCauley I'm your host for the summer series and I'm the graduate associate at the national data archive on child abuse and neglect. So the summer training series is hosted by the national data archive on child abuse and neglect. We are housed in the Bronfenbrenner center for translational research at Cornell University. And we are in charge of making this data publicly accessible and so part of our mission in that regards is to host summer training series like this that helps power participants understand and be able to use our data. Great so here's an overview of our summer if you've been with us the whole time you know that at the beginning we had an introduction to the archive, the data, and the services we have available. Then we went in-depth on our three different administrative data sources. We then had Frank,, Frank Edwards is an associate with our archive but is also professor at Rutgers. And he did a session on strategies for managing data especially for our large administrative data sets. Michael Dineen, one of our analysts, then came in and talked about how to link these three data sets. And send now we're going to be doing our concluding session where Frank has come back and he is going to be talking about the research project he did using some of our administrative data. So I know that a lot of us will kind of refer back to anyone who was here for the the strategies for data management session on August 7 and the have a lot in common and needs to use some pretty advanced strategies in order to manage that data. So I'm going to be passing it over to Frank who as I said is an analyst with us and is also a second-year professor at Rutgers. Hi everybody thanks for joining us today. And today I'm going to the presenting the findings of the paper that's recently appeared at Russell Sage foundation Journal of the social sciences. And we wanted to present this today to give you a sense of how you could use our administrative data to put together some interesting empirical analyses that can address important questions about both child maltreatments and the child welfare system. So today I'm going to give you an example of how you can use the NCANDS data that's the national child abuse and neglect data system to produce a county-year estimates of child abuse and neglect reports by type of reporters. So in this paper what I do is a use the NCANDS to to create a unit of analysis at the county and year level and to produce counts by different kinds of mandated reporters. Then I use multilevel models to explore how and why police report child abuse and neglect at different rates across places and over time. So the key research question for me in this analysis is who is reporting child abuse and neglect in different places, and does the way that police work in different places affect how likely they are to report child abuse and neglect to the child welfare system. So I want to emphasize that I welcome questions throughout so if at any point you have a question for me just go ahead and throw it in the chat box and I will try and respond to it as I go. Some of this stuff can get a little technical so I want to make sure that you feel free to to stop me as we're going if something is unclear. So again I assume you know if if you're interested in this analysis and interested in the code I used for the analysis it's all available on GitHub and I'll post a link in the chat at the end to the code I used. And the data are used for this is the annual NCANDS reports from I think 2004 two 2015. So you'd have to request those from NDACAN but if you have those and the code you could replicate the analysis. So I will proceed. The title of this paper is Family Surveillance, Race, Police, and the Reporting of Child Abuse and Neglect. So this analysis thinks about family surveillance as a social institution that emerges out of the actions of a lot of different organizations. So child welfare as I'm sure many of you know is unlike a lot of other state agencies in that it requires the participation of a really broad network of professionals and community members to conduct surveillance for child abuse and neglect out in the community, right? And the state statutes for who is a mandated reporter of child abuse and neglect do you vary somewhat and a few states have universal mandated reporting laws. But you know when we look at the primary categories of professionals who report to child to the child welfare services who report screened in reports because that's what with God reported in the NCANDS, about half of all child abuse and neglect reports nationally are provided by police and educational personnel education personnel teachers and other school personnel. Those alternate between the top category depending on the place and here. Teachers and police are responsible for about 20% of all reports each and then doctors and other medical personnel follow right behind that so those three categories of mandated professionals account for about half of all screened in CPS reports. So I'm interested in thinking about how these different agencies that are sort of external to the child welfare system can inform outcomes within the system, right? There's been some other research that's looked at schools, other research that's looked at medical professionals but here we want to look really closely that police. And so I'm asking whether variation in policing itself helps to explain variation in the intensity of child abuse and neglect reporting which I'm calling family surveillance here. And I'm also interested in thinking about whether inequalities and exposure to policing can help us understand racial and ethnic inequalities in upstream child welfare system outcomes, particularly things like foster care. We know that there's high levels of racial inequality in the criminal justice system so it may be the case that inequalities in exposure to police may help to explain inequalities in rates at which children are coming into the child welfare system across racial and ethnic groups. So, why might there be ties between police and child protection agencies beyond mandated reporting? In most jurisdictions there's mandated coordination between police departments and child welfare agencies for initial contact and for risk assessments. So police are often going out with caseworkers to conduct initial screenings when a report comes in. In a few jurisdictions this relationship is informal and discretionary based on perceptions of safety. Washington allows caseworkers to and supervisors to use discretion in whether they bring police along with them for initial screenings is my understanding of that policy. And there is a few counties in which police are the primary responder and can conduct initial safety screenings and there's a few counties in Florida and Arkansas in which police have taken over the primary responsibility of conducting child welfare investigations. So there's really deep ties between police and child welfare system systems across the country in addition to reporting. So the kind of theoretical model I'm thinking about here is wanting us to break down how a child abuse and neglect case that's recorded in the NCANDS data, right? And so we can think of this as crudely being a four step process, right? With step one being that a child or family must be observed by a potential reporter in order for a report to be generated, right? So someone with some propensity to report a case to the child welfare system must observe that family prior to a report being generated. That observer must cognitively classify that whatever they observed you know whether that's the behavior of a child, whether that's you know external symptoms or signs like bruises, whether that's you know symptoms of a in no psychological symptoms of of perhaps sexual maltreatment, right? That's the the the observer the you know potential reporter who is interacting with the child must kind of think ah this is a sign of child abuse or neglect, and must also decide that it's appropriate to call a state or local child abuse hotline in response to that sort of cognitive classification at which point CPS must decide to screen that case in, right? So there's a lot of institutional process that happens prior to a case coming in. I want us to kind of I, in my research I I try and push this away from thinking about you know the case of child abuse and neglect we we think of as automatically being brought to the attention of CPS and we know that's not the case. You know the the the case coming to the attention of CPS is conditional on contact with the potential reporter and the classification by that potential reporter of it being a case of abuse or neglect and the decision of that reporter to file a report. And then CPS's decision to screen it unless you're an automatic screen-in jurisdiction. So there's a lot of process that goes into a case rising to the level that it gets CPS attention and the variation exposure to mandated reporters or variation in ideas about what constitutes child abuse and neglect among different categories of reporters in addition to potential bias among mandated reporters and other community based reporters across class, race, sex, and other and disability otherwise might inform whether the juice to classify something us child abuse or neglect. Okay so just be explicit the sort of theory that I'm working from here suggests that inequalities in one system can drift across contacts to cause inequalities in another system, right? So in this case inequalities in policing may sort of migrate from the criminal justice system into the child welfare system through the vehicle of child abuse and neglect reporting. And so this could work in a few ways. One, that unequal policing could cause child welfare inequalities through unequal levels of detection, right? So if we are in a community in which police have frequent contact with children and families in the community they have more opportunities to observe families for potential signs of abuse and neglect and they have more opportunities to make a decision that's what they've seen looks like child abuse and neglect and deserves to be reported to CPS. Unequal policing might also cause child welfare inequalities because we know that when police touch members of communities, whether that's through an arrest or a conviction through the courts, that that experienced that that you know prior history of arrest, prior history of conviction, prior history of incarceration, these attach stigmas to members of the community, right? That are used as information by child welfare systems in establishing things like parental fitness and establishing things like you know potential danger to a child. And so the application of that criminal stigma might lead to inequalities in child welfare outcomes and we can think about whether you know that stigma is related to safety of course that's an important question. But for now we're kind of treating it as a neutral you know stigma ambivalent about whether that is related to child safety or not. Unequal policing might also cause child welfare inequalities because when you arrest a parent or caregiver you are obviously detaining that's person and and physically removing them from the environment and making it impossible for them to care for a child. This can be true of even short jail stints and you know the AFCARS does include variables for whether a child was placed into foster care as a result of parental incarceration. I think that may be an undercount but you know we can we can easily think of ways and we've heard anecdotal evidence from plenty of of you know folks on the ground that this is a routine way that some children are ending up in the system as parents are sort of incapacitated through the criminal justice system. And it might also strain the resources of families, right? So we might think of this as you know one effect being on the focal parent you know that parent who is arrested or stopped, but it could also strain the resources of the kinds of kin caregivers that often play a really important role in children's lives especially the lives of children of color and children of low income communities. You know people like grandparents, aunts and uncles may be playing supportive roles not only for a child who is in need of kinship care but also for potentially other relatives who may have criminal justice system contact. And providing that level of support to you know brothers, sisters, friends other other you know kin intictive can really strain both the emotional and financial resources of a family. So all these ways the criminal justice system might be kind of feeding into the inequalities we see happening in the child welfare system outcome stream. So I want to be clear this is a descriptive study we're not getting into a causal effect of whether you know police contact causes child welfare system inequalities. This is a county level analysis and we are not diving into microlevel relationships often does a particular parents exposure to arrest or incarceration or just police contact increase the likelihood that that parent or family will experience a CPS investigation. We don't currently have that's data at NDACAN and it would require linked administrative data across criminal justice and child welfare systems that some states have available but we don't have available for this for these national data. And there's also of course the problem of selection that while you know policing is not perfectly correlated with criminal or dangerous behavior there is some association there, right? So crime and child maltreatments are surely associated with each other, that is, you know people who know engage in violent behavior people who engage in substance abuse are likely to have higher levels of child maltreatment than those who do not and also are more likely to have contact with criminal justice systems than those who do not. So we're not dealing with selection here we're thinking of this as a purely descriptive, does variation in policing systematically with late to variation in police child maltreatment reporting with the important caveat that we're not getting at those microlevel relationships or the problem of selection into both crime and child maltreatment on the parent's level. So without further ado let's get to probably what's mostly of interest to folks here which is the data and methods. So again the focal measures for this study come from the NCANDS child file. I'm using data from 2000 to 2014, apologies for my confusion earlier about which years of the data were included in the analysis. And this paper relies on one focal variable in the NCANDS and that's reports source. So this is a I believe 13 category variable that tells us who made the initial report to a child welfare hotline. And it includes codes for law enforcement personnel which is what I'm focused on today, but it also includes codes for other categories of routine mandated reporters and codes for anonymous reporting, codes for family member reporting, codes for self reporting, right? It's a really useful variable to think about pathways into the child welfare system that in my opinion has not been widely has not been used widely enough in research. So you know if this kind of, this kind of research question as interesting to you I think there's a real opportunity to use this variable to ask questions about for example how often teachers are reporting child abuse or neglect, or how often doctors are reporting child abuse and neglect or how often members of the community are reporting. There's a lot of interesting questions we can take into with this variable. Here again I'm only focused on police. And what I'm doing is I'm taking the child file which is a report-child level, right? So for those of you who haven't worked with the NCANDS before it can work on as as two units of analysis, right? So it's at the report-child level meaning that a single child could have multiple screens in reports within a year in which case they'd have multiple rows in the data. So we have to think of this as you know working either at the reports level or the child level when we're working with some micro data and you have to be really careful about those decisions. But here I'm focused on the county and year so what I do is I aggregate the data to the county level for each annual file. And that gives me a count of reports filed by police at the county level for the all those counties that meet our threshold for inclusion that have I I don't recall the exact threshold right now but you have several at least several hundred reports of the child abuse and neglect in the county and so these are all of our you know mid- and high- population counties it's about four or 500 counties depending on the year. And so we get a time series here where for each county in the data we get from observation for each year and then we get a count of police child maltreatment reports by race and ethnicity because we're really interested I'm I'm really interested in understanding mechanisms of racial inequality. So we have a few data quality concerns though. For counties with small subpopulations, right? So for example we might be really interested in American Indian and Alaska native inequalities in child welfare and thinking about whether police might contribute to those inequalities, but in counties with relatively small populations of American Indian and Alaska natives if you know we could think of this you know if there is a as a probability problem. If there is a population of of 500 and the likelihood of being subject to a report in any given year is, say, 3%, right? The the the number of children in a given year who are subject to child welfare investigation is going to jump around a lot even if it stays within you know a standard deviation or two of that value. So it's going to appear to be more unstable than it really is so we need to think carefully about how to handle those kinds of unstable rate estimates. And that's something that multilevel models are quite good at handling. We also want to deal with missing data on race and report source. You know there is a small proportion of cases that are missing information on the focal child's race or ethnicity on the NCANDS report and there's a small proportion of cases that are missing data on the source of a child abuse and neglect report. Based on my examinations of the data and you want to do exploratory data analysis to make sure you meet the assumptions of a missing at random assumption, but I I believe that the data are missing at random conditional on the county and other covariates that we are able to include in the data so I used multiple imputation by chained equations to produce multiple sets of data for those missing values and use that as a source of uncertainty in the regression models. I'm going to show you in a little bit. Okay so let's keep going now recall this is the theoretical model introduced at the beginning, right? So we have this you know observation by potential mandated a potential reporter of child abuse and neglect, the classification decision which leads to a decision to call state or local hotline which leads to a CPS screening decision. But of course, we have a whole lot of confounding going on, right? So the actual you know behavior of the child, an injury or living condition, right? That is those kinds of external you know both bodily you know physical and psychological symptoms but also you know conditions of housing, you know various behavioral things observations of of parents and family behavior. These are all confounding each along the way. Of this process, right? That we can imagine ways in which you know actual symptoms of child or actual signals of child abuse and neglect could inform each of these processes along the way. And those are umeasured, right? We don't have good information other than the type of maltreatment alleged in the initial report and then substantiation decisions in the initial report. So we don't have a lot of data on what prompted either the initial observation by the reporter or the condition of the child or family when they observed them. And I also want to emphasize that we need to have the agency response recorded in the NCANDS in order for us to have a case. So there's a lot of confounding and there's also a potential issue in terms of quality of the data that gets recorded ultimately, in the NCANDS. And these are state reports to the federal to the Children's Bureau who of course do a substantial amount of work to harmonize and clean the data but it's really important to read that data documentation to understand the differences across some state reporting practices and differences within states over time. The documentation is your best friend in helping us to understand how to treat those data. Okay I'm being told that my mic isn't working, looks okay on my and okay sorry about that. So I think that might have been my AC coming on and cutting me out so let me just turn that off and that might help. Okay so anyway let's let's let's push forward. Okay so we have our focal outcomes so on the on the left-hand side of our regression models were going to have counts of police reports of child abuse and neglect by child, race, ethnicity. And on the right-hand side of our models we're going to have a series of predictors and controls from demographic data and from criminal justice data. So the crime and arrest data time using comes from the Uniform Crime Reporting System and so for those of you who don't know the UCR is the sort of most definitive data source we have on both criminal behavior and on arrest in the US for most of the country. It's a voluntary reporting system housed and collected by the Bureau of Justice statistics and the FBI. And it's been around for about 100 years. So there's a file called arrests Hoopes apologies arrests by age, sex, and race summarized yearly and for those who want to obtain the data it's available from the national archive of criminal justice data at the University of Michigan, ICPSR. So go check that out if you haven't seen the data before. So I used these data to derive county level arrest rates by year, race and and category of offense. So I'll kind of talk a little bit about the offense categories that I used in this study a little bit. We I use all arrests, violent arrests, and drug arrests, and quality of life arrests. And so the violent arrests include those categories that the UCR uses in their annual reports and those include homicide, assault, burglary and sexual assault. Drug offenses include all drug arrests. And quality-of-life offenses are a source of more diffuse set of arrestable offenses things like vagrancy and loitering and gambling, prostitution. The kinds of things we often associate with broken windows policing and quality of life policing. So I'm interested they're in thinking about that as a potential measure of low level arrests that may more directly pick up the decisions of police departments. Whereas enforcement of violent crime is is going to be a little more consistent across places conditional on levels of violence. So there is some missingness in the UCR as well, and so some agencies don't report some years and report others. And I use a combination of interpolation and multiple imputation to address those problems. And I also want to point out that arrest reporting of Latin X people is unreliable across places. So a lot of jurisdictions are not recording a field for a you know Hispanic arrests in their data and so we can't really compare easily arrests of Latinos and Latinas across places. And so the they are excluded from most analyses despite the fact that we have good data in the NCANDS we can't really compare it to the UCR. Other measures included here are full-time officers per capita so that's the number of cops on the police force in a county -and that's across all police agencies within the county so that would be both municipal and Sheriffs offices-, the number of police agencies in a county, the population by age, race, and year, and that's obtained from the seer, CDC's seer population data, and the officers per capita and number of police agencies comes from the LEMAS data  L E M A S, law enforcement management and something statistics that's also available at the NACJD. And so we look at population composition, child poverty, county metro type whether it's rural or urban and the kind of standard panel of controls that you would see in this sort of analysis. Okay so now to the models. Apologies to put a lot of Greek up on the screen but I think these models deserve a little bit of a talk through. So I'm using a Poisson multilevel model to model these outcomes with an over dispersion parameter. So a Poisson is a great model for count variables, right? That is variables where it's a discrete event that can happen a fixed number of times, you know and it's constrained to integer values, right? So it's a great model for count variables like that like those we're dealing with here and it's superior to I believe to to an ordinary least squares model for this kind of data because it allows for the kinds of over dispersion we often see in count data and the it's also constrained to always fit onto and integer scale. So it's it's a more natural kind of model for this sort of outcome we're looking at here. You could also consider a negative binomial for this sort of outcome or a gamma. Those those would also be good models to look into. It's a little easier to estimate a Poisson in those multilevel modeling packages than it is to model to fit a negative binomial. So here so we say that you know the the outcome of numbers of police reports is distributed Poisson and here on the second line of the equation note that the typical regression equation includes a the gamma which is a beta zero plus zeta and that's just our intercept term and that includes a you see for county "i" so that's our county level intercept. So each county in the data is going to get its own intercept and that's going to help us kind of a just for some of the between year fuzziness. We also have the the theta is just our typical panel of regression predictors those those in the model which I'll kind of come down here to this theta "i" "j" "k" to talk about in a second. And the epsilon is what's called an over dispersion term so that's an observation-level random effect in the data. So you can see that we are estimating to random effects here, one at the county level and one at the observation level. And then that's offset by the size of the focal child population, right? So the count can go no higher than the number of children that are in the population and that's the constraint I put on the model. So this theta "i" "j" "k" that's our sort of typical you know beta one plus beta two X plus beta three X, etc. here I'm using a method that allows us to simultaneously model both between jurisdiction differences and within jurisdiction changes over time. So this here we have the beta 1 j, j is year, so we have a national linear time term. In addition to having for each X in the for each control or predictor in the model we we have two betas, beta 2 x bar "i" "k" 1 and apologies that there's a lot of notation here. That is going to take the average value at the county level for that first measure X. So that could be arrests for drug offenses in let's say I am currently in Essex county New Jersey so that would be the number of drug arrests in Essex county New Jersey between 2000 and 2014, the mean value for that. Now I take that x bar "i" "k" 1 and I also difference it from those annual observations. So in addition to taking the mean value, I also look at the within county changes over time when that's the beta three and you know within the parentheses there we have x "i" "j" "k" 1 minus x "i" "k" 1 so it x bar "i" "k" 1. So that differences the individual year observations from the average observation each year and so that gives us both a between-county coefficient in beta 2 and then beta 3 a within-county coefficient. So the models are simultaneously estimating between and within county variation and that beta three parameter that beta three coefficient is going to be identical to what you'd get in a fixed effects model. So for those of you who really strongly prefer a fixed effects approach for these kinds of timeseries data, this approach has you know the flexibility of multilevel modeling which allows us to to to be a little more creative in how we specify models. It's it's allowing us to get the same kinds of results with get in a fixed effects set up but also get the between parameter had you lose when you to a fixed effects set up. That is you can't speak about between-unit variation when you're using that kind of a fixed effects approach but here we can. So far I'm happy to address any questions about the modeling or if any of this is unclear feel free to raise a question in the chat box. I know this is a bit technical but this is kind of a neat feature of the models that will let us do some interesting things going forward. So onward to the findings. So first I want to give you some descriptive overviews of what we see. This is suspected maltreatment types by child race and here we're sorting out police reports from all reports. So in each bar plot you'll see a racial or ethnic group with total on the top left. Then on the y-axis you'll see maltreatment type: sexual abuse, psychological maltreatment, physical abuse and neglect. On the y-axis you'll see the proportion of reports represented by each class, so it ranges from zero to about 0.75. And then the pink bar is for all reports and the blue bar is only sub setting it to police reports. And a few trends jumped out. Can see that for all groups African Americans, police are reporting neglect at or high a higher level than is than are other kinds of reporters. So the bulk of child maltreatment reports are neglect, and that's true for all reporters and true for police. For African-Americans they are reporting slightly lower levels of neglect. Interestingly police are reporting less physical abuse than are other kinds of reporters, and more psychological maltreatment and sexual abuse than other kinds of reporters. So this is just to kind of help us get our head around what kinds of child abuse and neglect police are reporting and how that compares to other reporters. And here is the rates of police maltreatment reporting by race over time. So this is maltreatment reports per thousand children in a county's population and we can see that this has been going up over time. The numbers per capita of police child maltreatment reports has steadily increased over time for all groups except for Asian Pacific Islanders and American Indians and Alaska natives for whom it's fluctuated quite a bit. But there is a pretty clear upward trajectory in the volume of cases that police have been reporting to the child welfare system. And when we look at this as a proportion of total reports filed so this is effectively dividing the counts of reports filed by police by the total number of reports in the county, we see a a slightly more dramatic story. Nationally in 2000, police were reporting about 15% of all child abuse and neglect cases that got screened in, and then in 2014 that's number rose to about 19% of all cases. When we look at African-Americans it rose from 16% to just about 20% in 2014. And then when we look at American Indians and Alaska natives, there's been a dramatic jump from about 14% all the way up to about 23% of all cases that get screened in being filed by police. So police are for all groups, filing somewhere between 15 and you know 23% of all child abuse and neglect reporting's. And again in the most recent year included in this analysis that number was about 19% of all child abuse and neglect cases that get screened in starting with a call from a police officer. So that between- and within- component of the model this is a way to help us think about this. This is a statistic called the coefficient of variation and it helps us think about how much of the variation in an outcome is between units and how much of it is within units. So counties are our units here. So the red the pink bars tell us how much of the variation across observations can we explain as a function of cross-county variation, and how much of it can we explain as variation within counties. And looking at total we have arrests on the left and reports on the on the right on the X axis and then the coefficient of variation is on the line. And again we're faceting it out by race ethnicity and note for Latin X peoples we don't have arrest data that we can use so those columns are blank and they will be excluded from the regression so I'll show you in a little bit. Okay so most of the variation in arrests is between counties, that is some counties have high levels of arrests, some counties have low levels of arrests and that kind of stays put over time. Those differences are relatively stable across counties we don't see a lot of counties moving from low arrest to high arrest, or high arrest to low arrest, right? There's not a lot of within-county variation but there's a whole lot of cross county variation. With child welfare reporting, there's slightly more within-county variation than there is with then there is with arrests but most of the variation is still between counties. That is, there's a lot more differences across counties then there are differences within counties over time. And that also is going to help have impact on how we estimate models. If there's not a lot of within-county variation in rates of police child maltreatment reporting or rates of arrest, then it's going to impact our ability to estimate within-county changes in that within within-county component of our model. That is we might not see much going on on the within-county component of the models because there's not much variance to explain their. And again indeed we'll see that when we get to the findings. Okay so these are the focal results from the regressions and instead of showing you a big table of regression parameters I'm going to show you these visuals that kind of estimate the simulated change in or expected maltreatment reports at different levels of arrest. So what we're seeing here the X is the expectation from our models from the models for a county at a total arrest rate that's at the average, right? And then the circle and bars are for what we expect to happen if we move up the arrest rate by one standard deviation. And then we'll break that out by different categories of offenses and by different racial and ethnic groups. And so we can see in the top panel that for total arrests, right? We expect somewhere about 1.8, 1.75 per thousand children to be reported by police in a county with an average total arrest rate. But when we move that up by one standard deviation we expect somewhere around 2.1 reports by police per thousand children, right? So counties with higher levels higher arrest rates have higher levels of reporting for all categories of children, for Asian Pacific Islanders, for African-Americans, or American Indians, and for white children, right? As we move down the each each panel is a different racial and ethnic group. Okay so let's add in our other arrest outcomes to see what else our arrest variables to see how they relate. So when we look at violent crime we see a similar pattern to that we see for total arrest rates. That is counties with more violent crime arrests have higher levels of police maltreatment reporting. The same is true for drugs. Counties with higher levels of drug arrests have higher levels of police maltreatment reporting. Note that the expected level of police maltreatment reporting for high drug arrests of American Indian children the gap there is a bit larger than it is for other groups. And for quality of life arrests we still see the same basic pattern that is counties with high levels of quality of life arrests tend to have higher levels of police child maltreatment reporting. Now for comparison I wanted to look at child poverty, right? So if we took to think about whether it's levels of arrest that are most influencing police maltreatment reporting or whether it is underlying conditions of the child population through things like child poverty and other well-being indicators. So here we can see that for all groups we don't see much of an association between child poverty rates and police maltreatment reporting. We see small positive relationships except for white children. For white where white child poverty is high we do see higher levels of police maltreatment reporting of white children. But for other racial and ethnic groups we don't see that same relationship. So that's the between-county component of the models. Here's the within-county component of the models. And I was you know kind of warning you earlier that there's not much within-county variance to explain so that standard deviation of a within-county change in arrests is quite low because you know most counties don't see changes in arrests but what if we did increase arrests by as much as we might ordinarily see when a county does have an increase in arrests. Well we don't see much difference in police child maltreatment reporting. Now that's not to say that there's no relationship there but there's not one that I'm able to pick up in this analysis. Okay so what are the implications of these results? First I think we need to think about the process of family surveillance as being inherently multi-institutional, right? That child welfare systems are not like police departments in that they do not have agents that are actively on the you know in communities monitoring for signs of child abuse and neglect and instead the have to rely on this kind of diffuse network of you know private and professional actors to engage in the sort of monitoring of children and families for the signs of child abuse and neglect. And so the rules that govern how those institutions work that is who they come in contact with and how they interact with those children and families are going to play a really important part in determining who comes into the child welfare system. And we need to be thinking about child welfare is something that doesn't be begin and end in the child welfare office but rather something that extends to all of those institutions that touched children and families and the decisions they make about who to investigate and when to call in a report. And police are central to that process. The way that services are organized locally the way that police decide to conduct their business locally plays a huge role in structuring who comes into contact with the system. And this analysis also shows that low levels of criminal justice contact might open up the possibility of family separation. When we recall that things like drug arrests and quality of life policing are still stably associated with levels of child abuse and neglect reporting by police it might be the case that arrests and police contact for low-level offenses opens up a family to the potentially very serious implications of involvement with the child welfare system. So we could think about police and CPS has kind of forming an overlapping institution that you know Vesla Weaver and Amy Lerman have called a "carceral lifeworld", right? That is they they they made a in the perceptions of community members in some interesting work from folks like Kelly Fong have shown that mothers in particular might be thinking about avoiding contact with the child welfare system in terms of making decisions about who to interact with and where to be. That we know from the quite a bit of research that police have similar effects on decision making in communities. And so we might want to think about how the joint role of police and the child welfare system might be shaping how low income mothers in particular and low income families in general think about interacting with the kinds of systems decides to make calls to child welfare hotlines. I also want to emphasize that police are not objective instruments. So we might think that you know a diffuse in a public health sense that a diffuse instrument for the surveillance of signs of child abuse and neglect that is having lots of potential you know nodes that could sort of send reports of child abuse and neglect up to CPS offices is a good thing, right? In that we want to make sure that we're getting those those we're we're having a high hit rate in terms of detecting actual child maltreatment out in the population but we know that police are not objective instruments of social surveillance, right? That the intensity and character of policing depends really heavily on where the police are working and that's a function of segregation, that's a function of race and ethnicity, that's a function of social class. We know that police are likely to interact with a low income African-American family in a very different way than they interact with an affluent white family, right? And that style of interaction could have relatively severe consequences for outcomes in the child welfare system. So I think we need to be cautious as we think about who we are tasking with child welfare surveillance and the implications that has for who comes into contact with this system. So you know again just to emphasize that policing may cause child welfare inequalities through varying levels of detection through the application of criminal stigma, through creating crises of care by removing caregivers, and by straining resources of the kin and victim kin who are often so important to taking care of vulnerable children. The policy implications here I'll kind of go quickly through but here's drug arrests are a big vector, here and this also gets us thinking about function creep in terms of whether we think it's appropriate for police to be the kinds of frontline first responders to child welfare abuse and neglect cases. I know some jurisdictions have advocated for this as a way to reduce the antagonism between families and child welfare caseworkers but I think this deserves a much deeper conversation. And I think we need to think seriously about whether there might be a sort of growing phenomenon of of what's been called legal cynicism and system avoidance in terms of interactions between children and families and the child welfare system in ways that look really similar to things we're seeing happening in the criminal justice system. You know obviously there's a lot of further questions and questions this research can't answer. First and most importantly we don't have causality here. Second the said these results may be sensitive to the UCR and we may need to think about replicating with other crime data. We also might think about other reports sources like teachers, like social service professionals, like medical professionals. And we also might think about the sensitivity of these surveillance instruments in terms of false positives and negatives in terms of the kinds of concerns I raised about thinking about police as a nonobjective instrument. So with that I look forward to your questions and thank you for attending today. Well thank you so much Frank that was a really really engaging presentation and I think it leaves a lot of us inspired about what you can do with this linked or non-linked administrative data especially if you have strong data management skills that can kind of you know grasp this data into a usable format. So that was really fantastic I'm going to give people a moment to start typing out any questions people have into the chat box but I'll just take take a moment to talk about our end of year survey. So this was the last presentation in our series so the the summer series is officially over. I will be sending out an email to everyone who participated in the next week or so with a survey. We're going to ask your opinions on the different sessions that you saw, what you really liked about this series what could be improved for next year and then we'll also be asking questions about areas where you'd like to see future topics. So the idea to do this summer is based on on looking at the entire administrative cluster was actually suggested by some of our participants of last summer. So if you have ideas about directions you want us to take it information that you still feel like you still need as a data user or potential data user please let us know. So we have our first question it says I missed the beginning so I might be asking a question that you covered but how did you a measure maltreatment report that was made by police? I was muted so this is actually a variable we have in the NCANDS. We have a variable for the source of a report so if you have access to the NCANDS child file you can easily look at who filed that report it's about a 13 category variables that includes lots of mandated reporters in addition to police. Perfect thank you Frank. Yup. Any other questions? Is it possible to measure a police report due to substance use crime? So we can kind of proxy that by looking at drug arrests, right? We could also look at we could also look at a report being filed by a police and there being allegations of a substance abuse-related maltreatment in the NCANDS. You you know want to make sure that you're looking carefully at the missingness in the maltreatment type variables in the NCANDS before you do that. But yet you could imagine interacting report source with with a the maltreatment type variables. Yes and I recommend anyone who's interested in pursuing research in this area to check out our codebooks on our website. They are pretty extensive as our user guides we create them all in-house so that's definitely a good place to start if you're trying to figure out exactly what we measure and how and just to reiterate what Frank said always look at the missingness. But there is quite a bit you can either find or proxy given the detail of the data. Other questions? Okay well I suppose we can leave it there. If you have any questions you'd like to follow up with I just put my email in the chat box I'm happy to address any questions you have about this study or other work with the NCANDS if you are pursuing work with these data. Wonderful so big thank you to Frank and for everyone for being here today and as I said please fill out our survey will be sending it in the next week and thank you so much for joining us for our second annual summer training webinar series.