In what kinds of communities do people on the sex offender registry live? An analysis of ten states

Abstract

This is the first large scale community-level study describing the characteristics of communities where Registered Sex Offenders (RSO’s) are more likely to live. This study presents RSO residence data from ten states, combined with census data. Zip code characteristics (e.g., income, race/ethnicity, percent of population under 18) were then used in bivariate analyses and negative binomial regression analyses to determine which community factors predicted RSO residency. Lower median household income predicted higher rates of RSO’s in nine of the ten states. These effects were large, with the rate of RSO’s dropping about two percentage points per $1000 in increased median household income. Other community characteristics were found to have smaller effects on a state by state basis.

1. Introduction

In this study, we are interested in understanding the kinds of communities where people required to register as sexual offenders (often known as “registered sex offenders” or RSO’s) live. In other words, we seek to provide information about the community characteristics which predict a higher density of RSO residents. This paper offers an analysis of the geographic distribution of RSO’s based on data gathered from ten online state sexual offender registries along with Census data from the American Community Survey (ACS). While most prior work on this subject has described the characteristics of individual RSO’s, this paper describes the characteristics of the communities in which RSO’s are most likely to reside. We also extend prior geographical work in scope, including ten states in our analyses, where most prior work has included only select counties or clusters of counties in a few states. We include data from Arizona, Florida, Georgia, Illinois, Maryland, Missouri, North Carolina, North Dakota and Oklahoma.

1.1. Background

Since the 1990′s, federal law has required those convicted of sexual offenses to register their residential location in public, searchable databases (McPherson, 2016). Thus, the term “registered sex offender” or RSO has entered the public and academic lexicon. According to the National Center for Missing and Exploited Children (NCMEC), there are over 850,000 RSO’s in the U.S. or about 2.64 per 1000 people (NCMEC, 2017), although some of the RSO’s counted in this estimate are only visible to law enforcement and not the general public. Ackerman, Harris, Levenson, and Zgoba (2011) estimated the number of RSO’s searchable to the public to be about two-thirds this number. In the last quarter century, substantial federal, state, and local policy and legislative work has been devoted to managing the perceived risk to public safety posed by RSO’s (Mustaine, Tewksbury, & Stengel, 2006). These policies have come to include not only reporting where RSO’s do live, but to also include restrictions on where RSO’s can live (Evans, Lytle, & Sample, 2014). Commonly, states and municipalities restrict RSO’s from living within a certain number of feet (typically between 500 and 2000) from schools, daycares, bus stops and parks (Savage & Windsor, 2018).

The stated primary intent of sex offender policies is the protection of children, with many national sex offender laws being named after child victims, including the Jacob Wetterling Act (1994) and Megan’s Law (1996) (McPherson, 2016). The degree to which these policies are evidence-based or have resulted in beneficial outcomes is doubtful (Ackerman & Furman, 2012; Wright, 2014). Most sexual assaults are not committed by RSO’s, and a recent study estimated the re-offense rate of RSO’s to be about 1% per year (Zgoba et al., 2013). Moreover, most residence restrictions are of questionable efficacy in keeping RSO’s away from children (Ackerman & Furman, 2012; Mogavero & Kennedy, 2017). At the same time, a range of collateral consequences for RSO’s have been documented, including employment and housing discrimination (Tewksbury, 2005), that have increased instability and decreased access to services for this population (Levenson, University, & Hern, 2007).

Who RSO’s are, where they live, and the degree to which their residential location should be restricted are topics of ongoing academic, policy and public interest. There have been two primary streams of scientific inquiry about RSO’s – those that collect and analyze data at the individual level of observation and those that compare rates of sex offenders across geographic areas. Studies using data at the individual level have yielded many valuable substantive (Ackerman & Sacks, 2018; Ackerman et al., 2011) and methodological (Ackerman, 2015) insights about who RSO’s are. Among the key findings are that RSO’s are about 98% male, with a median age of about 44 years, and that these RSO’s victims are mainly under 18 years old (74%), with a median age between 11 and 14 years, and female (87%).

As is common in criminal justice data, there is state-by-state variation in the meaning of the terminology used, the methods for categorizing people convicted of offenses, the data collected on each person, and the degree of public access granted to these data (National Research Council, 2014). For example, some states group RSO’s together and apply the same registration and residence restrictions to all, regardless of seriousness of offense or type of victim. Other states employ a “tier” system, assigning different requirements to different RSO’s based on perceived risk (Savage & Windsor, 2018). While federal law requires states to use a tier system to determine registration timeframes, each state uses its own labeling system for designating other tier-based requirements (Evans et al., 2014). The proportion of RSO’s at each level of the tier system is quite inconsistent across states (Ackerman et al., 2011), making cross-state comparisons meaningless. Moreover, there is a general tendency of states to record race but not ethnicity (Hispanic status), making racial epidemiology at the level of the RSO problematic beyond the documentation of a disproportionate representation of African Americans (Ackerman & Sacks, 2018). This is because failure to denote Hispanic status not only makes it impossible to estimate the number of Hispanics, it makes it impossible to estimate the number of non-Hispanic Whites, which is critical in many areas such as Texas and Arizona, in which Hispanics comprise over 30% of the population.

The second stream of research on who RSO’s are and where they live is at the community level – these studies look not at individual RSO’s but rates of RSO residence across different geographic areas. These studies have generally focused on collecting data from counties, metropolitan areas or within-state regions, often aggregating data at the tract level. They also tend to be more theoretical than epidemiological in nature, with a strong emphasis on conceptualizing aspects of community disorganization (Sampson, Raudenbush, & Earls, 1997). The most consistent finding in these studies has been higher concentrations of RSO’s in low-income areas (Hughes & Kadleck, 2008; Mustaine & Tewksbury, 2011; Socia, 2016). Because these studies place a premium on sophisticated measurement of variables associated with community disorganization, they have had “limited geographic coverage” (Socia, 2016, p. 750). Many studies look at only a single county. Indeed, one of the largest extant studies within this stream captures only a portion – 53 counties – of a single state (Tewksbury & Mustaine, 2008). The work which has been done to date is valuable and necessary, emphasizing internal validity in studies of selected geographic areas. We extend this work to the state level, allowing for greater generalizability through geographic coverage and diversity.

The current study is the first multi-state study attempting to understand the communities in which RSO’s live. Given the large and unavoidable differences in sex offender registries across states, we do not aggregate the data, but provide separate analyses for each of our ten states in parallel. In the simplest possible terms, the question posed by this study is “For each of these ten states, what are the characteristics of Zip Codes with higher rates of residential RSO’s?”

2. Materials and methods

There were two primary data sets used in this study. The first set consisted of ten downloaded state sex offender registries, all including Zip Codes of residence for RSO’s. The second data source was the American Community Survey 5-year files, accessed as Zip Code level files through Social Explorer.

2.1. State sex offender registry data

We were able to identify ten different states in which sex offender registries were available. This was an improvement over Ackerman et al. (2011), who were only able to identify six downloadable registries as of 2010. Ackerman et al. used a two-step process, utilizing downloadable registries for an initial core of states, then reporting on remaining states through a data-scraping approach. For the current study, we restrict ourselves to the use of existing registries, which we collected during 2018 and 2019. The decision not to employ a more general data-scraping approach was driven by two factors. First, downloading a full dataset exposes the study to fewer threats to data quality than data scraping (Ackerman et al., 2011). Second, focusing on ten states allowed for a detailed analysis capturing the state-by-state similarities and differences resulting from the tension between federal and state policies.

In this study, we began by searching all 50 states’ online sex offender registries. This was done in late 2018 and early 2019. We were able to directly download data files (in Excel spreadsheets or csv formats) for Georgia, Missouri, Illinois, Arizona and Florida. Maryland’s data were available as a Microsoft Word file. North Carolina, Oklahoma and Texas were available as text files and North Dakota was available as a .pdf file. For other states we could either not determine that downloadable files were available or the cost of downloading an entire file was prohibitive due to fees charged for each individual record downloaded. Files which were not in tabular format (e.g., text files or pdf files) were managed in word processing and spreadsheet programs in such a way as to preserve only the residential Zip Codes for non-incarcerated individuals in those states. Tabular datasets were in varying formats and the Missouri dataset was relational, being comprised of several parts.

For all states, data management included the following steps. First, we excluded individuals who either were incarcerated (some state registries include incarcerated persons while others do not), were not currently residents in that state (some registries included individuals convicted within the state who subsequently moved to another state), or had no address listed (e.g., absconders). Second, we removed duplicates from datasets in which multiple records existed for single individuals. Third, we looked to see if it was possible to determine the age of the victim. In five of the ten states (Florida, Missouri, North Carolina, Illinois and Texas), the registries provided enough information to determine if the victim was under the age of 18 at the time of the offense. There are three possible ways that this can be identified in the data. First, datasets can have the age of the victim at the time of the offense, either as age in years or as a minor/adult dichotomous variable. Second, datasets can have the current age of the victim and the date of the offense, allowing a calculation to be made of the victim age at the time of the offense. Third, when datasets lack variables specific to victim age, a text mining approach can be used. This involves searching the offense type for words such as “minor” or “child” or “underage.” In this study, we used the first two approaches almost exclusively. After performing text mining on datasets that had both an offense description and victim age variables, it became clear that the text mining approach could be unreliable. We found that in offenses with minor victims, the description of the offenses did not consistently indicate the victims’ minor status. For this reason, we classified minor victim status only in the five states which directly recorded that variable. The only exception to this was Illinois. For that state we used a text mining approach to fill in missing data for the 1.5% of RSO’s for which victim age was unavailable.

This process resulted in an individual level dataset (i.e., a “long” dataset) with one entry per RSO in each state. An additional long dataset was created for the five states for which victim age was known by keeping only RSO’s with minor victims. Both of these datasets were then transformed to a Zip Code level dataset (one row per Zip Code) with three variables: the Zip Code, the number of RSO’s (all 10 states), and the number of RSO’s with known minor victims (5 states only).

2.2. Census data

Census data was obtained using the five-year estimates from the American Community Survey (ACS). These files were used because they include the required variables of interest at the Zip Code level. We selected Census variables based on consideration of prior findings, relevant theory, and policy relevance. The avowed purpose of much sex offender legislation is to keep RSO’s away from children. We therefore selected variables that would allow us to evaluate the relative distribution of RSO’s in an area and the presence of children in those areas. We also selected measures of socioeconomic status of communities as prior findings have indicated that community poverty and RSO residence are associated. We have included urbanicity and race/ethnic composition as community characteristics in our models. These were not present in most prior models, and are impossible in most individual-level models as ethnicity is generally not recorded. Given the current and long overdue increase in emphasis on social, economic and environmental justice, we feel consideration of these factors is an obvious and needed way to move the field forward.

Lastly, we included the proportion of rental units given the relevance of housing access for this population and the likelihood that a released incarcerated population will move to areas with available rental units.

Variables included in our analysis were as follows: median household income in thousands; three dummy variables indicating the largest racial/ethnic population group (either Non-Hispanic White, Black, or Hispanic) within the Zip Code; percentage of all persons under age 18; percentage of rental housing units; and metropolitan status (coded as “1″ if any part of the Zip Code was within a county classified as part of a Central Metropolitan Statistical Area. Metropolitan status was computed by using county level 2017 ACS data merged with a Zip Code/county crosswalk table from the Office of Policy Development and Research (HUD, n.d.).

The ACS files we included in our analysis contained data collected between 2013 and 2017. These data are therefore centered on 2015. This means that our descriptions of communities are three years earlier than our sex offender registry data. We were not able to avoid this because of the lag time involved in collecting and posting ACS files. Fortunately, the variables we included in our study (e.g., median household income, population size) tend to be extremely stable in the five-year ACS dataset over a three-year period at the Zip Code level. We tested this assumption by downloading population counts and median household income counts from the 2013–2017 five-year estimates and the 2010–2014 five-year estimates. We found high correlations between the 2014 population and the 2017 population (r = 0.998) and between the 2014 median household income and the 2017 median household income (r = 0.966) for Zip Codes with at least 1000 residents (data not shown).

The final dataset was constructed by merging the sex offender registry datafile and the Census datafile at the Zip Code level. Rates of RSO’s per 1000 residents in each Zip Code were calculated for RSO’s in all ten states and also for RSO’s with minor victims in the selected five states. The final dataset is therefore geographic in nature, with each entry representing a single Zip Code.

A persistent consideration in analyzing Zip Code level data is the “Modifiable Areal Unit Problem” (MAUP) (Weschler, Ban & Li, 2019). In this context, “areal units” are simply the areas one seeks to use as a unit of observation. For example, a block group is a smaller areal unit than a tract, and a tract is a smaller areal unit than a county. The MAUP refers to the innate tension between the advantage of smaller areal units (i.e., increased within-area homogeneity) and the advantage of larger areal units (i.e., higher counts of the measured event providing a stable and reliable measure). Studies looking at child maltreatment (Aron et al., 2010; Lery, 2009) have shown that tract and Zip Code areal units work similarly with regard to rare data. These studies identified one problem, however, that Zip Codes with very small numbers of residents tend to be unreliable. These Zip Codes are common as a percentage of all Zip Codes, but include only a very small proportion of the total population.

In this study, we analyzed the data using Zip Code-level cutoffs as recommended by Aron et al. (2006). This process involved performing regression analyses to determine variance explained at different cutoff points. We determined that a 1000 resident cutoff was necessary to minimize error associated with inclusion of small areal units, and that setting the cutoff higher than 1000 residents did little to improve the explanatory power of the statistics used. Among nine of the ten states, less than 3% of all state residents were lost to analysis due to the cutoff and less than 5% of all RSO’s were omitted. In sparsely populated North Dakota, 23% of both residents and RSO’s were lost. With regard to loss of Zip Codes (not persons covered) North Dakota lost 73% of its total Zip Codes. Missouri and Oklahoma each lost about 35% and no other state lost more than 17%. As these numbers show, the “1000 person” cutoff resulted in some loss of numbers of Zip Codes analyzed, but only very small (<5%) losses of total residents or RSO’s covered, except for North Dakota.

2.3. Analytic strategy

We explored the data using descriptive as well as multivariate approaches. All statistics (including national descriptive statistics) were run with Zip Codes having fewer than 1000 residents omitted. First, we calculated total population figures for each state for all included Zip Codes; total numbers of RSO’s in those Zip Codes; total numbers of Zip Codes used in each state; and total numbers of Zip Codes that included at least one RSO (see Table 1). For each state, we then calculated, by Zip Code, median income; percentage of population under age 18; percentage of rental housing units; percentage of persons living in rental units; and percentage of persons living in a Zip Code in a county that is part of a central MSA. We also calculated the percentage of state residents and RSO’s living in Zip Codes in which Black residents outnumber Non-Hispanic White or Hispanic residents. Similarly, we calculated this for those in which Hispanic residents outnumber Non-Hispanic White or Black residents.

Table 1.

Community characteristics for full state and RSO populations.

	Population*		Number of Zip Codes		Median Zipcode Household Income ($)		Minor (%)		Rental Unit (%)		MSA (%)	In Black Plurality Communities (%)			In Hispanic Plurality Communities (%)
USA	316,456,661		22,676		56,750		22.99		31.39		87.5	8.01			14.70
	Population*		Number of Zip Codes		Median Zipcode Household Income ($)		Minor (%)		Rental Unit (%)		MSA (%)	In Black Plurality Communities (%)			In Hispanic Plurality Communities (%)
	State	RSO	State	RSO	State	RSO	State	RSO	State	RSO	State	RSO	State	RSO	State	RSO
AZ	6,770,732	7,245	331	305	51,763	41,485	24.20	26.45	34.73	43.41	93.76	93.01	0.00	0.00	26.37	46.53
FL	20,228,935	28,375	900	876	50,382	44,810	20.77	21.38	32.84	33.75	95.74	91.83	9.77	17.30	18.64	10.51
GA	10,132,441	18,052	614	608	51,290	42,967	25.36	24.93	33.41	37.22	86.81	78.69	25.53	34.93	3.19	1.90
IL	12,637,391	14,363	915	855	60,988	47,538	23.16	23.07	27.28	32.47	89.38	79.77	10.51	16.62	11.03	10.66
MD	7,540,552	5,625	413	321	82,351	64,969	23.13	22.84	31.76	35.16	96.70	93.76	25.67	36.26	4.81	2.83
MO	5,899,767	15,173	621	611	50,147	42,977	23.22	23.26	28.30	32.34	80.38	73.39	8.79	14.37	0.44	1.26
NC	9,987,105	11,853	683	656	48,354	45,223	22.85	23.25	31.87	34.26	79.92	75.36	15.79	19.65	0.14	0.19
ND	642,882	1,093	97	81	64,973	59,679	22.98	21.74	32.85	32.85	54.76	54.76	0.00	0.00	0.30	0.00
OK	3,794,841	4,040	413	381	48,769	43,777	24.54	24.65	30.97	28.80	70.97	64.23	2.11	2.95	3.65	7.23
TX	24,838,274	40,803	1254	1235	56,174	46,490	26.83	26.46	32.11	35.10	91.88	87.15	6.31	12.38	36.48	39.71

^*Interpretation: All numbers reflect included Zip Codes only. Median Zip Code income is the median resident’s Zip Code median household income. Minor (%) is the median resident’s proportion of minors in this Zip Code. Rental Unit (%) is the median resident’s proportion of rental units among all housing units in his Zip Code. MSA (%) is the percentage of residents in a Zip Code which is part of a central MSA. In Black Plurality Communities (%) is the percentage of residents living in Zip Codes with more Black residents than either White or Hispanic residents. In Hispanic Plurality Communities (%) is the percentage of residents living in Zip Codes with more Hispanic than either Black or White residents. “State” columns show median values for all state residents, “RSO” columns show median values for RSO residents only.

These statistics were computed as median values from our Zip Code level data while weighting by total population (in the “State”) column, or by number of RSO’s (The “RSO” column). This yields a value for each variable representing the median person (not Zip Code) in that state. So, for example, the median person in Arizona lives in a Zip Code with a median income of $51,763 (see Table 1 data for Arizona, Column 6) while the median RSO in Arizona lives in a Zip Code with a median income of $41,485 (Column 7).

To further investigate the relationship between rate of RSO’s and community characteristics, a Poisson regression was used to model the count of RSO’s per 1000 persons for each state. These models were run separately, first for all ten states using the number of RSO’s per 1000 residents as the outcome variable, and then for five states using number of RSO’s with minor victims per 1000 residents as the outcome variable. Models were run in SAS 9.4 using PROC GENMOD. Models were tested using a negative binomial and Poisson distribution approach to examine issues related to best fit for the type of outcome modelled. In all cases, there were two to three states that had overdispersion indicators that favored one over the other but in all models the Deviance and Pearson to DF ratios for Poisson were not large. Among the ten states modelling all RSOs, the Lagrange multiplier test in the negative binomial indicated a preference for Poisson for Illinois, North Carolina, and Oklahoma. When limited to minor victims, the test indicated a Poisson model for Texas and Illinois. Because the resulting model coefficients were similar, it was decided to opt for Poisson using the Pearson correction for over-dispersion rather than running different models for subsets of states (Allison, 2012). All models improved over the intercept only models. Comparisons of the deviance for the intercept only model with the model including predictors was used to assess whether the larger model was an improvement (Bartlett, 2014). The incidence rate ratios are an exponentiated function of the parameter estimates and are roughly interpreted as a percent change in the incidence rate. While we did choose to weight the descriptive data by population, we chose not to weight the multivariate analyses. This is because while the descriptive data are meant to provide a picture of each state as a whole, the multivariate analyses are meant to allow us the opportunity to better understand how different types of communities result in different rates of RSO residence within each state.

3. Results

Descriptive statistics can be found in Table 1. The population and number of RSO’s are listed for each state, along with counts of Zip Codes analyzed in each state and numbers of Zip Codes containing RSO’s. States ranged considerably in population, number of Zip Codes and RSO’s, with the largest numbers being in Texas (population of 24,838,274 in 1254 Zip Codes with 40,803 RSO’s) and the fewest being in North Dakota (population of 642,882 in 97 Zip Codes with 4040 RSO’s). Median Zip Code income was always considerably lower for RSO’s than for the general population in every state, ranging from 77.9% of the income of the general population in Illinois to 93.5% in North Carolina (percentages calculated from Columns 6 and 7 on Table 1).

Contrary to expectation and policy intent, there was very little variation in the median numbers of minors in Zip Codes for the general population and for the population of RSO’s (never more than 1.25 percentage points). In eight states, RSO’s did live in Zip Codes with more rental units, the exceptions being Oklahoma and North Dakota. Interestingly, RSO’s were less likely to live in Central MSA’s in all states except North Dakota. At the descriptive level, a stark racial imbalance was visible, with all states showing a higher percentage of RSO’s living in Zip Codes with more Black residents than Non-Hispanic White or Hispanic residents, except for Arizona and North Dakota, which had no such Zip Codes. Resident/RSO imbalance by ethnicity was inconsistent, with some states having notably higher proportions of RSO’s in Hispanic Zip Codes (Arizona and Oklahoma) while others showed the opposite (Florida and Georgia).

Multivariate models for each state can be seen in Tables 2a and 2b. Multivariate findings mirrored the bivariate findings with regard to the central importance of income. These findings fell within a surprisingly narrow band, showing a consistent effect across all states except North Dakota, which was nonsignificant for all variables in the model. Other states showed decreases in rates of RSO’s between 2.06% (Maryland) and 3.82% (Florida) for each $1000 increase in median household income.

Table 2a.

Poisson models predicting RSO residence rates (AZ-MD).

	Arizona	Florida	Georgia	Illinois	Maryland
Exponentiated Parameter Estimates (95% CI’s)
Median Household Income	−3.18 (−3.98 to −2.38)****	−3.82 (−4.33 to −3.31)****	−3.09 (−3.49 to −2.69)****	−3.58 (−3.99 to −3.17)****	− 2.06 (−2.36 to −1.74)****
Minors (%)	−1.94 (−3.29 to − 0.38)**	3.32 (2.06–4.60)****	−0.36 (−1.31 to 0.60)	−2.25 (−3.28 to −1.22)****	0.59 (−1.04 to 2.27)
Hispanics Largest Group in Zip†	20.18 (−9.20 to 59.07)	− 40.88 (−54.95 to − 22.43)****	− 40.20 (−65.02 to 2.22)	− 25.24 (−44.97 to 1.55)	−49.74 (−70.39 to 14.69)
Blacks Largest Group in Zip†	0	45.54 (17.88–79.70)****	3.89 (−9.29 to 18.99)	−10.03 (−27.20 to 11.20)	8.65 (−10.12 to 31.36)
Central Metropolitan Area	9.63 (−19.89 to 50.02)	−7.58 (−18.71 to 5.10)	5.05 (−5.81 to 17.16)	− 0.13 (−11.38 to 12.56)	−12.74 (−24.41 to 0.72)
Rental Residents (%)	1.28 (0.54–2.02)***	− 2.41 (−2.97 to −1.84)***	− 0.96 (−1.36 to − 0.56)***	−0.57 (−1.07 to − 0.06)*	− 0.45 (−1.01 to 0.13)
Model Deviance (df): Intercept Only	423.39 (304)	1433.89 (875)	655.76 (607)	1019.12 (854)	197.41 (320)
With Covariates	262.05 (298)	805.47 (864)	369.38 (601)	531.56 (848)	97.25 (314)
Chi Square Difference	161.34 (6)***	628.42 (11)***	286.38 (6)***	487.56 (6)***	100.16 (6)***

^*p ≤ .05

^**p ≤ .01

^***p ≤ .001

^****p ≤ .0001

^†Reference Group is those Zip Codes with Whites as largest group.

Table 2b.

Poisson models predicting RSO residence rates (MO-TX).

	Missouri	North Carolina	North Dakota	Oklahoma	Texas
Exponentiated Parameter Estimates (95% CI’s)
Median Household Income	−2.47 (−3.00 to −1.94)****	−2.36 (−3.02 to −1.72)***	0.18 (−0.91 to 1.29)	−2.37 (−3.21 to −1.53)****	−2.57 (−2.82 to −2.31)****
Minors (%)	−2.20 (−3.22 to −1.16)****	−1.03 (−2.62 to 0.59)	−2.97 (−6.34 to 0.54)	0.64 (−1.38 to 2.71)	−1.56 (−2.23 to −0.88)****
Hispanics Largest Group in Zip†	50.73 (−13.58 to 162.90)	−38.47 (−87.58 to 80.71)	0	117.08 (25.55–275.35)**	−10.39 (−19.11 to −0.72))*
Blacks Largest Group in Zip†	2.15 (−21.13 to 32.31)	0.47 (−18.63 to 23.28)	0	23.40 (−32.40 to 125.29)	47.45 (26.99–71.21)****
Central Metropolitan Area	−0.15 (−13.77 to 15.62)	−3.43 (−16.96 to 12.28)	1.49 (−26.57 to 40.28)	25.40 (2.74–53.04)*	8.67 (−0.74 to 18.96)
Rental Residents (%)	−2.47 (−3.00 to −1.94)****	−2.36 (−3.02 to −1.72)***	0.18 (−0.91 to 1.29)	−2.37 (−3.21 to −1.53)****	−2.57 (−2.82 to −2.31)****
Model Deviance (df): Intercept Only	941.90 (610)	445.57 (655)	51.82 (80)	344.78 (380)	1642.37 (1234)
With Covariates	641.95 (604)	349.28 (649)	48.76 (76)	280.77 (374)	1014.52 (1228)
Chi Square Difference	299.95 (6)***	96.29 (6)***	3.06 (4)	64.01 (6)***	627.85 (6)***

^*p ≤ .05

^**p ≤ .01

^***p ≤ .001

^****p ≤ .0001

^†Reference Group is those Zip Codes with Whites as largest group.

In four states (Arizona, Illinois, Missouri, and Texas) there were somewhat fewer RSO’s in Zip Codes with lower proportions of children in the population (around two percent fewer RSO’s per one percent fewer children as a proportion of all residents) while in one state (Florida) RSO’s were more commonly found in Zip Codes with more children. Surprisingly to our team, metropolitan status was nonsignificant in all but two states. In Florida, Zip Codes in counties which were part of central metropolitan statistical areas had 7.58% fewer RSO’s, while in Oklahoma, they had 25.40% more RSO’s. A higher proportion of rental properties was associated with fewer RSO’s in five states (Florida, Georgia, Illinois, Oklahoma, and Texas) but slightly higher numbers of RSO’s in Arizona. Racial findings diverged from bivariate models. In the bivariate results, RSO’s were more likely to live in communities in which Blacks were the largest group in all states except North Dakota and Arizona (both of which had no such Zip Codes). This difference became nonsignificant in the multivariate models for all states except Florida and Texas, which did show a higher number of RSO’s (+47.45% and +45.54% respectively) in such communities. While most states showed no significant difference for Zip Codes with a largest proportion of Hispanics, two states showed lower rates (Florida with −40.88% and Texas with −10.39%) while Oklahoma showed considerably higher rates (+117.08%).

Multivariate analyses were repeated for those five states for which minor victim status could be tracked (Florida, Illinois, Missouri, North Carolina, and Texas). Results are presented in Table 3. Findings were generally the same as presented in Tables 2a and 2b with all statistically significant variables for each state and the valence of each statistically significant relationship remaining the same.

Table 3.

Poisson models predicting RSO residence rates (minor victims only).

	Florida	Illinois	Missouri	North Carolina	Texas
Exponentiated Parameter Estimates (95% Cl’s)
Median Household Income	−3.94 (−4.48 to −3.40)****	−3.66 (−4.07 to −3.25)****	−2.53 (−3.06 to −1.99)****	−2.22 (−2.94 to −1.51)****	−2.56 (−2.82 to −2.29)****
Minors (%)	3.63 (2.30–4.99)****	−2.27 (−3.30 to −1.22)***	−1.93 (−2.97 to −0.89)****	−0.43 (−2.24 to 1.43)	−1.58 (−2.27 to −0.88)****
Hispanics Largest Group in Zip†	−42.14 (−56.73 to −22.64)	−26.96 (−46.55 to 0.21)	47.01 (−16.15 to 157.72)	−7.95 (−72.00 to 209.56)	−10.86 (−19.75 to −1.00)*
Blacks Largest Group in Zip†	46.42 (17.19–82.92)****	25.36 (−40.48 to 6.39)*	−9.12 (−30.72 to 19.21)	−2.42 (−23.02 to 23.69)	39.15 (18.95–62.78)****
Central Metropolitan Area	−5.64 (−20.48 to 11.99)	−0.22 (−11.42 to 12.41)	0.24 (−13.47 to 16.13)	21.81 (−2.50 to 44.74)	6.93 (−2.51 to 17.28)
Rental Residents (%)	−2.66 (−3.26 to −2.06)***	−0.74 (−1.24 to −0.22)*	0.01 (−0.52 to 0.54)	0.61 (−0.12 to 1.35)	−1.34 (−1.64 to −1.04)***
Model Deviance (df):Intercept Only	1291.65 (875)	950.83 (854)	847.44 (655)	333.08 (655)	1526.29 (1234)
With Covariates	733.43 (869)	512.84 (848)	589.25 (649)	294.69 (649)	960.67 (1228)
Chi Square Difference	558.22(6)***	437.99 (6)***	258.19 (6)***	38.39 (6)***	565.62 (6)***

^*p < .05

^**p < .01

^***p < .001

^****p < .0001.

^†Reference Group is those Zip Codes with Whites as largest group.

Finally, given the central importance of the relationship between median household income and the rate of RSO’s in a Zip Code, we provide visual scatterplots showing these relationships in each state (Fig. 1). It is important to recall that our study does not show all Zip Codes in a state, only those with at least 1000 residents. In addition, states were all forced to the same scale on both the X axis (Median Household Income) and Y axis (RSO rate per 1000 residents) to allow cross state comparison. For visual clarity some visual compression was used. Across all 6241 Zip Codes, 38 Zip Codes with median household incomes over $150,000 were shown as $150,000 and 42 Zip Codes with perpetrator rates over 10 were shown as 10. Clear asymptotic relationships (rapidly declining RSO’s resident rates with increasing Zip Code median household income) are visible for all states except for North Dakota and Oklahoma.

Fig. 1.

Scatterplot of Zip Codes by Rate of RSO Residence and Median Zip Code Household Income.

As a form of informal “post hoc” analysis of the generated distributions in Fig. 1, we can consider these distributions in the context of the severity of Residence Restriction (RR) rules in each state (Florida Action Committee, 2019). These rules set minimum distances from schools, parks or similar locations within which RSO’s may not live. Maryland has no restrictions, while Texas restrictions are on a case-by-case basis. North Dakota and Illinois have 500 foot radii, while Arizona, Florida, Missouri, North Carolina and Georgia have 1000 foot limits. There does not appear to be any consistent pattern, with the most restrictive state (Oklahoma) having a relatively scattered distribution, along with North Dakota, which has a relatively non-restrictive policy. Maryland, with no restrictions, may show a somewhat less defined relationship than states with the most common (1000 foot) established radii. In sum, a brief inspection of the scatterplots shows no clear and consistent association between severity of restrictions and the association between rates of RSO residence and median Zip Code household incomes.

4. Discussion

Our central finding, seen visually in Fig. 1 and statistically in Tables 1 and 2a/b, is the powerful inverse and somewhat curvilinear relationship between zipcode median household income and the rate of RSO’s in a given Zip Code. This relationship is relatively consistent across all states except for North Dakota and Oklahoma. In Fig. 1, it can be seen that living in a Zip Code with a median income above the national median household income ($57,652 for the 2013–2017 American Community Survey (Census Bureau, n.d.)) makes it very unlikely that there will be a high rate of residential RSO’s in your neighborhood. Very low-income neighborhoods lack such consistency, showing a wide distribution between low and high RSO density. The vast majority of high-density neighborhoods, however, are in low-income areas. With the exceptions of North Dakota and Oklahoma, surprisingly few outlier Zip Codes exist in any state – the two most prominent being in Illinois. These two Zip Codes (60602 and 60606) are adjacent Zip Codes in central downtown Chicago with a very broad income distribution.

Racial differences in the bivariate findings largely disappeared in multivariate models including income. This is a well-established phenomenon in the general child maltreatment literature (e.g. Kim & Drake, 2018; Putnam-Hornstein, Needell, King, & Johnson-Motoyama, 2013) and suggests that the increased presence of RSO’s in minority communities may largely be an artifact of the pervasive economic racial stratification in the United States. Increased rates of RSO’s in these communities may contribute to existing structural inequalities. For example, studies have found that proximity to an RSO may reduce property values (Kernsmith, Craun, & Foster, 2009; Linden & Rockoff, 2008). In this way, the concentration of RSO’s in poorer neighborhoods may serve as a further challenge to such neighborhoods in attracting more affluent residents and improving community economic status.

Other findings were modest and varied by state. Idiosyncratic differences in Zip Code boundaries or any number of other state-level geographic differences could lead to the inconsistencies we found. It is also possible that variations in state-level policies regarding RSO residence restrictions and classifications may make a difference. For example, while Missouri restricts all RSO’s from living within 1000 feet of schools, playgrounds, and parks, Oklahoma restricts them from living within 2000 feet. Texas restricts those with child victims on a case-by-case basis and has many municipalities with their own local residence restrictions. And Maryland has no residence restrictions at all (Savage & Windsor, 2018).

Strengths of the current study include the provision of the first ever large-scale community-level analysis of RSO residence, along with the ability to present data from ten different states. Limitations of the current study, particularly with regard to cross-state comparison, stem largely from the variations in sex offender policy in each state. In addition, more detailed analyses by race of the RSO were not possible due to the nature of current registries which tend to exclude ethnicity (e.g. Hispanic, not Hispanic) as a variable. Analyses by gender of the RSO were also precluded by the very low number of female RSO’s.

In summary, the current study suggests that very low-income communities are home to a vastly disproportionate number of RSO’s. At the other end of the income spectrum, higher income communities are almost uniformly insulated from the presence of RSO’s. It is impossible to say if this is a function of the fact that lower income communities have higher rates of sex offenders, who return to their home communities, or if the tremendously reduced financial opportunities of people who have been convicted of crimes or been incarcerated are responsible for the observed distribution. We suspect that both factors contribute.