Crime

The NEWS and NEWS-A (abbreviated version) consist of 67 and 54 items, respectively [12 (link)]. These are grouped into eight multi-item subscales (representing distinct constructs or latent factors) including perceived residential density; proximity to nonresidential land uses (land use mix – diversity); ease of access to nonresidential uses (land use mix – access); street connectivity; infrastructure for walking and cycling; aesthetics; traffic safety; and safety from crime. The first two subscales are not factor-analyzable and, hence, represent constructs rather than latent factors. Five single-item subscales (four in the NEWS-A) assess perceived major physical barriers to walking; hilly streets; difficult car parking in shopping areas; absence of cul-de-sacs; and presence of people being active in the neighborhood (not included in the NEWS-A). All subscales, with the exception of residential density and land use mix – diversity, are rated on a 4-point Likert scale. Residential density items are rated on a 5-point scale, and ratings are weighted relative to the average residential density that a specific item represents [11 (link)]. The weighted ratings are summed to create a perceived residential density score. Land use mix – diversity is assessed by the perceived walking proximity from home to various types of destinations, with responses ranging from 1- to 5-minute walking distance (coded as 5) to >30-min walking distance (coded as 1).

The original English [17 (link)] and Chinese translation [18 (link)] of the Neighborhood Environment Walkability Scale - Abbreviated (NEWS-A) were the source instruments used to develop the Neighborhood Environment Walkability Scale for Chinese Seniors (NEWS-CS). The original NEWS-A assesses perceived environmental characteristics, stemming in part from the urban planning literature, believed to influence walking and other forms of physical activity [16 (link),17 (link)]. It consists of 54 items. These are grouped into subscales including perceived residential density, proximity to non-residential land uses (land use mix - diversity), ease of access to non-residential uses (land use mix - access), street connectivity, infrastructure for walking and cycling, aesthetics, traffic safety and safety from crime. Four single items assess hilly streets, difficult car parking in shopping areas, absence of cul-de-sacs, and perceived major physical barriers to walking. All subscales and single items, with the exception of residential density and land use mix - diversity, are rated on a 4-point Likert scale. Residential density items use a 5-point scale, whereby ratings are weighted relative to the average residential density that a specific item represents. The weighted ratings are summed to create a perceived residential density score. Land use mix - diversity is assessed by the perceived walking proximity from home to various destinations, with responses ranging from 1- to 5-minute to >30-min walking distance. The NEWS-A has been shown to have acceptable levels of validity and reliability [16 (link)-18 (link)]. A study on Hong Kong adults reported high levels of test-retest reliability of the Chinese version of the NEWS-A [18 (link)].
The version of the NEWS-CS examined in this study encompasses all but two items of the NEWS-A, in their original or slightly modified form, and 24 additional items describing features of the environment relevant to the study setting and senior residents (see Development and translation section below).

Covariates including: age, sex, educational level (assessed based on total number of years of education categorized into three groups: 0 = 1–9 years, 1 = 10–12 years, 2 = > 12 years), relationship status (0 = single/widow, 1 = married/cohabiting), and perceived discrimination (assessed with the survey item: “Have you ever been discriminated against in a way that was highly distressing or disturbing because of your race, ethnic group, gender, sexual orientation, or religion?”: 0 = No, 1 = Yes) and/or hate crime (assessed with the survey item: “Have you ever been the victim of a hate crime?”: 0 = No, 1 = Yes) were measured at the baseline, that is, at the time of STAGE interview. Information on annual disposable income in 2016 were retrieved from the longitudinal integrated database for health insurance and labor market studies (LISA), Statistics Sweden [27 (link)]. Information on previous prescription of antidepressants (ATC: N06A) during the follow-up period 2006–2018 was retrieved from the Swedish Prescribed Drug Registry held at the National Board of Health and Welfare. Use of antidepressant was added as a covariate in the analyses as an indicator of common mental health disorders since the risk of depression has consistently been found elevated in sexual minority populations [11 (link)].

The dependent variables of crime counts are significantly skewed and overdispersed (i.e., the variance exceeds the mean). Thus, we analyze the spatial distribution of crime using negative binomial regression in Stata 17; a Poisson-based regression that effectively accounts for overdispersion via its alpha parameter [49 , 50 ]. While the Poisson distribution can be appropriately used to model certain count variables, for the present study, we find that Stata’s likelihood-ratio test, which tests the null hypothesis that the dispersion parameter (alpha) is equal to zero, is significant for all our models (p < 0.05). Negative binomial regression is therefore needed to account for overdispersion [49 , 50 ]. Yet at the same time, we acknowledge that ordinary least squares regression (OLS) is a viable alternative for modeling the spatial distribution of crime, especially given that a large majority of block groups do not have zero crime incidents, and therefore we have estimated ancillary models using OLS. To be clear, we are concerned with using the appropriate model(s) to analyze the spatial distribution of crime, however, it should be noted that negative binomial and OLS regression models are not representative of spatial regression. Therefore, we estimate ancillary spatial error models as a final robustness check (described in more detail below).
Geographic units such as block groups are not islands unto themselves [51 ], in fact, the conditions of spatially contiguous/adjacent units can very well shape what occurs in the focal unit—what is often referred to as a spillover effect. What this means for the current study is that crime in the focal block group is likely impacted by the amount of crime in nearby block groups [52 –54 (link)]. To account for this spatial dependence, we constructed a spatially lagged measure for each crime outcome using GeoDa software with first-order queen contiguity. Such a measure captures the average number of crime incidents among contiguous block groups in relation to the focal block group. We include a spatially lagged measure of crime (as a predictor) in our full models.
A general expression of the (full) negative binomial regression models that we estimate is as follows:
$$\mathrm{y}={\mathrm{B}}_1{\mathrm{P}\mathrm{O}\mathrm{W}}_{}+{\mathrm{B}}_2{\mathbf{S}\mathbf{D}}_{}+{\mathrm{B}}_3\mathbf{C}\mathbf{F}+{\mathrm{B}}_4\mathrm{S}\mathrm{L}\mathrm{y}+\mathrm{\alpha },$$
where y is the number of crime incidents, POW is the number of places of worship, SD is a matrix of the sociodemographic characteristic measures, CF is a matrix of the criminogenic facility measures, SLy is the average number of crime incidents in block groups adjacent to the focal block group (a spatially lagged measure), and α is an intercept.
While one approach for modeling crime across geographic units is to specify the population count as an exposure term (thereby estimating the outcome as a crime rate), we have instead modeled crime counts by including the population count as a predictor, given growing concerns over population count being the denominator of a calculated crime rate [e.g., see 18 , 55 (link)–58 (link)]. As anticipated, we detected minimal evidence of spatial autocorrelation in our full models as a result of including the spatially lagged measure of crime. Although the Moran’s I value was statistically significant in all instances, the maximum value was .08 (which is rather weak given that positive spatial autocorrelation ranges from 0 to 1). Furthermore, we assessed and found no evidence of multicollinearity issues based on variance inflation factors (VIF). The maximum VIF was 4.68, which does not exceed the commonly used cutoff of 10 [59 , 60 (link)].
In the results section, we present two models for both the violent and property crime outcomes (Table 2). We first estimate a baseline model that features our places of worship measure along with the sociodemographic characteristic measures, consistent with the modeling approach undertaken by certain prior studies [for example, see 6 (link), 8 (link), 12 ]. We then estimate a full model that additionally includes the measures of well-established criminogenic facilities and the spatially lagged measure of crime in order to determine the extent to which places of worship maintains a significant effect on crime (if at all). Crime and place researchers have called for analyses to integrate measures associated with social disorganization and routine activities theories simultaneously [for example see, 33 , 61 (link)]; therefore, our full model is consistent with this call.
In addition to discussing the observed effects in terms of their direction and statistical significance, we highlight the magnitude of these effects in relation to one another. We draw on an approach that determines the percent change in the expected crime count for a one standard deviation increase in the variable of interest using the following formula: (exp(β× SD)– 1) *100. This is a preferred approach because some of our independent variables drastically differ in terms of their scales [49: 492–493, 514–516.], most notably, the POW and facility measures are counts whereas the sociodemographic characteristic measures are percentages. Similar to previous crime and place studies [62 (link)–65 (link)] we utilize this approach to effectively compare the effect sizes of variables with substantively different scales.
On the other hand, we recognize that another common approach is to assess the magnitude of the effects using incident rate ratios (IRR). Specifically, an IRR denotes the percent increase or decrease for every one-unit increase in a predictor by multiplying the difference between the IRR and one by 100 where positive values yield a percent increase and negative values yield a percent decrease [49 ]. In Table 3, we compute the effect sizes using both approaches, although we base our inferences on the first approach because for the second approach, a one-unit increase may represent a very large increase for one predictor (e.g., DC Metro Station) and a very small increase for another predictor (e.g., population).

We have collected crime data from DC’s open data portal. These are official crime data coded and reported by the District of Columbia Metropolitan Police Department (MPD). MPD has provided crucial information on each crime incident from 2021, including the longitude–latitude coordinates of the crime, the date in which the crime occurred, and the type of Part 1 crime committed according to the Uniform Crime Reporting (UCR) program in the United States. Accordingly, we aggregated these data to their constituent block group and computed the number of incidents for the following crime types: murders, robberies, aggravated assaults with a gun, burglaries, larcenies, and motor vehicle thefts. Furthermore, we created an index of violent crimes (combining murders, robberies, and assaults) along with an index of property crimes (combining burglaries, larcenies, and motor thefts). Notably, our main models utilize the latter two indices as outcome measures, whereas some of the ancillary models assess each of the crime types (separately) that comprise both indices.