The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study

Vaccine xxx (2018) xxx–xxx

Contents lists available at ScienceDirect

Vaccine journal homepage: www.elsevier.com/locate/vaccine

The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study Ashley Gromis a, Kayuet Liu b,⇑ a b

Department of Sociology, Princeton University, United States Department of Sociology, UCLA, United States

a r t i c l e

i n f o

Article history: Received 14 May 2018 Received in revised form 7 September 2018 Accepted 18 September 2018 Available online xxxx Keywords: Vaccine refusals Personal belief exemptions Autism Race and ethnicity Social diffusion School vaccine requirements

a b s t r a c t The number of children entering schools without mandated vaccinations has increased in high-income countries due to the rise of nonmedical exemptions from school vaccination requirements. Herd immunity is threatened when unvaccinated children are concentrated in spatial pockets. Despite the role of vaccine-autism controversy in the current wave of the anti-vaccine movement, we do not know if exemption clusters are associated with local autism rates; it is often assumed that these clusters are merely the result of sociodemographic composition. This study uses data on the number of students with a Personal Belief Exemption reported by schools from 1992 to 2014 and unique data on the locations of children with an autism diagnosis in California to study the correlates of large exemption pockets. Our spatial analysis shows that the prevalence of autism is not associated with the locations of large pockets of vaccination exemptions. Likewise, the spatial distributions of socioeconomic factors and proximity to health care resources have limited roles in explaining these large exemption pockets. Racial/ethnic composition, however, has strong associations with the locations of the large pockets. Our results suggest that community-level interventions are needed to maintain herd immunity as exemption pockets are not merely the result of population composition. Ó 2018 Published by Elsevier Ltd.

1. Introduction Parental concerns over vaccine safety are suggested to have overtaken health care access as the primary obstacle to immunization in high-income countries [1]. In the United States, parents are required by law to provide documentation of immunization for enrollment to any school or childcare facility. However, increasing numbers of parents are seeking non-medical exemptions to vaccine requirements [2]. Non-medical exemptions differ from medical exemptions in that they are requested by parents on the basis of personal or religious beliefs rather than granted by medical professionals when vaccination may present adverse health risks. Parents who have sought non-medical exemptions are more likely to have expressed strong vaccine concerns or vaccine hesitancy [3]. Anti-vaccine beliefs have been shown to be difficult to change, and policy changes to tighten immunization requirements are often met with resistance [4,5]. Vaccine refusals threaten herd immunity, which requires immunization levels of approximately 90% [6].1 ⇑ Corresponding author at: 264 Haines Hall, 375 Portola Plaza, Los Angeles, CA 90095-1551, United States. E-mail address: [email protected] (K. Liu). 1 Higher immunization levels are required for extremely infectious diseases, e.g., pertussis, and lower levels for some less infectious diseases.

Although population vaccination levels continue to meet this threshold,2 the spatial clustering of children with non-medical exemptions [8–10] can compromise herd immunity locally. Concentrations of exempted children in certain areas increase the risk of vaccinepreventable disease outbreaks [11,12]. Additionally, private schools, particularly those that adopt alternative education methods, have been shown to have higher rates of PBEs than public schools [13,14]. When other interventions are ineffective or infeasible, preventing clusters of children with non-medical exemptions can preserve herd immunity. This study seeks to identify mechanisms underlying non-medical vaccine exemption clusters. First, we explore how the discredited association between vaccines and autism might have affected the spatial patterns of exemptions. Despite the importance of the discredited link between the Measles, Mumps and Rubella (MMR) vaccine and autism in the current wave of the anti-vaccine movement, this is the first study to use addresslevel data of children with autism diagnoses to study the relationship between prevalence of autism and clusters of vaccine exemptions. Like vaccine exemptions, autism diagnoses cluster spatially [15,16] and have a positive socioeconomic (SES) gradient [17,18].

2 For reference, the average non-medical exemption rate among kindergarteners in California was 2.54% in 2014, still far below 10% [7].

https://doi.org/10.1016/j.vaccine.2018.09.038 0264-410X/Ó 2018 Published by Elsevier Ltd.

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

2

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

Although many studies demonstrated that vaccines do not cause autism (see [19]), a substantial number of parents are still concerned about the alleged link [20,21]. Specifically, how worried parents are about autism and vaccines may depend on where they live: parents living in neighborhoods with higher rates of autism may be more concerned about vaccine safety than parents living in neighborhoods where autism is rare. Vicarious learning—learning derived indirectly from hearing or observing the experiences of family members, friends, neighbors or acquaintances – has been consistently shown to affect our perception of health risks [22,23]. Our address-level data on autism allow us to evaluate whether the clusters of vaccine exemption and autism diagnoses overlap spatially. Second, previous studies have shown that non-medical exemptions are more prevalent in predominately white, high SES neighborhoods [8–10,13,24–27]. Various mechanisms could account for the positive SES gradient of these exemptions, such the perceived risks of contagious diseases [28]. Despite the many findings on SES and non-medical exemptions, the relationship between SES and vaccine-related beliefs is unclear [29]. Low SES and ethnic minority parents have more vaccine safety concerns than high SES parents, even though they still vaccinate their children [20,30]. That said, a concentration of high SES individuals among the small number of parents who actually seek exemptions is sufficient to generate a positive SES gradient. This study seeks to move beyond identifying correlates of exemption clusters by asking: is the sorting of families by SES into certain neighborhoods sufficient to explain large pockets of exemptions that span multiple schools? Why do neighborhoods with residents of similar SES have drastically different exemption rates? Moreover, do some sociodemographic factors have a limited geographical span of influence while others are more important for broader clustering? Understanding the geographical span of the socio-demographic correlates of exemptions will shed light on the underlying mechanisms generating these clusters and has implications for how policy may better address preventing clusters of unvaccinated children.

2. Methods 2.1. Data 2.1.1. Personal beliefs exemptions This study uses the California Department of Public Health (CDPH) annual data on non-medical, Personal Belief Exemptions (PBEs) from school vaccination requirements from 1998 to 2014.3 Schools and licensed childcare centers are required to report exemptions to the CDPH.4 The CDPH releases this data for schools and childcare centers with 10 or more students enrolled. In January 2014, California policy for filing a PBE was changed to include a requirement that parents submit documents signed by a health care provider indicating that they had been made aware of the benefits and risk of vaccination as well as health risks of vaccine-preventable diseases.5 This was followed by a slight reduction in the percentage of kindergarteners entering school with PBEs for the 2014–15 school year.6 3 Although PBE status is indicative of lack of immunization or under immunization, our data do not tell us how many required vaccinations a child is lacking. See [31] for coverage estimates. 4 This data is available to be downloaded from the CDPH. http://www.shotsforschool.org/k-12/reporting-data/. Accessed May 11, 2018. 5 This law change (AB 2109) did create a new exemption for religious beliefs that did not require a signature by a health care provider, but of the new PBEs filed in the 2014–2015 school year, the majority (77%) were obtained through a Health Care Practitioner Counseled exemption [7]. 6 For the 2013–14 school year, 3.15% of California kindergarteners entered school with a PBE. In the 2014–15 school year, the comparable statistic was 2.54% [7].

Our regression analyses are conducted at the school level and the primary dependent variable is the school-level count of enrolled kindergarteners with PBEs between 1998 and 2014.7 We matched the CDPH PBE data to the California Department of Education (CDE) School Listing files to obtain the location of schools. This matching also provides indicators for school type (public noncharter, charter, or private), and the public school district the school is located in. School type is an important control as selection into private and charter schools8 with alternative pedagogical orientations may increase PBE rates [14,27]. Our spatial analyses focus on 2014 in particular to yield a conservative estimate of non-medical exemptions that were primarily driven by vaccine-related beliefs. Parents who filed these exemptions in 2014 would have been required to discuss vaccination with a health care provider before the exemption was obtained.9 2.1.2. Socioeconomic factors We construct a ‘‘500-child radius” around each school to capture the average household/neighborhood characteristics in its vicinity. The radius is calculated as the distance (in kilometers) from the geographic coordinates of the school to the residence of the 500th nearest child (see synthetic population description below). We chose 500 children as it is roughly the average enrollment size10 of an elementary school; additional tests indicate our results are fairly robust to the choice of buffer size.11 We use the 500-child radius to capture a school’s ‘‘catchment area” as we cannot directly observe school-specific enrollment areas.12 As the catchment areas of smaller schools can be similar to that of larger schools, a 500-child radius is used for each school, regardless of enrollment size. Using the same number of children in our calculations for all schools also ensures our estimates are not biased by small numbers. In the same way that children from the same block may attend different schools, the 500-child radius for two or more schools may also overlap.13 This flexible geographical unit also allows us to compare rural and urban areas. The 500-child radius around each school is calculated using a synthetic population we created yearly for the 1992–2010 period. The synthetic population is based on block-level population data of 3 to 9 year-olds using the 1990, 2000 and 2010 U.S. Censuses, supplemented by address-level California Birth Statistical Master Files (BMSF) data on all 8 million births in California from 1992 to 2007 and by ESRI Sourcebook [32]. The birth record data allows us to add new children to the synthetic population between the census years. A residential movement model based on observed moves of 2.8 million children was used to relocate some of the population

7

Individual-level data on PBEs are unavailable do to privacy concerns. Charter schools are independently run and can adopt alternative pedagogical orientations even though they are publicly funded. Admissions to charter schools are not constrained by residence in a particular school district. 9 This would not apply to the 23% of PBEs that were filed as religious exemptions (see [7] for more information. 10 The average public elementary school size in California is 596 in 2017. The number is lower when private schools are included. See https://www.publicschoolreview.com/average-school-size-stats/national-data. Accessed May 1, 2017. 11 We estimated spatial lag and spatial error models to examine how a range of reasonable buffer sizes may affect the results. Estimated coefficient effects did not vary significantly between models. Results available upon request. 12 The mean radius size for schools was 1.62 km (just over 1 mile) while the median radius size was 0.84 km (just over 0.5 mile). The radius size distribution is right skewed due to larger radius distances in rural areas. 13 A potential concern is that such overlapping radii would induce spatial autocorrelation and lead to underestimated standard errors. We estimated spatial Durbin models which explicitly measure the spillover effects of covariates. We also estimated these models with various nearest neighborhood sizes. Direct effects (school covariate effects on its own PBE rate) were similar to those presented in this analysis. Results available upon request. 8

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

in the intermittent years between the censuses. This synthetic population was previously used in an empirically-calibrated simulation to study the diffusion of autism among children in California (see [33] for futher details). For each child in the synthetic population, age and race/ethnicity are assigned based on census block-level data, or the BMSF data for children added in the intermittent years. We imputed mother’s highest year of schooling based on their distributions by zip code and racial groups from the birth record data. Logged median property values and population density were assigned based on estimates from linear interpolations14 of block-group level data from the three censuses and ESRI data from 2002 to 2004. Each child in this synthetic population is assigned a latitude and a longitude based on the centroid of his/her census block or the BMSF address. Using the synthetic population for this analysis, we can get both the kilometer distance to the 500th child and a demographic profile of the synthetic children within the 500 child radius for each school. We average across the SES characteristics we assigned to the children within a school’s radius to yield school-level measures of (1) mother’s years of education, (2) logged property values, (3) logged population density, and (4) percent non-Hispanic white. We are not able to consider other racial/ ethnic categories as their small numbers in certain areas yield unstable estimates. 2.1.3. Autism In addition to the socioeconomic measures described above, we use the 500-child radius to construct count measures that may be associated with PBE rates, e.g., the number of children with autism. California is the only U.S. state that uses a centralized system for both the assessment of and service provision to the majority of individuals with autism. In California, the Department of Developmental Disorders (DDS) routinely collects clinical and demographic information of clients with a diagnosis of autism.15 DDS evaluation is required for a wide range of service provision. Over 70% of children under age 10 with Autism are in the DDS system [34] and most children were evaluated every year. From the DDS we obtained address information for children with autism from 1992 to 2010, which provides a unique opportunity to study the associations between autism prevalence and vaccination exemptions. Further details of the data are described elsewhere [17]. For the following analyses, our measure of local prevalence is the count of children under age 10 with an autism diagnosis within the 500-child radius described above around a school in each year.16 2.1.4. Physicians Access to health care is traditionally a major determinant of immunization level. Access to physicians is measured as the count of general, family, and pediatric physicians within the 500-child radius around each school. These counts are created from yearly Directory of Physician data released by the American Medical Association from 1992 to 2012. This data includes office addresses for practitioners. As with our count measure of the number children with autism, we create these counts by first calculating the distance to the nearest physicians (using geocoded office locations)

14

Cubic interpolations yield similar estimates. The DDS does not provide services to people with Asperger Syndrome or Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS). 16 We do this by calculating each school/childcare’s distance to the 20 nearest children with autism. We then count how many of those children live within the school’s distance to the 500th nearest child. There were no schools that had more than 20 children with autism within the 500-child radius. A higher threshold would have been set if more than 20 children with autism could have been found within any school’s radius. 15

3

and then counting how many of these physicians fell within the 500-child radius around the school.17 2.1.5. Alternative health practitioners Exposure to alternative health practitioners has been suggested to increase non-medical exemptions [28,35,36]. To test the effect of proximity to alternative health practitioners, we construct annual counts of chiropractors and acupuncturists within a school’s 500-child radius using licensing records from the California Department of Consumer Affairs. These records include both the addresses of licensed practitioners, as well as the issuance and expiration dates for the license. A licensed practitioner was included in the count of practitioners in the 500-child radius of a school in any year in which their license was active. Practitioner counts are assigned to schools according to the same strategy as physicians. 2.1.6. Covariates used in supplementary models We construct two county-level measures to help unpack the effect of proximity to children with autism. First, awareness of vaccine-preventable diseases is expected to have the opposite effect of knowledge of autism. Immunization rates have been shown to increase after epidemics [37]. We obtained the number of pertussis cases in each county in California from 2000 to 2010 from the Immunization Branch of CDPH. The pertussis data do not cover the whole study period; we therefore incorporate pertussis incidence into a supplementary model as rates per 1000 population in the county in which the school was located. Secondly, it is possible that the saliency of autism is not about knowing a child with autism but its advocacy. To compare the effect of proximity to children with autism in the local community with autism advocacy efforts, we create county-level counts of autism advocacy organizations using data from the Internal Revenue Office’s records of tax-exempt organizations. Each school is assigned the count of active autism advocacy organizations within the county where it was located each year.18 We use a county-level measure instead of counting the number of autism advocacy organizations within the 500-child radius around a school. This is because the number of such organizations are relatively small,19 and we expect the awareness of these organizations to extend over a larger distance than knowledge of children with autism diagnoses in the local community. In addition to the two county-level measures, we use the enrollment information from the CDE to calculate white isolation indices [38] at the school district level.20 These indices measures the extent to which non-Hispanic white students within a school district are concentrated in certain schools. Examining the impact of grouplevel isolation helps determine whether the racial/ethnicity effect we observe operates at the individual or community level. Unfortunately, enrollment information by race/ethnicity is not available for private schools and they are not included in these models. Charter schools are considered public schools in California and are therefore included in this analysis.

17 We calculated each school/childcare’s distance to the 200 nearest physicians to make sure the counts included all physicians within the 500-child radius. As with the number of children with an autism diagnosis, we would have increased this number if necessary. 18 The Internal Revenue Office’s data includes dates active for each tax-exempt organization and we use these to determine which years the organization is included in the data. 19 There were 103 autism advocacy organizations active in one or more years included in this analysis. 20 Isolation index at the school board level is used for the eight school boards of the Los Angeles Unified School District due to its large size.

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

4

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

2.2. Analysis We conduct our analysis in three stages. First, we use a twolevel negative binomial regression model with random effects to yield the baseline estimates of covariate effects. The outcome is the annual PBE counts by school, with schools specified as the level-2 unit. All covariates are lagged 3 and 6 years for childcare centers and kindergartens, respectively, to account for the majority of vaccination decisions taking place before a child enters school.21 Expected count of PBEs (uij) is calculated as:

log uij




¼ log tij




þ a þ bxij þ czj þ tij ;

ble, however, that children in kindergarten have siblings enrolled in nearby childcare centers. We perform this same test with private schools as public school boundaries also do not structure their enrollment and it is less common to have one sibling go to a public school while the other attends a private school. 3. Results 3.1. Distributions of PBEs by school type The average PBE rates in California private schools and noncharter public schools included in our model in 2014 were 5.2% and 2.1%, respectively.23 The average PBE rate of charter schools was 7.5%, even higher than private schools. Yet, the distribution of PBEs across all types of schools is highly uneven. Fig. 1 plots the Lorenz curves—cumulative proportions of PBEs against the cumulative proportions of student population—by school type. It demonstrates that, within each school type, schools comprising 20% of students contribute as much as 70% of the PBEs.

where schools are indexed by i and year by j, log(tij) is the logged number of enrolled students (exposure), a is the global intercept, xij is a vector of lagged independent covariates measured by school and year with b estimated covariate effects, z is a set of indicator variables for year with c estimated effects, and tij is the error term, composed of mi  N(0,rm2) and eij  N(0,r2). In addition to the main model above, we estimate a three-level mixed effects Poisson model using PBE data from 2006 to 2014 to examine the effects of pertussis rates and counts of autism advocacy organizations, both measured at the county level. This model took the same general form as described above, except that county-level measures were included as covariates and random intercepts were included at both the school and county levels. We also estimate a three-level mixed effects Poisson model to examine the effect of non-Hispanic white isolation indices on PBE rates. In this model, random intercepts were included for both schools and school districts. The second stage, and the main focus of this study, uses Kulldorff Spatial Scan Statistic [39] to map clusters of high PBE rates among public schools, adjusting for the spatial distributions of the covariates. Maximum likelihood estimation identifies circular areas that encompass schools with an increased probability of PBEs compared to the state-wide PBE rate. Certainly, not all schools inside the clusters have high PBE rates. However, this method systematically identifies areas in which all children have a heightened risk of exposure to non-vaccinated children [40]. To test whether the spatial clusters detected are mainly driven by locations of charter schools, we also examined PBE clusters only among non-charter public schools. In the third stage, we perform an additional set of analyses to unpack possible explanations for the results obtained from the school-level and spatial cluster analyses. First, to uncouple a possible SES effect being picked up by the non-Hispanic white measure, we estimated school-level models for all school and childcare center types.22 Children in Head Start programs should represent a homogenous population with respect to SES as eligibility for Head Start generally depends on families living at or below the federal poverty line. Second, we test whether school-district factors may be responsible for the spatial clustering and help explain school-level effects. School district boundaries structure enrollment for non-charter public schools, but not childcare centers. School district boundaries then should not be responsible for clustering of PBEs in childcare centers. We test boundary effects by examining the effects of public school PBE clusters on nearby childcare centers. It is very possi-

We now turn to our main questions: are large PBE pockets spanning multiple schools the product of the spatial distributions of socio-demographic factors? Are local autism rates associated with the locations of these large PBE clusters? The shaded circular regions in Fig. 2A show where high-risk exemption areas are located without adjusting for any covariates: these areas have statistically significantly higher relative risk of PBEs than the non-shaded areas. These clusters of high PBE rates among public schools are not confined in certain areas but dis-

21 As our covariates are measured yearly between 1992 and 2010, we are able to create the lagged covariates by assigning covariate values to a current school year consistent with what was observed in the same school six years previously. For example, to create a 6 year lag, school PBE rates in 2010 were regressed on school covariates observed in 2004. Using different time lags, or no lag at all, does not substantially change the results (see Table A1 in Appendix for results using a 5-year lag; results from other models are available upon request). 22 School types include public (non-charter), private, and charter. Childcare types include daycares and Head Start centers.

23 CDPH reported an average PBE rate of 2.54% for all public schools and 5.33% for all private schools. Only kindergartens with 10 or more enrolled students are included in the CDPH immunization data. We also restricted our sample to schools that were able to be matched to the CDE School Listing file (96.9% of schools). These factors may account for the small differences in the PBE rates presented here compared to those reported by CDPH. 24 This finding may seem surprising as it appears to contradict traditional explanations linking lack of access to healthcare as a primary factor in undervaccination.

3.2. School-level correlates Table 1 shows the results of our two-level model examining PBEs among public school kindergartens between 1998 and 2014. Count of children with autism has the expected positive effect on PBEs at the school level. Effects of counts of chiropractors and acupuncturists are initially positive and statistically significant, but the effect of acupuncturists becomes insignificant once the other covariates are included in the model. PBE rates increase with physician density, which is consistent with a previous study [28].24 Maternal education and percentage non-Hispanic white population have the expected positive associations with PBEs. Property values are not associated with PBEs once the effects of other covariates are controlled for. We fit a supplementary model using only the PBE data from a more limited period (2006–2014) to evaluate the effect of proximity to autism advocacy organizations and pertussis incidence at the county level. Results of the three-level model show that the locations of such organizations have no statistically significant effect on PBEs (Table A2 in Appendix). The same model also shows that the association between pertussis outbreaks at the county level and PBEs is in the expected direction: an additional case of pertussis per 1000 among children aged 0–6 decreases PBE rates by 3.8%, net of other factors. However, the effect is only marginally significant statistically (p = 0.086). 3.3. PBE pockets

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

5

mainly driven by charter schools, we conducted the analysis again using only public non-charter schools (Fig. A2 in Appendix). Adjusted for the spatial distributions of all covariates, PBEs continue to be unevenly distributed across non-charter public schools, with some clusters in the same locations as when charter schools are included.

Fig. 1. Lorenz curves of PBE distributions in 2014 by School Type.

tributed all across the state, from the sparsely populated area in the north, to areas in the Central Valley, to coastal urban centers (e.g., San Francisco, Santa Clara, Santa Barbara, Los Angeles, and San Diego). Fig. 2B–H shows the spatial clustering of PBEs across schools after adjusting for each of the school-level covariates. If the spatial clustering of PBEs is completely driven by a covariate’s spatial distribution, the adjustment will eliminate the PBE clusters from the corresponding figure. 3.3.1. Autism Adjusting for spatial distribution of children with autism has almost no effect on the PBE pockets spanning multiple schools (Fig. 2B), even though proximity to children with autism has a statistically significant effect in the school-level model. The major spatial cluster of autism was located in Los Angeles [15]. In contrast, high PBE rates are found in other locations across California. 3.3.2. Proximity to health care practitioners Adjusting for the locations of chiropractors and acupuncturists has almost no effect on the large pockets of PBEs (Fig. 2C). Adjusting for counts of physicians similarly has negligible effect (Fig. 2D). 3.3.3. Socio-demographic factors Fig. 2E shows that controlling for maternal education has a surprisingly small impact on the multiple-school clusters despite its strong association with PBE rates in the school-level model. In contrast, Fig. 2F shows that while logged property values are not associated with PBEs at the school level, adjusting for their spatial distribution reduces the size of the clusters in the coastal urban areas. Controlling for logged population density also shrinks some of the clusters, especially in the north (Fig. 2G). Among the covariates, adjusting for the percentage of nonHispanic white has the most substantial impact in reducing both the number and sizes of PBE clusters (Fig. 2H). This strong percent white effect is consistently observed across years (Fig. A1 in Appendix) and therefore cannot be explained by idiosyncrasies of the 2014 data. Fig. 2I shows the effects of adjusting for all covariates. Multiple school clusters are still found (48% of remaining PBE clusters include more than one school),25 although they now cover much smaller areas. Among the multi-school clusters, 87.5% include at least one charter school. To test whether the remaining PBE clusters are 25 It is possible to have single-school clusters when one school has a statistically significantly higher risk of PBEs than the surrounding schools.

3.3.4. Exploring the percent white effect Why does percent non-Hispanic white have such a strong impact on reducing multiple school clusters while mother’s years of education, which has a substantial effect size in the schoollevel model, has almost no impact on reducing cluster sizes?26 In theory, any PBE-related spatial processes spanning multiple schools and correlated with percent non-Hispanic white could be responsible. The diffusion of vaccine skepticism over racially homophilous (i.e., similar) social networks is an example. We later discuss why there are reasons to expect that spatially embedded, dense social networks are particularly relevant to vaccine refusals. While our data do not allow us to directly test the social diffusion of vaccine skepticism, the following robustness checks attempt to address some of the alternative explanations. If the percent non-Hispanic white result in Fig. 2H is merely picking up unmeasured SES effects, we should expect it to have no effect on PBE rates among homogenous SES populations, like children enrolled in Head Start. However, Fig. 3A shows the opposite result— among all school types, percentage non-Hispanic white has the strongest effect on Head Start PBE rates. A 10% increase in percent non-Hispanic white in the area around a Head Start is associated with, on average, a 37% increase in PBEs, controlling for the effects of other covariates. So far, results suggest that residential segregation alone cannot explain the substantial percent white effect on PBE pockets. Yet, this does not mean that racial segregation does not matter. It just shows that segregation patterns alone do not explain why the effect of percent white spans larger areas than that of mother’s years of education. In fact, Table A4 in the Appendix confirms that school districts with greater isolation of non-Hispanic white children (i.e., a high isolation index) have higher PBE rates than those districts that are more racially mixed. The substantial effect of percent non-Hispanic white could also be picking up school district-specific effects. If this is the case, the influence on PBE rates should not cross school district boundaries. However, Fig. 3B shows child care centers near kindergarten PBE clusters are more likely to have higher PBE rates, controlling for the effects of socio-demographic variables. To ensure we were not just picking up the effect of kindergarteners having younger siblings enrolled in nearby childcare centers, we repeat the same analysis with private schools and still find PBE rates of private schools to be statistically significantly higher if they are close to public kindergarten PBE clusters (Fig. 3C). As private school enrollment is not defined by public school district boundaries, this suggest that the race/ethnicity effect is not simply a reflection of school district effects. 4. Discussion As of 2016, 47 U.S. states permitted religious and/or philosophical exemptions [41]. California began prohibiting PBEs in 2016. 26 This difference could be explained by residual segregation patterns. If families are more segregated by race/ethnicity than by education, it could explain why one covariate may have only localized effects while the other affects PBE clustering more broadly. To test this, we calculated Moran’s I, a spatial correlation statistic, at the Census block-group level for covariates. According to the Moran’s I test, California was similarly segregated by percent non-Hispanic white population and mother’s years of education (results available upon request). This indicates that this finding cannot be attributed to greater segregation by race/ethnicity alone.

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

6

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

Table 1 Correlates of PBEs among public school kindergartens, 1998–2014.a Year-adjustedb IRRc Children with Autism Acupuncturistsd Chiropractorsd Physicians Mother’s education Property values (ln) % whitee Pop. density (ln)

***

1.013 1.038*** 1.043*** 1.010*** 1.532*** 1.651*** 1.339*** 0.683***

95% C.I. 1.007 1.030 1.038 1.008 1.505 1.575 1.329 0.669

Fully-adjusted IRR 1.019 1.045 1.048 1.012 1.560 1.730 1.348 0.698

***

1.006 0.996 1.029*** 1.004*** 1.190*** 0.974 1.217*** 0.782***

95% C.I. 1.000 0.988 1.023 1.002 1.163 0.926 1.205 0.766

1.012 1.003 1.034 1.006 1.218 1.025 1.230 0.799

a Two-level random intercept negative binomial regression model. N = 89,550 school-years. Year indicator variables included in all models but not shown. All independent variables lagged 6 years. We estimated spatial lag models as robustness checks in light of potential spatial auto-correlation in the errors. Estimates and statistical significance of covariate effects do not differ substantially. We choose not include spatial lags in here because our aim is to map, rather than statistically control for, the unexplained spatial clustering of PBEs. b Adjusted for year indicator variables. c Incident Rate Ratio. d To ensure that proximity to alternative healthcare practitioners is not due to a spurious effect of proximity to business areas, we also construct measures of the densities of randomly selected licensed professionals, as well as the density of veterinarians, which should not be associated with beliefs about vaccination. Proximity to randomly selected licensed professionals did not have a statistically significant effect. Count of veterinarians has a significant, positive effect on PBEs, net of other variables (IRR = 1.02, p < 0.001). We thus cannot rule out that unmeasured neighborhood-level characteristics could also explain the positive associations between chiropractors and PBE rates. e Measured in 10% increments (deciles) for readability. *** p < 0.001.

Fig. 2. Unadjusted and adjusted spatial clusters of PBEs across public schools, 2014.

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

7

Fig. 3. Ruling out alternative explanations of the association between percent non-Hispanic white and PBE clusters across schools. (A) Effect of percent non-Hispanic white by school type; effects of distance to nearest public kindergarten in PBE cluster on PBE rates in (B) child care centers and (C) private schools. Note: Incidence Risk ratios (IRRs) estimated via negative binomial regression and adjusted for mother’s education, logged property values, logged population density, counts of physicians, chiropractors, acupuncturists, and children with autism, and yearly indicator variables. Public schools in PBE clusters only include those with statistically significantly higher relative risk of PBEs after adjusting for covariates. Schools (1998–2014): Public schools: N = 84,303; private schools: N = 27,378; charter schools: N = 5247; Childcare centers (2010–2014): Daycare centers: N = 36,850; Head Start centers: N = 6049.

However, disallowing PBEs may drive some parents to seek medical exemptions [42]. Banning PBEs could also concentrate unvaccinated children in independent study schools.27 It is often assumed that spatial clustering of behaviors is merely the result of population composition or uneven health care access; our findings suggest that it is not the case for PBEs. Moreover, the effects of sociodemographic factors are shown to have different geographical spans. The implication is that the spatial patterns of PBEs are likely the combined result of social processes operating at varying geographical scales. Interventions that go beyond the individual-level and address these social processes are needed to prevent exemption pockets. Specifically, the seemingly unexplained increase in autism diagnoses (but see [17]) has fueled popularity of the belief that vaccines can cause autism. Not surprisingly, parents who file PBEs were concerned about vaccine safety [3]. Nonetheless, this study shows that there is little overlap between local autism prevalence and the locations of the large pockets of PBEs. Instead, racial homogeneity, measured by percent non-Hispanic white, has a uniquely strong association with large PBE clusters while other socio-demographic factors show more local effects. We speculate that spatially embedded network interactions are behind this result. Social networks can substantially influence health decisions [43]. In particular, dense social networks, which are often racially homophilous, are crucial to the adoption of rare or risky behaviors [44]. Dense networks tend to be spatially clustered and thus can exacerbate the clustering of risky behaviors [45]. Such endogenous processes of social influence also explain why neighborhoods that are similar with regard to SES can have drastically different PBE rates. Anti-vaccine attitudes are notoriously difficult to change with individual-level interventions [46] but network interventions – interventions that leverage the influence of social network to affect behaviors [47] – may be able to change norms in communities and prevent exemption pockets (see [48] for an example of changing voting behavior using Facebook networks). Lastly, the charter school movement likely had an unintended consequence of further concentrating children with

27 See ‘‘California’s vaccine law: Opponents moving, home schooling to avoid controversial mandate” by Tracy Seipel published June 30, 2016 in Mercury News, and ‘‘Disneyland measles outbreak leaves many anti-vaccination parents unmoved” by Andrew Gumbel published January 25, 2015 in The Guardian.

vaccine exemptions. Future debates about school choice should also consider potential consequences for public health. This study is limited by the lack of individual-level data. Although robustness checks were used to evaluate some alternative explanations of our results on the effects of race/ethnicity, detailed network data are required to directly examine the diffusion of vaccine refusals and the role of race/ethnicity. While lacking individual-level details, the statewide data allow us to isolate the spatial correlates of vaccination exemptions at different geographical scales. Future research combining epidemiological and individual-level data can shed light on the specific mechanisms. Conflict of interest statement We have no conflict of interest to declare. Acknowledgement This research is funded by the UCLA Hellman Fellowship (2014– 2015). The authors benefited from facilities and resources provided by the California Center for Population Research at UCLA (CCPR), which receives core support (P2C-HD041022) from the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD). Appendix A. Supplementary material Supplementary data to this article can be found online at https://doi.org/10.1016/j.vaccine.2018.09.038. References [1] Dubé E et al. Vaccine hesitancy. Human Vacc Immunotherap 2013;9 (8):1763–73. [2] Omer SB et al. Nonmedical exemptions to school immunization requirements: secular trends and association of state policies with pertussis incidence. JAMA 2006;296(14):1757–63. [3] Gaudino JA, Robison S. Risk factors associated with parents claiming personalbelief exemptions to school immunization requirements: community and other influences on more skeptical parents in Oregon, 2006. Vaccine 2012;30 (6):1132–42. [4] Sadaf A, Richards JL, Glanz J, Salmon DA, Omer SB. A systematic review of interventions for reducing parental vaccine refusal and vaccine hesitancy. Vaccine 2013;31(40):4293–304.

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038

8

A. Gromis, K. Liu / Vaccine xxx (2018) xxx–xxx

[5] Constable C, Blank NR, Caplan AL. Rising rates of vaccine exemptions: problems with current policy and more promising remedies. Vaccine 2014;32(16):1793–7. [6] Hinman A. Eradication of vaccine-preventable diseases. Annu Rev Public Health 1999;20:211–29. [7] CDPH. 2014- 2015 Kindergarten immunization assessment results. California Department of Public Health Immunization Branch; 2015. [8] Omer SB et al. Geographic clustering of nonmedical exemptions to school immunization requirements and associations with geographic clustering of pertussis. Am J Epidemiol 2008;168(12):1389–96. [9] Lieu TA, Ray GT, Klein NP, Chung C, Kulldorff M. Geographic clusters in underimmunization and vaccine refusal. Pediatrics 2015. [10] Carrel M, Bitterman P. Personal belief exemptions to vaccination in California: A spatial analysis. Pediatrics 2015. [11] Eames KT. Networks of influence and infection: parental choices and childhood disease. J Roy Soc Interface/Roy Soc 2009;6(38):811–4. [12] Salathe M, Bonhoeffer S. The effect of opinion clustering on disease outbreaks. J Roy Soc, Interface/Roy Soc 2008;5(29):1505–8. [13] Richards JL et al. Nonmedical exemptions to immunization requirements in California: a 16-year longitudinal analysis of trends and associated community factors. Vaccine 2013;31(29):3009–13. [14] Brennan JM et al. Trends in personal belief exemption rates among alternative private schools: Waldorf, Montessori, and Holistic Kindergartens in California, 2000–2014. Am J Public Health 2017;107(1):108–12. [15] Mazumdar S, King M, Liu K-Y, Zerubavel N, Bearman P. The spatial structure of autism in California, 1993–2001. Health Place 2010;16(3):539–46. [16] Mazumdar S, Winter A, Liu KY, Bearman P. Spatial clusters of autism births and diagnoses point to contextual drivers of increased prevalence. Soc Sci Med 2013;95:87–96. [17] Liu K, King M, Bearman PS. Social influence and the autism epidemic. Am J Sociol 2010;115(5):1387–434. [18] Fountain C, Bearman PS. Risk as social context: immigration policy and autism in California. Sociol Forum 2011;26 (2 %L 0000). [19] Taylor LE, Swerdfeger AL, Eslick GD. Vaccines are not associated with autism: an evidence-based meta-analysis of case-control and cohort studies. Vaccine 2014;32(29):3623–9. [20] Freed GL, Clark SJ, Butchart AT, Singer DC, Davis MM. Parental vaccine safety concerns in 2009. Pediatrics 2010;125(4):654–9. [21] Stefanoff P et al. Tracking parental attitudes on vaccination across European countries: The Vaccine Safety, Attitudes, Training and Communication Project (VACSATC). Vaccine 2010;28(35):5731–7. [22] Weinstein ND. Effects of personal experience on self-protective behavior. Psychol Bull 1989;105(1):31–50. [23] Schwarzer R. Optimism, vulnerability, and self-beliefs as health-related cognitions: A systematic overview. Psychol Health 1994;9(3):161–80. [24] Safi H et al. Vaccine policy and Arkansas childhood immunization exemptions: a multi-year review. Am J Prev Med 2012;42(6):602–5. [25] Atwell JE et al. Nonmedical vaccine exemptions and pertussis in California, 2010. Pediatrics 2013;132(4):624–30. [26] Imdad A et al. Religious exemptions for immunization and risk of pertussis in New York State, 2000–2011. Pediatrics 2013;132(1):37–43.

[27] Birnbaum MS, Jacobs ET, Ralston-King J, Ernst KC. Correlates of high vaccination exemption rates among kindergartens. Vaccine 2013;31(5):750–6. [28] Walker ET, Rea CM. Pediatric care provider density and personal belief exemptions from vaccine requirements in california kindergartens. Am J Public Health 2016;106(7):1336–41. [29] Brown KF et al. Factors underlying parental decisions about combination childhood vaccinations including MMR: A systematic review. Vaccine 2010;28 (26):4235–48. [30] Shui IM, Weintraub ES, Gust DA. Parents concerned about vaccine safety: differences in race/ethnicity and attitudes. Am J Prev Med 2006;31(3):244–51. [31] Buttenheim AM et al. MMR vaccination status of children exempted from school-entry immunization mandates. Vaccine 2015;33(46):6250–6. [32] ESRI. Community Sourcebook – America with ArcReader (2002-2004 editions). ESRI: Redlands, CA; 2002–2004. [33] Liu K, Bearman PS. Focal points, endogenous processes, and exogenous shocks in the autism epidemic. Sociol Methods Res 2015;44(2):272–305. [34] Croen LA, Grether JK, Hoogstrate J, Selvin S. The changing prevalence of autism in California. J Autism Dev Disord 2002;32(3):207–15. [35] Salmon DA et al. Factors associated with refusal of childhood vaccines among parents of school-aged children: A case-control study. Arch Pediatr Adolesc Med 2005;159(5):470–6. [36] Campbell JB, Busse JW, Injeyan HS. Chiropractors and vaccination: A historical perspective. Pediatrics 2000;105(4):E43. [37] Bauch CT, Bhattacharyya S. Evolutionary game theory and social learning can determine how vaccine scares unfold. PLoS Comput Biol 2012;8(4):e1002452. [38] Massey DS, Denton NA. The dimensions of residential segregation. Soc Forces 1988;67(2):281–315. [39] Kulldorff M. Spatial Scan Statistics: models,calculations, and applications. In: Glaz J, Balakrishnan N, editors. Scan statistics and applications. Birkhäuser Boston: Statistics for Industry and Technology; 1999. p. 303–22. [40] Buttenheim A, Jones M, Baras Y. Exposure of California kindergartners to students with personal belief exemptions from mandated school entry vaccinations. Am J Public Health 2012;102(8):e59–67. [41] Yang TY, Silverman R. Legislative prescriptions for controlling nonmedical vaccine exemptions. J Am Med Assoc 2015;313(3):247–8. [42] Delamater PL, Leslie TF, Yang Y. Change in medical exemptions from immunization in california after elimination of personal belief exemptions. JAMA 2017;318(9):863–4. [43] Centola D. The spread of behavior in an online social network experiment. Science 2010;329(5996):1194–7. [44] Granovetter M. Threshold models of collective behavior. Am J Sociol 1978;83 (6):1420–43. [45] Centola D, Macy M. Complex contagions and the weakness of long ties. Am J Sociol 2007;113(3):702–34. [46] Horne Z, Powell D, Hummel JE, Holyoak KJ. Countering antivaccination attitudes. Proc Natl Acad Sci 2015;112(33):10321–4. [47] Valente TW. Network interventions. Science 2012;337(6090):49–53. [48] Bond RM et al. A 61-million-person experiment in social influence and political mobilization. Nature 2012;489(7415):295–8.

Please cite this article in press as: Gromis A, Liu K. The roles of neighborhood composition and autism prevalence on vaccination exemption pockets: A population-wide study. Vaccine (2018), https://doi.org/10.1016/j.vaccine.2018.09.038