Sociodemographic Predictors of Vaccination Exemptions on the Basis of Personal Belief in California

Abstract

Objectives. We examined the variability in the percentage of students with personal belief exemptions (PBEs) from mandatory vaccinations in California schools and communities according to income, education, race, and school characteristics.

Methods. We used spatial lag models to analyze 2007–2013 PBE data from the California Department of Public Health. The analyses included school- and regional-level models, and separately examined the percentage of students with exemptions in 2013 and the change in percentages over time.

Results. The percentage of students with PBEs doubled from 2007 to 2013, from 1.54% to 3.06%. Across all models, higher median household income and higher percentage of White race in the population, but not educational attainment, significantly predicted higher percentages of students with PBEs in 2013. Higher income, White population, and private school type significantly predicted greater increases in exemptions from 2007 to 2013, whereas higher educational attainment was associated with smaller increases.

Conclusions. Personal belief exemptions are more common in areas with a higher percentage of White race and higher income.

Vaccines are credited with virtually eliminating many diseases in the United States, improving life expectancy, and lowering health care costs. Reduced financial barriers to obtaining vaccinations, strong support from health care professionals, and laws requiring specific vaccinations for school entrance have contributed to improving and sustaining high rates of vaccine uptake in the United States.¹ However, the tremendous success of immunizations in controlling vaccine-preventable diseases has resulted in diminishing fear of these illnesses among parents. Coupled with an increase in fear of vaccines themselves, this has led to falling vaccination rates in many areas of the country,^2–4 as increasing numbers of parents seek and obtain waivers from vaccination requirements.^5,6

The outbreak of measles in several states in 2014–2015⁷ has shone a bright light on the potential consequences of offering nonmedical vaccination exemptions.⁸ Nonmedical exemptions allow individuals to opt out of school-entry and other requirements for vaccinations on the basis that they hold a sincere philosophical or religious belief opposing vaccination.⁹ Previous research has demonstrated that states with nonmedical exemptions have higher incidence of pertussis⁹ and that children with a nonmedical exemption are at increased risk of acquiring and transmitting measles and pertussis.^10,11

Use of nonmedical exemptions has increased over time, especially in states that make them easy for parents to obtain.¹² In California, the percentage of schoolchildren with a nonmedical or “personal belief” exemption (PBE; defined as including both philosophical and religious objections), climbed from 0.77% in the 2000–2001 school year to nearly 3.15% by 2013–2014¹³—well above the median rate of 1.8% across all states.¹⁴ Media commentators on this phenomenon have suggested that the number of PBEs in California had doubled since 2007 and pointed to wealthy, highly educated parents as the primary drivers.^15,16 Although some evidence suggests that higher exemption rates are associated with higher population proportions of Whites, college graduates, and higher-income households,¹⁷ few rigorous, multiyear studies have investigated data beyond 2007, hindering efforts to understand and respond to recent rises in PBEs.

We investigated the sociodemographic predictors of variation in PBE rates across schools and communities in California from 2007 to 2013. We also explored whether schools are becoming more or less similar over time in their percentages of students with PBEs after we controlled for socioeconomic and geographic factors.

METHODS

We obtained publicly available, annual PBE data for kindergarteners in the 2007–2008 through 2013–2014 school years for California schools with more than 10 enrolled kindergarteners from the California Department of Public Health. The PBEs for each school represent the number of kindergarteners who obtained a philosophical or religious exemption for any of the required vaccinations. Information for each school also included the address, whether it was public or private, and the number of children enrolled in kindergarten. We calculated the yearly percentage (proportion × 100) of kindergarteners with PBEs for each year, from 2007 to 2013. We estimated average annual percentage change in the PBE from 2007 to 2013 by using linear regression. We geocoded schools by using their address information.

Our analytical sample consisted of 6911 schools for PBE percentage in 2013. From the California Department of Public Health’s data set, we excluded 400 schools that did not actually report PBEs or enrollment in 2013, 57 that could not be geocoded, and 13 that were missing covariate data. For annual percentage change, the analytical sample contained 6207 schools, as we removed an additional 704 schools because of an incomplete record of PBE data from 2007 to 2013.

We ran spatial lag models by using 4 different dependent variables: school-level PBE percentage in 2013, region-level PBE percentage in 2013, average annual percentage change in school-level PBE from 2007 to 2013, and average annual percentage change in region-level PBE over that period. We log-transformed the 2013 PBE variable because of nonnormality and the large number of schools with a PBE percentage of zero. For the region-level models, we constructed and separately tested the dependent variables at 3 different geographic levels: US Census block group, US Census tract, and public school district.

For the 2013 models, explanatory variables included school characteristics and sociodemographic characteristics (Table 1). We used logged population density to model rurality and we used the ratio of per-capita income to median earnings as a measure of income inequality. The change-over-time models also included the percentage of PBEs in 2007 as a covariate to both model change as a function of initial PBE levels and to detect whether temporal convergence across schools or regions occurred between 2007 and 2013.¹⁸

TABLE 1—

Characteristics of Schools and US Census Block Groups: California, 2013

Characteristics	No. (%) or Mean ±SD
School (n = 6 911 schools)
% kindergarteners with vaccination personal belief exemptions
2007a	1.54 ±4.421
2013	3.062 ±6.331
Average annual % change, 2007–2013a	0.193 ±0.189
No. of enrolled kindergarteners, 2013	76.39 ±43.76
Ownership type, 2013
Public	5 569 (80.58)
Private	1 342 (19.42)
Regional (n = 6 190 block groups)b
Median household income, $, 2008–2012	70 521 ±34 579
Per capita income-to-earnings ratio, 2008–2012	0.806 ±0.278
% White population, 2010	39.46 ±26.08
% of population with college degree, 2008–2012	28.85 ±21.07
Population density, 2010 (people per km2)	2 079 ±2 452

^an = 6207.

^bRegional characteristics are the means and SDs of the block groups containing at least 1 school.

We drew data on region-level median household income and educational attainment from the 2008–2012 American Community Survey 5-year estimate, and we obtained race/ethnicity data from the 2010 US Census. To account for uncertainty about the area from which each school’s students were drawn, we tested the covariates at 3 different geographic levels (block group, tract, and school district) in separate models. We overlaid regional boundary data with school locations to assign a block group, tract, and school district to each school (details provided in Appendix Figure A, available as a supplement to the online version of this article at http://www.ajph.org). In the school-level models, we directly assigned the region-level community attributes to the individual schools. For the region-level models, we aggregated the attributes for all schools located within the same region to create a single observation representing all of the schools in the region (a weighted mean, weighted by enrollment size).

We observed strong correlations (Pearson coefficient ≥ 0.5) among the explanatory variables median household income, White population percentage, and percentage of the population with a college degree. To address the problem of multicollinearity that would result from including all of these important predictor variables in the model, we created new variables representing the prevalence of White race and college degrees via bivariate regressions with income.¹⁹ In this approach, the residuals from the bivariate regressions represent the variation in the White and college degree variables that is unexplained by the correlation with income. These residuals are good proxies for the original variables—the correlations with the original variables are high (for the 3 geographic levels, Pearson coefficients ranged from 0.89–0.98 for the White variable and 0.64–0.72 for the college degree variable). By making these residuals the explanatory variables in the regression models, race and education effects can be estimated coherently and without multicollinearity.

We employed spatial lag models because diagnostics of ordinary least squares models indicated significant spatial autocorrelation within the regression residuals (Moran I, P < .05) and a lag relationship.²⁰ The spatial lag model controls for spatial autocorrelation among observations by modeling the dependent variable as a function of both the explanatory variables and the dependent variable values of neighboring observations. For the region-level analysis, we defined neighbors (in binary terms) as any regions sharing a boundary. For the school-level analysis, we defined neighbors as schools within 100 kilometers to capture regional similarities among schools. We employed an inverse distance squared weighting scheme to assign the strength of the neighbor relationships among schools (i.e., we assigned nearby locations a stronger relationship than distant locations).

The residuals from the spatial lag models showed statistically significant heteroskedasticity; we adjusted standard errors for the parameter estimates accordingly.^21,22 The maximum observed variance inflation factor for any variable (over all models) was 2.027, indicating that the parameter estimates were not biased because of multicollinearity.¹⁹

We performed analyses with R version 3.1.1 software (R Foundation, Vienna, Austria) and ArcGIS 10.2 software (ESRI Inc, Redlands, CA).

RESULTS

The overall PBE percentage for students entering California kindergartens doubled from 2007 to 2013, from 1.54% to 3.06% (Table 1). The PBEs were nearly twice as common in private schools (5.43%) as in public schools (2.88%) in 2013.

Figure 1 illustrates the overall, state-level relationships between PBEs and key variables from the analysis. In general, 2013 PBE percentages were higher in regions with higher income, education, and White population. For average annual percentage change in the PBE (Appendix Figure B, available as a supplement to the online version of this article at http://www.ajph.org), similar relationships are observed but appear less pronounced.

FIGURE 1—

Boxplots of Mandatory Vaccination Personal Belief Exemption Percentage Stratified by (a) Median Household Income, (b) Population With a College Degree, (c) Race/Ethnicity, and (d) Kindergarten Enrollment: California, 2013

Note. PBE = personal belief exemption. The x-axis variables are stratified by quantiles. Boxplots show results for block groups: the mean value (asterisks), median value (heavy line), upper and lower quartiles (top and bottom extent of the box), and range (the whiskers, bound by the true minimum and maximum or 1.5 times the length of the interquartile range).

Figure 2 maps 2013 PBE percentage by using hexagon-based units to show regional patterns rather than focusing on individual schools. Regions with high PBE percentages are found in far-Northern California, east of the Central Valley in Northern California, parts of the Bay Area, areas around Santa Cruz and Santa Barbara, and areas in Los Angeles, Orange, and San Diego counties. The PBE percentages generally appear to be low in the central point of large urban areas, increasing in the suburban and exurban areas surrounding them, and in the Central Valley west and south of Fresno. For average annual percentage change between 2007 and 2013, the regions with the highest increases (≥ 6% from 2007 to 2013) are distributed fairly evenly throughout California, whereas reduced exemption rates over time appear to be primarily found in rural regions.

FIGURE 2—

Mandatory Vaccination Personal Belief Exemption Percentages: California, 2013

Note. School-level data have been generalized to hexagon spatial units for display purposes. Hexagons with no schools located within their boundaries are not mapped.

Standardized parameter estimates for the block group models are presented in Table 2 (tract and school district model results are provided in Appendix Tables A and B, available as a supplement to the online version of this article at http://www.ajph.org). The estimates can be interpreted as the change in the outcome variable (in SDs) for a 1-SD increase in the explanatory variable, which permits comparison of the magnitude of effects across variables. Median household income and percentage White population were the strongest predictors of 2013 PBE percentages and change in PBE percentages over 2007 to 2013. Both variables had a positive, significant effect (P < .001) in nearly all model specifications. To provide an illustrative interpretation of the magnitude of these effects, we focus on the results of the school-level block group model.

TABLE 2—

Association Between School and Regional Characteristics and 2013 Mandatory Vaccination Personal Belief Exemption Percentages and 2007–2013 Personal Belief Exemption Percentage Change: California

	PBE % 2013, School Level		PBE % 2013, Region Level		PBE % Change 2007–2013, School Level		PBE % Change 2007–2013, Region Level
Characteristic	b (95% CI)	P	b (95% CI)	P	b (95% CI)	P	b (95% CI)	P
% PBE, 2007					−0.167 (−0.254, −0.079)	< .001	−0.226 (−0.314, −0.137)	< .001
Kindergarten enrollment	−0.012 (−0.038, 0.013)	.349	0.043 (0.018, 0.069)	< .001	−0.005 (−0.035, 0.025)	.748	0.007 (−0.022, 0.036)	.633
Private school	−0.007 (−0.037, 0.023)	.637	0.018 (−0.012, 0.048)	.238	0.100 (0.061, 0.139)	< .001	0.089 (0.051, 0.127)	< .001
Income	0.175 (0.155, 0.196)	< .001	0.199 (0.177, 0.221)	< .001	0.061 (0.034, 0.087)	< .001	0.074 (0.044, 0.105)	< .001
Income inequality	−0.005 (−0.032, 0.021)	.691	−0.002 (−0.032, 0.027)	.877	−0.012 (−0.045, 0.020)	.446	−0.011 (−0.045, 0.024)	.547
White population	0.363 (0.338, 0.389)	< .001	0.458 (0.431, 0.485)	< .001	0.214 (0.183, 0.245)	< .001	0.268 (0.232, 0.304)	< .001
Education	−0.001 (−0.026, 0.024)	.953	0.013 (−0.014, 0.039)	.347	−0.048 (−0.077, −0.019)	.001	−0.044 (−0.078, −0.011)	.01
Population density	−0.015 (−0.043, 0.013)	.308	−0.023 (−0.053, 0.007)	.126	−0.034 (−0.069, 0.002)	.062	0.001 (−0.021, 0.024)	.909
Spatial lag (ρ)	0.265 (0.227, 0.303)	< .001	0.072 (0.050, 0.094)	< .001	0.125 (0.079, 0.171)	< .001	0.051 (0.025, 0.076)	< .001

Note. CI = confidence interval; PBE = personal belief exemption. Table shows results from the block group spatial lag models.

Because of the log transformation of the dependent variable, 2013 PBE percentages, the marginal effects of income and White race were not consistent over the range of those variables; detailed marginal effects are presented in Appendix Table C (available as a supplement to the online version of this article at http://www.ajph.org). The marginal effect of income ranged from a 0.155% increase in the 2013 PBE percentage for an increase from $25 000 to $50 000 to a nearly 0.259% increase as income increased from $125 000 to $150 000 (1 PBE per 386 students). The effect of moving from the 10th to the 90th percentile of observations on income (from $25 000 to $135 000) was 1 additional PBE per 116 students in 2013. A similar move for race (from 5% to 85% White) was associated with a 2.66% increase in PBE, or 1 additional PBE per 38 students. An increase in percentage White population from 20% to 35% yielded 1 additional PBE per 280 students in 2013, and an increase from 65% to 80% yielded 1 additional PBE per 139 students.

In the models predicting change over time, a $25 000 increase in median household income resulted in a 0.032% annual increase (1 additional PBE per 3125 students per year). A 15% increase in the percentage White population was associated with a 0.099% annual increase in the PBE percentage, or 1 additional PBE per 1010 students per year.

Educational attainment did not independently predict 2013 PBEs. More educated populations had slower rates of change in PBE percentages from 2007 to 2013 (P ≤ .01). For example, in the school-level block group model, a 10% increase in the percentage of the population with a college degree was associated with a 0.025% decrease in the annual rate of growth from 2007 to 2013.

Private school type had a positive relationship with PBE rate of change (P ≤ .001), but not absolute 2013 percentages. We observed strong evidence of convergence in PBE percentages over time: having a low percentage in 2007 was associated with a greater rate of change from 2007 to 2013. Notably, of the 3767 schools with zero PBEs in 2007, only 55% remained at zero PBEs in 2013. The effects of the other explanatory variables were generally small and inconsistent over both models (PBE percentage and change over time).

DISCUSSION

We found that areas of California with higher household income and proportion White population are associated with higher overall PBE percentages as well as greater increases in PBEs from 2007 to 2013. In contrast to some previous studies,¹⁷ we did not find an independent predictive effect of educational attainment level once we controlled for those characteristics. Although the marginal effects of income and race were modest in magnitude, the overall PBE percentage doubled from 2007 to 2013, and more than 17 000 PBEs were issued in California in 2013.

At that level, if all or most of the exemptors indeed are not vaccinated, herd immunity could be lost in some areas where exemptors cluster.^23,24 Although California’s overall vaccination rates are stable, some suburban pockets of the state have rates hovering near 50%.¹⁴ More than 25% of schools in California have measles immunization rates for kindergarteners below the 92% to 94% recommended to maintain herd immunity.¹³

Qualitative research has closely linked PBEs with the concept of socioeconomic “privilege,”²⁵ and quantitative analyses have examined characteristics of children who did not receive vaccines (because of a PBE or other reasons).^26–28 A small number of studies have sought to quantify associations between PBEs and sociodemographic characteristics.¹⁷ Omer et al. examined spatial clustering of nonmedical exemptions in Michigan from 1991 to 2004, using Census tract as the unit of analysis.²⁹ Their multivariate analysis found that Census tracts with higher population density and larger average family size were more likely to be in a high-exemption cluster, but no other sociodemographic characteristics achieved statistical significance.

Carrel and Bitterman used 2001–2013 school-level PBE data to examine clusters of PBEs for kindergartners in California.³⁰ They found many such clusters in wealthy coastal areas with relatively large percentages of White students and a higher proportion of private or charter schools. They used spatial regression methods, as we did, but split their data and analysis into public and private schools. Because data were available for private schools only for some of their model’s explanatory variables, there was no reliable way to determine race and other aspects of the sociodemographic composition of these schools. Furthermore, they did not include neighborhood characteristics, except for urbanicity of the school’s Census tract.

Finally, Richards et al. used California PBE data to examine trends in PBEs over 1994 through 2009.³¹ Their multivariate analysis determined that schools with higher PBE percentages were located in Census tracts that were rural, lower income, had a higher proportion of White population, and had a higher proportion of the population college educated. The PBEs were also more common in private schools than public schools.

Our study extends and advances previous work in several ways. These innovations may account for the different findings we observed. First, the study by Richards et al. used generalized estimating equations, whereas we employed spatial lag models to account for the spatial autocorrelation we observed in the PBE data. Second, our estimation method did not model marginal effects as constant across the entire range of covariate values. Economic theory suggests that the effect of a $10 000 rise in income, for example, is not the same at the poverty level and 5 times above it. Third, to address multicollinearity, Richards et al. reportedly removed covariates that contributed to it, yet retained 3 variables—household income, education, and White race—that we found to be highly correlated in this data set. We detangled the correlation among these variables, which enabled us to examine their independent effects without eliminating any of these crucial constructs or sacrificing the validity of estimates. Finally, we tested specifications employing block group and school district as well as tract. This check provides reassurance of the robustness of our findings.

Limitations

Our study has limitations. Our ecological analysis provides a valuable perspective on the association between community factors and PBEs, but cannot provide information regarding some individual- and school-level factors that may be associated with the decision to obtain a PBE. For example, some previous work suggests that school policies and attitudes of school officials may affect exemption rates.^32,33

Another limitation is the imperfect information regarding the service area for each school, especially private schools, whose students are not drawn from defined school districts. As noted previously, to address this uncertainty, we conducted our analysis at both the school and region level and examined multiple definitions of the community associated with each (block groups, tracts, and school districts). In light of the consistency of the relationships observed across model specifications, we find no evidence that this issue affected our results.

A third limitation is the exclusion of some schools from our analytical sample. The California Department of Public Health excludes schools with fewer than 10 kindergarteners from its public use data file, and we excluded a small percentage (5.4%) of other schools from our analysis because of missing data. Whether these exclusions biased our findings cannot be estimated.

Another potential source of bias is the temporal mismatch between the PBE data (2007–2013) and the population covariate data (2010, 2008–2012). This limitation was unavoidable also because of data availability. However, major shifts in the socioeconomic and demographic characteristics of the populations residing in geographic units as large as block groups, tracts, and school districts generally do not occur over short time periods.

Conclusions

Our findings underscore the potential value of a targeted, tailored approach to messaging about the importance and safety of vaccines. Recommendations to target interventions to communities with high PBE rates^27,29,31 are sensible in light of the wide variation in these rates among communities within states and the differing sociodemographic makeup of high- and low- PBE communities. Focusing on selected communities may conserve scarce public health resources, and tailoring efforts may ensure that the recipients perceive them as “locally relevant.” Tailoring could consist both of adjusting the nature and tone of the messaging and identifying influential stakeholders in each community.

Some have reasoned from findings that high-PBE communities are better educated that public health strategies should focus on disseminating more scientific data on vaccine safety and the consequences of vaccine-preventable illnesses.²⁷ Our results call into question the reported link between high-PBE communities and higher average educational attainment, and other research also points to the need for messages that extend beyond providing vaccine safety data. For example, although there is little doubt that misperceptions of vaccine risks drive vaccine refusals, also important may be beliefs among upper-income, White parents that protective parenting techniques are effective substitutes for immunizations.²⁶

Our study also has implications for ongoing state efforts to eliminate PBEs or stiffen procedural requirements for obtaining them. Following the multistate measles outbreak of 2014–2015 that began in Disneyland, California legislators passed legislation repealing PBEs in the state.³⁴ Other states may follow California’s lead.^35,36 Alternatively, they may continue to impose new procedural hurdles to obtaining a PBE—a strategy with demonstrated efficacy in reducing PBE rates.¹² Although such laws arguably burden hourly workers and other disadvantaged persons who cannot easily attend physician counseling appointments or navigate other procedural requirements, our study and others suggest that such persons may be relatively uncommon among PBE seekers.

The attitudes and beliefs that lead parents to seek PBEs are now fairly well understood. The scientific case controverting them is persuasive. The task of bringing this evidence to refusers in an effective manner, however, is difficult and complex. Our study suggests that approaches tailored to sociodemographic factors may do the most good.