Increasing vaccine hesitancy has presented a unique public health challenge over recent years. The most substantial evidence of this trend has been the hesitation—and, in some areas, resistance—associated with the COVID-19 vaccine. While the initial vaccine release was met with excitement by much of the American population, the rate at which first doses were administered nationally began to slow down by the late summer of 2021 and uptake of subsequent boosters has been even lower than that of the primary 2-dose series.1 1 2 , 3 4 , 5 6 7 , 8 9
In response to the pressing need to encourage vaccination, public health entities at the local, state, and national levels crafted numerous innovative outreach strategies. For example, Alabama ran a statewide competition on popular social media platform TikTok, asking contestants to share why they chose to get vaccinated.10 11
One technique to identify subgroups within a population is cluster analysis, a machine learning technique that classifies data points into groups based on shared features.12 13–16 16–18
In this article, we provide an overview and results of an approach that marries statistical and behavioral science methods to develop tailored outreach strategies that encourage COVID-19 primary dose vaccine uptake in the state of Missouri. We combined statewide individual- and population-level data to group Missouri's census tracts into manageable, distinct groups based on shared likeness across social determinants of health and behavioral data points. The resulting clusters, or “community segments,” provided a more nuanced view of typologies within Missouri for the state's public health department to better tailor outreach.
Our team of public health and behavioral science practitioners used the vaccination model originated by Brewer and colleagues19 19 individual factors (ie, perceived risk, trust, safety concerns), social factors (ie, social norms, information sharing, rumors), and practical issues (ie, convenience, cost). Using this model, we hypothesized primary barriers to vaccination—both motivational and practical—for each segment and developed corresponding intervention recommendations to address each segment's barriers and promote behavior change.
Methods
Data sources
We obtained data from 3 sources: ShowMeVax Immunization Information System from the state of Missouri20 21 22 *
Step 1: Variable preparation and selection
We pulled and aggregated variables from our data sources at the census tract level. Given the number of candidate variables for a model was quite large, variable or feature selection was used to remove variables that were either unreliable and weakly associated with a target variable or collinear with other variables. Variable selection for inclusion in cluster models was performed through a literature review of factors likely to influence vaccination and 2 statistical approaches: ordinary least squares (OLS) linear regression and geographically weighted regression (GWR). Regression techniques were chosen in lieu of other data science techniques because of their explanatory power and geographic sensitivity.
We then performed a series of regressions using vaccination as the dependent variable starting with baseline OLS models. Variables most strongly associated with vaccination (ie, significant at 60% level or with a negative or positive relationship of 90%) were retained in the GWR. Conducting the GWR enabled a more nuanced approach by identifying variables that were driving spatial variations within the 9 emergency management regions across Missouri and pinpointing variables that were spatially fixed or significantly associated with vaccination across all 9 regions. The GWR similarly used vaccination as the dependent variable at both the statewide and regional level, with resulting P values outputted for each model.
A bivariate correlation matrix was outputted for those variables with significant explanatory power to eliminate multicollinearity effects, reducing the variable pool to 14 variables. The resulting list was then appended with 4 variables that have been historically linked to vaccine uptake per the literature review (Table 1 ).
TABLE 1 -
Summary of Clustering Variables and Their Variable Selection Source(s)
Variable
Source
Vaccination status
Literature
Health behavior
Literature
Religion
Literature
Age
Literature
Religion
Literature
Uninsured rate
OLS
Disability
OLS
Census tract density
GWR
Veteran status
GWR
Households receiving transfer payments
GWR
Substandard housing
GWR
Housing: Mobile home
GWR, OLS
Foreign born
GWR, OLS
Census tract mean hours worked
GWR, OLS
Occupation
GWR, OLS
Percent Black and Latinx
GWR, OLS
Households with children
GWR, OLS
Education (bachelor's degree)
OLS, GWR
Income
OLS, GWR
Abbreviations: GWR, geographically weighted regression; OLS, ordinary least squares.
Step 2: Clustering of census tracts
A combination of dimensionality reduction and clustering models was performed at the census tract level across Missouri using the 18 selected variables as inputs. To test which models fit our data distribution best, 2 dimensionality reduction algorithms (UMAP, t-SNE) and 3 clustering models (GMM, k -means, hierarchical) were deployed. Dimensionality reduction is a common practice when analyzing data sets with very large numbers of variables (ie, dimensions) and involves transforming data from a high- to a low-dimensional space. Cluster analysis refers to a series of quantitative approaches to grouping objects (eg, census tracts) based on likeness or similarity. The specific techniques were chosen to empirically test a mix of established (eg, t-SNE) and emerging methods (eg, UMAP) within the data science field. These models produced 32 potential combinations of statistical clusters, with potential cluster totals ranging from k = 3 to k = 15.
The 32 potential combinations were then evaluated through a 2-step process. First, quantitative methods, including assessing silhouette scores23
Next, the remaining 16 combinations were reviewed to determine the combination that produced distinct clusters that both fit with existing COVID-19 patterns (eg, higher-income groups often being more vaccinated) and were manageable enough for a public health leader to feasibly act upon. A team of 6 individuals with backgrounds in public health, data science, and behavioral science reviewed crosstabs of the model combinations across more than 100 demographic and behavioral characteristics. For each combination, each cluster was compared with the rest of the clusters in the model output to determine whether it was distinct enough to benefit from unique public health outreach. Geographic dispersion was also reviewed for each potential combination to determine which clusters tracked most closely with local understanding of population patterns. The final dimensional reduction technique and cluster combination was selected because of its geographic diversity, the distinctiveness in clusters across demographic and behavioral variables, and an actionable number of resulting segments from a public health intervention standpoint (see Supplemental Digital Content Table S1, available at https://links.lww.com/JPHMP/B158
Step 3: Post–cluster development of community segments
To equip local public health officials with more action-oriented data on their local communities, each census tract cluster from the final model was developed into a “community segment.” First, segments were mapped geospatially to understand the location of each segment across the state. These insights provided context into each segment's environment, as vaccine attitudes, social influences, and practical barriers vary greatly by geography.
Next, a detailed profile was created for each segment using descriptive statistics. In line with the Brewer model, we selected variables that provided more insight into practical issues and motivational factors to describe each segment (see Supplemental Digital Content Table S2, available at https://links.lww.com/JPHMP/B159
Finally, customized messaging and outreach strategies were developed on the basis of each segment's practical barriers and motivational influences, leveraging best practices from the behavioral science and public health literature.24 , 25
Results
Ten community segments, UMAP k -means = 10, were selected on the basis of cluster analyses of 18 variables, and each was assigned a letter label A through J (Table 2 ). Geospatial patterns emerged when viewing the segments across the state. For instance, segments B, C, and G were concentrated in the most urban parts of the state while segments A and E were predominately rural. Areas with greater population density (eg, Kansas City, Saint Louis, and Springfield) also had a more diverse distribution of clusters than the less-populated, rural parts of the state (Figure ).
FIGURE: Geographic Distribution of the 10 Segments Across the State of MissouriThis figure is available in color online (
www.JPHMP.com ).
TABLE 2 -
Sample of Descriptive Statistics, Barriers, and Interventions Across the 10 Segments
Segment
% Missouri Population
# People
First-Dose Vaccination Rate
Example Barrier to Vaccination
Example Intervention or Messaging
A
12%
602 000
44%
Social norms against vaccination
Make it discreet : Appeal to residents concerned about social backlash by offering flexible and discrete vaccination options (eg, drive-through, at-home).
B
6%
315 000
74%
Low perceived COVID-19 risk
Emphasize positive norms to apply social pressure : Convey that getting vaccinated is widely accepted and practiced in their community (“Most people in your neighborhood have gotten vaccinated so far”).
C
4%
215 000
57%
Non–English-speaking
Help navigate the system : Utilize ambassadors who speak the language (eg, promotoras , who are lay health workers who work with Spanish-speaking communities) to help navigate the system via home visits, help signing up for appointments, and serving as navigators to connect to social support services).
D
16%
824 000
68%
Friction
Make it easy : Engage PCPs and pediatricians to provide vaccinations at routine appointments. Encourage clinics to craft appointment messages to patients ahead of routine checkups to instill ownership (eg, “A COVID-19 vaccine has been reserved for you at your upcoming appointment”).
E
13%
652 000
49%
Perceived loss of control/personal freedom
Provide a sense of control : Combat perceived loss of control by framing vaccination as a personal choice and giving individuals ample options surrounding vaccination, eg, providers can encourage patients to choose their vaccine type and use commitment devices such as “vaccinate later.”
F
9%
430 000
62%
Lack of trust in health care system
Rebuild trust : Recruit health care professionals of color to address historical mistrust head-on and discuss reasons to trust the vaccine (eg, improved racial representation in the clinical trials, their personal experience with getting vaccinated).
G
6%
310 000
53%
Lack of health insurance
Make it accessible : Include vaccine scheduling with clinical and social services (eg, upon hospital admission, at the emergency department, annual visits, food banks, unemployment centers, WIC locations).
H
14%
687 000
56%
Distrust of government and the pharmaceutical industry
Keep it local with trusted messengers : Collaborate with and train trusted community leaders (eg, veterans, active-duty members, pediatricians) to serve as ambassadors, having one-to-one conversations using motivational interviewinga
I
10%
492 000
52%
Health literacy
Educate with empathy : Create easy-to-understand materials that are tailored to the literacy level and culture of the community, avoiding overly technical language.
J
10%
528 000
62%
Inertia
Bundle it : Work with local businesses and schools to host joint flu and COVID-19 vaccine drives on-site.
Abbreviations: PCP, primary care physician; WIC, Special Supplemental Nutrition for Women, Infants, and Children.
In addition to the geographic patterns, trends emerged in associations between vaccine uptake and demographic, socioeconomic, and behavioral characteristics. Segments B, D, J, and F had the highest vaccination rates, ranging from 62% to 74%. Relative to the state average and other segments with lower vaccination rates, these clusters typically had higher educational attainment and income levels, slightly more diversity across racial groups, and low unemployment, with residents generally more likely to work in white-collar professions, suggesting residents in these communities may be experiencing fewer practical barriers to vaccination. However, certain metrics across demographics and social values illustrate their distinctness. For example, segments F and D had above-average first-dose vaccination rates (62% and 68%, respectively) but differed in their racial distribution and religiosity: 17% of segment F was Black compared with only 1% of segment D, and 30% of segment F highly prioritized religion, twice that of segment D (Table 3 ).
TABLE 3 -
Sample of Key Preferences, Values, and Access Factors Across the 10 Segments
a
Segment
Age 65+ y
Less Than High School Education
Black
Latinx
Born Outside US
Veteran
Household Income <$40 000
Disability
Uninsured
No Car Access
Non-Native English Speaker
Lack Broadband Access
High Religiosity
Personal Health
A
28%
56%
0%
1%
1%
11%
48%
20%
16%
5%
1%
25%
31%
7%
B
23%
28%
7%
3%
10%
6%
27%
12%
12%
8%
7%
11%
19%
5%
C
20%
55%
8%
8%
9%
8%
72%
18%
21%
13%
9%
22%
34%
5%
D
25%
19%
1%
2%
6%
7%
7%
9%
9%
3%
5%
6%
14%
4%
E
25%
48%
0%
1%
1%
10%
26%
14%
14%
3%
1%
18%
25%
6%
F
22%
38%
17%
3%
7%
7%
45%
14%
16%
9%
4%
14%
30%
5%
G
21%
49%
73%
2%
4%
7%
78%
18%
22%
21%
3%
26%
42%
5%
H
23%
40%
1%
2%
3%
10%
30%
14%
14%
5%
3%
13%
23%
5%
I
25%
55%
2%
2%
3%
9%
57%
20%
18%
8%
2%
23%
31%
6%
J
23%
30%
2%
2%
4%
9%
13%
11%
11%
3%
3%
8%
19%
5%
a For additional questions about data, contact the corresponding author.
Segments C, G, H, and I had average to slightly below-average vaccination rates, ranging from 52% to 57%, but experienced some of the largest increases in vaccine uptake in the 3 months preceding this analysis. These segments had a mix of income that skewed lower, had higher proportions of racial and ethnic minority populations, and were more likely to be uninsured and utilize community health clinics, hospitals, or emergency departments for health care. Unlike the first group of segments, residents in these segments appeared to face both practical and motivational barriers. For instance, 78% of segment G's population had an income below $40 000 per year and 21% lacked access to a car, signaling potential practical obstacles in accessing the vaccine. Also, 73% were Black and 22% were uninsured, potentially deterred by mistrust of the health care system and historical, institutional racism.8 26
The remaining segments—segments A and E—had the lowest overall vaccination rates (44% and 49%, respectively) and showed minimal recent uptake over the preceding 3 months. These segments' residents had relatively lower levels of education, were more likely to live in rural areas, and were more likely to support religious causes. Residents experienced several practical barriers, such as lack of reliable broadband access; however, their primary barriers to vaccination were hypothesized to be motivational, given trends of vaccine resistance observed in rural parts of the state and among populations with lower educational attainment. Still, these segments varied slightly across a few metrics, such as income: 48% of segment A's population had an income below $40 000 per year compared with 26% of segment E.
Recommended interventions to boost vaccine uptake were developed according to these practical and motivational obstacles paired with each segment (Table 2 ), in addition to outreach and messaging best practices in the behavioral science and public health literature. For instance, given segment C had several practical barriers including transportation and lack of linguistic inclusion, paired with their steady recent vaccine uptake, recommendations focused not only on making vaccination as accessible as possible for those interested in getting vaccinated but also on building trust for those still on the fence. Strategies included establishing local mobile clinics to mitigate transportation and/or mobility barriers and recruiting linguistic peer ambassadors to help navigate the system (eg, help signing up for appointments and navigating social support services).
On the other hand, given segment E's below-average vaccination rate, fewer apparent practical barriers, and a plateau of recent vaccine uptake, recommendations focused on fostering vaccine motivation through collaborating with and training trusted community leaders to serve as vaccine ambassadors and combatting perceived loss of control by providing ample choice and options around vaccination (eg, timing, type of vaccination). For segments with above-average vaccine uptake, such as segment B, strategies focused more on leveraging the power of social norms—highlighting that most of these communities were vaccinated. Additional detailed results can be found in the Community Segments Playbook available on the Missouri DHSS Web site.†
Discussion
Despite the significant potential impact, few public health efforts currently harness the latest data science and behavioral science tools that afford a nuanced understanding of and concrete approaches to addressing the health needs faced by different communities. In this applied research, we combined robust, statewide individual- and population-level data to group Missouri's census tracts into 10 manageable, distinct groups, based on shared likeness across an array of social determinants of health and behavioral data points. The resulting community segments provided a view into typologies within Missouri to better tailor outreach for the COVID-19 primary vaccine. For each segment, we hypothesized primary barriers to vaccination—both practical and motivational—and developed sets of recommended public health interventions to address these barriers and promote behavior change.
While corresponding outreach recommendations that accounted for statewide heterogeneity were developed for each segment, the more detailed understanding of the community types also offered public health leaders a means of identifying which areas may be the ripest for intervention. Behavioral science research suggests that lessening the intention-action gap can lead to positive health behavior change27 24 , 28–30
This approach has the potential to not only inform hyper-local outreach but also strengthen statewide collaboration between local public health leaders at the county and city levels. An immediate application of the current analysis involved using the distribution of community segments across the jurisdictions of each of Missouri's 114 local public health agencies (LPHAs) to serve as the organizing factor in creating strategy “working groups.” Public health leaders from LPHAs with similar community segments in their county were grouped together and met virtually to share lessons learned throughout the pandemic and brainstorm innovative ways to better engage their populations—and resulting community segments—in getting vaccinated. In a survey following working group implementation, 92% of local public health leaders said they had either already started, or planned to, enact a new outreach strategy based on something they had learned from the sessions. Programs by LPHAs included offering vaccine education and administration at a local bingo event and enhancing a mobile vaccine clinic's location strategy. Ultimately, these working groups provided a novel venue for public health practitioners to collaborate in data-driven ways and strategize how to boost vaccination rates in their communities. A timeline delineating steps of the behavioral segmentation analysis, from variable preparation to working group implementation, can be found in Supplemental Digital Content Table S3 (available at https://links.lww.com/JPHMP/B160
Overall, this research and its application demonstrate the potential impact of combining data science and behavioral science methods to provide local health leaders with actionable insights into the barriers to vaccination their residents may be facing. By identifying such barriers, pinpointing the geographies in which they were most prevalent, and leveraging behavioral science best practices, the analysis enabled a shift away from a one-size-fits-all approach to enacting new outreach strategies tailored to the specific needs of local communities.
Methodological Limitations
This work has several potential limitations. Given clustering was performed using data at the census tract level as opposed to the individual level, the characteristics of each segment are estimates of communities, not metrics representative of each resident in a census tract. Individuals in each community segment may vary from the average in any of these groups and for any of the sociodemographic and behavioral variables used. However, clustering at the census tract level allows for an actionable public health approach and enables agencies to prioritize specific communities within their service area.
In addition, our data sources introduced limitations due to the nature of the variables and the timing at which they were developed. Several of the variables from HealthPrism are based on predictive models. There was some overlap in variables used to build the predictive variables and the raw variables used for clustering (eg, race), but multicollinearity concerns were mitigated through correlation analysis during variable selection. Other HealthPrism variables are reliant on survey data and credit card expense data, which introduce potential selection and response biases. Finally, ACS variables were from 2019 and it is possible the dynamics of these metrics may have changed in the communities surveyed during the time since data capture. Such considerations notwithstanding, the benefit of clustering and community segmentation on “big data” is the ability to detect meaningful patterns among communities in the absence of complete data precision.
Finally, given the dynamic nature of the COVID-19 pandemic, vaccine attitudes and intentions may change over time. This analysis was conducted at a point in time; however, as the pandemic evolves and the nature of vaccine hesitancy shifts, the process should be reconducted: selecting variables that are associated with current vaccine attitudes, completing cluster analysis on recent data, and confirming the stability of the segmentation or adjusting as needed.
Implications for Policy & Practice
Enabling practitioners to understand subgroups within populations and develop tailored interventions to meet the needs of each group connects to precision public health, a concept coined by Martin Khoury, to develop evidence-based approaches that deliver “the right intervention to the right population at the right time.”31 (p1) ,32
Formation of working groups based on the distribution of community segments in each jurisdiction could also be applied to other priorities. The groups enabled collaboration with practitioners outside the state's regions—fostering idea sharing in novel settings. Grouping leaders together not based on geography, but based on typologies within jurisdictions, could generate more coordination and innovation across state and local public health. We recommend identifying local and state “champions” to shepherd the process forward and garner engagement.
References
1. Centers for Disease Control and Prevention. CDC COVID Data Tracker: vaccinations in the US.
https://covid.cdc.gov/covid-data-tracker/#vaccinations_vacc-people-onedose-pop-5yr . Published 2021. Accessed June 2, 2022.
2. Rossen LM, Branum AM, Ahmad FB, Sutton P, Anderson RN. Excess deaths associated with COVID-19, by age and race and ethnicity—United States, January 26-October 3, 2020. MMWR Morb Mortal Wkly Rep. 2020;69(42):1522–1527.
3. Hughes MM, Wang A, Grossman M, et al. County-level COVID-19 vaccination coverage and social vulnerability—United States, December 14, 2020-March 1, 2021. MMWR Morb Mortal Wkly Rep. 2021;70(12):431–436.
4. McElfish PA, Willis DE, Shah SK, Bryant-Moore K, Rojo MO, Selig JP. Sociodemographic determinants of COVID-19 vaccine hesitancy, fear of infection, and protection self-efficacy. J Prim Care Community Health. 2021;12:21501327211040746.
5. Choi KH, Denice PA, Ramaj S. Vaccine and COVID-19 trajectories. Socius Sociol Res Dyn World. 2021;7:1–11.
6. Leonhardt JM, Ridinger G, Rong Y, Talaei-Khoe A. Invincibility threatens vaccination intentions during a pandemic. PLoS One. 2021;16(10):e0258432.
7. Bogart LM, Dong L, Gandhi P, et al. COVID-19 vaccine intentions and mistrust in a national sample of Black Americans. J Natl Med Assoc. 2022;113(6):599–611.
8. Braveman PA, Arkin E, Proctor D, Kauh T, Holm N. Systemic and structural racism: definitions, examples, health damages, and approaches to dismantling. Health Aff (Millwood). 2022;41(2):171–178.
9. Bolsen T, Palm R. Politicization and COVID-19 vaccine resistance in the U.S. Prog Mol Biol Transl Sci. 2022;188(1):81–100.
10. Alabama Department of Public Health (ADPH). Vaccinate Alabama TikTok Contest.
https://www.alabamapublichealth.gov/covid19vaccine/tiktok.html . Accessed April 18, 2022.
11. Ramos AK, Duysen E, Carvajal-Suarez M, Trinidad N. Virtual outreach: using social media to reach Spanish-speaking agricultural workers during the COVID-19 pandemic. J Agromed. 2020;25(4):353–356.
12. Han J, Kamber M, Pei J. Data Mining: Concepts and Techniques. Waltham, MA: Morgan Kaufmann Publishers; 2012.
13. Watts P, Buck D, Netuveli G, Renton A. Clustering of lifestyle risk behaviours among residents of forty deprived neighbourhoods in London: lessons for targeting public health interventions. J Public Health (Oxf). 2016;38(2):308–315.
14. Busch V, Van Stel HF, Schrijvers AJ, De Leeuw JR. Clustering of health-related behaviors, health outcomes and demographics in Dutch adolescents: a cross-sectional study. BMC Public Health. 2013;13:1118.
15. Conry MC, Morgan K, Curry P, et al. The clustering of health behaviours in Ireland and their relationship with mental health, self-rated health and quality of life. BMC Public Health. 2011;11:692.
16. Li Y, Jasani F, Su D, et al. Decoding nonadherence to hypertensive medication in New York City: a population segmentation approach. J Prim Care Community Health. 2019;10:2150132719829311.
17. Taylor CR, Alexander GR, Hepworth JT. Clustering of U.S. women receiving no prenatal care: differences in pregnancy outcomes and implications for targeting interventions. Matern Child Health J. 2005;9(2):125–133.
18. Grant RW, McCloskey J, Hatfield M, et al. Use of latent class analysis and
k -means clustering to identify complex patient profiles. JAMA Netw Open. 2020;3(12):e2029068.
19. Brewer NT, Chapman GB, Gibbons FX, Gerrard M, McCaul KD, Weinstein ND. Meta-analysis of the relationship between risk perception and health behavior: the example of vaccination. Health Psychol. 2007;26(2):136–145.
20. Missouri Department of Health & Senior Services. What is ShowMeVax?
https://health.mo.gov/living/wellness/immunizations/showmevax/index.php . Accessed March 1, 2022.
21. US Census Bureau. Selected housing characteristics, 2019 American Community Survey 1-year estimates data profiles.
https://data.census.gov/ . Published 2019. Accessed April 18, 2022.
23. Rousseeuw PJ. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J Comput Appl Math. 1987;20:53–65.
24. Brewer NT, Chapman GB, Rothman AJ, Leask J, Kempe A. Increasing vaccination: putting psychological science into action. Psychol Sci Public Interest. 2017;18(3):149–207.
25. Barsade S, Chatman J, Duckworth A, et al. COVID-19 Vaccination Uptake Behavioral Science Task Force: Final Report—February 23, 2021. Philadelphia, PA: University of Pennsylvania; 2021.
26. Njoku A, Joseph M, Felix R. Changing the narrative: structural barriers and racial and ethnic inequities in COVID-19 vaccination. Int J Environ Res Public Health. 2021;18(18):9904.
28. The Behavioural Insights Team. Practicalities are the most significant impediments to people getting a COVID vaccine—and the easiest to address.
https://www.bi.team/blogs/practicalities-are-the-most-significant-impediments-to-people-getting-a-covid-vaccine-and-the-easiest-to-address . Accessed April 18, 2022.
29. Huang L, Yu R. How to (actually) change someone's mind. Harv Bus Rev.
https://hbr.org/2020/07/how-to-actually-change-someones-mind . Published 2020. Accessed May 5, 2022.
30. Milkman K. How to Change: The Science of Getting From Where You Are to Where You Want to Be. New York, NY: Penguin; 2021.
31. Khoury MJ, Iademarco MF, Riley WT. Precision public health for the era of precision medicine. Am J Prev Med. 2016;50(3):398–401.
32. Arnold C. Is precision public health the future—or a contradiction? Nature. 2022;601(7891):18–20.
