Tracking COVID-19 Vaccination Expectancies and Vaccination Refusal in the United States

Abstract

To identify factors that predict COVID-19 vaccination refusal and show how expectancies affect vaccination acceptance for non-vaccinated adults, we used a monthly repeated cross-sectional sample from June/2021 to October/2021 to collect data on vaccination behaviors and predictor variables for 2,116 US adults over 50 years of age. Selection bias modeling – which is required when data availability is a result of behavioral choice – predicts two outcomes: (1) no vaccination vs. vaccination for the entire sample, and (2) the effects of expectancy indices predicting vaccination Refuser vs. vaccination Accepters for the unvaccinated group. Vaccine refusers were younger, less educated, endorsed common misconceptions about the COVID-19 epidemic, exhibited higher psychological reactance, and were Black. Vaccination expectancies were related to vaccination refusal in the unvaccinated eligible group: negative expectancies increased vaccine refusal while positive expectancies decreased it. We conclude that behavior-related expectancies (as opposed to more stable psychological traits) are important to identify because they are often modifiable and provide a point of intervention, not just for COVID-19 vaccination acceptance but for other positive health behaviors.

Both the SARS-CoV-2 virus and resistance to vaccination are a worldwide phenomenon (Bults, Beaujean, Richardus, & Voeten, 2015; Harrison & Wu, 2020; Lazarus et al., 2021; MacDonald, Butler, & Dubé, 2018; Robinson, Jones, Lesser, & Daly, 2021; Sallam, 2021). During the pre-COVID-19 vaccine period, about 39% of Americans reported they would not get the vaccine if it were available (Tyson, Johnson, & Funk, 2020), a refusal rate similar to earlier pandemics (e.g., Kok et al., 2010; Mesch & Schwirian, 2014; Quinn et al., 2013). Pre-vaccine research naturally focused on the likelihood of/intention toward vaccination using demographic predictors (Malik, McFadden, Elharake, & Omer, 2020; Robinson et al., 2021; Szilagyi et al., 2021) although additional explanatory variables like vaccine features, anticipated preventive effectiveness, political ideology, and belief in conspiracy fantasies were sometimes added (Čavojová, Šrol, & Ballová Mikušková, 2020; Earnshaw et al., 2020; Kreps et al., 2020; Lin, Tu, & Beitsch, 2021; Ruiz & Bell, 2021; Stroope, Kroeger, Williams, & Baker, 2021; Y. Wang & Liu, 2022; Woko, Siegel, & Hornik, 2020).

However, as vaccines became available, a focus on vaccination behavior is preferable even though the intention-behavior correlation is substantial for many health outcomes (McEachan, Conner, Taylor, & Lawton, 2011; McEachan et al., 2016). This study uses expectancy value theory augmented with demographics and identity-related variables to predict vaccine behavior during the 6-month period after vaccines were available in the US. We first review the expectancy value concept and summarize the role it plays in published reasoned action studies (Fishbein & Ajzen, 2010) that predicted vaccination intention. Then we discuss the analysis issues that arise when vaccination acceptance (not intention) is the behavioral outcome and why selection bias modeling is the appropriate analysis methodology. We discuss the implications of our results from a perspective of designing effective health promoting messaging using expectancy value logic as operationalized in theories of behavior change.

The Expectancy Value Approach

Expectancy Value (EV) is a framework for examining health risk assessments and predicting health behaviors. It assumes that the outcome is a function of the importance of the behavioral outcome (the “V”) and the respondent’s subjective probability of the likelihood of attaining the outcome (the “E”) if the target behavior is performed (Ajzen & Fishbein, 2000). Expectancies are incorporated in theories that assume the risks and benefits of performing the target behavior are considered in determining prospective intention and actual behavioral performance. For example, in the Transtheoretical Model (Prochaska, 2008), expectancies are the “pros” and “cons” of performing the behavior. In the Reasoned Action Approach, expectancies are used in two ways. Semantic differential items like “Bad-Good,” “Foolish-Wise” or “Ineffective-Effective” are general evaluations used to describe the target behavior defined by the intention measure. However, expectancies are also measured as “behavioral beliefs” concerning specific outcomes resulting from performing the target behavior. In the Health Belief Model (Becker, 1974; Janz & Becker, 1984; Rosenstock, 1966) expectancies are the “benefits” and “barriers” to performing the target behavior. The Health Belief Model (Blue & Valley, 2002; Limbu, Gautam, & Pham, 2022; Wong, Alias, Wong, Lee, & AbuBakar, 2020) and the Reasoned Action Approach (Agarwal, 2014; Bleakley et al., 2021; Montano, 1986; Norman, Wilding, & Conner, 2020) have been used to predict vaccination acceptance behaviors in relation to both the H1N1 and the SARS-CoV-2 virus. Note however that while expectancies are often incorporated within behavioral theories, expectancies can be used as stand-alone predictors of intention and behaviors such as eating disorders (Hohlstein, Smith, & Atlas, 1998), media use (Galloway, 1981), drinking and driving (Grube & Voas, 1996), sexual behavior (Fromme, Katz, & Rivet, 1997; Hull & Bond, 1986), consumption of alcohol (Hittner, 1997; Hull & Bond, 1986; Jones, Corbin, & Fromme, 2001; Spijkerman, van den Eijnden, Vitale, & Engels, 2004) and sex and alcohol use (Fromme, D’Amico, & Katz, 1999; O’Hare, 2001).

Reasoned Action, Expectancies, and COVID-19 Vaccination Intention

Now that COVID-19 vaccination intention has now been added to this list of expectancy value applications, reviewing reason action research on vaccination intention highlights the systematic use of expectancy-type variables. Bleakley et al. (Bleakley, Hennessy, et al., 2022) investigated vaccination intention in US adults over 50 years of age across three ethnic groups (White, Black, and Hispanic). For all groups, attitude toward vaccination was the best theoretical predictor compared to normative pressure and behavioral control items. Lueck and Spiers (Lueck & Spiers, 2020) used a reasoned action elicitation study (e.g., Bleakley, Maloney, et al., 2022) and an online survey. In the elicitation, behavioral expectancies (e.g., protecting respondents from infection; uncertainty about vaccine efficacy), normative expectancies concerning important significant others who support vaccination (e.g., co-workers and friends; persons with pre-existing conditions) and control expectancies (e.g., having to pay for vaccination; having to travel to a vaccination site) were identified. Survey data analysis showed that “instrumental attitudes” was the single best predictor of vaccination intention. Akther and Nur (Akther & Nur, 2022) used a reasoned action model with attitude expectancies (e.g., “I feel injecting COVID-19 vaccine is a good idea”) and subjective norm items to predict vaccination intention along with the expectancy “precursor variables” (Hennessy et al., 2010) of perceived usefulness and perceived ease of use of the vaccine. In their SEM analysis, attitude had the highest standardized coefficient predicting vaccine acceptance. A similar study was performed by Baeza-Rivera et al. (Baeza-Rivera, Salazar-Fernández, Araneda-Leal, & Manríquez-Robles, 2021) analyzing survey data with a reasoned action SEM model predicting vaccine intention directly from beliefs about vaccine effectiveness, conspiracy beliefs, and injunctive normative beliefs (demographic and political orientation items were precursors). Expectancy beliefs about effectiveness had the largest direct effect on intention. Kim et al. (Kim, Lee, Ihm, & Kim, 2022) used a reasoned action SEM to predict vaccination intention with the proximal reasoned action predictors (i.e., semantic differential attitudes, injunctive norms, descriptive norms, and perceived behavioral control) with types of vaccine-related misinformation as precursors. Attitude had the second largest direct effect on intention (.44) and injunctive norm was first (.49) based on the standardized regression coefficient. Seddig et al. (Seddig, Maskileyson, Davidov, Ajzen, & Schmidt, 2022) used a reasoned action SEM analysis to predict vaccine intention. In their model with just the proximal predictors of intention, attitude had the largest effect with a standardized slope of .80. Finally, Chu and Liu (Chu & Liu, 2021) used SEM to predict vaccine intention using reasoned action items along with perceived risk measures (another type of behavioral expectancy). Their attitude measure was semantic differential expectancies predicted by four kinds of behavioral expectancies: anticipated benefits to the individual (e.g., “If I get the vaccines, I will be less likely to get COVID-19), anticipated benefits to the community (e.g., “COVID-19 vaccines protect the health of my community”), anticipated vaccine safety concerns (e.g., “Not enough research done on COVID-19 vaccines”), and anticipated financial concerns (e.g., “My insurance may not cover COVID-19 vaccine”). Their vaccine attitude variable had the largest total effect on vaccination intention. The results of all these studies strongly support the use of expectancy variables in predicting vaccination intention.

From Vaccination Intention to Vaccination Behavior: the Role of Selection Bias

Although an individual’s vaccination intention can legitimately change over time, modeling longitudinal vaccination status raises important analysis issues. As persons in the general population are vaccinated, fewer respondents in each wave of data collection are eligible for vaccination and these “later adopters” (who evolve into “vaccine refusers” if they remain unvaccinated) may be quite different than those initially vaccinated. Whenever data availability (e.g., groups of unvaccinated respondents versus groups of vaccinated respondents) change endogenously in this way, non-vaccinated status is not a random process but a result of a behavioral choice similar to deciding to get HIV test (Miller, Hennessy, Wendell, Webber, & Schoenbaum, 1996). Ignoring the selection process that produces the analysis sample (here, the group of unvaccinated respondents for whom the expectancy effects will be estimated) inevitably lead to a correlation between the predictor variables in the analysis equation and the error term of that equation: an example of “simultaneous equation bias” (Heckman, 1978).^[1]

Heckman’s regression method corrects the bias by simultaneously including an equation that accounts for the selection process directly (Cuddeback, Wilson, Orme, & Combs-Orme, 2004). One equation (the “selection” model) predicts the presence of relevant data (i.e., here differentiating between unvaccinated and vaccinated persons in the total sample) using a set of exogenous variables. The second equation (the “analysis” model) performs the analysis of the selected dependent variable using all the data that are available after selection (here, the remaining group of unvaccinated eligibles). Detailed reviews of the selection modeling approach are available elsewhere (Certo, Busenbark, Woo, & Semadeni, 2016; Infante-Rivard & Cusson, 2018).

Our research questions are:

• What at the relevant predictors of vaccination acceptance? We use established predictive models for vaccination intention but predict actual vaccination status. The correspondence between vaccination intention and vaccination acceptance is unknown, but as noted above, meta-analysis of the intention-behavior correlation is high for many health behaviors.

• Are expectancy value beliefs effective in predicting vaccination status for sample respondents who have still not been vaccinated at the end of the data collection period? In other words, do expectancies differentiate between those who refuse to get vaccinated (Refusers) and those who were vaccinated or plan on doing so (Acceptors) in the vaccination eligible group?

Methods and Measures

Data Collection

This research is a component of a study of US adults 50 years of age or older. Because older persons have greater susceptibility to SARS-CoV-2 virus due to age, pre-existing conditions, or immuno-suppression, research on vaccination acceptance in this group is important to control infection and reduce COVID-19 related death. Online survey data were collected by the research firm SSRS (www.ssrs.com) over 12 waves from October 2020 (wave 1) to September 2021 (wave 12)^[2] to track COVID-19 beliefs and behaviors as the epidemic progressed. Because infection presents great risk to health for older individuals, wave 1 consisted of a large sample survey of respondents 50 years of age or more (N = 2284) and a smaller sample of adults between 18 and 49 years of age. The sample used quotas for racial/ethnic group to achieve equal sample sizes for White, Black, and Hispanic adults. Wave 2 was a follow-up data collection on the same respondents (Bleakley et al., 2021). Waves 3 through 11 used a repeated cross-sectional independent sample design (n = approximately 225 for each wave) to track changes in COVID-19 related beliefs and behaviors. A volunteer web panel was used to identify respondents who reported their age as 50 or older and self-reported being White, non-Hispanic, Black, non-Hispanic or Hispanic to complete the survey (web panelists who completed the survey received a 5 dollar give card). In each wave, quotas (approximately 75 respondents per race/ethnicity group) were employed to reach similar numbers of respondents in each group. Wave 12 was designed to complement the wave 1 survey and collected data from a larger sample (White: n = 295; Black: n = 351; Hispanic: n = 381). All surveys were approved by the Institutional Review Board of the University of Delaware and participating respondents received a 5$ gift card for compensation.

The survey took approximately 30-45 minutes to complete. The survey system implemented several quality assurance procedures including a “speeder trap” to aid in identifying those who do not appear to be giving the survey the attention it deserves to complete a quality survey. In addition, respondents with an interview length less than 10 minutes or respondents who skipped or left blank more than 20 percent of the questions were removed. In the data used here, few respondents were eliminated in this way (wave 6 had he maximum of 4 eliminated respondents for these reasons).

Data were weighted to provide nationally representative and projectable estimates within each race and ethnicity group (White non-Hispanic; Black non-Hispanic; Hispanic). The weighting ensures that the demographic profile of the sample matches the profile of the target population. The parameters used in the post stratification weighting were age (50-64, 65+), sex (M/F), education level attained (Less Then College Degree, Bachelor’s Degree or Higher), and census region (Northeast, Midwest, South, West). The demographic benchmarks were obtained from the 2018 American Community Survey conducted by the US Census (https://www.census.gov/programs-surveys/acs). Data were weighted within race (White, non-Hispanic; Black, non-Hispanic; Hispanic; Other, non-Hispanic). The weights were trimmed at the 4th and 96th percentiles to prevent individual respondents having too much influence on the survey-derived estimates given the sample size of each group.

Data Pooling Tests for Data Waves 6-11

The independent cross-sectional sampling for waves 3-11 enabled survey items to be added or deleted in response to changes in the COVID-19 social, epidemiological, and political context. For example, in waves 1-4 most of the focus was on mask-wearing perceptions and behaviors but in wave 5 and later, the focus changed to vaccination perceptions and behaviors. However, because the monthly samples in waves 3-11 were small, data pooling tests to verify pooling of wave-specific data to maximize statistical power is necessary. Therefore, using data from waves 6 through 11 only (i.e., after the survey focus switched to vaccination behaviors and beliefs), we performed data pooling tests on predictor variables that should not be affected by COVID-19 mask-wearing or vaccination behavior but would only show variation through sampling error. These tests take the pooling variable of interest and compare two models: a constrained data set that assumes the same mean and SD for each wave and a free model that allows each wave to have a different mean and SD (that is, their actual values in the data for that wave). As a sensitivity test for false positives, we used percent male, age, a media bias index, an index of belief in scientific conspiracies: demographic and trait variables that should not change over time. None of these comparisons were significantly different using a likelihood ratio test (p for percent male: .19; p for age: .14; p for media bias: .97; p for science conspiracies: .20, all these tests have a df of 10). The p for education was .014, but this variable is a four-category ordinal measure with a skewed distribution toward lower educational levels. As a sensitivity test for false negatives, we used the percent vaccination as a pooling variable, an outcome that should be very sensitive to wave of the study: this p value was < .05 as expected. Thus, are confident in our use of pooled data from waves 6-11.

Measures

Vaccination status was a result of self-reports at each wave. Those reporting “I have been fully vaccinated” were the Vaccinated group. If respondents reported that “I do not want to get the vaccine” they were classified in the Refuser group. The two other alternatives (“I have completed one shot but still need to receive my second one” or “I am eligible to receive the vaccine and want to get it but have been unable to get an appointment” defined the Accepter group. The Refuser and Acceptor group defined the eligible group of unvaccinated respondents. Behavioral expectancies were derived from earlier studies on vaccination intention (Agarwal, 2014; Montano, 1986; Su, Chengbo, & Mackert, 2019) and are from (Bleakley, Hennessy, et al., 2022). The expectancy items had the stem: “How likely do you think each of the following are. If I were to get the coronavirus (COVID-19) vaccine it would:” with the items “Prevent me from getting sick with coronavirus (COVID-19)”, “Protect others from getting coronavirus (COVID-19)”, “Make me feel safer”, “Make me feel sick or give me side effects”, “Make me feel like a ‘guinea pig”, “Make others think of me as a role model”, “Inspire others to get the vaccine”, and “Improve my community’s trust in medical research” coded from −3 to +3 from “extremely unlikely” to “extremely likely”. The items were categorized into 3 belief indices because the items create serious multicollinearity issues when used individually (i.e., the determinant of the complete belief items correlation matrix is .01, a sign of problems when used as predictors in a regression (Rockwell, 1975)). In addition, a three factor logistic confirmatory factor analysis showed a moderately good fit (RMSEA: .078, CFI: .95, Tucker-Lewis Index: .92) even though underlying beliefs usually are not be treated as items with a common cause, i.e., effect indicators, but rather like elements of an index, i.e., causal indicators (Hennessy, Bleakley, & Fishbein, 2012). Thus, we used the three expectancy indices in the analysis below: the first three items defined the positive expectancy index with a polychoric alpha (Gadermann, Guhn, & Zumbo, 2012; Zumbo, Gadermann, & Zeisser, 2007) of .85. The average of the next two items defined the negative expectancy index. The last three items defined the community expectancy index, polychoric alpha: .89. Scientific conspiracy beliefs (from Bleakley, Hennessy, et al., 2022) had the stem: “There is often debate about whether or not the public is told the whole truth about various important issues. Please indicate the degree to which you believe each statement is likely to be true”. The items were: “The spread of certain viruses and/or diseases is the result of the deliberate concealed efforts of some organization(s)”, “Technology with mind-control capacities is used on people without their knowledge”, and “Experiments involving new drugs or technologies are routinely carried out on the public without their knowledge or consent” coded −2 to +2 from “definitely not true” to “definitely true” (polychoric alpha: .86). Covid-19 misinformation items (also from Bleakley, Hennessy, et al., 2022) had the stem: “How much do you agree or disagree that:” and the items were “Coronavirus (COVID-19) has affected most countries more negatively than the United States”, “The coronavirus (COVID-19) pandemic is mostly over in the United States”, “The flu is more lethal than coronavirus (COVID-19)”, “Getting vaccinated for COVID-19 could be riskier to your health than getting COVID-19”, “Some of the coronavirus (COVID-19) vaccines will alter people’s DNA”, “The official count of COVID-19 deaths in the United States has been over-reported”, and “The ingredients used to make the COVID-19 vaccine are harmful to an individual’s health” coded −3 to +3 from “strongly disagree” to “strongly agree” (polychoric alpha: .85). Covid-19 media bias (from Bangerter et al., 2012) items had the stem: “ Please indicate how much you agree or disagree with the following statements:” with the items “Media have exaggerated the risk posed by coronavirus (COVID-19)”, “Media provide information to avoid outbreaks and health problems caused by coronavirus (COVID-19)” (item reversed), and “One cannot trust what one hears in the media about coronavirus (COVID-19)” coded −2 to +2 from “strongly disagree” to “strongly agree” (polychoric alpha: .75). Respondent descriptors were self-reports of biological sex, age in years, education (3 group classification) and political party identification (Democratic, Republican, Independent) and the respondent’s ethnic group based on SSRS quota sampling selection procedures.

Statistical Analysis

Bar graphs and tabulations track univariate changes over time. Selection bias modeling was performed using Stata’s (StataCorp, 2019) Heckman selection bias modeling module for dichotomous outcomes using probit regression: heckprob. The components of the two Heckman equations were as follows. Selection Model: the dependent variable is non-vaccinated status using all the available data. Non-vaccinated status is a function of respondent sex, age, education, political party identification, wave^[3], and because of the quota sample design, respondent ethnicity. COVID-19 specific predictors are one index related to COVID-19 misinformation, one index concerning belief in scientific conspiracies, and one index reflecting COVID-19 media bias because these kinds of media and knowledge related perceptions are known to reduce COVID-19 prevention intention (Dees & Berman, 2013; Hornik et al., 2021; Romer & Jamieson, 2020; Stroope et al., 2021; Y. Wang & Liu, 2022) and thus should be positively related to unvaccinated status. Analysis Model: The dependent variable is Refuser status (compared to Acceptor status) in the eligible group of unvaccinated respondents. The predictors are the three indices of the behavioral expectancies as well as wave of the study (to adjust for changing percentages of the three vaccination groups over time). If expectancy value theory holds in the eligible group, the indices for positive and community outcomes (community outcomes were framed as positively valued), should reduce Refuser status. If the index refers to negative outcomes, it should increase the probability of Refuser status for the vaccination eligible group.

Results

In the sample of 2117 respondents, 1627 were vaccinated (77%). In the eligible group (n = 490), 61% were Refusers and 39% were Accepters. Table 1 shows the information on the predictors for each race group. On average, the three COVID-19 belief indices (science conspiracies, misperceptions, media bias) were negative (i.e., not endorsed).

Table 1:

Respondent Characteristics (N = 2117)

	Average or Percent	95% Confidence Interval
Race (%)		Lower	Upper
White	65.0	64.2	65.7
Black	62.8	62.2	63.4
Hispanic	61.5	60.8	62.2
Male (%)
White	47.3	42.7	51.8
Black	42.3	38.0	46.6
Hispanic	47.1	42.4	51.8
Vaccinated (%)
White	77.5	73.3	81.7
Black	68.5	64.5	72.6
Hispanic	75.3	71.4	79.2
Accepter (%)
White	6.3	3.8	8.9
Black	12.0	9.0	15.0
Hispanic	13.0	9.7	16.2
Refuser (%)
White	16.2	12.5	19.8
Black	19.5	16.1	23.0
Hispanic	11.7	8.8	14.6
Scientific Conspiracy Index
White	−0.4	−0.5	−0.3
Black	−0.2	−0.3	−0.1
Hispanic	−0.2	−0.3	−0.2
Misinformation Index
White	−1.0	−1.1	−0.8
Black	−1.1	−1.2	−1.0
Hispanic	−0.8	−0.9	−0.7
Media Bias Index
White	−0.2	−0.3	−0.1
Black	−0.5	−0.6	−0.5
Hispanic	−0.2	−0.3	−0.1
Education (%)	High School or Less	Some College	College Degree or More
White	19.3	34.3	46.4
Black	25.17	42.7	32.2
Hispanic	20.8	40.6	38.6
Political Party (%)	Democrat	Republican	Independent
White	34.6	37.6	27.8
Black	82.9	2.6	14.5
Hispanic	47.6	26.5	25.8

Figure 1 shows how the distribution of the vaccine classification groups changed over time from wave 6 to wave 12. Total vaccinated percentages rose over time (from 43% to around 80%) while the Accepter group fell, and starting at wave 8, the majority of the vaccination eligibles are Refusers, not Accepters. Because of the increasing influence of Refusers over time, the two positive expectancy indices move from a positive average to a negative average while the negative expectancy index remained stable. This may suggest that the positive expectancy indices will perform poorly in influencing the transition from Refuser to Accepters status in the unvaccinated eligible group.

Figure 1:

Vaccine Groups and Average Expectancy Indices by Wave of Study

Selection and Analysis Model Results for Behavioral Expectancies

Table 2 shows the selection and the analysis model results. The selection results predicting unvaccinated status show that age ( b = −0.03, t = −5.13, p < .05), higher education levels (some college: b = −0.20, t = −2.13, p = 0.03; college degree or more: b = −0.54, t = −5.24, p < .05), and being a Democrat (b = −0.31, t = −2.86, p < .05) significantly decreased unvaccinated status. Wave of the study was negatively related to unvaccinated status (the signs of all the coefficients are negative) because over time more people are vaccinated and fewer are unvaccinated, consistent with the data shown in Figure 1. Factors that decreased COVID-19 vaccination likelihood include being Black (b = 0.53, t = 4.53, p < .05), belief in scientific conspiracies (b = 0.11, t = 2.24, p = 0.03) and endorsement of COVID-19 related misinformation beliefs (b = 0.39, t = 8.72, p < .05). Perceived media bias had a negative effect on vaccination (b = 0.02) but did not attain statistical significance (p = 0.75). The selection model fit the data well, the pseudo-R² was 0.43 (Hu, Shao, & Palta, 2006; McKelvey & Zavoina, 1975; Veall & Zimmermann, 1996).

Table 2:

Results of Heckman Regression Model

Predictors	Analysis Model for Expectancy Effects for Refuser vs. Accepter Status	Selection Model for Unvaccinated vs. Vaccinated Choice
	Slope (t-value) p value	Slope (t-value) p value
Positive Expectancy Index	−0.42
	(−5.07)
	0.00
Negative Expectancy Index	0.24
	(3.38)
	0.00
Community Expectancy Index	−0.26
	(−2.57)
	0.01
Wave 7	0.130
	(0.45)
	0.65
Wave 8	0.50
	(1.50)
	0.12
Wave 9	1.03
	(2.87)
	0.01
Wave 10	0.22
	(0.62)
	0.53
Wave 11	0.96
	(2.38)
	0.02
Wave 12	0.21
	(0.73)
	0.48
Male Sex		−0.01
		(−1.86)
		0.06
Age		−0.03
		(−5.13)
		0.00
Some College		−0.20
		(−2.13)
		0.03
College Degree Plus		−0.54
		(−5.24)
		0.00
Scientific Conspiracy Index		0.11
		(2.24)
		0.03
Misinformation Index		0.390
		(8.72)
		0.00
Media Bias Index		0.02
		(0.29)
		0.75
Democrat		−0.31
		(−2.86)
		0.00
Republican		0.01
		(0.11)
		0.91
Black		0.530
		(4.53)
		0.00
Hispanic		−0.01
		(−0.11)
		0.92
Wave 7		−0.57
		(−3.36)
		0.00
Wave 8		−0.74
		(−4.19)
		0.000
Wave 9		−1.29
		(−7.39)
		0.00
Wave 10		−1.00
		(−5.18)
		0.000
Wave 11		−1.32
		(−6.99)
		0.00
Wave 12		−1.27
		(−9.14)
		0.00
Intercept	0.06	2.54
	(0.24)	(6.64)
	0.81	0.000
Error correlation between Selection and Analysis Equation	−0.18
N	490	2117
# McKelvey & Zavoina R2	.65	.43

Notes: For education, “some high school or less” is the comparison. For race, “white” is the comparison. For study wave, “wave 6” is the comparison. For party identification, “Independent” is the comparison.

^#Because the Heckman model estimates both equations at the same time, it is not possible to get an equation level pseudo-R² directly. They were estimated from two independent logistic regressions on the same sample and the identical model variables as shown here: the coefficients were similar, often only differing in the second decimal place due to the correlation between the two equations in the Heckman model which is not present with the two independent logistic regressions

The analysis model predicted Refuser vs. Accepter status for the group of eligible unvaccinated respondents only. These results show that the positive outcome index (b = −0.42, t = −5.07, p < .05) and the community index (b = −0.26, t = −2.57, p = 0.01) decreased the likelihood of Refuser status while negative outcome index increased the likelihood of refusal (b = 0.24, t = 3.38, p < .05). The analysis equation fit the data well: the pseudo-R² is 0.65. Thus, COVID-19 expectancies operated as expected even in this selected sample of vaccination eligible group that became more vaccine hesitant over time.

Study Limitations

The sample used here is a composite of a representative sample and a volunteer convenience sample: this may bias the results in unknown ways. Clearly the focus on older adults assumes that reducing short-term negative public health consequences is more immediately important than population representativeness, but this may be a reasonable decision given the greater dangers of infection for older adults. Nonetheless, younger vaccine refusers, for example, may have quite different expectancy-based motivations for their behavior than older ones. A longitudinal repeated measures design would be able to track individual variation in intention over time, but for the purposes here, repeated measures may not be an advantage because vaccinated status is not reversible in the same way as is vaccination intention.

Discussion

The selection model answered the question: who are COVID-19 vaccine refusers? Results showed that vaccine refusers were younger, less educated, endorsed scientific conspiracy beliefs and common misconceptions about the epidemic, and were Black. Our results for predicting vaccination refusers are consistent with the published narrative review of predictors of vaccination intenders (Y. Wang & Liu, 2022) which suggests that the correlation between vaccination intention and vaccination behavior is high.

The analysis model answers the question: do expectancies about outcomes decrease vaccination refusal for vaccine eligible persons? The analysis equation results showed that consistent with expectancy theory, underlying behavioral beliefs about the outcomes of vaccination were related to vaccination refusal: negative expectancy outcomes increased vaccine refusal and, conversely, positive outcomes decreased refusals. In other words, expectancies still were related to vaccination status even though over half of the eligible respondents were Refusers and not Accepters.

One alternative way to understand vaccination behavior might be to investigate the role of ideological orientation on vaccination choice. We can produce the adjusted marginal probabilities of both unvaccinated (based on the selection model) and Refuser status (based on the analysis model) for any kind of respondent distinctions. Figure 2 shows these applications based on a seven-level political orientation item (coded “Very Liberal” to “Very Conservative”). Figure 2(A) shows that for the unvaccinated choice, Liberals had the lowest probability and Conservatives the highest. Figure 2(B) has the corresponding results for Refuser vs. Accepter status. Because this vaccine eligible group is dominated by respondents who are Refusers, the non-vaccination outcome probabilities are much higher overall especially for the two highest levels of “conservative” (we omit the confidence intervals due to small sample sizes for the unvaccinated eligible group). Thus, we find that political traits are important for audience segmentation and accurate political category targeting.

Figure 2:

(A) Probability of Unvaccinated Status and (B) Probability of Refuser Status Given Eligible Group by Political Ideology

Notes: VL: Very Liberal L: Liberal SL: Somewhat Liberal MOD: Moderate SC: Somewhat Conservative C: Conservative VC: Very Conservative. (A) estimated from selection model results in Table 2. (B) Estimated from analysis model results in Table 2.

For intervention message design, however, our model specification is preferable. Respondents’ expectancies (i.e., “pros or cons”, “underlying beliefs”, or “benefits and barriers” depending on the specific theory) are a basis for intervention development to induce behavior change. In the reasoned action approach, changes in underlying beliefs – specific outcome expectancies – determine intention to change the corresponding specific behavior (Yzer, 2013). Thus, reasoned action based interventions are designed to reinforce the belief predictors that are positively associated with intention and counter-argue the belief predictors that are negatively associated with intention (Bleakley, Piotrowski, Hennessy, & Jordan, 2013; Gregorio-Pascual & Mahler, 2019; Hennessy, Bleakley, Mallya, & Romer, 2014; Massi Lindsey, 2017; Norman et al., 2018).

Unfortunately, political ideology and other global respondent characteristics while perhaps associated with behavior, generally do not have counter-arguable elements: these kinds of variables are identity-driven and more like personality traits or decision-making strategies (Bernhard & Freeder, 2020; Schaffner & Streb, 2002; X. T. Wang, 2008). For example, pro-vaccination messages will have difficulty counter-arguing a decision-making heuristic such as “I am very conservative, therefore I oppose vaccinations” but may have fewer problems intervening against such COVID-19 vaccination myths as “COVID-19 vaccinations implant microchips into people to control them” (Skafle, Nordahl-Hansen, Quintana, Wynn, & Gabarron, 2022), which can be counterargued through a variety of persuasion strategies. Thus, for intervention message design, specific expectancies should be the components of vaccination promoting messages while the approach demonstrated in Figure 2 may play an important role in identifying specific message-resistant or message-favorable social groups (i.e., target audiences) during message development and message testing formative research.

Conclusion

As some US adults resist SARS-CoV-2 vaccination, explaining vaccination refusal is a public health priority. In this paper we used data collected over a 6-month period after vaccine availability to track vaccine uptake and refusal using selection bias modeling and expectancy theory. Expectancies are incorporated into theories of behavioral change because individuals are assumed to assess the advantages and disadvantages associated with performing a specific behavior. The results here show that expectancies were operating even though over half of the unvaccinated respondents were Refusers and not Accepters, validating the assumption that expectancies are important to identify because they are modifiable and provide a point of intervention, not just for COVID-19 vaccination acceptance, but also for other volitional health promoting behaviors.