In early 2020, the COVID-19 pandemic wreaked havoc worldwide, triggering rapid vaccine development efforts. Despite federal, state, and workplace vaccination mandates, many individuals made judgments against COVID-19 vaccination, leading researchers to study the psychology underlying individual vaccination preferences and what might differentiate the framework for judgment between individuals who were fully vaccinated against COVID-19 and those who were not (henceforth referred to as vaccine uptake). A better understanding of these differences in judgment may highlight targets for public messaging and education to increase the incidence of choosing vaccination.

Multiple studies have sought to predict an individual’s intention to receive a COVID-19 vaccine or specific variables underlying vaccination choices or mitigation strategies [1-7], but few have predicted vaccine uptake. One such study used 83 sociodemographic variables (with education, ethnicity, internet access, income, longitude, and latitude being the most important predictors) to predict vaccine uptake with 62% accuracy [8], confirming both the importance and limitations of these variables in prediction models. Other studies have compared demographic groups between vaccinated and nonvaccinated persons; Bulusu et al [9] found that young adults (aged 18-35 years), women, and those with higher levels of education had higher odds of being vaccinated. In a study of >12 million persons, the largest percentage of those who initiated COVID-19 vaccination were White, non-Hispanic women between the ages of 50 and 64 years [10]. Demographic variables are known to affect how individuals judge what is rewarding or aversive [11,12] yet are not themselves variables quantifying how individuals make judgments that then frame decisions.

Judgment reflects an individual’s preferences, or the variable extent to which they approach or avoid events in the world based on the rewarding or aversive effects of these events [13-15]. The definition of preference in psychology differs from that in economics. In psychology, preferences are associated with “wanting” and “liking” and are framed by judgments that precede decisions, which can be quantified through reinforcement reward or incentive reward tasks [12,16-21]. In economics, preferences are relations derived from consumer choice data (refer to the axioms of revealed preference [22]) and reflect choices or decisions based on judgments that place value on behavioral options. Economist Paul Samuelson noted that decisions are “assumed to be correlative to desire or want” [23]. In this study, we focused on a set of variables that frame judgment, with the presumption that judgments precede choices [12,20]. Variables that frame judgment can be derived from tasks using operant key-pressing tasks that quantify “wanting” [24-33] or simple rating tasks that are analogous to “liking” [20,34]. Both operant keypress and rating tasks measure variables that quantify the average (mean) magnitude (K), variance (σ), and pattern (ie, Shannon entropy [H]) of reward and aversion judgments [35]. We refer to this methodology and the multiple relationships between these variables and features based on their graphical relationships as relative preference theory (RPT; Figure 1) [18,36]. RPT has been shown to produce discrete, recurrent, robust, and scalable relationships between judgment variables [37] that produce mechanistic models for prediction [33], and which have demonstrated relationships to brain circuitry [24-27,30] and psychiatric illness [28]. Of the graphs produced for RPT, 2 appear to resemble graphs derived with different variables in economics, namely, prospect theory [38] and the mean-variance function for portfolio theory described by Markowitz [39]. Given this graphical resemblance, it is important to note that RPT functions quantifying value are not the same as standard representations of preference in economics. Behavioral economic variables such as loss aversion and risk aversion [38,40-51] are not to be interpreted in the same context given that both reflect biases and bounds to human rationality. In psychology, they are grounded in judgments that precede decisions, whereas in economics, they are grounded in consumer decisions themselves. Going forward, we will focus on judgment-based loss aversion, representing the overweighting of negative judgments relative to positive ones, and judgment-based risk aversion, representing the preference for small but certain assessments over larger but less certain ones (ie, assessments that have more variance associated with them) [38,40-51]. Herein, loss aversion and risk aversion refer to ratings or judgments that precede decisions.

A number of studies have described how risk aversion and other judgment variables are important for individual vaccine choices and hesitancies [52-58]. Hudson and Montelpare [54] found that risk aversion may promote vaccine adherence when people perceive contracting a disease as more dangerous or likely. Trueblood et al [52] noticed that those who were more risk seeking (as measured via a gamble ladder task) were more likely to receive the vaccine even if the vaccine was described as expedited. Wagner et al [53] described how risk misperceptions (when the actual risk does not align with the perceived risk) may result from a combination of cognitive biases, including loss aversion. A complex theoretical model using historical vaccine attitudes grounded in decision-making has also been proposed to predict COVID-19 vaccination, but this model has not yet been tested [59]. To our knowledge, no study has assessed how well a model comprising variables that reflect reward and aversion judgments predicts vaccine uptake.

Given the many vaccine-related issues that occurred during the COVID-19 pandemic (eg, vaccine shortages, hospital overload, and vaccination resistance or hesitancy), it is critical to develop methods that might improve planning around such shortcomings. Because judgment variables are fundamental to vaccine choice, they provide a viable target for predicting vaccine uptake. In addition, the rating methodology used to quantify variables of judgment is independent of methods quantifying vaccine uptake or intent to vaccinate, limiting response biases within the study data.

In this study, we aimed to predict COVID-19 vaccine uptake using judgment, demographic, and COVID-19 precaution (ie, behaviors minimizing potential exposure to COVID-19) variables using multiple machine learning algorithms, including logistic regression, random forest, and balanced random forest (BRF). BRF was hypothesized to perform best given its potential benefits with handling class imbalances [60], balancing both recall and specificity, and producing Gini scores that provide relative variable importance to prediction. In this study, the need for data imbalance techniques was motivated by the importance of the specificity metric, which would reflect the proportion of participants who did not receive full vaccination; without balancing, the model might not achieve similar recall and specificity values. When there is a large difference between recall and specificity, specificity might instead reflect the size of the minority class (those who did not receive full vaccination). In general, random forest approaches have been reported to have benefits over other approaches such as principal component analysis and neural networks, in which the N-dimensional feature space or layers (in the case of neural networks) are complex nonlinear functions, making it difficult to interpret variable importance and relationships to the outcome variable. To provide greater certainty about these assumptions, we performed logistic regression in parallel with random forest and BRF. The 3 machine learning approaches used a small feature set (<20) with interpretable relationships to the predicted variable. Such interpretations may not be achievable in big data approaches that use hundreds to thousands of variables that seemingly add little significance to the prediction models. Interpretation was facilitated by (1) the Gini importance criterion associated with BRF and random forest, which provided a profile of the judgment variables most important for prediction; and (2) mediation and moderation analyses that offered insights into statistical mechanisms among judgment variables, demographic (contextual) variables, and vaccine uptake. Determining whether judgment variables are predictive of COVID-19 vaccine uptake and defining which demographic variables facilitate this prediction presents a number of behavioral targets for vaccine education and messaging—and potentially identifies actionable targets for increasing vaccine uptake.

More broadly, the prediction of vaccine uptake may aid (1) vaccine supply chain and administration logistics by indicating areas that may need more or fewer vaccines, (2) targeted governmental messaging to locations with low predicted uptake, and (3) preparation of areas that may experience high cases of infection that could ultimately impact health care preparedness and infrastructure. The proposed method could also be applied to other mandated or government-recommended vaccines (eg, influenza and human papillomavirus) to facilitate the aforementioned logistics. Locally, vaccine uptake prediction could facilitate local messaging and prepare health care institutions for vaccine rollout and potential hospital overload. Nationally, prediction might inform public health officials and government communication bodies that are responsible for messaging and vaccine rollout with the goal of improving vaccine uptake and limiting infection and hospital overload.

Similar recruitment procedures for a smaller population-based study have been described previously [61-63]. In this study, participants were randomly sampled from the general US population using an email survey database used by Gold Research, Inc. Gold Research administered questionnaires in December 2021 using recruitment formats such as (1) customer databases from large companies that participate in revenue-sharing agreements, (2) social media, and (3) direct mail. Recruited participants followed a double opt-in consent procedure that included primary participation in the study as well as secondary use of anonymized, deidentified data (ie, all identifying information was removed by Gold Research before retrieval by the research group) in secondary analyses (refer to the Ethical Considerations section for more detail). During consent procedures, participants provided demographic information (eg, age, race, and sex) to ensure that the sampled participants adequately represented the US census at the time of the survey (December 2021). Respondents were also presented with repeated test questions to screen out those providing random and illogical responses or showing flatline or speeder behavior. Participants who provided such data were flagged, and their data were removed.

Because other components of the survey required an adequate sample of participants with mental health conditions, Gold Research oversampled 15% (60,000/400,000) of the sample for mental health conditions, and >400,000 respondents were contacted to complete the questionnaire. Gold Research estimated that, of the 400,000 participants, >300,000 (>75%) either did not respond or declined to participate. Of the remaining 25% (100,000/400,000) who clicked on the survey link, >50% (52,000/100,000) did not fully complete the questionnaire. Of the ≥48,000 participants who completed the survey (ie, ≥48,000/400,000, ≥12% of the initial pool of queried persons), those who did not clear data integrity assessments were omitted. Participants who met quality assurance procedures (refer to the following section) were selected, with a limit of 4000 to 4050 total participants.

Eligible participants were required to be aged between 18 and 70 years at the time of the survey, comprehend the English language, and have access to an electronic device (eg, laptop or smartphone).

All participants provided informed consent, which included their primary participation in the study as well as the secondary use of their anonymized, deidentified data (ie, all identifying information removed by Gold Research before retrieval by the research group) in secondary analyses. This study was approved by the Northwestern University institutional review board (approval STU00213665) for the initial project start and later by the University of Cincinnati institutional review board (approval 2023-0164) as some Northwestern University investigators moved to the University of Cincinnati. All study approvals were in accordance with the Declaration of Helsinki. All participants were compensated with US $10 for taking part. Detailed survey instructions have been published previously [61-63].

Three additional quality assurance measures were used to flag nonadhering participants: (1) participants who indicated that they had ≥10 clinician-diagnosed illnesses (refer to Figure S1 in Multimedia Appendix 1 [18,33,36,64-68] for a list), (2) participants who showed minimal variance in the picture-rating task (ie, all pictures were rated the same or the ratings varied only by 1 point; refer to the Picture-Rating Task section), and (3) inconsistencies between educational level and years of education and participants who completed the questionnaire in <800 seconds.

Data from 4019 participants who passed the initial data integrity assessments were anonymized and then sent to the research team. Data were further excluded if the quantitative feature set derived from the picture-rating task was incomplete or if there were extreme outliers (refer to the RPT Framework section). Using these exclusion criteria, of the 4019 participants, 3476 (86.49%) were cleared for statistical analysis, representing 0.87% (3476/400,000) of the initial recruitment pool. A flowchart of participant exclusion is shown in Figure 2.

Participants were asked to report their age, sex, ethnicity, annual household income, marital status, employment status, and educational level. Participants were asked to report whether they had received the full vaccination (yes or no responses). At the time of the survey, participants were likely to have received either 2 doses of the Pfizer or Moderna vaccine or 1 dose of the Johnson & Johnson vaccine as per the Centers for Disease Control and Prevention guidelines. Participants were also asked to respond yes (they routinely followed the precaution) or no (they did not routinely follow the precaution) to 4 COVID-19 precaution behaviors: mask wearing, social distancing, washing or sanitizing hands, and not gathering in large groups (refer to Tables S1 and S2 in Multimedia Appendix 1 for the complete questions and sample sizes, respectively). In addition, participants completed a picture-rating task at 2 points during the survey (refer to the Picture-Rating Task section).

A picture-rating task was administered to quantify participants’ degree of liking and disliking a validated picture set using pictures calibrated over large samples for their emotional intensity and valence [69,70]. Ratings from this task have been mathematically modeled using RPT to define graphical features of reward and aversion judgments. Each feature quantifies a core aspect of judgment, including risk aversion and loss aversion. Judgment variables have been shown to meet the criteria for lawfulness [37] that produce mechanistic models for prediction [33], with published relationships to brain circuitry [24-27,30] and psychiatric illness [28]. A more complete description of these judgment variables and their computation can be found in the RPT Framework section and in Table 1.

For this task, participants were shown 48 unique color images from the International Affective Picture System [69,70]. A total of 6 picture categories were used: sports, disasters, cute animals, aggressive animals, nature (beach vs mountains), and men and women dressed minimally, with 8 pictures per category (48 pictures in total; Figure 1A). These images have been used and validated in research on human emotion, attention, and preferences [69,70]. The images were displayed on the participants’ digital devices with a maximum size of 1204 × 768 pixels. Below each picture was a rating scale from −3 (dislike very much) to +3 (like very much), where 0 indicated indifference (Figure 1A). While there was no time limit for selecting a picture rating, participants were asked to rate the images as quickly as possible and use their first impression. Once a rating was selected, the next image was displayed.

Ratings from the picture-rating task were analyzed using an RPT framework. This framework fits approach and avoidance curves and derives mathematical features from graphical plots (Figures 1B-1D). These methods have been described at length in prior work and are briefly described in this section [11,18,33,36]. More complete descriptions and quality assurance procedures can be found in Multimedia Appendix 1.

At least 15 judgment variables can be mathematically derived from this framework and are psychologically interpretable; they have been validated using both operant keypress [9,25-27] and picture-rating tasks [11,34]. The 15 judgment variables are loss aversion, risk aversion, loss resilience, ante, insurance, peak positive risk, peak negative risk, reward tipping point, aversion tipping point, total reward risk, total aversion risk, reward-aversion trade-off, trade-off range, reward-aversion consistency, and consistency range. Loss aversion, risk aversion, loss resilience, ante, and insurance are derived from the logarithmic or power-law fit of mean picture ratings (K) versus entropy of ratings (H); this is referred to as the value function (Figure 1B). Peak positive risk, peak negative risk, reward tipping point, aversion tipping point, total reward risk, and total aversion risk are derived from the quadratic fit of K versus the SD of picture ratings (σ); this is referred to as the limit function (Figure 1C). Risk aversion trade-off, trade-off range, risk aversion consistency, and consistency range are derived from the radial fit of the pattern of avoidance judgments (H₋) versus the pattern of approach judgments (H₊); this is referred to as the trade-off function (Figure 1D). Value (Figure S2A in Multimedia Appendix 1), limit (Figure S2B in Multimedia Appendix 1), and trade-off (Figure S2C in Multimedia Appendix 1) functions were plotted for 500 randomly sampled participants, and nonlinear curve fits were assessed for goodness of fit, yielding R², adjusted R², and the associated F statistic for all participants (Figure S2D in Multimedia Appendix 1). Only the logarithmic and quadratic fits are listed in Table S3 in Multimedia Appendix 1. Each feature describes a quantitative component of a participant’s reward and aversion judgment (refer to Table 1 for abbreviated descriptions and Multimedia Appendix 1 for complete descriptions). Collectively, the 15 RPT features will be henceforth referred to as “judgment variables.” The summary statistics for these variables can be found in Table S3 in Multimedia Appendix 1.

Of the 400,000 persons queried by Gold Research, Inc, 48,000 (12%) completed the survey, and 3476 (0.87%) survived all quality assurance procedures. Participants were predominately female, married, and White individuals; employed full time with some college education; and middle-aged (mean age 51.40, SD 14.92 years; Table 2). Of the 3476 participants, 2525 (72.64%) reported receiving a full dose of a COVID-19 vaccine, and 951 (27.36%) reported not receiving a full dose. Participants who indicated full vaccination were predominately female, married, White individuals, and retired; had some college education; and were older on average (mean age 54.19, SD 14.13 years) when compared to the total cohort. Participants who indicated that they did not receive the full vaccine were also predominately female, married, and White individuals. In contrast to those who received the full vaccination, those not fully vaccinated were predominately employed full time, high school graduates, and of average age (mean age 43.98, SD 14.45 years; median age 45, IQR 32-56 years) when compared to the total cohort. Table 2 summarizes the demographic group sample size percentages for the total cohort, those fully vaccinated, and those not fully vaccinated.

When comparing percentages between vaccination groups, a higher percentage of male individuals were fully vaccinated, and a higher percentage of female individuals were not fully vaccinated (Table 2). In addition, a higher percentage of married, White and Asian or Pacific Islander, and retired individuals indicated receiving the full vaccine when compared to the percentages of those who did not receive the vaccine (Table 2). Conversely, a higher percentage of single, African American, and unemployed individuals indicated not receiving the full vaccine (Table 2).

Because judgment variables and demographic variables (age, income, and educational level) were important predictors, we evaluated post hoc whether demographics statistically mediated or moderated the relationship between each of the 15 judgment variables and binary responses to COVID-19 vaccination.

For primary mediations, age significantly mediated the statistical relationship between 11 judgment variables and vaccine uptake (α<.05; Table 8), income mediated 8 relationships α< <.05; Table 8), and educational level mediated 9 relationships (α<.05; Table 8). In total, 7 judgment variables overlapped across the 3 models: loss resilience, ante, insurance, peak positive risk, peak negative risk, risk aversion trade-off, and consistency range. Of these, 5 significantly differed between vaccine uptake (those fully vaccinated and those not): loss resilience, ante, insurance, peak positive risk, and peak negative risk (Table 3). Thus, 2 judgment features did not differ by vaccine uptake but were connected with uptake by significant mediation.

For the secondary mediation analyses, 5 judgment variables mediated the statistical relationship between age and vaccine uptake; these variables overlapped with the 11 findings of the primary mediation analyses. Furthermore, 4 judgment variables mediated the statistical relationship between income and vaccine uptake; these variables overlapped with the 8 findings of the primary mediation analyses. Finally, 4 judgment variables mediated the statistical relationship between educational level and vaccine uptake; these variables overlapped with the 9 findings of the primary mediation analyses. In all secondary analyses, approximately half of the judgment variables were involved in mediation as compared to the doubling of judgment variable numbers observed in the primary mediation analyses. In the secondary mediation analyses, the same 4 judgment variables were found in both primary and secondary mediation results, indicating a mixed mediation framework.

From the moderation analyses, only 2 interactions out of a potential 45 were observed. Age interacted with risk aversion trade-off, and income interacted with loss resilience to statistically predict vaccine uptake (α<.05; Table 8). The 2 moderation results overlapped with the mediation results, indicating mixed mediation-moderation relationships [78,80,81].

Relatively few studies have sought to predict COVID-19 vaccine uptake using machine learning approaches [8,59]. Given that a small set of studies has assessed the psychological basis that may underlie vaccine uptake and choices [6,52,53,56,58,59,83], but none have used computational cognition variables based on reward and aversion judgment to predict vaccine uptake, we sought to assess whether variables quantifying human judgment predicted vaccine uptake. This study found that 7 demographic and 15 judgment variables predicted vaccine uptake with balanced and moderate recall and specificity, moderate accuracy, high AUROC, and high precision using a BRF framework. Other machine learning approaches (random forest and logistic regression) produced higher accuracies but lower specificities, indicating a lower prediction of those who did not receive the vaccine. The BRF also had challenges predicting the negative class, as demonstrated by the relatively low NPV despite having higher specificity than random forest and logistic regression. Feature importance analyses from both BRF and random forest showed that the judgment variables collectively dominated the Gini importance scores. Furthermore, demographic variables acted as statistical mediators in the relationship between judgment variables and vaccine uptake. These mediation findings support the interpretation of the machine learning results that demographic factors, together with judgment variables, predict COVID-19 vaccine uptake.

Those who were fully vaccinated had lower values for loss aversion, ante, peak positive risk, peak negative risk, total reward risk, and total aversion risk, along with higher values for risk aversion, loss resilience, insurance, and trade-off range (refer to Table 1 for variable descriptions). Lower loss aversion corresponds to less overweighting of bad outcomes relative to good ones [84] and a potential willingness to obtain a vaccine with uncertain outcomes. A lower ante suggests that individuals are less willing to engage in risky behaviors surrounding potential infection, which is also consistent with the 4 other judgment variables that define relationships between risk and value (peak positive risk, peak negative risk, total reward risk, and total aversion risk). In participants who indicated full vaccination, lower peak positive risk and peak negative risk were related to individuals having a lower risk that they must overcome to make a choice to either approach or avoid, as per the decision utility equation by Markowitz [39,71]. The lower total reward risk and total aversion risk indicate that the interactions between reward, aversion, and the risks associated with them did not scale significantly; namely, higher reward was not associated with higher risk, and higher negative outcomes were not associated with the uncertainty of them. For these participants, the ability of the vaccine to increase the probability of health and reduce the probability of harm from illness did not have to overcome high obstacles in their vaccine choice. Higher risk aversion in vaccinated participants suggests that these participants viewed contracting COVID-19 as a larger risk and, therefore, were more likely to receive the full dose. These findings are consistent with those of a study by Lepinteur et al [58], who found that risk-averse individuals were more likely to accept the COVID-19 vaccination, indicating that the perceived risk of contracting COVID-19 was greater than any risk from the vaccine. Hudson and Montelpare [54] also found that risk aversion may promote vaccine adherence when people perceive contracting a disease as more dangerous or likely. Higher loss resilience in the vaccinated group was also consistent with the perspective that vaccination would improve their resilience and act as a form of insurance against negative consequences. The higher trade-off range suggests that vaccinated individuals have a broader portfolio of preferences and are more adaptive to bad things occurring, whereas a lower trade-off indicates a restriction in preferences and less adaptability in those who did not receive the vaccine.

When testing these judgment variables (with demographic and COVID-19 precaution behavior variables) in a BRF framework to predict vaccine uptake, we observed a high AUROC of 0.79, where an AUROC of 0.8 is often the threshold for excellent model performance in machine learning [85,86]. The similarity of our reported recall and specificity values with the BRF suggests a balance between predicting true positives and true negatives. The high precision indicates a high certainty in predicting those who were fully vaccinated. The BRF model was successful in identifying those who received the full vaccine (positive cases; indicated by high precision and moderate recall) and those who did not (negative cases; indicated by the specificity). However, NPV was low, indicating a higher rate of false prediction of those who did not receive a full dose counterbalanced by a higher specificity that reflects a higher rate of predicting true negatives. These observations are reflected in the moderate accuracy, which measures the number of correct predictions. A comparison of random forest, logistic regression, and BRF revealed that random forest and logistic regression models produced less balance between recall (high) and specificity (low), which could be interpreted as a bias toward predicting the majority class (ie, those who received the vaccine). That being said, the NPV for BRF was lower than that for random forest and logistic regression, where a low NPV indicates a low probability that those predicted to have not received the vaccine truly did not receive the vaccine when taking both classes into account. Together, the results from all 3 machine learning approaches reveal challenges in predicting the negative class (ie, those who did not receive the vaccine). Overall, the 3 models achieved high accuracy, recall, precision, and AUROC. BRF produced a greater balance between recall and specificity, and the outcome of the worst-performing metric (ie, NPV) was still higher than the specificities for the random forest and logistic regression models.

Of the 3 prediction algorithms, random forest and BRF had very similar Gini importance results, whereas logistic regression elevated most demographic variables and a minority of judgment variables. This observation could be due to the large variance in each of the judgment variables, which could present challenges for achieving a good fit with logistic regression. In contrast, the demographic and COVID-19 precaution variables had low variance and could be more easily fit in a linear model, hence their significance in the logistic regression results. In comparison to logistic regression, decision trees (eg, BRF and random forest) use variable variance as additional information to optimize classification, potentially leading to a higher importance of judgment variables over most demographic and all COVID-19 precaution variables.

Focusing on the model with balanced recall and specificity (ie, the BRF classifiers [with and without COVID-19 precaution behaviors]), the top predictors were 3 demographic variables (age, income, and educational level), with distributions that varied by vaccine uptake in manners consistent with those of other reports. Namely, older individuals, those identifying as male and White individuals, and those who indicated a higher income and educational level corresponded to those who were or intended to be vaccinated [2,5,87]. Despite their saliency, these 3 variables together only contributed 23% to the prediction, corresponding to approximately one-third of the contribution from the 15 judgment variables (63%-69%). The individual Gini importance scores for the 15 judgment variables only ranged from 0.039 to 0.049 but were the dominant set of features behind the moderate accuracy, high precision, and high AUROC. The 18% difference between the accuracy and precision measures suggests that variables other than those used in this study may improve prediction, including contextual variables that may influence vaccine choices. Variables may include political affiliation [7], longitude and latitude [8], access to the internet [8], health literacy [54], and presence of underlying conditions [9]. Future work should seek to include these types of variables.

In the second BRF classifier, the 4 COVID-19 precaution behaviors only contributed 6.6% to the prediction. This low contribution could be due to these variables being binary, unlike the other demographic variables, which included a range of categories. In addition, COVID-19 precaution behaviors are specific to the context of the COVID-19 pandemic and do not promote interpretation beyond their specific context. The 15 judgment variables represent a contrast to this as they are empirically computed from a set of functions across many picture categories. An individual with higher risk aversion will generally tolerate higher amounts of uncertainty regarding a potential upside or gain as opposed to settling for what they have. This does not depend on what stimulus category they observe or the stimulus-response condition. Instead, it is a general feature of the bounds to their judgment and is part of what behavioral economists such as Kahneman consider as bounds to human rationality [84].

The Gini score plots were clear sigmoid-like graphs (Figure 3), with only 3 of the 7 demographic variables ranking above the judgment variables. This observation was consistent in both BRF classifiers (with and without COVID-19 precaution behaviors), raising the possibility of a statistically mechanistic relationship among the top 3 demographic variables, the 15 judgment variables, and vaccine uptake. Indeed, we observed 28 primary mediation effects and 13 secondary mediation effects in contrast to 2 moderation relationships, which also happened to overlap with mediation findings, suggesting mixed mediation-moderation relationships [81,88]. The observation that most judgment variables were significant in mediation relationships but not in moderation relationships argues that prediction depended on the directional relationship between judgment and demographic variables to predict vaccine uptake. Furthermore, there were more significant primary mediations (when judgment variables were the IVs) compared to secondary mediations, suggesting the importance of judgment variables as IVs and demographic variables as mediators. Mathematically, judgment variables (IVs) influenced vaccine uptake (DV), and this relationship was stronger when demographic variables were added to the equation. The 13 secondary mediations all overlapped with the 28 primary mediations, where demographic variables were IVs and judgment variables were mediators, suggesting that demographic variables influenced vaccine uptake (DV) and that this relationship became stronger with the addition of judgment variables. This overlap of primary and secondary mediations for 4 of the judgment variables suggests that both judgment and demographic variables influenced the choice of being vaccinated within a mixed mediation framework because adding either one of them to the mediation model regressions made the relationships stronger [49]. The lack of moderation results and a considerable number of overlapping primary and secondary mediation results imply that the relationship between judgment variables and vaccine uptake did not depend purely on their interaction with age, income, or educational level (ie, moderation) but, instead, depended on the direct effects of these 3 demographic variables to strengthen the relationship between judgment variables and vaccine uptake. This type of analysis of statistical mechanisms is helpful for understanding contextual effects on our biases and might be important for considering how best to target or message those with higher loss aversion, ante, peak positive risk, peak negative risk, total reward risk, and total aversion risk (ie, in those who were not fully vaccinated).

The developed model is automatable and may have applications in public health. The picture-rating task can be deployed on any smart device or computer, making it accessible to much of the US population or regional populations. The ratings from this task can be automatically processed, and the results can be stored in local or national databases. This method of data collection is novel in that persons cannot bias their responses as the rating task has no perceivable relation to vaccination choices. Government and public health bodies can access these data to determine predicted vaccine uptake rates locally or nationally, which can be used to (1) prepare vaccine rollouts and supply chain demand, (2) prepare health care institutions in areas that may experience low vaccine adherence and potentially higher infection rates, and (3) determine which areas may need more targeted messaging to appeal to specific judgment profiles. For use case 3, messaging about infection risks or precaution behaviors could be framed to address those with lower risk aversion, who, in this study, tended to forgo vaccination. Given that such individualized data would not be available a priori, it would be more plausible to collect data from similarly sized cohorts in geographic regions of concern to obtain regional judgment behavior profiles and, thus, target use cases 1 to 3. Further development of this model with different population samples might also improve our understanding of how certain judgment variables may be targeted with different types of messaging, offering a means to potentially improve vaccine uptake. This model might also be applied to other mandated or recommended vaccines such as those for influenza or human papillomavirus, ultimately improving preparation and messaging efforts. However, future work would be needed to model these varying vaccine choices.

Given the use of demographic variables in the proposed model, specific demographic populations could be assessed or considered for messaging. If particular demographic groups are predicted to have a low vaccine uptake rate, messaging can be targeted to those specific groups. For example, we observed that a higher percentage of female individuals were not fully vaccinated when compared to male individuals. This could be related to concerns about the COVID-19 vaccine affecting fertility or pregnancy. To improve uptake in this population, scientifically backed messaging could be used to confirm the safety of the vaccine in this context. Lower rates of vaccination have been reported in Black communities, which was also observed in this study. Researchers have identified targetable issues related to this observation, which include engagement of Black faith leaders and accessibility of vaccination clinics in Black communities, to name a few [89].

In summary, this model could be used to predict vaccine uptake at the local and national levels and further assess the demographic and judgment features that may underlie these choices.

This study has a number of limitations that should be considered. First, there are the inherent limitations of using an internet survey—namely, the uncontrolled environment in which participants provide responses. Gold Research, Inc, and the research team applied stringent exclusion criteria, including the evaluation of the judgment graphs given that random responses produce graphs with extremely low R² fits (eg, <0.1). This was not the case in our cohort of 3476 participants, but this cannot perfectly exclude random or erroneous responses to other questionnaire components. Second, participants with mental health conditions were oversampled to meet the criteria for other survey components not discussed in this paper. This oversampling could potentially bias the results, and future work should use a general population sample to verify these findings. Third, demographic variability and the resulting confounds are inherent in population surveys, and other demographic factors not collected in this study may be important for prediction (eg, religion and family size). Future work might consider collecting a broader array of demographic factors to investigate and include in predictive modeling. Fourth, we used a limited set of 7 demographic variables and 15 judgment variables; however, a larger set of judgment variables is potentially computable and could be considered for future studies. There is also little information on how post–COVID-19 effects, including socioeconomic effects, affect COVID-19 vaccination choices.

To our knowledge, there has been minimal research on how biases in human judgment might contribute to the psychology underlying individual vaccination preferences and what differentiates individuals who were fully vaccinated against COVID-19 from those who were not. This population study of several thousand participants demonstrated that a small set of demographic variables and 15 judgment variables predicted vaccine uptake with moderate to high accuracy and high precision and AUROC, although a large range of specificities was achieved depending on the classification method used. In an age of big data machine learning approaches, this study provides an option for using fewer but more interpretable variables. Age, income, and educational level were independently the most important predictors of vaccine uptake, but judgment variables collectively dominated the importance rankings and contributed almost two-thirds to the prediction of COVID-19 vaccination for the BRF and random forest models. Age, income, and educational level significantly mediated the statistical relationship between judgment variables and vaccine uptake, indicating a statistically mechanistic relationship grounding the prediction results. These findings support the hypothesis that small sets of judgment variables might provide a target for vaccine education and messaging to improve uptake. Such education and messaging might also need to consider contextual variables (ie, age, income, and educational level) that mediate the effect of judgment variables on vaccine uptake. Judgment and demographic variables can be readily collected using any digital device, including smartphones, which are accessible worldwide. Further development and use of this model could (1) improve vaccine uptake, (2) better prepare vaccine rollouts and health care institutions, (3) improve messaging efforts, and (4) have applications for other mandated or government-recommended vaccines.