We conducted an online cross-sectional survey between 11th to 30th of August 2022. The survey questionnaire was built using Redcap (Supplementary Table S1). Penn State Institutional Review Board approved the study protocol. Please review Supplementary File S2 for the details about pilot study and preparatory phase.

We used an NIH-funded non-profit web-based portal “ResearchMatch” to recruit survey participants (26). ResearchMatch offers access to adult volunteers of all diversities, including ethnicity, age, race, sex, and US states. The survey inclusion criteria were the age of 18 years or above and US residency. This was an unpaid survey. Implied consents were obtained at the beginning of the survey, and the responses were stored in the Redcap database.

We collected data on the participants' socio-demographic characteristics, such as age, race, sex, level of education, family income, and occupation (Table 1). Survey participants were characterized based on their age (five groups in ascending order), race (eight categories), sex (three categories), level of education (five levels from “high school” to “PhD.” in ascending order), family income (four categories in ascending order), and healthcare worker (yes vs. no). The participants' residence states were also regrouped into ten US Health and Human Services (HHS) regions (27). We asked whether the study subjects have children aged between 5 and 18 years in their families. Children were stratified into three subgroups based on their age: 5–8 years, 9–12 years, and 13–18 years. We also documented the COVID-19 and Flu vaccination status (Yes vs. No) of survey participants and family children. Finally, post-vaccination side effects experienced by the participants and their children were classified and quantified on a 0–10 severity Likert scale (Table 2). SPSS 28 and SAS 9.4 were used for data analysis.

Our data source was the survey responses captured in RedCap. Most of the answers were either in “Yes” vs. “No” dichotomous variables (for example, COVID-19 vaccine received, healthcare worker), categorical variables (for example, race and US state of residence), or on various Likert scales. We used χ² tests to determine the difference in vaccine acceptance rates among the children of various age groups.

The selection of the participants was a concern in terms of selection bias, since we could only access a pre-defined cohort. “Researchmatch” offers access and the opportunity to recruit study participants across the country, which was not necessarily a representative of US population demographics (Supplementary File S2). Nevertheless, the study participants' relatively uniform distribution across the country perhaps mitigated selection bias to some extent.

Since the percentage of COVID-19 vaccine acceptance among participants was unknown, we assumed that half of the study subjects would be pro-vaccination. Considering the adult US population of 332,403,650 in early 2023 (28), a minimum of 1,068 survey participants would be required to achieve a confidence level of 95% with a margin of error of 3%, with a pre-defined acceptance rate of 50% (29).

Pediatricians' COVID-19 vaccine recommendation: We asked about pediatricians' opinions and recommendations encouraging COVID-19 vaccination. The responses were characterized into five categories: strong recommendation, recommendation, neutral opinion, recommended against, and no discussion, respectively. We assigned scores of “5′ and “4” on a “1–5” Likert scale as strong and favorable recommendations, respectively. A score of “3” was assigned for no conversation and “2” for neutral opinion, considering pediatricians' neutral opinion could be perceived as less encouraging to the families than having no conversation. Finally, a score of “1” was designated if the pediatrician advised against the COVID-19 vaccine.

Internal consistency among the responses indicating study participants' decision to vaccinate family children of different age groups (“No” vs. “Yes” responses were captured numerically as 0 and 1), and pediatricians' recommendation (1–5 scale) to vaccinate children of three different age groups were estimated using Cronbach's alpha.

We used two methods to estimate the influence of various independent variables (socio-demographic and personal characteristics, including pediatricians' opinions) on the study participants' decision to vaccinate the children in the family (Tables 3, 4). Several respondents had more than one child in the family and had diverse opinions about vaccination for different age groups. Thus, the respondents were categorized into positive, negative, and mixed vaccine acceptance categories.

i) Multinomial logistic regression model: We used logistic regression since the dependent variable (participants' vaccine acceptance) was categorically distributed (Tables 3, 4), while independent variables were comprised of both categorical and continuous variables. The strength of the model or goodness of fit was estimated by the likelihood ratio test (χ², significance) and McFadden's pseudo R². Most influential predictors were identified based on the value of “−2 Log Likelihood of Reduced Model”, or an estimation of how much the model would change if that variable was removed (30). It is also defined as model fitting criteria.

We used the χ² test to compare vaccine acceptance among children of different age groups (Table 5). The subjects were divided into “vaccine-compliant” vs. “vaccine-hesitant” categories based on the COVID-19 vaccination status of their children. Participants with vaccine acceptance for any of the family children were considered vaccine-compliant. The χ² tests were conducted on the aggregated data and compared the subgroups.

All the participants were asked about their concerns about immunizing under-five children, irrespective of having children in the family, and the concerns or deterrents were classified into five categories such as “inadequate information about vaccine safety”, “vaccine's efficacy in disease prevention”, “children are not at risk”, “cardiac side effects”, and “vaccination is unnecessary since the child had COVID-19′.

We received 2,771 survey responses. Among them, 2,622 participants completed the survey and were included in the study analysis. A considerable number of respondents (44.1%) had one or more children aged five years or above. Survey participants were often mothers (54.23%) or grandparents (29.26%), and less frequently fathers (11.46%) of the child. Participants were predominantly aged between 24 and 44 years (34.9%), white (84.5%), female (75.9%), holding a graduate (28%) or master's degree (26.4%), financially belonging to the middle class (56.2%), non-healthcare worker (76.6%), and residing in HHS Region 4 Atlanta (25.1%) or Region 5 Chicago (22.2%) (Table 1). The comparison between the demographics of study participants and US national trend is described in details in Supplementary File S2 (data sheet 4).

More than 70% of the participants indicated that the pediatrician discussed vaccinating children against COVID-19 and had an affirmative opinion (score of 4 or 5 on 1–5 Likert scale). Pediatricians’ recommendation score (mean ± SD) for the youngest age group (5–8 years) was 3.06 ± 1.90, lower compared to the score for 9–12 years (3.35 ± 1.83) and 13–18 years (3.32 ± 1.82) age groups.

Survey participants' decision to vaccinate multiple family children of different age groups demonstrated significant internal consistency (Cronbach's alpha = 0.894). Likewise, individual pediatricians' recommendations were consistent for children of various age groups who belonged to the same caregiver (Cronbach's alpha = 0.964).

Logistic regression analysis demonstrated that the model was significant vs. the null (likelihood-ratio χ²= 514.57, p < .001, pseudo R²= .440). Both the models (logistic regression and MLPNN) identified pediatricians' recommendations and participants' COVID-19 vaccination status, closely followed by the side effect experienced, as the most influential predictors of vaccine acceptance for children (Table 3). Several other predictors also showed statistically significant impact, (Table 3). Parameter estimates are displayed in Table 4. MLPNN model also demonstrated significant prediction power with correct prediction rates of 82.9% and 81.9% for the training and testing models, respectively. (Details of input, hidden, and output layers of MLPNN are demonstrated in Supplementary File S3).

About 28% of the adults (survey participants), compared to 15% of children, reported moderate (3–6 on a “1–10” scale) or severe side effects (7–10 on “0–10” severity scale) following COVID-19 vaccination. Both adults and children reported fever as the most common side effect, followed by local pain/irritation and gastrointestinal upset. However, children less frequently experienced side effects than adults (Table 2).

Most of the study subjects received the COVID-19 vaccine (87.7%) and the influenza vaccine (82.7%), while COVID-19 and flu vaccine coverages for family children were 71.01% and 75.27%, respectively. COVID-19 vaccine acceptance was significantly lower (χ² test: p-values <0.05) for children aged 5–8 years (63.40%) than in the older age groups (69.64% and 79.67% for 9–12 years and 13–18 years old cohort, respectively) (Table 5).

We asked about potential concerns about vaccinating under-five children. Half of the participants (49.6%) expressed concern about inadequate information about vaccine safety, while 18.2% had concerns about the vaccine's efficacy in disease prevention. Lesser number of participants chose other reasons such as “children are not at risk” (15.79%), potential cardiovascular side effects of vaccine (12.51%), and children already had COVID-19 (7.74%) as a deterrent against vaccinating under-five children. In addition, one-third of the families (34.6%) with multiple adult members disagreed with themselves about their decision to vaccinate children.