The NIS-ACM uses the random digit dialed cellular telephone sample of the National Immunization Survey (NIS) to survey adults in the United States, stratified by state and selected local areas and U.S.-affiliated jurisdictions [13]. The NIS is used to monitor vaccination coverage, including routine child and teen vaccinations, as well as influenza and COVID-19 vaccinations. The NIS-ACM collects demographic and immunization data on all eligible persons ≥18 years. This paper focuses on modeling COVID-19 vaccine uptake propensity of US adults aged ≥18 years using 12 months of NIS-ACM data collected from April 22, 2021, to April 30, 2022. NIS-ACM data are split into approximate monthly data files for weighting and analytic use. Monthly sample sizes for the NIS-ACM ranged from 39,316 to 79,189 with an average of 65,536 per month. Survey weights were also calibrated by age and sex to state-level vaccine administration data reported to CDC by jurisdictions as of the middle of the monthly data collection period [14]. Response rates for the monthly datasets ranged from 17.2 % to 22.1 %.

This activity was reviewed from an ethics perspective and approved by CDC, deemed not research, and was conducted consistent with applicable federal law, ethics, and CDC policy (45C.F.R. part 46.102(l) (2), 21C.F.R. part 56; 42 U.S.C. §241(d); 5 U.S.C. §552a; 44 U.S.C. §3501). The activities that generated the NIS data set on which these secondary analyses were performed had been approved by the Centers for Disease Control and Prevention and the National Opinion Research Center at the University of Chicago Institutional Review Board. The NIS-ACM respondents provide verbal consent to participate. Personal identifiable information is not collected from the survey participants.

Using the R regression tree package rpms [15], we modeled COVID-19 vaccine initiation (receipt of first dose of the two-dose Pfizer-BioNTech or Moderna COVID-19 vaccine series or a single dose of the Janssen COVID-19 vaccine (Johnson & Johnson)), primary series completion (completing two-dose Pfizer-BioNTech or Moderna COVID-19 vaccine series or having received the single dose of the Janssen COVID-19 vaccine), and first monovalent booster dose completion propensities as a linear effect of time in months using the NIS-ACM monthly data. We built three regression trees using rpms: 1) modeling the cumulative propensity of COVID-19 vaccine initiation and the linear relationship over time of vaccine uptake controlling for demographic and geographic characteristics during May 2021 (time = 0) through April 2022 (time = 11); 2) modeling both the linear relationship over time controlling for demographic and geographic characteristics and the propensity of primary series completion from May 2021 (time = 0) through April 2022 (time = 11); and 3) modeling both the linear relationship over time controlling for demographic and geographic characteristics and the propensity of first booster dose completion (from October 2021 (time = 0) through April 2022 (time = 6).

The following demographic and geographic variables were controlled for in all models: age, sex, race/ethnicity, income/poverty, reporting having received a non-COVID 19 vaccine in the past 2 years, health insurance status, education, essential worker status, Department of Health and Human Services region, receiving a health care worker recommendation for a vaccine, work/school requirement, metropolitan statistical area (MSA), reporting previously having had COVID-19, and SVI of county of residence.

All input variables were categorical variables. Missingness (e.g., not reported) within categorical variables can be treated as valid values in the model using the rpms package, which minimizes the loss of observations due to data incompleteness. The first tree model was built based on all 786,429 observations (unique persons) that had no missing value on first dose of the COVID-19 vaccine. The second tree model was built based on the 669,455 persons (of the 786,429 who initiated receiving a COVID-19 vaccine) with complete data on completion of the primary series of COVID-19 vaccine. The third tree model was built based on 365,504 out of 443,743 observations (who had completed the primary series) with complete data on first booster completion, beginning in October 2021. Table 1 presents the detailed monthly sample size and breakdown by COVID-19 vaccine initiation, primary series completion, and first booster dose completion.

All trees were constructed using a p-value cutoff of 0.0001 and a minimum bin size (end node size) equivalent to 10 % [16] of the dose completion dataset; COVID vaccine initiation model minimum bin size = 78,643; primary series completion model minimum bin size = 66,946; and first booster dose model minimum bin size = 36,550. The rpms algorithm uses a permutation test described in Toth (2020) to determine if each split is significant. Once the algorithm fails to find a significant split given the p-value cutoff and minimum bin size, the tree stops splitting. The subgroups identified at each split are referred to as regression tree nodes. The mutually exclusive subgroups identified at the bottom of the tree are referred to as end nodes or leaves. The first node includes the entire data set before splitting and is referred to as the root node and is labeled Node 1. The first left child node of the root node is labeled Node 2 and the right child node of the root node is labeled Node 3. This leads to every node getting a unique node label.

In the first of the two regression tree models described above, shown in Fig. 1 below, the linear effect of time on COVID-19 vaccine initiation propensity is presented. The second regression tree, shown in Fig. 2 below, presents the linear effect of time on COVID-19 primary series completion propensity, and the third regression tree, shown in Fig. 3, presents the linear effect of time on COVID-19 first booster dose propensity. All three regression tree models control for demographic and geographic characteristics. The y-intercept (β₀) represents the average reported vaccination rate for respondents at time zero, and the slope (β₁) is the effect on reported vaccination rate for every additional month. Table 2 presents an overview of the demographic and geographic characteristics included in the model, and the significant demographic and geographic characteristic relationships in Models One, Two, and Three.

The first model indicated that persons who reported having received non-COVID-19 vaccines in the past 2 years (see Fig. 1 first split to the right, Nodes 6 & 7, the cumulative intercept β₀ = 0.81) were more likely to report having received a COVID-19 first dose (by about 31 percentage points), than those not having received a non-COVID-19 vaccine in the past 2 years or those who did not report their vaccination status (see Fig. 1 first split to the left, Nodes 5, 8, & 9, the cumulative intercept β₀ = 0.50). Persons who reported having received non-COVID-19 vaccines in the past 2 years aged ≥40 years or persons not reporting age were most likely to report having received aCOVID-19 first doseinitially with a one percentage point increase over time (see Fig. 1, Node 7, β₁═0.01).; and those aged 18–39 years saw a steady increase of about 3 percentage points every month (β₁═0.03) over 11 months, taking them from an initial first dose coverage of 65 % (β₀ = 0.65) to 98 % (see Fig. 1, Node 6).

For those not having received a non-COVID-19 vaccine in the past 2 years (or not reporting their vaccination status), those reporting not having a school or work requirement were less likely to report being vaccinated, especially those aged 18–49 years (or not reporting age) (see Fig. 1, Node 9). Those aged ≥50 years not having received a non-COVID-19 vaccine in the past two years (or not reporting their vaccination status) and reporting not having a school or work requirement had a higher first dose COVID-19 vaccine initiation propensity than those aged 18–49 years (or those not reporting age) (see Fig. 1, Node 8 & 9, β₀). Those not reporting the vaccination status or reporting not having received a non-COVID-19 vaccine in the past 2 years and reporting a work or school requirement (or not reporting whether they had a work/school requirement) saw first dose COVID-19 vaccine initiation propensities and increases over time like those for adults aged18 to 39 years who reported having received non-COVID-19 vaccines in the past 2 years (see Fig. 1, Nodes 5 & 6, β₀ &β₁).

Due to the recommended timing lag (at least 21 days between first and second primary dose of Pfizer COVID-19 vaccine and 28 days between first and second primary dose of Moderna COVID-19 vaccine) between first dose and primary series completion, model intercepts (β₀) in the second model have negative initial starting values for some groups, since both the first and second model use the same initial start month of April 2021 (Fig. 2). The second model indicated that persons who reported having had COVID-19 or did not report whether they previously had COVID-19 were initially more likely to report having completed the COVID-19 primary series (highest initial percentage of reported series completion β₀ = 0.03), but were slightly less likely over time to report having completed the COVID-19 primary series (lowest increasing percentage point β₁ = 0.10), compared with persons who reported not having had COVID-19.

Among those reporting not having previously had COVID-19, having received non-COVID-19 vaccines (or declining to report), and having no school or work COVID-19 vaccination requirement (or declining to report), those who reported having income level greater than $75,000 (or not reporting income) were the least likely to complete the COVID-19 primary series initially (lowest initial percentage β₀ = −0.07), but over time had the steepest positive trend for completing the COVID-19 primary series (highest increasing percentage point β₁ = 0.12), compared with other persons.

Among those who reported never having had COVID-19 and having received non-COVID-19 vaccines (or declining to report), those who reported having had a work or school requirement were more likely to report having completed the primary dose initially (β₀ = 0.01), compared with all other groups who never had COVID-19.

The third model indicated that older persons had higher first booster doses initially and higher uptake over time, with those aged ≥50 years having the highest initial first dose propensity and a nine-percentage point increase in propensity every month (See Fig. 3, Node 5, β₁ = 0.09).

Those aged ≥50 years (or not reporting age) who reported having received non-COVID-19 vaccines in the past 2 years were more likely to report having received a COVID-19 booster dose initially (by about 22 percentage points), compared with those aged 18–49 years (see Fig. 3, Nodes 6 & 7, β₀). Those who did not report or reported not having received non-COVID-19 vaccines in the past 2 years_, ages ≥50 years (or not reporting age) were more likely to report having received a COVID-19 booster dose initially (by about 11 percentage points), compared with those aged 18–49 years (see Fig. 3, Nodes 4 & 5, β₀).

In our analyses, persons reporting having received a non-COVID-19 vaccine in the past 2 years were more likely to initiate receiving a COVID-19 vaccine than those not reporting having received a non-COVID 19 vaccine. This is consistent with findings from other surveys on adult vaccination assessment that have indicated that adults who obtained other recommended vaccines, e.g., human papillomavirus (HPV) or annual influenza vaccine, were more likely to obtain COVID-19 vaccination [17–19]. As has been reported in other settings in which making employment contingent upon vaccination against influenza has resulted in dramatic uptake of influenza vaccination, we found that persons not reporting having received non-COVID-19 vaccines in the past 2 years but reporting having a work or school vaccination requirement were more likely to initiate receiving a COVID-19 vaccine than those without a work/school requirement [20–21]. Similarly, among persons not reporting having received non-COVID-19 vaccines in the past 2 years and not reporting having a work or school vaccination requirement, those who were aged ≥50 years were more likely to initiate receiving a COVID-19 vaccine than younger adults. This finding reinforced previous findings in the literature that identified older age as a common predictor of COVID-19 vaccine uptake [22–23].

Persons aged ≥50 years were most likely to get a first booster dose. This was consistent with CDC recommendations at that time that persons who received their primary series who were aged ≥65 and persons who resided in long term care settings should receive a COVID-19 booster ≥6 months after completing their primary vaccination series [24]. Persons aged 18–49 years reporting having received non-COVID-19 vaccines in the past 2 years were more likely to get a first booster dose than those not reporting having received a non-COVID-19 vaccine. Among persons who reported having received non-COVID-19 vaccine in the past 2 years, those who were aged ≥50 years were more likely to get a first booster dose than those aged 18–49 years.

Overall, the models demonstrated that: 1) older adults were more likely to initiate receiving a COVID-19 vaccine, complete their primary series (among those not previously having had COVID-19 and not having received any non-COVID-19 vaccinations in the previous 2 years), and get a first booster dose than younger persons; 2) persons reporting having received a non-COVID-19 vaccine in the past 2 years were more likely to initiate receiving a COVID-19 vaccine, complete their primary series dose (among those not reporting previously having COVID-19), and get a first booster dose than those not reporting having received a non-COVID-19 vaccine in the past 2 years; and 3) persons reporting having work or school COVID-19 vaccination requirements (or refusing to report work or school requirements) were more likely to initiate receiving a COVID-19 vaccine (conditional on not receiving or not reporting having received a non-COVID-19 vaccinations in the previous 2 years) and complete their primary series dose, given reporting not having previously had COVID-19 and reporting having had a previous non-COVID-19 vaccination in the previous 2 years (or refusing to report), than those reporting no work or school requirements.

From a methodological perspective, the choice of the R regression tree package rpms was made for its advantages as an approach to modeling the COVID-19 vaccine initiation, primary dose completion, and first booster dose propensities as a linear effect of time in months using the NIS-ACM monthly data. While we could have used logistic regression to model vaccination propensities, regression trees are able to handle missing data automatically and are able to identify moderating and interaction effects and can be used to easily describe subgroups with higher or lower propensities [10,11,24–29]. Unlike most prior studies using logistic regression models, using rpms enabled the examination of trends over time in the relationships between demographic and geographic factors and vaccination propensities, e.g., among persons having received a recent non COVID-19 vaccination, those aged 18 to 39 years were less likely to initiate receiving a COVID-19 vaccine compared with persons 40 years or older in April 2021 (the beginning of the study period), but over time had a steeper positive trend for vaccine initiation; similarly, persons with income >$75,000 were least likely to complete the COVID-19 primary series at the beginning of the study period, but over time had the steepest positive trend for completion, compared with persons in other income categories.

In sum, the rpms package can be used to identify key group differences both in terms of means and in terms of intercepts and slopes when a linear term (e.g., time) is added to the model. The rpms regression tree package recursively partitions the data, fitting a specified linear model on each node separately. The rpms recursive partitioning algorithm splits variables and selects splits using a randomized permutation test procedure that provides unbiased variable selection, adjusts for complex sample design, and produces design-consistent coefficients [30].

Five limitations might be considered when interpreting these findings. First, NIS-ACM has a low response rate (<23 %). As noted in NIS-ACM reports elsewhere, although survey weights were calibrated to COVID-19 vaccine administration data to mitigate at least some of the possible bias from incomplete sample frame (for example, sampling did not cover the landline-only and phoneless populations), nonresponse, and misclassification of vaccination status, bias estimates might remain after weighting [31]. Second, COVID-19 vaccination was self-reported and might be subject to recall or social desirability bias. Third, the NIS-ACM sampled noninstitutionalized adults via cellular telephone; therefore, adults who were nursing home residents or were incarcerated might not have been represented. Fourth, because these findings rely on self-reported data, survey estimates of COVID-19 vaccination coverage might differ from reported vaccine administration data and other data reported at https://covid.cdc.gov/covid-data. Fifth, the findings of this analysis are based on predictive modeling, and do not take into account the complex causal pathways that lead to vaccination decisions by individuals. All the categories of the variables or options for each survey items included in the model were predetermined by the NIS-ACM survey. In addition, due to potential multicollinearity, there may have been a common set of demographic factors (such as race, education level, geographic location/residency) associated with receipt of COVID-19 in our data that were not identified as significant factors in the current tree model [32–33]. In the supplemental tables A-1 to A-3, we have displayed details on the nodes/splits of each tree model by all factors included in the present study.