Estimating the Effectiveness of Rotavirus Vaccine Schedules

Abstract

Background:

Important questions exist regarding the comparative effectiveness of alternative childhood vaccine schedules; however, optimal approaches to studying this complex issue are unclear.

Methods:

We applied methods for studying dynamic treatment regimens to estimate the comparative effectiveness of different rotavirus vaccine (RV) schedules for preventing acute gastroenteritis-related emergency department visits or hospitalization. We studied the effectiveness of six separate protocols: one- and two-dose monovalent rotavirus vaccine (RV1); one-, two-, and three-dose pentavalent rotavirus vaccine (RV5); and no RV vaccine. We used data on all infants to estimate the counterfactual cumulative risk for each protocol. Infants were censored when vaccine receipt deviated from the protocol. Inverse probability of censoring-weighted estimation addressed potentially informative censoring by protocol deviations. A non-parametric group-based bootstrap procedure provided statistical inference.

Results:

The method yielded similar 2-year effectiveness estimates for the full-series protocols; weighted risk difference estimates comparing unvaccinated children to those adherent to either full-series (two-dose RV1, three-dose RV5) corresponded to four fewer hospitalizations and 12 fewer emergency department visits over the 2-year period per 1,000 children. We observed dose–response relationships, such that additional doses further reduced risk of acute gastroenteritis. Under a theoretical intervention to fully vaccinate all children, the 2-year risk differences comparing full to observed adherence were 0.04% (95% CI: 0.03%, 0.05%) for hospitalizations and 0.17% (95% CI: 0.14%, 0.19%) for emergency department visits.

Conclusions:

The proposed approach can generate important evidence about the consequences of delaying or skipping vaccine doses, and the impact of interventions to improve vaccine schedule adherence.

Introduction

Poor adherence to recommended vaccines schedules is a growing clinical and public health problem. Although the majority of parents in the U.S. vaccinate their children according to the immunization schedule recommended by the Advisory Committee on Immunization Practices (ACIP), there is a small but growing subpopulation of vaccine-hesitant parents who refuse or delay vaccine doses due to safety concerns (1, 2). Other reasons for poor adherence to vaccine schedules include barriers to care and age restrictions (e.g., ACIP-recommended maximum age for the last dose of rotavirus vaccine is 8 months and 0 days).(3) Of the recommended vaccines in the early childhood schedule, adherence to the rotavirus vaccine schedule is particularly poor (4), despite high rotavirus-related healthcare utilization and costs.(5, 6) Yet, little is known about the real-world effectiveness or safety of these alternative childhood vaccine schedules.

Better estimates of the effects of non-adherence to multi-dose vaccine schedules are needed to help prioritize development of effective adherence improvement strategies for vaccination and to ultimately reduce vaccine-preventable diseases. For vaccine schedule research, optimal approaches to estimating the effectiveness of different schedules are not yet established (7). The feasibility of randomizing infants to various vaccine schedules is limited due to the lack of equipoise regarding the risk of infectious outcomes under no vaccination, partial schedules, and full schedules. To understand the real-world effectiveness of vaccine schedules, non-experimental studies can be useful. However, non-experimental studies of vaccine schedules are challenging due to the tremendous variability in real-world patterns of vaccine receipt, represented by millions of distinct patterns of vaccination, resulting from the high number of recommended vaccines, formulations, and timing of doses (1, 7).

One approach to studying the effects of vaccine schedules is to determine the schedule that each child followed during a defined schedule ascertainment period, and then compare the risk of outcomes that occur after the ascertainment period between children on various schedules. However, requirement of a schedule ascertainment interval is problematic for several reasons. First, it does not permit study of outcomes that occur during the interval, an important issue for studies of many childhood vaccines, such as the rotavirus vaccines, that are intended to prevent outcomes during the time period when they are administered. Also, restricting the study to children who remain under observation for the entire schedule ascertainment interval may impair generalizability, particularly in U.S. healthcare databases in which risk of disenrollment is related to many patient-level variables (8). Finally, if the schedule itself affects the probability of being observed for the full schedule ascertainment period, selection bias may result (9).

To address these issues, we propose to estimate the effectiveness of childhood vaccine schedules as defined by vaccine type and the number and timing of administered doses, using an inverse probability of censoring weighted estimation approach to account for potentially informative schedule deviations. Our method adapts an analytic approach originally developed to study protocol adherence in randomized trials (10). The general approach has been used to study the effect of dynamic treatment regimens (11-13), and to study the population attributable effect of preventing treatment discontinuation (14).

Here, we apply the approach in the context of a study of various rotavirus vaccine (RV) schedules for the prevention of acute gastroenteritis -related hospitalization and emergency department (ED) visits during early childhood. The effectiveness of full- and partial-series completion with the two-dose monovalent rotavirus vaccine, Rotarix (RV1; GlaxoSmithKline Biologicals, Rixensart, Belgium), and the three-dose pentavalent rotavirus vaccine, RotaTeq (RV5; Merck & Co., Inc., Whitehouse Station, New Jersey, USA), have been wellestablished (5, 15-19). We estimate the effect of varying levels of adherence to RV schedules on the cumulative risk of acute gastroenteritis -related hospitalization and emergency department (ED) visits during the first 2 years of life. We also estimate the population attributable effects representing the potential decrease in acute gastroenteritis-related hospitalizations and ED visits from a theoretical intervention that shifts the entire population from their observed real-world adherence to full adherence of the recommended RV schedule.

METHODS

Data source

We used data (2006-2016) from the IBM^® MarketScan^® Commercial Database, which captures patient-level data on inpatient, outpatient, and prescription drug claims from employers and health plans that insure employees and their dependents. The database contains data for several million individuals annually in the U.S. (20). This study using deidentified data was considered exempt from human subject review by the Institutional Review Board at Washington University.

Guideline Recommendations for Rotavirus Vaccination

We used ACIP recommendations for rotavirus vaccination to inform our study design. ACIP recommends routine vaccination of U.S. infants with rotavirus vaccine. ACIP does not express a preference for either RV5 or RV1. RV5 is to be administered in a three-dose series (at 2, 4, and 6 months of age). RV1 is to be administered in a two-dose series (at 2 and 4 months of age). (3) ACIP recommendations were used to define timely RV dose #1 (38 to 104 days of age), RV dose #2 (66 to 240 days of age), and dose #3 (94 to 240 days of age). Timely doses allowed a 4-day grace period for early vaccination (1) and required administration ⩾4 weeks apart (3).

Study population

We identified all U.S. infants born between 1 January 2009 and 19 November 2016 who received their first dose of diphtheria–tetanus–pertussis (DTaP) vaccine between 38 to 104 days of age (1, 21). We required at least one dose of DTaP vaccine to restrict the study population to infants who received the 2-month vaccine with the highest coverage rate (22), since vaccine recipients and non-recipients may differ with respect to unmeasured potential confounders.(23) We required continuous enrollment from birth until DTaP dose #1. Birth dates were determined by ICD-9 codes for live-born infants (V30-39). We excluded infants who received rotavirus vaccine prior to 38 days of age (i.e., the minimum ACIP-recommended age allowing a 4-day grace period); were diagnosed with intussusception or severe combined immunodeficiency prior to DTaP vaccine dose #1 (24); or resided in states with universal Vaccines for Children (VFC) programs (eTable 1) (25).

Exposure

The exposure, receipt of rotavirus vaccination dose #1, was assessed on the date of DTaP dose #1 (i.e., index date). The index date was anchored on the date of DTaP dose #1 because DTaP vaccine has the highest coverage rate among all vaccines that ACIP recommends for concomitant administration with rotavirus vaccine dose #1 (22). We allowed a 4-day grace period such that infants who received dose #1 of RV 4 days before DTaP dose #1 were classified as receiving the exposure, as long as RV was received from 38 to 104 days of age. We identified a vaccine administration history including rotavirus vaccine formulation (RV1, RV5), dose number, date, and age at administration using Current Procedural Terminology codes (eTable 2) (26). Rotavirus vaccination has been validated in commercially-insured U.S. infant populations (positive predictive value range, 87.1% to 88.5%).(27, 28)

Outcome

We identified acute gastroenteritis-related inpatient visits and ED visits using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) and ICD-10-CM codes (eTable 3). We focused on gastroenteritis-coded events since rotavirus-coded events underestimate the true burden of rotavirus disease because of lack of routine laboratory testing and coding (29). The follow-up period began the day after the index date and ended at the earliest of: the outcome of interest, protocol non-adherence, 720 days after the index date, or end of study (31 December 2016).

Covariates

We identified baseline covariates from birth until the index date. Demographic and clinical characteristics included maternal age at birth, birth month (30), birth year (30), number of siblings, health plan type, provider type, infant overnight hospitalizations, geographic region, state, and metropolitan statistical area (defined by U.S. Census Bureau as a geographical region with high population density at its core and close economic ties throughout the area) (31).

Statistical analysis

For each child, we created six copies of the longitudinal dataset to correspond to six RV protocols of interest: (P1) unvaccinated with RV1 or RV5; (P2) partial one-dose RV5 series; (P3) partial two-dose RV5 series; (P4) full three-dose RV5 series; (P5) partial one-dose RV1 series; and (P6) full two-dose RV1 series. Within each protocol-specific copy, a child who did not follow a given protocol at baseline (i.e., did not meet eligibility criteria for that protocol based on RV receipt or non-receipt on the index date) was censored at a randomly generated time between days 0 and 1. Then, within each protocol-specific copy, a child was deemed non-adherent to a protocol and subsequently censored when RV receipt or non-receipt first deviated from the protocol; detailed definitions of protocol non-adherence are presented in eTable 4. Figure 1 presents the schematic for eight illustrative children. For example, children A and B were adherent to the unvaccinated protocol (P1) because they did not receive RV by the index date. During follow-up, child A was adherent to the protocol because they did not receive RV, and was therefore administratively censored at 720 days after the index date. Alternatively, child B was censored due to protocol non-adherence upon receipt of a RV dose not specified by the protocol. Children C thru H were each adherent to two protocols - partial RV1 series (P5) and complete RV1 series (P6). During follow-up, children C and F were adherent to protocols P5 and P6, respectively, and were therefore both administratively censored 720 days after the index date. Child D was censored from P5 due to protocol non-adherence upon receipt of an early RV dose (i.e., RV1 dose #2 was administered <4 weeks after RV1 dose #1). Child E was censored from P5 due to protocol non-adherence upon receipt of a different RV formulation (i.e., RV5). Child G was censored from P6 due to protocol non-adherence upon receipt of a RV dose not specified by the protocol. Child H was censored from P6 due to protocol non-adherence upon no receipt of RV dose #2 by 241 days of age.

Figure 1.

Schematic of study design with eight illustrative children for estimation of the effect of rotavirus vaccine schedule on the prevention of acute-gastroenteritis hospitalization and emergency department visits. The children are followed over time until first occurrence of either an outcome or censoring event. The asterisks indicate censoring due to protocol non-adherence. For example, children A and B were adherent to the unvaccinated protocol (P1) because they did not receive RV by the index date. During follow-up, child A was adherent to the protocol because he/she did not receive RV, and was therefore administratively censored at 720 days after the index date. Alternatively, child B was censored due to protocol non-adherence upon receipt of a RV dose not specified by the protocol. More detailed descriptions of illustrative children and definitions of protocol non-adherence are presented in the Methods section and eTable 4.

We calculated means and frequencies of baseline covariates by index rotavirus vaccine status (i.e., RV unvaccinated, RV1 vaccinated, RV5 vaccinated). We plotted standardized mean differences of baseline covariates in the unweighted population to determine the magnitude of imbalance of observed covariates between each vaccinated group (RV1, RV5) versus the unvaccinated group (32). For each vaccine-schedule protocol g, we calculated the two-year cumulative risk functions of acute gastroenteritis-related hospitalization or ED visits had all individuals in the population followed protocol g. At 30, 90, 150, and 720 days after the index date, we calculated the risk difference for each outcome comparing the risk under each protocol to the risk under the ‘never vaccinate’ protocol. We also computed estimates of direct vaccine effectiveness (VE) as (1 – risk ratio) x 100 (33). For each outcome, we computed 95% CIs for the estimates of cumulative risk, risk difference, and VE using a non-parametric bootstrap based on 200 resamples (34, 35).

Our approach to estimating the cumulative risk under the scenario in which all children initiated a given vaccine schedule involved applying g specific inverse probability of censoring weights to each child copy g to account for informative protocol deviations (10-13, 36, 37). Additional details on this method can be found in Robins and Finkelstein (10). We estimated the cumulative incidence of the outcome using the following estimator:

$$\widehat{Pr}(T(g)<t)=\frac{1}{n}\sum _i\frac{I(T_{g,i}^{\ast }<t){\Delta }_i}{{\widehat{\mathbf{Pr}}}_g\left(\Delta =1\mid X_i,A_i,T_{g,i}^{\ast }\right)}$$

where n is the size of the study population, $T_{g,i}^{\ast }$ is the minimum of the time of the outcome (T_i) or censoring (C_i,g) for person-copy i under protocol g, Δ_i is an indicator that takes the value 1 if person-copy i is not censored, X_i is a vector of baseline covariates, and T_k(g) is the value of T_k that child k would have had if, possibly counter to fact, they had followed protocol g. We note that inverse probability of treatment weights are not needed because each vaccine schedule protocol group contains exactly the same children at baseline. Rather, children who do not follow a given protocol at baseline are censored prior to day 1 and the remaining children are weighted to represent them using the inverse probability of censoring weights.

The probability of remaining uncensored under each protocol, ${\widehat{\text{Pr}}}_g\left(\Delta =1\mid X_i,A_i,T_{g,i}^{\ast }\right)$, was estimated with a Cox proportional hazards model and the Breslow estimator for the baseline hazard function (38), with the model fit separately within levels of each vaccine schedule protocol. For the adjusted estimates, several covariates were included in the model: birth month, birth year, gender, maternal age, number of siblings, overnight hospitalization, health plan type, network of provider type, residence in a metropolitan statistical area, and geographic region.

In addition, we estimated the population attributable effect, which compared the observed risk of the outcome in the study population with the risk that would have been observed under a counterfactual exposure distribution in which the entire population received the preventive intervention (i.e., full-series of RV5 or RV1) (39, 40). Thus, we compared the two-year cumulative risk functions under two exposure distributions: 1) real-world adherence to the recommended schedule; and 2) full adherence to the recommended schedule, where the recommended schedule was defined as timely receipt of the full three-dose series of RV5 or the full two-dose series of RV1. To generate the 2-year cumulative risk function assuming real-world adherence to the RV1 or RV5 full-series, we censored individuals who did not receive a first dose of either RV or deviated from the protocol corresponding to their first RV dose (e.g., an individual who received RV1 at baseline was censored if they did not adhere to the full RV1 schedule). To generate the 2-year cumulative risk function assuming full adherence to the RV1 or RV5 full series, we repeated the same estimation approach but did not censor individuals for protocol non-adherence.

RESULTS

We identified 1,147,699 eligible infants. On the index date, 9.7% remained unvaccinated with RV, 16.0% had received RV1 dose #1, and 74.3% had received RV5 dose #1. eTable 5 presents baseline characteristics of the cohort. Compared to unvaccinated children, those vaccinated at index with RV1 or RV5 were more likely to have younger mothers, be born later in the study period, reside in the South, and have a pediatrician as primary provider, as indicated by standardized mean differences >0.10 (eFigure 1) (41).

Figure 2 illustrates the distribution and timing of infant censoring from each vaccine schedule protocol due to protocol deviation. For the unvaccinated protocol, most infants (90.3%) were censored immediately upon RV receipt by the index date. For all RV5 protocols, 25.7% of infants were censored immediately due to RV1 receipt or non-receipt of RV by the index date. For the one-dose RV5 protocol, most infants (69.0%) were censored between 30 and 90 days since index, predominantly for receipt of RV dose #2 at the 4-month well-child visit. For the two-dose RV5 protocol, more than half of infants (51.5%) were censored between 120 and 180 days since index, predominantly for receipt of RV dose #3 at the 6-month well-child visit. For the three-dose RV5 protocol, censoring rarely occurred during follow-up until 150 to 200 days since index when 7.8% of infants were censored, predominantly due to non-receipt of RV5 dose #3 by the last eligible age (i.e., 241 days of age). Patterns of censoring due to protocol deviation were similar in RV1 and RV5 protocols.

Figure 2.

Timing of deviation from vaccine schedule protocol. Within each protocol-specific copy, a child who did not follow a given protocol at baseline (i.e., did not meet eligibility criteria for that protocol based on RV receipt or non-receipt on the index date) was censored at a randomly generated time between days 0 and 1.

By design, the RV1 and RV5 protocols accrued an increasing number of events and person-time during follow-up with each additional dose required by the protocol (Tables 1 and 2). For example, the one-dose RV5 protocol accrued 686 events during 162,196 person–years; the two-dose RV5 protocol accrued 1124 events during 323,558 person–years; and the three-dose RV5 protocol accrued 2455 events during 895,946 person–years. The RV1 protocols followed a similar pattern.

Table 1.

Two-year Inverse Probability of Censoring-Weighted Effect Estimates of Acute Gastroenteritis-related Hospitalization by Protocol, United States, 2009-2016 ^a,b

Protocol		No. of events	No. of person– years at risk	Risk,% (95% CI) b	Risk Difference, % (95% CI) b	Vaccine Effectiveness (%) (95% CI) c
P1	Unvaccinated series	495	101,254	0.88 (0.79, 0.97)	0.00
P2	Partial RV5 series (1 dose)	686	162,196	0.80 (0.62, 1.02)	−0.08 (−0.30, 0.13)	9 (−14, 33)
P3	Partial RV5 series (2 doses)	1,124	323,558	0.61 (0.52, 0.71)	−0.27 (−0.40, −0.15)	31 (19, 43)
P4	Full RV5 series (3 doses)	2,455	895,946	0.47 (0.45, 0.49)	−0.40 (−0.50, −0.31)	46 (40, 52)
P5	Partial RV1 series (1 dose)	185	38,710	0.55 (0.37, 0.83)	−0.33 (−0.57, −0.08)	37 (10, 64)
P6	Full RV1 series (2 doses)	553	186,040	0.49 (0.45, 0.54)	−0.38 (−0.49, −0.28)	44 (36, 52)

^aEstimates adjusted for month of birth, year of birth, gender, maternal age, number of siblings, overnight hospitalization, health plan type, network of provider type, residence in a metropolitan statistical area, and geographic region.

^b95% confidence intervals estimated using a non-parametric bootstrap (N=200).

^cVaccine effectiveness calculated as (1 – risk ratio) x 100.

CI indicates confidence interval; RV1, rotavirus monovalent vaccine; RV5, rotavirus pentavalent vaccine.

Table 2.

Two-year Inverse Probability of Censoring-Weighted Effect Estimates of Acute Gastroenteritis-related Emergency Department Visits by Protocol, United States, 2009-2016^a,b

Protocol		No. of events	No. of person– years at risk	Risk,% (95% CI) b	Risk Difference, % (95% CI) b	Vaccine Effectiveness (%) (95% CI) c
P1	Unvaccinated series	2,416	99,818	4.36 (4.17, 4.57)	0.00
P2	Partial RV5 series (1 dose)	2,695	161,729	4.57 (4.04, 5.16)	0.20 (−0.38, 0.79)	−5 (−18, 9)
P3	Partial RV5 series (2 doses)	5,585	321,969	4.23 (3.95, 4.54)	−0.13 (−0.49, 0.23)	3 (−5, 11)
P4	Full RV5 series (3 doses)	15,297	885,993	3.15 (3.09, 3.21)	−1.22 (−1.43, −1.00)	28 (24, 32)
P5	Partial RV1 series (1 dose)	699	38,563	4.07 (3.42, 4.84)	−0.30 (−1.03, 0.43)	7 (−10, 23)
P6	Full RV1 series (2 doses)	3,235	184,002	3.15 (3.04, 3.27)	−1.21 (−1.43, −0.99)	28 (24, 32)

^aEstimates adjusted for month of birth, year of birth, gender, maternal age, number of siblings, overnight hospitalization, health plan type, network of provider type, residence in a metropolitan statistical area, and geographic region.

^b95% confidence intervals estimated using a non-parametric bootstrap (N=200).

^cVaccine effectiveness calculated as (1 – risk ratio) x 100.

CI indicates confidence interval; RV1, rotavirus monovalent vaccine; RV5, rotavirus pentavalent vaccine.

The 2-year weighted cumulative risk, cumulative risk difference, and VE estimates of acute gastroenteritis-related hospitalization or ED visits are presented in Tables 1-2 and Figure 3. Unadjusted estimates are presented in eTables 6-7. In analyses of hospitalizations, the weighted 2-year cumulative risk estimates and VE estimates were similar for children who received the full series of either vaccine formulation, suggesting similar effectiveness. Specifically, the weighted 2-year cumulative risk estimate was 0.47% (95% CI 0.45%, 0.49%) for the full 3-dose RV5 series and 0.49% (95% CI 0.45%, 0.54%) for the full two-dose RV1 series; the VE estimate was 46% (95% CI 40%, 52%) for the full three-dose RV5 series and 44% (95% CI 36%, 52%) for the full two-dose RV1 series (Table 1; Figure 3). The weighted risk difference estimates comparing unvaccinated children to those adherent to either full series were similar, corresponding to four fewer hospitalizations over the 2-year period per 1,000 children.

Figure 3.

Estimates of 2-year inverse probability of censoring-weighted cumulative risk functions of A) acute gastroenteritis-related hospitalizations and B) acute gastroenteritis-related emergency department visits, by rotavirus vaccine schedule. Black lines denote RV5 protocols. Medium grey lines denote RV1 protocols. Light gray lines denote unvaccinated protocol.

For both vaccine formulations, we observed a dose–response for each additional dose in the series wherein the reduction in risk of hospitalizations increased with each additional dose. For example, the weighted risk difference estimates comparing unvaccinated children to those who received 1, 2, or 3 doses of RV5 corresponded to 1, 3, or 4 fewer hospitalizations over the 2-year period per 1,000 children. Results for ED visits followed a similar pattern. The weighted risk difference estimates comparing unvaccinated children to those adherent to either full-series similar corresponded to 12 fewer ED visits over the 2-year period per 1,000 children (Table 2). For analyses of both hospitalizations and ED visits, the 95% CIs were widest for the partial one-dose RV1 series and the partial one-dose RV5 series, reflecting that infants were most likely to be artificially censored due to protocol deviation for these protocols.

The 30-, 90-, and 150-day weighted cumulative risk, cumulative risk difference, and VE estimates of acute gastroenteritis-related hospitalization or ED visits are presented in eTables 8-9. In analyses of emergency department visits, we did not observe differences in effectiveness by protocol at 30-, 90-, or 150 days after index date. Similarly, in analyses of hospitalizations, we did not observe differences in effectiveness by protocol at 30 days after index date. However, at 90 days after index date, two protocols (two-dose RV5 and three-dose RV5) indicated higher effectiveness versus the unvaccinated protocol, corresponding to 0.3 fewer hospitalizations per 1,000 children. Later, at 150 days after index date, three protocols indicated higher effectiveness compared to the unvaccinated protocol. Specifically, children who received two doses of RV5, three doses of RV5, and two doses of RV1 had weighted risk difference estimates that corresponded to 0.5, 0.6, and 0.5 fewer hospitalizations per 1,000 children, respectively.

Table 3 and Figure 4 present the population attributable effects representing the potential decrease in hospitalizations and ED visits from a theoretical intervention that shifts the entire population from their observed real-world adherence to full adherence of the recommended schedule (e.g., individuals on partial 1-dose RV1 series shift to full 2-dose RV1 series). Unadjusted estimates are presented in eTable 10. Under real-world adherence, 23,821 children experienced an ED visit during 1,309,279 person–years of follow-up. Under a theoretical intervention to fully vaccinate all children, the 2-year risk of ED visits was 3.15%, yielding a risk difference comparing full to observed adherence of 0.17% (95% CI: 0.14%, 0.19%). Under the same theoretical intervention, the 2-year risk of hospitalizations was 0.48%, yielding a risk difference comparing full to observed adherence of 0.04% (95% CI: 0.03%, 0.05%).

Table 3.

Two-year Inverse Probability of Censoring-Weighted Effect Estimates of Acute Gastroenteritis-related Emergency Department Visits and Hospitalizations Under Different Exposure Distributions, United States, 2009-2016

Exposure distribution a	No. of events	No. of person– years at risk	Risk,% (95% CI) b, c	Risk Difference, % (95% CI) b,c
Hospitalizations
Real-world adherence to recommended schedule	3,963	1,325,252	0.52 (0.50, 0.54)	0.04 (0.03, 0.05)
Full adherence to recommended schedule	3,008	1,081,828	0.48 (0.46, 0.50)	0.00
Emergency Department Visits
Real-world adherence to recommended schedule	23,821	1,309,279	3.32 (3.27, 3.36)	0.17 (0.14, 0.19)
Full adherence to recommended schedule	18,532	1,069,837	3.15 (3.10, 3.20)	0.00

^aRecommended schedule defined as full three-dose RV5 series or full two-dose RV1 series.

^bEstimates adjusted for month of birth, year of birth, gender, maternal age, number of siblings, overnight hospitalization, health plan type, network of provider type, residence in a metropolitan statistical area, and geographic region.

^c95% confidence intervals were estimated using a non-parametric bootstrap (N=200).

CI indicates confidence interval; RV1, rotavirus monovalent vaccine; RV5, rotavirus pentavalent vaccine.

Figure 4.

Estimates of 2-year cumulative risk functions of A) acute gastroenteritis-related hospitalizations and B) acute gastroenteritis-related emergency department visits under different exposure distributions of real-world adherence versus full adherence. The 2-year cumulative risk function assuming full- and real-world adherence to the RV1 or RV5 full-series were generated by censoring and not censoring individuals for protocol non-adherence, respectively. Light gray lines denote full adherence. Dark gray lines denote real-world adherence.

DISCUSSION

In a national cohort of commercially insured children, we estimated the effects of multi-dose rotavirus vaccine schedules. We found similar 2-year effect estimates for the prevention of acute gastroenteritis-related hospitalization or ED visits for the full two-dose series of RV1 and the full three-dose series of RV5. VE estimates were low (maximum VE = 46%), likely due to non-specificity of the outcome (i.e., AGE rather than rotavirus). Our results also revealed a dose–response relationship wherein receipt of each additional dose further reduced risk of AGE-related hospitalization or ED visits. In addition, we estimated the population attributable effects of a theoretical intervention that shifted the entire population from their observed real-world adherence to full adherence of the preventive intervention (i.e., full series of RV5 or RV1). Use of this approach quantified the additional number of ED visits and hospitalizations that could be averted with an intervention to achieve optimal vaccine schedule adherence.

We have outlined an approach for estimating the effects of multi-dose vaccine schedules. The approach that we employed uses inverse probability of censoring weights to account for potentially informative protocol deviations, and is motivated by previous work describing methods to estimate the effects of dynamic treatment regimes (11). Our approach is suitable when at least one of the comparison groups is defined by receipt of multiple vaccine doses at different points in time.

Our findings are subject to several limitations. First, our analysis may be subject to uncontrolled confounding. The unadjusted and weighted estimates were similar, indicating little confounding or selection bias by measured factors. However, our analysis only accounted for baseline covariates, thus it is possible that unmeasured time-varying confounders may bias the results. For example, rotavirus vaccination is universally recommended in the U.S. for all infants except those with very rare contraindications (i.e., life-threatening allergic reaction to any component of rotavirus vaccine, intussusception, severe combined immunodeficiency syndrome) (24). We excluded infants with contraindications during the baseline period, but did not account for development of contraindications during follow-up, which could result in bias if these patients have differential risk of the outcome. Other potential time-varying covariates may include use of antibiotics or development of other infections during follow-up. Instead of controlling for such factors in a censoring model, an alternative approach would be to explicitly incorporate time-varying covariates into the protocols, wherein children who develop contraindications would not be censored for non-receipt of timely doses, because non-receipt would be medically indicated. In a similar study on the effect of preventable treatment discontinuation, medically indicated treatment discontinuation was not considered “preventable.”(14) Since contraindications for rotavirus vaccination are very rare,(42, 43) the bias resulting from occurrence of these events during follow-up is likely small.

Second, our study was conducted under the assumption of no interference (i.e., the potential outcomes of one individual are assumed to be unaffected by the exposure of other individuals). This assumption does not hold because the risk of infection in one person depends on the distribution of vaccination in other members of the population (44-46). However, recent simulation results have shown that the bias due to interference may be small under even moderate levels of interference (47).

Third, the consistency of our results relies on correct model specification. We specified a single semiparametric model for the probability of censoring due to not following a protocol at baseline and for not adhering to a protocol during follow-up. We assume the same set of covariates are sufficient to account for informative censoring due to either reason; therefore, this approach will remain valid if the composite censoring model is correctly specified. However, the mechanisms underlying these sources of censoring likely differ, thus separate models may be warranted for each source of censoring. Nevertheless, the similarity of our adjusted and unadjusted results suggests that the resulting bias may be small.

Despite these limitations, our approach of using inverse probability of censoring weights to account for informative protocol deviations overcomes limitations of previous non-experimental studies that implement traditional epidemiologic approaches to study vaccine schedules. Specifically, our approach precludes the need for a schedule ascertainment window, which enables inclusion of all eligible infants at their 2-month visit, even if they are censored or experience an outcome prior to completing the planned schedule. By not conditioning on a pre-specified length of follow-up, our study achieves more appropriate confounding control and avoids problems with selection bias and generalizability.

In summary, we have applied an approach that may prove useful for studying the effectiveness and safety of different vaccine schedules using observational data. We were able to estimate the daily reduction in risk of healthcare utilization outcomes due to each additional dose in the RV schedule series. The approach demonstrates a method to evaluate the effects of delaying or skipping vaccine doses on health-related outcomes, and the possible impact of interventions to improve recommended vaccine schedule adherence. Application of this method can be used to generate much-needed evidence to inform decisions regarding childhood vaccine schedules for a wide range of antigens, a topic of substantial clinical and policy importance.