Adult

The findings and conclusions in report are those of the authors and do not necessarily represent the official position of the Disease Control and Prevention. ABSTRACT We estimated vaccine effectiveness for prevention of influenza-associated hospitalizations among adults during the 2018-2019 influenza season. Adults admitted with acute respiratory illness to 14 hospitals of the US Hospitalized Adult Influenza Vaccine Effectiveness Network and testing positive for influenza were cases; patients testing negative were controls. Vaccine effectiveness was estimated using logistic regression and inverse probability of treatment weighting. We analyzed data from 2863 patients with mean age of 63 years. Adjusted VE against influenza A(H1N1)pdm09-associated hospitalization was 51% (95%CI 25, 68). Adjusted VE against influenza A(H3N2) virus-associated hospitalization was −2% (95%CI −65, 37) and differed significantly by age, with VE of −130% (95% CI −374, −27) among adults 18 to ≤56 years of age. Although vaccination halved the risk of influenza-A(H1N1)pdm09-associated hospitalizations, it conferred no protection against influenza A(H3N2)-associated hospitalizations. We observed negative VE for young-and middle-aged adults but cannot exclude residual confounding as a potential explanation. Here, we report final VE estimates for prevention of influenza-associated hospitalization among adults for the HAIVEN study during the 2018-2019 US influenza season, with a focus on reducing potential bias in VE estimates by improving control of confounding arising from baseline differences in vaccinated and unvaccinated study participants. for additional respiratory viruses, including human coronavirus coronavirus OC43, coronavirus NL63, coronavirus 229E, parainfluenza viruses 1–4, human metapneumovirus, enterovirus, and adenovirus. Estimated VE differed significantly by age group, and we observed a statistically significant VE of −130% against influenza A(H3N2) among the youngest age group (18 to ≤56 years); VE was most markedly negative for participants between the ages of 18 and 40 years. Our findings are broadly similar to those reported from studies conducted in Canada and Europe[2, 3, 4], which reported negative VE against influenza A(H3N2) of genetic group 3C.3a. A potential explanation for this finding is provided by Skowronski et al.’s hypothesis that childhood priming following the 1968 influenza A(H3N2) pandemic provided immunity among imprinted individuals, protecting them as adults from subgroup 3C.3a viruses that were similar in 2018-2019 to the imprinting childhood viruses, and that receipt of antigenically-mismatched vaccine negatively interfered with this immunity leaving vaccinees at greater risk of 3C.3a infection[2]. However, the immunologic processes underlying this interference are unknown, and a limitation of our study is the lack of serologic markers with which to examine correlation of antibody titers with disease risk or protection.


INTRODUCTION
The 2018-2019 US influenza season was characterized by extensive circulation of influenza A viruses, with influenza A(H1N1)pdm09 virus circulation beginning in October, followed by a wave of influenza A(H3N2) virus circulation from February to May. Influenza A(H1N1)pdm09 viruses were well-matched to the vaccine virus strain, but the majority of influenza A(H3N2) viruses were of the 3C.3a genetic group and antigenically distinct from the genetic group 3C.2a1 vaccine virus strain.
Influenza vaccination was moderately protective against outpatient influenza illness caused by influenza A(H1N1)pdm09 viruses, with estimated vaccine effectiveness (VE) of 44%, but offered no protection against the predominant drifted influenza A(H3N2) viruses [1]. Several studies reported negative VE against influenza A(H3N2) infections among young-to middle-aged adults during the 2018-2019 Northern hemisphere influenza season [2,3,4], including preliminary findings of the U.S.
It is important to understand vaccine protection against serious outcomes such as hospitalizations in addition to protection offered against milder, outpatient illnesses. The observational test-negative design is the standard design for studies of influenza VE and estimates VE by comparing odds of vaccination among patients with acute respiratory illness (ARI) who test positive for influenza with those who test negative for influenza [7,8]. However, interpreting VE estimates for prevention of influenza hospitalization from this non-randomized design is complicated by the complexity of patients typically enrolled in hospital-based studies, who often have comorbidities and other features that give rise to systematic differences between vaccinated patients and the unvaccinated comparison group [9,10]. Additionally, the use of controls with non-influenza ARI may lead to confounding of VE if these patients are more likely to be vaccinated due to their chronic conditions [11][12][13]; hence, adequate adjustment and control for these characteristics is particularly important for estimation of VE in the inpatient setting.
Downloaded from https://academic.oup.com/jid/advance-article/doi/10.1093/infdis/jiaa772/6041424 by guest on 21 December 2020 M a n u s c r i p t Here, we report final VE estimates for prevention of influenza-associated hospitalization among adults for the HAIVEN study during the 2018-2019 US influenza season, with a focus on reducing potential bias in VE estimates by improving control of confounding arising from baseline differences in vaccinated and unvaccinated study participants.

METHODS
Participants. Participants were adults (aged ≥18 years) hospitalized for ARI presenting with new or worsening cough or sputum production of ≤10 days' duration at 14 hospitals comprising the HAIVEN study during the 2018-2019 influenza season. Eligibility criteria, enrollment procedures, and ARI definitions have been described previously [9]. Briefly, adults admitted to the hospital with ARI were identified by daily review of electronic medical records (EMRs). Eligible participants or surrogates gave written consent, and institutional review boards at participating hospitals and CDC approved study procedures. Participants or proxies were interviewed to collect information about demographics, influenza vaccination, illness characteristics and subjective assessment of frailty.
Diagnosis codes from the index hospitalization and medical encounters in the year prior to enrollment were extracted from EMRs and used to calculate the Charlson Comorbidity Index [14] and identify presence of specific comorbidities known to increase risk of serious influenza complications ("high risk conditions")*15+. Clinical outcomes were extracted from EMRs. Influenza vaccine receipt was documented using EMRs, state immunizations registries, or plausible patient self-report from the enrollment interview (self-report was considered plausible if timing and location of vaccination were provided).
Specimen collection and laboratory methods. Nasal and throat specimens (at 7 hospitals) or nasopharyngeal specimens (at 7 hospitals) were obtained from patients and tested for influenza and respiratory syncytial virus (RSV) by molecular assays. A subset of specimens for which respiratory Statistical methods. Patients who tested positive for influenza were cases and patients who tested negative for influenza were controls. Characteristics of cases and controls were compared using absolute standardized mean difference (SMD) between groups [16] and conventional tests of differences using χ 2 tests or Fisher exact tests for categorical variables and t-test or Wilcoxon ranksum test for continuous variables. Characteristics of vaccinated and unvaccinated participants were compared similarly. For subgroup analyses, we defined age groups by tertiles of age among the sample overall (18 to ≤56, 56 to ≤69, >69 years).
Patients were defined as vaccinated if they had documented or plausible self-report of vaccination ≥14 days before illness onset; we excluded individuals who received the vaccine 1-13 days before illness onset or who could not report either location and approximate timing of vaccination and had no documented vaccination. Frailty score was defined as the sum of the dichotomized subjected assessments of frailty as previously described [9,17].
Estimation of VE. VE was estimated separately for influenza A(H1N1)pdm09 and influenza A(H3N2).
Participants enrolled outside the period of local subtype-specific circulation were excluded, leaving differing numbers of controls in the analytic datasets. Patients co-infected with influenza and another respiratory virus were excluded. The small number of cases precluded estimation of VE against influenza B-associated hospitalizations. VE was estimated in all subgroups using multivariate logistic regression with influenza case status as the outcome and vaccination as the predictor of interest, with VE = 1-[adjusted odds ratio (OR) for vaccination] x 100%. A c c e p t e d M a n u s c r i p t To adjust for confounding, for all analyses, the data were first balanced by baseline characteristics that differed substantially between vaccinees and nonvaccinees using propensity score models and inverse probability of treatment weighting (IPTW) [18]. Details of the propensity score model and VE regression model building strategies are provided in the Supplement. Briefly, characteristics with large baseline differences between vaccinated and unvaccinated participants (SMD >≈ 0.20) and also associated with case status were considered for inclusion in the propensity score model. Because the largest baseline difference was in regard to influenza vaccination habit (proportion of participants reporting receipt of influenza vaccine always/almost always) and prior season vaccination may confound current season VE [19], this characteristic was prioritized and balanced for all analyses. Additional characteristics associated with both vaccination and case status were balanced as feasible, while maintaining balance on vaccination habit. Time-varying characteristics (including calendar time of illness onset relative to peak of subtype-specific case onset date and days between onset and specimen collection) and baseline characteristics that remained unbalanced after weighting were included in the regression model as adjustment variables. Goodness of fit between alternative models was compared with the Akaike Information Criterion (AIC); 95% confidence intervals excluding the null value were considered statistically significant. In subgroup analyses, we examined VE by tertiles of age (18 to ≤56, 56 to ≤69, >69 years) and by influenza vaccination habit (always/almost always vs never/rarely). In addition, we examined VE against influenza A(H3N2) as a function of age specified as a natural cubic spline. Analyses were performed in SAS version 9.4 (SAS Institute, Cary NC) and R version 3.4 (R Group, Vienna).
In addition to the primary analyses, we conducted a bias indicator analysis to assess if VE results could be attributable to additional bias [20,21]. For this analysis, we replicated the primary analysis with identical methods and using as "cases" those patients who tested positive for ≥1 non-influenza respiratory virus as a negative control outcome. The bias indicator analysis was restricted to patients A c c e p t e d M a n u s c r i p t who received clinically ordered testing for multiple respiratory viral pathogens and excluded influenza-positive cases. Because influenza vaccination should not influence infection with a noninfluenza virus, observing an association between vaccination and this outcome would suggest that residual confounding may have affected results. Failure to observe an association would suggest (but not prove) that residual confounding was minimal. Significantly negative or positive bias-indicator estimates would suggest under-estimation or over-estimation of VE in our primary analysis, respectively. To increase comparability between primary and bias analysis datasets, bias indicator analyses were conducted separately for influenza A(H3N2) and influenza A(H1N1)pdm09. group, similar to national trends [22]. Patient characteristics by case status. Average participant age was 63 years (range 18 to 102 years).

Enrollment
Fifty-six per cent and 26% percent of participants were female and of Black race, respectively.
Influenza A(H3N2) cases tended to be older than controls (mean age 64 vs 61 years; p=0.06), and this pattern was similar regardless of vaccination status (Table 1). Comorbidities were common among both influenza A(H3N2) cases and controls, with cardiopulmonary conditions and metabolic A c c e p t e d M a n u s c r i p t disorders (including diabetes) the most prevalent; mean number of high risk conditions was >4 in both groups. However, influenza A(H3N2) cases tended to have fewer comorbid conditions and fewer hospitalizations in the past year than controls. Specifically, control patients had significantly more cardiovascular, lung and immunosuppressive disorders than did influenza A(H3N2) cases.
Controls were slightly more likely than cases to have been admitted to tertiary care hospitals (56% vs 48%). Influenza A(H3N2) cases in the youngest age group (18 to ≤56 yr) had considerable comorbidity and, although less frail, were more likely to have been hospitalized in the year prior to enrollment than older influenza A(H3N2) cases (55% vs 46%; Supplemental Table 1).
Influenza A(H1N1)pdm09 cases (mean age 60 yr) tended to be younger than controls (mean age 62 yr) ( Table 2). As with influenza A(H3N2), cardiorespiratory and metabolic comorbidities were common among influenza A(H1N1)pdm09 cases, but cases tended to have fewer comorbid conditions and hospitalizations than controls. Controls were more likely than influenza A(H1N1)pdm09 cases to have been admitted to tertiary care hospitals. Cases in the youngest age group (18 to ≤56 yr) were more likely to have been hospitalized in the prior year and to have been admitted to tertiary care hospitals compared to older cases (Supplemental Table 2).
Patient characteristics by vaccination status. Compared to unvaccinated participants, vaccinees were older (64 vs 56 yr) and more likely to be of non-Black race (78% vs 63%), to have most specific types of comorbidities and more comorbidities overall (mean Charlson Comorbidity Index of 3.2 vs 2.7), to use home oxygen (30% vs 19%), and to have been hospitalized in the past year (63% vs 55%) (p<0.05 for all; Table 3). Differences in prevalence of comorbidities between the vaccinated and unvaccinated groups were more pronounced in the youngest age group (18 to ≤56 yr). Unvaccinated participants were more likely to have been admitted to tertiary care hospitals (58% vs 53% among vaccinees). Vaccinated participants were less likely to be current tobacco users (22% vs 36%) and far Downloaded from https://academic.oup.com/jid/advance-article/doi/10.1093/infdis/jiaa772/6041424 by guest on 21 December 2020 A c c e p t e d M a n u s c r i p t more likely than unvaccinated participants to report always or almost always receiving seasonal influenza vaccine (90% vs 28%).
Vaccine effectiveness against influenza A(H3N2)-associated hospitalizations. One-hundred and ninety four of 247 (79%) influenza A(H3N2) cases were vaccinated compared with 1455/2057 (71%) controls (Table 4) Fully-adjusted VE against influenza A(H3N2)-associated hospitalization as -15% (95%CI -79, 25) when estimated using baseline characteristics as model covariates instead of IPTW-balanced data (Supplemental Table 3). The largest confounders of the association between vaccination status and influenza risk were age, study site, and influenza vaccination habit. Regression models using IPTW-  Table 5). After balancing baseline characteristics and adjusting for covariates as in the primary analysis, VE against hospitalization associated with non-influenza respiratory viral pathogens was −47 (95%CI −115, 0). Bias indicator analysis results were similar when stratified by age tertile, with the exception of the oldest age tertile (>69 years of age), for which the bias analysis indicated a null result.
Vaccine effectiveness against influenza A(H1N1)pdm09-associated hospitalizations. One-hundred and eighteen of 225 (52%) influenza A(H1N1)pdm09 cases were vaccinated compared with 1568/2222 (71%) controls ( A c c e p t e d M a n u s c r i p t among those >69 years of age; however, this interaction was not statistically significant. Among participants who self-reported never or rarely receiving influenza vaccine, estimated VE against influenza A(H1N1) was 58% (95%CI −7, 84). Among those who reported always/almost always receiving influenza vaccine, estimated VE was 40 (95%CI −4, 65). When estimated using baseline characteristics as model covariates instead of IPTW-balanced data, fully-adjusted VE against influenza A(H1N1)pdm09-associated hospitalization as 50% (95%CI 25, 67) (Supplemental Table 4). A c c e p t e d M a n u s c r i p t Estimated VE differed significantly by age group, and we observed a statistically significant VE of −130% against influenza A(H3N2) among the youngest age group (18 to ≤56 years); VE was most markedly negative for participants between the ages of 18 and 40 years. Our findings are broadly similar to those reported from studies conducted in Canada and Europe [2,3,4], which reported negative VE against influenza A(H3N2) of genetic group 3C.3a. A potential explanation for this finding is provided by Skowronski et al.'s hypothesis that childhood priming following the 1968 influenza A(H3N2) pandemic provided immunity among imprinted individuals, protecting them as adults from subgroup 3C.3a viruses that were similar in 2018-2019 to the imprinting childhood viruses, and that receipt of antigenically-mismatched vaccine negatively interfered with this immunity leaving vaccinees at greater risk of 3C.3a infection [2]. However, the immunologic processes underlying this interference are unknown, and a limitation of our study is the lack of serologic markers with which to examine correlation of antibody titers with disease risk or protection.

Bias indicator analysis for VE against influenza
An alternative explanation is that uncontrolled confounding, selection bias, or chance was responsible for the observed elevation in risk among adult vaccinees, which was most pronounced among the youngest age group. Our relatively small sample of sixty-six young adult cases suffered as much baseline comorbidity as older cases, alluding to their medical complexity and suggesting they likely had characteristics associated with baseline differences in vaccination status that may confound VE estimates. A strength of our study is control for this confounding by systematically balancing baseline participant characteristics within each subgroup. Despite this, we observed that influenza vaccinees had higher risk of hospitalization with non-influenza respiratory pathogens in our bias indicator analyses, suggesting residual confounding may have affected our primary VE results, biasing them lower. A c c e p t e d M a n u s c r i p t VE against influenza A(H3N2) in both the primary analysis and bias indicator analysis was negative in the youngest age group and essentially null in the oldest age group. This may suggest that bias influencing the primary VE results was more pronounced among the younger age group and implying greater confidence in the results for the older age group, for which a finding of null VE was not unexpected given that most circulating viruses were antigenically drifted from the vaccine virus. A potential source of this bias could be a selection bias that allowed preferential recruitment of vaccinated patients with respiratory viral infections in the younger age groups. To manifest as the bias we observed, preferential recruitment would have occurred among vaccinated individuals relatively more than unvaccinated individuals, in the younger age groups relative to the older groups. It is possible that the threshold for hospitalization may have been lower among vaccinated individuals, who tended to have greater underlying comorbidity. If influenza and other respiratory virus infection was also associated with greater likelihood of hospitalization, it is possible, although speculative, that selection bias related to threshold for hospital admission contributed to the bias we observed. For example, immunosuppressed patients account for about a third of all HAIVEN influenza-positive participants and rates of vaccination tend to be higher in this population. Because fever is a common presenting sign of influenza, febrile immunosuppressed patients with influenza, who also have high rates of vaccination, may have a lower threshold for admission [23]. Additional research is needed to further examine the potential for unmeasured confounding and selection bias, and methodologic development of techniques for controlling for unmeasured confounding in the context of influenza VE studies is needed [24]. In particular, the use of a negative control outcome as a bias indicator method and the attributes of an adequate negative control outcome should be further investigated [25]. The utility of positivity for another respiratory virus as a negative control outcome could be limited if influenza infection provided non-specific protection against noninfluenza respiratory viruses, which has been observed for some viruses [26][27][28]. Nonetheless, the bias indicator analysis raises concerns that confounding could not be adequately controlled in this A c c e p t e d M a n u s c r i p t study. Simulation and sensitivity analyses [29,30] to examine the potential magnitude of bias could also be useful next steps.
Although we do not know with certainty the cause of our observed finding of negative VE among the youngest age group, it is worth noting that negative VE results for one virus are unlikely to influence current recommendations for yearly vaccination since vaccination is protective against other influenza viruses and the negative VE result is confined to a specific subgroup and possibly due to measurement bias.
Our results highlight the benefit conferred by currently available influenza vaccines for prevention of influenza A(H1N1)pdm09-associated hospitalizations and provide additional evidence that vaccineinduced protection against influenza A(H3N2) is consistently lower [31]. Biological mechanisms related to early childhood imprinting may potentially contribute to this diminished vaccine effectiveness [32], especially among older individuals who are repeatedly vaccinated [33]. While current vaccines can prevent substantial morbidity and mortality from influenza, particularly for influenza A(H1N1)pdm09 [34,35], better vaccines are urgently needed for prevention of influenza A(H3N2)-associated illness and hospitalization. M a n u s c r i p t A c c e p t e d M a n u s c r i p t ACIP = Advisory Committee on Immunization Practices; ICD10 = International Classification of Diseases, 10th Edition, SD = standard deviation a reported as mean and standard deviation for continuous variables and as number and percent for categorical variables b p value for test of difference across case and control groups based on χ 2 statistic for categorical variables, t-test for difference of means of normally distributed continuous variables, non-parametric log rank test for non-normally distributed continuous variables c standardized absolute mean differences >0.30 are shown in red and >0.10 shown in bold; standardized differences ≤0.10 are generally considered negligible differences between groups d restricted to patients with clinical respiratory viral panel e excludes patients coinfected with a non-influenza respiratory virus f excludes patients coinfected with influenza g age groups defined by tertile of age distribution of all subjects h derived from ICD10 codes associated with inpatient and outpatient medical encounters in the year prior to enrollment admission i defined by patient self report during enrollment interview j plausible self report defined as affirmative self report of current season vaccination including known or approximate date and location of vaccination k self report defined as affirmative self report of current season vaccination including known or approximate date l influenza-like illness defined as self-reported subjective fever/feverishness plus cough or sore throat    A c c e p t e d M a n u s c r i p t 27 Footnotes Table 3 ACIP = Advisory Committee on Immunization Practices; ICD10 = International Classification of Diseases, 10th Edition, SD = standard deviation a vaccination defined as documented and/or plausible self report b excludes patients coinfected with influenza c derived from ICD codes associated with inpatient and outpatient medical encounters in the year prior to admission d defined by patient self report during enrollment interview e influenza-like illness defined as self-reported subjective fever/feverishness plus cough or sore throat f p value for test of difference across case and control groups based on χ 2 statistic for categorical variables, t-test for difference of means of normally distributed continuous variables, non-parametric log rank test for non-normally distributed continuous variables g standardized absolute mean differences >0.10 shown in bold; standardized differences ≤0.10 are generally considered negligible differences between groups A c c e p t e d M a n u s c r i p t 28 A c c e p t e d M a n u s c r i p t 29