Social determinants of influenza hospitalization in the United States

Background Influenza hospitalizations result in substantial morbidity and mortality each year. Little is known about the association between influenza hospitalization and census tract‐based socioeconomic determinants beyond the effect of individual factors. Objective To evaluate whether census tract‐based determinants such as poverty and household crowding would contribute significantly to the risk of influenza hospitalization above and beyond individual‐level determinants. Methods We analyzed 33 515 laboratory‐confirmed influenza‐associated hospitalizations that occurred during the 2009‐2010 through 2013‐2014 influenza seasons using a population‐based surveillance system at 14 sites across the United States. Results Using a multilevel regression model, we found that individual factors were associated with influenza hospitalization with the highest adjusted odds ratio (AOR) of 9.20 (95% CI 8.72‐9.70) for those ≥65 vs 5‐17 years old. African Americans had an AOR of 1.67 (95% CI 1.60‐1.73) compared to Whites, and Hispanics had an AOR of 1.21 (95% CI 1.16‐1.26) compared to non‐Hispanics. Among census tract‐based determinants, those living in a tract with ≥20% vs <5% of persons living below poverty had an AOR of 1.31 (95% CI 1.16‐1.47), those living in a tract with ≥5% vs <5% of persons living in crowded conditions had an AOR of 1.17 (95% CI 1.11‐1.23), and those living in a tract with ≥40% vs <5% female heads of household had an AOR of 1.32 (95% CI 1.25‐1.40). Conclusion Census tract‐based determinants account for 11% of the variability in influenza hospitalization.


| INTRODUCTION
Influenza epidemics result in substantial morbidity and mortality each year in the United States with annual deaths ranging from approximately 3300-49 000. [1][2][3][4][5] Although annual vaccination is fundamental in preventing and controlling influenza, 6 vaccination coverage is suboptimal and is further impeded by racial and ethnic disparities. 7,8 To guide vaccination programs and reduce disease burden, it is essential to understand the risk factors for severe influenza outcomes.
Interventions and policies that focus on both individuals and their neighborhoods may be more effective in improving health outcomes than those that focus exclusively only on individuals. 9,10 As most disease surveillance systems do not capture sufficient individual socioeconomic characteristics, using census tract data can provide important neighborhood-level dynamics that impact an individual's risk for influenza to characterize neighborhood-level determinants. [11][12][13][14][15][16][17][18] Census tract-based determinants such as percent living below poverty, household crowding, and female head of household have been associated with influenza hospitalization. [19][20][21][22] To date, researchers have incorporated census tract-based determinants into evaluations of univariate associations with influenza hospitalization, which do not account for whether census tractbased determinants have an independent effect beyond the effect of individual-level factors. In community-level data, population characteristics of individuals living within a census tract tend to be correlated, and hence, the assumption of independence among observations is violated. Multilevel analyses take into account hierarchical data structures 23,24 and enable us to account for this lack of independence, detect multivariate associations, incorporate covariates at both the individual and the geographic level, and model interactions between variables. These models have been used to evaluate health disparities and to describe the relationship between geographic exposures and a wide variety of health outcomes such as diabetes, immunization, obesity, and cancer mortality. [25][26][27][28][29][30][31] We conducted a multilevel analysis to identify individual and census tract-based determinants associated with influenza hospitalization using population-based surveillance data collected from 14 sites across the United States. We hypothesized that census tract-based determinants such as poverty and household crowding would contribute significantly to the risk of influenza hospitalization above and beyond individual-level determinants.

| Data sources
We used Influenza Hospitalization Surveillance Network (FluSurv-NET) 32   if receipt of influenza vaccine occurred at least 14 days prior to hospitalization. Multiple sources, including the medical record, state vaccination registry, the patient's primary care provider, and interview of the patient or proxy, were used to obtain vaccination history.
Data collection was determined by the CDC to be for routine public health surveillance purposes, and thus was not subject to institutional review board (IRB) approval for human research protections.
Participating sites submitted the study to their state and local IRBs for review as required.

| Study population
In this analysis, we included children and adults with laboratory-

| Geocoding and census tract determinants
Each patient's address was geocoded to a latitude and longitude point and linked to a census tract using geocoding software such as ArcMap. 33 Each site participating in FluSurv-NET was responsible for geocoding its own data. Census tracts from geocoded data were then merged to the US Census Bureau's American Community Survey  (Table S1).

| Outcome and covariates
Our primary outcome variable was hospitalization due to laboratoryconfirmed influenza during the study period. As our goal was to estimate the association of neighborhood and individual-level characteristics on influenza hospitalization rates, we included covariates at both the individual level and the census tract level. Individual-level covariates included age (<5, 5-17, 18-49, 50-64, >65 years), race (White, African American, other), gender, and ethnicity (Hispanic or non-Hispanic). Census tract-based determinants for the model were chosen a priori based on existing literature, biological plausibility, and whether these factors were found to be associated with influenza hospitalization in a prior study using Tennessee data. 19 When two variables were found to be highly correlated (Spearman's correlation coefficient >0.70), only one was included to avoid inflation of variances. Spearman correlation values between census tract-based determinants are presented in the eAppendix (Table S2). Census tractbased determinants used in the analysis were percent below poverty

| Exclusions
For our analysis, we excluded those without a hospitalization date and those with a missing laboratory diagnostic test for influenza. For the purposes of geocoding, we excluded addresses that could not be geocoded to rooftop accuracy. Additionally, due to new surveillance areas being incorporated over time, for consistency purposes we limited our sample to only those census tracts that remained common through our study period of interest.

| Data analyses
We calculated descriptive statistics by site and influenza season.
Categorical variables were summarized by frequency and percentage whereas median and interquartile ranges (IQR) were used to describe continuous variables. We computed crude site-specific annual incidence rates as well as age-standardized incidence rates by individual factors as well as census tract-based determinants. The age-standardized rates were calculated using the 2000 U.S. Standard Population. Health disparity measures such as risk ratio (RR), risk difference (RD), relative index of inequality (RII), and the Concentration Curve for individual factors age, race, gender, and ethnicity and census tract-based determinant were calculated by site and are presented in the eAppendix (Table S3-S8). 36 These methods and interpretation have been previously published by Tennessee. 19 The dataset used for the analysis of associations had a three-level cell structure in which individuals were nested in census tracts, which were further nested within sites. A cell is defined as a unique combination of age, race, gender, ethnicity, season, census tract, and site for the popula- To address whether the association between percent poverty and influenza hospitalization was modified based on population density, we included an interaction of the two terms in the model. In addition, we evaluated whether using population density of children less than 5 years old would yield identical results to the overall population density. Finally, to explore the impact of individual-level covariates such as the presence of asthma, underlying medical conditions, and influenza vaccination status, we refit the model separately on subsets of patients that were indicated to be affirmative for that variable. The model was also fit on the subset of hospitalized patients who did not get vaccinated for influenza.

| Descriptive statistics of influenza-positive patients
Exclusion details are provided via a flowchart in Figure 1

| Multilevel model
In the unconditional model that included only the site and tract within site as random effects, the intraclass correlation that describes the cluster effect was computed to be 8% for site and tract and 11% for site alone. This indicated the presence of a moderate clustering effect by site and that 11% of the variability in influenza hospitalization could be explained by census tract-based determinants. This left 89% of the variability to be accounted for by individual-level characteristics.
All individual-level demographics were significantly associated with influenza hospitalization with the highest adjusted odds ratio (AOR) of 9.20 (95% CI 8.72-9.70) for adults 65 years and older vs children 5-17 years (  Excluded for the following reasons: 1. Missing report date (n = 1) 2.
Inconsistent tracts through the 5-year surveillance period (n = 882). Tracts were removed from sites RI, CT, MI, NM, and OH 4.

| Sensitivity analyses
As part of sensitivity analyses, we evaluated whether the association between percent poverty and influenza hospitalization was modified based on population density by including an interaction of the two terms in the model. We also evaluated whether using population density of children less than 5 years old would result in the same results as including overall population density. The model results remained similar in inference when using population density of children less than 5 years old in place of overall population density. The interaction between population density and percent of people living below poverty was not statistically significant and hence was removed from the model for parsimony and ease of interpretation. We conducted stratified analyses on subsets of the data that included asthma, underlying medical conditions, and influenza vaccination status. The model was also fit on the subset of patients that indicated that they did not get vaccinated for influenza. Among people diagnosed with asthma, those who were not vaccinated for influenza as well as those   (Table S13).

| DISCUSSION
In this study, we adopted a multilevel modeling approach in order to Area-based socioeconomic determinants have been shown to be strongly associated with a wide range of health problems, and hence, our finding that socioeconomic determinants play a fundamental role in influenza hospitalization is not alarming. Although the exact underlying mechanisms remain unclear, we hypothesize that socioeconomic determinants probably shape individual behavior by influencing characteristics such as individual education, their income, medical insurance status, education, access to care, and health-related behaviors as well. Factors such as crowding may also increase exposure to the influenza virus. It is likely that there exist possible different perceptions of the seriousness of influenza and hence delays in seeking care that may be related to the outcome as well. It can also be hypothesized that high level of health literacy may be associated with areas with higher percent of educated.
In conclusion, although the strongest associations were observed with respect to individual-level characteristics such as age and race, census tract-based determinants also were associated with hospitalization for influenza. Identifying and targeting areas for specific prevention and control interventions (eg, vaccination) could help reduce some disparities in influenza outcomes for areas with high percent of poverty, household crowding, and female heads of households.
T A B L E 3 Adjusted odds ratio and its associated 95% confidence interval from the multilevel regression model of influenza hospitalization adjusting for individual-level (age, gender, race, and ethnicity) and census tract-based determinants (female head of household, percent below poverty, percent crowding, population density, percent insured, and percent college-educated)