Background
Health inequalities, or the differential experience of health between subpopulations, pose major challenges for the control and prevention of illnesses. People of low socioeconomic status (SES) and people in high-poverty neighborhoods experience many worse health outcomes than higher SES populations [
1]. In the case of transmissible infections, researchers have shown a link between SES and respiratory infections including influenza [
2‐
8]. At the household-level, income and educational level are strong predictors of respiratory infection [
8‐
10]. Recent studies have shown a strong, persistent link between area-level poverty and influenza hospitalization rates. High-poverty census tracts in New Haven County, CT (NHC) had significantly higher influenza hospitalization rates during multiple influenza seasons compared to lower-poverty tracts [
9,
10]. In New York City, influenza hospitalization rates during the first wave of the 2009 H1N1 pandemic were significantly higher in high-poverty areas than in low-poverty areas [
11].
Household- and area-level density are correlated with higher disease rates [
3,
9], but whether these factors are sufficient to generate the observed area-level inequalities in influenza remains unknown. In NHC, influenza hospitalization rates were correlated not only with the poverty level, but also with crowding in census tracts [
9]. On the other hand, an analysis of 305 administrative units in the UK suggested that population density and household crowding were insufficient to explain differential influenza rates during the 1918 influenza pandemic [
12], and an analysis of influenza mortality in 10 US cities suggested that population density was only weakly correlated with mortality during the 1918 pandemic [
13]. It remains unclear if population structure (population density and age-structure) is sufficient to generate the spatially differentiated patterns of influenza hospitalization and mortality that have been observed in surveillance data.
The Connecticut Emerging Infections Program (CT-EIP) conducts surveillance for hospitalized, lab-confirmed influenza cases. The program geo-codes each case to its residential census tract, allowing an analysis of influenza hospitalization rate by census tract poverty levels. This surveillance showed that adult and child influenza hospitalization rates increased as the proportion of people living below the federal poverty level in census tracts increased [
9,
10]. This relationship was seen in multiple years from 2007 to 2011, but the mechanism that generates the unequal pattern of disease remains unknown. Because the distribution of household income differs by neighborhood poverty, exposure or susceptibility differences by household income may manifest as area-level inequalities by census tract poverty.
Differential exposure to virus, susceptibility to disease, and access to healthcare have been proposed as possible mechanisms that could generate unequal levels of influenza illness and death [
5,
14‐
16]. If, in fact, higher hospitalization rates are a reflection of higher incidence rates, differential exposure due to larger household size, greater population density, and younger age-structure could explain unequal hospitalization rates. Attack rates were higher in households classified as lower SES in multiple cities during the 1918 influenza pandemic [
17]. Community-based surveillance for acute respiratory infection in Tecumseh also suggested a higher incidence rate correlated with lower household income [
8].
Once exposed, differential susceptibility due to differential vaccination, chronic condition prevalence, and smoking rates could further impact incidence rate inequalities [
14]. Vaccination and smoking rates differ by household income [
18,
19], as do rates of chronic conditions such as obesity [
20], diabetes [
21], and asthma [
22]. In NHC, hospitalized cases from high-poverty tracts had lower vaccination rates and higher rates of chronic conditions [
9], suggesting that cases were more susceptible to disease in these tracts. Alternatively, hospitalization rate inequalities could be manifestations of differential disease severity due to underlying chronic conditions and delays in healthcare access.
With so many potential drivers of inequalities, it is important to first ask if population structure is sufficient to generate the observed inequalities. Because influenza is typically documented only when ill people contact the healthcare system, i.e. in only a subset of cases, influenza surveillance data are not ideal to test if population structure is sufficient to generate differential incidence rates that reflect the magnitude of observed hospitalization rate inequalities. To address this question, we require innovative methods such as spatially explicit agent-based models.
Agent-based models using populations of agents with realistic spatial and demographic characteristics can serve as a counterfactual laboratory in which to test competing hypotheses regarding the possible causes of influenza inequalities [
23,
24]. Here, we ask if population structure is sufficient to generate the observed inequalities in NHC influenza rates by area-level poverty. By population structure, we mean population density, age structure, and average household size in the residential census tract; and contact rates in each location (workplace, neighborhood, school, household). We use simulation models that include demographic factors (age structure, household size, and population density) and neighborhood, household, school, and workplace mixing to address this question. We present a temporal analysis of attack rates by poverty, and a validation of model findings using CT-EIP data. Agent-based models are especially useful in examining the role of individual-level rules in generating population-level patterns. In this case, we examined whether individual contact rates or household income-based infection susceptibility were able to generate the pattern of area-level disparities observed in NHC. We discuss our findings in relation to factors that could lead to differential area-level influenza rates and propose additional hypotheses for future research.
Discussion
Though area-level inequalities in influenza hospitalization and mortality rates have been reported [
9‐
12], the causes remain unclear. Knowledge of the causes and their relative explanatory power will allow us to design interventions that have the greatest impact on reducing inequalities. Spatially explicit simulation methods allow us to test the role of hypothesized social factors in generating area-level inequalities. We used agent-based models with realistic populations to test the sufficiency of population structure in generating the observed area-level inequalities in influenza hospitalization rates in NHC. An epidemic simulation model that included demographic factors (age structure, household size, and population density), as well as neighborhood, household, school, and workplace mixing, accounted for 33 % of the positive correlation between census tract poverty level and observed influenza rates. Higher poverty areas had not only higher simulated infection rates, but also earlier and steeper increases in infection rates—an emergent property of the model that we corroborated using surveillance data. To our knowledge, this is the first study to use simulations to probe the causes of observed inequalities in influenza disease patterns.
Other studies have reported the ability of simulation models to generate spatially patterned influenza rates, but did not compare their results to observed surveillance data. Stroud
et al. found that simulated influenza attack rates in southern California census tracts were significantly correlated with average household size, the proportion of population 5-18y age, population density, and per capita income [
46]. Dalziel et al. showed that a model based on intra-city mobility patterns of individuals was sufficient to generate unequal influenza rates between Canadian cities [
47]. We need to compare their results to surveillance data from California and Canada in order to determine if model-generated patterns of influenza incidence rates are comparable to those observed in reality.
Temporal trends in census tracts
Our simulation suggested that high-poverty areas reach higher attack rates earlier during an epidemic than do lower-poverty areas. We examined the CT-EIP data for temporal differences in hospitalization rates by census tract poverty and saw that indeed, the data qualitatively validate our model findings. This result has important implications for preparedness and planning. Interventions should be rolled out in high-poverty areas before low-poverty areas. During the (H1N1) 2009 pandemic, vaccine first became available in October [
48]. If the epidemic progresses at different rates in high- and low-poverty areas, it becomes important to increase availability of vaccine in high-poverty areas first. Past models have shown that targeting vaccine to high-poverty counties when vaccine is limited could reduce disease overall [
36]. The impact of targeting non-pharmaceutical interventions to high-poverty areas on reducing influenza inequalities should also be examined.
Individual agent-parameters that generate area-level inequalities
In our baseline model, all agents were equally susceptible to influenza. Under these conditions, increasing the individual contact rates in neighborhoods to higher than the calibrated value did not increase the incidence rate ratio between high- and low-poverty neighborhoods. As we increase the contact rate in the neighborhood, the proportion of infections in the neighborhood goes up as expected, with a resulting decrease in the proportion of infections occurring in other locations (compared to the calibrated contact rates in our baseline scenario). Our models then do not reflect observations and past assumptions regarding place-based infection rates [
39,
49], precluding a sensitivity analysis with a larger range of neighborhood contact rate values. We expect that attack rates result from contacts not only in the residential tract but also in workplaces, schools, and households. In addition, intensity of contacts is lower in neighborhoods compared to the intensity in other locations (workplaces, schools, households). The extent of disparity that the model generates is therefore an emergent property of the model.
A systematic variation in household-level differences in susceptibility suggests that when the highest income households were substantially less susceptible to influenza (14 % as susceptible as the lowest income households), our model generated the area-level inequality that has been observed in NHC. Researchers have reported differential incidence and susceptibility by household SES [
8,
17]. The relationship between income and susceptibility may not, however, be linear, as we assumed in our model. In order to probe the relationship between income and susceptibility, models may need to incorporate more detailed data on how vaccine uptake and smoking behavior, as well as chronic disease prevalence differ with household income, and how these behavioral and biological factors modulate susceptibility to infection.
The role of virus transmissibility
Our model suggests that area-level inequalities are more pronounced with lower transmissibility of the virus. This is consistent with previous studies reporting that differential attack rates due to population structure were more pronounced when simulating pathogens with lower transmissibility [
47]. Based on overall hospitalization rates from the CT-EIP [
41] and RR
EIP between the highest and lowest poverty category tracts in NHC [
50], it appears that influenza seasons with lower overall hospitalization rate in the population correspond with higher RR
EIP (signifying higher inequalities) [
9,
50]. A likely explanation is that the greater the transmissibility, the higher the attack rate in all poverty categories, including the lowest poverty tracts, leading to lower inequalities.
Limitations of the model
The current model does not account for disease symptoms or severity. This may limit our ability to model hospitalization rate inequalities that result from unequal disease
severity between high- and low-poverty neighborhoods even if attack rate differences do not completely reflect the inequality observed. Future studies should include more detailed models of individualized symptoms. [
51] A determinant of disease severity is underlying chronic disease. Obesity and chronic conditions such as diabetes and asthma increase the risk of hospitalization due to influenza [
52,
53]. Similarly, stress, which is higher in low-income populations [
54], increases influenza severity once infected [
55]. In post-hoc analyses, we found that taking into account differential influenza hospitalization rates between asthmatics and non-asthmatics explained only a small additional proportion of observed inequalities in adult AR
EIP. Future studies should account for uncontrolled asthma and other chronic conditions, and stress.
In addition, neighborhood-level factors, such as access to healthcare, may interact with household income to result in differential delays in health care access for influenza symptoms. This would result in cases being more severe in high-poverty neighborhoods at the time they first contact the healthcare system. Low-income people (unemployed adults and those without insurance) as well as those with underlying chronic disease are more likely to delay seeking healthcare when they have influenza symptoms. [
56] Differential healthcare-seeking behavior may thus add to differential attack rates to close the gap between observed and model-generated inequalities in influenza hospitalization rates in NHC.
These same factors may also increase model-generated inequalities in children’s AR
SIM, which in our current model did not approach that observed in reality. On the other hand, the finding that inequalities in children’s AR
SIM fell short of that observed in NHC may point to a limitation in the modeled contact mixing structure among children. A strength of our simulated populations is that they include realistic school sizes and locations [
29], but there is little literature on contact probabilities in schools based on school size. An important use of models is to guide additional data collection [
24,
58]. Contact-mixing data in schools should be collected, and such information, if available, should be incorporated into ABMs in the future.
Conclusions
Our agent-based model provides a counterfactual laboratory to test hypotheses for the production of inequalities. It is difficult to compare alternative hypotheses for the generation of influenza disparities using observational techniques because of the large number of factors that would need to be controlled for in order to make conclusions. For example, in order to test the hypothesis that population structure is sufficient to account for influenza attack rate disparities, we would need to factor in household size, the number and contacts of school-going children, employment status and workplace contacts of adults, and population density and contacts in neighborhoods, in addition to controlling for alternative hypotheses such as stress and health behaviors. Given the small number of clinically confirmed influenza cases and the dynamics of infectious agent spread, observational methods are not ideal for such studies. As pointed out by Epstein [
58], and more recently by Marshall and Galea [
24], agent-based models are especially well-suited to the exploration of area-level inequalities because of their characteristic features: heterogeneity of agents and spatially explicit agent interactions. We have utilized these features of agent-based models to provide the first test of a proposed theoretical mechanism (crowding or population structure) for the generation of inequalities in influenza disease.
Results from the current study suggest that population structure in our simulations accounts for 33 % of the observed area-level inequalities in adult influenza hospitalization rates, leaving 67 % of the inequality to be explained by other factors. Future models should quantify the capacity of individual behavioral and biological factors to generate influenza inequalities thus allowing us to prioritize interventions aimed at factors with the greatest explanatory power.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
SK conceptualized the study. SK and JG designed the experiments. JG and DDG developed the model. SK carried out experiments. SK and KNP analyzed data and wrote the paper. JJG and JLH edited the paper. All authors read and approved the final manuscript.