Introduction

Health disparities are differences in health outcomes between social groups, and they persist in all modern public health settings. Health disparities may be the result of health inequalities, which are caused by biological or cultural variations, or by health inequities, which are driven by unfair factors and are avoidable with policy action [1]. There is extensive evidence that social factors, including education, employment, income, race, and ethnicity have a distinct influence on how healthy a person is: the lower an individual’s socioeconomic position, the higher their risk of poor health for both chronic and infectious diseases in low-, middle-, and high-income settings alike [2]. There is also a role played by geographic context: the spatial distribution of disparity in health cannot be explained by variation in social factors alone [3]. As the divide in health disparities grows wider across the world and within countries, it is imperative that we continue to understand how social determinants impact health, and how this is reflected geographically [4]. Here, we integrate empirical insights from past studies to characterize the impact of social determinants on the dynamics and spatial heterogeneity of an infectious disease case study, influenza.

Influenza is a respiratory-transmitted infectious disease that occurs in annual epidemics in temperate regions that can have severe outcomes, especially in young children and elderly individuals [5]. Several studies have demonstrated social differences in influenza morbidity and mortality [6–11]. The most impoverished areas have been shown to experience twice the influenza hospitalizations compared to regions with the lowest rates of poverty [12], and low education has been shown to be positively associated with influenza hospitalization rates [13]. Past work has even shown that socioeconomic factors played a significant role in the morbidity and mortality caused by the 1918 influenza pandemic [14–16]. The proposed determinants of disparities in influenza burden include a number of socio-behavioral and healthcare-based dimensions [17, 18]. In particular, influenza vaccine coverage and healthcare access are higher in areas with increased levels of education and household income [19, 20]. Additionally, low socioeconomic status (SES) individuals have been shown to experience increased susceptibility to respiratory infections due to increased stress [21, 22] and have less access to paid sick leave, resulting in less school and workplace sickness absenteeism, defined as remaining home due to illness [23, 24]. Lastly, it has been proposed that the social patterns of low SES populations affect their influenza risk: larger household sizes and higher population density may lead to higher infection risk [25, 26], while a less robust social network might result in decreased exposure, but also less support during recovery if infected [18].

Mathematical modeling studies of social disparities in influenza burden have used a simulation approach [27–29] and have focused on the effects of material deprivation (i.e. lack of access from income, education, and employment) or social deprivation (i.e. lack of social cohesion and support due to small household sizes, single parenting, divorce or widowing). Such studies are important in uncovering the mechanistic explanations of influenza disparities, but have been limited in their geographical extent, or by the use of proxy measures. For example, [27, 29] consider phenomenological variation in social contact rates without empirical evidence linking vulnerable groups to that variation, thus limiting insights on the mechanisms that lead to influenza disparities; [28, 29] focus on dynamics within specific cities, limiting generalizability.

Surveillance-based statistical studies of influenza disparities have been spatial in nature and have highlighted the challenges of disease surveillance under these disparities. Surveillance systems gather the data that shapes our understanding of influenza dynamics, and in the US and most European countries, influenza-like illness (ILI) surveillance occurs through reporting by sentinel healthcare providers. Such sentinel surveillance systems have been resource-efficient means of collecting high quality data, but they do not reliably capture data for all populations, since they are dependent on health care accessibility, health care seeking behavior, and other reporting issues [30, 31]. As a result, studies that rely on healthcare data for characterizing rates of ILI sometimes find decreasing rates of disease with increasing social deprivation [18]. While this negative association may be the result of lower exposure in impoverished areas (as suggested by [18]), it is likely that there exist spatial and social heterogeneities in surveillance caused by healthcare utilization. Indeed, Scarpino et al. have shown that the most impoverished areas are blindspots in the US influenza sentinel surveillance system, ILINet, and models based on these data make the best predictions in affluent areas, while making the worst predictions in impoverished locations [32]. To better understand and respond to influenza epidemics and pandemics, we must improve our capability to detect and monitor outbreaks in at-risk populations.

In this work, we (a) develop data-driven epidemiological models to assess how social and healthcare-based determinants impact population-level influenza transmission in a controlled manner; and (b) develop statistical ecological models from large-scale disease data to estimate latent influenza burden in vulnerable populations in the United States. We hypothesize that low SES populations bear a disproportionate burden of influenza infection, and that a combination of social, economic and health factors cause this disparity. We aim to identify geographic areas where burden is highest in low SES populations to provide hotspots for additional surveillance. As health disparities widen, it is imperative that we develop an empirically-driven mechanistic understanding of the determinants of health disparities, and capture disease burden in at-risk populations. Such insights can allow for improved influenza forecasting, resource allocation and targeted intervention design.

Results

Here, we have evaluated the impact of social and healthcare-based mechanisms on driving influenza disparities. We achieved this through three main steps. First, we used population network modeling to generate contact networks that represent realistic social contact based on SES and modeled influenza transmission on these realistic contact networks, incorporating hypothesized drivers of influenza inequities. This framework better allowed us to understand the role that SES-driven variation plays in determining influenza dynamics and allowed us to disentangle the effects of multiple proposed drivers of influenza transmission among those of differing SES. Second, we ecologically investigated the impacts of low SES on influenza. We estimated low SES ILI incidence ratios at the county-level in half of the states in the US using a spatial inferential model, accounting for transmission trends identified in the prior epidemiological model experiments, variation in social, economic and health factors, and measurement biases. This model allowed us to identify areas which may be currently overlooked by influenza surveillance systems. These findings also indicate potential SES-based factors associated with disproportionate burden at the population level, which could guide future public health efforts to reduce socioeconomic health disparities.

Contact patterns vary by socioeconomic status

Contact patterns have been demonstrated to vary by socioeconomic status [18], but we have lacked social contact networks that explicitly incorporate these differences. To enable testing of hypotheses about social contact trends, we used an egocentric exponential random graph model (ERGM) to simulate networks with realistic social contact patterns based on socioeconomic status (measured by education level, [33]) from the POLYMOD social contact survey, a large social contact survey conducted across Europe [34]. The fitted network model is consistent with the contact heterogeneity in the data (Fig 1A), and all individual-level attributes (i.e. age, sex, contact location, and education level) are significant in predicting contact structure (S1 Table). Additionally, we incorporated varying levels of low SES individuals into the networks to investigate hypotheses in populations with varying SES composition, from 20% low SES to 60% low SES. The resulting networks are consistent in network structure based on degree and assortative degree (number of contacts with those of the same attribute/number in network with that attribute) by SES-status (Fig 1B). Thus, networks with increased representation of low SES individuals maintain the same SES-based contact patterns as the POLYMOD data. Importantly, the network model captures variation in contact structure by SES. In particular, low SES individuals have lower mean degree and variation in degree (Fig 1C), but have higher SES-assortative degree compared to those of higher SES (Fig 1D).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The characteristics of the networks generated from the ERGM model based on POLYMOD data.

A: The degree distribution of the POLYMOD data (light green) compared to 10 simulated networks (dark green). B: The Kolmogorov-Smirnov (KS) statistic to evaluate the dissimilarity of the ERGM-simulated networks to the POLYMOD data as additional low education individuals are added to the network. KS statistics compare the dissimilarity of the overall degree distribution (dark green), the degree distribution of low SES nodes (light blue, solid), the degree distribution of high SES nodes (dark blue solid), the assortative degree (e.g. the low SES contacts of low SES nodes) for low SES nodes (light blue, dashed), and the assortative degree for high SES nodes (dark blue, dashed). Low KS values indicate similar distributions. C: The degree distribution of low SES nodes (light blue) and high SES nodes (dark blue) in 10 simulated networks. D: The relative assortative degree distribution (e.g. number of low SES contacts of low SES nodes/number of low SES nodes) of low SES nodes (light blue) and high SES nodes (dark blue) in 10 simulated networks.

https://doi.org/10.1371/journal.pcbi.1008642.g001

Inequities increase low SES influenza transmission

There appears to be variation in contact trends dependent on socioeconomic status, thus it is important to consider how this network structure impacts epidemiological dynamics. To assess the role of social and healthcare-based heterogeneity, we integrated into an epidemiological network model of influenza transmission five key hypothesized drivers of disparities in influenza burden: a) social contact differences, or fewer social contacts and higher assortativity (as represented in our empirically-informed contact network model); b) low vaccine uptake; c) low healthcare utilization, which results in less access to influenza antivirals; d) high susceptibility, which results from stressful environmental factors; and e) low sickness absenteeism from school or work. S4 Table explains how these drivers are incorporated into the influenza transmission simulation, and includes other epidemiological parameters. The “low” parameters apply to low SES individuals, and “high” parameters apply to medium and high SES individuals. Then, we simulate influenza with these parameters randomly distributed across the population at the same rate as a positive control. Fig 2A shows the infection burden of low SES individuals (i.e. the ratio of the number low SES infections and the number of all infections) in the presence of each factor, combined with social cohesion (included in the network structure). Each factor results in a significant increase in the low SES infection burden in the presence of SES-based variation in parameters (dark green) compared the the random distribution of parameters (light green), and the effect is most pronounced when all the factors occur simultaneously. In contrast, the epidemic size (i.e. the ratio of the number of infections and the population size) for the positive control is larger than the SES-heterogeneous treatments, for all treatments (with the exception of the increased stress treatment) (S28 Fig). A sensitivity analysis of the low SES parameter values demonstrates that our findings of increased infection of low SES individuals when they experience transmission-increasing mechanisms, compared to when the transmission-increasing mechanisms are randomly distributed, hold for all considered values (S31–S34 Figs). Thus, wherever there is a disparity in these parameters, our findings hold, regardless of the magnitude of the disparity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Results of epidemiological simulations on ERGM networks with SES-driven social and healthcare-based differences.

A) All of the proposed SES-driven differences result in an increase in infection of low SES individuals (dark green, right of paired violin plots), compared to simulations where the differences are randomly distributed throughout the population (light green, left of paired violin plots). This difference is most pronounced when all of the mechanisms occur together. These simulations were performed on a network composed of 60% low SES, but the results are consistent across networks with different SES compositions. B) In all networks, when all SES-driven differences are present, low SES individuals (mean percent of infected population that is low SES shown in light blue dots) are disproportionately infected, relative to the expectation (light blue dashed line). High SES individuals are disproportionately underinfected compared the expectation (dark blue dots compared to dark blue dashed line).

https://doi.org/10.1371/journal.pcbi.1008642.g002

This combination of results can be explained by the role that low SES individuals play in the network. On the one hand, low SES individuals have lower mean degree (Fig 1C). When these low degree individuals experience transmission-increasing mechanisms, this results in a smaller epidemic size, compared to the scenario where high SES, and high degree, individuals experience the same mechanisms. Thus, when SES-driven processes that increase transmission affect low SES individuals, it results in a smaller overall epidemic. On the other hand, low SES individuals have high assortativity with other low SES individuals (Fig 1D). Thus, when health disparities increase transmission for low SES individuals, they are more likely to infect other low SES individuals that are also experiencing these mechanisms, resulting in increased spread among this assortative group. This result highlights the need for surveillance and research focused on low SES populations, as the emergent high infection burden of low SES, at-risk individuals could be overlooked due to lower epidemic sizes when aggregated.

Next, we consider how low SES infection burden scales with an increasingly large low SES population. We find that epidemic size increases with an increasing proportion of low SES individuals, and this effect appears to be driven by increasing infection of low SES individuals as they make up a larger component of the network (S27 Fig). Indeed, low SES individuals experience a disproportionately large infection burden when all SES-based factors occur (Fig 2B). Additionally, high SES individuals experience a disproportionately small infection burden in the presence of the same factors. This distribution of infection also is consistent over multiple influenza seasons, with partial polarized immunity to reinfection. Though the overall epidemic size changes due to prior immunity (S29 Fig), low SES individuals are disproportionately infected in each season (S30 Fig).

Low SES infection burden is spatially heterogeneous, and high in the southeastern US

Our results thus far characterize the mechanistic role that social and healthcare-based factors play on influenza burden in low SES populations in data-driven controlled experiments. Here, we aim to characterize how macroscopic factors impact influenza dynamics in low-SES populations, integrating our theoretical findings with population-level data. For population-level influenza data, we used medical claims of ILI at the county level in 25 states in the US, based on sufficient data availability. This data stream has been demonstrated to provide enhanced surveillance opportunities for influenza-like illness [31, 35]. However, when we compare county ILI burden with the proportion of the county’s population of low SES, we find a negative relationship, indicating lower levels of influenza in counties with larger low SES populations (S35(A) Fig). This pattern is counter to our previous mechanistic model findings and to past small scale studies, suggesting that there may be measurement biases in these surveillance data (S35(B) Fig).

To better estimate influenza burden in low SES populations, we fit a Bayesian spatial hierarchical model that accounts for measurement biases and borrows information from spatial covariates pertaining to low SES individuals and the mechanistic modeling experiments. Here, we define low SES ILI as an incidence ratio, where low SES ILI cases are normalized by the number of 1000 visits, to account for spatio-temporal variation in database coverage and healthcare-seeking. Thus, the model outcome data is the rate of ILI healthcare visits per 1000 healthcare visits within each county. Our model estimates of this low SES ILI incidence ratio show a positive relationship with low SES population size (S35(C) Fig), and allow us to consider spatial disparities in influenza burden. Fig 3 shows the county-level map of the low SES ILI incidence ratio. This map highlights areas with a high incidence ratio among low SES individuals in the southeastern United States, which is a region where low socioeconomic status is prevalent. This also demonstrates that there are significant levels of heterogeneity both within and between states. These estimates can guide targeting of improved surveillance and steps to alleviate the influenza burden in low SES populations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. County-level map of model estimates of low SES ILI incidence ratio per 1,000 people.

Lower values are represented in light blue, and higher values are represented in darker blue. States in white were omitted due to lack of covariate data. Some county covariate data in included states was imputed based on surrounding neighbors, where missing. Unimputed findings are available in S37 Fig.

https://doi.org/10.1371/journal.pcbi.1008642.g003

To validate our findings, we grouped our model estimates by county-level poverty rates, and compared the incidence ratio to prior population-level studies that correlate influenza rates and poverty levels, though these studies do not focus on low SES individuals, so the comparison is not direct. We find increasing low SES ILI incidence in areas with increasing levels of poverty, which agrees with trends in [11, 12] (S40 Fig).

Susceptibility and sickness absenteeism differences may be associated with ILI in low SES populations

Fig 4 shows the coefficient estimates and credible intervals resulting from the Bayesian spatial hierarchical model. Levels of poor health among low SES individuals, as a measure of susceptibility to infection, are positively related with low SES ILI incidence. Thus, areas with higher reports of poor health among low SES individuals are associated with higher burden of ILI among low SES populations. Also, access to sickness absenteeism among low SES individuals, represented by the number of low SES students that are absent for more than 10 days in a school year, is negatively related to low SES ILI incidence ratios. Thus, areas where more low SES students are able to be absent are associated with lower rates of low SES ILI.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Mean model coefficient estimates and credible intervals.

Points are colored by what process each covariate represents (black: measurement bias, red: susceptibility, orange: social contact differences, green: sickness absenteeism, blue: vaccination, purple: healthcare utilization). Each process covariate is specific to low SES populations (e.g.“adult vaccination” is only vaccination rates of low SES adults).

https://doi.org/10.1371/journal.pcbi.1008642.g004

Discussion

Increased infectious disease prevalence among lower socioeconomic status populations has been observed in many settings. What has been missing, however, is a better understanding of the mechanisms that drive this disparity. We used a mechanistic epidemiological network model which allowed us to assess the impacts of SES-based social and healthcare-based differences on influenza in controlled experiments. This highlighted the role played by all mechanisms in tandem to produce disproportionate disease burden in low SES populations. To address the gap that exists in our surveillance of ILI and to estimate the spatial distribution of influenza disparities, we then used a Bayesian spatial hierarchical model to estimate population-level low SES ILI at a fine spatial scale across the United States, accounting for disproportionate infection of low SES individuals, measurement biases, and county-level factors hypothesized to be associated with influenza and SES. Our results shine light on the spatial distribution of respiratory-transmitted disease health disparities.

In our epidemiological model, disease transmission occurs over the contact network structure, which accounts for heterogeneity in contact patterns by SES. While past work has integrated contact heterogeneity by other socio-demographic characteristics such as age and occupation [36, 37], SES-based contact heterogeneity has not been integrated into contact network models for epidemiological purposes. Epidemiological simulations on the SES-heterogeneous network reveals that each hypothesized factor leads to increased infection of low SES individuals. Additionally, we find that communities with larger low SES populations experience larger epidemics, which is in agreement with prior studies [10–12]. The proposed drivers are not mutually exclusive, so this reveals potential effects that could not be identified in past studies that investigate the impact of a single SES-based mechanism or impacts that might be aggregated in observational studies. We note that these experiments also include SES-based variation in social cohesion (i.e. SES-based contact heterogeneity in the population model), so the effect shown in Fig 2A is the result of both mechanisms combined. In S28 Fig, we also illustrate the impact of each mechanism independent of social cohesion.

Our efforts to consider the impacts of low SES on influenza spatial heterogeneity generated county-level maps of ILI incidence in low SES populations. Our findings identify pockets of high ILI burden in low SES populations across the United States, and represent a first step in filling the gap that exists in all healthcare-based surveillance. The model also produced a set of estimates for the effect of each hypothesized ecological measure. We find that low sickness absenteeism and high susceptibility are significantly associated with influenza in low SES populations. This supports our previous finding that multiple mechanisms compound to result in disproportionate low SES influenza burden. Low SES absenteeism is here measured by student absenteeism, which may not be a perfect measure of sickness absenteeism or paid sick leave access. However, other fine-scale data was lacking, and a student’s ability to be absent is related to a parent’s ability to be home to care for the child, and differences in access to paid sick leave by SES have been related to student sickness absenteeism levels due to influenza [38–41]. To validate our findings, we compared our model estimates to previous estimates of influenza incidence ratios, stratified by poverty level. This is not a direct comparison, as previous studies present the incidence ratios for the entire population, not just for low SES individuals within those populations. Our results show more consistently high incidence ratios compared to the larger increases between poverty rates in prior studies. We attribute this to the incorporation of the measurement process into our models, which accounts for undersurveillance of low SES infection, whereas healthcare access and healthcare seeking differences may have missed low SES cases in prior studies. Ideally, data on respiratory infection of low SES individuals would be available at a fine spatial scale to more directly assess the validity of our models, but the lack of such a dataset highlights the need for future surveillance and data collection that focuses attention on lower SES populations.

Our work has several limitations. The network structure of our epidemiological model is based on one social survey from 2007 in Europe, and may be less representative of the United States today. Additionally, survey data was not collected for the SES of the contacts of survey participants, which required us to make assumptions which could affect our results about SES assortativity. Additional social contact data collection across the United States that accounts for SES heterogeneity would be useful for future studies given the large socioeconomic inequality in the country [42, 43]. In our spatial ecological model, we assume that disproportionate burden in low SES populations remains constant over influenza seasons. While there may be variation in the dynamics of ILI among low SES populations over time, time-varying data on our covariates is currently unavailable and we do expect socioeconomic factors and health behaviors to remain relatively consistent across seasons. Future work could focus on temporal variation in low SES ILI dynamics. Additionally, our spatial ecological model is only able to provide estimates for half of the states in the US, and the states are mostly on the coasts. This highlights the need for more data collection pertaining to low SES individuals, not only for epidemiological data, but also for a wide variety of other topics to provide covariate data and to create a better understanding of at-risk populations.

A main limitation of the spatial inferential model is the identifiability of separate effects. Here, we have identified possible associations in our model, but this is only the start to disentangling the factors that contribute to health inequities. The lack of significant association with the other incorporated covariates does not indicate that these are not important to inequities in influenza transmission in low SES populations. These impacts may be obscured by several issues. The covariate data may be impacted by its own biases, insufficient sample sizes, and other limitations. When ubiquitous systemic inequities go unaccounted for in data collection and processing, the signal of low SES individuals may be obscured. We aimed to counteract this by only using covariate data specific to low SES populations, but this was parsed out from data collected for the whole population that included demographic data, identifying potentially lower SES individuals. Next, there may be other factors relating to increased influenza transmission that may not be identifiable when focused on mechanistic explanations, and the model may not be able to parse synergistic factors. Our network epidemic model demonstrates that multiple factors of inequity can compound one another non-linearly, and statistically identifying individual effects remains a challenge due to lack of data and statistical limitations. Further attention to systemic inequities in health and epidemiology will be necessary to move this problem forward.

As the divide in health inequities, driven by income inequality and systemic racism, grows wider across the United States, we propose the use of infectious disease case studies to improve our understanding of this challenging problem. We suggest that we move beyond studies based on proxy measures such as income and education which may provide an incomplete picture, and dig into the mechanisms that may be at the root of inequities. Furthermore, we advocate for the prioritization of capabilities to detect and monitor outbreaks in at-risk populations so that we may prevent exacerbation of inequities. Addressing health inequities in respiratory infectious disease burden is an important step towards social justice in public health, and ignoring them promises to pose a serious threat to the entire population. Indeed, the damaging impacts of health inequities for respiratory infectious diseases have already been highlighted in the COVID-19 pandemic [44]. Our results suggest that (a) the effect of overlapping behavioral and social factors is synergistic and reducing this intersectionality can significantly reduce inequities; and (b) health disparities are expressed geographically and targeting public health efforts spatially may be an efficient use of resources to abate inequities. Further attention to the mechanisms and processes that lead to health inequities, and specifically health inequities that may be overlooked by our currently surveillance systems, will be important to identifying actionable steps to mitigate negative health outcomes in the future.

Materials and methods

In this study, we use (1) a mechanistic network epidemiological model to assess influenza transmission in the presence of individual-level socioeconomic status (SES)-based social and healthcare-based variation; and (2) an inferential spatial model to geographically localize influenza-like illness (ILI) burden among low SES populations in the presence of population-level variation in social and health indicators. Data and code for the implementation of these methods are available at [45].

Supporting information

Acknowledgments

We thank Håvard Rue for his development of and assistance with the R-INLA package.