Socioeconomic bias in influenza surveillance

Samuel V. Scarpino, James G. Scott, Rosalind M. Eggo, Bruce Clements, Nedialko B. Dimitrov, Lauren Ancel MeyersPLOS Computational Biology 2020 · DOI

TL;DR

This paper reveals a critical blind spot in current US influenza surveillance systems: they perform poorly in low-income communities, leading to less accurate predictions for those most at risk. By integrating traditional and novel data sources, the authors demonstrate that while overall surveillance can be improved, the socioeconomic disparity in data quality remains a significant challenge.

What Problem Does This Paper Address?

Accurate influenza surveillance is crucial for effective public health interventions, especially for individuals in low socioeconomic brackets who are at higher risk of severe complications. However, the primary US surveillance system, ILINet, relies on outpatient healthcare providers, which are often less accessible to lower-income populations. This creates a potential bias, where surveillance data may not adequately represent the disease burden in these vulnerable communities. The paper aims to evaluate the adequacy of existing and emerging surveillance data sources (ILINet, BioSense 2.0, and Google Flu Trends) for forecasting influenza hospitalizations across different socioeconomic strata, specifically focusing on zip codes categorized by poverty levels.

Methods

The researchers developed a flexible statistical framework to integrate multiple surveillance data sources for short-term influenza hospitalization predictions. They employed generalized additive models (GAMs) using a Poisson distribution to model hospitalization counts. Key to their approach was the use of B-splines with a lasso penalty for regularization, allowing them to model potentially nonlinear relationships between predictors and outcomes while preventing overfitting.

They partitioned zip codes into four poverty quartiles and fitted separate models for each. To evaluate predictive performance, they used out-of-sample Root Mean Squared Error (ORMSE), normalized by population size. To assess the statistical significance of performance differences across poverty groups, they conducted a permutation test. They also explored the synchrony of influenza hospitalization patterns within poverty quartiles to rule out asynchrony as a cause for poor performance in disadvantaged areas.

Datasets

The study utilized data primarily from six counties in the Dallas-Fort Worth, Texas region between 2007 and 2012. The datasets included:

  • Weekly BioSense 2.0 data: Extracted from an online repository, representing percent of emergency department (ED) visits for Upper Respiratory Infection (URI). County-level aggregates were used.
  • ILINet data: Weekly estimates of influenza-like illness (ILI) cases reported by primary care providers, provided by the Texas Department of State Health Services. County-level aggregates were used.
  • Google Flu Trends (GFT) data: Daily estimates of ILI patients per 100,000 people based on internet search terms. Historical data from Google.org was used, including state-level and city-level time series.
  • Hospital discharge records: Filtered for influenza-related principal diagnostic codes (ICD-9 487.*) from the Texas Health Care Information Collection (THCIC), aggregated by week and patient zip code.

Demographic data for poverty and age distributions were sourced from the 2010 Census and the 2011 American Community Survey.

Key Insights and Ideas

  • The "Blind Spot" in Surveillance: The most striking finding is that zip codes in the highest poverty quartile are a significant blind spot for ILINet, and this gap is not fully ameliorated by integrating next-generation data sources like BioSense 2.0 and Google Flu Trends.
  • Socioeconomic Disparity in Prediction Accuracy: Predictive models consistently performed best in the most affluent zip codes and worst in the most impoverished ones, regardless of the data sources used. This highlights a direct socioeconomic bias in the accuracy of influenza surveillance.
  • Synchrony Doesn't Explain the Gap: Contrary to a potential hypothesis, the poorest zip codes exhibited more synchrony in influenza hospitalization patterns than less impoverished quartiles. This suggests the issue is not due to disparate timing of outbreaks but rather data quality or coverage.
  • A New Socioeconomic Inequity: The study identifies a previously unrecognized socioeconomic inequity: a reduced capability to detect and monitor outbreaks in at-risk populations, which hinders effective public health interventions.
  • Beyond Timeliness and Accuracy: While new data sources promise improved timeliness and accuracy, they may not inherently fix the critical data gaps in underserved communities.

Key Results

  • Models made the best predictions in the most affluent 25% of zip codes (0-7.5% poverty) and the worst predictions in the most impoverished 25% (21.2-48.1% poverty).
  • These differences in prediction errors between the highest and lowest poverty quartiles were statistically significant (P < 0.0001).
  • The study found a positive correlation between poverty and influenza hospitalization rates in populations under age 65, consistent with existing literature.
  • While BioSense 2.0 showed slightly better performance across poverty quartiles, it did not eliminate the disparity.

Limitations and Future Work

The authors acknowledge several limitations:

  • Geographic Restriction: The study was limited to the Dallas-Fort Worth region, and findings may not generalize to the entire USA.
  • Proxy Data: BioSense 2.0 data used Upper Respiratory Infection (URI) as a proxy for influenza, as influenza-specific diagnoses were not accessible. While expected to improve predictions, it might not fully close the poverty gap.
  • Unconsidered Data Sources: Other potentially valuable data sources, such as public health laboratory data, school absenteeism records, or social media data, were not included.
  • Model Selection: The use of a lasso penalty prevented the application of standard model selection criteria (like AIC) for comparing model performance.
  • Individual-Level Data: The lack of individual-level socioeconomic data for ILINet, BioSense 2.0, and GFT prevented direct assessment of underrepresentation.

Future work could involve:

  • Testing Other Data Sources: Exploring data sources that might offer better coverage of at-risk populations.
  • Subsampling Strategies: Developing methods for subsampling data to improve predictive performance in low-income areas, especially for internet-derived signals.
  • Adapting Surveillance Frameworks: Adapting surveillance design frameworks to specifically address and remedy critical data gaps and biases in underserved communities.
  • Investigating Causes of Disparity: Further research into the underlying factors driving disparities in sanitation, crowding, vaccine coverage, or prevalence of underlying conditions.

Summary updated

BibTeX
@article{scarpino2020socioeconomic,
  author = {Samuel V. Scarpino and James G. Scott and Rosalind M. Eggo and Bruce Clements and Nedialko B. Dimitrov and Lauren Ancel Meyers},
  title = {{Socioeconomic bias in influenza surveillance}},
  journal = {PLOS Computational Biology},
  year = {2020},
  doi = {10.1371/journal.pcbi.1007941},
}
Figure