Addressing Selection Biases within Electronic Health Record Data for Estimation of Diabetes Prevalence among New York City Young Adults: A Cross-Sectional Study

Abstract

Introduction: There is growing interest in using electronic health records (EHRs) for chronic disease surveillance. However, these data are convenience samples of in-care individuals, which are not representative of target populations for public health surveillance, generally defined, for the relevant period, as resident populations within city, state, or other jurisdictions. We focus on using EHR data for estimation of diabetes prevalence among young adults in New York City, as rising diabetes burden in younger ages call for better surveillance capacity.

Methods: This article applies common nonprobability sampling methods, including raking, post-stratification, and multilevel regression with post-stratification, to real and simulated data for the cross-sectional estimation of diabetes prevalence among those aged 18-44 years. Within real data analyses, we externally validate city- and neighborhood-level EHR-based estimates to gold-standard estimates from a local health survey. Within data simulations, we probe the extent to which residual biases remain when selection into the EHR sample is non-ignorable.

Results: Within the real data analyses, these methods reduced the impact of selection biases in the citywide prevalence estimate compared to gold standard. Residual biases remained at the neighborhood-level, where prevalence tended to be overestimated, especially in neighborhoods where a higher proportion of residents were captured in the sample. Simulation results demonstrated these methods may be sufficient, except when selection into the EHR is non-ignorable, depending on unmeasured factors or on diabetes status.

Conclusions: While EHRs offer potential to innovate on chronic disease surveillance, care is needed when estimating prevalence for small geographies or when selection is non-ignorable.

Keywords: diabetes mellitus; electronic health records; prevalence; selection bias; surveillance.

Figure 1. Simulation study directed acyclic graph with baseline OR associations. Observed diabetes within those selected into the EHR sample; scenario 1 (orange): modified the level of misclassification of the auxiliary variable W compared with the unobserved variable U at levels equivalent to 10%, 30%, 50%, 70% and 90% misclassification; scenario 2 (purple): modified the association between diabetes and selection at OR levels of 0.33, 0.67, 1.0, 1.5 and 3.0. DM, diabetes mellitus; EHR, electronic health record; HIS, Hispanic; NHB, non-Hispanic Black; OTH, Other race.

Figure 2. Characterisation of the NYU Langone patient sample and comparison of NYU EHR-based to gold standard diabetes prevalence estimates for young adults aged 18–44 years by New York City PUMA neighbourhood. (A) Proportion of general population captured within the EHR sample by NYC PUMA, calculated by dividing NYU Langone patient counts by the total NYC PUMA population estimates from the American Community Survey 2019 5-year data, obtained through IPUMS USA. (B) Comparison of NYU EHR-based to gold standard diabetes prevalence estimates. Each point represents a PUMA neighbourhood. EHR estimates are defined using NYU Langone Health 2019 data. The gold standard estimate is defined using NYC CHS 2015–2020 data. (C) Comparison of relative bias in NYU EHR-based prevalence estimates versus proportion of the general population captured within the EHR sample. Relative bias is calculated as the per cent change between the gold standard and EHR-based prevalence estimate for each NYC PUMA neighbourhood. CHS, Community Health Survey; EHR, electronic health record; MLRP, multilevel regression with post-stratification; NYC, New York City; NYU, NYU Langone Health; PUMA, Public Use Microdata Area.

Figure 3. Mean relative bias in the EHR-based estimates versus true diabetes prevalence by simulation scenario. Error bars represent SD in mean relative bias across simulations. (A) Scenario 1 modified the level of misclassification of the auxiliary variable W compared with the unobserved variable U; (B) scenario 2 modified the association between diabetes and selection (OR_DM). EHR, electronic health record; MLRP, multilevel regression with post-stratification.