479 - Using Insurance Status as a Proxy for Socioeconomic Status in Clinical Research: A Bias Analysis of Confounder Misclassification

Publication Number: 4470.479

Michael Monuteaux, Boston Children's Hospital, Boston, MA, United States; Catherine Coughlin, Boston Children's Hospital, Boston, MA, United States; Lukas Gaffney, Boston Children's Hospital, Boston, MA, United States; Joel H. Hudgins, Brown Emergency Medicine, Milton, MA, United States

Background: Socioeconomic status (SES) is often considered a confounding variable in observational pediatric research. The use of insurance status as a proxy for SES in multivariable analysis can result in confounder misclassification and biased results.

Objective: To estimate the magnitude and impact of confounder misclassification in multivariable regression analyses when using insurance status as a proxy for SES, compared to using direct measures of SES (income and education).

Design/Methods: We conducted a Monte Carlo simulation study. We used insurance status (public (low) vs private (high)) and a composite of income and education (low vs high) as the misclassified and gold standard (GS) confounder measurements, respectively. We used the nationally representative Medical Expenditure Panel Survey (2022) to estimate the test characteristics of insurance status (ie, the magnitude of misclassification). Using simulation datasets with a binary outcome, binary exposure, and both SES measurements, we compared odds ratios (ORs) with GS adjustment versus ORs with misclassified adjustment. The test characteristics were held constant, while sample size (n's ranging from 250 to 3000) and true population effect size (ES; ORs ranging from 1.5-5) were varied. 5000 repetitions were estimated per sample size-effect size combination. The primary outcomes were Percent Adjustment (% of confounding bias removed using the misclassified confounder, relative to total confounding bias) and Percent Correct (% of repetitions where both the GS and misclassified results were consistent with the population, showing a significantly increased risk for outcome among exposed).

Results: The sensitivity and specificity were 33% and 94%, respectively. The median Percent Adjustment across all n/ES combinations plateaued at 89%, increasing as the n and ES increased. However, the medians for all ES values ≤3.5 failed to reach 89%. Simulations with ES of 4 and 5 plateaued when the sample size reached 2000 and 750, respectively. The Percent Correct also increased as both the n and ES increased, plateauing at 100%. However, the Percent Correct ranged from 4-76% for ES≤2 (eg, only 4% of 5000 simulated datasets provided the correct effect estimate when n=200 and ES=1.5).

Conclusion(s): The impact of confounder misclassification on multivariable analyses resulting from the use of insurance status as a proxy for SES varies considerably with sample size and effect size. Researchers should be aware of this misclassification when using insurance to adjust for SES, especially in smaller studies.

Table 1: Median (interquartile range) Bias Removed Score across simulated samples (n=5,000 samples per simulation), stratified by sample size and effect size
Bias Removed Score = percent bias removed, defined as: (ORc - ORpa) / (ORc - ORca) x 100, where ORc = crude odds ratio; ORpa = partially adjusted odds ratio (ie from model with misclassified confounder); ORca = completely adjusted odds ratio (ie from model with gold standard confounder)
Scores>100 indicate misclassification over-correction (i.e., ORpa

Table 2: Proportion of simulated samples (n=5000 samples per simulation) where the exposure effect estimates from both gold standard and misclassified models is consistent with population (ie, significantly increased risk for outcome among exposed), stratified by sample size and effect size