Potential overestimation of the influence of unmeasured confounding in non-randomized studies: a comparison of bias-analysis methods Open Access

Brady, Sydney (Spring 2020)

Permanent URL: https://etd.library.emory.edu/concern/etds/2227mq765?locale=en


Bias is a common concern among non-randomized public health studies. Thus, it is important to accurately estimate the magnitude of bias affecting a study. There are several methods that have been used to quantify the degree of bias, such as calculating an E-value or producing bias-adjusted estimates of association. We compared different approaches for addressing uncontrolled confounding to estimates of association obtained from the systematic removal of potential confounders in an analysis of the association between hypertension and cardiovascular disease (CVD) mortality.

We studied 16,220 National Health and Nutrition Examination Study (NHANES III) participants who were successfully linked to the 2015 Public-use Linked Mortality File. The final study population excluded those with prior confirmed history of CVD. We fit Cox regression models to estimate the association between hypertension and CVD mortality. Confounders we considered were age, race/ethnicity, sex, BMI, alcohol use, diet, hypercholesterolemia, health insurance, education, diabetes, exercise, tobacco use, and household income. Each confounder assessment was fit with a separate Cox model.

The crude association between hypertension and CVD mortality was HR=10.83, whereas the fully-adjusted association was HR=1.03. The ratio of single confounder-adjusted HRs to fully adjusted HRs ranged from 0.96 to 2.48. The E-value for the fully-adjusted model was 1.22 and the relative risk due to confounding computed by bias analysis methods ranged from 0.93 to 2.40.

The crude association between hypertension and CVD mortality was substantially confounded. Bias adjustment methods consistently overestimated the strength of confounding by each variable treated independently, presumably because these methods do not account for the covariance between adjustment variables. Use of E-values as bounds on confounding overestimated the actual strength of confounding by all adjustment variables.

Table of Contents

1. Introduction (6-7)

2. Methods (8-13)

3. Results (14-17)

4. Discussion (18-19)

5. References (20-21)

About this Master's Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
Subfield / Discipline
  • English
Research Field
Committee Chair / Thesis Advisor
Last modified

Primary PDF

Supplemental Files