January 1, 1989American Journal of Epidemiology
The Impact of Confounder Selection Criteria on Effect Estimation
Authors
Discussion
Loading...
Member takes
Implication
Simulation study demonstrates substantial bias from significance testing in variable selection, indicating change-in-estimate criteria provide superior effect estimation.
Key Points
- To evaluate how different statistical strategies for selecting confounders influence the bias and mean squared error of effect estimates in epidemiologic analyses.
- Conducted Monte Carlo simulations modeling various combinations of sample sizes, exposure prevalence, and degrees of confounding.
- Compared selection strategies including significance testing of confounder-outcome relationships at various alpha levels against change-in-estimate criteria.
- Conventional significance testing using strict alpha levels (such as 0.05) frequently fails to control confounding, introducing substantial bias into effect estimates.
- Using broader cutoff levels (such as alpha levels of 0.20 or higher) or direct change-in-estimate criteria minimizes mean squared error and residual confounding.
Cite This Study
MICKEY et al. (1989) studied this question.