Related Experiment Videos
Modeling and variable selection in epidemiologic analysis
1Division of Epidemiology, University of California, School of Public Health, Los Angeles 90024.
American Journal of Public Health
|March 1, 1989
Summary
This study addresses challenges in multivariate modeling for epidemiologic data, emphasizing improved model and variable selection methods. It warns against standard stepwise regression, recommending regression diagnostics and direct confounding estimation for valid results.
Area of Science:
- Epidemiology
- Biostatistics
- Statistical Modeling
Background:
- Multivariate modeling of epidemiologic data presents significant challenges.
- Model selection is a critical yet complex aspect of this process.
Purpose of the Study:
- To provide an overview of problems in multivariate modeling of epidemiologic data.
- To examine proposed solutions, focusing on model and variable selection.
- To offer guidance on best practices for robust statistical analysis.
Main Methods:
- Review of existing problems and proposed solutions in multivariate modeling.
- Examination of model selection processes, including model form, variable entry, and variable form.
- Discussion of regression diagnostic procedures and goodness-of-fit tests.
Main Results:
- Conventional stepwise regression algorithms can yield invalid estimates and tests.
- Regression diagnostics are crucial for selecting appropriate model and variable forms.
- Direct estimation of confounding is superior to significance-testing algorithms for variable selection.
Conclusions:
- Model and variable forms should be chosen using regression diagnostics alongside goodness-of-fit tests.
- Variable selection requires careful consideration beyond automated algorithms to ensure validity.
- Model assumptions must be rigorously evaluated against data and prior information before effect estimation.