Related Experiment Videos
Cautions on the reanalysis of epidemiologic databases.
J E Michalek1, D Mihalko, R C Tripathi
1Epidemiology Division, USAF School of Aerospace Medicine, Brooks AFB, TX 78235-5301.
Statistics in Medicine
|June 1, 1989
Summary
Reanalyzing medical databases for new questions can lead to biased results if not done carefully. Adjusting for the original study design and testing for data interactions can significantly reduce this bias.
Area of Science:
- Statistics
- Epidemiology
- Medical Data Analysis
Background:
- Medical databases are frequently reanalyzed in statistical and epidemiologic research.
- Reanalyses can be confirmatory or explore new research questions not originally intended.
Purpose of the Study:
- To demonstrate that uncritical reanalysis of designed studies can yield biased results.
- To analytically show how ignoring design variables in reanalysis introduces bias.
Main Methods:
- Utilized a log-linear model for analytical demonstration.
- Expressed bias as a function of second- and third-order interactions.
- Investigated the impact of preliminary testing for interactions.
Main Results:
- Reanalysis of designed studies without adjusting for the design variable produces biased outcomes.
- The magnitude of bias is directly related to second- and third-order data interactions.
- Preliminary testing for interactions effectively reduces reanalysis bias.
Conclusions:
- Uncritical reanalysis of medical databases poses a risk of biased findings.
- Statistical adjustment for study design and interaction testing are crucial for valid reanalysis.