Readers should systematically assess methods used to identify, measure and analyze confounding in observational cohort studies

Objective — To describe techniques used to address confounding in published observational studies.

Study Design and Setting — A systematic literature review identified studies using administrative or registry data to investigate health effects of drug therapies. Studies published from January 2001 to December 2005 came from BMJ, New England Journal of Medicine, Lancet, Annals of Internal Medicine, and JAMA. A structured abstraction form was used to collect information about confounding.

Results — The search identified 29 studies. Twenty-two studies (76%) had 10,000 or more subjects and 18 (62%) used a mortality outcome. None mentioned use of a literature search to identify confounders, however, 28 (97%) listed confounders included, and 26 (90%) listed confounders not included in the study. Eighteen (62.1%) discussed the validity of confounder data. Most (22, or 76%) studies included a table with the distribution of confounders but none used effect size to assess imbalance between comparison groups. Almost all studies used regression techniques (28, or 97%); fewer used stratification (16, or 55%) or matching (four, or 14%) to address confounding. Eleven (40%) studies discussed sensitivity analyses.

Conclusion — Published cohort studies routinely include a list of potential confounders but there is room for improvement in confounder identification, measurement, and analysis.