Reporting on Statistical Methods To Adjust for Confounding

Summaries for Patients are a service provided by Annals to help patients better understand the complicated and often mystifying language of modern medicine.

Summaries for Patients are presented for informational purposes only. These summaries are not a substitute for advice from your own medical provider. If you have questions about this material, or need medical advice about your own health or situation, please contact your physician. The summaries may be reproduced for not-for-profit educational purposes only. Any other uses must be approved by the American College of Physicians-American Society of Internal Medicine.

The summary below is from the full report titled "Reporting on Statistical Methods To Adjust for Confounding: A Cross-Sectional Survey." It is in the 15 January 2002 issue of Annals of Internal Medicine (volume 136, pages 122-126). The authors are M Müllner, H Matthews, and DG Altman.

What is the problem and what is known about it so far?

Research projects often try to show that a condition (such as age) has a strong association with a health care outcome (such as death from heart disease). Confounding occurs when the health care outcome is influenced by factors other than the one that a researcher thinks is the most important. These factors are called confounders. For example, suppose that a researcher was studying whether pregnant women who drink coffee are more likely to have small babies than are women who do not drink coffee. Now suppose that the researcher found that the babies of coffee drinkers weighed less than the babies of women who did not drink coffee. However, the researcher failed to consider the possibility that women who drink a lot of coffee are also more likely to drink alcohol and use tobacco. In this case, alcohol and tobacco are potential confounders. In fact, drinking alcohol and smoking can lead to babies who weigh less at birth. If these behaviors are more common in coffee drinkers than in non–coffee drinkers, researchers say that these factors may "confound" the relationship between coffee drinking and birth weight. An analysis that uses special statistical methods to account for potential confounders might conclude that coffee drinking itself does not result in low birth weight. Adjustment for confounding is very important because it can prevent people from drawing incorrect conclusions about what information is important in patient care.

Why did the researchers do this particular study?

To determine how frequently researchers describe how they adjusted for confounders in their analysis.

Who was studied?

The researchers studied 537 original research articles published in 34 medical journals in January 1998.

How was the study done?

The researchers examined each article and collected information about whether and how it reported the methods that the original researchers used to adjust for confounding.

What did the researchers find?

Of the 537 articles examined, 169 stated that the researchers adjusted for confounding. Many of the 169 articles that adjusted for confounding provided too little information to decide whether the researchers had done so correctly. Articles that had an author affiliated with a department of statistics, epidemiology, or public health were much more likely to provide a full description of their adjustment for confounding.

What were the limitations of the study?

The study included only articles from a single month in one year and sampled only 34 journals.

What are the implications of the study?

Researchers frequently do not publish enough detail about the methods they use to adjust for confounders. Researchers should provide adequate information about confounding, and journal editors should insist that they do.