Mortality and Poststroke Depression
To the Editor: The conclusions of Dr. Jorge and colleagues regarding the association between use of antidepressants and all-cause mortality following stroke are suspect because of the possibility of uncontrolled confounding. The authors reported that the association between antidepressant use and mortality remained significant after they controlled for age, stroke type, diabetes mellitus, and relapsing depression in a logistic regression model. However, there are other variables (reported in their Table 1), such as marital status, which were associated with treatment group and mortality and might therefore confound the association between treatment group and mortality. In addition, there are variables that are associated with mortality (reported in their Table 4), such as obesity and atrial fibrillation, which might confound the association if they were also associated with treatment group. The only variables that were included in the logistic model were those that were associated with treatment group and outcome (mortality) at the p<0.05 level.
However, that is an inappropriate criterion to use in assessing confounding (1–3). Confounding depends only on the magnitude of the associations between a covariate and both treatment and outcome, not on whether these associations are due to chance (2, 3). Further, a p value from a statistical test reflects both the strength of an association as well as the sample size, but confounding is a type of bias that is related to only the strength of the associations between the confounder and the outcome and the confounder and the treatment (2). When the size of the sample is small, as in the study by Dr. Jorge and colleagues, differences in baseline characteristics, such as marital status, might not be statistically significant but might still be large enough to exert a substantial effect on the outcome.
The appropriate way to assess confounding is to use the “change in estimate” criterion: potential confounders are included in the model if their deletion results in a substantial change in the magnitude of the association between the treatment and the outcome (1, 3). Since the association between antidepressant use and mortality was only marginally significant (p<0.03) in the multivariate model with four confounders, my concern about the effects of other potential confounders is not merely an academic quibble. It would be useful if Dr. Jorge and colleagues had reported both the crude odds ratios and the odds ratios adjusted for all confounders (and their 95% confidence intervals) by using the more appropriate “change in estimate” method for determining which variables should be included in the multiple logistic regression model.
1. Sonis J: A closer look at confounding. Fam Med 1998; 30:584–588Medline, Google Scholar
2. Rothman KJ, Greenland S: Modern Epidemiology, 2nd ed. Philadelphia, Lippincott-Raven, 1998, pp 256–258Google Scholar
3. Savitz DA: Interpreting Epidemiologic Evidence. New York, Oxford University Press, 2003, p 144Google Scholar