Letters Section Editor: Stephen J. Lurie,
MD, PhD, Senior Editor.
In Reply: In response to Dr Ahmed and colleagues,
it is important to note that, when generating the propensity scores on which
patients were matched, we chose as the outcome measure in the logistic regression
whether or not a patient was transfused, and we therefore selected variables
that might be related to that event. We were not selecting variables that
were predictors of mortality. The propensity score was used to remove, or
to some extent control for, confounding and selection biases for transfusion.
The propensity score for transfusion was determined without regard to mortality.
Variables were not weighted in any manner. Logistic regression yields a predicted
probability of group membership (in our case, transfused and nontransfused)
based on the best combination of predictors. More detailed explanations of
propensity scoring and its application in observational studies can be found
Jean-Louis Vincent. Blood Transfusions and Mortality Among Critically Ill Patients. JAMA. 2003;289(10):1242–1243. doi:10.1001/jama.289.10.1242-a