The pitfalls of inappropriate methods

The use of digitalis has long been plagued by controversy. In the last few years, its efficacy in the prolonged treatment of heart failure among specific groups of patients has been questioned.' The current controversy stems from a claim by Moss et al.2 that the use of digitalis may increase mortality among infarction patients with heart failure and complex arrhythmias. Other investigators,3-5 on examining their own data, have found no evidence that digitalis is harmful (table 1). All of these claims, however, have been based not on the results of reliably controlled studies, but on inferences from data bases collected for other purposes. These investigators have used various statistical methods that are purported to overcome the biases arising when observational data are used to assess the effects of treatments. The fundamental problem, of course, is that not only may the treatment affect both the patient's condition and outcome, but conversely various features of the patient's condition may affect the choice of treatment. Features recorded before the institution of a treatment can at least be allowed for in analysis. But unrecorded features cannot be allowed for, and may instead lead to bias and to a mistaken conclusion that a treatment caused a patient's death when in fact it was used because the patient was at risk of death. The classic example is the old joke that, "The