In randomized clinical trials of treatment for ischemic heart disease that compare medical with surgical treatment, many persons initially assigned to medical therapy eventually receive surgical intervention. For example, in the three major trials of bypass grafting for stable angina, crossover rates from medical to surgical therapy were approximately 25% at 5 years. For this reason, the classic intent-to-treat analyses have been criticized for their inability to evaluate the "true" effect of treatment. In this article we emphasize the concept of "initial treatment" as it applies to intent-to-treat analyses and examine four proposed alternative methods of analysis based on adherence with survival data from the Veterans Administration Cooperative Study to illustrate the concepts. The alternative methods include (1) censoring crossovers when treatment changes, (2) transferring crossovers from the original to the new treatment group when treatment changes, (3) excluding all crossovers from analysis, and (4) counting crossovers from the date of randomization in the treatment ultimately received group. We point out the biases attendant on analyses based on adherence and reaffirm the validity of intent-to-treat analysis.