It is an honor to be a discussant to the Morris Hansen Lecture, and a pleasure to be discussing Don Rubin's talk. Dr. Rubin has clarified over the years many of the deep issues relating to causal inference. Let me start with a story. About 20 years ago when I was teaching at UCLA, I was eating breakfast one morning at my kitchen table, and my two-and-a-half year-old daughter was in the next room, lying on her back and kicking the wall with her feet. I told her to stop, which she did for a few seconds, and then be gan again. I told her to stop again, and that I really meant it. The kicking stopped for a longer period this time, maybe 30 seconds, and then started up again. Just then the Whittier-Narrows earthquake hit, 5.9 on the Richter scale. Our 50-year-old house started shaking like crazy. As I was running into the next room to get my daughter, I ran into her running into the kitchen screaming sorry, Daddy, I'm sorry. I didn't mean to do it! Which brings me to my first point: causal inference can be tricky. Causal inference can be tricky not just for small chil dren, but for epidemiologists and biostatisticians, too. As an example, consider hormone-replacement therapy for postmenopausal women. Dozens of observational studies (including case-control studies and cohort stud ies) had suggested a 40-50% reduction in coronary heart disease (Stampfer and Colditz, 1991). However, the recently reported results of the Women's Health Initiative trial demonstrated that the treatment had an elevated incidence of coronary heart disease (Manson et al., 2003). Now the statisticians who worked on these epidemiologic studies thought they were mak ing a valid causal inference. In fact, many women took estrogen replacement therapy partly because they be lieved that it would offer cardiovascular benefits. How