11.07.2015 Views

2DkcTXceO

2DkcTXceO

2DkcTXceO

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

90 Passion for statisticsIn my experience, most students pick from a small list of suggestions bytheir chosen advisor. Indeed, that is the way I started my research. Norm hadsuggested the development of a sequential testing method based on partiallikelihood (Cox, 1975). I am sure Norm foresaw applications in clinical trials.He probably had some confidence I could do the hard math involved.Being a conscientious student, I started off on this problem but I was verysoon sidetracked onto related problems. I am not sure if I would recommendthis lack of focus to others. However, I do think you have to love what youare doing, as there is no other way to succeed in the long run. I did not lovethe problem he gave me. It seemed to be technically difficult without beingdeep, and not really a chance to grow intellectually.In the end I wrote a thesis on my own topic, about three steps removedfrom Norm Breslow’s proposed topic. I started by studying partial likelihood,which was then a very hot topic. But as I read the papers I asked myself this—whatisthejustificationforusingthispartiallikelihoodthingbeyonditsease of computation? A better focused student would have hewn to Norm’soriginal suggestion, but I was already falling off the track. Too many questionsin my mind.I assure you my independence was not derived from high confidence. Onthe contrary, no matter my age, I have always felt inferior to the best of mypeers. But I am also not good at following the lead of others — I guess I likemarching to the beat of my own drummer. It does not guarantee externalsuccess, but it gives me internal rewards.One risk with research on a hot topic is that you will be scooped. As itturns out, the justification, on an efficiency basis, for Cox’s partial likelihoodin the proportional hazards model was on Brad Efron’s research plate aboutthat time. So it was a lucky thing for me that I had already moved on to anolder and quieter topic.The reason was that the more I read about the efficiency of likelihoodmethods, the less that I felt like I understood the answers being given. Itall started with the classic paper by Neyman and Scott (1948) which demonstratedsevere issues with maximum likelihood when there were many nuisanceparameters and only a few parameters of interest. I read the papers that followedup on Neyman and Scott, working forward to the current time. I haveto say that I found the results to that time rather unsatisfying, except formodels in which there was a conditional likelihood that could be used.My early exposure to mixture models provided me with a new way tothink about consistency and efficiency in nuisance parameter problems, particularlyas it related to the use of conditional and partial likelihoods. I putthese models in a semiparametric framework, where the nuisance parameterswere themselves drawn from a completely unknown “mixing distribution.” Inretrospect, it seems that no matter how much I evolved in my interests, I wasstill drawing strength from that 1975 consulting project.My research was mostly self-directed because I had wandered away fromNorm’s proposed topic. I would report to Norm what I was working on, and

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!