Michael Chernick

19,793 reputation

21539

bio	website
	location
	age
visits	member for	1 year
	seen	Nov 14 '12 at 14:43
stats	profile views	6,106

I am a biostatistician at the Lankenau Institute for Medical Research where I work on lab experiments, clinical trials and other medical research. I have a PhD in Statistics from Stanford University. I have published books on bootstrap and biostatistics and have written or coauthored many articles in statistics, mathematics and medical journals. I am an ASA Fellow and am also a member of ENAR, the IMS, the Bernoulli Society and the Royal Statistical Society. I like teaching and mentoring and playing chess with my son (who usually beats me).

	19,793 reputation	bio	website		visits	member for	1 year
21539 badges		location			seen	Nov 14 '12 at 14:43

summary answers questions tags badges favorites bounties reputation activity

4,661 Actions

suggestions reviews revisions comments badges posts accepts all

Sep 17	comment	How do I find peaks in a dataset? All I am saying is that there could be situations where other approaches are better. It may be a parmetric approach like harmonic regression or some other method. Of course with any time series modeling I would be assessing the goodness of fit to the data and maybe even look at prediction accuracy depending on how I plan to apply the model.
Sep 17	comment	How do I find peaks in a dataset? @whuber You may be right but that is opinion. I am not making any claims about what is better. I especially have no idea what would work better for the OP in this question. It is my opinion that harmonic regression is a viable approach in time series analysis and there are situations where it would work better than a nonparametric approach. At this point I have read completely the answer you gave to the OP. My only criticism is that although you mention the advantages of your approach you do not acknowledge any disadvantages.
Sep 17	comment	How do I find peaks in a dataset? Regarding this time series example, I think that it is possible in practice to have a good idea that your time series is periodic and harmonic regression modeling would make sense. There is nothing wrong with taking a conservative nonparametric approach. All I am saying is that such an approach is not always better than harmonic regression and I think it is important to make that point. This argument can be made for many different problems, e.g. choosing a Weibull model for survival over the Kaplan-Meier estimate, fitting a gamma density to sample data instead of a kernel density estimate.
Sep 17	comment	How do I find peaks in a dataset? That is theory. In practice parametric models can be good approximations to reality. In that case the parametric estimate (say mle) is more efficient than the nonparametric estimate. Also the parametric confidence intervals will be better because they will be tighter. But many times you don't know how good the parametric model is for your example. In such cases you have to decide between conservativism (being safe) with the nonparametric approach or being bold (and possibly wrong) using the parametric approach.
Sep 17	comment	How do I find peaks in a dataset? Nonparametric methods have the advantage of not requiring modeling assumption, in this case no assumption of periodicity. My statement about parametric approaches being better than nonparametric approaches when the modeling assumptions hold should be very familiar to you. I don't need to argue about parametric assumptions never holding exactly. That is an opinion that I basically agree with. But I am talking about something like Pitman efficiency. Nonparametric estimates are not as efficient as parametric estimates when the model is "correct".
Sep 17	comment	How do I find peaks in a dataset? @whuber I said that you would fit the model then since the modle is a sum of sines and cosines the function is periodic the peaks occur when both the first derivative is zero and the the second derivative at the zero point is decreasing. That is what I meant when i said that you take the first and second derivative tests. Now you can solve to find all the solutions but if you have one peak the others are one period and multiple periods away from the solution you have. My point is not to claim any superiority of the method. I just want to point out that there is no free lunch.
Sep 17	answered	Formal definition of random assignment
Sep 17	answered	Interpretation of results in non-inferiority study
Sep 17	answered	How do I find peaks in a dataset?
Sep 17	revised	Present value model of a stock exchange: how to deal with zero dividends? added 1 characters in body
Sep 17	comment	Estimation/identification simple problem What is N.I.D.? Normal and identically distributed or normal and independent and identically distributed or something else??
Sep 17	answered	Likelihood vs conditional distribution for Bayesian analysis
Sep 17	accepted	Where did the term “learn a model” come from
Sep 17	comment	Where did the term “learn a model” come from +1 I am giving an upvote because although a little off the point of the question it is related and interesting and did get 3 upvotes. Keep it in.
Sep 17	comment	Present value model of a stock exchange: how to deal with zero dividends? It doesn't matter. Take a=0.1.
Sep 17	comment	Present value model of a stock exchange: how to deal with zero dividends? Yes. I didn't say I would recommend it. But you are free to pick your transformation. It will have the shape of a log and the transformed values depend on the choice of a, of course.
Sep 16	answered	Present value model of a stock exchange: how to deal with zero dividends?
Sep 16	answered	Weighted geometric mean vs weighted mean
Sep 16	comment	Adding training examples to Bayesian classifier reduces accuracy I think the answer to all your questions is yes possibly. To have a better idea would require knowing more details about your data.
Sep 16	comment	What is the canonical example which show advantage of robust linear regression over LS linear regression? What are the degrees of freedom for the t? If df is > 5 the tails are not very large and the difference from a standard normal are not that great. Why not try a Cauchy? Why mix the t with a normal? Try all errors Cauchy. Did you pick a model where the slope(s) of the regression curve (surface) is (are) large? You could also pick a few points that would have errors that would make them highly influential for the least square regression parameters or the bivariate correlation between a predictor and the response.