Oscar Kempthorne: From Observation to Inference, 1991 part 4

I think you've given expert witnesses a court cases i also sent you i was involved with so threatened Shockley I see friend you know the the great physicist you know Nobel and so on and I think it's true that leaked is an ailment of Nobel Prize winners yes they think they can understand anything they can analyze anything and shocked me was surely doing this so you know this he was thinking that if you see correlation you can infer causation and his thesis was something like this that now shall I put it that amongst bite people invite people the proportion of white ancestry they such and such the portion of black ancestry is such an such a very environment lower so you know white ancestry causes high intelligence and he was riding that horse so they had this court case in atlanta and so I they called me in as an expert witness against chocolate my expert witness Singh was primarily to the effect that Shockley was incompetent so anyway curiously enough he won his case I think but he received damages of wonder so don't get much out of it on so yeah I'd like to ask you about new aspects of statistical genetics and beyond my metrics there's a subject that has generated a lot of interest called biotechnology what do you think istics what role do you think statistics has to play in this subject girl I think statistics will have a massive role which I believe the biotechnologists do not understand they the ideas are the ideas of so I say classical genetics 1920 of nineteen thirty that you one looked at inheritance and then as I put it now you look at sons offspring and parents and then they shared certain ailments who deficiencies well and so this led on certain especial attributes eye color or whatever and this led to sort of gene mapping mm-hmm they you know and was the techniques for doing this were very limited but with Watson quick then and then massive development since people can look at the whole chromosome or set of chromosomes and then the someone discovered scissors the things that call scissors to chop this up into bits and then there are techniques these differ in molecular weight and so one can put this these chocolate chopped up stuff in the whatever the word is some sort of jelly an electric field starts gel electrophoresis and it won't can just sort out this stuff then one can pick it up so to speak yes I multiply it it's the latest thing which I don't understand there's a a great big advance called polymerase chain reaction yes in which one can get a little piece of DNA and one to multiply it they untied this so one has enough stuff to do all sorts of analyses surely variability in all these process so this is and when I see variability i think immediately of statistics oh yeah you think the same that's right because the year it seems to me that the people who are working on this don't really understand the complexity because i think i'm right i don't know if I've lost a few digits I think in humans there are two billion nucleotide pairs yes so and you know it's either seed CG is one possibility or 80 is the other whatever doesn't matter what the names are so this is like a factorial experiment with two bitten factors it is so you know how does one find out this and but you know apparently they do find they find out some things they've finally nailed down cystic fibrosis for instance uh-huh so so the signals are fairly strong in in the maps that they're making so you know they are finding out to some things goodness so can I move on now to a very important topic and that is statistical inference statistical inference I think it's appropriate that this last part of our conversation be about inference can you tell us what inference is and whether statistics is approaching the problem in the right way well inference I suppose tackles questions of this will I be alive in the year 2000 now there is uncertainty on that and there's variability so one way of expressing this variability is to bring in probability as frequency calculus mm-hmm so so there we are you see now because in order to use this frequency calculus one has to have a statistical model they infect a stochastic model now what goes on in the world is that the stochastic model is produced out of nowhere yes no one is he the only way to get it appropriate statistical model or stochastic model is by data analysis mm-hmm but you know the only people who do data analysis are so quote dirty unquote applied statisticians oh but so well I don't actually think that's all that fear it's not affair no I don't think that's a that's a fair comment I think many things go into the modeling and a lot of prior thoughts and conceptions about the physics chemistry or biology of the problem going to it and in fact a bayesian would say when they build their prize they don't build them from completely from from no knowledge at all in fact the the mechanism behind the process goes into the building of a prior it's very important for it well you know perhaps you are right I guess you will forgive my certain skepticism or I said I've you know I've listened to the I've listened to the big guns you know lindley and I've listened to savage the foundations of statistics yes which goes is nothing about the foundations of sticks at all you know there's nothing about data analysis and huh models amount of thin air the umpires well you got priors by introspection now that is that isn't much of e that is not much of an approach so to speak well everybody introspect right I think you get priors from your experience I think perhaps a baby is born with a flat prior and then gradually as that baby grows the experience accumulate accumulate and knowledge about processes around them forth the priors to become somewhat Pete oh well you see this is the this is the problem with statistics that it does not approach is matter of inference properly you know I i bought the encyclopedia of philosophy and so look there's a ton of stuff in there but these philosophers are absolutely useless far as I'm concerned well there's no point well they don't address any real problems okay right so let me follow up on bayesian statistics given that we're there results from bayesian statistics seem to differ from results from classical statistics in applications where there is little available information about the process on the study perhaps a lot of your work has been in data rich situations where one gets a similar answer regardless of the prior chosen do you think that's easy because I guess I object to the process of using a prior I think you once said you have an axiom in statistics that one should not be evasion one should not be a button okay so you know so it's a philosophical objection well I should it's philosophical but but you know I think the field of statistics is failing by not addressing the basic questions and they and you know I read these I read the basins and you know they applier my prior comes in out of this he and Jeffries you know how did bang and trying to book to find invariant priors you know good parts from a mathematical point of view and he failed and so the process is wrong