Oscar Kempthorne: From Observation to Inference, 1991 part 5

but how would you perform inference based on just 2 or 3 observations I went now too about very data poor situations well I guess I haha I would give up on it but you have you would then ask the scientists where those two or three data came from oh yeah one often deals particularly in an exploration oil exploration where one datum would cost an enormous amount of money a million dollars and often deals with situations where only two or three data are available and they certainly don't want to give up on it because they'd spent three million dollars sure but then you see in that situation there is a huge richness of physical scientist knowledge right which has brought to bear on it so the I guess I suspect that the use of the Bayesian process really doesn't matter you know but there is no doubt that you know Maya in my opinion statistics is failing and is failing rather miserably have you thought about how it might recover haha I've been trying for the last 30 years but I've known much success I think I think the end thing is that G what we are hoping for we seem to be hoping for it is a deductive theory mm-hmm but you know this g one cannot some others something inherently wrong in thinking about a deductive theory for inductive processes because in that case if one could have that then everything would be deduction ok so you know they're all sorts of and complexities I'd like to broaden this conversation a little into inference in general your third book probability statistics and data analysis by Kempthorne and folks treated a number of inference issues from a fresh perspective what contribution to that book made to improve I don't know if it really makes anything but the idea that is pursued there is to make a differentiation between Neyman Pearson statistics which is the standard thing all the theoretical profession make a distinction between the Neyman Pearson and I guess data analysis inference which I mean was somewhat fish Aryan Fisher was the I guess the the greatest day our analyst of all time of all time I guess okay but I do well one topic I know you've addressed is that our probability models often assume data can take any value on a continuum oh yes and imprecision of measurement makes this assumption unrealistic what effect does this have on inference and what can we do about it well I don't know this you know I've been hopped-up bugged about this for 30 years and know people say all Oh Kempthorne is off riding his horse again I'm still writing it right because well take for instance fiducial inference which fisher fought apparently was the answer but producing oil inference cannot be done with discrete data so if all data are discrete then I do soils is out mm-hmm but you know I'd like to talk a little now about hypothesis testing and significance tests oh well can-can I think this relates to what you were saying about name and pearson inference can you contrast the two and tell which one we should be using well the fisher and karl pearson before produce significance tests okay and the great one was chi-squared goodness of fit well in there are all sorts of uncertainties and ambiguities about this so Mammon and pearson tried to fix it and in order to fix it they produced a theory of testing hypotheses which is really a theory of accept reject rules so it seems to me and I that this theory of accept reject rules just doesn't really address the problem of inference but then because neither does neither does the fish Aryan theory you know the models in the in the well it's not going to confuse issues here the the fishery and theory sort of produces a model but then it it tries to obtain a rash and measure of a rational degree of reluctance towards a hypothesis mm-hmm and it seems to me that is really what the game is about rather than you know accepting or rejecting mm-hmm so well a bayesian would put a prior on the hypothesis yeah and use the posterior probability oh yes the hypothesis that is true but where does he where does he get the the model parametric model and where does he get the fire now if you couldn't nail those things down then what would be in business okay I think introspection is probably not the whole answer and we have we've talked about that just three or four minutes ago well we've come to a the end of what I think is a very interesting hour of discussion with you Kemp it's been great fun and enormously interesting I know that your colleagues friends and relatives are going to take pleasure from seeing vintage Kempthorne but equally importantly those who have only read your written words will surely enjoy and benefit from the extra dimensions that we've seen today Thank You Oscar Kempthorne well and thank you know aggressive for working on this and making this thing go because on my own I would not have been able to make it go so thank you you you