Tobacco Documents Online

!.': 2w~.~.3# a

he trained statistician has social \scientists and psychologists are concerned. versations with the users of the methods. researchers have had no control over what they are doing. Without . .. . . . . _ . . .. ... . ... ... . .. . .. . . ..-_., i ..~.i. }~~5~ to an assortment of statistical methods in an effort to have the computer come ,.. . . .,. „ .. ..y ~~ --.. . . .. ,,. :.~ . . ...' .. . ~. . :, ; . . i.... ...~-. : :.~ . . ~...~.. .. . . ~. . . . :' , ... . -.; . ..~.~ ~ ,.. .~. ~ ... ~ 4, `. .a specific hypothesis to prove or disprove, they have subjected their rnaterial up with one or more hypotheses proved. It is really a good thing that they Someone had to do it. them and the reader what they have to face up to. When they removed what appeared to be the critical variables, they came .up with one set of values. reader, how fickle variables can be in a contrived study of this kind. When they put them back, with another. „ This has shown them, as well as the deserves admiration. The possibility of the massive complicated mathema- If they had started with a modest model they could have studied the ana- tomy of their findings so to speak. The heroic proportions of the attempt tics of which a giant computer is capable, make.it like a child who has matured w zt`,7„ too fast: The user has to be aware of its pitfalls as well as its advantages.

.~r him to the correct population with minimum error, usually on the basis o ,; F_., .. . . . . . . :• ,-.. . -, ~.;;,:•, ma have corne from one p p , , ~ . . . ,may of several o ulations, the problem is to allocate _ , . . are brought to bear upon, problems of classification. Given that an indi,vidual medical integration by experience. Discriminant analysis is the b , . ,. multiple measurernents on the individual and a prior set of similar measure-`- ments on individuals whose origin is ltnown. Discriminant :T into two different classes as a rule, not into multiple classes. There is an ~ effort to carry it to multipie classes in this study, but it becomes less reliable ~~ The technique has not yet been~ develioped for this procedure. 1005136147 . .. =t~ No statistician has been involved in this study. It shows in everything that is said and done. Nor have they had proper advice from medical, consultants.' Multivariate analysis is a powerful tool. Cornfield's several papers on its use are well done, and demonstrate that you can't play fast and loose with this-.technique. Because many variables are available is not in itself sufficient reason to apply multivariate analyses techniques.

,of cotton will reflect a long linear increase ~~ - y one of several met o s. Then you have a cyciical variation Index eth d m o s n€= ,.~ are available to remove that. Then you have a seasonal trend which you . f'.... S~ U~ ~tN Then you have your natural variation inherent in the price of cotton They are all causal, or parts of the cause It is not ossibl t . p e o en say one is secular and one is' causal. W What the authors are trying to show is that the sudden availability of cigarettes at the first world war made an entirely different cohort. But they M are not using cohort methods, nor have they clarified their secular theory. Factor analysi~s is a method that qualified statisticians will not touch. Cattell, the psychologist, used it guardedly and with many warnings as to its

'S At 'r "1'' . . . - . . . . ~ .. .. ti- ~. ;,The discussion oftossing out outliers is another case in point. 10-1, Supplement 4, August 1968: , ~, •,_ . `:'In these cases the rejection of extreme values is often recommended, and it may often happen that gross errors are thus rejected. As a statis ~tical measure, however, the rejection of observations is too crude to be :defended; and unless there are other reasons for rejection thanrnere divergence from the majori~ty, it would be more philosophical to accept ~nar, these extreme values, not as gross errors, but as indications that the s distribution of errors is not normal" (Fisher, R. A. :"On the Mathema- 5.,. .tical Foundations of Theoretical Statistics", Phil. Trans. Roy. Soc. A RULE REGARDING REJECTION .rK -Many of the methods developed by the lateSir Ronald Fisher were based upon the gaussian ("normal") curve, and in view of that it is •,< :;.,h interesting to read what he said more than 45 years ago about "the prac-_ -.'tical difficulty which observers often find, that a few extreme observations appear to dominate the mean. " His comment was as follows: up a discussion of the problem with the emphatic words: "A point is never to be excluded on statistical grounds alone" (Statistical Adjustment of Data, New York, Wiley; London, Chapman and Hall, 1943, p. 171). About 20 years after Fisher wrote this, W. Edwards Deming summed to decide when we ought to seek for something additional to the factors The essence of Deming's discussion was that we ought not to -seek a', cutoff point in order to discard an observation but as a criterion by which

T - ~ ... , z~ ..-•;~. ~ ~y-. into the mathematics of the powerful tools available for his use. . . . . - r< ..,.Ail.r, These men are essentiall social s' t t th h e h . ; ~ .,.. ... .. ... . •.. . y~ .. .ki . r qturements are met in this study. a ~sf~~ None of these three e y ci n is oug ~ one as an M. D. degree . . . _ .. y t•..'~,, and ose an Throat Service of the V. A. Outpatient Clinic heads the E e Ea N in Boston h h b , as een involved in solving the methodological problems inherent in I .,,," ;+„A ;.,-,, , E..a: __ ..,,_ _ . . . .. - . . . nist or other clinicians in the picture. There is no mention of a modest work-'Ik . . . . `~. . .A . ~ ,~ . ~. li.~~.,... .. . . . . - . ~ _ .. - ,. _. ., ... ~ ' . . ' . ~ ' i{l.,Y~~.,'`~'l..u •~.1~-*,~..'y j2. .able constrained hypothesis from which to. start with, just a few variables, as .~ - - ,.,..c. _ _ _. _ . _ .. , ., .c?K~"*'sAN .~.~.5, .. ~..i'~L.' :i~ ~~ r1•a +~K,4J.)SJC+~!c '1~~@'.'6L Y.. . a. ~ 4 ... ~ , •A'~+. ~ A~i'-'M~LM Ci:~.v~+ K'~~.{'.~ ~ ~ - - • - 3 •. l~ , . . ...'".t:.'.c{. ` A qualified and modest computer expert who will seek and depend on li"'j, the expertise of the above two consultants before he tosses dozens ~ . ;~. Y .1. :~ :,, , ° i. • ,- : . ., '~ , -, .....~.y .~ .-'' ,' "-. ;~:.., r !,~_t~, lr s~}r x.> ~r.~,''.•r .ui,j~ techniques ad-lib without'any knowledge of the relative imraortances of variables into the machine and attacks them by applying powerful.; . . . . .. .. . . . , l .. . . , . .. .: .. . . ..:_y. . . . of the variables themselves, nor of their interrelationships `'&~ >• .:: • ~ .;~~~_~~

---