6 INTRODUCTION 



one could make certain that there were no hereditary differences between 

 the controls (the ones getting the complete diet) and the experimentals 

 (the ones getting the thiamme-deficient diet). One might postulate that 

 the thiamine-hee diet does not have as attracive a taste as the one with 

 thiamine, and the experimental animals simply eat less food, fail to 

 gro^v. and develop the deficiency symptoms because they are partially 

 starved. This source of error can be avoided by "pair-feeding," by pairing 

 in some arbitrary way each control and experimental animal, then 

 weighing the food eaten each day by each experimental animal and 

 giving only that much food to the corresponding control member of the 

 pair. 



One of the more useful methods of detecting cause and effect rela- 

 tionships is the method of concomitant variation. If a variation in the 

 amount of one gnen factor produces a parallel variation in the effect, 

 the factor may be the cause. Thus, if several gioups of rats were given 

 diets with varying amounts of thiamine, and if the amount of protection 

 against beriberi varied directly with the amount of thiamine in the diet, 

 one could be reasonably sure that thiamine deficiency is the cause of 

 beriberi. 



It must be emphasized that it is seldom that we can be more than 

 "reasonably sure" that X is the cause of Y. As more experiments and 

 observations lead to the same result, the probability increases that X is 

 the cause of Y. When experiments or observations can be made quanti- 

 tative, when their results can be counted or measured in some way, the 

 methods of statistical analysis provide a means for calculating the proba- 

 bility that Y follows X simply as a matter of chance. Scientists are usu- 

 ally satisfied that there is some sort of cause and effect relationship 

 between X and Y if they can sho^v that there is less than one chance in 

 a huntired that the observed X-Y relationship could be due to chance 

 alone. A statistical analysis of a set of data can never give a flat yes or 

 no to a question; it can state only that something is \ery probable or 

 very improbable. It can also tell an investigator approximately how 

 many more times he must repeat the experiment to shoA\ ^vith a given 

 probability that Y is caused by X. 



The proper design of experiments is a science in itself, and one for 

 which only general rules can be made. In all experiments, the scientist 

 must ever be on his guard against bias in himself, bias m the suoject, 

 bias in his instrument and bias in the wav the experiment is designed. 



Each experiment must include the proper control group (indeed 

 some experiments require several kinds of control groups) . The control 

 group is one treated exactly like the experimental group in all respects 

 but one, the factor whose effect is Ijeins: tested. The use of controls in 

 medical experiments raises the difficult question of the moral justifica- 

 tion of \\ithholding treatment from a patient \\"ho might be benefited 

 by it. If there is sufficient evidence that one treatment is indeed better 

 than another, a physician -would hardly be justified in further experi- 

 mentation. However, the medical literature is full of treatments no^v 

 known to be useless or even detrimental, which were used for man) 

 years, only to be abandoned finally as experience showed that they were 



