I. AVOID FRUITLESS EXPERIMENTS 31 



also quite misleading to treat poor observations statistically. Infinite 

 experimental skill would make statistical analysis unnecessary, but 

 even infinite statistical skill would possess little value if only bad ex- 

 perimental observations were available. 



Errors, by their very nature uncertain things, may be classified 

 into several categories. The expression "standard error" refers to a 

 mathematical concept employed as a measure of variation in connec- 

 tion with the Gaussian error curve. It is applied to that class of 

 errors that are presumed small and accidental in nature, and that 

 possess a given symmetrical distribution about a mean value. The 

 usual statistical formulas are strictly applicable to this class of errors 

 only. Other categories are known as constant errors, systematic 

 errors, and blunders. An example will bring out the distinction. In 

 Millikan's measurements on the electron by the oil drop method, the 

 value of the charge seemed to change with the size of the oil drop. 

 This apparent variation was discovered later to be due to a "syste- 

 matic" error, which was introduced into the calculations through the 

 use of an inadequate equation for the rate of fall of tiny spherical 

 bodies in air; this error was eliminated before Millikan published his 

 final results in 1917. However, it was not discovered until 1928, 

 when Birge carried out a careful recomputation from the original 

 data, that the published value for the electronic charge was in error 

 because its initial calculation involved an obsolete value for the ve- 

 locity of light and, further, was based on the mistaken assumption 

 that the international volt is identical with the absolute volt. These 

 two "constant" errors amounted to —0.004 and were nearly as large 

 as the standard error. (The "probable error" was stated in this ex- 

 periment as ±0.0038.) Ten years later (1938-39) improved viscosity 

 determinations for air were found to exceed the best value available 

 to Millikan by moi-e than one-half of one per cent, thus revealing 

 another "constant" error. Although the most acceptable value for 

 the electronic charge has changed through the years, it should be noted 

 that the experimental data with the original estimate of statistical 

 error retain their worth in spite of the large corrections made from time 

 to time as "constant" errors have been discovered. 



Large errors, or shall we say blunders, come in still another cate- 

 gory. Frequently they occur during the recording of observations 

 or they may appear in publications as typographical errors. Sta- 

 tistical criteria do not apply, obviously, but sometimes an error of 

 this type may be revealed during mathematical analysis of the data. 



