According to table 2, the size distribution of the fish yielding 

 "agreements" on a given age, when compared v.'ith that of fish assigned 

 to that same age by revision of readings that had originally disagreed 

 ("revised disagreements"), gave P values high enough to indicate that 

 the tviTo distributions had been of the same population in the statistical 

 sense. On the other hand, vfhen a parallel comparison was made v/ith ad- 

 jacent year classes, the P values were all so low that there could be 

 no question that the fish v/ere of different populations in the statistical 

 sonse. It appears,' then, that the final decisions on the disagreements 

 v;ere not determined by chance, but must have been, on the whole, correc- 

 tions of errors in the first reading. 



It is, of course, not possible to Imov/ the absolute error of the 

 final determinations. Among the "agreements" can be erroneous readings 

 that agree by chance; but for scales of ordinary difficulty the number 

 of these is probably lov/, being something less than one percent of the 

 casesl/. In addition, there must be errors among the joint readings. 



There may be more of these than are found among "agreements," since 

 only the more difficult scales tend to be the subjtct of disagreement. 

 P\irtherraore , the final decisions in the joint readings may be determined 

 often by the domination of one of the participants. Finally, there are 

 the errors resulting from scale faults, referred to above. 



Errors from all these causes are not measurable by any knoinm methods. 

 The material and methods at hand provide merely this incomplete measure: 

 If B disagrees with A, A is not held to be in error if a joint reading 

 concurs vfith him rather than with B. Hence, an index of a reader's error 

 is the number of age determinations v/ith v;^hich there is disagreement in 

 both a parallel reading, and in a subsequent joint reading, expressed as 

 percentage of the number of specimens ^aged. For the three tests discussed 

 above, this index of error, by ages and by three readers, A, B, and C, 

 totalled as given in table 3. 



■=' This is based on the average knovm error of each man being close to 

 seven percent. The probability of both misreading any one scale on the 

 average, assuming purely ramdom error, ,is (0.07)^ = .OOI4.9. 



103 



