R. A. Fisher, H. G. Thornton, and W. A. Mackenzie 339 



a worse fit by chance from normal data is -765. Grouping together the 

 consecutive positive and negative errors, it only falls to -571. There is 

 again no significant deviation of the distribution in this range from 

 expectation. 



Table IX 



Comparison of observed and expected distribution ofx^, 5-plate data 



Of the values above 11, three he between 12 and 13, and in discussing 

 the evidence for epidemics we shall assume that these are normal sets, 

 and that all those above 13 are exceptions. When we take the evidence 

 of epidemic incidence into account, it is found that the only four sets 

 above 13 which might reasonably be considered normal all occur in 

 epidemic periods, and that the same is true of one out of the three between 

 12 and 13. This therefore (No. 160, see Fig. 2) is probably also an 

 exception. 



The conclusions to be drawn from the 4-plate and from the 5-plate 

 data, thus confirm each other at every point. In both groups the sets 

 having exceptionally high variabihty may be identified in almost every 

 case with certainty. The majority of both groups, about 124 of the 

 4-plate sets, and about 117 of the 5-plate sets, are evidently true samples 

 of the Poisson Series. Both groups show an excess of cases of small 

 variability, but it is not possible to specify the actual sets affected by 

 this; it is evident that this cause, hke that which produces high varia- 

 bility, is sporadic and not systematic in its action; it afTects a certain 

 number of sets in a definite manner, leaving the majority unaffected. 

 This effect, whatever be its nature, is more clearly brought out in the 



