39 



result not of a single year's weather, but of the 

 prevailing climate of the district. 



2. To repeat the same experiment on a large number of 

 farms, so that the average yield may indicate not the 

 result of a single soil, but the average result of the 

 soils of the region explored. 



3. To repeat the same experiments on small plots on an 

 apparently uniform piece of land, and so to obtain 

 some estimate of the experimental errors of field 

 experimentation. 



The difficulties encountered by the first method are great 

 expense, delay in arriving at definite conclusions, cumulative 

 effect of soil heterogeneity and uncertainty to what extent the 

 discrepancies between different years are ascribable to weather 

 differences and to what extent to experimental errors. 



The difficulties of the second method are expense in the 

 absence of widespread and intelligent support from the farmers, 

 unrepresentative character of the weather of a single season and 

 uncertainty to what extent discrepancies between different farms 

 are ascribable to soil differences, to experimental errors or to 

 weather differences. 



The third method possesses the advantage of attempting not 

 merely to " average out," but to evaluate the causes of variation ; 

 by itself it makes no attempt to study the variations due to soil 

 and weather, but deliberately aims at evaluating the experi- 

 mental errors and so of obtaining a result of known accuracy. 

 The principal difficulty encountered has been the marked hetero- 

 geneity often found on apparently uniform pieces of land. The 

 soil (heterogeneity has often not merely detracted from the 

 accuracy of the results, but has vitiated the estimates of error in 

 such a way that the degree of accuracy of the results is in reality 

 unknown. 



These difficulties of the method of experimentation may be 

 overcome by the replication of small plots. A valid estimate of 

 accuracy may be achieved by arranging the plots in the field so 

 that they conform to the requirements of the statistical theory 

 used in the reduction of the data. To this end, definite rules 

 may now be laid down. The lowering of the experimental error 

 may be achieved to a greater extent than has hitherto been 

 attempted by the systematic adoption of the principle of local 

 control, by which plots to be compared are set out on land of 

 comparatively similar quality, without vitiating the estimate of 

 the experimental error calculated from the totality of the results. 



Testing these new principles of procedure upon the results 

 of uniformity trials, such as that of Mercer and Hall (1910), it 

 appears that when small plots (l/200th acre) are practicable, 

 the comparative values of, say, five different treatments or 

 varieties may be obtained from an acre of land with errors 

 within 1 per cent., and moreover with known accuracy. The 

 actual arrangement may be varied to meet other requirements, 

 but for small plot work with four ? five or six treatments to be 



