This is essentially the same pattern as for replication


 s = oajn.

 Data from uniformity trials on farms collected and analysed by Hector Barreto seem to confirm the
 relationship between plot standard deviation and plot area. Hence, we should expect the usual statistical
 benefits of using smaller plots, allowing better control of the random variation, to apply on farms.

 There are, however, other considerations. One is that the researcher (with the farmer) may not be able to
 use an appropriate form of blocking because of inadequate time to examine the particular farm situation
 and likely pattern of yield variation. Hence, the experimental design will not be efficient in controlling
 random variation and small plot benefits will be dissipated.

 A second consideration is that small plots may seem quite unrealistic to the farmer and he may not apply
 the care of management which he would to larger plots.

 Another consideration is that the total area of an experiment on farm is less than one on-station, and it is
 also true that the absolute level of variability, per area, on farms is greater than on station. Hence, it is
 inevitable that precision of results on farms will be reduced compared with that expected from station
 experiments. The disappointment with the level of precision on farms, ignoring information about expected
 precision which can be calculated as in section 2.1, may be partly responsible for the belief that small plots
 are the cause of poor precision in OFR.

 The choice of plot size for future OFR experiments must continue to be a matter of judgement. If the level
 of control of variation through the researcher's knowledge of the particular farm conditions, expressed in
 careful choice of blocking, is good, then smaller plots may be appropriate. If such control is not feasible or
 if farmer preference and difficulty of management make small plots unacceptable, then we may have to use
 larger plots.

 An almost inevitable consequence of using larger plots is that the number of plots per farm will be reduced
 and it is then possible that the precision of information on a singlefarm will be inadequate to produce
 useful conclusionsfor thatfarm. The implication would then be that we have to expect that most
 information would derive from analysis of the total (multiple farm) experiment. This further implies that
 the design of the whole experiment be considered as a whole rather than repeating an identical experiment
 at a number of farms. It is, of course, possible more generally that designing the whole experiment, with
 non-identical site-experiments, will be beneficial.

 Certainly the use of larger plots does not necessitate a reduction in the total number of treatments
 (corresponding to the total number of questions asked). We may well need to use subsets of the total set of
 treatments at each farm and the design of experiments with different treatment subsets is discussed later.
 We must be prepared to design OFR Experiments to fit the particular resources and questions according to
 the principles of Sections 2.1 2.3.

 4. Managing Variation

 4.1. Information About Variation at the Farm Level

 In some situations it may not be possible for the experimenter to make any assessment of the variability
 within the area made available for experimentation by the farmer. This necessarily makes designing
efficient experiments much more difficult. If it can be assumed that the person responsible for local
arrangements for the experiment both understands the concept of blocking and is capable of identifying the
likely patterns of variation between plots then it is reasonable to use blocking in the design. In such a
situation, where the experimenter cannot control, directly, the decisions about plot and block design I think