Basic Principles 7
as possible. The trial is then partitioned into small units, or plots. These
are harvested individually and the individual yields, together with the
location of the plot from which they came, are recorded. These can then
be used to construct a yield map of the site, such as that shown in Figure 1.1.
The raw yield map may be smoothed to remove minor irregularities and
some of the normal plant-to-plant variation. This smoothing can be done
in a number of ways, of which one of the easiest to use is running medians
of three plots. For this purpose the yield of a plot is replaced by the median
of three adjacent plots in the same row. This procedure is called a 3M
smooth in the row. The 3M row smooth for the yields in Figure I .1 is
shown in Figure 1.2. The next step is to do a 3M smooth in the columns
of the map that has been smoothed by rows. The yield map has now been
smoothed in two directions. The resulting smooth map is presented in
Figure 1.3. The last step is to delineate areas of equal yield. This can be
done by using different colors, shading, or crosshatching to develop a
smooth yield contour map of the area. The final map of the sample data
is presented in Figure 1.4.
Such a map can be of help in deciding on the layout of the field plan.
It can also be used in determining the shape and orientation of the plots
and in grouping the plots into blocks of similar plots. Further, information
from a uniformity trial can be used to determine the optimum size of plot
as well. We consider this aspect further when we look at plot size.
1.5 EXPERIMENTAL ERROR MEASUREMENT
AND CONTROL
Modern experimental procedures include a number of features that permit
the crop scientist to measure and control experimental error. These features
include (1) replication, (2) randomization, and (3) blocking.
1.5.1 Replication
It is sometimes possible for a scientist to obtain an estimate of experimental
error from a previous experiment involving similar treatments on similar
plots. The preferred procedure, however, is to obtain an estimate of error
from the experiment itself. To do this the scientist must repeat, or replicate,
each treatment in the experiment. That is, each treatment must appear on
more than one plot in the trial.
There are a number of reasons for replication in an experiment:
1. It provides an estimate of experimental error because it provides
observations on several plots receiving the same treatment. In this
way the variation among plots treated alike can be measured.
8 Chapter 1
488 406 491 479 438 367
488 392 457 438 296 510
457 359 436 396 503 337
440 403 388 450 420 384
325 365 369 343 361 376
347 462 393 390 368 437
424 387 408 354 386 402
426 384 346 389 375 375
448 454 347 374 321 403
476 430 490 428 435 452
416 445 440 415 449 442
387 382 319 393 438 455
470 418 353 402 470 410
446 C'l£
352 394 427 435
410 393 385 411 446 441
388 420 389 419 414 348
303 333 389 396 473 371
315 328 365 426 455 375
Figure 1.1 Site map of yield data (g/plot) of wheat from a uniformity trial.
Basic Principles 9
488 488 479 479 438 367
488 457 438 438 438 510
457 436 396 436 396 337
440 403 403 420 420 364
325 365 365 361 361 376
347 393 393 390 390 437
424 408 387 386 386 402
426 384 384 375 375 375
448 448 374 347 374 403
476 476 430 435 435 452
416 440 440 440 422 422
387 382 382 393 438 455
470 418 402 402 410 410
446 426 394 394 427 435
410 393 393 411 441 441
388 389 419 414 414 348
303 333 389 396 396 371
314 328 365 426 426 375
Figure 1 .2 Yield data from Figure 1.1 smoothed by running medians of 3 in rows (3M
smooth in rows).
10 Chapter 1
488 488 479 479 438 367
488 457 438 438 438 367
457 436 396 420 396 376
440 403 396 420 396 376
347 393 393 390 390 384
347 393 387 386 386 402
424 393 387 386 386 402
426 408 384 375 375 402
448 448 384 375 375 1 403
448 448 430 435 435 422
416 440 430 435 435 452
416 418 402 402 422 422
446 418 394 394 427 435
446 418 394 402 427 435
410 393 393 411 427 435
388 389 393 411 414 371
315 333 365 414 414 371
315 328 365 426 426 375
Figure 1.3 Data from Figure 1.2 smoothed further by running medians of 3 in columns
(3M smooth in columns).
Basic Principles 11
g g ~ ~ § D
48 - 50 44 47 41 - 43 38 - 40 35 - 37 32 - 34
Figure 1.4 Smoothed variability map from a unifonnity trial with wheat (yield in
dglplot).
12 Chapter 1
2. It increases precision by reducing standard errors. Recall that the
standard error of a mean, s1, is given by
Sy = fi
where s2 is the sample variance and n is the number of observations
(replications) on which the mean is based. Clearly, as the number
of replications, n, is increased, the standard error, sy, decreases.
3. It can broaden the base for making inferences. As replication is
increased a wider variety of plots can be brought into the experiment
This can lead to a greater range of conditions over which the results
of the experiment can be applied.
Before considering the factors that determine the number of replications,
we need to recall two relations:
1. sy = Js 2/n, as given previously.
2. L(J.L, - J,L 2) = (Y1 - Yz) ::!: taJ2s 2/n, where
L(J.1 1 - J,L 2) = (1 a)IOO% confidence interval estimate of the
true difference between the means of treatments I,
J.L,and 2, J.12
Y~> y2 = sample means for treatments 1 and 2
ta = two-tailed t at the a probability level
s2 = sample variance per plot
n = number of replications
Thus the number of replications determine, in part, the precision with
which the true means of the treatments can be estimated and the magnitude
of the differences that can be detected. With this background we see that
some of the things that determine the number of replications are:
1. Pattern and magnitude of variability in the soils of the experimen-
tal site
2. Size of the difference to be detected
3. Required significance level
4. Number of treatments
5. Amount of resources that can be devoted to the experiment
6. Limitations in cost, labor, time, and so on.
1.5.2 Randomization
In addition to replication there is another factor that a plant scientist must
consider. Suppose that a wheat breeder had decided to use five plots of a
Basic Principles 13
new variety and five plots of an old one to test the difference between the
two varieties. Suppose, further, that he had laid out a strip of 10 plots and
assigned varieties as shown in Figure 1.5.
Now, suppose that unknown to the breeder, there is a fertility gradient
ranging from low at one end of the strip to high at the other. In each pair
of plots in the field plan the new variety has been assigned to the plot with
the higher fertility level. An observed difference between varieties will
reflect a fertility difference in favor of the new variety as well as a varietal
difference if it exists. To avoid this problem the treatments should be
assigned to the plots in such a way that no treatment is favored by being
assigned consistently to the better plots. To do this the breeder should resort
to a process called randomization.
With proper randomization any plot is equally likely to be assigned to
any treatment. In some experimental designs randomization is restricted in
certain ways, but in no design is it completely eliminated. We shall look
at the details of the randomization procedure for each type of experimental
design when we consider the individual designs.
There are at least two reasons for randomization:
I. To eliminate bias. Randomization assures that no treatment is
favored or discriminated against by the way it is assigned to the plots.
2. To ensure independence among observations. This is required for
valid significance tests and interval estimates.
1.5.3 Blocking
Another feature of good experimental technique is blocking. The procedure
here is to group the plots, or experimental units, into blocks, or groups, of
homogeneous units. Treatments are then assigned at random to the plots
within the blocks. The basic goal is to remove block-to-block variation
from experimental error. Treatment comparisons are then made within the
blocks on plots that are as nearly alike as possible.
0 N 0 N 0 N 0 N
LOW FERTILITY ---->~ HIGH
Figure 1.5 Plot layout for new (N) and old (0) varieties assigned systematically to a
strip of plots with a fertility gradient.
14 Chapter 1
For example, the field under consideration for an experiment might
contain two types of soil. We could select plots on each soil type and
randomly assign treatments to plots within soil types. In this case soil type
would be a blocking factor and the difference between soil types would
be removed from experimental error in the analysis of the data. Of course,
criteria other than soil type can be used as blocking factors. The key to
successful blocking is to maximize the variation among blocks and to
minimize the variation among plots within blocks. We shall consider the
concept of blocking in more detail as we look at specific experimental
designs.
There are a number of reasons for blocking:
1. It can increase precision. Differences among blocks are removed
from experimental error in the analysis of the experimental data.
2. Treatment comparisons are more uniform. Because comparisons are
made within blocks of uniform plots the observed differences more
nearly reflect true differences among treatments.
3. It can sometimes increase the information from an experiment.
Blocking permits the scientist to sample a wider range of conditions
and hence can broaden the scope of inferences.
Note: Contrary to usage in some areas of agricultural research, blocks
and replications are not the same thing. A block is simply a group of plots,
whereas a replication is a repetition of a treatment in an experiment.
1.5.4. Reduction of Experimental Error
We have indicated earlier that one of the goals of sound research procedure
is to reduce experimental error and thus increase precision. Recall that the
standard error, Sy, of a mean is a measure of the precision with which the
mean is estimated. The smaller is the standard error the more precise is
the estimate. Recall, again, that
Sy= l
where s2 is the error variance per plot and n is the number of observations
(replications) per treatment. We can reduce Sy in the following ways:
l. Increase the size of the experiment. Increase the size by increasing
the replication or by increasing the number of treatments. We are
attempting to reduce sy by increasing n or by increasing the degrees
of freedom with which s2 is computed, in this way, reducing s9• As
the size of the experiment is increased, however, more heterogeneous