ASSESSMENT OF RISK
ATTACHED TO RECOMMENDATIONS
Training Working Document No. 2
C15 v (
ASSESSMENT OF RISK
ATTACHED TO RECOMMENDATIONS
Training Working Document No. 2
Prepared by
Roger Mead
Consultant
in collaboration
with CIMMYT staff
CIMMYT
Lisboa 27
Apdo. Postal 6641,
06600 M6xico, D.F., Mexico
PREFACE
This is one of a new series of publications from CIMMYT entitled Training Working
Documents. The purpose of these publications is to distribute, in a timely fashion,
trainingrelated materials developed by CIMMYT staff and colleagues. Some Training
Working Documents will present new ideas that have not yet had the benefit of extensive
testing in the field while others will present information in a form that the authors have
tested and found useful for teaching. Training Working Documents are intended for
distribution to participants in courses sponsored by CIMMYT and to other interested
scientists, trainers, and students. Users of these documents are encourage to provide
feedback as to their usefulness and suggestions on how they might be improved. These
documents may then be revised based on suggestions from readers and users and
published in a more formal fashion.
CIMMYT is pleased to begin this new series of publications with a set of six documents
developed by Professor Roger Mead of the Applied Statistics Department, University of
Reading, United Kingdom, in cooperation with CIMMYT staff. The first five documents
address various aspects of the use of statistics for onfarm research design and analysis,
and the sixth addresses statistical analysis of intercropping experiments. The documents
provide onfarm research practitioners with innovative information not yet available
elsewhere. Thanks goes out to the following CIMMYT staff for providing valuable input
into the development of this series: Mark Bell, Derek Byerlee, Jose Crossa, Gregory
Edmeades, Carlos Gonzalez, Renee Lafitte, Robert Tripp, Jonathan Woolley.
Any comments on the content of the documents or suggestions as to how they might be
improved should be sent to the following address:
CIMMYT Maize Training Coordinator
Apdo. Postal 6641
06600 Mexico D.F., Mexico.
Document 2A
THE ASSESSMENT OF THE RISK ATTACHED TO RECOMMENDATIONS
THEORY
1. Preliminary Comments
It is assumed that we would like to recommend a change from treatment 0 to treatment 1. The evidence for
this intention is a number of separate pieces of experimental information about the size of the advantage to
be derived from this change. If the evidence includes estimates of the size of the advantage from a single
set of trials which is accepted as totally representative of the situations in which the recommendation might
be applied, then the methodology of the "Representation of Risk" report (Mead, 1990a) is relevant.
Normally the requirements for such a set of trials would be that they be representative both of the
population of possible locations and of the population of possible years.
The use of average yields for a range of locations, each averaged over a few years, might be regarded as an
adequate basis for assessing the risk appropriate to the recommendation, in the sense of representing the
variation across locations of average benefit over a number of years. That is, the risk would be for average
performance over years for a single location rather than for performance in a single year at a single
location.
The situation which we attempt to solve here is the assessment of risk for a single year at a single location
when we do not have adequate direct information about the population variation for both years and
locations. The information which is required is the two or three variance components for years, for
locations and, if necessary, for year x location interaction. We could subsume the interaction component in
the year variation if we redefine year variation as yearwithinlocation variation, unless we can believe that
interaction variation is negligible (year to year differences are consistent across locations within
experimental error).
2. Variances and Variance Components.
Following Byerlee [following Binswanger and Barah, (1980)] define
a2(s) as the variance between sites/locations (averaging out the differences between years);
a2(y) as the variance between years (averaging out the differences between locations);
a2 as the interaction year x site/location variance, incorporating variation within a site/location.
If we are to think in terms of a combined 02 and a2(y) we can further define
o2(yws) as the variance between years within site/locations.
The total variance for a prediction for a single site/location for a single (future) year is either
02 + a2(s) + a2(y)
or
a2(s) + o2(yws).
We assume that we will always have sufficient information about the variation between sites within (one or
more) years. This is equivalent to
o2 + o2(s).
It is clear that we can obtain, directly from data on a set of trials, separable information about all three
variance components if, and only if, the set includes at least several site/locations repeated over at least
several years. This requirement appears to be rarely, if ever, satisfied and if the assessment of risk is to be
accorded a high priority this situation may have to be reconsidered.
If we cannot estimate all three of o2(s), o2(y), and o2 directly we must find some way of predicting them
from each other. We could, for example, demonstrate (or assume) that the primary causes of variation in
the three forms were similar and examine the corresponding variance components for the causative
variables. Thus, if we could derive a relation between yield difference and, say, rainfall (total for a period,
or some combination of rainfall information then we could examine the variation of yield across sites (not
necessarily the same as trial locations) and years.
3. A Practical Approach and Examples
Two quite substantial sets of data have been analysed as case histories. In addition a set of simulation data
for ten sites for ten years, with rainfall data for the same 100 combinations has been constructed, and
subsamples of data from the total data set have been analysed to provide further illustration of the possible
procedures. I hope that by using all these examples I can demonstrate an overall philosophy for assessing
risk and also assess the potential and implications of the implementation of that philosophy.
We consider four distinct situations:
3.1 Direct Estimation of Variance Components.
There is data on the yield difference between two alternative treatments for all, or most of, the
combinations of a considerable number of years and a considerable number of sites. We can estimate the
three variance components and their sum and thence calculate risk probabilities for a recommendation to
change from one treatment to the other. We shall return to the consideration of how much data is required
to obtain adequate information about the variance components. In this situation we do not have, nor would
intend to use, information on ancillary variables.
This situation is rare (or possibly nonexistent). The analysis of the complete simulated data set (Document
2D) demonstrates the estimation of the variance components.
3.2 Estimate from Total Sample Variance (No Components)
If, in (3.1) the evidence for consistent site or year differences is not compelling then we can treat a set of
data for various site/year combinations as a representative sample from the population of possible site/year
combinations. We can then estimate the variance of that population from the sample variance (ignoring site
and year differences) and hence calculate the risk probabilities attached to the recommendation for change.
This would work well in situations where the domain for experimentation was homogeneous or where the
pattern of variation of yield differences across years was completely different for different sites. The data
from N x P trials in Poza Rica (Document 2B) appears to represent this situation well, though the possible
identification of a consistent difference between site groups would lead to the recommendation being for a
subarea of the district in which trials were performed.
This approach, of assuming that the complete sample of trial results is representative of the population for
which the recommendation is required, is always a possible option. It is only convincing when the absence
of systematic site and year variation (over and above the interaction variation) has been demonstrated.
Nevertheless it will quite often offer an estimate of the population variance which is based on a good
number of degrees of freedom and could be used in preference to other estimates based on more credible
models but inadequate data.
3.3 Predict Year Variation from Regression.
If the data sample contains too few years to be able to deduce any useful estimate of the year variance
component there may be information in the data sample about a relationship between yield difference and
an explanatory variable. If this relationship is quite strong and there is additional information about the
variation over years (and possibly also sites) of the explanatory variable then it may be feasible to construct
estimates of the year (and other) variance components.
The procedure would involve extracting whatever information is available from the data about the variance
components of field difference. Then the regression of yield difference on the explanatory variable is
calculated. If the regression relationship is well estimated then we could assume that the variation of yield
differences can be predicted from the variation of the explanatory variable. If the regression equation is
yield difference, y = A + Bx
and the variance of the estimated regression coefficient is
var(B) = (Residual Mean Square)/ (Corrected SS of x),
then the variance of y across years and sites is
var (y) = B2 (variance of x) + (Mean of x)2 var(B).
To split the variation of the predicted y values for different site year combinations into variance
components we assume that the proportional division of the variation of y into variance components is the
same as the corresponding division of the variation of x. This is a dubious assumption but if the regression
relationship is strong then it may be a reasonable approximation.
Rainfall is one potential explanatory variable which might produce a good regression and for which there
will often be supplementary information over a range of years and sites. The rainfall measure could be an
annual total or a period total or even a set of rainfall measures for several periods. Another possibility as an
explanatory variable could be mean farm yield for the standard farm practice.
The data example from Ghana (Document 2C) provides an example where the possibility of using rainfall
as an explanatory variable was developed by identifying, for each experimental site, the nearest point from
which rainfall information was available. Analysis of the data, however, did not discover a useable
regression relationship either with the rainfall measure or with the average yield of the two treatments. The
only possibility for this data set was to revert to (3.1) and assume that the set of 27 experimental conditions
could provide a valid estimate of the population variance for the recommendation prediction.
Three examples of typical structures of experimental information providing the basis for a recommendation
are drawn from the simulated data (Document 2D) and show how the method could be applied and also
illustrate the variability of the resulting estimate of the variance components and of the combined variance.
3.4 Using Computer Models
The other possibility is to use a computer modelling system to predict the variability in yield difference. If
a suitable computer model is available and there is sufficient data to validate the model for the local
Recommendation Domain, then the model predictions can be calculated for a range of site year
combinations for which the necessary input information for the model is available. No example of such a
procedure is available.
4.Conclusions
Of the real data examples, I believe the risk probability assessments from the N x P trials at Poza Rica are
probably reliable. They are based on a wide range of years and of sites and the lack of evidence for
systematic site or year differences (omitting site group 7) is quite compelling. Although the two years of
the Ghana data were very different, I would not be confident in using the risk probability assessments
because of having only the two years. Nevertheless the risk assessments are better than nothing.
The simulation results make the difficulties in estimating variance components very clear. We can quantify
this difficulty by considering the Chisquare distribution. For a variance, or variance component, estimated
on v degrees of freedom, the divisors to construct 90% confidence limits for the true variance are
calculated from the 5% and 95% points of the Chisquare distribution. Some values are shown.
Degrees of freedom Divisors ( of sample variance) for
of sample variance Lower limit Upper limit
4 0.18 2.37
6 0.27 2.10
9 0.37 1.88
12 0.44 1.75
15 0.48 1.67
20 0.54 1.57
Thus even for the estimates of o2(s) and o2(y) from the complete simulated data sample, based on 9 df, the
uncertainty about the value of the variance component is between
Sample variance /1.88 and Sample variance /0.37.
I believe that in most cases the estimation of the predictive variance of the yield difference on which a
recommendation is based will most effectively be based on the total sample of data values of the yield
difference. Where the data sample allows the estimation of variance components to be attempted using the
methods illustrated in documents 2B, 2C or 2D then this should be attempted. However for all estimates
the precision of variance estimates should be assessed and the resulting calculated risk probabilities
interpreted in the context of that precision.
The conclusions for the planning of verification experiments prior to making a recommendation are that the
number and distribution of trials must be large.
Document 2B
DATA FROM 30 NxP TRIALS AT POZA RICA
1. Introduction
The trials are part of a larger set performed over two decades on farms in the Poza Rica region. Each trial
has two replicates of twelve treatments arranged in two randomised complete blocks of twelve plots per
block. The twelve treatments were all combinations of four levels of Nitrogen (0, 50, 100, 150 kgN/ha)
with three levels of Phosphorus (0, 40, 80kg P205/ha).
The 30 trials are spread between 1976 and 1986, in two seasons (A and B) as shown:
Season A Season B
1976 1 trial I trial
1979 1 trial 2 trials
1980 3 trials I trial
1981 4 trials
1982 3 trials 2 trials
1983 3 trials 5 trials
1984 1 trial 2 trials
1986 1 trial
The trials are also spread geographically with some repetition of villages and some other possible
clustering. Based on partial information seven groups of locations are defined as follows:
Group 1 Papatlarillo, Copal;
Group 2 El Palmar, Mahuapan, Mamey, Ind.Nat.II;
Group 3 Jiliapa, Zaragoza;
Group 4 La Reforma, Tihuatlan;
Group 5 Zapotalillo;
Group 6 Cardel, Zapotalillo (Calvo, Hemrndez);
Group 7 Tierra Blanca, Sabanillas.
A summary paper for 19 trials in 1973 to 1977 examining average responses for each season and using net
benefit analysis suggests that the important treatments are NOPO, NIPO, NIPl and N2PO.
For 26 of the trials, results from the preliminary analysis were available providing Error Mean Squares and
F ratios for N and P main effects and for the interaction N x P. The general pattern of the evidence
indicated that both main effects were important (11 and 8 trials with 5% significant F ratios for N and P
respectively and almost all F ratios for P greater than 1). There were three significant interaction F ratios
(two of them from experiments with very low error mean squares) but only nine of the 26 interaction F
ratios were greater than 1. The overall evidence is, I believe, quite clear that a main effects model is
adequate for these trials, not only overall but also for individual trials.
2. The Distribution of Treatment Differences
With a sample which includes a wide range of both years and locations, such as the present set of trials, it
is reasonable to hope that the sample is sufficiently representative, simultaneously, of future years and
locations. Therefore we first examine the distribution, over the 30 sample points, of treatment differences
expecting that the mean and variance of the sample of treatment differences will be adequate for estimating
risk from a recommendation based on the mean difference. We shall check the possibility of consistent
year and location differences after this initial analysis.
If a main effects model is the correct summary for each trial then the best estimates of differences of yield
between treatment combinations will be derived from the main effect estimates and not from comparisons
of the particular combinations. This follows not from the results from individual experiments but from the
overall pattern of results, and is an important consequence of being able to consider a substantial set of
experiments.
2.1 Estimates Based on a Fitted Quadratic Response
Calculation of fitted response models can always be achieved through the use of a multiple regression
program. The model to be fitted can be determined by comparing the residual mean squares for alternative
possible models. A general model would be
Yield difference = a + bl*N + cl*N2 + b2*P +c2*P2 +d*NP
in which the bl*N, cl*N2,b2*P and c2*P2 terms are components of main effects and the d*NP term is part
of the interaction.
Since we have decided that the interaction term is negligible we omit the d*PN term. The estimation of
quadratic response terms for N and P can now be calculated separately for each factor. Instead of using
multiple regression we can estimate the linear and quadratic terms directly from the main effect means for
the two factors. This is quicker than using a regression program when we have only three or four levels of
the factor.
For the response to P, for which there are three levels the quadratic will fit the mean yields for PO, PI and
P2 exactly. Our purpose in fitting quadratic responses is to estimate yield differences between levels of
each factor. We do not therefore need to calculate the linear and quadratic terms from the mean yields and
then calculate estimates of the mean yield differences between levels from the linear and quadratic terms,
since we would simply get back to where we had started.
The main effect estimate of PIPO for any level of N will therefore be the difference between the mean
yields for PI and for PO. This difference will be based on 8 observations per mean and, for each experiment
the estimate will have a variance
2 o2/8
compared with the variance for NOPINOPO which is
2 Y2/2
(where o2 is the random variance estimated by the experimental error mean square).
There are four levels of N and the quadratic response has to be estimated from the four mean yields for NO,
NI, N2 and N3. This is done by finding appropriate contrasts of the mean yields. If the response model is
written
yield = a + b(N Nmean) + c(N Nmean)2,
and we code the levels of N as 0,1,2 and 3 so that Nmean = 1.5, then the mean yields for NO, NI, N2 and
N3 are
NO = a 1.5b + 2.25c
N1 = a 0.5b + 0.25c
N2 = a +0.5b + 0.25c
N3 = a +1.5b + 2.25c.
By manipulation of these equations we can produce
N3 NO= 3b
N2 N1 = b.
Hence, by simple regression, the estimate for b is
b = (3(N3NO) + (N2NI)) /10.
Similarly, from the equations for NO, NI, N2 and N3
N3 N2 = b + 2c
N1 NO = b 2c.
Hence, by subtraction, the estimate for c is
c = ( N3 N2 NI +NO )/4.
Finally NI + N2 = 2a + 0.5 c
so that the estimate for a is
a = ( N2 +N c/2 ) /2.
During the course of this derivation, which can also be developed from the general statistical theory of
treatment contrasts, we have derived expressions for NI NO and for N2 NI. Hence the main effect
estimates of NINO and of N2NI, based on the quadratic response are
Estimate of (NiNO), Est(Q) = b 2c
Estimate of (N2N1), Est(Q) = b.
The estimates of b and c are linear combinations of mean yields for NO, N1, N2 and N3 and the variance of
each mean yield is
o2/6.
Hence the variances of b and c are
Variance(b) = o2/6 (32 + 32 + 12 + 12)/102 = o2/30
Variance(c) = o2/6 (12 + 12 + 12 + 12)/42 = o2/24
and the variances of the estimates of NINO and N2N1 based on the quadratic response are
Variance (Est(Q) NINO) = o2(1/30 + 4/24) = 02/5
Variance (Est(Q) N2N1) = 02/30
compared with the variances for NIPONOPO and N2PONIPO of
202/2.
Note the improved precision first from the use of main effects instead of individual treatment comparisons,
and then from regression estimates of the response. The estimate of the linear term is particularly precise.
3. Comparisons of Effect Estimates
To illustrate the benefits of improved information from the use of estimates of treatment mean differences
based on the main effects fitted quadratic response, we shall calculate estimates both from our fitted
quadratic response and from direct comparison of mean yields for treatment comparisons.
3.1 Estimates of N1PONOPO.
The estimates of NIPONOPO directly from the two experimental treatment means and from Est(Q) are
calculated for each experiment (the quadratic response being also calculated for each experiment). The
results are listed below and plotted in Figure 1.
Site/year Direct Est(Q) Site/year Direct Est(Q)
1 +0.06 +0.47 2 0.55 +0.06
3 +1.71 +1.16 4 +0.84 +0.74
5 +0.43 +0.44 6 0.14 0.14
7 +0.54 +0.22 8 +0.05 +0.50
9 +1.20 +0.96 10 0.45 +0.35
11 +0.27 +0.18 12 0.22 +0.29
13 +0.90 +0.83 14 +0.73 +0.52
15 0.01 +0.83 16 0.55 +0.14
17 +0.16 +0.24 18 0.45 +0.10
19 0.12 +0.23 20 0.34 0.07
21 0.14 +0.48 22 0.23 +0.04
23 0.03 0.42 24 +0.39 +0.33
25 +0.34 +0.31 26 0.04 +0.15
27 +1.37 +1.18 28 +0.25 +0.08
29 +0.19 +0.47 30 +0.66 +0.23
Since the average experimental error mean square for the trials is about 0.35 the experimental standard
errors for these estimates of these estimates are approximately
S.E. (direct) = 4(2 (0.35)/2) = 0.59
S.E. (Est(Q)) = 4( 0.35/5) = 0.26.
These express the average withinexperiment precision with which each single value is estimated. The
improved precision of the estimates based on the fitted quadratic response is reflected in the reduced
variation of those values shown in figure 1.
The means and variances for the two samples of 30 values for NIPONOPO are
Direct Est(Q)
Mean +0.227 +0.363
Variance 0.324 0.136
Standard deviation 0.570 0.369
Experimental standard error 0.59 0.26
The variance and standard deviation summarize the variation of the effect estimates between the 30
experiments. Note that the standard deviation of the sample of 30 directly estimated values is almost the
same as the approximate experimental standard error. This implies that using the imprecise direct estimates
there is little evidence of variation of the treatment difference across site/year combinations. In contrast the
standard deviation for the variation of Est(Q) across sites is clearly greater than the approximate
experimental standard error.
3.2 Estimates of N2PON1PO
In exactly the same manner as in 3.1, the estimates of N2PONIPO are calculated both directly from the two
experimental treatment means and from Est(Q). The results are listed below and shown in Figure 2.
Site/year Direct Est(Q) Site/year Direct Est(Q)
1 +0.39 +0.29 2 +1.10 +0.38
3 0.41 +0.53 4 +0.20 +0.38
5 +0.12 +0.06 6 +0.62 +0.14
7 0.43 +0.32 8 +1.49 +0.26
9 0.04 +0.20 10 0.35 +0.11
11 +0.25 +0.24 12 +0.65 +0.09
13 +0.06 +0.31 14 +0.20 +0.52
15 +0.58 +0.31 16 +0.86 +0.16
17 +0.22 +0.16 18 +0.07 +0.02
19 +0.67 +0.05 20 0.08 +0.03
21 +0.81 +0.26 22 0.11 +0.66
23 0.00 0.12 24 +0.15 +0.13
25 +0.45 +0.25 26 +0.50 +0.07
27 +0.26 +0.54 28 +0.23 +0.16
29 +0.62 +0.19 30 +0.44 +0.03
The experimental standard errors for these values are approximately
S.E. (Direct) = 4(2(0.35)/2) = 0.59
S.E. (Est(Q)) = (0.35/30) = 0.11.
The means and variances for the two samples of 30 values for N2PONIPO are
Direct Est(Q)
Mean +0.317 +0.224
Variance 0.188 0.032
Standard Deviation 0.433 0.179
Experimental standard error 0.59 0.11
Note again that for the direct form of estimate the standard deviation across site/years is smaller than would
be expected from the experimental standard error but that for the Est(Q) the standard deviation across
site/years is larger than the experimental standard error.
A further aspect of the relative imprecision of the estimates obtained directly from the comparison of
treatments is seen when we look at both the NIPONOPO and N2PONIPO estimates. The mean values of
the direct estimates are +0.227 for NIPONOPO and +0.317 for N2PONIPO. It would be most surprising if
the second increase were really greater than the first. In contrast, the Est(Q) mean values are +0.362 for
NIPONOPO and +0.224 for N2PONIPO, showing the expected diminishing returns, as N is increased.
33 Estimates of NIP1NOP1
The estimates of NIP1NIPO calculated directly from the experimental treatment means and from Est(Q)
are listed below and plotted in Figure 3.
Site/year Direct Est(Q) Site/year Direct Est(Q)
1 +1.01 +0.64 2 +0.75 +0.06
3 0.49 +0.48 4 +1.08 +1.39
5 0.51 +0.44 6 0.46 +0.12
7 +0.56 +0.99 8 +1.12 +0.23
9 0.44 0.04 10 0.05 +0.26
II 0.22 +0.31 12 +0.70 +0.31
13 0.25 0.21 14 +0.21 +0.54
15 +0.68 +0.25 16 +0.55 +0.19
17 +0.05 +0.12 18 +0.70 +0.38
19 +0.33 +0.40 20 +0.45 +0.26
21 +0.48 +0.33 22 +0.51 +0.45
23 +1.02 +1.02 24 +0.27 +0.23
25 +1.07 +0.47 26 0.13 0.01
27 +0.77 +0.90 28 0.19 +0.20
29 +1.57 +0.53 30 +0.63 +0.36
The experimental standard errors of these values are approximately
S.E. (Direct) = 4( 2(0.35)/2) = 0.59
S.E. (Est(Q)) = 4(2(0.35)/8) = 0.30.
The means and variances for the two samples of 30 values for N1PINIPO are
Direct Est(Q)
Mean +0.392 +0.387
Variance 0.313 0.105
Standard deviation 0.560 0.324
Experimental standard error 0.59 0.30
Once again we note that the variation of the direct estimate across site/years is slightly less than would be
expected from the experimental standard errors and the Est(Q) varies slightly more than expected from its
experimental standard error.
The large experimental standard errors, particularly for the less precise direct estimates, emphasize the
difficulty of assessing variation of treatment difference across site/year combinations. Using the better
Est(Q) form of estimates of the effects of changing treatments we can be more confident of the existence of
variation across site/years and have a better estimate of the risk attached to the recommendation to change
fertilizer level.
4. Differences between Sites and between Years
To investigate the possible consistency of differences between site groups and years we consider the NIPO
NOPO effect estimated by Est(Q).
Site Group
1 2 3 4 5 6 7
Year/season
1976 A +0.47
1976 B +0.06
1979 A +0.50
1979 B +0.18 +0.96
1980 A +0.33 +0.31
+0.15
1980 B 0.42
1981 B +0.04 +0.23 +0.48 0.07
1982 A +0.52 +0.83 +0.83
1982 B +0.29 +0.35
1983 A +0.47 +0.08
+0.23
1983 B +1.16 0.14 +0.22
+0.74
+0.44
1984 B +1.18 +0.24 +0.14
1986 B +0.10
We could calculate a complete analysis (nonorthogonal) for site groups, seasons and years. Instead we
look at each factor separately, initial, calculating separate analyses of variance, to assess the variation
between and within (i) site groups, (ii) seasons and (iii) years.
SS df MS
(i) Between Site group 1.1812 6 0.197
Within Site groups 2.7573 23 0.120
(ii) Between Seasons 0.0823 1 0.082
Within Seasons 3.8561 28 0.138
(iii) Between Years 0.9254 7 0.132
Within Years 3.0131 22 0.137
Only the effect of sites seems at all interesting.
If (using the sweeping method described in detail in Mead, 1990e) we subtract the site group means from
the individual data values for that site group we can examine the season and year differences after allowing
for site group differences.
Site groups
1 2 3 4 5 6 7
Year/season
1976 A +0.13
1976 B 0.45
1979 A +0.05
1979 B 0.36 +0.60
1980 A 0.18 +0.17
0.36
1980 B 0.30
1981 B 0.30 0.13 +0.03 +0.05
1982 A +0.18 +0.32 +0.38
1982 B 0.07 0.10
1983 A 0.07 0.28
0.31
1983 B +0.62 0.28 0.14
+0.20
0.10
1984 B +0.65 +0.10 +0.26
1986 B 0.35
Calculating means for seasons and for years shows virtually no difference between seasons but some
differences between years. We therefore ignore seasons amalgamating the seasons results within years and
continue to sweep out differences between years and between site groups until the final residuals shown are
reached, as follows:
Site groups
1 2 3 4 5 6 7
Years
1976 +0.26 0.26
1979 0.46 +0.53 0.06
1980 0.03 +0.37 0.12
0.21
1981 0.26 0.02 +0.10 +0.18
1982 0.00 +0.20 0.18 +0.23
0.25
1983 0.04 0.18 0.22
0.28 0.08
+0.65
+0.23
0.07
1984 +0.29 0.21 0.07
1986 0.00
The sum of squares of these final residuals is (0.262+ 0.262 + 0.462 + +0.072 + 02 ) =1.9064.
The summary analysis of variance is therefore
Source SS df MS
Between site groups (ignoring years) 1.1812 6 0.197
Between years + seasons 0.8509 8 0.106
(adjusted for site differences)
(calculated by subtraction)
Residual 1.9064 15 0.127
As suggested by the initial analyses of variance, neither years nor seasons show any significance and this
confirms that the only possible consistent pattern is that between site groups. Reverting to the initial
analysis for site groups the original F ratio is 1.64 (on 6 and 23 df) some way below significance at 10%.
The group means are
group 1
group 2
group 3
group 4
group 5
group 6
group 7
+0.34
+0.51
+0.54
+0.14
+0.36
+0.45
0.14
These continue to stimulate some interest with the lowest group (7) being extreme to the North, group 4
being nearest geographically to group 7, and group 2 being the most extreme to the South. If we omit group
7 the analysis of variance is
SS df MS
Between site groups 0.4059 5 0.0812
Within site groups 2.5972 21 0.1237
Any suggestion of between site group differences has gone and it might therefore be of interest considering
the risk probabilities not only for the complete sample of 30 experiments, but also for the reduced sample
of 27 omitting the three sites in group 7.
5. Assessment of Risk
5.1 Risk Based on the Complete Sample of 30 Experiments.
We shall consider each of the treatment comparisons, NIPONOPO, N2P0NIPO and NIPINIPO. Using the
costs quoted in the previous summary paper the marginal changes in total costs that vary and the marginal
net benefits can be calculated. We start with the table of mean yield changes and their standard deviations
across site/year combinations (the methodology here is similar to that in Mead, 1990f.)
Change N1PONOPO N2PON1PO N1PIN1PO
Yield Difference
Mean +0.363 +0.224 +0.387
Standard deviation 0.369 0.179 0.324
Now these marginal increases in yield are converted to net benefits using a multiplier of $2160 for yields
and the marginal costs from the previous calculations (using peso costs prevailing inn 1977).
Marginal TCV $359 $359 $259
Marginal Net Benefits
Mean $425 $125 $577
Standard deviation $797 $387 $700
Marginal Rate of Return
Mean 1.18 0.35 2.23
Standard deviation 2.22 1.08 2.70
On the mean values both NIPONOPO and NIPINIPO give acceptable MRR over 100%. The N2P0NIPO
does not.
The risk probabilities are calculated from the Normal probability distribution using the mean and standard
deviation for the relevant variable.
Net Benefits
The probability of a net benefit greater than zero from the change NIPONOPO is calculated from the
normal deviate
giving
z = (0 425)/797 = 0.53
Prob(greater than zero) = 0.702
(from Normal distribution tables);
For the change NIP1NIPO
z = (0 577)/700 = 0.82
giving
Prob(greater than zero) = 0.794.
Marginal Rate of Return
The probability of a MRR greater than 100% (1.0) is calculated in the same way.
For the change NIPONOPO
z = (1.0 1.18)/2.22 = 0.08
Prob(greater than 100%) = 0.532;
For the change NIPINIPO
z = (1.0 2.23)/2.70 = 0.45
Prob(greater than 100%) = 0.675.
Finally, note that because we have used estimates based on main effects, the assessment of the change
NOPINOPO will be exactly the same as that for NIP1NIPO.
5.2 Risk Probabilities Omitting Sites in Group 7
It is interesting to examine the effect on the risk probability calculations of omitting the three experiments
which were quite some distance North of the rest of the experiments. Although the justification for this
omission is based on the analysis for the NIPONOPO values only, we will examine the effect on the risk
probabilities for both NIPONOPO and NIPINIPO.
The revised figures are
Change NIPONOPO N1P1.N1PO
Yield Difference
Mean +0.416 +0.376
Standard Deviation 0.340 0.330
Marginal TCV $359 $259
Marginal Net Benefits
Mean $540 $553
Standard Deviation $734 $713
Marginal rate of Return
Mean 1.50 2.14
Standard Deviation 2.04 2.75
Net Benefits
For the change NIPONOPO
z = (0 540)/734 = 0.74
Prob (greater than zero) = 0.770.
For the change NIPINIPO
z = (0 553)/713 = 0.78
Prob (greater than zero) = 0779.
Marginal Rate of Return
For the change NIPONOPO
z = (1.0 1.50)/2.04 =0.24
Prob (greater than 100%) = 0.595.
For the change NIP1NIPO
z = (1.0 2.14)/2.75 = 0.41
Prob (greater than 100%) = 0.659.
The restriction to 27 experiments, and to a smaller recommendation domain, has improved the risk
probabilities for the NIPONOPO change, which was the effect which suggested the omission. There is little
affect on results for the NIPINIPO change.
Document 2C
GHANA FERTILIZER PLACEMENT TRIALS
This is a summary of an initial look at data provided with potential for a recommendation. The
recommendation to be made in this instance was that a change of treatment would make no difference to
yield, while reducing costs.
Data was available from 27 experiments, 11 in 1982 and 16 in 1983. From each experiment the following
information was extracted:
Site name
Year
Yield for treatment TIPI (averaged over types of N)
Yield for treatment T2P2 (averaged over types of N)
Difference T2P2 TIPI
Mean of TIP1 and T2P2
Rainfall at nearest source of information
The comparison between TIPI and T2P2 was a comparison of previous recommendation with prospective
recommendation. The calculated data is shown
Site Year TIP1 T2P2 Diff Mean Rainfall(dist)
Agogo 1982 6.24 7.05 +0.81 6.64 676 (30miles)
Mampong 1982 4.96 4.80 0.16 4.88 676 (25miles)
Suromani 1982 5.84 5.35 0.49 5.60 754 (30miles)
Wenchi 1982 3.93 4.13 +0.20 4.03 754 (35miles)
Abetifi 1982 4.78 6.75 +1.97 5.76 835 (Omiles)
Akuase 1982 4.28 4.17 0.11 4.22 1215 (60miles)
Enyan D. 1982 2.96 3.42 +0.46 3.19 915 (20miles)
Gomoa Adam 1982 2.65 3.33 +0.68 2.99 1075 (20miles)
Simbrofo 1982 3.38 2.71 0.67 3.04 915 (30miles)
KwamiKrom 1982 3.72 2.68 1.04 3.20 935 (mean)
Dzolokpuita 1982 2.40 2.06 0.34 2.23 739 (50miles)
Anwomaso 1983 0.60 0.56 0.04 0.58 635 (Omiles)
Ejura 1983 4.70 4.33 0.37 4.52 606 (50miles)
Techiman 1983 0.63 0.82 +0.19 0.72 606 (20miles)
Yamfo 1983 0.16 0.21 +0.05 0.18 606 (10miles)
Pepease 1983 3.80 4.08 +0.28 3.94 677 (10miles)
Konko 1983 3.59 3.11 0.48 3.35 388 (Omiles)
Jukwa 1983 3.38 3.48 +0.10 3.43 387 (45miles)
Bawjiase 1983 1.73 1.78 +0.05 1.76 387 (15miles)
Logba 1983 5.15 4.58 0.57 4.86 369 (20miles)
Okadzakrom 1983 5.12 5.06 0.06 5.09 523 (80miles)
Dzalele 1983 2.24 2.64 +0.40 2.44 298 (30miles)
Matse 1983 4.63 4.21 0.42 4.42 298 (30miles)
Damongo 1983 2.18 2.12 0.06 2.15 750 (60miles)
Tampion 1983 0.88 1.01 +0.13 0.94 485 (15miles)
Salaga 1983 2.14 2.22 +0.08 2.18 750 (90miles)
Navrongo 1983 4.47 5.29 +0.82 4.88 485 (60miles)
Initial inspection of the data did not suggest any substantial difference T2P2 TIPI, nor that the difference
values were related to rainfall or, indeed, to site mean yield. Graphs of difference against rainfall, and of
difference against mean site yield supported these impressions. A graph of mean site yield against rainfall
confirmed the suspicion that these two variables were also unrelated.
A multiple regression of difference on rainfall and on mean yield confirmed the lack of relationships. The
fitted model was
Difference = 0.338 +0.067(mean yield) +0.00025 (rainfall).
The analysis of variance was
Source SS df MS
Regression 0.3945 2 0.1972
Residual 8.5694 24 0.3571
Total 8.9639 26 0.3448
In the absence of any relationship on which to base indirect assessment of yeartoyear variability, the only
possible summary of the variability must be derived from the sample of 27 year x site combinations.
Within the two sets of sites some pairs of sites were judged to be close enough to be thought of as similar.
The difference values classified by site and year are shown
1982
Site type 1
Site type 2
Site type 3
Site type 4
Site type 5
Site type 6
Site type 7
+0.81
0.16
+0.20
+1.97
+0.68
+0.46
0.34
1983
0.04
0.37
+0.19
+0.05
+0.28
+0.10
+0.05
0.57
0.06
0.42
There does seem to be some consistency of result within a site type. The analysis of variance for sites and
years has to be calculated for both orders of fitting and the results are
Source SS df MS
Between years (ignoring types) 1.4376 1 1.4376
Between types (adjusted for years) 3.2147 6 0.5358
Error 1.2633 9 0.1404
Total 5.9156 16
Between types (ignoring years) 3.4789 6 0.5798
Between years (adjusted for types) 1.1734 1 1.1734
Error 1.2633 9 0.1404
The analysis confirms the impression from the initial tabulation that there are consistent type differences,
and also a year difference, which give mean squares larger by a factor between 3 and 4 than that for the
interaction. However with only two years we cannot say anything useful about the variance component for
years. Therefore, unless some relationship of the form searched for earlier is discovered, we must rely on
the total sample of 27 values being representative of the population for which the recommendation is
sought.
The data are therefore appropriately summarised by the mean and variance of the entire sample.
Mean difference = +0.052
Variance = 0.3446
(from the total sum of squares in the Regression ANOVA)
S.E.(mean) = 0.113
The best estimate of risk for a future siteyear combination is obtained by assuming that the difference is
normally distributed with mean +0.052 and standard deviation 0.59 ( = squareroot of 0.3446).
Thus the risk probability that the treatment difference for T2P2 TIPI is less than zero is calculated from
the normal deviate
z = (0 0.052)/0.59 = 0.09
Prob (difference less than zero) = 0.46.
For the risk probability of a difference less than 0.5 (that is of a disadvantage from the recommendation of
0.5)
z = (0.5 0.052)/0.59 = 0.93
Prob (difference less than 0.5) = 0.176.
For the risk probability of a difference less than 1.0
z = (1.0 0.052)/0.59 = 1.78
prob (difference less than 1.0) = 0.04.
Document 2D
SIMULATED SITE x YEAR DATA AND ANALYSIS
1. Simulation Model
A simulation set of data for 100 site x year combinations was constructed to illustrate and test methods of
assessing the risk attached to recommendations. The data set was constructed according to the following
model.
Treatment Site Year
= Mean + + + B (Rain 7) + Error
Yield Effect Effect Effect
The model is defined for a treatment yield effect rather than a treatment yield because we are considering
the situation where a recommendation to change from one practice to a different one is intended and we
therefore wish to assess the variability of the yield difference resulting from the change. The dependence
on rainfall is included to represent some relation with an additional variable for which data would be
available for other years and sites. The numerical values used in the simulation were as follows:
Overall mean = 30
Site effects ranging between 10 and +10
Year effects ranging between 10 and +10
Rainfall values ranging between 3 and 14 (with some patterns of
consistency between sites and years)
Rainfall effect coefficient (B) = 2
Errors normally distributed with mean zero and standard deviation of 10.
The units were based, very loosely, on values in a large scale verification trial from Ghana (G.Edmeades)
in which yield differences (measured in (bags/acre)*10) ranged between 10 and +100 (the range here is a
bit less being between 14 and +69). The rainfall units envisaged are 100mm. Site and year variation is
pure guesswork but is chosen to be large enough to make a substantial contribution to the overall variation
in the sample.
Although these efforts have been made to clothe the simulation results in a degree of realism, it is
important to recognize that the actual scale of values used is completely irrelevant. The question being
asked is to what extent the known properties of the complete sample can be estimated from small
subsamples with particular structures.
2. Simulated sample values
The resulting simulation values which are used in the rest of this paper are shown.
Rainfall
Year
1 2 3 4 5 6 7 8 9 10
6 3 9
9 3 6
12 6 8
8 7 7
11 4 9
13 11 8
12 10 9
7 7 7
10 7 10
9 8 10
6 7 11 8 5
5 8 13 8 3
4 4 9 11 7
5 3 10 10 4
7 7 12 13 3
9 11 13 11 9
11 8 12 14 8
9 7 10 9 7
8 12 14 12 11
10 11 11 9 7
Treatment Yield Effect
Year
1 2 3 4 5 6 7 8 9 10
30 33 2 23 33 12 38 39 23 7
11 34 18 9 32 23 42 63 19 1
33 28 22 12 26 11 22 34 8 17
28 28 6 14 32 14 16 23 5 11
41 31 22 10 31 41 41 48 44 30
61 50 34 32 50 28 37 69 46 29
35 50 39 29 24 33 40 55 48 12
17 27 32 20 35 20 35 28 38 21
58 37 24 22 63 45 40 62 30 41
45 43 18 35 34 34 33 34 41 26
Site
Site
3. Analysis of the Complete Sample
The analysis of variance for the 10 sites x 10 years is shown.
Source SS df MS
Sites 9394 9 1044
Years 7433 9 825
Error(Interaction) 7869 81 97
Total 24696 99 249
Variance components are estimated as follows:
02 = 97
o2(s) = (1044 97)/ 10 = 95
a2(y) = (825 97)/10 =73.
The variance of a predicted value for a future site and year combination is
a2 + a2(s) + o2(y) = 97 +95 +73 = 265.
Note that for this balanced sample this would be well estimated by the variance for the total sample = 249.
4. A Sample of 7 Sites in 2 Years with 4 Common Sites
Seven sites were selected randomly from year 9 and a second random sample of seven from year 10, with
the restriction that only four sites should appear in both samples (i.e. all 10 sites were included at least
once). The resulting sample values for yield effect and rainfall are shown.
Yield Effect Rainfall
Year Year
9 10 9 10
Site Site
1 7 1 5
2 19 2 8
3 8 17 3 11 7
4 5 11 4 10 4
5 44 5 13
6 29 6 9
7 48 12 7 14 8
8 38 21 8 9 7
9 41 9 11
10 41 10 9
Examining the data we can see that we have two relatively extreme, and different years and a wide range of
rainfall levels. We assume that additionally rainfall data are available from the common sites (3,4,7,8) for
eight years (1,2,3,4,5,6,7,8) as shown.
Year
1 2 3 4 5 6 7 8
Site
3 5 12 6 8 9 5 3 10
4 3 8 7 7 7 5 3 10
7 9 12 10 9 11 11 8 12
8 7 7 7 7 10 9 7 10
We employ three forms of analysis. First analysis of variance of the observed data to estimate site, year and
error variances. Second the regression of yield on rainfall. Finally the analysis of variance for the rainfall
data to estimate site, year and error variances.
For the initial analysis of variance information about year and error variation is available only from the
common sites.
Year
9 10
Site
3 8 17
4 5 11
7 48 12
8 38 21
Source SS df MS
Sites 1490 3 497
Years 450 1 450
Error 512 3 171
The analysis of variance for the whole sample (which could be obtained directly from a computer analysis)
is:
Source SS df MS
Sites 3150 9 350
Years 450 1 450
Error 512 3 171
The estimates of components of variance are
o2 = 171 (on 3 df)
o2(s) = (350 171)/1.4 = 128 (on 9 df but using 3 df estimate of 02)
a2(y) = (450 171)/4 = 70 (on I df)
The divisors used in calculating the variance components are (for sites) the average replication per site, and
(for years) the replication of common sites, which are the only ones providing information about the year
variation. Inevitably the estimates of variance components are very imprecise based on so few df and
observations and we have been lucky in our sample in that these variance component estimates are not far
from those calculated earlier for the complete sample (97,95 and 73).
The regression of yield effect on rainfall gives the following results:
Corrected SS for rainfall = 101
Corrected SS for yield = 4112
Corrected sum of products = 462
Regression line: y = 4.58 (rainfall) 18.11.
Analysis of variance:
Regression SS = 2113 on I df
Residual SS = 1999 on 12 df Residual MS = 167
Variance of regression coefficient = 167/101 =1.65.
If we propose to use the rainfall data for other year x site combinations to estimate year and site variation
then we shall be assuming a relationship
y = A + B (rain)
and the consequent expression for the variance for y
Variance (y) = B2 variance(rain) + (mean rain)2 variance(B).
We can split the variance(rain) into components for sites, years and error, but the same is not feasible for
Variance(B).
The analysis of variance for rainfall is:
Source SS df MS
Sites 70.84 3 23.61
Years 82.97 7 11.85
Error 44.91 21 2.14
Total 198.72 31 6.41
The estimates of variance components are
02 = 2.14 (on 21 df)
o2(s) = 2.68 (on 3 df)
o2(y) = 2.43 (on 7 df)
These variance components are for rainfall and to make them relevant to the variance of yield effects we
multiply by the ratio of the total variances for yield effects (predicted from rainfall) and for rainfall
Variance (y) = (4.58)2 6.41 + (7.91)2 1.65 = 237
giving the ratio = 237/6.41 = 37.0
giving estimates of variance components for yield effect variation, based on the regression model, of
02 = 2.14 x 37 = 79 (on 21 df)
o2(s) = 2.68 x 37 = 99 (on 3 df)
(2(y) = 2.43 x 37 = 90 (on 7 df).
These estimates are based on the assumption that the split of the variability of rainfall into site, year and
error components is similar to that for yield effect variation. This assumption is dubious, to put it mildly,
but seems to be the only way of utilizing the regressionbased information on variation.
If we combine the two sets of variance component estimates, weighting them by their degrees of freedom
we get the following:
Estimated from Data Regression Combined
02 171 (3df) 79 (21df) 90
y2(s) 128 (9df) 99 (3df) 121
o2(y) 70 (ldf) 90 (7df) 87
The resulting predicted variance for a future site and year combination is
02 + 42(s) + o2(y) = 90 + 121 + 87 = 298.
This is really very respectably close to the "correct" value of 2651from the complete sample. I suspect that
this is a distinctly lucky sample.
5. A Sample of 8 Sites in 2 Years with 6 Common Sites
As in example 3 two samples of 8 sites from years 7 and 8 were selected randomly with the constraint that
only six common sites were included. The additional rainfall data available is for years 1 to 6 of the
common sites (1,4,6,8,9,10).
Sites
Yield effects
Years
9 10
Sites
Rainfall
Years
9 10
7 11
13
Additional rainfall data
Year
1 2 3 4 5 6
Site
1 4 6 3 9 10 6
4 3 8 7 7 7 5
6 8 13 11 8 12 9
8 7 7 7 7 10 9
9 7 10 7 10 13 8
10 9 9 8 10 8 10
The analysis of variance for the yield effect data is calculated in two stages, first for the common sites and
then for the full set of 16 observations.
Common sites Full
Source SS df SS df MS
Years 261 1 261 1 261
Sites 1597 5 2559 9 284
Error 544 5 544 5 109
Total 2402 11 3364 15 224
The SS for Years and Error from the first analysis are transferred to the second analysis and the Sites SS in
the second analysis is calculated by subtraction from the Total SS in that second analysis.
The regression analysis is calculated as shown:
Corrected SS for rainfall= 152
Corrected SS for yield = 3364
Corrected sum of products = 538
y = 3.54(rain) +4.45
Regression SS = 1904 on I df
Residual SS =1460 on 14 df Residual Mean Square = 104
Variance of regression coefficient = 104/152 =0.68
The analysis of variance for the additional rainfall data is shown
Source SS df MS
Years 50.22 5 10.04
Sites 78.89 5 15.78
Error 66.45 25 2.66
Total 195.56 35 5.59
The variance component estimates for rainfall are
a2 = 2.66
a2(s) = 2.19
o2(y) = 1.23
Variance(y) = B2 variance (rain) + (mean rain)2 variance (B)
= (3.54)2 5.59 + (8.11)2 0.68 = 115.
Hence, ratio for scaling up variance component estimates is 115/5.59 = 20.5, and the estimated variance
components for yield effects, based on the regression are
02 = 20.5 x 2.66 = 55 (on 25 df)
a2(s) = 20.5 x 2.19 = 45 (on 5 df)
a2(y) = 20.5 x 1.23 = 25 (on 5 df).
The weighted averages of the two sets of estimates of variance components are
Estimated from Data Regression Combined
o2 109 (5 df) 55 (25 df) 64
o2(s) 109 (9 df) 45 (5 df) 85
a2(y) 25(1 df) 25 (5 df) 25
The resulting predicted variance for a future site year observation is
02 + o2(s) + o2(y) = 64 +85 +25 = 174.
Not so good as the first example, but still not too bad an estimate of the "correct" value of 265.
6. A Sample of 4,4 and 8 Sites in 3 Years
The sample data, randomly selected from the first three years, to include two sites common to all three
years, is shown:
Yield effects
Years
1 2 3
30 33
34
28
Sites
1
2
3
4
5
6
7
8
9
10
Rainfall
Years
1 2 3
4 6
9
3
The spread of rainfall values is not so wide and there seems to be a lot of inconsistency of site and year
differences in the yield effects data. It is likely that this data set will prove more difficult.
Sites
1
2
3
4
5
6
7
8
9
10
The additional rainfall data is from sites 1,2,4 and 8 for years 4 to 10:
Years
4 5 6 7 8 9 10
Site
1 9 10 6 7 11 8 5
2 6 12 58 13 8 3
4 7 7 5 3 10 10 4
8 7 10 9 7 10 9 7
The analysis of variance for the yield effects data has two forms because of the nonorthogonality of sites
and years (even within those sites occurring in more than one year). The sums of squares given here are for
fitting years (ignoring year differences) and for sites (adjusting for year differences).
Source SS df MS
Years 726 2 363
Sites 846 9 94
Error 785 4 143
Total 2357 15 157
Immediately there is a problem in estimating the variance component since our estimate of the site variance
would be negative. With small df (for Error) this can happen quite often. We shall assume that the estimate
of the site variance component is zero and estimate the error variance from the combined sums of squares
for Sites and Error:
02 = 125 (13 df)
a2(s) = 0
o2(y) = (363 125) /3.33 = 71 (2 df).
The regression calculations are
Corrected SS for yields = 2357
Corrected SS for rainfall = 90
Corrected sum of products = 204
y = 2.27 (rain) +10.22
Regression SS = 463 on I df
Residual SS = 1894 on 14 df
Residual MS = 135
Variance of Regression coefficient = 135/90 = 1.5.
The analysis of variance for the additional rainfall data is shown
Source SS df MS
Years 117.21 6 19.54
Sites 13.43 3 4.48
Error 47.07 18 2.62
Total 177.71 27 6.58
The estimates of variance components for rainfall are
a2 = 2.62 (on 18 df)
o2(s) = 0.27 (on 3 df)
o2(y) = 4.23 (on 6 df).
The scalingup factor is
((2.27)2 6.58 + (7.71)2 1.50 )/ 6.58 = 18.7.
The scaledup estimates of variance components are
02 = 49 (on 18 df)
o2(s) = 5 (on 3 df)
o2(y) = 79 (on 6 df).
The weighted averages of the variance component estimates are therefore
Estimated from Data Regression Combined
02 125 (13 df) 49 (18 df) 81
o2(s) negative 5 (3 df) 5
S2(y) 66(2 df) 79 (6 df) 76
The resulting predicted variance for a future site year combination is
o2 + o2(s) + o2(y) = 81 +5 +76 = 162
Given the failure to get any useful estimate of between site variance this is not as bad as might have been
expected but it is considerably less than the "correct" value.
7. Summary
The sets of estimated values of variance components are
Full data 7 sites 8 sites 4,4 and 8 sites
2 years 2 years 3 years
a2 97 90 64 81
a2(s) 95 121 85 5
o2(y) 73 87 25 76
Combined 265 298 174 162
With two obvious exceptions the individual estimates of variance components are reasonably close to the
"correct" values from the full data. Nevertheless these results do emphasize the wellknown difficulty of
estimating variance components accurately from small df.
To illustrate the problems in estimating variance components, the 100 values in the total sample were split
into four sets of 5 sites by 5 years. An analysis of variance was calculated for each set and the three
variance components calculated. The results were
Sites (1,2,3,4.5) (1,2,3,4,5) (6,7,8,9,10) (6,7,8,9,10)
Years (1,2,3,4,5) (6,7,8,9,10) (3,4,5,6,7) (1,2,8,9,10)
a2 33 102 88 116
a2(s) 3 163 negative 64
a2(y) 85 167 14 53
Combined 121 432 102 233
These results make those from our three examples look quite good. The methods used to predict the
variance for treatment yield differences, although they have rather dodgy aspects, seem to give reasonable
estimates, given the amount of initial data. We probably cannot hope to get better estimates of the
combined variance without considerably more information.
References
Mead, R., 1990a, Representation of Risk. CIMMYT Training Working Document IB.
Binswanger, H.P., and B.C. Barah. 1980. Yield risk, risk aversion and genotype selection: Conceptual
issues and approaches. ICRISAT Research Bulletin Series No. 2.
Mead, R., 1990e, Sweeping Methods for Analysis of Variance. CIMMYT Training Working Document 3C.
Mead, R., 1990f, Precision of Net Benefit Analysis. CIMMYT Training Working Document IA.
I I'll lim.ll M,11:1, mid vvlwll 1llIj)f4)VtIIwlll
