and the variance random element are the two main parameters describing the response pattern
of a cropping system Piepho, 1998. Different approaches to stability analyis differ in how the
random term is further partitioned.
The measurement of yield stability over time involves at least three components: 1 the rela-
tionship of yield with the local environment, 2 the average yield level and 3 the variability of
the yield Mead et al., 1986. Univariate stability measures are divided into two broad categories
that use either variance or regression methods. The most important variance parameters are the
within-treatment mean square s
2
, the coefficient of variation CV, the ecovalence W
2
Wricke, 1962 and the stability variance s
2
Shukla, 1972. The two latter parameters measure the
contribution of the treatment to the treatment × environment interaction Callaway and Francis,
1993. There are several approaches that combine the mean and variance of a system. Recently
Kang 1993 developed a yield stability YS statistic that combines yield and stability of per-
formance into a single selection criterion. It is often found that variance is related to the mean.
For this reason, the square root of the variance is usually standardised by the mean, which leads to
the coefficient of variance CV.
A common approach to stability analysis is to regress the performance of the system onto an
environmental index computed as the mean of all observations in an environment. The index may
be taken as a measure of the productivity of an environment. Regression techniques used to de-
velop stability parameters are based on linear slope and deviation from that slope. Systems
where the regression has a relatively large slope show an above-average response to improved en-
vironmental conditions, as indicated by the envi- ronmental index. The regression approach was
first suggested by Yates and Cochran 1938, fol- lowed later by Finlay and Wilkinson 1963 and
Eberhart and Russell 1966. Piepho 1997 dis- cussed the regression method in a mixed model
context. A stable system has been defined as one that changes least in response to changes in the
environment. Finlay and Wilkinson 1963 sug- gested that slopes with b B 1.0 indicated better
adaptation to poor environments, while genotypes with b \ 1.0 are best used in superior environ-
ments. A cropping system with an estimate of b equal to unity shows an average response to envi-
ronmental conditions, as measured by the envi- ronment mean. But it is not obvious why a system
with an average response should be particularly stable Dyke et al., 1995. Several statistical and
biological limitations of the regression method were reported by Crossa 1990, Guertal et al.
1994 and Piepho 1998.
Alternatively, stability may be assessed in terms of the risk of a system falling below a certain level
or the risk of one system being outyielded by another Piepho, 1998. In many situations, e.g. in
subsistence farming, risk considerations are more important than the concern about variability
alone.
In agricultural research the analysis of yield stability has been largely confined to multienvi-
ronmental trials of crop cultivars. The idea of applying methods of stability analysis to cropping
systems is not novel Willey, 1979; Mead and Riley, 1981. Hildebrand 1984, Raun et al.
1993 and Guertal et al. 1994 employed stability analysis in long-term experiments to evaluate fer-
tilisation treatments. Piepho 1998 emphasizes that methods for comparing the stability of culti-
vars can also be used for comparing different agronomic treatments in general, of which culti-
vars are but a special case.
The aim of the present paper was a to evalu- ate the effect of various crop sequences and fertil-
isation treatments on maize and wheat yields in comparison with maize and wheat monocultures
on the basis of almost 40 years of data, and b to use conventional analysis of variance as well as
the variance and regression methods of stability analysis to characterise the effect of experimental
treatments on yield stability.
2. Materials and methods
The long-term experiment was set up at Mar- tonva´sa´r N 47°21, E 18°49, Hungary in 1961,
with the same treatments applied to the same plot year after year. The soil of the experimental area
was a humous loam of the chernozem type with forest residues, slightly acidic in the ploughed
layer, with poor supplies of available phosphorus and good supplies of potassium. The major char-
acteristics of the experimental soil in different treatments are summarized in Table 1. The parti-
cle size mm distribution in the 0 – 90 cm level was: \ 0.25: 6 – 7; 0.25 – 0.05: 19 – 24; 0.05 – 0.01:
32 – 34; 0.01 – 0.005: 2 – 7; 0.005 – 0.001: 16 – 20; B 0.001: 13 – 21. The carbonate content was ex-
tremely high below 30 cm. The N and humus contents exhibited parallel trends and the humus
content was only higher than 1 to a depth of 60 cm. The quantity of clay declined with depth,
being 21 in the 0 – 30 cm layer, 16 in the 30 – 60 cm layer and 13 in the 60 – 90 cm layer. The clay
mineral composition was 27 – 42 illite and 21 – 38 smectite in the 0 – 90 cm layer, below which
smectite became dominant.
The total precipitation for the study period 1961 – 1998 averaged 539 mm, which ranged
from 386 to 760 mm. Differences in precipitation contributed to different yield responses, as shown
by the various year × treatment interactions. Rain during the reproductive period of wheat and
maize has traditionally been assumed to decisively influence grain yield. In the dry years the average
rainfall during the vegetation period Apr. – Sept. was 243 mm and the total annual rainfall 474
mm, compared to 346 and 576 mm, respectively, in the wet years. There was a significant positive
correlation between the grain yield of maize and the amount of precipitation during the vegetation
period.
2
.
1
. Treatments The crop rotation experiment is a two-factorial
split-plot with four replications. The main plots are the crop sequences and the subplots are the
fertiliser treatments, in a randomized design. The subplot size is 7 m × 7 m = 49 m
2
and the main plot size is 49 × 5 fertiliser treatments = 245 m
2
. The main plots consist of seven crop sequences,
namely maize and wheat monocultures, three di- cultures, one triculture and a Norfolk crop
rotation:
1. Maize monoculture 2. Wheat monoculture
3. 3 years alfalfa – 5 years maize 4. 3 years alfalfa – 5 years wheat
5. 2 years wheat – 2 years maize 6. 3 years alfalfa – 3 years maize – 2 years wheat
7. Maize – spring barley – peas – wheat
Each cycle of Treatments 3, 4 and 6 takes 8 years, while that of Treatments 5 and 7 takes 4
years, so the full experimental cycle is 8 years. Four complete cycles of 8 years, or eight cycles of
Table 1 Long-term effect of crop rotation and fertilisation treatments on major properties in the 0–20 cm soil layer at Martonva´sa´r
a
Treatments pH
KCl
AL-P
2
O
5
ppm Humus
AL-K
2
O ppm Crop sequences
1. Maize monoculture 2.81 d
5.91 ab 53.3 c
272.5 b 317.5 a
76.0 ab 5.79 ab
2. Wheat monoculture 3.24 ab
264.9 b 43.3 c
5.51 b 3.39 a
3. 3 years alfalfa–5 years maize 3.31 a
6.06 ab 4. 3 years alfalfa–5 years wheat
40.7 c 267.7 b
72.7 b 5. 2 years wheat–2 years maize
335.3 a 6.33 ab
2.99 cd 6.01 ab
6. 3 years alfalfa–3 years maize–2 years wheat 255.8 b
41.2 c 3.07 bc
2.96 cd 6.56 a
92.8 a 313.0 a
7. Norfolk crop rotation Fertilisation treatments
6.23 a A: Unfertilised control
25.8 d 3.05 b
251.2 c 309.1 a
67.3 b 6.25 a
B: Farmyard manure+NPK 3.21 a
5.93 b C: Recycled crop residues+NPK
57.7 c 3.11 ab
283.3 b 5.89 b
3.08 b 291.9 b
D: NPK equivalent to that extracted by the crop 57.2 c
312.0 a 91.9 a
5.82 b 3.10 ab
E: NPK for high yield level
a
Data followed by the same letter within a column or group of treatments do not differ significantly at the P50.05 level.
4 years have taken place since the start of the experiment. The proportions of maize and wheat
were 25.0, 37.5, 50.0, 62.5 and 100 depending on the type of crop sequence. Because of the different
number of replications within the year the effects of the various crop sequences are compared with
the monocultures. The various crops represent the treatments and it follows from the structure of the
experiment that the plots being compared do not always contain the same crops in the same years.
However, due to the long period which has passed since the experiment was set up, it is possible to
compare the treatments over an adequate number of years. The subplots in the experiment represent
five different fertilisation treatments:
1. Control, without fertiliser 2. 60 t ha
− 1
farmyard manure every 4 years + NPK supplementation
3. 5 t ha
− 1
straw or 7 t ha
− 1
maize stalks each year + NPK supplementation
4. NPK fertiliser equivalent to that extracted by the crop
5. NPK for a yield of 15 t ha
− 1
maize and 10.5 t ha
− 1
wheat.
2
.
2
. Analysis of 6ariance The response of a system is governed by a
systematic component m
i
and a random compo- nent e
ij
. In the usual analysis of variance ANOVA setting, the random component e
ij
is partitioned into an environmental main effect and
interaction Piepho, 1998: y
ij
= m
i
+ u
j
+ e
ij
where y
ij
is the performance of the ith system in the jth environment; m
i
is the mean expected value, i.e. the systematic effect of the ith system;
u
j
is a random environmental main effect and e
ij
is a random residual comprising both system × envi-
ronment interaction and error. In the rotation experiment the analysis is based
on the combined evaluation of yield data from the maize and wheat monocultures and from the vari-
ous crop sequences for comparable years. Thus, depending on the type of crop sequence, the com-
bined analysis could be carried out for 10, 15, 18 or 23 years for maize and 8, 9, 20 or 23 years for
wheat. The analysis of variance was carried out for
each experimental year and the homogeneity of error variance for the different years was studied
using the Bartlett test Sva´b, 1981. The structure of the combined analysis of variance over years
was determined on the basis of Gomez and Gomez 1984. This combined variance analysis
provided an overview of the magnitude of varia- tion between the experimental years, the variation
between the treatments and especially the treat- ment × year interaction.
For evaluating the effect of crop rotation and fertilisation on maize and wheat grain yield the
hierarchical analysis of variance in the MSTAT-C program was used. The hierarchical analysis of
variance was computed between crop rotation within fertilisation.
2
.
3
. Stability analysis The stability analysis of the experimental treat-
ments was conducted using the variance and re- gression methods. Among the variance methods,
the ecovalence index W
2
of Wricke 1962, the stability variance index s
2
of Shukla 1972 and the yield stability index YS of Kang 1993 were
calculated using the STABLE model elaborated by Kang and Magari 1995. The stability vari-
ance index s
2
of Shukla 1972 is very closely related to the ecovalence index W
2
of Wricke 1962 Piepho, 1998. The regression model can
be written Piepho, 1998 as y
ij
= m
i
+ b
i
u
j
+ d
ij
where b
i
is a regression coefficient corresponding to the ith system; u
j
is an effect of the jth environ- ment and d
ij
is a random deviation from the regression line.
Linear regression analysis was carried out ac- cording to the method of Finlay and Wilkinson
1963, without the use of data transformation. Stability analysis is the linear regression of treat-
ment yield on the year environment mean yield average yield of all comparable treatments in a
given year. The effect of crop sequences and fertilisation treatments in different environments
is characterised by parallel or intersecting straight lines, after Raun et al. 1993.
Steps to determine differences in the slope and intercept components for linear equations from
the stability analysis were derived from Mead et al. 1993. The regression test for differences be-
tween level regressions was used from the MSTAT-C statistical program.
The data were processed on an IBM-compatible computer using the MSTAT-C and SPSS 9.0 for
Windows programs MSTAT-C, 1991; SPSS for Windows, 1999.
3. Results
