SOME STATISTICAL ASPECTS OF PARTITIONING GENO-

TYPE-ENVIRONMENTAL COMPONENTS OF VARIABILITY

G. K. SHUKLA
A.R.C. Unit of Statistics, 21 Buccleuch Place, University of Edinburgh,
Edinburgh EH8 9LN
Received 1 7.xii.7 1

1. INTRODUCTION
RECENTLY, Freeman and Perkins (1971) examined some of the existing
methods of partitioning genotype-environmental component of variability
and their statistical validity. They have considered the usual practice of
calculating the regression of genotype means on the environmental means
calculated by taking the average of all genotypes in that environment, first
used by Yates and Cochran (1938) and later used by Finlay and Wilkinson
(1963) and since then used by several other authors of which references have
been cited by Freeman and Perkins (1971). They have shown the statistical
invalidity of using such regressions and their sums of squares for testing
homogeneity. They have further suggested that some genotypes should be
taken in each environment (not included in calculating the mean of each
genotype) as a measure of environment and regression of genotype means
should be calculated on the independent measure of environment thus
making the procedure statistically more valid.
In the first part of this paper we have shown how by looking at the model
in a different way one could draw statistically valid conclusions of certain
hypotheses from the same analysis of Yates and Cochran (1938). In the
second part of this paper we have suggested a method of estimating a
component of genotype-environmental interaction corresponding to each
genotype, thus giving a better measure of genotype stability. This paper
is concerned with the presentation of practical methods rather than with
statistical theory (to which references are given); it does not itself contain
much that is new, but a more general statistical treatment is being prepared
and will be published elsewhere.

2. REGREssIoN OF GENOTYPE MEANS ON ENVIRONMENTAL MEANS

We have used mostly the same model and notation as used by Perkins
and Jinks (1968) for t genotypes, s environments and r replications of each
genotype within each environment and the model could be represented as,
= ,a+d1+e+gJ+eJ, (1)
where is the grand mean, d1(i = 1, ..., t) the additive genetic contribution
of the ith genotype, €1(j = 1, ..., s) the additive environmental contribution
of the jth environment, g5 the genotype-environment interaction of the ith
genotype in the jth environment and eJk(k = 1, .., r) is the residual vari-
ation contributed by the kth replicate of the ith genotype in the jth environ-
ment. We shall assume that
E(e15) = 0; V(e5) = (2)
237
238 G. K. SHUKLA

e
for all i and j where is the mean of e over r replications and a is the
within environment error variance for the mean of r replications. We shall
further assume that environment effects e/s are random effects with popu-
lation mean zero and variance o. A random sample of s environments
has been selected from an infinite population of environments. We shall
estimate cr as usual by o, where
= (,IJk—.9J)2Istr(r— 1), (3)
i jic
with st(r — 1) degrees of freedom. Hereafter we shall work with jY, (the mean
of r replications of the ith genotype at the jth environment). In the present
paper we shall confine ourselves to the genotype-environment (G x E) part
of the analysis of variance.
Working with the means, the model in (1) can be written
= (4)
Putting
= e+ 'ij and c = ]j + ej.
we obtain from (4)
= (5)
Putting b = b —b' where b' b/t we obtain from (5)
= i+d+ (1 +h') +be2+;2. (6)
The model in (5) is reparameterised in (6) in such a way that = 0.
When all b's are 0 (or b are equal) then the model in (6) becomes as in
(7),
= p+d1+e(l+b')+1. (7)
Thus the problem of testing the equality of all b"s becomes the problem
of testing model (7) against model (6). This is equivalent to testing the
presence of the non-additivity term b1e in (6) when E are taken as fixed
effects. The test for presence of non-additivity of this type was given by
Mandel (1961), which is a generalisation of Tukey (1949). Same results
hold good even if are taken as random effects. He has estimated b by
b and calculated the sum of squares due to non-additivity (S) as follows:

S (8)
and "Balance" = Interaction Sum of Square (G x E) — S.
The sum of squares in (8) is similar to that calculated by Yates and
Cochran (1938) and the same as that due to heterogeneity in regression in
Freeman and Perkins (1971) table 2, with z5 in Freeman and Perkins'
notations replaced by The estimated regression coefficients, b,
here are similar to those given by Yates and Cochran (1938) except that
we have regressed deviation of genotype means from the environmental
 GENOTYPE-ENVIRONMENTAL COMPONENTS 239

means rather than genotype means. The sum of squares S indeed is a ratio
of quartic terms and quadratic terms in they's.
In the absence of interaction (g15 = 0 for all i and j) it can be shown
that S/o and "Balance "/a are both independently distributed as x2 °"
(t— 1) and (t— l)(s—2) degrees of freedom, respectively. However, in the
presence of interaction SJcr and "Balance "/o are not distributed as x2
even if all b1 = 0, though they are independently distributed of each other.
In the presence of interaction the appropriate test statistic for all b = 0 will
be F' as given in (9).

F' — S/(t—l) 9
"Balance "/(t— l)(s—2)
F' will be distributed as F on (1— 1) and (t — 1) (s —2) degrees of freedom.
The same test was proposed by Perkins and Jinks (1968) and this statistic
gives correct probability level. Equality of any two b1's could be tested by
doing the similar analysis for any two genotypes of interest. The above
argument can be generalised for other arrangements of genotypes within
and between environments and also in the presence of non-orthogonality
in the data (Milliken and Graybill, 1970).
To use the 's in the usual sense of regression would not be valid, but
they give rough guidance about the relation of genotype means to environ-
mental means. The 's are biased estimators of the b's. In general this
bias will be small when r is large but could be corrected in the way suggested
by Tai (1971). The partition of sum of squares into two components is also
only approximate, but this may be quite satisfactory for practical purposes.
More efficient estimates of b's and e5's, and their sumof squares could be
obtained by fitting the model in (6) by a non-linear least squares method as
suggested by Elston (1961) and Tai (1971) and approximate tests could be
obtained. Under these circumstances, independent measure of environment
based on more genotype means may not be worthwhile.

3. COMPONENTS OF INTERACTION SUM OF SQUARES

The characterisation of genotypes on the basis of regression coefficients

may not be very effective when only a small fraction of the interaction sum
of squares (G x E) can be attributed to heterogeneity among the regressions.
It might be then of great interest to partition G x E into I components, one
corresponding to each genotype, as mentioned by Baker (1969). Put
= V.
Let us further assume that
E(v) = 0; V(v1) = cr; E(v1, v1.) = 0 for i i' orj
V(g5) = 2;
E(ge1) = 0; i = 1, ..., t. (10)
Then,
=

In the above expression cr could be taken as the sum of two components,

viz, within environmental variance (o) and between environmental variance
(a2) of the ith genotype (after correcting for additive common effect of
240 G. K. SHUKLA

environment €3), and we shall name it the "stability variance" of the ith
genotype. We shall call a genotype stable if its stability variance (a) is
equal to within environmental variance (o-) which means that (42 = 0.
Relatively large values of a will indicate more instability of genotype.
Estimation of a is analogous to the problem of estimating heterogeneous
variances in a two-way classification when they change in one way considered
by Ehrenberg (1950) and later by Russell and Bradley (1958). Rao (1970)
has further generalised the above prcoedure for any classification and also
considered some optimum properties of the above estimators. Without
going into detail, we give the unbiased estimators of a, denoted by a, as

[t(t— 1) (91—y. —9.1—f-fl..)2

(s— 1)(t— 1)(t—2)
—
-9.,+5..)]
[t(t — I) (u11-ü1.)2— (u-ü1.)z] (11)
(s— l)(t- l)(t-2)
where
u =—J and un/s.

They are obtained as linear combinations of squares of residuals

therefore, they are independent of /L, d and variance of c. It is not difficult

to verify that they are unbiased estimators of o. Under the assumption of
symmetrical distribution of o's, Rao (1970) has proved that on average they
have minimum variance among all possible quadratic unbiased estimators
(MINQUE) of o. It is not difficult to see that their mean is the same as
the mean sum of squares (G x E). Therefore, by multiplying each ó by
(t — 1) (s — 1) ft we shall obtain t components of G xE, one corresponding to
each genotype. These components are not statistically independent; as
they are differences of two sums of squares, they can be negative, but
negative estimates of variances are not uncommon in variance components
problems.
The essential difference between the present approach and Baker's
(1969) approach is this that his method estimates ((t_2)cr+52)/t while the
above method estimates a where
=
- i =1 a.
The same is true for deviation from regression component.
The variance of o is not only a function of a but of variances of other
genotypes o(j i) taken in trial. Such estimators are only available when
3. The problem of testing homogeneity of a's has been considered by
Russell and Bradley (l958),Johnson (1962), Han (1969) and Shukla (1971).
The method proposed by Shukla could be easily applied for testing the
homogeneity of all the variances or any pair of them.
It might be of some interest to test whether certain genotypes are stable
or not. Johnson (1962) suggested a test criterion based on the ratio â/ô.
 GENOTYPE-ENVIRONMENTAL COMPONENTS 241

It is difficult to derive the exact distribution of â, but when t is large the

variance of â is approximated as

V(â) 2o/(s—l). (12)

The above expression helps us in obtaining an approximate distribution.
When t is large, (s— l)â/a will be approximately distributed as on (s— 1)
degrees of freedom; thus, under the hypothesis that (42 = 0, F* will have
an approximate F distribution on (s— 1) and st(r— 1) degrees of freedom
where
F* = (13)
When & is negative or less than â then (42 may be taken equal to zero
as usual.

4. FURTHER EXTENSION OF MODEL

For further progress in the interpretation of instability, we shall reconsider
the model in (5). To keep the treatment general, we replace by z1in the
non-additive term bE and rewrite the model as
= (14)
where z is a measure of some characteristic of jth environment; by taking
deviation from the mean we can make z5 = 0, and (4 = b'z.
We shall further assume that
V(5) = s; (i = 1, ..., t),
and then discuss the method of estimation of s. The usual estimator of b,
by the method of unweighted least squares, can be obtained as

= (u—u1.)1
(15)

Using methods as in Section 3, unbiased estimators of s for extended

model in (14) could be obtained as

where
(t_2)(s_2) [s_ t(t-l)] (16)

S=
j1 (u15_u._z)2.
It is apparent that the model in (14) is just the extension of the model in
(7) to take into account a covariate ;. The estimators obtained in (16)
are quadratic (in "s) estimators of s and have the properties of MINQUE
estimators. When t is large, the variance of . can be approximated by

V() = (17)
(s-2)
 242 G. K. SHUKLA
and their distribution can be approximated by x2 on (s —2) degrees of
freedom. An approximate significance test against d is possible as in (13).
If some of the genotypes become stable after taking the covariate into
account, it may be inferred that the instability was introduced by the linear
effect of the covariate and such information may be useful. The above
approach could also be extended to more than one covariate.

5. RELATIONSHIP BETWEEN REGRESSION APPROACH AND THE

"STABILITY VARIANCE" APPROACH
The definition of stability is similar to Baker (1969) and Eberhart and
Russell (1966). A significant departure of the regression of a genotype from
zero will be indicated by a relatively high "stability variance ", but a
regression coefficient of zero need not mean that the particular genotype is
stable. A zero regression will be obtained if there is no linear relationship
between genotype mean and environmental mean, yet the "stability
variance" (ox) may be greater than o.
Once some of the genotypes are found unstable, it may be of interest to
examine further the reasons for instability. The approach of Section 4 may
be followed if observations are available on covariates which are likely to
affect the genotypes differentially. We can examine the effect on stability
variance of linear regression on environmental means by the method in the
previous section. To examine any effect of differential fertility we have
used z1 = 9•—5• as used by many other authors mentioned in Sections 1
and 2. It must be noted here that the estimators of s obtained by putting
in (16) will not be quadratic estimators ofy's and therefore the
optimum properties described in Section 4 may not hold. Again as men-
tioned earlier the effect of departure from optimality may be small when
u is large. The effect of such differential regression on the stability of
genotypes could be tested as above under the assumption that 's are con-
stant. The estimation of individual s is analogous to what Perkins and
Jinks (1968) have suggested by the mean sum of squares /(s—2) and
Baker (1969) by deviation from regression sum of squares but the above
approach has an advantage as they are unbiased estimates of s (free from
any other nuisance parameters) and the mean of is the same as the mean
sum of squares of departure from regressions (" Balance ") and this could
be taken as equivalent to dividing the "Balance" into components corres-
ponding to each genotype.
Recently Tai (1971) has worked with the above problem. The difference
between our method and his method is this, that he has considered the
model under certain side conditions on the interaction and we have not
imposed any such conditions on them. It would not be very justifiable to
impose any condition on interaction while estimating the individual com-
ponent. The definition of stability is also different. According to his
definition of stability one should have b = — I and s = o. Our definition
of stability coincides with his definition of average stability ( = 0; X =
in Tai, 1971, notations). By our definition of stability we only mean that
the performance of a genotype is sum of additive genotypic effect, additive
environmental effect and a random error without any interaction between
genotype and environment.
 GENOTYPE-ENVIRONMENTAL COMPONENTS 243

Prediction of the expected performance can also be made with reasonable

accuracy for a given environment, if either the interaction is not present or
most of it can be accounted by linear regression term (Jinks and Perkins,
1970).

6. NUMERICAL EXAMPLE
For illustration purposes we have considered the data analysed by Yates
and Cochran (1938). We shall only consider the part of the table dealing
with GxE.
TABLE I
Variety x Place totals over the years

APlaces

4 Total
Varieties 1 2 3 5 6
Manchusia l6l7 2470 1854 2187 1653 1546 l1327
Svansota l877 2575 1824 l833 1389 l438 10936
Velvet 200l 2629 1949 2202 1658 1463 l1902
Tribi l969 3392 2712 2663 l512 1936 14184
Peatland l825 2538 2192 2O05 1844 l901 l2305
Total 9289 13604 10531 10890 8056 8284 60654

Table 2 gives the values of u obtained from table 1.

TABLE 2
u1j's and regression coefficients
Places
r
Varieties 2 3 4 5 6
Manchusia —2408 —2508 —2522 09O 4l8 — 1108 —0156
Svansota l92 —l4•58 —2822 —3450 —2222 —21•88 —0Ol4
Velvet 14•32 —918 — 1572 24O 468 — 1938 —0O54
Tribi l112 67l2 6058 485O —992 2792 0609
Peatland ....328 — 1828 8•58 — l730 2328 2442 —0385

TABLE 3
Ut5 —tjZJ

Places

Varieties 2 3 4 5 6
Manchusia — 26 14 — 14l8 —2390 3.34 —223 — I677
Svansota 169 — 1360 — 28 10 ..3428 —2279 — 2239
Velvet 1343 —5•41 —1526 324 246 —2 135
Tribi 2l1l 2455 55.44 3899 l508 5017
Peatland 9.59 863 1183 —1129 7.47 1037

To obtain the component of variances on the same unit as the sum of

squares (units of single plot) in the Analysis of Variance (table 4) we have
divided them by 6.
Comparison of d's with â shows that Tribi and Peatland are unstable.
Further regression analysis shows that Tribi remains unstable, though, its
variability has reduced considerably. Peatland becomes stable after taking
covariate into consideration. Similar conclusions were drawn by Yates and
244 G. K. SHUKLA
Cochran (1938) but the above type of analysis in general may be advan-
tageous.
TABLE 4
Analysis of variance table and" stability variances " (units of single plot)
Source D.F. S.S. M.S. F
Places 5 7072 92 — —
Varieties 4 177028 — —
VxP 20 147784 7389 —

— — 2588 1.11
— — 19•60 084
— — 2273 098
— — 22553 9.69**
— — 7568 3.25**

Source D.F. S.S. M.S. F

Heterogeneity 4 77316 19329 —
Balance 16 70469 4404 —

— — 34•10 l•46
— — 4048 174
— — 4278 l84
— — 797O 3.42*
— — 2327 1.00

216 — 2328

7. SUMMARY
1. The usual regression approach of explaining genotype-environment
interaction has been considered by using a non-additive model and the
statistical validity of the analysis has been discussed.
2. Alternative approach of dividing genotype-environmental interaction
into components, one corresponding to each genotype has been proposed
and the optimum properties have been discussed.
3. The alternative approach has been extended to take into account a
covariate.
4. The relationship of new approach to the regression approach has
been discussed.
5. A numerical example has been given as an illustration.
Acknowledgment.—I am very grateful to Professor D. J. Finney and Mr H. D. Patterson
for their valuable suggestions and help. I am grateful to referees for their valuable sugges-
tions.

8. REFERENCES
BAKER, R. j. 1969. Genotype-environment interaction in yield of wheat. Can. J. Plant Sci.,
49, 743-751.
EBERHART, S. A., AND RUSSELL, W. L. 1966. Stability parameters for comparing varieties.
Crop Sci., 6, 36-40.
 GENOTYPE-ENVIRONMENTAL COMPONENTS 245

EHRENBERG, A. s. c. 1950. The unbiased estimation of heterogeneous error variances.

Bioinetrika, 37, 347-357.
ELSTON, K. c. 1961. On additivity in the analysis of variance. Biometrics, 17, 209-219.
FINLAY, K. W., AND WILKINSON, G. N. 1963. The analysis of adaption in a plant breeding
programme. Aust. 3. Agric. Ret., 14, 742-754.
FREEMAN, 0. H., AND PERKINS, JEAN 55. 1971. Environmental and genotype-environmental
components of variability. VIII. Relations between genotypes grown irs different
environments and measure of these environments. Heredity, 26, 15-23.
JINKS, J. L., AND PERKINS, JEAN, M. 1970. Environmental and genotype-environmental
components of variability. VII. Simultaneous prediction across environments and
generations. Heredity, 25, 475-480.
HAN, C. p 1969. Testing the homogeneity of variances in a two-way classification. Bio-
metrics, 25, 153-158.
JOHNSON, N. L. 1962. Some notes on the investigation of heterogeneity in interaction.
Trabajos de Estadistics, XIII, 183-199.
MANDEL, j. 1961. Non-additivity in two-way analysis of variance. 3. Am. Stat. Assoc., 56,
878-888.
MILLIKEN, 0. A., AND GRAYBILL, F. A. 1970. Extension of general linear hypothesis. 3. Am.
Stat. Assoc., 65, 797-807.
PERKINS, JEAN M., AND JINK5, j. L. 1968. Environmental and genotype environmental
components of variability. III. Multiple lines and crosses. Heredity, 23, 2 39-256.
RAO, C. K. 1970. Estimation of heteroscedastic variances in linear models. 3. Am. Stat.
Assoc., 65, 161.172.
RUSSELL, T. S., AND aRADLEY, R. A. 1958. One-way variances in a two-way classification.
Biometrika, 45, 111-129.
SHUKLA, 0. K. 1971. An invariant test for the homogeneity of variances in a two-way
classification. To appear in Biometrics.
TAI, 0. C. C. 1971. Genotype stability analysis and its application to potato regional trials.
Crop Science, 11, 184-190.
TUKEv, j. w. 1949. One degree of freedom of non-additivity. Biometrics, 5, 232-242.
YATES, F., AND COCHRAN, w. o. 1938. The analysis of group experiments. 3. Agric. Sri.,
28, 556-580.