Crop Science 43:549-555 (2003)
© 2003 Crop Science Society of America
CROP BREEDING, GENETICS & CYTOLOGY
Prediction of Cultivar Performance Based on Single- versus Multiple-Year Tests in Soybean
Weikai Yan and
Istvan Rajcan*
Dep. of Plant Agriculture, Crop Sci. Bldg., Univ. of Guelph, Guelph, ON, N1G 2W1, Canada
* Corresponding author (irajcan{at}uoguelph.ca)
 |
ABSTRACT
|
|---|
Because of the omnipresent genotype x year or genotype x location x year interactions in crop performance trials, it is commonly believed that multiple-year data should be used in selecting cultivars for the next year. An implicated but rarely tested hypothesis is that multiple-year data are more predictive than single-year data of cultivar performance in the next year. Yield data of the 1991 to 2000 Ontario Soybean Variety Trials in the 2800 Crop Heat Unit (CHU) area were used to study the power of single-year, multiple-location trials in predicting cultivar performances in the following year, and to see if data from multiple-year trials are more predictive. Mixed models were used to estimate best linear unbiased predictions (BLUP) of tested genotypes on the basis of single- or multiple-year trials, and the t-statistic of BLUP (tBLUP) was used as a measure of cultivar performance. Results indicated that a single-year, multiple-location trial had sufficient power for identifying genotypes that would perform well or poorly in the next year. Two to four years' data gave only slightly better predictions of next-year performances than single-year data but allowed more genotypes to be evaluated conclusively. The tBLUP of genotype effects based on 2 yr of multiple-location trials should be used as a basis for soybean cultivar selection and recommendation in the 2800 CHU area of Ontario.
Abbreviations: BLUP, best linear unbiased prediction • CHU, crop heat units • E, environment main effect • G, genotype main effect • GE, genotype x environment interaction • GL, genotype x location interaction • GLY, genotype x location x year interaction • GY, genotype x year interaction • L, location main effect • tBLUP, t-statistics of BLUP • Y, year main effect • OOPSCC, Ontario Oil and Protein Seed Crop Committee • OSVT, Ontario Soybean Variety Test
|
INTRODUCTION
|
|---|
REGIONAL PERFORMANCE TRIALS are conducted annually for all major crops throughout the world to help growers select cultivars for the next year. Such a decision would be straightforward were there no genotype x environment (GE) interaction (Gauch and Zobel, 1996). GE interaction is, however, almost omnipresent, which complicates the decision making.
GE interaction in multiple-location and multiple-year trials can be dissected into genotype x location interaction (GL), genotype x year interaction (GY), and genotype x location x year three-way interaction (GLY) (Comstock and Moll, 1963; Annicchiarico and Perenzin, 1994). Presence of GL within a single year necessitates multiple-location trials; presence of GY warrants multiple-year trials; and presence of GLY requires both multiple-year and multiple-location trials. Since GE interaction is almost omnipresent, and as yearly variation is typically the largest source of yield variation, it is commonly believed that the greater the number of years a genotype is tested, the more reliable its evaluation will be. An extended belief is that results based on more years of performance trials are more predictive of cultivar performance in the next year, which becomes a dogma for cultivar recommendation. Thus, it is recommended in regional performance trial reports, almost without exception, that cultivar selection should be based on multiple-location trials across multiple years. Surprisingly, the hypothesis that multiple-year data give better prediction of the next year's performance has been tested by only a few researchers (Cross and Helm, 1986; Gellner, 1989; Bowman, 1998). On the other hand, it is not feasible for researchers to make decisions on the basis of many years of testing because few genotypes (except the checks) are tested in many years and many genotypes are withdrawn from the trials if they do not perform well in the first year.
This project was set up to address the following questions: (i) what is the predictive power of single-year multiple-location trials in cultivar selection; (ii) are data from multiple-year trials more predictive than those from a single-year trial; and (iii) what is the merit of using multiple-year data?
|
MATERIALS AND METHODS
|
|---|
Data Source
Yield data from the 1991 to 2000 Ontario Soybean Variety Test (OSVT) in the 2800 crop heat unit (CHU) (OMAFRA, 1993) area of Ontario, Canada, were used in this study. The OSVT is an official annual test conducted by the Ontario Oil and Protein Seed Crop Committee (OOPSCC) and supported by Ontario Ministry of Agriculture and Food and the Ontario Soybean Growers. The OSVT assumes the functions of both soybean [Glycine max (L.) Merr.] registration trials and performance trials and is conducted across Ontario covering cultivars from 2300 to 3400 CHU. The cultivars in the 2800 CHU test belong approximately to relative maturity groups from 0.4 to 1.5. The 2800 CHU trials included four test locations, namely, Exeter, St. Pauls, Woodstock, and Winchester.
Geographically, the first three locations are in southwestern Ontario, whereas Winchester is in eastern Ontario. Although there were strong crossover GL interactions each year, the interaction pattern varied considerably across years (Yan and Rajcan, 2002). Each year, 60 to 113 adapted cultivars or breeding lines from public and/or private breeding programs were tested. Many more entries were tested in recent years, as compared with the earlier years (Table 1). Although the same sets of genotypes were tested at all locations within a single year, the genotypes varied greatly with the year. In general, about 50% or more of the entries were removed each year; and only one cultivar was tested in all 10 yr (Table 1). A total of 526 genotypes were tested during the 10-yr period. Except for the check cultivars, which were determined by the OOPSCC, cultivar sponsors were solely responsible for the entering of their cultivars into the tests. Also, in some years, as a result of poor germination, data from only three locations were collected. Consequently, the dataset was highly unbalanced, with 83.3% missing cells.
View this table:
[in this window]
[in a new window]
|
Table 1. Number of genotypes that are commonly tested in the 2800 Ontario Soybean Variety Test during the period from 1991 to 2000.
|
|
At each location, a lattice design (before 1998) or a nearest neighbor design (1998 and later) with four replicates was used. The experiments were planted according to local practice with planting rate about 50 seeds m-2. The harvested plot size was 8.25 m2 (four 5.5-m rows with 37.5-cm spacing between rows). Mean grain yield was computed in accordance with experimental design for each genotype at each location. The analysis reported here is based on mean values.
Statistical Analysis
The following model was used in this analysis:
where Yijk is the mean yield of genotype i in year j at location k, µ is the grand mean, gi is the main effect of genotype i, yj is the main effect of year j, lk is the main effect of location k, and (yl)jk is the interaction between year j and location k. The term eijk is the residual associated with genotype i, year j, and location k and consists of genotype x environment interaction (including GL, GY, and GLY) confounded with experimental error. All effects except the grand mean are assumed random, with 0 mean and normally distributed variances. Restricted maximum likelihood was used to estimate the variance components of various factors and the random effects, i.e., the best linear unbiased predictors (BLUP) of the genotypes. Likelihood ratio was used to test the significance of the estimated variance components (Littell et al., 1996). The SAS procedure PROC MIXED (SAS Institute, 1996) was used to obtain the variance components and the random effects. The SAS statements were:
Since the purpose was to predict the next-year performance of the genotypes, only genotypes tested in the year in which the cultivar performance was to be predicted were included in the analysis. For example, if data from 1991 to 1995 were jointly analyzed to generate predictions for cultivar performance in 1996, only genotypes present in 1996 would be included in the analysis.
When analysis was based on multiple-year data, the model included a fixed effect for the grand mean and random effects for genotype, year, location, and year x location interaction. The genotype x environment interaction (including GL, GY, and GLY) was implicitly stated in the SAS statements and confounded with the experimental error. When analysis was based on single-year data, the model included a fixed effect for the grand mean and random effects for genotype and location. The genotype x location interaction was implicitly stated and confounded with the experimental error. The option "DFM = SATTERTH" was used to estimate degrees of freedom for appropriate hypothesis testing.
Relevant outputs of the above SAS statements include two aspects: (i) the variance component for each random factor (genotype, location, etc.), which gives information on the relative importance of the variation sources and (ii) the estimated random effect (i.e., BLUP relative to the grand mean) for each level of all random factors. Our interest was in the BLUP of the genotype main effects. The estimated BLUP for a genotype, under the proposed model, is the predicted yield of the genotype relative to the grand mean. Whether the BLUP of a genotype is significantly higher or lower than the grand mean is tested by the t-statistics of BLUP (tBLUP), which is the ratio of BLUP to the associated prediction error, which is inversely related to the number of data points from which the BLUP was estimated. Yan et al. (2002) indicated that tBLUP is highly correlated with, and is more informative than, BLUP per se as it provides an intuitive and convenient measure of superiority. For two-tail test, a t 
2.0 indicates superiority over the average, and a t 
-2 indicates inferiority to the average at the 0.05 probability level. For one-tail test, the threshold |t| value can be reduced to 1.67 for the same level of significance. Based on this threshold level, the genotypes can be easily classified as superior (t 1.67), inferior (t -1.67), and intermediate (1.67 < t > -1.67) groups.
The predictive power of single-year, multiple-location trials was measured by the correlation coefficient between the tBLUP estimated from one year and that from the following year across genotypes. Similarly, the predictive power of multiple-year trials is measured by the correlation coefficient between tBLUP estimated from data from a number of previous years and data from the current year across genotypes. For example, the predictive power of a 5-yr test for 2000 is measured by the correlation coefficient between tBLUP based on data from 1995 to 1999 and that based on data from 2000 across the genotypes tested in 2000.
|
RESULTS
|
|---|
Variance Components
The location main effect (L) was a much greater source of variation than genotype main effect (G) and genotype x location interaction (GL) (confounded with experimental error) for all years except 1995 and 1996, and GL was always greater than G (Table 2). The large yearly GL interaction, which causes cultivar rank changes at different locations (Yan and Rajcan, 2002), warrants multiple-location trials for cultivar evaluation.
View this table:
[in this window]
[in a new window]
|
Table 2. Variance components of genotype (G), location (L), and genotype x location (GL) interaction in individual years.
|
|
When all 10 yr of data were analyzed together, the year x location interaction was found to be the most important source of variation for soybean yield, accounting for 55% of the total variation. The second largest source of variation was the year main effect (15%), whereas the location effect was only 2% (Table 3). The genotype main effect, G, explained 10% of the total yield variation, and GE, confounded with the experimental error, explained 15% of the total variation. However, the likelihood ratio test indicates that only the variance components of genotype and year x location interaction were significant (Table 3). The variance components of year and location main effects were not significant because of relatively small number of degrees of freedom and the large number of missing cells.
Theoretically, GE should be partitioned into GL, GY, and GLY. This partition was not possible in our analysis, however, because of the large number of missing cells. Therefore, the GE interactions were implicitly confounded with the experimental error. This treatment was justified by a previous study, which revealed that the yearly GL interaction was largely random and was not repeated across years (Yan and Rajcan, 2002). Consequently, the observed GE across years was likely a random GLY three-way interaction. Random GE cannot be exploited by dividing the target area into megaenvironments; it must be avoided through testing in multiple environments representative of the target environment (Yan and Hunt, 1998; Yan and Rajcan, 2002). Large GLY three-way interaction calls for multiple-year and multiple-location trials for reliable cultivar evaluation.
Predictive Power of a Single-year Trial
The predictive power of a single-year trial is measured by the correlation coefficient between cultivar performance (i.e., tBLUP) in one year and that in the next year across genotypes. Except for the correlation between 1991 and 1992, the correlation coefficients ranged from 0.41 to 0.71 (Table 4), all being highly significant (P < 0.01).
View this table:
[in this window]
[in a new window]
|
Table 4. Correlation coefficients (r) between the tBLUP values obtained from 1-yr, multiple-location trials and those from the next year across genotypes in common.
|
|
The predictive power of a single-year trial may be better appreciated when cultivars are divided into superior (tBLUP > 1.67, scored "1"), inferior (tBLUP < -1.67, scored "-1"), and intermediate (1.67 < tBLUP > -1.67, scored "0") groups. Table 5 displays only genotypes that were tested for three or more years. The scores of a genotype were largely consistent across years. There were changes between 1 and 0 or between 0 and –1, but a switch between 1 (superior) and -1 (inferior) from one year to the next did not occur except for genotypes OAC Libra and RCAT 9002 between 1991 and 1992 (Table 5), and this was because of the exceptionally low correlation between the two years (Table 4). In general, if a cultivar performed well (scored ‘1’) in one year, it was likely to perform well in the following year; its performance may be intermediate, but it will not be poor. Likewise, if a cultivar performs poorly (scored ‘-1’) in one year, it is likely to perform poorly again in the next year; it may perform better but not among the best. Therefore, although the correlation between consecutive years of performance was only modest (Table 4), 1-yr data appeared to have sufficient power for selection of superior cultivars and for culling of inferior cultivars (Table 5).
View this table:
[in this window]
[in a new window]
|
Table 5. Genotype scores obtained for individual years from 1991 to 2000. Only genotypes that were tested in three or more years are shown.
|
|
Are Multiple-Year Data Better Predictors of Cultivar Performance?
If multiple-year data predict cultivar performance better than single-year data, tBLUP values estimated from multiple-year data should have a larger correlation coefficient to that estimated in the next year. The analysis did not generate strong support for this assumption, however. The prediction of cultivar performance for 2000 by a single-year test (1999) was r = 0.57; the correlation increased to 0.63 when it was based on two years' data (1998 and 1999). The prediction for performance in 2000 did not improve by including data from additional years (Table 6). For four of the seven years (Table 6) 2-yr data gave the best prediction. For predicting cultivar performance in 1995, data from 1994 alone were better predictors than when data from earlier years were included. In general, 2 yr of data seemed to predict better than 1-yr data but inclusion of data from additional years did not substantially improve the prediction.
View this table:
[in this window]
[in a new window]
|
Table 6. Correlation coefficients between tBLUP based on one to nine previous years and that based on the present year.
|
|
Merit of Using Multiple-Year Data
The cultivars that scored "1" performed significantly above the grand mean, and those that scored "-1" performed significantly below the grand mean. The progress of all other genotypes, with a score of "0," however, was not conclusive. These genotypes may have truly yielded similar to the grand mean, or they may be inadequately evaluated (too few data points), which causes the BLUP to shrink toward the grand mean. One merit of using data from more than one year is that more cultivars can be decisively evaluated. For example, when the multiple-location trial of 2000 alone was used, 30 out of 112 cultivars were judged as superior or inferior. When two-year data (1999 and 2000) were used, 35 cultivars were judged as superior or inferior. Forty-four cultivars were decisively evaluated when 4 yr of data (from 1997–2000) were used (Table 7). In general, the larger the number of years the dataset contains, the larger the number of cultivars that can be conclusively evaluated (Table 7). The increase in the number of genotypes that can be decisively evaluated, however, becomes trivial beyond 3 yr.
|
DISCUSSION
|
|---|
One conclusion from this study is that data from single-year, multiple-location soybean trials had sufficient power to identify superior and inferior genotypes, and 2 yr of data gave slightly better prediction of next year's performance. This conclusion is consistent with three previous studies that investigated the predictive power of single vs. multiple year trials for next-year cultivar performance. Recognizing that growers are interested in finding a selection strategy that will maximize their production the next year, Cross and Helm (1986) compared strategies of corn [Zea mays L.] hybrid selection based on yield data from one, two, or three previous years' tests at single or multiple locations. The criterion was yield advantage of selected hybrids relative to the mean yield of all entries in the next year. Prediction based on data from one or two previous years was better than that based on data from the previous three years. Using performance trial data of spring wheat [Triticum aestivum L] and oats [Avena sativa L.], Gellner (1989) compared eight predictors of cultivar performance for their ability to identify the five top-yielding cultivars in the subsequent year. Three of these predictors were mean yields of respective cultivars obtained from one, two, or three previous years of multiple-location trials. He concluded that prediction of next-year performance on the basis of data from the previous year alone was equally good as that based on data from three previous years.
Using a balanced subset of data from North Carolina performance trials, Bowman (1998) compared probabilities of predicting the top varieties in a subsequent year based on single and 2-yr data for barley [Hordeum vulgare L.], corn, cotton [Gossypium hirsutum L.], oats, soybean, and wheat. He concluded that it was appropriate to use one-year multiple-location trial data to select midseason corn hybrids and 2-yr, multiple-location trial data in selecting the other crops. The finding that 1-yr performance trial data had sufficient power in identifying superior and inferiors genotypes justifies the common practice by breeders and cultivar sponsors of withdrawing their cultivars from the test based on a single-year trial.
Bowman (1998) was not able to compare predictions of cultivar performance by data from more than two years since very few cultivars were tested in more than three consecutive years. For the same reason, Cross and Helm (1986) and Gellner (1989) only examined predictions using data from up to 3 yr. We were able to investigate predictions based on data from up to 9 yr because of the use of mixed models. Mixed models allow the use of unbalanced data and give BLUP of cultivars. This allowed us to examine the merits of using multiple-year data and conclude that although a single-year, multiple-location test has enough power to identify the best cultivars, use of multiple-year data allow more genotypes to be evaluated conclusively.
With fewer data points, BLUP has the property of shrinking toward the grand mean, which makes it particularly suitable for cultivar evaluation based on unbalanced data (Piepho, 1994; DeLacy et al., 1996). Panter and Allen (1995) also concluded that BLUP was superior to least square means in predicting soybean cultivar performance. Instead of using BLUP, we used the t-statistic of BLUP, i.e., tBLUP, as a measure of cultivar performance. The two measures are highly correlated; tBLUP has an advantage over BLUP per se, however, in that it is also an intuitive measure of reliability of cultivar performance relative to the grand mean, as t-statistics are directly related to the probability of being different from the grand mean. Thus, tBLUP provides a convenient means for classifying tested genotypes into superior, inferior, and intermediate groups (Yan et al., 2002).
Alternatively, three different yield-stability statistics have been used to recommend cultivars for production in three maturity zones in Minnesota (Pazdernik et al., 1997). The stability characteristics studied consisted of two nonparametric statistics based on cultivar means and Kang's yield stability statistic, YSi (Kang, 1993). It was concluded that the stability statistics, in addition to cultivar means, could be used by consultants, cultivar testing personnel, and breeders to recommend the most appropriate cultivars for maxiumum yield and protein concentrations (Pazdernik et al., 1997). Our study, however, was aimed at using single vs. multiple-year, multiple-location yield trial data to select cultivars in the following year, reflecting a practice that is widely used by breeders and extension specialists.
The results of our study are in full agreement with a previous study of historical data from the Ontario Winter Wheat performance trials (Yan, unpublished). Winter wheat performance trials in Ontario have been conducted at eight to 10 locations across the province, a much wider region than the 2800 CHU area reported here for soybean. Nevertheless, similar conclusions were reached with regard to the predictive power of single- versus multiple-year trials. In conclusion, when multiple-year performance trial data are available, we recommend the use of tBLUP based on data from 2 yr of performance as a basis for cultivar evaluation and recommendation. Estimates of tBLUP based on single-year performance are also sufficient to select superior cultivars and to discard inferior genotypes.
Received for publication February 25, 2002.
|
REFERENCES
|
|---|
- Annicchiarico, P., and M. Perenzin. 1994. Adaptation patterns and definition of macro-environments for selection and recommendation of common wheat genotypes in Italy. Plant Breed. 113:197–205.
- Bowman, D.T. 1998. Using crop performance data to select hybrids and varieties. J. Prod. Agric. 11:256–259.
- Comstock, R.E., and R.H. Moll. 1963. Genotype-environment interactions. p. 164–196. In W.D. Hanson and H.F. Robinson (ed.) Statistical genetics and plant breeding. National Academy of Sciences and National Research Council, Washington DC.
- Cross, H.Z., and J.L. Helm. 1986. Hybrid maize selection strategies based on state yield trials. J. Agron. Educ. 15:110–113.
- DeLacy, I.H., K.E. Basford, M. Cooper, J.K. Bull, and C.G. McLaren. 1996. Analysis of multi-environment trials—a historical perspective. p. 39–124. In M. Cooper and G.L. Hammer (ed.) Plant adaptation and crop improvement. CAB Int., Wallingford, Oxon, UK.
- Gauch, H.G., and R.W. Zobel. 1996. AMMI analysis of yield trials. p. 85–122. In M.S. Kang and H.G. Gauch, Jr. (ed.) Genotype-by-environment interaction. CRC Press, Boca Raton, FL.
- Gellner, J.L. 1989. Predicting superior yielding spring wheat an oat cultivars using past yield data. Agron. J. 81:194–197.[Abstract/Free Full Text]
- Kang., M.S. 1993. Simultaneous selection for high yield and stability in crop performance trials: Consequences for growers. Agron. J. 85: 754–757.[Abstract/Free Full Text]
- Littell, R.C., G.A. Milliken, W.W. Stroup, and R.D. Wolfinger. 1996. SAS system for mixed models. SAS Institute Inc., Cary, NC.
- Ontario Ministry of Agriculture. Food and Rural Affairs (OMAFRA). 1993. Crop Heat Units for Corn and Other Warm-Season Crops in Ontario. [Online], available at: http://www.gov.on.ca/OMAFRA/english/crops/facts/93-119.htm. [modified 3 July 2001; verified 13 September 2002]. Government of Ontario, ON, Canada.
- Panter, D.M., and F.L. Allen. 1995. Using best linear unbiased predictions to enhance breeding for yield in soybean. II Selection of superior crosses from a limited number of yield trials. Crop Sci. 35:405–410.[Abstract/Free Full Text]
- Pazdernik, D.L., L.L. Hardman, and J.H. Orf. 1997. Agronomic performance and stability of soybean vareieties grown in three maturity zones of Minnesota. J. Prod. Agric. 10:425–430.
- Piepho, H.P. 1994. Best linear unbiased prediction (BLUP) for regional yield trials: A comparison to additive main effects and multiplicative interaction (AMMI) analysis. Theor. Appl. Genet. 89:647–654.
- SAS Institute. 1996. SAS/STAT user's guide. SAS Institute Inc., Cary, NC.
- Yan, W., and L.A. Hunt. 1998. Genotype-by-environment interaction and crop yield. Plant Breed. Rev. 16:135–178.
- Yan, W., L.A. Hunt, P. Johnson, G. Stewart, and X. Lu. 2002. On-farm strip trials versus replicated performance trials for cultivar evaluation. Crop Sci. 42:385–392.[Abstract/Free Full Text]
- Yan, W., and I. Rajcan. 2002. Biplot analysis of test sites and trait relations of soybean in Ontario. Crop Sci. 42:11–20.[Abstract/Free Full Text]
This article has been cited by other articles:
|
|
|
|
 
H.-P. Piepho and J. Mohring
Selection in Cultivar Trials--Is It Ignorable?
Crop Sci.,
December 2, 2005;
46(1):
192 - 201.
[Abstract]
[Full Text]
[PDF]
|
|
|
TOP
ABSTRACT
INTRODUCTION
