Previous PageTable Of ContentsNext Page

Is There a Case for Discarding Low-Yielding Soybean Performance Trials?

Daryl T. Bowman

Crop Science Dep., Box 8604, North Carolina State University, Raleigh, NC 27695-8604, daryl_bowman@ncsu.edu

Abstract

Low-yielding crop performance trials may add little to overall performance data because they may contribute to crossover interactions that would reduce predictability. This study was designed to examine the ramifications of low-yielding tests. Data from 1983 - 1992 North Carolina Official Maturity Group V Soybean Cultivar Trials were examined. Two-year cultivar yield data were compared with the following single-year data to determine predictability with the entire data set and partial data sets that excluded tests with mean yields 1 standard deviation (SD) and 1.5 SD below the overall average. Average yields were 2755 kg/ha with yields 1 SD below average at 2016 kg/ha and 1.5 SD below at 1680 kg/ha. Correlations, as indicators of predictability, did not improve with the deletion of low-yielding tests; it is thought this may be due to few crossover interactions. This study also shows that it was more critical to remove less precise tests than low-yielding tests when attempting to improve predictability.

Media Summary

Crop performance trials are occasionally subject to below average growing conditions but elimination of these sites did not improve predictability in a set of soybean trials.

Keywords

Predictability, crossovers, stress

Introduction

Crop performance trials are conducted throughout the world for many agronomic and horticultural crops. Performance trials aim to provide unbiased data to assist breeders, growers, and extension personnel in making decisions regarding recommendations or release of inbred and hybrid cultivars. Performance trials need to predict performance accurately, because they tell a grower which cultivars should perform well under similar environmental conditions. The ability of trials to predict performance is contingent upon accurate sampling in target environments and minimization of cultivar by environment interactions. Predictability is defined as the ability to predict performance as evidenced by the magnitude of the correlation between past and future performance. Cultivar by environment interactions of any type diminish the predictability of the data. Atypical data sets tend to contribute to genotype by environment interactions and lower predictability.

Atypical data sets would, for example, include low-yielding tests that result from drought or other types of stress and tests with low precision; it is assumed in this discussion that locations (test sites) have been determined to represent the target area and are typical of production practices in the area. Low-yielding tests tend to have high coefficients of variation (CV) that have been used as a criteria to remove them from publication.

The inclusion of data from low-yielding trials often minimizes the differences among cultivars and makes it difficult to choose the best. It has also been argued that low-yielding tests provide little useful information because significant differences often cannot be detected.

If cultivar rank performance is similar for high and low-yielding sites, then removal of low-yielding sites may have no impact on predictability. However, if relative cultivar performance varies with productivity level, i.e. there is a crossover interaction, then inclusion of low-yielding sites may reduce predictability. This could be due to a crossover interaction or to the fact that it may not be possible to detect differences among cultivars near the crossover point. The purpose of this study was to discuss the ramifications of discarding low-yielding trials on ability to choose among cultivars.

Materials and Methods

To determine the impact of removing low-yielding tests, overall mean yield was calculated from a ten-year data set (1983-1992 North Carolina Official Soyean Cultivar Trials for MGV). Environments which had means lower than one standard deviation (SD) below the overall 10 year mean and 1.5 SD below the overall mean were identified and eliminated in separate data sets. These partial data sets were used to calculate alternative yield and rank correlations to provide an estimate of predictability relative to the entire data set. The correlations were calculated from data pooled over locations and years where appropriate. Two-year averages by cultivar were correlated with the following one-year average to determine predictability.

Results and Discussion

The mean yield over the entire data set was 2755 kg/ha with 1 SD of 739 kg/ha and 1.5 SD of 1075 kg/ha. Thus data from environments with mean yields of 2016 kg or less per hectare were deleted in Table 1 for the 1.0 SD criterion while environments of 1680 kg or less per hectare were deleted for the 1.5 SD criterion. The 1.0 SD criterion with a minimum limit of 2016 kg/ha is the midpoint between the economic yield threshold of 1882 kg/ha and the mean cross-over point of 2150 kg/ha. Twenty-five of the 110 environments or 23% fit the 1.0 criterion (Table 1) while only 10 environments or 9% fit the 1.5 SD criterion; these are higher numbers than that reported by Allen et al. (1978) which may be attributed to different rainfall patterns. The number of environments eliminated ranged from 1 to 6 (6 to 38% of the data) in the 1.0 SD criterion whereas the number of environments eliminated at the 1.5 SD criterion ranged from 0 to 3 (0 to 25% of the data) (Table 1).

The total number of environments in the data sets ranged from 10 to 17 (Table 1). Six of 11 environments in the 1985-87 data set yielded 1 SD less than the overall mean; most of these were in 1985 which was a dry year. In fact, correlations could not be run using the 1.0 SD criterion for that particular data set since there were no data from 1985 to correlate (Table 2).

In most cases, eliminating data from the low-yield environments did not improve predictability in terms of yield and rank correlations (Table 2). Mean yield correlations were 0.46, 0.39, and 0.41 for the entire data set, 1.0 SD criterion, and 1.5 SD criterion, respectively. Mean rank correlations were the same for both the 1.0 SD and 1.5 SD criterion (0.40) which were numerically lower than the entire data set (0.46).

Using the criteria for discarding trials with questionable precision as defined by Bowman and Rawlings (1995), one test in 1984 had high error variance and should not have been reported. Deletion of this test data did not change correlations for the 1983, 1984, 1985 data set, 0.72 vs. 0.74 and 0.67 vs. 0.64 for yield and rank correlations, respectively. However it did substantially improve correlations for the 1984, 1985, 1986 data set, 0.27 vs. 0.46 and 0.09 vs. 0.34 for yield and rank correlations, respectively. These data showed that it was more critical to eliminate less precise tests than low-yielding tests when attempting to improve predictability.

Conclusions

Examination of the data sets included in this study did not substantiate the deletion of low-yield trials on the basis of predictability alone. Even though crossover interactions did occur with some of the cultivars at low yield levels (approximately equal to 1.0 SD below the overall average yield), elimination of data from low-yielding trials did not improve correlations. By sampling more environments, one may actually be improving predictability. The argument could be made that by sampling more environments one reduces the cultivar variance, thus true means are estimated more accurately when low-yielding tests are included even though low yields do not contribute much to the overall means.

The data in this study support the strategy of Bradley et al. (1988) of including all environments in their data sets when making decisions regarding cultivar release; accuracy at individual locations is sacrificed by reducing replicates in order to sample more environments, thereby improving predictability. Moreover, if all low-yielding tests were eliminated there might be no data in some years. For example, as previously mentioned all tests in 1985 had yields less than 1.0 SD below the overall average. Two-year averages over locations would provide the best predictions of future performance (Bowman, 1998) since the single year 1985 would confuse growers.

Table 1. Environment and entry numbers of MG V soybean data sets from the North Carolina Official Variety Trials.

 

Number of common entries

Total number of environments

Number of environments below average

Years

   

1 SD

1.5 SD

1983-85

9

12

5

3

1984-86

14

10

4

2

1985-87

11

11

6

2

1986-88

13

12

2

0

1987-89

14

16

3

0

1988-90

17

17

2

1

1989-91

17

17

2

1

1990-92

16

15

1

1

Table 2. North Carolina MG V soybean yield and rank correlations of two-year subsets with following one-year data using the entire set versus deleting environments with yields 1 SD and 1.5 SD below overall average.

 

Yield Correlations

Rank Correlations

Years

Entire data set

1 SD

1.5 SD

Entire data set

1 SD

1.5 SD

1983-85

0.72*

.

0.64

0.67*

.

0.52

1984-86

0.27

0.12

0.21

0.09

0.05

0.02

1985-87

0.03

0.23

-0.09

0.23

0.25

-0.04

1986-88

0.61*

0.53*

0.61*

0.61*

0.58*

0.61*

1987-89

0.51

0.50

0.51

0.65*

0.52*

0.65*

1988-90

0.74**

0.69**

0.71**

0.70**

0.74**

0.72**

1989-91

0.48*

0.42

0.44

0.41

0.39

0.42

1990-92

0.30

0.27

0.27

0.28

0.26

0.26

Mean

0.46

0.39

0.41

0.46

0.40

0.40

*,**Significant at the 0.05 and 0.01 levels of probability, respectively.

References

Allen F L, Comstock R E, and Rasmusson D C (1978). Optimal environments for yield testing. Crop Sci. 18,747-751.

Bowman D T (1998). Using crop performance data to select hybrids and varieties. J. Prod. Agric. 11,256-259.

Bowman D T, and Rawlings J O (1995). Establishing a rejection procedure for crop performance data. Agron. J. 87,147-151.

Bradley J P, Knittle K H, and Troyer AF (1988). Statistical methods in seed corn product selection. J. Prod. Agric. 1,34-38.

Previous PageTop Of PageNext Page