Cross-validation procedure

Cross-validation for estimation of AMMI models

THe original dataset is split into two datasets: training set and validation set. The 'training' set has all combinations (genotype x environment) with N-1 replications. The 'validation' set has the remaining replication. The splitting of the dataset into modeling and validation sets depends on the design informed. For Completely Randomized Block Design (default), and alpha-lattice design (declaring block arguments), complete replicates are selected within environments. The remained replicate serves as validation data. If design = 'RCD' is informed, completely randomly samples are made for each genotype-by-environment combination (Olivoto et al. 2019). The estimated values considering naxis-Interaction Principal Component Axis are compared with the 'validation' data. The Root Mean Square Prediction Difference (RMSPD) is computed. At the end of boots, a list is returned.

IMPORTANT: If the data set is unbalanced (i.e., any genotype missing in any environment) the function will return an error. An error is also observed if any combination of genotype-environment has a different number of replications than observed in the trial.

Usage

cv_ammi(
  .data,
  env,
  gen,
  rep,
  resp,
  block = NULL,
  naxis = 2,
  nboot = 200,
  design = "RCBD",
  verbose = TRUE
)

Arguments

.data: The dataset containing the columns related to Environments, Genotypes, replication/block and response variable(s).
env: The name of the column that contains the levels of the environments.
gen: The name of the column that contains the levels of the genotypes.
rep: The name of the column that contains the levels of the replications/blocks. AT LEAST THREE REPLICATES ARE REQUIRED TO PERFORM THE CROSS-VALIDATION.
resp: The response variable.
block: Defaults to NULL. In this case, a randomized complete block design is considered. If block is informed, then a resolvable alpha-lattice design (Patterson and Williams, 1976) is employed. All effects, except the error, are assumed to be fixed.
naxis: The number of axis to be considered for estimation of GE effects.
nboot: The number of resamples to be used in the cross-validation. Defaults to 200.
design: The experimental design. Defaults to RCBD (Randomized complete Block Design). For Completely Randomized Designs inform design = 'CRD'.
verbose: A logical argument to define if a progress bar is shown. Default is TRUE.

Value

An object of class cv_ammi with the following items: * RMSPD: A vector with nboot-estimates of the Root Mean Squared Prediction Difference between predicted and validating data.

RMSPDmean: The mean of RMSPDmean estimates.
Estimated: A data frame that contain the values (predicted, observed, validation) of the last loop.
Modeling: The dataset used as modeling data in the last loop
Testing: The dataset used as testing data in the last loop.

References

Olivoto, T., A.D.C. L\'ucio, J.A.G. da silva, V.S. Marchioro, V.Q. de Souza, and E. Jost. 2019. Mean performance and stability in multi-environment trials I: Combining features of AMMI and BLUP techniques. Agron. J. 111:2949-2960. doi:10.2134/agronj2019.03.0220

Patterson, H.D., and E.R. Williams. 1976. A new class of resolvable incomplete block designs. Biometrika 63:83-92.

Author

Tiago Olivoto tiagoolivoto@gmail.com

Examples


# \donttest{
library(metan)
model <- cv_ammi(data_ge,
                env = ENV,
                gen = GEN,
                rep = REP,
                resp = GY,
                nboot = 5,
                naxis = 2)
#> Validating 1 of 5 sets |========                                 | 20% 00:00:02 
Validating 2 of 5 sets |================                         | 40% 00:00:03 
Validating 3 of 5 sets |=========================                | 60% 00:00:04 
Validating 4 of 5 sets |=================================        | 80% 00:00:06 
Validating 5 of 5 sets |=========================================| 100% 00:00:07 

# }