tebv

Trial-estimated breeding values: best linear unbiased predictors of entries in single and multi-environment breeding trials

Build Status	License

A streamlined interface to calculate the breeding values of entries in breeding trials.

Installation

devtools::install_github("jeffersonfparil/tebv")

Approach and principle

At the moment the approach will be to create a more or less homogenous interface for univariate analysis with plans to extend it to multivariate analysis.

We fit multiple plausible linear models via:

• canonical Henderson's equations (efficient if $n \ge p$) and,
• Newton-Raphson transformations (more efficient if $n \lt p$). We then select the best fitting model based on which one on average has higher log-likelihood, lower AIC and lower BIC.

We fit 3 basic experimental design models:

• completely randomised design (CRD),
• randomised complete/incomplete block design (RBD), and
• spatial design with row and column effects (SPAT).

These models are implemented for:

• single environment trials, and
• multiple environment trials.

We use open-source tools as much as possible. This means we prefer sommer over ASReml. I would like to believe that being open-source where many other people can review and correct your code is a much better way of having confidence on the tool than having a lot of previous studies published using the tool.

Naming convention

I like types, hence R being not really strongly typed I like to prefix my variable names with the type I wish that variable to remain to be throughout its lifetime:

• Generally no prefix for scalars, except for indexes (i.e., idx_) and booleans (i.e., bool_)
• vec_ for vectors (a single lowercase letter can be used especially in the context of linear algebra, e.g. $y$, $\beta$ amd $\epsilon$ in $y = X\beta + \epsilon$)
• mat_ for matrices (a single uppercase letter can be used especially in the context of linear algebra, e.g. $X$ in $\left( X'X \right)^{-1}$)
• arr_ for arrays with more than 2 dimensions
• list_ for lists
• df_ for data frames
• fn_ for functions
• mod_ for models

The following are the names of the input data frame and its fields:

• The main input data frame per function is always called df.
• In univariate models, the trait is always called y with the corresponding function parameter trait, e.g. trait="y".
• Environmental factor which may refer to different trials across time and space is always called environ with the corresponding function parameter env, e.g. env="environ".
• Complete or incomplete blocks factor is always called block with the corresponding function parameter block, e.g. block="block".
• Row factor is always called row with the corresponding function parameter row, e.g. row="row".
• Column factor is always called col with the corresponding function parameter col, e.g. col="col".

Model definitions:

• There should always be an explicit 1 to indicate the intecept in the model.
• The residual variance-covariance matrix should always be explicit even when inpedently and identically distributed, e.g. rcov(units).

Documentation

Write documentation or comments into your code a much as you can especially on top of your function definitions.

devtools::document()

Unit tests

Value modularity and write tests for each function definition.

See the main tests function and each Rscript for individual unit tests:

devtools::check()

Or per module as:

devtools::load_all()
library(sommer)
library(testthat)
source("tests/testthat/test-helpers.R")
source("tests/testthat/test-univariate_gx1.R")
source("tests/testthat/test-univariate_gxe.R")

Test new models in tests/testthat/test-helpers.R and adjust parsing accordingly via: R/helpers.R::fn_henderson_vs_newtonraphson_fit().