agridat (version 1.16)

perry.springwheat: Multi-environment trial of wheat cultivars introduced 1860-1982.

Description

Yields of wheat cultivars introduced 1860-1982. Grown in 20 environments.

Usage

data("perry.springwheat")

Format

A data frame with 560 observations on the following 6 variables.

yield

yield, kg/ha

gen

genotype/cultivar factor, 28 levels

env

environment factor, 20 levels

site

site factor

year

year, 1979-1982

yor

year of release, 1860-1982

Details

Twenty-eight of the most significant wheat cultivars of the past century in Western Australia, were grown in 20 field trials over 4 years in the Central and Eastern wheat-belt of Australia.

At the Wongan Hills site there were separate early and late sown trials in 1979 and 1980. Later sowing dates generally have lower yields.

Note: Although not indicated by the original paper, it may be that the Merredin site in 1979 also had early/late sowing dates.

Examples

# NOT RUN {
data(perry.springwheat)
dat <- perry.springwheat

library(lattice)
xyplot(yield~yor|env, dat, type=c('p','r'), xlab="year of release",
       main="perry.springwheat")

# Show a table of sites*year
# library(latticeExtra)
# useOuterStrips(xyplot(yield~yor|site*factor(year), dat,
#                       type=c('p','r')))


# Perry reports a rate of gain of 5.8 kg/ha/year.  No model is given.
# We fit a model with separate intercept/slope for each env
m1 <- lm(yield ~ env + yor + env:yor, data=dat)
# Average slope across environments
mean(c(coef(m1)[21], coef(m1)[21]+coef(m1)[22:40]))
## [1] 5.496781

# ----------------------------------------------------------------------------

# }
# NOT RUN {
  # Now a mixed-effects model.  Fixed overall int/slope.  Random env int/slope.

  # First, re-scale response so we don't have huge variances
  dat$y <- dat$yield / 100

  require(lme4)
  # Use || for uncorrelated int/slope.  Bad model.  See below.
  # m2 <- lmer(y ~ 1 + yor + (1+yor||env), data=dat)
  ## Warning messages:
  ## 1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  ##   Model failed to converge with max|grad| = 0.55842 (tol = 0.002, component 1)
  ## 2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  ##   Model is nearly unidentifiable: very large eigenvalue
  ##  - Rescale variables?;Model is nearly unidentifiable: large eigenvalue ratio
  ##  - Rescale variables?

  # Looks like lme4 is having trouble with variance of intercepts
  # There is nothing special about 1800 years, so change the
  # intercept -- 'correct' yor by subtracting 1800 and try again.
  dat$yorc <- dat$yor - 1800
  m3 <- lmer(y ~ 1 + yorc + (1+yorc||env), data=dat)
  
  # Now lme4 succeeds.  Rate of gain is 100*0.0549 = 5.49
  fixef(m3)
  ## (Intercept)        yorc
  ##  5.87492444  0.05494464

  # asreml3
  require(asreml)
  m3a <- asreml(y ~ 1 + yorc, data=dat, random = ~ env + env:yorc)

  require(lucid)
  vc(m3)
  ##      grp        var1 var2     vcov   sdcor
  ##      env (Intercept) <NA> 11.61    3.407
  ##    env.1        yorc <NA>  0.00063 0.02511
  ## Residual        <NA> <NA>  3.551   1.884

  vc(m3a)
  ##           effect component std.error z.ratio      con
  ##      env!env.var  11.61     4.385        2.6 Positive
  ## env:yorc!env.var   0.00063  0.000236     2.7 Positive
  ##       R!variance   3.551    0.2231      16   Positive

# }
# NOT RUN {
# ----------------------------------------------------------------------------

# }