Randomized Block Mixed Linear Model
Fits a mixed linear model by REML. The linear model must contain only one random factor apart from the unit errors.
randomizedBlock(formula, random, weights=NULL, only.varcomp=FALSE, data=list(), subset=NULL, contrasts=NULL, tol=1e-6, maxit=50, trace=FALSE) randomizedBlockFit(y, X, Z, w=NULL, only.varcomp=FALSE, tol=1e-6, maxit=50, trace=FALSE)
- formula specifying the fixed model.
- vector or factor specifying the blocks corresponding to random effects.
- optional vector of prior weights.
- logical value, if
TRUEcomputation of standard errors and fixed effect coefficients will be skipped
- an optional data frame containing the variables in the model.
- an optional vector specifying a subset of observations to be used in the fitting process.
- an optional list. See the
- small positive numeric tolerance, passed to
- maximum number of iterations permitted, passed to
- logical value, passed to
TRUEthen working estimates will be printed at each iteration.
- numeric response vector
- numeric design matrix for fixed model
- numeric design matrix for random effects
- optional vector of prior weights
This function fits the model $y=Xb+Zu+e$ where $b$ is a vector of fixed coefficients and $u$ is a vector of random effects.
Write $n$ for the length of $y$ and $q$ for the length of $u$.
The random effect vector $u$ is assumed to be normal, mean zero, with covariance matrix $\sigma^2_uI_q$ while $e$ is normal, mean zero, with covariance matrix $\sigma^2I_n$.
If $Z$ is an indicator matrix, then this model corresponds to a randomized block experiment.
The model is fitted using an eigenvalue decomposition which transforms the problem into a Gamma generalized linear model.
Note that the block variance component
varcomp is not constrained to be non-negative.
It may take negative values corresponding to negative intra-block correlations.
However the correlation
varcomp/sum(varcomp) must lie between
Missing values in the data are not allowed.
This function is equivalent to
lme(fixed=formula,random=~1|random), except that the block variance component is not constrained to be non-negative, but is faster and more accurate for small to moderate size data sets.
It is slower than
lme when the number of observations is large.
This function tends to be fast and reliable, compared to competitor functions which fit randomized block models, when then number of observations is small, say no more than 200.
However it becomes quadratically slow as the number of observations increases because of the need to do two eigenvalue decompositions of order nearly equal to the number of observations.
So it is a good choice when fitting large numbers of small data sets, but not a good choice for fitting large data sets.
- A list with the components:
varcomp vector of length two containing the residual and block components of variance. se.varcomp standard errors for the components of variance.
fixed.estimates=TRUEthen the components from
lm.fitare also returned.
Venables, W., and Ripley, B. (2002). Modern Applied Statistics with S-Plus, Springer.
# Compare with first data example from Venable and Ripley (2002), # Chapter 10, "Linear Models" library(MASS) data(petrol) out <- randomizedBlock(Y~SG+VP+V10+EP, random=No, data=petrol) cbind(varcomp=out$varcomp,se=out$se.varcomp)