Linear Models with Correlated Errors

Now that we have a sense of how to model spatial variation, the next step is to incorporate that into a linear model. The starting point is the linear mixed model. In RCBD design, often the treatments are treated as fixed and the block effect as random.

$$Y_ij = \mu + \alpha_i + \beta_j + \epsilon_{ij}$$

$Y_ij$ is the independent variable
$\mu$ is the overall mean
$\alpha_i$ is the effect due to the $i^{th}$ treatment
$\beta_j$ is the effect due to the $j^{th}$ block
$\epsilon_{ij}$ are the error terms distributed as $N ~\sim (0,\sigma)$

Here is an expanded version of the last term:

$$ \epsilon_{ij} ~\sim N \Bigg( 0, \left[ { \begin{array}{ccc} \sigma & \cdots & 0 \\
\vdots & \ddots & \vdots \\
0 & \cdots & \sigma \end{array} } \right] \Bigg) $$

This is a mathematically representation of iid, independent and identically distributed, an assumption of linear models. When there is spatial autocorrelation, observations closer to one another are correlated, so the off-diagonals in the variance-covariance matrix are not zero.

Spatial models seek to mathematically model this covariance so it is properly accounted for during hypothesis testing and prediction.

Code for this section

The following scripts build upon work done in previous section(s).

R

library(emmeans); library()
# (nlme and gstat should already be loaded)
library(spaMM) # for running `corMatern()`

# standard linear model
nin_lme <- lme(yield ~ gen, random = ~1|rep,
              data = Nin,
              na.action = na.exclude)
              
# extract the esimated marginal means for variety
preds_lme <- as.data.frame(emmeans(nin_lme, "gen"))

# use information from the variogram fitting for intialising the parameters
nugget <- Nin_variofit_gau$psill[1] 
range <- Nin_variofit_gau$range[2]  
sill <- sum(Nin_variofit_gau$psill) 
nugget.effect <-  nugget/sill

# initalise the covariance structure (from the nlme package)
cor.gaus <- corSpatial(value = c(range, nugget.effect), 
                  form = ~ row.length + col.width, 
                  nugget = T, fixed = F,
                  type = "gaussian", 
                  metric = "euclidean")

# update the rcbd model
nin_gaus <- update(nin_lme, corr = cor.gaus)
# extract predictions for 'gen'
preds_gaus <- as.data.frame(emmeans(nin_gaus, "gen")

# a similar procedure can be follow for other models
# but we are going to take a shortcut and not specify the parameters

# exponential
cor.exp <- corSpatial(form = ~ row.length + col.width, 
                      nugget = T, fixed = F)

nin_exp <- update(nin_lme, corr = cor.exp)
preds_exp <- as.data.frame(emmeans(nin_exp, "gen"))

# Matern structure
cor.mat <- corMatern(form = ~ row.length + col.width, 
                     nugget = T, fixed = F)
nin_matern <- update(nin_lme, corr = cor.mat)
preds_mat <- as.data.frame(emmeans(nin_matern, "gen")

SAS

proc mixed data=alliance ;
	class entry rep;
	model yield = entry ;
	random rep;
	lsmeans entry/cl;
	ods output LSMeans=NIN_RCBD_means;
	title1 'NIN data: RCBD';
run;

proc mixed data=alliance maxiter=150;
	class entry;
	model yield = entry /ddfm=kr;
	repeated/subject=intercept type=sp(gau) (Row Col) local;
	parms (11) (22) (19);
	lsmeans entry/cl;
	ods output LSMeans=NIN_Spatial_means;
	title1 'NIN data: Gaussian Spatial Adjustment';
run;

Previous
Next