Spatial analysis of a field experiment - Barley

Introduction

In this section we illustrate the ASReml syntax for performing spatial and incomplete block analysis of a field experiment. There has been a large amount of interest in developing techniques for the analysis of spatial data both in the context of field experiments and geostatistical data (see for example, Cullis and Gleeson, 1991; Cressie, 1991; Gilmour et al., 1997). This example illustrates the analysis of 'so-called' regular spatial data, in which the data is observed on a lattice or regular grid. This is typical of most small plot designed field experiments. Spatial data is often irregularly spaced, either by design or because of the observational nature of the study. The techniques we present in the following can be extended for the analysis of irregularly spaced spatial data, though, larger spatial data sets may be computationally challenging, depending on the degree of irregularity or models fitted.

The data we consider is taken from Gilmour et al. (1995) and involves a field experiment designed to compare the performance of 25 varieties of barley. The experiment was conducted at Slate Hall Farm, UK in 1976, and was designed as a balanced lattice square with replicates laid out as shown in the Table. The data fields were Rep, RowBlk, ColBlk, row, column and yield. Lattice row and column numbering is typically within replicates and so the terms specified in the linear model to account for the lattice row and lattice column effects would be Rep.latticerow Rep.latticecolumn. However, in this example lattice rows and columns are both numbered from 1 to 30 across replicates (see Table). The terms in the linear model are therefore simply RowBlk ColBlk. Additional fields row and column indicate the spatial layout of the plots.

The ASReml input file is presented below. Three models have been fitted to these data. The lattice analysis is included for comparison in PATH 3. In PATH 1 we use the separable first order autoregressive model to model the variance structure of the plot errors. Gilmour et al. (1997) suggest this is often a useful model to commence the spatial modelling process. The form of the variance matrix for the plot errors (R structure) is given by σ2Σ = σ2(Σccross Σr) where Σc and Σr are 15 by 15 and 10 by 10 matrix functions of the column (φc) and row (φr) autoregressive parameters respectively.

Gilmour et al. (1997) recommend revision of the current spatial model based on the use of diagnostics such as the sample variogram of the residuals (from the current model). This diagnostic and a summary of row and column residual trends are produced by default with graphical versions of ASReml when a spatial model has been fitted to the errors. It can be suppressed, by the use of the -n option on the command line. We have produced the following plots by use of the -g22 option.

Field layout of Slate Hall Farm experiment

Column - Replicate levels
Row123456789101112131415
1111112222233333
2111112222233333
3111112222233333
4111112222233333
5111112222233333
6444445555566666
7444445555566666
8444445555566666
9444445555566666
10444445555566666
Column - Rowblk levels
Row123456789101112131415
11111111111111112121212121
22222212121212122222222222
33333313131313132323232323
44444414141414142424242424
55555515151515152525252525
66666616161616162626262626
77777717171717172727272727
88888818181818182828282828
99999919191919192929292929
10101010101020202020203030303030
Column - Colblk levels
Row123456789101112131415
1123456789101112131415
2123456789101112131415
3123456789101112131415
4123456789101112131415
5123456789101112131415
6161718192021222324252627282930
7161718192021222324252627282930
8161718192021222324252627282930
9161718192021222324252627282930
10161718192021222324252627282930

 Slate Hall example
   Rep 6      # Six replicates of 5x5 plots in 2x3 arrangement
   RowBlk 30  # Rows within replicates numbered across replicates
   ColBlk 30  # Columns within replicates numbered across replicates
   row 10     # Field row
   column 15  # Field column
   variety 25
   yield
 barley.asd !skip 1 !DOPATH 1
 !PATH 1 # AR1 x AR1
 y ~ mu var
 1 2
 15 column AR1 0.1   # Second field is specified so ASReml can sort
 10 row AR1 0.1      # records properly into field order

 !PATH 2 # AR1 x AR1 + units
 y ~ mu var !r units
 1 2
 15 column AR1 0.1
 10 row AR1 0.1

 !PATH 3 # incomplete blocks
 y ~ mu var !r Rep Rowblk Colblk

 !PATH 0
 predict variety !TWOSTAGEWEIGHTS
Abbreviated ASReml output file is presented below. The iterative sequence has converged to column and row correlation parameters of (.68377,.45859) respectively. The plot size and orientation is not known and so it is not possible to ascertain whether these values are spatially sensible. It is generally found that the closer the plot centroids, the higher the spatial correlation. This is not always the case and if the highest between plot correlation relates to the larger spatial distance then this may suggest the presence of extraneous variation (see Gilmour et al., 1997), for example. Figure presents a plot of the sample variogram of the residuals from this model. The plot appears in reasonable agreement with the model.

The next model includes a measurement error or nugget effect component. That is the variance model for the plot errors is now given by σ2Σ = σ2(Σccross Σr + ψI150) where ψ is the ratio of nugget variance to error variance (σ2). The abbreviated output for this model is given below. There is a significant improvement in the REML Loglikelihood with the inclusion of the nugget effect (see Table).


Figure 1 Sample variogram of the residuals from the AR1 cross AR1 model for the Slate Hall data
 # AR1 x AR1
 #
    1 LogL=-739.681     S2=  36034.        125 df    1.000     0.1000     0.1000
    2 LogL=-714.340     S2=  28109.        125 df    1.000     0.4049     0.1870
    3 LogL=-703.338     S2=  29914.        125 df    1.000     0.5737     0.3122
    4 LogL=-700.371     S2=  37464.        125 df    1.000     0.6789     0.4320
    5 LogL=-700.324     S2=  38602.        125 df    1.000     0.6838     0.4542
    6 LogL=-700.322     S2=  38735.        125 df    1.000     0.6838     0.4579
    7 LogL=-700.322     S2=  38754.        125 df    1.000     0.6838     0.4585
    8 LogL=-700.322     S2=  38757.        125 df    1.000     0.6838     0.4586
  Final parameter values                        1.0000    0.68377    0.45861


  Source                Model  terms     Gamma     Component    Comp/SE   % C
  Variance                150    125   1.00000       38756.6       5.00   0 P
  Residual            AR=AutoR    15  0.683767      0.683767      10.80   0 U
  Residual            AR=AutoR    10  0.458607      0.458607       5.55   0 U

                                    Wald F statistics
      Source of Variation           NumDF     DenDF    F-inc             Prob
    8 mu                                1      12.8   850.88            <.001
    6 variety                          24      80.0    13.04            <.001

 # AR1 x AR1 + units
    1 LogL=-740.735     S2=  33225.        125 df    :   2 components constrained
    2 LogL=-723.595     S2=  11661.        125 df    :   1 components constrained
    3 LogL=-698.498     S2=  46239.        125 df
    4 LogL=-696.847     S2=  44725.        125 df
    5 LogL=-696.823     S2=  45563.        125 df
    6 LogL=-696.823     S2=  45753.        125 df
    7 LogL=-696.823     S2=  45796.        125 df

  Source                Model  terms     Gamma     Component    Comp/SE   % C
  units                   150    150  0.106154       4861.48       2.72   0 P
  Variance                150    125   1.00000       45796.3       2.74   0 P
  Residual            AR=AutoR    15  0.843795      0.843795      12.33   0 U
  Residual            AR=AutoR    10  0.682686      0.682686       6.68   0 U

                                    Wald F statistics
      Source of Variation           NumDF     DenDF    F-inc             Prob
    8 mu                                1       3.5   259.81            <.001
    6 variety                          24      75.7    10.21            <.001
The lattice analysis (with recovery of between block information) is presented below. This variance model is not competitive with the preceding spatial models. The models can be formally compared using the BIC values for example.
 # IB analysis
    1 LogL=-734.184     S2=  26778.        125 df
    2 LogL=-720.060     S2=  16591.        125 df
    3 LogL=-711.119     S2=  11173.        125 df
    4 LogL=-707.937     S2=  8562.4        125 df
    5 LogL=-707.786     S2=  8091.2        125 df
    6 LogL=-707.786     S2=  8061.8        125 df
    7 LogL=-707.786     S2=  8061.8        125 df

           - - - Results from analysis of yield - - -

           Approximate stratum variance decomposition
  Stratum     Degrees-Freedom   Variance      Component Coefficients
  Rep                    5.00    266657.        25.0     5.0     5.0     1.0
  RowBlk                24.00    74887.8         0.0     4.3     0.0     1.0
  ColBlk                23.66    71353.5         0.0     0.0     4.3     1.0
  Residual Variance     72.34    8061.81         0.0     0.0     0.0     1.0

  Source                Model  terms     Gamma     Component    Comp/SE   % C
  Rep                       6      6  0.528714       4262.39       0.62   0 P
  RowBlk                   30     30   1.93444       15595.1       3.06   0 P
  ColBlk                   30     30   1.83725       14811.6       3.04   0 P
  Variance                150    125   1.00000       8061.81       6.01   0 P

                                    Wald F statistics
      Source of Variation           NumDF     DenDF    F-inc             Prob
    8 mu                                1       5.0  1216.29            <.001
    6 variety                          24      79.3     8.84            <.001
Finally, we present portions of the .pvs files to illustrate the prediction facility of ASReml . The first five and last three variety means are presented for illustration. The overall SED printed is the square root of the average variance of difference between the variety means. The two spatial analyses have a range of SEDs which are available if the !SED qualifier is used. All variety comparisons have the same SED from the third analysis as the design is a balanced lattice square. The F-statistic statistics for the spatial models are greater than for the lattice analysis. We note the F-statistic for the AR1 cross AR1 + units model is smaller than the F-statistic for the AR1 cross AR1.
 Predicted values of yield
 #AR1 x AR1
  variety             PredictedVlue StandardEror Ecode
        1.0000              1257.9763        64.6146 E
        2.0000              1501.4483        64.9783 E
        3.0000              1404.9874        64.6260 E
        4.0000              1412.5674        64.9027 E
        5.0000              1514.4764        65.5889 E
          .                     .                .
       23.0000              1311.4888        64.0767 E
       24.0000              1586.7840        64.7043 E
       25.0000              1592.0204        63.5939 E
  SED: Overall Standard Error of Difference   59.05

 #AR1 x AR1 + units
  variety             PredictedVlue StandardEror Ecode
        1.0000              1245.5843        97.8591 E
        2.0000              1516.2331        97.8473 E
        3.0000              1403.9863        98.2398 E
        4.0000              1404.9202        97.9875 E
        5.0000              1471.6197        98.3607 E
          .                     .                .
       23.0000              1316.8726        98.0402 E
       24.0000              1557.5278        98.1272 E
       25.0000              1573.8920        97.9803 E
  SED: Overall Standard Error of Difference   60.51

 # IB
  Rep                  is ignored in the prediction
  RowBlk               is ignored in the prediction
  ColBlk               is ignored in the prediction

  variety             PredictedVlue StandardEror Ecode
        1.0000              1283.5870        60.1994 E
        2.0000              1549.0133        60.1994 E
        3.0000              1420.9307        60.1994 E
        4.0000              1451.8554        60.1994 E
        5.0000              1533.2749        60.1994 E
          .                     .                .
       23.0000              1329.1088        60.1994 E
       24.0000              1546.4699        60.1994 E
       25.0000              1630.6285        60.1994 E
  SED: Overall Standard Error of Difference   62.02
Notice the differences in SE and SED associated with the various models. Choosing a model on the basis of smallest SE or SED is not recommended because the model is not necessarily fitting the variability present in the data.

Summary of models for the Slate Hall data
REML number of
model log-likelihood parameters F-statistic SED
AR1 cross AR1 -700.32 3 13.04 59.0
AR1 cross AR1 + units -696.82 4 10.22 60.5
IB -707.79 4 8.84 62.0

The predict statement included the qualifier !TWOSTAGEWEIGHTS. This generates an extra table in the .pvs file which we now display for each model.
  Predicted values with Effective Replication assuming
  Variance= 38754.26
  Heron:    1  1257.98      22.1504
  Heron:    2  1501.45      20.6831
  Heron:    3  1404.99      22.5286
  Heron:    4  1412.57      22.7623
  Heron:    5  1514.48      21.1830
    .       .      .          .
  Heron:   25  1592.02      26.0990

  Predicted values with Effective Replication assuming
  Variance= 45796.58
  Heron:    1  1245.58      23.8842
  Heron:    2  1516.24      22.4423
  Heron:    3  1403.99      24.1931
  Heron:    4  1404.92      24.0811
  Heron:    5  1471.61      23.2995
    .       .      .          .
  Heron:   25  1573.89      26.0505

  Predicted values with Effective Replication assuming
  Variance= 8061.808
  Heron:    1  1283.59      4.03145
  Heron:    2  1549.01      4.03145
  Heron:    3  1420.93      4.03145
  Heron:    4  1451.86      4.03145
  Heron:    5  1533.27      4.03145
    .       .      .         .
  Heron:   25  1630.63      4.03145
The value of 4 for the IB analysis is clearly reasonable given there are 6 actual replicates but this analysis has used up 48 degrees of freedom for the rowblk and colblk effects. The values from the spatial analyses are similar but much higher reflecting the gain in accuracy from the spatial analysis.
  • Back

    Return to start