1. Home
  2. GLRTEST procedure

GLRTEST procedure

Calculates likelihood tests to assess the random terms in a generalized linear mixed model (R.W. Payne).

Options

PRINT = string tokens
Controls printed output (tests); default test
SELECTIONstring tokens Specifies information to print with the tests (aicsicbic, critical); default crit
CRITICAL = variate Saves the index number that follows the prefix in the filename
GLSAVE = pointer Save structure of the original analysis from GLMM; default * uses the save structure from the most recent GLMM analysis

Parameters

TERMS = formula
Random terms to be tested; default is to test them all
TESTSTATISTICscalar or pointer to scalars Test statistics for each term
DF = = scalar or pointer to scalars Degrees of freedom of the test statistics
AIC = = scalar or pointer to scalars Akaike information coefficients for each term
sIC = = scalar or pointer to scalars Schwarz (Bayesian) information coefficients for each term

Description

GLRTEST can be used after a GLMM analysis to assess the effect of dropping random terms from the model. It uses the REML deviances to do this. In the GLMM algorithm, REML is used to analyse the adjusted dependent variate z, with the variate of iterative weights, defined by the generalized linear model. These depend on the current fitted values, and change at each iteration until convergence. (See the Method section of the GLMM procedure for more details.) The REML deviance is taken from the analysis of the final adjusted z-variate with the final iterative weights.

GLRTEST saves the deviance from the original analysis using VKEEP, and the final adjusted z-variate and variate of iterative weights using GLKEEP. It then does REML analyses with these variates, omitting each random term, saving their deviances, and calculating their differences from the original deviance. Akaike and Schwarz (Bayesian) information coefficients are obtained using the VAIC procedure.

Note that, for compatibility, it is important to use the same adjusted z-variate and the same iterative weights as in the original analysis. With the alternative, of doing GLMM analyses removing each random term, we would be taking deviances from REML analyses with their own adjusted z-variates and weights, which could be very different from those in the original analysis. So we would be comparing REML analyses with different models, different response variates and different weights, which would not provide a valid comparison. Of course this does mean that the results pertain to the REML analysis rather than to the GLMM analysis itself. So they should be used as guidance rather than as a definitive test. Often, however, the random terms will have been defined by the design of the investigation. The tests will then be used more as an indication of the effectiveness of the design than to decide whether to omit terms from the analysis.

By default, GLRTEST produces tests for every random term. However, you can use the TERMS parameter to request tests for a specific set of terms.

The default is to print the tests, but you can set option PRINT=* to suppress this. The additional information to be printed with the tests is controlled by the SELECTION option, with settings:

aic Akaike information coefficients;
sic Schwarz (Bayesian) information coefficients;
bic synonym for sic, and
critical critical values (default).
 

If the variance components are unconstrained, the critical values are from a chi-square distribution with one degree of freedom. Alternatively, if they are constrained to be positive, the asymptotic distribution of test is a 50:50 mixture of chi-square distributions with zero and one degree of freedom. Essentially this means that the critical values are from a chi-square distribution with one degree of freedom but at double the probability level. See, for example, Lee, Nelder & Pawitan 2006, Section 6.5. The CRITICAL option can save three critical values, in a variate with units for probabilities of 0.05, 0.001 and 0.001.

The TESTSTATISTIC parameter can save the statistics. the DF parameter can save their numbers of degrees of freedom. (These will always be equal to one, but the parameter is included for compatibility with the HGFTEST and HGRTEST procedures.) The AIC and SIC parameters can save the Akaike and Schwarz (Bayesian) information coefficients, respectively. If you are making a test for a single term, you can supply a scalar for each of these parameters. However, if you have several terms, you must supply a pointer which will then be set up to contain as many scalars as there are terms.

Options: PRINTSELECTIONCRITICALGLSAVE.
Parameters: TERMSTESTSTATISTICDFAICSIC.

See also

Procedures: GLDISPLAY, GLKEEP, GLMM, GLPLOT, GLPERMTEST, GLPREDICT, GLTOBITPOISSON.
Commands for: Regression analysis.

Example

CAPTION   'GLRTEST example',\
          !t('Data from an experiment on Great Knott, Rothamsted;',\
          'see West, J.S., Fitt, B.D.L., Leech, P.K., Biddulph, J.E.,',\
          'Huang, Y.-J. &, Balesdent, M.-H. (2002).',\
          'Effects of timing of ~italic{Leptosphaeria maculans}',\
          'ascospore release and fungicide regime on phoma leaf spot',\
          'and phoma stem canker development on winter oilseed rape',\
          '(~italic{Brassica napus}) in southern England.',\
          'Plant Pathology, 51, 454–463.'); STYLE=meta,plain
SPLOAD    '%data%/GtKnott2000.gsh'
GLMM      [PRINT=model,components,wald,deviance; DISTRIBUTION=binomial;\
          LINK=logit; DISPERSION=*; FIXED=Cultivar*Fungicide;\
          RANDOM=Block/Wholeplot] LMplants; NBINOMIAL=Nplants
GLRTEST   [SELECTION=aic,sic,critical]
Updated on February 7, 2023

Was this article helpful?