PROC GLIMMIX non-convergence problem solutions

Tips and Strategies for Mixed Modeling with SAS/STAT® Procedures Kathleen Kiernan, Jill Tao, and Phil Gibbs, SAS Institute Inc., Cary, NC, USA

ABSTRACT Inherently, mixed modeling with SAS/STAT® procedures, such as GLIMMIX, MIXED, and NLMIXED is computationally intensive. Therefore, considerable memory and CPU time can be required. The default algorithms in these procedures might fail to converge for some data sets and models. This paper provides recommendations for circumventing memory problems and reducing execution times for your mixed modeling analyses. This paper also shows how the new HPMIXED procedure can be beneficial for certain situations, as with large sparse mixed models. Lastly, the discussion focuses on the best way to interpret and address common notes, warnings, and error messages that can occur with the estimation of mixed models in SAS software.

http://support.sas.com/resources/papers/proceedings12/332-2012.pdf

glimmix data=xxx METHOD=RSPL  ITDETAILS;
class  xxx;
model xxx=  xxx
/dist=binomial link=logit s ddfm=kr;
random int / subject = xxx;
NLOPTIONS MAXITER=100;
run;

Tables available from PROC GLIMMIX

Output Added:
-------------
Name: ModelInfo
Label: Model Information
Template: Stat.Glimmix.ModelInfo
Path: Glimmix.ModelInfo
-------------

Output Added:
-------------
Name: ClassLevels
Label: Class Level Information
Template: Stat.Glimmix.ClassLevels
Path: Glimmix.ClassLevels
-------------

Output Added:
-------------
Name: NObs
Label: Number of Observations
Template: Stat.Glimmix.NObs
Path: Glimmix.NObs
-------------

Output Added:
-------------
Name: Dimensions
Label: Dimensions
Template: Stat.Glimmix.Dimensions
Path: Glimmix.Dimensions
-------------

Output Added:
-------------
Name: OptInfo
Label: Optimization Information
Template: Stat.Glimmix.OptInfo
Path: Glimmix.OptInfo
-------------

Output Added:
-------------
Name: IterHistory
Label: Iteration History
Template: Stat.Glimmix.IterHistory
Path: Glimmix.IterHistory
-------------
NOTE: Convergence criterion (GCONV=1E-8) satisfied.
NOTE: At least one element of the gradient is greater than 1e-3.

Output Added:
-------------
Name: ConvergenceStatus
Label: Convergence Status
Template: Stat.Glimmix.ConvergenceStatus
Path: Glimmix.ConvergenceStatus
-------------

Output Added:
-------------
Name: FitStatistics
Label: Fit Statistics
Template: Stat.Glimmix.FitStatistics
Path: Glimmix.FitStatistics
-------------

Output Added:
-------------
Name: CovParms
Label: Covariance Parameter Estimates
Template: Stat.Glimmix.CovParms
Path: Glimmix.CovParms
-------------

Output Added:
-------------
Name: ParameterEstimates
Label: Solutions for Fixed Effects
Template: Stat.Glimmix.ParameterEstimates
Path: Glimmix.ParameterEstimates
-------------

Output Added:
-------------
Name: Tests3
Label: Type III Tests of Fixed Effects
Template: Stat.Glimmix.Tests3
Path: Glimmix.Tests3
-------------

Bonferroni correction for power analysis with more than two groups

I'm trying to find a textbook reference for the following procedure written explicitly in the context of power analysis.  Please let me know if you know (k u e k a w a AT gmail com).

When there are more than two conditions in the experiment design, the alpha level, one of the parameters that go into power analysis, can be divided by the number of contrasts.  If there are three groups (control, treatment 1, treatment 2), there are three contrast points:

C vs T1, C vs T2, and T1 vs T2.

The typical alpha level is 0.5, so you can do:

0.5 / 3 = 0.16

and use that in the power analysis software.

If only two contrasts are important for your purpose:

0.5 / 2 = 0.25

Reference:

page 24 of

https://medschool.vanderbilt.edu/cqs/files/cqs/media/2010Ayumi.pdf

Bonferroni correction:

https://en.wikipedia.org/wiki/Bonferroni_correction

SAS PROC GLIMMIX method=

The default technique is METHOD=RSPL, corresponding to maximizing the residual log pseudo-likelihood with an expansion about the current solutions of the best linear unbiased predictors of the random effects. In models for normal data with identity link, METHOD=RSPL and METHOD=RMPL are equivalent to restricted maximum likelihood estimation, and METHOD=MSPL and METHOD=MMPL are equivalent to maximum likelihood estimation.

***

The following SAS Usage Note:

  http://support.sas.com/kb/37107 

  http://support.sas.com/kb/40724

 

provide information on testing covariance parameters when using PROC MIXED and PROC GLIMMIX.

 

SAS Proc POWER examples

Comparison of two independent groups:

proc power;
twosamplemeans test=diff
groupmeans = 0 | .2
stddev = 1
npergroup = .
power = .8;
run;

Comparison of dependent data (paired)

proc power;
pairedmeans test=diff
meandiff = .2
corr = 0.5
stddev = 1
npairs = .
power = .8;
run;

Comparison of proportions

proc power;
twosamplefreq test=pchi
groupproportions = (.65 .70)
nullproportiondiff = 0
power = .80
npergroup =.;
run;

Logit coefficients from logistic regression model

How do we interpret logic coefficients estimated by logistic regression model?  The following is a hypothetical result:

log(p/1-p) = 0.3 + 0.2*Male + 0.4*TREATMENT

One use of this result is to see if Male effect and GPA effect are statistically significant.  We also want to know the meaning of values, such as 0.2 and 0.4.  Because the left side of equation is a complex mathematical construct, it is not immediately clear what 0.2 or 0.4 means.

<Under construction>

Effect size by Cohen

Effect size of:

.2 Small effect

.5 Medium effect

.8 Large effect

Reference:

Cohen, J. Statistical power for the behavioral sciences (2nd ed.). Hillsdale, NJ: Erlbaum (1988).

Page 5 of  http://www.wmich.edu/evalphd/wp-content/uploads/2010/05/Effect_Size_Substantive_Interpretation_Guidelines.pdf

Quotation:

"Cohen’s benchmarks Cohen (1988) attempted to address the issue of interpreting effect size estimates relative to other effect sizes. He suggested some general definitions for small, medium, and large effect sizes in the social sciences. However, Cohen chose these quantities to reflect the typical effect sizes encountered in the behavioral sciences as a whole -- he warned against using his labels to interpret relationship magnitudes within particular social science disciplines or topic areas. His general labels, however, illustrate how to go about interpreting relative effects. Cohen labeled an effect size small if d = .20 or r = .10. He wrote, "Many effects sought in personality, social, and clinical-psychological research are likely to be small . . . because of the attenuation in validity of the measures employed and the subtlety of the issue frequently involved" (p. 13). Large effects, according to Cohen, are frequently "at issue in such fields as sociology, economics, and experimental and physiological psychology, fields characterized by the study of potent variables or the presence of good experimental control or both" (p. 13). Cohen suggested large magnitudes of effect were d = .80 or r = .50. Medium-sized effects were placed between these two extremes, that is d = .50 or r = .30. A caution against using Cohen’s benchmarks as generic descriptors of the magnitude of effect size is implied above. Because some areas, like education, are likely to have smaller effect sizes than others, using Cohen’s labels may be misleading."