[R] Do Users of Nonlinear Mixed Effects Models Know Whether Their Software Really Works?

From: dave fournier <otter_at_otter-rsch.com>
Date: Fri 14 Oct 2005 - 06:19:24 EST

           Do Users of Nonlinear Mixed Effects Models Know

                  Whether Their Software Really Works?


   Lesaffre et. al. (Appl. Statist. (2001) 50, Part3, pp 325-335)    analyzed
   some simple clinical trials data using a logistic random effects    model. Several packages and methods MIXOR, SAS NLMIXED were employed.    They reported obtaining very different parameter estimates and    P values for the log-likelihood with the different packages and    methods. We thought it would be interesting to revisit this example    using the AD Model Builder random effects module which we feel is    the most stable software available for this problem at this time.

              http://otter-rsch.com/admodel.htm

   You can get Table 2 from Lesaffre et al at

           http://otter-rsch.com/downloads/other/lesaffre.pdf

   The data and more information are available    from the publisher.

           http://www.blackwellpublishers.co.uk/rss/Volumes/Cv50p3.htm

   We considered three questions:

     1.) What are the  estimates using the Laplace approximation
         for integrating out the random effects.

     2.) What are the exact MLE's.

     3.) How well does hypothesis testing (likelihood-ratio) using
         the Laplace approximation compare with the exact MLE's.

   We first fit the data using ADMB-RE's Laplace approximation    option.

   Laplace approximation estimates:

      # Number of parameters = 4  log-likelihood = -629.817
                       value      std dev  P value
            b_1     -2.3321e+00 7.6973e-01  < 0.0024
            b_2     -6.8795e-01 6.6185e-01    0.298
            b_3     -4.6134e-01 4.0000e-02  < 0.001
          sigma      4.5738e+00 7.0970e-01


    The parameter of interest here the treatment effect b_2 which is the    parameter reported in Lesaffre et. al.

    To calculate the exact MLE we fit the model using 100 point adaptive     Gaussian integration. The ADMB-RE results were:

    Gaussian integration estimates:

      # Number of parameters = 4 log-likelihood = -627.481

          name         value      std dev    P value
           b_1     -1.4463e+00 4.2465e-01   < 0.001
           b_2     -5.2225e-01 5.5716e-01     0.348
           b_3     -4.5150e-01 3.6663e-02   < 0.001
         sigma      4.0137e+00 3.8083e-01

   Of the estimates reported in Lesaffre et al. in table 2 only the    50 point quadrature for the program MIXOR appear to be correct    for both the log-likelihood value and the parameter estimates    while the authors concluded that the SAS NLMIXED parameter estimates    they obtained were correct. So even though these authors were looking    for pathological behaviour and were presumably very careful, and their    paper was presumably peer-reviewed, they came to the wrong conclusion    using SAS NLMIXED.

   How do we know that our exact MLE's are correct? To confirm our    results we used our parameter estimates as initial values in the SAS    NLMIXED procedure using 100 point adaptive quadrature. The procedure    returned our values, that is it agreed that these are the maximum    likelihood estimates. However we verified that to get these estimates    from the SAS NLMIXED procedure one must begin with fairly good    starting values. In contrast the ADMB-RE procedure is very insensitive    to the starting values used. Our conclusion is that while SAS NLMIXED    might work for this very simple problem it probably begins to break    down when the problem is a bit more difficult.

   The ADMB-RE software is more stable because it calculates exact higher    oreder derivatives by automatic differentiation for use in its    optimization procedure and calculations while other packages do not.

   Gauss-Hermite integration for the random effects can    be used for this model because the Hessian for the random effects is    diagonal which permits one dimensional integration over the random    effects to great accuracy. However this procedure does not scale well    to problems where the Hessian is not diagonal. Suppose that it takes    a 20 point quadrature to obtain reliable parameter estimates with a    diagonal Hessian. Then with a block diagonal Hessian where the blocks    are of size 4x4 it would take 160,000 points.

   Results using R

   We fit the model using what appear to be the currently available    procedures in R. The two routines lmer (lme4 package) and glmmPQL    (MASS library) were tried.

   The call

>> lmer(y ~ treat + time + (1|subject),data=lesaffre,family=binomial)

   resulted in a warning message from lme4() but both routines produced    the same results.

     Generalized linear mixed model fit using PQL
     Formula: y ~ treat + time + (1 | subject)
        Data: lesaffre
      Family: binomial(logit link)
           AIC      BIC    logLik deviance
      1305.859 1333.628 -647.9295 1295.859
     Random effects:
          Groups        Name    Variance    Std.Dev.
         subject (Intercept)      6.8059      2.6088
     # of obs: 1908, groups: subject, 294
     Estimated scale (compare to 1)  0.9091945
     Fixed effects:
                  Estimate Std. Error  z value Pr(>|z|)
     (Intercept) -0.626214   0.264996  -2.3631  0.01812 *
     treat       -0.304660   0.360866  -0.8442  0.39853
     time        -0.346605   0.026666 -12.9979  < 2e-16 ***
     sigma        2.608
     ---
     Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
     Warning message:
     optim or nlminb returned message ERROR:
     ABNORMAL_TERMINATION_IN_LNSRCH
      in: LMEopt(x = mer, value = cv)


   The R routine correctly identifies the treatment effect as not    significant. However the parameter estimates are poor.

   Likelihood Ratio Testing

   Accurate calculation of the log-likelihood value is desirable so that    hypothesis testing can be carried out using likelihood ratio tests.

   However as noted above the use of Gaussian integration is not    practical for many nonlinear mixed models. We were interested in    seeing how well the use of the approximate log-likelihood values    produced by ADMB-RE's Laplace approximation option would perform.

   We consider the alternative model with an extra interaction term    (b_4) from Lesaffre et al.

   Here are the results for the laplace approximation:

     # Number of parameters = 5  log-likelihood = -627.809
         name       value      std dev       P vlaue
         b_1      -2.5233e+00 7.8829e-01    < 0.002
         b_2      -3.0702e-01 6.8996e-01      0.655
         b_3      -4.0009e-01 4.7059e-02    < 0.001
         b_4      -1.3726e-01 6.9586e-02      0.044
        sigma      4.5783e+00 7.2100e-01

   and the exact parameter estimates by 100 point Gaussian    integration.

     # Number of parameters = 5  log-likelihood = -625.398
         name       value      std dev       P value
          b_1     -1.6183e+00 4.3427e-01    < 0.001
          b_2     -1.6077e-01 5.8394e-01      0.783
          b_3     -3.9100e-01 4.4380e-02    < 0.001
          b_4     -1.3679e-01 6.8013e-02      0.044
         sigma     4.0131e+00 3.8044e-01

    The log-likelihood differences are 2.01 for the Laplace     approximation and 2.08 for Gaussian integration.     Since the 95% point for hypothesis testing is 1.92     use of either model results in acceptance of the interaction     term.

   Conclusions

   With the exception of AD Model Builder random effect module none of    the packages tested appear to function reliably for this problem.    SAS NLMIXED was beginning to exhibit symptoms of instability which    would probably render it unreliable on more difficult problems. We    can see no reason for using "quasi-likelihoods" to fit nonlinear    mixed models when ADMB-RE can fit the models by maximum likelihood    with all the advantages that ensue.

   Note

   We realize that there are many other packages out there. We would    welcome results for other packages. If we can find a serious    competitor to AD Model Builder then we could move on to comparing    the relative performance on more difficult models.

        Cheers,

        Dave Fournier

-- 
David A. Fournier
P.O. Box 2040,
Sidney, B.C. V8l 3s3
Canada
http://otter-rsch.com


-- 
Internal Virus Database is out-of-date.

______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Received on Fri Oct 14 06:20:26 2005

This archive was generated by hypermail 2.1.8 : Fri 03 Mar 2006 - 03:40:44 EST