specification for glmmPQL

Hello Dr. Bates and group,

I understand, the attached data file did not accompany my original
message. I have listed below the code used to create that file.

data.1 <- data.frame(subject  = factor(rep(c("one", "two", "three", "four",
                                             "five", "six", "seven",
"eight"),
                                           each = 4),
                                       levels = c("one", "two", "three",
                                                  "four", "five", "six",
                                                  "seven", "eight")),
                     day      = factor(rep(c("one", "two", "three", "four"),
                                           times = 8),
                                       levels = c("one", "two", "three",
                                                  "four")),
                     expt     = rep(c("control", "treatment"), each = 16),
                     response = c(58, 63, 57, 54, 63, 59, 61, 53, 52, 62,
                                  46, 55, 59, 63, 58, 59, 62, 59, 64, 53,
                                  63, 75, 62, 64, 53, 58, 62, 53, 64, 72,
                                  65, 74))

mtrx.1 <- matrix(apply(data.1[, -4], 2, function(x)
                 rep(x, 100 - data.1$response)), ncol = 3, byrow = F)
mtrx.2 <- matrix(apply(data.1[, -4], 2, function(x)
                 rep(x, data.1$response)), ncol = 3, byrow = F)

data.2 <- data.frame(subject  = factor(c(mtrx.1[,1], mtrx.2[,1]),
                                       levels = c("one", "two", "three",
                                                  "four", "five", "six",
                                                  "seven", "eight")),
                     day      = factor(c(mtrx.1[,2], mtrx.2[,2]),
                                       levels = c("one", "two", "three",
                                                  "four")),
                     expt     = factor(c(mtrx.1[,3], mtrx.2[,3]),
                                       levels = c("control", "treatment")),
                     response = factor(c(rep("yes", nrow(mtrx.1)),
                                         rep("no", nrow(mtrx.2))),
                                       levels = c("yes", "no")))

#-------------------------------------------------------------------------------#

(fm.4 <- lmer(response ~ expt + (1|subject), data.2, binomial))

(fm.4a <- lmer(response ~ expt + (1|subject), data.2, binomial, method = "AGQ"))

Douglas Bates wrote:

On 9/4/05, Andrew R. Criswell <r-stats at arcriswell.com> wrote:

Hello All,

I have a question regarding how glmmPQL should be specified. Which of
these two is correct?

summary(fm.3 <- glmmPQL(cbind(response, 100 - response) ~ expt,
                       data = data.1, random = ~ 1 | subject,
                       family = binomial))

summary(fm.4 <- glmmPQL(response ~ expt, data = data.2,
                       random = ~ 1 | subject, family = binomial))

One might think it makes no difference, but it does.

I have an experiment in which 8 individuals were subjected to two types
of treatment, 100 times per day for 4 consecutive days. The response
given is binary--yes or no--for each treatment.

I constructed two types of data sets. On Rfile-01.Rdata (attached here)
are data frames, data.1 and data.2. The information is identical but the
data are arranged differently between these two data frames. Data frame,
data.1, groups frequencies by subject, day and treatment. Data frame,
data.2, is ungrouped.

I don't think your attached .Rdata file made it through the various
filters on the lists or on my receiving email.  Could you send me a
copy in a separate email message?

Consistency of these data frames is substantiated by computing these
tables:

ftable(xtabs(response ~ expt + subject + day,
            data = data.1))
ftable(xtabs(as.numeric(response) - 1 ~ expt + subject + day,
            data = data.2))

If I ignore the repeated measurement aspect of the data, I get, using
glm, identical results (but for deviance and df).

summary(fm.1 <- glm(cbind(response, 100 - response) ~ expt,
                   data = data.1, family = binomial))

summary(fm.2 <- glm(response ~ expt, data = data.2,
                   family = binomial))

However, if I estimate these two equations as a mixed model using
glmPQL, I get completely different results between the two
specifications, fm.3 and fm.4. Which one is right? The example which
accompanies help(glmmPQL) suggests fm.4.

summary(fm.3 <- glmmPQL(cbind(response, 100 - response) ~ expt,
                       data = data.1, random = ~ 1 | subject,
                       family = binomial))

summary(fm.4 <- glmmPQL(response ~ expt, data = data.2,
                       random = ~ 1 | subject, family = binomial))

Thank you,
Andrew

On 9/4/05, Andrew R. Criswell <r-stats at arcriswell.com> wrote:
 

Hello Dr. Bates and group,

I understand, the attached data file did not accompany my original
message. I have listed below the code used to create that file.

data.1 <- data.frame(subject  = factor(rep(c("one", "two", "three", "four",
                                            "five", "six", "seven",
"eight"),
                                          each = 4),
                                      levels = c("one", "two", "three",
                                                 "four", "five", "six",
                                                 "seven", "eight")),
                    day      = factor(rep(c("one", "two", "three", "four"),
                                          times = 8),
                                      levels = c("one", "two", "three",
                                                 "four")),
                    expt     = rep(c("control", "treatment"), each = 16),
                    response = c(58, 63, 57, 54, 63, 59, 61, 53, 52, 62,
                                 46, 55, 59, 63, 58, 59, 62, 59, 64, 53,
                                 63, 75, 62, 64, 53, 58, 62, 53, 64, 72,
                                 65, 74))

mtrx.1 <- matrix(apply(data.1[, -4], 2, function(x)
                rep(x, 100 - data.1$response)), ncol = 3, byrow = F)
mtrx.2 <- matrix(apply(data.1[, -4], 2, function(x)
                rep(x, data.1$response)), ncol = 3, byrow = F)

data.2 <- data.frame(subject  = factor(c(mtrx.1[,1], mtrx.2[,1]),
                                      levels = c("one", "two", "three",
                                                 "four", "five", "six",
                                                 "seven", "eight")),
                    day      = factor(c(mtrx.1[,2], mtrx.2[,2]),
                                      levels = c("one", "two", "three",
                                                 "four")),
                    expt     = factor(c(mtrx.1[,3], mtrx.2[,3]),
                                      levels = c("control", "treatment")),
                    response = factor(c(rep("yes", nrow(mtrx.1)),
                                        rep("no", nrow(mtrx.2))),
                                      levels = c("yes", "no")))

#-------------------------------------------------------------------------------#
   

Thanks for sending the data.

In your first message you said that you got completely different
results from glmmPQL when fitting the two models.  When I fit these
models with glmmPQL I got quite similar parameter estimates.  The
reported log-likelihood or AIC or BIC values are quite different but
these values apply to a different model (the list weighted linear
mixed model used in the PQL algorithm) and should not be used for a
glmm model in any case.

The fm4 results from lmer in the lme4 package (actually lmer is now in
the Matrix package but that is only temporary) confirm those from
glmmPQL.  The model fm3 when fit by lmer provides different standard
errors but that is because the weights are not being appropriately
adjusted in lmer.  We will fix that.

In general I think it is safest to use the long form of the data as in
your data.2.

Here are the results from lmer applied to the long form.  The results
from the Adaptive Gauss-Hermite Quadrature (AGQ) method are preferred
to those from the PQL method because AGQ is a more accurate
approximation to the log-likelihood of the GLMM model.  In this case
the differences are minor.

The log-likelihood reported here is an approximation to the
log-likelihood of the GLMM model.

 

(fm.4 <- lmer(response ~ expt + (1|subject), data.2, binomial))
   

Generalized linear mixed model fit using PQL 
Formula: response ~ expt + (1 | subject) 
  Data: data.2 
Family: binomial(logit link)
     AIC     BIC    logLik deviance
4298.026 4322.31 -2145.013 4290.026
Random effects:
    Groups        Name    Variance    Std.Dev. 
   subject (Intercept)    0.015835     0.12584 
# of obs: 3200, groups: subject, 8

Estimated scale (compare to 1)  0.9990621 

Fixed effects:
             Estimate Std. Error z value  Pr(>|z|)
(Intercept)    0.30764    0.08075  3.8098 0.0001391
expttreatment  0.21319    0.11473  1.8582 0.0631454
 

(fm.4a <- lmer(response ~ expt + (1|subject), data.2, binomial, method = "AGQ"))
   

Generalized linear mixed model fit using AGQ 
Formula: response ~ expt + (1 | subject) 
  Data: data.2 
Family: binomial(logit link)
     AIC      BIC    logLik deviance
4298.023 4322.306 -2145.011 4290.023
Random effects:
    Groups        Name    Variance    Std.Dev. 
   subject (Intercept)    0.015855     0.12592 
# of obs: 3200, groups: subject, 8

Estimated scale (compare to 1)  1.007675 

Fixed effects:
             Estimate Std. Error z value  Pr(>|z|)
(Intercept)    0.30811    0.08075  3.8156 0.0001358
expttreatment  0.21352    0.11473  1.8611 0.0627322


 

Douglas Bates wrote:

   

On 9/4/05, Andrew R. Criswell <r-stats at arcriswell.com> wrote:

     

Hello All,

I have a question regarding how glmmPQL should be specified. Which of
these two is correct?

summary(fm.3 <- glmmPQL(cbind(response, 100 - response) ~ expt,
                      data = data.1, random = ~ 1 | subject,
                      family = binomial))

summary(fm.4 <- glmmPQL(response ~ expt, data = data.2,
                      random = ~ 1 | subject, family = binomial))

One might think it makes no difference, but it does.

I have an experiment in which 8 individuals were subjected to two types
of treatment, 100 times per day for 4 consecutive days. The response
given is binary--yes or no--for each treatment.

I constructed two types of data sets. On Rfile-01.Rdata (attached here)
are data frames, data.1 and data.2. The information is identical but the
data are arranged differently between these two data frames. Data frame,
data.1, groups frequencies by subject, day and treatment. Data frame,
data.2, is ungrouped.

       

I don't think your attached .Rdata file made it through the various
filters on the lists or on my receiving email.  Could you send me a
copy in a separate email message?


     

Consistency of these data frames is substantiated by computing these
tables:

ftable(xtabs(response ~ expt + subject + day,
           data = data.1))
ftable(xtabs(as.numeric(response) - 1 ~ expt + subject + day,
           data = data.2))

If I ignore the repeated measurement aspect of the data, I get, using
glm, identical results (but for deviance and df).

summary(fm.1 <- glm(cbind(response, 100 - response) ~ expt,
                  data = data.1, family = binomial))

summary(fm.2 <- glm(response ~ expt, data = data.2,
                  family = binomial))

However, if I estimate these two equations as a mixed model using
glmPQL, I get completely different results between the two
specifications, fm.3 and fm.4. Which one is right? The example which
accompanies help(glmmPQL) suggests fm.4.

summary(fm.3 <- glmmPQL(cbind(response, 100 - response) ~ expt,
                      data = data.1, random = ~ 1 | subject,
                      family = binomial))

summary(fm.4 <- glmmPQL(response ~ expt, data = data.2,
                      random = ~ 1 | subject, family = binomial))

Thank you,
Andrew