ds.glmerSLMA | R Documentation |
ds.glmerSLMA
fits a Generalized Linear Mixed-Effects Model
(GLME) on data from one or multiple sources with pooling via SLMA (study-level meta-analysis).
ds.glmerSLMA(
formula = NULL,
offset = NULL,
weights = NULL,
combine.with.metafor = TRUE,
dataName = NULL,
checks = FALSE,
datasources = NULL,
family = NULL,
control_type = NULL,
control_value = NULL,
nAGQ = 1L,
verbose = 0,
start_theta = NULL,
start_fixef = NULL,
notify.of.progress = FALSE,
assign = FALSE,
newobj = NULL
)
formula |
an object of class formula describing the model to be fitted. For more information see Details. |
offset |
a character string specifying the name of a variable to be used as an offset. |
weights |
a character string specifying the name of a variable containing prior regression weights for the fitting process. |
combine.with.metafor |
logical. If TRUE the estimates and standard errors for each regression coefficient are pooled across studies using random-effects meta-analysis under maximum likelihood (ML), restricted maximum likelihood (REML) or fixed-effects meta-analysis (FE). Default TRUE. |
dataName |
a character string specifying the name of a data frame that contains all of the variables in the GLME formula. For more information see Details. |
checks |
logical. If TRUE |
datasources |
a list of |
family |
a character string specifying the distribution of the observed
value of the outcome variable around the predictions generated by the linear predictor.
This can be set as |
control_type |
an optional character string vector specifying the nature of a parameter
(or parameters) to be modified in the |
control_value |
numeric representing the new value which you want to allocate the
control parameter corresponding to the |
nAGQ |
an integer value indicating the number of points per axis for evaluating the adaptive
Gauss-Hermite approximation to the log-likelihood. Defaults 1, corresponding to the Laplace approximation.
For more information see R |
verbose |
an integer value. If |
start_theta |
a numeric vector of length equal to the number of random effects. Specify to retain
more control over the optimisation. See |
start_fixef |
a numeric vector of length equal to the number of fixed effects (NB including the intercept).
Specify to retain more control over the optimisation. See |
notify.of.progress |
specifies if console output should be produced to indicate progress. Default FALSE. |
assign |
a logical, indicates whether the function will call a second server-side function (an assign) in order to save the regression outcomes (i.e. a glmerMod object) on each server. Default FALSE. |
newobj |
a character string specifying the name of the object to which the glmerMod object
representing the model fit on the serverside in each study is to be written. This argument is
used only when the argument |
ds.glmerSLMA
fits a generalized linear mixed-effects model (GLME)
- e.g. a logistic or Poisson regression model including both fixed and random effects -
on data from single or multiple sources.
This function is similar to glmer
function from lme4
package in native R.
When there are multiple data sources, the GLME is fitted to convergence
in each data source independently. The estimates and standard errors returned
to the client-side which enable cross-study pooling using Study-Level Meta-Analysis (SLMA).
The SLMA used by default metafor
package
but as the SLMA occurs on the client-side (a standard R environment), the user can choose
any approach to meta-analysis. Additional information about fitting GLMEs
using glmer
function can be obtained using R help for glmer
and the lme4
package.
In formula
most shortcut notation allowed by glmer()
function is
also allowed by ds.glmerSLMA
.
Many GLMEs can be fitted very simply using a formula like:
y~a+b+(1|c)
which simply means fit an GLME with y
as the outcome variable (e.g.
a binary case-control using a logistic regression model or a count or a survival
time using a Poisson regression model), a
and b
as fixed effects, and c
as a random effect or grouping factor.
It is also possible to fit models with random slopes by specifying a model such as
y~a+b+(1+b|c)
where the effect of b
can vary randomly between groups defined by c
.
Implicit nesting can be specified with formulas such as: y~a+b+(1|c/d)
or y~a+b+(1|c)+(1|c:d)
.
The dataName
argument avoids you having to specify the name of the
data frame in front of each covariate in the formula.
For example, if the data frame is called DataFrame
you avoid having to write:
DataFrame\$y ~ DataFrame\$a + DataFrame\$b + (1 | DataFrame\$c)
.
The checks
argument verifies that the variables in the model are all defined (exist)
on the server-site at every study
and that they have the correct characteristics required to fit the model.
It is suggested to make checks
argument TRUE if an unexplained
problem in the model fit is encountered because the running process takes several minutes.
In the family
argument can be specified two types of models to fit:
"binomial"
: logistic regression models
"poisson"
: poisson regression models
Note if you are fitting a gaussian model (a standard linear mixed
model) you should use ds.lmerSLMA
and not ds.glmerSLMA
.
For more information you can see R help for lmer
and glmer
.
In control_type
at present only one such parameter can be modified,
namely the tolerance of the convergence criterion to the gradient of the log-likelihood
at the maximum likelihood achieved. We have enabled this because our practical experience
suggests that in situations where the model looks to have converged with sensible parameter
values but formal convergence is not being declared if we allow the model to be more
tolerant to a non-zero gradient the same parameter values are obtained but formal
convergence is declared. The default value for the check.conv.grad
is 0.001
(note that
the default value of this argument in ds.lmerSLMA
is 0.002
).
In control_value
at present (see control_type
)
the only parameter this can be is the convergence tolerance check.conv.grad
. In
general, models will be identified as having converged more readily if the value set
for check.conv.grad
is increased from its default value (0.001
). Please note
that the risk of doing this is that the model is also more likely to be declared
as having converged at a local maximum that is not the global maximum likelihood.
This will not generally be a problem if the likelihood surface is well behaved but if
you have a problem with convergence you might usefully compare all the parameter
estimates and standard errors obtained using the default tolerance (0.001
) even though
that has not formally converged with those obtained after convergence using the higher
tolerance.
Server function called: glmerSLMADS2
Many of the elements of the output list returned by ds.glmerSLMA
are
equivalent to those returned by the glmer()
function in native R. However,
potentially disclosive elements
such as individual-level residuals and linear predictor values are blocked.
In this case, only non-disclosive elements are returned from each study separately.
The list of elements returned by ds.glmerSLMA
is mentioned below:
coefficients
: a matrix with 5 columns:
: the names of all of the regression parameters (coefficients) in the model
: the estimated values
: corresponding standard errors of the estimated values
: the ratio of estimate/standard error
: the p-value treating that as a standardised normal deviate
CorrMatrix
: the correlation matrix of parameter estimates.
VarCovMatrix
: the variance-covariance matrix of parameter estimates.
weights
: the vector (if any) holding regression weights.
offset
: the vector (if any) holding an offset.
cov.scaled
: equivalent to VarCovMatrix
.
Nmissing
: the number of missing observations in the given study.
Nvalid
: the number of valid (non-missing) observations in the given study.
Ntotal
: the total number of observations
in the given study (Nvalid
+ Nmissing
).
data
: equivalent to input parameter dataName
(above).
call
: summary of key elements of the call to fit the model.
Once the study-specific output has been returned, the function returns the number of elements relating to the pooling of estimates across studies via study-level meta-analysis. These are as follows:
input.beta.matrix.for.SLMA
: a matrix containing the vector of coefficient
estimates from each study.
input.se.matrix.for.SLMA
: a matrix containing the vector of standard error
estimates for coefficients from each study.
SLMA.pooled.estimates
: a matrix containing pooled estimates for each
regression coefficient across all studies with pooling under SLMA via
random-effects meta-analysis under maximum likelihood (ML), restricted maximum
likelihood (REML) or via fixed-effects meta-analysis (FE).
convergence.error.message
: reports for each study whether the model converged.
If it did not some information about the reason for this is reported.
DataSHIELD Development Team
## Not run:
## Version 6, for version 5 see Wiki
# Connecting to the Opal servers
require('DSI')
require('DSOpal')
require('dsBaseClient')
builder <- DSI::newDSLoginBuilder()
builder$append(server = "study1",
url = "http://192.168.56.100:8080/",
user = "administrator", password = "datashield_test&",
table = "CNSIM.CNSIM1", driver = "OpalDriver")
builder$append(server = "study2",
url = "http://192.168.56.100:8080/",
user = "administrator", password = "datashield_test&",
table = "CNSIM.CNSIM2", driver = "OpalDriver")
builder$append(server = "study3",
url = "http://192.168.56.100:8080/",
user = "administrator", password = "datashield_test&",
table = "CNSIM.CNSIM3", driver = "OpalDriver")
logindata <- builder$build()
# Log onto the remote Opal training servers
connections <- DSI::datashield.login(logins = logindata, assign = TRUE, symbol = "D")
# Select all rows without missing values
ds.completeCases(x1 = "D", newobj = "D.comp", datasources = connections)
# Fit a Poisson regression model
ds.glmerSLMA(formula = "LAB_TSC ~ LAB_HDL + (1 | GENDER)",
offset = NULL,
dataName = "D.comp",
datasources = connections,
family = "poisson")
# Clear the Datashield R sessions and logout
datashield.logout(connections)
builder <- DSI::newDSLoginBuilder()
builder$append(server = "study1",
url = "http://192.168.56.100:8080/",
user = "administrator", password = "datashield_test&",
table = "CLUSTER.CLUSTER_SLO1", driver = "OpalDriver")
builder$append(server = "study2",
url = "http://192.168.56.100:8080/",
user = "administrator", password = "datashield_test&",
table = "CLUSTER.CLUSTER_SLO2", driver = "OpalDriver")
builder$append(server = "study3",
url = "http://192.168.56.100:8080/",
user = "administrator", password = "datashield_test&",
table = "CLUSTER.CLUSTER_SLO3", driver = "OpalDriver")
logindata <- builder$build()
# Log onto the remote Opal training servers
connections <- DSI::datashield.login(logins = logindata, assign = TRUE, symbol = "D")
# Fit a Logistic regression model
ds.glmerSLMA(formula = "Male ~ incid_rate +diabetes + (1 | age)",
dataName = "D",
datasources = connections[2],#only the second server is used (study2)
family = "binomial")
# Clear the Datashield R sessions and logout
datashield.logout(connections)
## End(Not run)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.