ds.glmSLMA {dsBaseClient}R Documentation

ds.glmSLMA calling glmSLMADS1, glmSLMADS2


Fits a generalized linear model (glm) on data from a single or multiple sources with pooled co-analysis across studies being based on study level meta-analysis


ds.glmSLMA(formula = NULL, family = NULL, offset = NULL,
  weights = NULL, combine.with.metafor = TRUE, dataName = NULL,
  checks = FALSE, maxit = 30, datasources = NULL)



Denotes an object of class formula which is a character string describing the model to be fitted. Most shortcut notation for formulas allowed under R's standard glm() function is also allowed by ds.glmSLMA. Many glms can be fitted very simply using a formula such as: "y~a+b+c+d" which simply means fit a glm with y as the outcome variable with a, b, c and d as covariates. By default all such models also include an intercept (regression constant) term. If all you need to fit are straightforward models such as these, you do not need to read the remainder of this information about "formula". But if you need to fit a more complex model in a customised way, the following text gives a few additional pointers: As an example, the formula: "EVENT~1+TID+SEXF*AGE.60" denotes fit a glm with the variable "EVENT" as its outcome with covariates TID (in this case a 6 level factor [categorical] variable denoting "time period" with values between 1 and 6), SEXF (also a factor variable denoting sex and AGE.60 (a quantitative variable representing age-60 in years). The term "1" forces the model to include an intercept term which it would also have done by default (see above) but using "1" may usefully be contrasted with using "0" (as explained below), which removes the intercept term. The "*" between SEXF and AGE.60 means fit all possible main effects and interactions for and between those two covariates. As SEXF is a factor this is equivalent to writing SEXF+AGE.60+SEXF1:AGE.60 (the last element being the simple interaction term representing the product of SEXF level 1 [in this case female] and AGE.60). This takes the value 0 in all males (0 * AGE.60), and the same value as AGE.60 (1 * AGE.60) in females. If the formula had instead been written as: "EVENT~0+TID+SEXF*AGE.60" the 0 would mean do NOT fit an intercept term and, because TID happens to be a six level factor this would mean that the first six model parameters which were originally intercept+TID2+TID3+TID4+TID5+TID6 using the first formula will now become TID1+TID2+TID3+TID4+TID5+TID6. This is mathematically the same model, but conveniently, it means that the effect of each time period may now be estimated directly. For example, the effect of time period 3 is now obtained directly as the coefficient for TID3 rather than the sum of the coefficients for the intercept and TID3 which was the case using the original formula.


This argument identifies the error distribution function to use in the model. At present ds.glm has been written to fit family="gaussian" (i.e. a conventional linear model with normally distributed errors), family="binomial" (i.e. a conventional unconditional logistic regression model), and family = "poisson" (i.e. a Poisson regression model - of which perhaps the most commonly used application is for survival analysis using Piecewise Exponential Regression (PER) which typically closely approximates Cox regression in its main estimates and standard errors. At present the gaussian family is automatically coupled with an 'identity' link function, the binomial family with a 'logistic' link function and the poisson family with a 'log' link function. For the majority of applications typically encountered in epidemiology and medical statistics, one these three classes of models will typically be what you need. However, if a particular user wishes us to implement an alternative family (e.g. 'gamma') or an alternative family/link combination (e.g. binomial with probit) we can discuss how best to meet that request: it will almost certainly be possible, but we may seek a small amount of funding or practical in-kind support from the user in order to ensure that it can be carried outin a timely manner


A character string specifying the name of a variable to be used as an offset. An offset is a component of a glm which may be viewed as a covariate with a known coefficient of 1.00 and so the coefficient does not need to be estimated by the model. As an example, an offset is needed to fit a piecewise exponential regression model. Unlike the standard glm() function in native R, ds.glmSLMA() only allows an offset to be set using the <offset> argument, it CANNOT be included directly in the formula via notation such as "y~a+b+c+d+offset(offset.vector.name)". So in ds.glmSLMA this model must be specified as: formula="y~a+b+c+d", ..., offset="offset.vector.name" and ds.glmSLMA then incorporates it appropriately into the formula itself.


A character string specifying the name of a variable containing prior regression weights for the fitting process. Like offset, ds.glmSLMA does not allow a weights vector to be written directly into the glm formula.


This argument is Boolean. If TRUE (the default) 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).


A character string specifying the name of an (optional) dataframe that contains all of the variables in the glm formula. This avoids you having to specify the name of the dataframe in front of each covariate in the formula e.g. if the dataframe is called "DataFrame" you avoid having to write: "DataFrame$y~DataFrame$a+DataFrame$b+DataFrame$c+DataFrame$d" Processing stops if a non existing data frame is indicated.


This argument is a boolean. If TRUE ds.glmSLMA then undertakes a series of checks of the structural integrity of the model that can take several minutes. Specifically it 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 a GLM. The default value is FALSE and so it is suggested that the argument <checks> is only made TRUE if an unexplained problem in the model fit is encountered.


A numeric scalar denoting the maximum number of iterations that are permitted before ds.glm declares that the model has failed to converge. Logistic regression and Poisson regression models can require many iterations, particularly if the starting value of the regression constant is far away from its actual value that the glm is trying to estimate. In consequence we choose to set maxit=30 - but depending on the nature of the models you wish to fit, you may wish to be alerted more quickly than this if there is a delay in convergence, or you may wish to allow MORE iterations.


specifies the particular opal object(s) to use, if it is not specified the default set of opals will be used. The default opals are always called default.opals. This parameter is set without inverted commas: e.g. datasources=opals.em or datasources=default.opals If you wish to specify the second opal server in a set of three, the parameter is specified: e.g. datasources=opals.em[2]. If you wish to specify the first and third opal servers in a set specify: e.g. datasources=opals.em[c(2,3)]


ds.glmSLMA specifies the structure of a generalized linear model (glm) to be fitted separately on each study/data source. The model is first constructed and disclosure checked by glmSLMADS1. This aggregate function then returns its output to ds.glmSLMA which processes the information and uses it in a call to the second aggregate function glmSLMADS2. This call specifies and fits the required glm in each data source. Unlike glmDS2 (called by the more commonly used generalized linear modelling client-side function ds.glm) the requested model is then fitted to completion on the data in each study rather than iteration by iteration on all studies combined. At the end of this SLMA fitting process glmSLMADS2 returns study-specific parameter estimates and standard errors to the client. These can then be pooled using random effects (or fixed effects) meta-analysis - eg using the metafor package. This mode of model fitting may reasonably be called study level meta-analysis (SLMA) although the analysis is based on estimates and standard errors derived from direct analysis of the individual level data in each study rather than from published study summaries (as is often the case with SLMA of clinical trials etc). Furthermore, unlike common approaches to study-level meta-analysis adopted by large multi-study research consortia (eg in the combined analysis of identical genomic markers across multiple studies), the parallel analyses (in every study) under ds.glmSLMA are controlled entirely from one client. This avoids the time-consuming need to ask each study to run its own analyses and the consequent necessity to request additional work from individual studies if the modelling is to be extended to include analyses not subsumed in the original analytic plan. Additional analyses of this nature may, for example, include analyses based on interactions between covariates identified as having significant main effects in the original analysis. From a mathematical perspective, the SLMA approach (using ds.glmSLMA) differs fundamentally from the usual approach using ds.glm in that the latter is mathematically equivalent to placing all individual-level data from all sources in one central warehouse and analysing those data as one combined dataset using the conventional glm() function in R. However, although this may sound to be preferable under all circumstances, the SLMA approach actually offers key inferential advantages when there is marked heterogeneity between sources that cannot simply be corrected with fixed effects each reflecting a study or centre-effect. In particular, fixed effects cannot simply be used in this way when there there is heterogeneity in the effect that is of scientific interest.


Many of the elements of the output list returned by ds.glmSLMA from each study separately are precisely equivalent to those returned by the glm() function in native R. However, potentially disclosive elements such as individual-level residuals and linear predictor values are blocked. The return results from each separate study appear first in the return list with the full set of results from each study presented in a block and the blocks listed in the order in which the studies appear in <datasources>. As regards the elements within each study the most important elements are included last in the return list because they then appear at the bottom of a simple print out of the return object. In reverse order, these key elements are listed below. In addition to the elements reflecting the primary results of the analysis, ds.glmSLMA also returns a range of error messages if the model fails indicating why failure may have occurred and in particular detailing any disclosure traps that may have been

coefficients:- a matrix in which the first column contains the names of all of the regression parameters (coefficients) in the model, the second column contains the estimated values of the coefficients (called estimates), the third the corresponding standard errors, the fourth the ratio corresponding to the value of each estimate divided by its standard error and the fifth the p-value treating that ratio as a standardised normal deviate (a simple Wald test).

family:- indicates the error distribution and link function used in the glm

formula:- see description of formula as an input parameter (above)

df.resid:- the residual degrees of freedom around the model

deviance.resid:- the residual deviance around the model

df.null:- the degrees of freedom around the null model (with just an intercept)

dev.null:- the deviance around the null model (with just an intercept)

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 (enters glm with a coefficient of 1.00)

cov.scaled:- equivalent to VarCovMatrix

cov.unscaled:- equivalent to VarCovMatrix but assuming dispersion (scale) parameter is 1

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)

dispersion:- - the estimated dispersion parameter: deviance.resid/df.resid for a gaussian family multiple regression model, 1.00 for logistic and poisson regression

call:- - summary of key elements of the call to fit the model

na.action:- - chosen method of dealing with NAs. Usually, na.action=na.omit indicating any individual (or more strictly any "observational unit") that has any data missing that are needed for the model is exluded from the fit, even if all the rest of the required data are present. These required data include: the outcome variable, covariates, or any values in a regression weight vector or offset vector. As a side effect of this, when you include additional covariates in model you may exclude extra individuals from the analysis and this can seriously distort inferential tests based on assuming models are nested (eg likelihood ratio tests).

iter:- the number of iterations required to achieve convergence file for the glm() function in native R.

input.beta.matrix.for.SLMA:- a matrix containing the vector of coefficient estimates from each study. In combination with the corresponding standard errors (see input.se.matrix.for.SLMA) these can be imported directly into a study level meta-analysis (SLMA) package such as metafor to generate estimates pooled via SLMA

input.se.matrix.for.SLMA:- a matrix containing the vector of standard error estimates for coefficients from each study. In combination with the coefficients (see input.beta.matrix.for.SLMA) these can be imported directly into a study level meta-analysis (SLMA) package such as metafor to generate estimates pooled via SLMA

SLMA.pooled.estimates:- if the argument <combine.with.metafor> = TRUE, ds.glmSLMA also returns 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)

there are a small number of more esoteric items of information returned by ds.glmSLMA. Additional information about these can be found in the help


Paul Burton for DataSHIELD Development Team

[Package dsBaseClient version 5.0.0 ]