| ds.glmPredict {dsBaseClient} | R Documentation |
Applies native R's predict.glm() function to a serverside glm object previously created using ds.glmSLMA.
ds.glmPredict(
glmname = NULL,
newdataname = NULL,
output.type = "response",
se.fit = FALSE,
dispersion = NULL,
terms = NULL,
na.action = "na.pass",
newobj = NULL,
datasources = NULL
)
glmname |
is a character string identifying the glm object on serverside to which predict.glm is to be applied. Equivalent to <object> argument in native R's predict.glm which is described as: a fitted object of class inheriting from 'glm'. |
newdataname |
is a character string identifying an (optional) dataframe on the serverside in which to look for new covariate values with which to predict. If omitted, the original fitted linear predictors from the original glm fit are used as the basis of prediction. Precisely equivalent to the <newdata> argument in the predict.glm function in native R. |
output.type |
a character string taking the values 'response', 'link' or 'terms'. The value 'response' generates predictions on the scale of the original outcome, e.g. as proportions in a logistic regression. These are often called 'fitted values'. The value 'link' generates predictions on the scale of the linear predictor, e.g. log-odds in logistic regression, log-rate or log-count in Poisson regression. The predictions using 'response' and 'link' are identical for a standard Gaussian model with an identity link. The value 'terms' returns either fitted values or predicted values on the link scale based not on the whole linear predictor but on separate 'terms'. So, if age is modelled as a five level factor, one of the output components will relate to predictions (fitted values or link scale predictions) based on all five levels of age simultaneously. Any simple covariate (e.g. not a composite factor) will be treated as a term in its own right. ds.glmPredict's <output.type> argument is precisely equivalent to the <type> argument in native R's predict.glm function. |
se.fit |
logical if standard errors for the fitted predictions are required. Defaults to FALSE when the output contains only a vector (or vectors) of predicted values. If TRUE, the output also contains corresponding vectors for the standard errors of the predicted values, and a single value reporting the scale parameter of the model. ds.glmPredict's <se.fit> argument is precisely equivalent to the corresponding argument in predict.glm in native R. argument is equivalent to the <type> argument in native R's predict.glm function. |
dispersion |
numeric value specifying the dispersion of the GLM fit to be assumed in computing the standard errors. If omitted, that returned by summary applied to the glm object is used. e.g. if <dispersion> is unspecified the dispersion assumed for a logistic regression or Poisson model is 1. But if dispersion is set to 4, the standard errors of the predictions will all be multiplied by 2 (i.e. sqrt(4)). This is useful in making predictions from models subject to overdispersion. ds.glmPredict's <dispersion> argument is precisely equivalent to the corresponding argument in predict.glm in native R. |
terms |
a character vector specifying a subset of terms to return in the prediction. Only applies if output.type='terms'. ds.glmPredict's <terms> argument is precisely equivalent to the corresponding argument in predict.glm in native R. |
na.action |
character string determining what should be done with missing values in the data.frame identified by <newdataname>. Default is na.pass which predicts from the specified new data.frame with all NAs left in place. na.omit removes all rows containing NAs. na.fail stops the function if there are any NAs anywhere in the data.frame. For further details see help in native R. |
newobj |
a character string specifying the name of the serverside object to which the output object from the call to ds.glmPredict is to be written in each study. If no <newobj> argument is specified, the output object on the serverside defaults to the name "predict_glm". |
datasources |
specifies the particular 'connection object(s)' to use. e.g. if you have several data sets in the sources you are working with called opals.a, opals.w2, and connection.xyz, you can choose which of these to work with. The call 'datashield.connections_find()' lists all of the different datasets available and if one of these is called 'default.connections' that will be the dataset used by default if no other dataset is specified. If you wish to change the connections you wish to use by default the call datashield.connections_default('opals.a') will set 'default.connections' to be 'opals.a' and so in the absence of specific instructions to the contrary (e.g. by specifiying a particular dataset to be used via the <datasources> argument) all subsequent function calls will be to the datasets held in opals.a. If the <datasources> argument is specified, it should be set without inverted commas: e.g. datasources=opals.a or datasources=default.connections. The <datasources> argument also allows you to apply a function solely to a subset of the studies/sources you are working with. For example, the second source in a set of three, can be specified using a call such as datasources=connection.xyz[2]. On the other hand, if you wish to specify solely the first and third sources, the appropriate call will be datasources=connections.xyz[c(1,3)] |
Clientside function calling a single assign function (glmPredictDS.as) and a single aggregate function (glmPredictDS.ag). ds.glmPredict applies the native R predict.glm function to a glm object that has already been created on the serverside by fitting ds.glmSLMA. This is precisely the same as the glm object created in native R by fitting a glm using the glm function. Crucially, if ds.glmSLMA was originally applied to multiple studies the glm object created on each study is based solely on data from that study. ds.glmPredict has two distinct actions. First, the call to the assign function applies the standard predict.glm function of native R to the glm object on the serverside and writes all the output that would normally be generated by predict.glm to a newobj on the serverside. Because no critical information is passed to the clientside, there are no disclosure issues associated with this action. Any standard DataSHIELD functions can then be applied to the newobj to interpret the output. For example, it could be used as the basis for regression diagnostic plots. Second, the call to the aggregate function creates a non-disclosive summary of all the information held in the newobj created by the assign function and returns this summary to the clientside. For example, the full list of predicted/fitted values generated by the model could be disclosive. So although the newobj holds the full vector of fitted values, only the total number of values, the total number of valid (non-missing) values, the number of missing values, the mean and standard deviation of all valid values and the 5 are returned to the clientside by the aggregate function. The non-DataSHIELD arguments of ds.glmPredict are precisely the equivalent to those of predict.glm in native R and so all detailed information can be found using help(predict.glm) in native R.
ds.glmPredict calls the serverside assign function glmPredictDS.as which writes a new object to the serverside containing output precisely equivalent to predict.glm in native R. The name for this serverside object is given by the newobj argument or if that argument is missing or null it is called "predict_glm". In addition, ds.glmPredict calls the serverside aggregate function glmPredictDS.ag which returns an object containing non-disclosive summary statistics relating either to a single prediction vector called fit or, if se.fit=TRUE, of two vectors 'fit' and 'se.fit' - the latter containing the standard errors of the predictions in 'fit'. The non-disclosive summary statistics for the vector(s) include: length, the total number of valid (non-missing) values, the number of missing values, the mean and standard deviation of the valid values and the 5 the output always includes: the name of the serverside glm object being predicted from, the name - if one was specified - of the dataframe being used as the basis for predictions, the output.type specified ('link', 'response' or 'terms'), the value of the dispersion parameter if one had been specified and the residual scale parameter (which is multipled by sqrt(dispersion parameter) if one has been set). If output.type = 'terms', the summary statistics for the fit and se.fit vectors are replaced by equivalent summary statistics for each column in fit and se.fit matrices which each have k columns if k terms are being summarised.
Paul Burton, for DataSHIELD Development Team 13/08/20