# Generates 1-, 2-, and 3-dimensional contingency tables with option of assigning to serverside only and producing chi-squared statistics Creates 1-dimensional, 2-dimensional and 3-dimensional tables using the `table` function in native R. ```r ds.table( rvar = NULL, cvar = NULL, stvar = NULL, report.chisq.tests = FALSE, exclude = NULL, useNA = "always", suppress.chisq.warnings = FALSE, table.assign = FALSE, newobj = NULL, datasources = NULL, force.nfilter = NULL ) ``` ## Arguments - `rvar`: is a character string (in inverted commas) specifying the name of the variable defining the rows in all of the 2 dimensional tables that form the output. Please see 'details' above for more information about one-dimensional tables when a variable name is provided by but and are both NULL - `cvar`: is a character string specifying the name of the variable defining the columns in all of the 2 dimensional tables that form the output. - `stvar`: is a character string specifying the name of the variable that indexes the separate two dimensional tables in the output if the call specifies a 3 dimensional table. - `report.chisq.tests`: if TRUE, chi-squared tests are applied to every 2 dimensional table in the output and reported as "chisq.test_table.name". Default = FALSE. - `exclude`: this argument is passed through to the `table` function in native R which is called by `tableDS`. The help for `table` in native R indicates that 'exclude' specifies any levels that should be deleted for all factors in rvar, cvar or stvar. If the argument does not include NA and if the argument is not specified, it implies = "always" in DataSHIELD. If you read the help for `table` in native R including the 'details' and the 'examples' (particularly 'd.patho') you will see that the response of `table` to different combinations of the and arguments can be non-intuitive. This is particularly so if there is more than one type of missing (e.g. missing by observation as well as missing because of an NaN response to a mathematical function - such as log(-3.0)). In DataSHIELD, if you are in one of these complex settings (which should not be very common) and you cannot interpret the output that has been approached you might try: (1) making sure that the variable producing the strange results is of class factor rather than integer or numeric - although integers and numerics are coerced to factors by `ds.table` they can occasionally behave less well when the NA setting is complex; (2) specify both an argument e.g. exclude = c("NaN","3") and a argument e.g. useNA= "no"; (3) if you are excluding multiple levels e.g exclude = c("NA","3") then you can reduce this to one e.g. exclude = c("NA") and then remove the 3s by deleting rows of data, or converting the 3s to a different value. - `useNA`: this argument is passed through to the `table` function in native R which is called by `tableDS`. In DataSHIELD, this argument can take two values: "no" or "always" which indicate whether to include NA values in the table. For further information, please see the help for the argument (above) and/or the help for the `table` function in native R. Default value is set to "always". - `suppress.chisq.warnings`: if set to TRUE, the default warnings are suppressed that would otherwise be produced by the `table` function in native R whenever an expected cell count in one or more cells is less than 5. Default is FALSE. Further details can be found under 'details' and the help provided for the argument (above). - `table.assign`: is a Boolean argument set by default to FALSE. If it is FALSE the `ds.table` function acts as a standard aggregate function - it returns the table that is specified in its call to the clientside where it can be visualised and worked with by the analyst. But if is TRUE, the same table object is also written to the serverside. As explained under 'details' (above), this may be useful when some elements of a table need to be used to drive forward the overall analysis (e.g. to help select individuals for an analysis sub-sample), but the required table cannot be visualised or returned to the clientside because it fails disclosure rules. - `newobj`: this a character string providing a name for the output table object to be written to the serverside if is TRUE. If no explicit name for the table object is specified, but is nevertheless TRUE, the name for the serverside table object defaults to `table.newobj`. - `datasources`: a list of `DSConnection-class` objects obtained after login. If the the default set of connections will be used: see datashield.connections_default . If the is to be specified, it should be set without inverted commas: e.g. datasources=connections.em or datasources=default.connections. If you wish to apply the function solely to e.g. the second connection server in a set of three, the argument can be specified as: e.g. datasources=connections.em[2]. If you wish to specify the first and third connection servers in a set you specify: e.g. datasources=connections.em[c(1,3)]. - `force.nfilter`: if is non-NULL it must be specified as a positive integer represented as a character string: e.g. "173". This the has the effect of the standard value of 'nfilter.tab' (often 1, 3, 5 or 10 depending what value the data custodian has selected for this particular data set), to this new value (here, 173). CRUCIALLY, the `ds.table` function only allows the standard value to be INCREASED. So if the standard value has been set as 5 (as one of the R options set in the serverside connection), "6" and "4981" would be allowable values for the argument but "4" or "1" would not. The purpose of this argument is for the user or developer to force the table to fail the disclosure control tests so the he/she can see what then happens and check that it is behaving as anticipated/hoped. ## Returns Having created the requested table based on serverside data it is returned to the clientside for the analyst to visualise (unless it is blocked because it fails the disclosure control criteria or there is an error for some other reason). The clientside output from `ds.table` includes error messages that identify when the creation of a table from a particular study has failed and why. If table.assign=TRUE, `ds.table` also writes the requested table as an object named by the argument or set to 'newObj' by default. Further information about the visible material passed to the clientside, and the optional table object written to the serverside can be seen under 'details' (above). ## Details The `ds.table` function selects numeric, integer or factor variables on the serverside which define a contingency table with up to three dimensions. The native R `table` function basically operates on factors and if variables are specified that are integers or numerics they are first coerced to factors. If the 1-dimensional, 2-dimensional or 3-dimensional table generated from a given study satisfies appropriate disclosure-control criteria it can be returned directly to the clientside where it is presented as a study-specific table and is also included in a combined table across all studies. The data custodian responsible for data security in a given study can specify the minimum non-zero cell count that determines whether the disclosure-control criterion can be viewed as having been met. If the count in any one cell in a table falls below the specified threshold (and is also non-zero) the whole table is blocked and cannot be returned to the clientside. However, even if a table is potentially disclosive it can still be written to the serverside while an empty representation of the structure of the table is returned to the clientside. The contents of the cells in the serverside table object are reflected in a vector of counts which is one component of that table object. The true counts in the studyside vector are replaced by a sequential set of cell-IDs running from 1:n (where n is the total number of cells in the table) in the empty representation of the structure of the potentially disclosive table that is returned to the clientside. These cell-IDs reflect the order of the counts in the true counts vector on the serverside. In consequence, if the number 13 appears in a cell of the empty table returned to the clientside, it means that the true count in that same cell is held as the 13th element of the true count vector saved on the serverside. This means that a data analyst can still make use of the counts from a call to the `ds.table` function to drive their ongoing analysis even when one or more non-zero cell counts fall below the specified threshold for potential disclosure risk. Because the table object on the serverside cannot be visualised or transferred to the clientside, DataSHIELD ensures that although it can, in this way, be used to advance analysis, it does not create a direct risk of disclosure. The argument identifies the variable defining the rows in each of the 2-dimensional tables produced in the output. The argument identifies the variable defining the columns in the 2-dimensional tables produced in the output. In creating a 3-dimensional table the ('separate tables') argument identifies the variable that indexes the set of two dimensional tables in the output `ds.table`. As a minor technicality, it should be noted that if a 1-dimensional table is required, one only need specify a value for the argument and any one dimensional table in the output is presented as a row vectors and so technically the variable defines the columns in that 1 x n vector. However, the ds.table function deals with 1-dimensional tables differently to 2 and 3 dimensional tables and key components of the output for one dimensional tables are actually two dimensional: with rows defined by and with one column for each of the studies. The output list generated by `ds.table` contains tables based on counts named "table.name_counts" and other tables reporting corresponding column proportions ("table.name_col.props") or row proportions ("table.name_row.props"). In one dimensional tables in the output the output tables include _counts and _proportions. The latter are not called _col.props or _row.props because, for the reasons noted above, they are technically column proportions but are based on the distribution of the variable. If the argument is set to TRUE, chisq tests are applied to every 2-dimensional table in the output and reported as "chisq.test_table.name". The argument defaults to FALSE. If there is at least one expected cell counts < 5 in an output table, the native R function returns a warning. Because in a DataSHIELD setting this often means that every study and several tables may return the same warning and because it is debatable whether this warning is really statistically important, the argument can be set to TRUE to block the warnings. However, it is defaulted to FALSE. ## Author(s) Paul Burton and Alex Westerberg for DataSHIELD Development Team, 01/05/2020