Generating Synthetic Versions of Sensitive Microdata for Statistical Disclosure Control

A tool for producing synthetic versions of microdata containing confidential information so that they are safe to be released to users for exploratory analysis. The key objective of generating synthetic data is to replace sensitive original values with synthetic ones causing minimal distortion of the statistical information contained in the data set. Variables, which can be categorical or continuous, are synthesised one-by-one using sequential modelling. Replacements are generated by drawing from conditional distributions fitted to the original data using parametric or classification and regression trees models. Data are synthesised via the function syn() which can be largely automated, if default settings are used, or with methods defined by the user. Optional parameters can be used to influence the disclosure risk and the analytical quality of the synthesised data. For a description of the implemented method see Nowok, Raab and Dibben (2016) .


News

synthpop 1.4-1

NEW FEATURES

  • New version of utility.gen() and print.utility.gen(). These now implement the methods described in Snoke et. al. and include the following:
    • appropriate method of getting the null distribution is selected based on synthesis details;
    • warning messages if CART models fail to split;
    • allow a seed value to be set and stored for resampling methods;
    • printing Z scores from logit models using the average fit over different syntheses.
  • New function multinom.synds() to fit multinomial models to synthetic data using the multinom() function from the package nnet.

CHANGES

  • Default cp parameter in syn.cart() changed to 1e-8.
  • Default print.tables parameter in utility.tab() changed to TRUE.
  • All syn() result components that are lists, e.g. cont.na, have their elements named.

BUG FIXES

  • Call to classIntervals() in utility.tab() uses style = "fisher" as default to avoid problems for variables with a small number of unique values.
  • utility.tab() computes breaks for grouping from the combined data rather than just from the observed. This avoids problems when synthetic values are outside the range of the observed ones.
  • In compare.fit.synds() the quantity real.varcov needs to be multiplied by sigma^2 when fitting.function is "lm".

synthpop 1.4-0

NEW FEATURES

  • Vignette on inference from fitted models (we are grateful to Joerg Drechsler for his comments).
  • New function syn.satcat allows saturated categorical models to be fitted from all possible interactions of the predictor variables.

CHANGES

  • utility.tab() returns p-values for utility measures rather than standardised versions of the measures.
  • smooth.vars parameter added to sdc() function which allows smoothing of numeric variables in the synthesised dataset.
  • If models = TRUE in syn(), for logreg and ployreg coefficients of the fitted model are returned.
  • Output from print.summary.fit.synds() is labelled differently according to whether population.inference is TRUE or FALSE (see vignette on inference). Also it now includes p-values and stars as are shown for lm() and glm().
  • Warning messages are given in summary.fit.synds() when inference is attempted from an invalid model where synthsis has not conditioned on any variables that have been left unsynthesised.
  • The msel parameter of print.summary.fit.synds() prints a table of estimates rather than a lsiting of each fit in detail.
  • Several changes in compare.fit.synds() are described in detail in the new vignette on inference ('Inference from fitted models in synthpop'). These include extensions of the lack-of-fit tests and confidence interval overlap measures for different options.

BUG FIXES

  • replicated.uniques() for a single variable.
  • syn.cart() for logical variables without missings (thanks to bug report by Ruben Arslan).
  • syn() can be used without loading the package (thanks to reported issue by Ruben Arslan)
  • Missing data factor level as a stratum in syn.strata()
  • seed for syn.strata().
  • print.fit.synds() for syn.strata() object.

synthpop 1.3-2

CHANGES

  • For consistency reasons tab.utility() changed to utility.tab() and utility.synds() to utility.gen().
  • utility.gen() under major revision and temporarily unavailable.
  • Revision of utility.tab() function: a measure of fit proposed by Voas and Williams and one proposed by Freeman and Tukey are calculated; continous variables are categorised using classIntervals() function.
  • Revision of compare.fit.synds() function: new lack-of-fit measures and mean values for existing analysis-specific utility measures (confidence interval overlap and absolute standardized difference between coefficient); return.result parameter replaced by print.coef with slightly different functionality - analysis-specific utility measures are always printed but you can choose whether to print or not model estimates.

BUG FIXES

  • compare.synds() for integer variables without missing values (thanks to bug report by Joerg Drechsler).

synthpop 1.3-1

CHANGES

  • Update of the vignette.

BUG FIXES

  • compare.synds() for variables with NA values in observed but not in synthetic data returns correct value (0) for NA category in synthetic data.
  • Invalid times argument corrected (lists of numbers coerced to numbers).

synthpop 1.3-0

NEW FEATURES

  • Storing results of CART models when models set to TRUE.
  • Function syn.strata() for stratified synthesis.
  • Function multi.compare() for multivariate comparison of synthesised and observed data.
  • Synthesising method "nested" for a variable nested within another variable.
  • Tabular utility function tab.utility() for comparing contingency tables from observed and synthesized data.
  • Parameter uniques.exclude for the sdc() function, which can be used to remove some variables from the identification of uniques.
  • Function replicated.uniques() returns a number of unique individuals in the original data set ($no.uniques).

CHANGES

  • Synthetic values of collinear variables are derived based on the one that is synthesised first and their method is set to "collinear". They do not have to be removed prior to synthesis.
  • Synthesising method for constant variables is set to "constant" and the variables are not removed from the synthesised data set when drop.not.used = TRUE.
  • Default synthesising method changed to "cart".
  • Default minnumlevels changed to -1 (during synthesis numeric variables are not changed to factors regardless of the number of distinct values).
  • Coefficient estimates and their confidence intervals are ploted in the same order as they are presented in a tabular form.
  • No message on the seed value used (it is stored in the result object).
  • Formula of the model to be fitted using glm.synds() or lm.synds() can be specified outside the function.
  • Massage for sdc() on number of replicated uniques also when it is equal to zero.
  • Maximum number of iterations for a multinomial model used in polyreg and polr method increased to 1000 (maxit parameter). Message if the limit is reached.
  • write.syn() saves complete synds object into a file synobject_filename.RData.
  • Error on exceeding maxfaclevels in not generated if method for the factor is set to "sample" or "nested".
  • For constant variables method is changed to constant.
  • Year format for variables ymarr and ysepdiv in SD2011 dataset changed from yy to yyyy.

BUG FIXES

  • Types and placement of special signs that are allowed in rules have been extended and include e.g. initial and closing round bracket.
  • compare.synds() provides output for logical variables.
  • Synthesis of logical variables with missing values.
  • Message about a change of method for a variable without predictors.
  • Check for filetype in write.syn()

synthpop 1.2-1

BUG FIXES

  • No calling var(x) on a factor x (in checks).
  • No contrasts attribute for factors synthesised using parametric method.
  • Misspelled vector name (nlevels) replaced with a correct one (nlevel).

synthpop 1.2-0

NEW FEATURES

  • A new function utility.synds() for distributional comparison of synthesised data with the original (observed) data using propensity scores.
  • New measures for comparing model estimates based on synthesised and observed data implemented in compare.fit.synds() function: standardized differences in coefficient values(coef.diff) and confidence interval overlap (ci.overlap).

CHANGES

  • No dependency on coefplot package.
  • Default for drop.not.used changed to FALSE.

synthpop 1.1-1

CHANGES

  • Both variable names and their column indices can be used in visit.sequence.
  • Arguments rules, rvalues, cont.na, semicont, smoothing, event, denom are specified as named lists, e.g. rules = list(marital = "age < 18") and do not have to be specified for all variables.
  • Optional arguments can be passed to synthesising functions by specifying funname.argname arguments, e.g. ctree.minbucket = 5; they are function-specific; minbucket removed from arguments.
  • Smoothing is possible for numeric variables when synthesised with the method 'sample'.
  • compare() is a generic function with two methods (for class synds and fit.synds); it replaced two separate functions.
  • New argument return.plot for compare() method for class fit.synds.
  • New argument msel for compare() method for class synds, which allows comparison for pooled or selected data set(s). Results for multiple synthetic data sets can be plotted on the same graph.
  • New argument nrow for compare() method for class synds; nrow and ncol determine number of plots per screen.
  • Argument plot.na for compare() method for class synds is no longer required and missing data categories for numeric variables are ploted on the same plot as non-missing values.
  • Argument object of lm.synds() and glm.synds() functions changed to data.
  • print() method for class fit.synds gives by default combined coefficient estimates only.
  • summary() method for class fit.synds gives combined coefficient estimates and their standard errors.
  • summary() method for class synds with multiple synthetic data sets provides by default summaries that are calculated by averaging summary values for all synthetic data copies.
  • Argument obs.data of compare.fit.synds() function changed to data.
  • Method surv.ctree and cart.bboot changed to survctree and cartbboot.

BUG FIXES

  • denom and event for variables with missing data.
  • maxfaclevels can be increased.
  • Continuous variables with missing data when zero is a non-missing value.
  • Synthesis of a single variable (with or without auxiliary predictors) now works.

synthpop 1.1-0

NEW FEATURES

  • Function sdc() for statistical disclosure control of the synthesised data set(s); function replicated.uniques() to determine which unique units in the synthesised data set(s) replicates unique units in the original data set.
  • Function read.obs() to import original data sets form external files.
  • Function write.syn() to export synthetic data sets to external files and create a text file with information about the synthesis.
  • syn() has new semicont parameter that allows to define spike(s) for semi-continuous variables in order to synthesise them separately.
  • lognorm, sqrtnorm and cubertnorm methods for synthesis by linear regression after natural logarithm, square root or cube root transformation of a dependent variable.
  • seed argument for syn() function.

CHANGES

  • Revised output of summary.fit.synds() and compare.fit.synds(); standard errors of Z scores corrected (se(Z.syn)) (thanks to Joerg Drechsler).
  • Figures for compare.fit.synds() and compare.synds() functions plotted using ggplot2 functions.
  • period.separated or alllowercase naming convention has been adopted and parameter names populationInference, visitSequence, predictorMatrix, contNA, defaultMethod, printFlag and nlevelmax have been changed to population.inference, visit.sequence, predictor.matrix, cont.na, default.method, print.flag and minnumlevels respectively.
  • Default for drop.pred.only changed to FALSE.

BUG FIXES

  • Rounding procedure (thanks to bug report by Joerg Drechsler).
  • Warning about extra disregarded argument family in compare.fit.synds().

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.

install.packages("synthpop")

1.4-3 by Beata Nowok, 3 months ago


Browse source code at https://github.com/cran/synthpop


Authors: Beata Nowok, Gillian M Raab, Joshua Snoke and Chris Dibben


Documentation:   PDF Manual  


Task views: Official Statistics & Survey Methodology


GPL-2 | GPL-3 license


Imports graphics, stats, utils, rpart, party, foreign, plyr, proto, polspline, randomForest, classInt

Depends on lattice, MASS, methods, nnet, ggplot2


See at CRAN