replace_high_risk: Replace high-risk cases in synthetic data

In simongrund1/robosynth: Robust Synthetic Data Generation with Data-Augmented Multiple Imputation

Replace high-risk cases in synthetic data

Description

Replaces values of high-risk cases in synthetic data sets with values from similar cases as determined by nearest-neighbor matching.

Usage

replace.high.risk(data, R, threshold, k = 1, match.variables, replace.variables, n.replace, include = TRUE)

Arguments

data	list of synthetic data sets as returned by extract (or similar).
R	numeric: matrix of record-level risk estimates as returned by compute.risk (or similar).
threshold	numeric: the threshold, above which cases are considered "high-risk"
k	integer: number of neighbors from which replacements are sampled. The default is to replace with the nearest neighbor.
match.variables	character: names of variables to be used in the search for nearest neighbors.
replace.variables	character: names of variables to be replaced.
n.replace	integer: (maximum) number of cases to replace in each data set.
include	logical: if TRUE (the default), high-risk cases are included in the search for nearest neighbors. Otherwise, only low-risk cases are considered possible neighbors.

Details

This function can be used to replace cases that are at a high risk of identification disclosure with the values of similar cases in a list of synthetic data sets. For each target case, replacement values are sampled randomly from its $k$ nearest neighbors, which have the most similar (or identical) configuration on categorical variables and similar (in terms of squared distance) values on continuous variables.

The primary purpose of this function is to reduce the risk of (identification) disclosure by replacing the values of a *small* number of high-risk cases on selected variables. The number of cases that are considered "high-risk" is determined by threshold, and the number of replacements is determined by n.replace. If the number of replacements is lower than the number of high-risk cases, then a random selection of high-risk cases will be replaced.

The result of the replacement algorithm is a list of synthetic data sets with a (usually) lower disclosure risk but also (usually) lower utility. For this reason, it is good practice to replace only a small proportion of high-risk cases and to re-compute the risk of disclosure in the resulting data sets (see compute.risk).

Value

An object of class robosynth.list (a list of data frames).

Author(s)

Simon Grund

See Also

extract, compute.risk

Examples

# create masked copies
sociosexuality <- within(sociosexuality, {

  m_sex <- mask.categorical(sex, probability = .80)
  m_sexpref <- mask.categorical(sexpref, probability = .60)
  m_age <- mask.continuous(age, reliability = .90)

})

# combine synthesis and masking models
models <- combine.models(

  synthesis.model(sex ~ 1, type = "binary"),
  synthesis.model(sexpref ~ 1 + sex, type = "categorical"),
  synthesis.model(age ~ 1 + sex + sexpref, type = "continuous"),

  masking.model(m_sex ~ sex, type = "binary"),
  masking.model(m_sexpref ~ sexpref, type = "categorical"),
  masking.model(m_age ~ age, type = "continuous"),

  data = sociosexuality

)

# run synthesis
syn <- synthesize(models = models, m = 5, iter = 5)

# extract list of synthetic data sets
synlist <- extract(syn)

# compute risk for matching by "age"
risk <- compute.risk(synlist, original = sociosexuality, synthetic = "age")
summary(risk, threshold = .20, show.cases = 3)

# replace values in "age" using cases with similar "sex", "sexpref", and "age"
synlist.rep <- replace.high.risk(synlist, R = risk$R, threshold = .20, match.variables = c("sex", "sexpref", "age"), replace.variables = "sex", n.replace = 3)

# re-compute risk
risk.rep <- compute.risk(synlist.rep, original = sociosexuality, synthetic = "age")
summary(risk.rep, threshold = .20, show.cases = 3)

simongrund1/robosynth documentation built on March 20, 2022, 6:15 p.m.