The same model that was used for the synthesised data set is fitted to the
observed data set. The coefficients with confidence intervals for the
observed data is plotted together with their estimates from synthetic data.
When more than one synthetic data set has been generated (object$m>1)
combining rules are applied. Analysis-specific utility measures are used to
evaluate differences between synthetic and observed data.
# S3 method for fit.synds
compare(object, data, plot = "Z",
print.coef = FALSE, return.plot = TRUE, plot.intercept = FALSE,
lwd = 1, lty = 1, lcol = c("#1A3C5A","#4187BF"),
dodge.height = .5, point.size = 2.5,
population.inference = FALSE, ci.level = 0.95, ...)# S3 method for compare.fit.synds
print(x, print.coef = x$print.coef, ...)
An object of class compare.fit.synds which is a list with the
following components:
the original call to fit the model to the synthesised data set.
a data frame including estimates based on the observed
data: coefficients (Beta), their standard errors (se(Beta))
and Z scores (Z).
a data frame including (combined) estimates based on
the synthesised data: point estimates of observed data coefficients
(B.syn), standard errors of those estimates (se(B.syn)),
estimates of the observed standard errors (se(Beta).syn), Z scores
estimates (Z.syn) and their standard errors (se(Z.syn)).
Note that se(B.syn) and se(Z.syn) give the standard errors
of the mean of the m syntheses and can be made very small by
increasing m (see the vignette on inference for more details).
a data frame containing standardized differences between the coefficients estimated from the original data and those calculated from the combined synthetic data. The difference is standardized by dividing by the estimated standard error of the fit from the original. The corresponding p-values are calculated from a standard Normal distribution and represent the probability of achieving differences as large as those found if the model use for synthesis is compatible with the model that generated the original data.
Mean absolute standardized difference (over all coefficients).
a data frame containing the percentage of overlap between
the estimated synthetic confidence intervals and the original sample
confidence intervals for each parameter. When population.inference = TRUE
overlaps are calculated as suggested by Karr et al. (2006). Otherwise
a simpler overlap measure with intervals of equal length is calculated.
Mean confidence interval overlap (over all coefficients).
lack-of-fit measure from all m synthetic data
sets combined, calculated as follows, when object$incomplete = FALSE.
The vector of mean differences (diff) between the coefficients
calculated from the synthetic and original data provides
a standardised lack-of-fit = t(diff) %*% V^(-1) t(diff), where
%*% represents the matrix product and V^(-1) is the
inverse of the variance-covariance matrix for the mean coefficients
from the original data. If the model used to synthesize the data is
correct this quantity, which is a Mahalanobis distance measure, will
follow a chi-squared distribution with degrees of freedom, and thus
expectation, equal to the number of parameters (p) in the fitted
model. When object$incomplete = TRUE the variance-covariance matrix
of the coefficients is estimated from the differences between the m
estimates and the lack-of-fit statistic follows a Hotelling's T*2
distribution and the lack-of-fit statistic is referred to an F(p, m - p).
p-value for the combined lack-of-fit test of the NULL hypothesis that the method used for synthesis retains all relationships between variables that influence the parameters of the fit.
ggplot of the the coefficients with confidence
intervals for models based on observed and synthetic data.
If return.plot was set to FALSE then ci.plot
is NULL.
a logical value determining whether tables of estimates for the original and synthetic data should be printed.
the number of synthetic versions of the original (observed) data.
the number of coefficients in the fitted model (including an intercept).
whether methods for incomplete synthesis due to Reiter (2003) have been used in calculations.
whether intervals as decribed by Karr et al. (2016) have been calculated.
an object of type fit.synds created by fitting a model
to synthesised data set using function glm.synds or
lm.synds.
an original observed data set.
values to be plotted: "Z" (Z scores) or "coef"
(coefficients).
a logical value determining whether tables of estimates for the original and synthetic data should be printed.
a logical value indicating whether a confidence interval plot should be returned.
a logical value indicating whether estimates for intercept should be plotted.
the line type.
the line width.
line colours.
size of vertical shifts for confidence intervals to prevent overlapping.
size of plotting symbols used to plot point estimates of coefficients.
a logical value indicating whether intervals for inference to population quantities, as decribed by Karr et al. (2006), should be calculated and plotted. This option suppresses the lack-of-fit test and the standardised differences since these are based on differences standardised by the original interval widths.
Confidence interval coverage as a proportion.
additional parameters passed to ggplot.
an object of class compare.fit.synds.
This function can be used to evaluate whether the method used for
synthesis is appropriate for the fitted model. If this is the case the
estimates from the synthetic data of what would be expected from the original
data xpct(Beta) xpct(Z) should not differ from the estimates from
the observed data (Beta and Z) by more than would be expected from
the standard errors (se(Beta) and se(Z)). For more details see the
vignette on inference.
Karr, A., Kohnen, C.N., Oganian, A., Reiter, J.P. and Sanil, A.P. (2006). A framework for evaluating the utility of data altered to protect confidentiality. The American Statistician, 60(3), 224-232.
Nowok, B., Raab, G.M and Dibben, C. (2016). synthpop: Bespoke creation of synthetic data in R. Journal of Statistical Software, 74(11), 1-26. tools:::Rd_expr_doi("10.18637/jss.v074.i11").
Reiter, J.P. (2003) Inference for partially synthetic, public use microdata sets. Survey Methodology, 29, 181-188.
summary.fit.synds
ods <- SD2011[,c("sex","age","edu","smoke")]
s1 <- syn(ods, m = 3)
f1 <- glm.synds(smoke ~ sex + age + edu, data = s1, family = "binomial")
compare(f1, SD2011)
compare(f1, SD2011, print.coef = TRUE, plot = "coef")
Run the code above in your browser using DataLab