lmdp: OLS differential privacy

In georgieevans/PrivacyUnbiased: Unbiased Estimates from Privacy Protected Data

Description

This function works similarly to lm. It takes a formula and data and returns an lmdp object containing bias corrected OLS coefficients and standard errors. The output can be summarised by inputting it as an argument to summary.lmdp. See an overview at bit.ly/PrivUexample.

Usage

lmdp(formula, data, bootstrap_var = FALSE, nsims_var = 500,
  noise = NULL)

Arguments

formula	An lm style formula.
data	The data to estimate the model on. The first row of data should contain the DP standard error associated with that column unless noise argument is not NULL.
bootstrap_var	If FALSE, then the variance is estimated via simulation. If TRUE then the variance covariance matrix is estimated via bootstrap methods. Default is FALSE unless model contains interaction terms/squared terms or fewer than 10000 obsesrvations
nsims_var	Number of bootstrap samples/simulations. Default is 500
noise	Set a default differentially private standard error for every column of the data matrix

Value

Returns an object of class lmdp containing:

b	Inconsistent OLS coefficient estimate
b_vcov	Estimate of variance covariance of b
beta_tilde	Consistent estimates of coefficients, \tilde{β}
beta_tilde_vcov	Estimate of variance covariance of \tilde{β}
var_sims	Full set of simulated/bootstrap estimates of b and \tilde{β}
Sigma_sq_hat	Estimate of σ^2
vc_pos_def	Indicator variable = 1 if covariance estimate was PD. NA if bootstrap used
boot	Indicator variable = 1 if bootstrap was used to estimate variance
est_vc	Variance-covariance matrix used in variance simulation. NA if bootstrap used
X	Matrix of covariates
Y	Dependent variable vector
formula	Model formula

Examples

## Not run: data(dp_data)
## Not run: lmdp_test <- lmdp(Y ~ X1 + X2 + X3, data = dp_data)
## Not run: summary(lmdp_test)

georgieevans/PrivacyUnbiased documentation built on May 22, 2021, 1:01 p.m.