psc_algebraic_d: Propensity Score Calibration with Extra D Variable to Relax...

In vandomed/meuc: Measurement Error and Unmeasured Confounding

Description

Implements the "algebraic" version of propensity score calibration as described by Sturmer et al. (Am. J. Epidemiol. 2005), but with an extra D variable in the MEM, which allows the error-prone propensity score to be included in the TDM rather than assumed uninformative of the outcome.

Usage

psc_algebraic_d(all_data = NULL, main = NULL, internal = NULL,
  external = NULL, y_var, x_var, d_var, gs_vars, ep_vars,
  tdm_family = "gaussian", ep_data = "validation", delta_var = TRUE,
  boot_var = FALSE, boots = 100, alpha = 0.05)

Arguments

all_data	Data frame with data for main study and validation study.
main	Data frame with data for the main study.
internal	Data frame with data for internal validation study.
external	Data frame with data for the external validation study.
y_var	Character string specifying name of Y variable.
x_var	Character string specifying name of X variable.
d_var	Character string specifying name of D variable.
gs_vars	Character vector specifying names of variables for gold standard propensity score.
ep_vars	Character vector specifying names of variables for error-prone propensity score.
tdm_family	Character string specifying family of true disease model (see glm).
ep_data	Character string controlling what data is used to fit the error-prone propensity score model. Choices are "validation" for validation study data, "all" for main study and validation study data, and "separate" for validation data for first step and main study data for second step.
delta_var	Logical value for whether to calculate a Delta method variance-covariance matrix. May not be justified theoretically because propensity scores are estimated, but also may perform better than bootstrap in certain scenarios, e.g. if bootstrap is prone to extreme estimates.
boot_var	Logical value for whether to calculate a bootstrap variance-covariance matrix.
boots	Numeric value specifying number of bootstrap samples to use.
alpha	Significance level for percentile bootstrap confidence interval.

Details

The true disease model is a GLM:

g[E(Y)] = beta_0 + beta_x X + beta_g G + beta_gstar Gstar

where G = P(X|C,Z), with C but not Z available in the main study, and Gstar = P(X|C). Logistic regression models are used to model P(X = 1|Z,C) and P(X = 1|C):

logit[P(X = 1)] = gamma_0 + gamma_z^T Z + gamma_c^T C logit[P(X = 1)] = gamma_0^* + gamma_c^*T C

A linear model is used to map fitted error-prone propensity scores Gstar to the expected gold standard propensity score G logistic regression):

E(G) = lambda_0 + lambda_d D + lambda_x X + lambda_g Gstar

There should be main study data with (Y, X, D, C) as well as external validation data with (X, D, C, Z).

References

Sturmer, T., Schneeweiss, S., Avorn, J. and Glynn, R.J. (2005) "Adjusting effect estimates for unmeasured confounding with validation data using propensity score calibration." Am. J. Epidemiol. 162(3): 279-289.

vandomed/meuc documentation built on May 12, 2019, 6:17 p.m.