icsw (version 1.0.0)

compliance.score: Estimate compliance scores using covariates.

Description

Estimate probability that units are compliers using generalized linear models for probability of being a complier or always-taker and for being an always-taker conditional on being a complier or always-taker. These compliance scores can be used in inverse probability weighting to estimate average treatment effects. In the case of one-sided non-compliance, this estimation is considerably simpler.

Usage

compliance.score(D, Z, W, weights = NULL,
  link = qnorm, inv.link = pnorm, genoud = TRUE,
  num.iter = ifelse(genoud, 200, 10000),
  one.sided = FALSE)

Arguments

D

Binary treatment of interest.

Z

Binary instrument.

W

Matrix of covariates for compliance model.

weights

Observation weights.

link

Link function applied for linear models. Defaults to probit link function.

inv.link

Inverse link function (i.e., mean function) applied for linear models. Defaults to probit mean function.

genoud

Whether to use global optimization via genetic optimization from package rgenoud. Default, and highly recommended because the likelihood need not be concave. Otherwise, use the BOBYQA algorithm for constrained optimization from package minqa.

num.iter

Number of iterations of optimization routine.

one.sided

Whether non-compliance is one-sided (logical). When compliance is one-sided, the previous four arguments are ignored, and the compliance scores are estimated with probit regression.

Value

Vector of estimated probabilities of being a complier (i.e., compliance scores).

Details

A unit \(i\) is a complier if \(D_{i1} > D_{i0}\), where \(D_{i1}\) and \(D_{i0}\) are the potential treatments for unit \(i\) when \(Z\) is set to 1 and 0. This is a latent (unobserved) characteristic of individual units, since each unit is only observed with one value of \(Z\).

By default this function uses genetic optimization via genoud because the loss function for the complier scores is not necessarily convex.

References

Peter M. Aronow and Allison Carnegie. (2013). Beyond LATE: Estimation of the average treatment effect with an instrumental variable. Political Analysis.

See Also

Used by icsw.tsls.

Examples

# NOT RUN {
# Load example dataset, see help(FoxDebate) for details
data(FoxDebate)

# Matrix of covariates
covmat <- with(FoxDebate, cbind(partyid, pnintst, watchnat, educad, readnews, 
                                gender, income, white))

# Estimate compliance scores with covariates, assuming (default) 
#   case of two-sided non-compliance
cscoreout <- with(FoxDebate, compliance.score(D = watchpro, Z = conditn,
                                              W = covmat))

# Extract vector of estimated compliance scores
cscore <- cscoreout$C.score
summary(cscore)

# }