larf {LARF}

Local Average Response Functions for Instrumental Variable Estimation of Treatment Effects

Description

The function provides instrumental variable estimation of treatment effects when both the endogenous treatment and its instrument are binary. Applicable to both binary and continuous outcomes.

Usage

larf(formula, treatment, instrument, data, method = "LS",  
     AME = FALSE, optimizer = "Nelder-Mead", zProb = NULL)

Arguments

formula

specification of the outcome model in the form like either y ~ x1 + x2 or y ~ X where X is a matrix containing all the covariates excluding the treatment. Also support multi-part formulas (Zeileis and Croissant, 2010). For example, y + d ~ x1 + x2 | z, where d represents the treatment and z the instrument.

treatment

A vector containing the binary treatment.

instrument

A vector containing the binary instrument for the endogenous treatment.

data

an optional data frame. If unspecified, the data will be taken from the working environment.

method

the estimation method to be used. The default is “LS", standing for least squares. “ML", standing for maximum likelihood, is an alternative.

AME

whether average marginal effects (AME) should be reported. The default is FALSE, in which case marginal effects at the means (MEM) are reported.

optimizer

the optimization algorithm for the ML method. It should be one of “Nelder-Mead", “BFGS", “CG", “L-BFGS-B", “SANN", or “Brent". See optim in R for more detail.

zProb

a vector containing the probability of receiving the treatment inducement (i.e., instrument = 1) that have been estimated by semiparametrical methods.

Details

larf is the high-level interface to the work-horse function larf.fit. A set of standard methods (including print, summary, coef, vcov, fitted, resid, predict) can be used to extract the corresponding information from a larf object.

The function provides instrumental variable estimation of treatment effects when both the endogenous treatment and its instrument (i.e., the treatment inducement) are binary. The method (Abadie, 2003) involves two steps. First, pseudo-weights are constructed from the probability of receiving the treatment inducement. By default the function estimates the probability by a Probit regression. But it also allows users to employ the probability that has been estimated by semiparametric methods. Second, the pseudo-weights are used to estimate the local average response function of the outcome conditional on the treatment and covariates. The function provides both least squares and maximum likelihood estimates of the conditional treatment effects.

Value

coefficients

Estimated coefficients.

SE

Standard errors of the estimated coefficients.

MargEff

Estimated marginal effects, available only for binary outcomes.

MargStdErr

Standard errors of the estimated marginal effects, available only for binary outcomes.

vcov

Variance covariance matrix of the estimated coefficients.

fitted.values

Predicted outcomes based on the estimated model. They are probabilities when the outcome is binary.

Note

We derived part of the code from the Matlab code written by Professor Alberto Abadie, available at http://www.hks.harvard.edu/fs/aabadie/larf.html. We thank Onur Altindag and Behzad Kianian for helpful suggestions on improving the computation.

Author(s)

Weihua An, Departments of Sociology and Statistics, Indiana University Bloomington, [e-mail].
Xuefu Wang, Department of Statistics, Indiana University Bloomington, [e-mail].

References

Abadie, Alberto. 2003. "Semiparametric Instrumental Variable Estimation of Treatment Response Models." Journal of Econometrics 113: 231-263.
An, Weihua and Xuefu Wang. 2016. "LARF: Instrumental Variable Estimation of Causal Effects through Local Average Response Functions." Journal of Statistical Software 71(1): 1-13.
Zeileis, Achim and Yves Croissant. 2010. "Extended Model Formulas in R: Multiple Parts and Multiple Responses." Journal of Statistical Software 34(1): 1-13. http://www.jstatsoft.org/v34/i01/.

See Also

larf.fit, c401k

Examples

data(c401k)
attach(c401k)

## Not run: 
# Continuous outcome. Treatment effects of participation in 401(k) 
# on net family financial assest
est1 <- larf(nettfa ~ inc + age + agesq + marr + fsize, treatment = p401k, 
instrument = e401k, data = c401k)
summary(est1)

# Nonparametric estimates of the probability of 
# receiving the treatment inducement
library(mgcv)
firstStep <- gam(e401k ~ s(inc) + s(age) + s(agesq) + marr + s(fsize), 
data=c401k, family=binomial(link = "probit"))
zProb <- firstStep$fitted
est2<- larf(nettfa ~ inc + age + agesq + marr + fsize, treatment = p401k, 
instrument = e401k, data = c401k, zProb = zProb)
summary(est2) 

# Binary outcome. Treatment effects of participation in 401(k) 
# on participation in IRA
est3 <- larf(pira ~ inc + age + agesq + marr + fsize, treatment = p401k, 
instrument = e401k, data = c401k)
summary(est3) 

## End(Not run)