| Type: | Package |
| Title: | Estimate IV-Optimal Individualized Treatment Rules |
| Version: | 0.1.0 |
| Author: | Bo Zhang |
| Maintainer: | Bo Zhang <bozhan@wharton.upenn.edu> |
| Description: | A method that estimates an IV-optimal individualized treatment rule. An individualized treatment rule is said to be IV-optimal if it minimizes the maximum risk with respect to the putative IV and the set of IV identification assumptions. Please refer to <doi:10.48550/arXiv.2002.02579> for more details on the methodology and some theory underpinning the method. Function IV-PILE() uses functions in the package 'locClass'. Package 'locClass' can be accessed and installed from the 'R-Forge' repository via the following link: https://r-forge.r-project.org/projects/locclass/. Alternatively, one can install the package by entering the following in R: 'install.packages("locClass", repos="http://R-Forge.R-project.org")'. |
| License: | GPL-3 |
| Encoding: | UTF-8 |
| LazyData: | true |
| RoxygenNote: | 7.1.0 |
| Depends: | R (≥ 2.10) |
| Suggests: | locClass |
| Imports: | stats, nnet, randomForest, dplyr, rlang |
| NeedsCompilation: | no |
| Packaged: | 2020-09-03 19:15:08 UTC; ASUS |
| Repository: | CRAN |
| Date/Publication: | 2020-09-11 08:40:03 UTC |
Estimate an IV-optimal individualized treatment rule
Description
IV_PILE estimates an IV-optimal individualized treatment
rule given a dataset with estimated partial identification intervals
for each instance.
Usage
IV_PILE(dt, kernel = "linear", C = 1, sig = 1/(ncol(dt) - 5))
Arguments
dt |
A dataframe whose first column is a binary IV 'Z', followed by q columns of observed covariates, a binary treatment indicator 'A', a binary outcome 'Y', lower endpoint of the partial identification interval 'L', and upper endpoint of the partial identification interval 'U'. The dataset has q+5 columns in total. |
kernel |
The kernel used in the weighted SVM algorithm. The user may choose between 'linear' (linear kernel) and 'radial' (Gaussian RBF kernel). |
C |
Cost of violating the constraint. This is the parameter C in the Lagrange formulation. |
sig |
Sigma in the Gaussian RBF kernel. Default is set to 1/dimension of covariates, i.e., 1/q. This parameter is not relevant for linear kernel. |
Value
An object of the type wsvm, inheriting from svm.
Examples
## Not run:
# It is necessary to install the package locClass in order
# to run the following code.
attach(dt_Rouse)
# Construct an IV out of differential distance to two-year versus
# four-year college. Z = 1 if the subject lives not farther from
# a 4-year college compared to a 2-year college.
Z = (dist4yr <= dist2yr) + 0
# Treatment A = 1 if the subject attends a 4-year college and 0
# otherwise.
A = 1 - twoyr
# Outcome Y = 1 if the subject obtained a bachelor's degree
Y = (educ86 >= 16) + 0
# Prepare the dataset
dt = data.frame(Z, female, black, hispanic, bytest, dadsome,
dadcoll, momsome, momcoll, fincome, fincmiss, A, Y)
# Estimate the Balke-Pearl bound by estimating each constituent
# conditional probability p(Y = y, A = a | Z, X) with a multinomial
# regression.
dt_with_BP_bound_multinom = estimate_BP_bound(dt, method = 'multinom')
# Estimate the IV-optimal individualized treatment rule using a
# linear kernel, under the putative IV and the Balke-Pearl bound.
iv_itr_BP_linear = IV_PILE(dt_with_BP_bound_multinom, kernel = 'linear')
## End(Not run)
Rouse (1995) dataset
Description
Variables of the dataset is as follows:
- educ86
Years of education since 1986.
- twoyr
Attending a two-year college immediately after high school.
- female
Gender: 1 if female and 0 otherwise.
- black
Race: 1 if African American and 0 otherwise.
- hispanic
Race: 1 if Hispanic and 0 otherwise.
- bytest
Test score.
- dadsome
Dad's education: some college.
- dadcoll
Dad's education: college.
- momsome
Mom's education: some college.
- momcoll
Mom's education: college.
- fincome
Family income.
- fincmiss
Missingness indicator for family income.
- tuition2
Average state two-year college tuition.
- tuition4
Average state four-year college tuition.
- dist2yr
Distance to the nearest two-year college.
- dist4yr
Distance to the nearest four-year college.
Usage
data(dt_Rouse)
Format
A data frame with 4437 rows and 16 columns.
Source
ss
Estimate the Balke-Pearl bound for each instance in a dataset
Description
estimate_BP_bound estimates the Balke-Pearl bound for
each instance in the input dataset with a binary IV, observed
covariates, a binary treatment indicator, and a binary outcome.
Usage
estimate_BP_bound(dt, method = "rf", nodesize = 5)
Arguments
dt |
A dataframe whose first column is a binary IV 'Z', followed by q columns of observed covariates, followed by a binary treatment indicator 'A', and finally followed by a binary outcome 'Y'. The dataset has q+3 columns in total. |
method |
A character string indicator the method used to estimate each constituent conditional probability of the Balke-Pearl bound. Users can choose to fit multinomial regression by setting method = 'multinom', and random forest by setting method = 'rf'. |
nodesize |
Node size to be used in a random forest algorithm if method is set to 'rf'. The default value is set to 5. |
Value
The original dataframe with two additional columns: L and U. L indicates the Balke-Pearl lower bound and U is the Balke-Pearl upper bound.
Examples
attach(dt_Rouse)
# Construct an IV out of differential distance to two-year versus
# four-year college. Z = 1 if the subject lives not farther from
# a 4-year college compared to a 2-year college.
Z = (dist4yr <= dist2yr) + 0
# Treatment A = 1 if the subject attends a 4-year college and 0
# otherwise.
A = 1 - twoyr
# Outcome Y = 1 if the subject obtained a bachelor's degree
Y = (educ86 >= 16) + 0
# Prepare the dataset
dt = data.frame(Z, female, black, hispanic, bytest, dadsome,
dadcoll, momsome, momcoll, fincome, fincmiss, A, Y)
# Calculate the Balke-Pearl bound by estimating each constituent
# conditional probability p(Y = y, A = a | Z, X) with a random
# forest.
dt_with_BP_bound_rf = estimate_BP_bound(dt, method = 'rf', nodesize = 5)
# Calculate the Balke-Pearl bound by estimating each constituent
# conditional probability p(Y = y, A = a | Z, X) with a multinomial
# regression.
dt_with_BP_bound_multinom = estimate_BP_bound(dt, method = 'multinom')
Estimate the partial identification bound as in Siddique (2013, JASA) for each instance in a dataset
Description
estimate_Sid_bound estimates the partial identification bound
for each instance in the input dataset with a binary IV, observed
covariates, a binary treatment indicator, and a binary outcome according
to Siddique (2013, JASA).
Usage
estimate_Sid_bound(dt, method = "rf", nodesize = 5)
Arguments
dt |
A dataframe whose first column is a binary IV 'Z', followed by q columns of observed covariates, followed by a binary treatment indicator 'A', and finally followed by a binary outcome 'Y'. The dataset has q+3 columns in total. |
method |
A character string indicator the method used to estimate each constituent conditional probability of the partial identification bound. Users can choose to fit multinomial regression by setting method = 'multinom', and random forest by setting method = 'rf'. |
nodesize |
Node size to be used in a random forest algorithm if method is set to 'rf'. The default value is set to 5. |
Value
The original dataframe with two additional columns: L and U. L indicates the lower bound and U the upper bound as in Siddique 2013
Examples
attach(dt_Rouse)
# Construct an IV out of differential distance to two-year versus
# four-year college. Z = 1 if the subject lives not farther from
# a 4-year college compared to a 2-year college.
Z = (dist4yr <= dist2yr) + 0
# Treatment A = 1 if the subject attends a 4-year college and 0
# otherwise.
A = 1 - twoyr
# Outcome Y = 1 if the subject obtained a bachelor's degree
Y = (educ86 >= 16) + 0
# Prepare the dataset
dt = data.frame(Z, female, black, hispanic, bytest, dadsome,
dadcoll, momsome, momcoll, fincome, fincmiss, A, Y)
# Calculate the Siddique bound by estimating each constituent
# conditional probability p(Y = y, A = a | Z, X) with a random
# forest.
dt_with_Sid_bound_rf = estimate_Sid_bound(dt, method = 'rf', nodesize = 5)
# Calculate the Siddique bound by estimating each constituent
# conditional probability p(Y = y, A = a | Z, X) with a multinomial
# regression.
dt_with_Sid_bound_multinom = estimate_Sid_bound(dt, method = 'multinom')