| Title: | Some Invalid Some Valid Instrumental Variables Estimator |
|---|---|
| Description: | Selects invalid instruments amongst a candidate of potentially bad instruments. The algorithm selects potentially invalid instruments and provides an estimate of the causal effect between exposure and outcome. |
| Authors: | Hyunseung Kang <[email protected]> |
| Maintainer: | Hyunseung Kang <[email protected]> |
| License: | GPL-2 |
| Version: | 1.4 |
| Built: | 2024-11-16 09:15:30 UTC |
| Source: | https://github.com/cran/sisVIVE |
Computes a K-fold cross validation for
cv.sisVIVE(Y, D, Z, lambdaSeq, K = 10, intercept = TRUE, normalize = TRUE)cv.sisVIVE(Y, D, Z, lambdaSeq, K = 10, intercept = TRUE, normalize = TRUE)
Y |
A numeric vector of outcomes |
D |
A numeric vector of exposures |
Z |
A numeric matrix of instruments, with each column referring to one instrument |
lambdaSeq |
A numeric vector of lambdas to cross-validate from. Cross-validation will be performed only on these sequence of lambdas. You can either supply lambdaSeq or nLambda. See Details |
K |
Number of cross-validation folds |
intercept |
A logical declaring whether the intercept be included. Default is TRUE |
normalize |
A logical declaring whether the columns of Z should be scaled with variance 1. Default is TRUE |
Performs K-fold cross validiation to select lambda and returns the "best" lambda based on this cross-validation. If lambdaSeq is unspecified, the algorithm defaults to using the sequence of lambdas selected by sisVIVE. If lambdaSeq is specified, the algorithm will only evaluate its cross-validation on the specified lambdaSeq.
A list is returned, which contains the estimates of alpha, beta, and the set of invalid instruments for the "best" lambda chosen by cross validation
lambda |
"best" lambda as chosen by cross validation |
estCVError |
Estimated cross-validated error at this lambda |
alpha |
Estimate of alpha at the said lambda |
beta |
Estimate of beta, the causal effect of exposure on outcome, at the said lambda |
whichInvalid |
Estimate of set of invalid instruments at the said lambda |
Hyunseung Kang
library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = cv.sisVIVE(Y,D,Z,K=10)library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = cv.sisVIVE(Y,D,Z,K=10)
predict.sisVIVE can estimate a set of invalid instruments or the causal effect of exposure on outcome for any particular lambda(s)
## S3 method for class 'sisVIVE' predict(object, lambda, type = c("coefficients", "instruments"), ...)## S3 method for class 'sisVIVE' predict(object, lambda, type = c("coefficients", "instruments"), ...)
object |
A fitted sisVIVE object |
lambda |
A numerical vector of lambda(s) to obtain estimates. The length of this vector can be one. |
type |
If type="coefficients", predict returns estimated alpha and betas (i.e. the causal effect of exposure on outcome) for supplied lambda(s). If type = "instruments", predict returns estimated set of invalid instruments for supplied lambda(s) |
... |
Any other argument for predict.lars |
predict() estimates the alphas and betas for given sequence of lambda(s) or set of invalid instrument.
lambda |
Sequence of lambda(s) supplied by the user |
alpha |
If type="coefficients", alpha is the estimated direct effect of the instruments on the outcome. Each row refers to an estimated alpha for one lambda value and each column refers to the corresponding column in Z |
beta |
If type="coefficients", beta is the estimated causal effect of exposure on outcome. Each entry refers to an estimated beta for one lambda value |
instruments |
If type="instruments",provides a character vector of invalid instruments. Each entry refers to column indices of Z that are estimated to be invalid. |
Hyunseung Kang
summary, and cv.sisVIVE
library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = sisVIVE(Y,D,Z) estCoef1 = predict(result,lambda = 10,type="coefficients") estCoef2 = predict(result,lambda = seq(1,100,10),type="coefficients") estInst1 = predict(result,lambda = 10,type="instruments") estInst2 = predict(result,lambda = seq(1,100,10),type="instruments")library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = sisVIVE(Y,D,Z) estCoef1 = predict(result,lambda = 10,type="coefficients") estCoef2 = predict(result,lambda = seq(1,100,10),type="coefficients") estInst1 = predict(result,lambda = 10,type="instruments") estInst2 = predict(result,lambda = seq(1,100,10),type="instruments")
Selects invalid instruments from a candidate of instruments and estimates the causal effect of exposure on outcome
sisVIVE(Y, D, Z, intercept = TRUE, normalize = TRUE)sisVIVE(Y, D, Z, intercept = TRUE, normalize = TRUE)
Y |
A numeric vector of outcomes |
D |
A numeric vector of exposures |
Z |
A numeric matrix of instruments, with each column referring to one instrument |
intercept |
A logical declaring whether the intercept be included. Default is TRUE |
normalize |
A logical declaring whether the columns of Z should be scaled with variance 1. Default is TRUE |
Given a candidate of instruments, the algorithm selects invalid instruments and estimates the causal effect of exposure on outcome by leveraging LARS. The entire solution path, which includes the estimate of the causal effect as well as the estimated set of invalid instruments, is computed for lambda.
A "sisVIVE" object is returned, for which print, plot, summary, and predict methods exist.
alpha |
Estimates of alpha, the causal effect of the instrument on the outcome. Each row represents a lambda value and each column corresponds to the estimate of alpha corresponding to that same column in Z |
beta |
Estimate of beta, the causal effect of the exposure on the outcome. Each entry represents an estimate of beta for a lambda value |
whichInvalid |
Estimate of which instruments are invalid. Each entry refers to the column indexes of the Z matrix which were deemed to be invalid for a lambda value |
lambda |
Lambda used in the algorithm |
larsObject |
A LARS object |
Y |
A numeric vector of outcomes |
D |
A numeric vector of exposures |
Z |
A numeric matrix of instruments, with each column referring to one instrument |
Dhat |
A numeric vector of predicted exposure given instruments |
Hyunseung Kang. We thank Frank Windmeijer for his suggestion to improve the code.
For LARS: Efron, Hastie, Johnstone and Tibshirani (2003) "Least Angle Regression" (with discussion) Annals of Statistics
summary and predict methods for sisVIVE, and cv.sisVIVE which performs cross-validation to choose the tuning parameter
library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = sisVIVE(Y,D,Z) summary(result)library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = sisVIVE(Y,D,Z) summary(result)
Produces a summary of the sisVIVE object, including the estimates of the causal effect of exposure on outcome and the estimated set of invalid instruments for sequence of lambdas
## S3 method for class 'sisVIVE' summary(object,...)## S3 method for class 'sisVIVE' summary(object,...)
object |
a sisVIVE object |
... |
additional arguments for summary generic |
Each row summarizes an estimated causal effect of exposure on outcome and the estimated set of invalid instruments for one lambda
Each row summarizes an estimated causal effect of exposure on outcome and the estimated set of invalid instruments for one lambda. For the estimated set of invalid instruments, each number is a column index number for Z that is estimated to be invalid.
Hyunseung Kang
predict and cv.sisVIVE
library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = sisVIVE(Y,D,Z) summary(result)library(MASS) library(lars) n = 1000; L = 10; s= 3; Z <- matrix(rnorm(n*L),n,L) error <- mvrnorm(n,rep(0,2),matrix(c(1,0.8,0.8,1),2,2)) intD = rnorm(1); ZtoD = rnorm(L,0,1); ZtoY = c(rnorm(s),rep(0,L-s)); DtoY = 1; intY = rnorm(1) D = intD + Z %*% ZtoD + error[,1] Y = intY + Z %*% ZtoY + D * DtoY + error[,2] result = sisVIVE(Y,D,Z) summary(result)