Package 'ivmodel'

Statistical Inference and Sensitivity Analysis for Instrumental Variables Model

Description

The package fits an instrumental variables (IV) model of the following type. Let $Y$, $D$, $X$, and $Z$ represent the outcome, endogenous variable, p dimensional exogenous covariates, and L dimensional instruments, respectively; note that the intercept can be considered as a vector of ones and a part of the exogenous covariates $X$.The package assumes the following IV model

$Y = X \alpha + D \beta + \epsilon, E(\epsilon | X, Z) = 0$

It carries out several IV regressions, diagnostics, and tests associated with the parameter $\beta$ in the IV model. Also, if there is only one instrument, the package runs a sensitivity analysis discussed in Jiang et al. (2015).

The package is robust to most data formats, including factor and character data, and can handle very large IV models efficiently using a sparse QR decomposition.

Details

Supply the outcome $Y$, the endogenous variable $D$, and a data frame and/or matrix of instruments $Z$, and a data frame and/or matrix of exogenous covariates $X$ (optional) and run ivmodel. Alternatively, one can supply a formula. ivmodel will generate all the relevant statistics for the parameter $\beta$.

The DESCRIPTION file:

Package:	ivmodel
Type:	Package
Title:	Statistical Inference and Sensitivity Analysis for Instrumental Variables Model
Version:	1.9.1
Date:	2023-04-08
Author:	Hyunseung Kang, Yang Jiang, Qingyuan Zhao, and Dylan Small
Maintainer:	Hyunseung Kang <[email protected]>
Description:	Carries out instrumental variable estimation of causal effects, including power analysis, sensitivity analysis, and diagnostics. See Kang, Jiang, Zhao, and Small (2021) <https://muse.jhu.edu/article/804372> for details.
Imports:	stats,Matrix,Formula,reshape2,ggplot2
License:	GPL-2 | file LICENSE
LazyData:	true
RoxygenNote:	7.2.3
NeedsCompilation:	no
Suggests:	testthat
Repository:	https://mrcieu.r-universe.dev
RemoteUrl:	https://github.com/hyunseungkang/ivmodel
RemoteRef:	HEAD
RemoteSha:	825fa71c52961bc8b1066f6acb1bd24e93abcbba

Index of help topics:

AR.power                Power of the Anderson-Rubin (1949) Test
AR.size                 Sample Size Calculator for the Power of the
                        Anderson-Rubin (1949) Test
AR.test                 Anderson-Rubin (1949) Test
ARsens.power            Power of the Anderson-Rubin (1949) Test with
                        Sensitivity Analysis
ARsens.size             Sample Size Calculator for the Power of the
                        Anderson-Rubin (1949) Test with Sensitivity
                        Analysis
ARsens.test             Sensitivity Analysis for the Anderson-Rubin
                        (1949) Test
CLR                     Conditional Likelihood Ratio Test
Fuller                  Fuller-k Estimator
IVpower                 Power calculation for IV models
IVsize                  Calculating minimum sample size for achieving a
                        certain power
KClass                  k-Class Estimator
LIML                    Limited Information Maximum Likelihood Ratio
                        (LIML) Estimator
TSLS.power              Power of TSLS Estimator
TSLS.size               Sample Size Calculator for the Power of
                        Asymptotic T-test
balanceLovePlot         Create Love plot of standardized covariate mean
                        differences
biasLovePlot            Create Love plot of treatment bias and
                        instrument bias
card.data               Card (1995) Data
coef.ivmodel            Coefficients of the Fitted Model in the
                        'ivmodel' Object
coefOther               Exogenous Coefficients of the Fitted Model in
                        the 'ivmodel' Object
confint.ivmodel         Confidence Intervals for the Fitted Model in
                        'ivmodel' Object
distributionBalancePlot
                        Plot randomization distributions of the
                        Mahalanobis distance
fitted.ivmodel          Extract Model Fitted values in the 'ivmodel'
                        Object
getCovMeanDiffs         Get Covariate Mean Differences
getMD                   Get Mahalanobis Distance
getStandardizedCovMeanDiffs
                        Get Standardized Covariate Mean Differences
icu.data                Pseudo-data based on Branson and Keele (2020)
iv.diagnosis            Diagnostics of instrumental variable analysis
ivmodel                 Fitting Instrumental Variables (IV) Models
ivmodel-package         Statistical Inference and Sensitivity Analysis
                        for Instrumental Variables Model
ivmodelFormula          Fitting Instrumental Variables (IV) Models
model.matrix.ivmodel    Extract Design Matrix for 'ivmodel' Object
para                    Parameter Estimation from Ivmodel
permTest.absBias        Perform a permutation test using the sum of
                        absolute biases
permTest.md             Perform a permutation test using the
                        Mahalanobis distance
residuals.ivmodel       Residuals from the Fitted Model in the
                        'ivmodel' Object
vcov.ivmodel            Calculate Variance-Covariance Matrix (i.e.
                        Standard Error) for k-Class Estimators in the
                        'ivmodel' Object
vcovOther               Variance of Exogenous Coefficients of the
                        Fitted Model in the 'ivmodel' Object

Author(s)

Hyunseung Kang, Yang Jiang, Qingyuan Zhao, and Dylan Small

Maintainer: Hyunseung Kang <[email protected]>

References

Anderson, T. W. and Rubin, H. (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. Annals of Mathematical Statistics 20, 46-63.

Andrews, D. W. K., Moreira, M. J., and Stock, J. H. (2006). Optimal two-side invariant similar tests for instrumental variables regression. Econometrica 74, 715-752.

Card, D. Using Geographic Variation in College Proximity to Estimate the Return to Schooling. In Aspects of Labor Market Behavior: Essays in Honor of John Vanderkamp, eds. L.N. Christophides, E.K. Grant and R. Swidinsky. 201-222. National Longitudinal Survey of Young Men: https://www.nlsinfo.org/investigator/pages/login.jsp

Fuller, W. (1977). Some properties of a modification of the limited information estimator. Econometrica, 45, 939-953.

Moreira, M. J. (2003). A conditional likelihood ratio test for structural models. Econometrica 71, 1027-1048.

Sargan, J. D. (1958). The estimation of economic relationships using instrumental variables. Econometrica , 393-415.

Wang, X., Jiang, Y., Small, D. and Zhang, N. (2017), Sensitivity analysis and power for instrumental variable studies. Biometrics 74(4), 1150-1160.

Examples

data(card.data)
# One instrument #
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
card.model1IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
card.model1IV

# Multiple instruments
Z = card.data[,c("nearc4","nearc2")]
card.model2IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
card.model2IV

---

Power of the Anderson-Rubin (1949) Test

Description

AR.power computes the power of Anderson-Rubin (1949) test based on the given values of parameters.

Usage

AR.power(n, k, l, beta, gamma, Zadj_sq, 
         sigmau, sigmav, rho, alpha = 0.05)

Arguments

n	Sample size.
k	Number of exogenous variables.
l	Number of instrumental variables.
beta	True causal effect minus null hypothesis causal effect.
gamma	Regression coefficient for effect of instruments on treatment.
Zadj_sq	Variance of instruments after regressed on the observed variables.
sigmau	Standard deviation of potential outcome under control. (structural error for y)
sigmav	Standard deviation of error from regressing treatment on instruments.
rho	Correlation between u (potential outcome under control) and v (error from regressing treatment on instrument).
alpha	Significance level.

Value

Power of the Anderson-Rubin test based on the given values of parameters.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T.W. and Rubin, H. (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. Annals of Mathematical Statistics, 20, 46-63.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

# Assume we calculate the power of AR test in a study with one IV (l=1) 
# and the only one exogenous variable is the intercept (k=1). 

# Suppose the difference between the null hypothesis and true causal 
# effect is 1 (beta=1).
# The sample size is 250 (n=250), the IV variance is .25 (Zadj_sq =.25).
# The standard deviation of potential outcome is 1(sigmau= 1). 
# The coefficient of regressing IV upon exposure is .5 (gamma= .5).
# The correlation between u and v is assumed to be .5 (rho=.5). 
# The standard deviation of first stage error is .4 (sigmav=.4). 
# The significance level for the study is alpha = .05.

# power of Anderson-Rubin test:
AR.power(n=250, k=1, l=1, beta=1, gamma=.5, Zadj_sq=.25, 
         sigmau=1, sigmav=.4, rho=.5, alpha = 0.05)

---

Sample Size Calculator for the Power of the Anderson-Rubin (1949) Test

Description

AR.size computes the minimum sample size required for achieving certain power of Anderson-Rubin (1949) test for giving value of parameters.

Usage

AR.size(power, k, l, beta, gamma, Zadj_sq, 
        sigmau, sigmav, rho, alpha = 0.05)

Arguments

power	The desired power over a constant.
k	Number of exogenous variables.
l	Number of instrumental variables.
beta	True causal effect minus null hypothesis causal effect.
gamma	Regression coefficient for the effect of instrument on treatment.
Zadj_sq	Variance of instruments after regressed on the observed variables.
sigmau	Standard deviation of potential outcome under control (structural error for y).
sigmav	Standard deviation of error from regressing treatment on instruments
rho	Correlation between u (potential outcome under control) and v (error from regressing treatment on instrument).
alpha	Significance level.

Value

Minimum sample size required for achieving certain power of Anderson-Rubin (1949) test.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T.W. and Rubin, H. (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

# Assume we performed an AR test in a study with one IV (l=1) and the 
# only one exogenous variable is the intercept (k=1). We want to know 
# the minimum sample size for this test to have an at least 0.8 power.

# Suppose the difference between the null hypothesis and true causal 
# effect is 1 (beta=1).
# The IV variance is .25 (Zadj_sq =.25).
# The standard deviation of potential outcome is 1(sigmau= 1). 
# The coefficient of regressing IV upon exposure is .5 (gamma= .5).
# The correlation between u and v is assumed to be .5 (rho=.5). 
# The standard deviation of first stage error is .4 (sigmav=.4). 
# The significance level for the study is alpha = .05.

# minimum sample size required for Anderson-Rubin test:
AR.size(power=0.8, k=1, l=1, beta=1, gamma=.5, Zadj_sq=.25, 
        sigmau=1, sigmav=.4, rho=.5, alpha = 0.05)

---

Anderson-Rubin (1949) Test

Description

AR.test computes the Anderson-Rubin (1949) test for the ivmodel object as well as the associated confidence interval.

Usage

AR.test(ivmodel, beta0 = 0, alpha = 0.05)

Arguments

ivmodel	ivmodel object
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is 0.
alpha	The significance level for hypothesis testing. Default is 0.05.

Value

AR.test returns a list containing the following components

Fstat	The value of the test statistic for testing the null hypothesis $H_0: \beta = \beta_0$ in ivmodel
df	degree of freedom for the test statistic
p.value	The p value of the test under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel
ci	A matrix of two columns, each row contains an interval associated with the confidence interval
ci.info	A human-readable string describing the confidence interval

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T.W. and Rubin, H. (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
AR.test(foo)

---

Power of the Anderson-Rubin (1949) Test with Sensitivity Analysis

Description

ARsens.power computes the power of sensitivity analysis, which is based on an extension of Anderson-Rubin (1949) test and allows IV be possibly invalid within a certain range.

Usage

ARsens.power(n, k, beta, gamma, Zadj_sq, sigmau, sigmav, rho, 
             alpha = 0.05, deltarange = deltarange, delta = NULL)

Arguments

n	Sample size.
k	Number of exogenous variables.
beta	True causal effect minus null hypothesis causal effect.
gamma	Regression coefficient for effect of instruments on treatment.
Zadj_sq	Variance of instruments after regressed on the observed variables.
sigmau	Standard deviation of potential outcome under control (structural error for y).
sigmav	Standard deviation of error from regressing treatment on instruments.
rho	Correlation between u (potential outcome under control) and v (error from regressing treatment on instrument).
alpha	Significance level.
deltarange	Range of sensitivity allowance. A numeric vector of length 2.
delta	True value of sensitivity parameter when calculating the power. Usually take delta = 0 for the favorable situation or delta = NULL for unknown delta.

Value

Power of sensitivity analysis for the proposed study, which extends the Anderson-Rubin (1949) test with possibly invalid IV. The power formula is derived in Jiang, Small and Zhang (2015).

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T.W. and Rubin, H. (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63.
Wang, X., Jiang, Y., Small, D. and Zhang, N (2017), Sensitivity analysis and power for instrumental variable studies, (under review of Biometrics).

See Also

See also ivmodel for details on the instrumental variables model.

Examples

# Assume we calculate the power of sensitivity analysis in a study with
# one IV (l=1) and the only exogenous variable is the intercept (k=1). 

# Suppose the difference between the null hypothesis and true causal 
# effect is 1 (beta=1).
# The sample size is 250 (n=250), the IV variance is .25 (Zadj_sq =.25).
# The standard deviation of potential outcome is 1(sigmau= 1). 
# The coefficient of regressing IV upon exposure is .5 (gamma= .5).
# The correlation between u and v is assumed to be .5 (rho=.5). 
# The standard deviation of first stage error is .4 (sigmav=.4). 
# The significance level for the study is alpha = .05.

# power of sensitivity analysis under the favorable situation, 
# assuming the range of sensitivity allowance is (-0.1, 0.1)
ARsens.power(n=250, k=1, beta=1, gamma=.5, Zadj_sq=.25, sigmau=1, 
     sigmav=.4, rho=.5, alpha = 0.05, deltarange=c(-0.1, 0.1), delta=0)

# power of sensitivity analysis with unknown delta, 
# assuming the range of sensitivity allowance is (-0.1, 0.1)
ARsens.power(n=250, k=1, beta=1, gamma=.5, Zadj_sq=.25, sigmau=1, 
     sigmav=.4, rho=.5, alpha = 0.05, deltarange=c(-0.1, 0.1))

---

Sample Size Calculator for the Power of the Anderson-Rubin (1949) Test with Sensitivity Analysis

Description

ARsens.size computes the minimum sample size required for achieving certain power of sensitivity analysis, which is based on an extension of Anderson-Rubin (1949) test and allows IV be possibly invalid within a certain range.

Usage

ARsens.size(power, k, beta, gamma, Zadj_sq, sigmau, sigmav, rho, 
            alpha = 0.05, deltarange = deltarange, delta = NULL)

Arguments

power	The desired power over a constant.
k	Number of exogenous variables. =
beta	True causal effect minus null hypothesis causal effect.
gamma	Regression coefficient for effect of instruments on treatment.
Zadj_sq	Variance of instruments after regressed on the observed covariates.
sigmau	Standard deviation of potential outcome under control (structural error for y).
sigmav	Standard deviation of error from regressing treatment on instruments.
rho	Correlation between u (potential outcome under control) and v (error from regressing treatment on instruments).
alpha	Significance level.
deltarange	Range of sensitivity allowance. A numeric vector of length 2.
delta	True value of sensitivity parameter when calculating power. Usually take delta = 0 for the favorable situation or delta = NULL for unknown delta.

Value

Minimum sample size required for achieving certain power of sensitivity analysis for the proposed study, which extends the Anderson-Rubin (1949) test with possibly invalid IV. The power formula is derived in Jiang, Small and Zhang (2015).

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T.W. and Rubin, H. (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63.
Wang, X., Jiang, Y., Small, D. and Zhang, N (2017), Sensitivity analysis and power for instrumental variable studies, (under review of Biometrics).

See Also

See also ivmodel for details on the instrumental variables model.

Examples

# Assume we performed a sensitivity analysis in a study with one 
# IV (l=1) and the only exogenous variable is the intercept (k=1). 
# We want to calculate the minimum sample size needed for this 
# sensitivity analysis to have an at least 0.8 power.

# Suppose the difference between the null hypothesis and true causal 
# effect is 1 (beta=1).
# The IV variance is .25 (Zadj_sq =.25).
# The standard deviation of potential outcome is 1(sigmau= 1). 
# The coefficient of regressing IV upon exposure is .5 (gamma= .5).
# The correlation between u and v is assumed to be .5 (rho=.5). 
# The standard deviation of first stage error is .4 (sigmav=.4). 
# The significance level for the study is alpha = .05.

# minimum sample size for sensitivity analysis under the favorable 
# situation, assuming the range of sensitivity allowance is (-0.1, 0.1)
ARsens.size(power=0.8, k=1, beta=1, gamma=.5, Zadj_sq=.25, sigmau=1, 
    sigmav=.4, rho=.5, alpha = 0.05, deltarange=c(-0.1, 0.1), delta=0)

# minimum sample size for sensitivity analysis with unknown delta, 
# assuming the range of sensitivity allowance is (-0.1, 0.1)
ARsens.size(power=0.8, k=1, beta=1, gamma=.5, Zadj_sq=.25, sigmau=1, 
    sigmav=.4, rho=.5, alpha = 0.05, deltarange=c(-0.1, 0.1))

---

Sensitivity Analysis for the Anderson-Rubin (1949) Test

Description

ARsens.test computes sensitivity analysis with possibly invalid instruments, which is an extension of the Anderson-Rubin (1949) test. The formula for sensitivity analysis is derived in Jiang, Small and Zhang (2015).

Usage

ARsens.test(ivmodel, beta0 = 0, alpha = 0.05, deltarange = NULL)

Arguments

ivmodel	ivmodel object.
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel
alpha	The significance level for hypothesis testing. Default is 0.05.
deltarange	Range of sensitivity allowance. A numeric vector of length 2.

Value

ARsens.test returns a list containing the following components

ncFstat	The value of the test statistic for testing the null hypothesis $H_0: \beta = \beta_0$ in ivmodel
df	degree of freedom for the test statistic
ncp	non-central parameter for the test statistic
p.value	The p value of the test under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel
ci	A matrix of two columns, each row contains an interval associated with the confidence interval
ci.info	A human-readable string describing the confidence interval
deltarange	The inputted range of sensitivity allowance.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T.W. and Rubin, H. (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63.
Wang, X., Jiang, Y., Small, D. and Zhang, N. (2017), Sensitivity analysis and power for instrumental variable studies, (under review of Biometrics).

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
ARsens.test(foo, deltarange=c(-0.03, 0.03))

---

Create Love plot of standardized covariate mean differences

Description

balanceLovePlot creates a Love plot of the standardized covariate mean differences across the treatment and the instrument. Can also display the permutation quantiles for these quantities. This function is used to create Figure 3a in Branson and Keele (2020).

Usage

balanceLovePlot(X, D, Z, permQuantiles = FALSE, alpha = 0.05, perms = 1000)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
D	Indicator vector for a binary treatment (must contain 1 or 0 for each unit).
Z	Indicator vector for a binary instrument (must contain 1 or 0 for each unit).
permQuantiles	If TRUE, displays the permutation quantiles for the standardized covariate mean differences.
alpha	The significance level used for the permutation quantiles. For example, if alpha = 0.05, then the 2.5% and 97.5% permutation quantiles are displayed.
perms	Number of permutations used to approximate the permutation quantiles.

Value

Plot of the standardized covariate mean differences across the treatment and the instrument.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
	data(icu.data)
	#the covariate matrix is
	X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
	#the treatment
	D = icu.data$icu_bed
	#the instrument
	Z = icu.data$open_bin
	#make the Love plot with permutation quantiles
	## Not run: balanceLovePlot(X = X, D = D, Z = Z, permQuantiles = TRUE, perms = 500)

---

Create Love plot of treatment bias and instrument bias

Description

biasLovePlot creates a Love plot of the bias across the treatment and the instrument. Can also display the permutation quantiles for these quantities. Note that the bias is different for the treatment than for the instrument, as discussed in Equation (3) of Branson and Keele (2020). This function is used to create Figure 3b in Branson and Keele (2020).

Usage

biasLovePlot(X, D, Z, permQuantiles = FALSE, alpha = 0.05, perms = 1000)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
D	Indicator vector for a binary treatment (must contain 1 or 0 for each unit).
Z	Indicator vector for a binary instrument (must contain 1 or 0 for each unit).
permQuantiles	If TRUE, displays the permutation quantiles for the biases.
alpha	The significance level used for the permutation quantiles. For example, if alpha = 0.05, then the 2.5% and 97.5% permutation quantiles are displayed.
perms	Number of permutations used to approximate the permutation quantiles.

Value

Plot of the bias across the treatment and the instrument.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
  data(icu.data)
  #the covariate matrix is
  X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
  #the treatment
  D = icu.data$icu_bed
  #the instrument
  Z = icu.data$open_bin
  #make the Love plot with permutation quantiles
  ## Not run: biasLovePlot(X = X, D = D, Z = Z, permQuantiles = TRUE, perms = 500)

---

Card (1995) Data

Description

Data from the National Longitudinal Survey of Young Men (NLSYM) that was used by Card (1995).

Usage

data(card.data)

Format

A data frame with 3010 observations on the following 35 variables.

id

subject id

nearc2

indicator for whether a subject grew up near a two-year college

nearc4

indicator for whether a subject grew up near a four-year college

educ

subject's years of education

age

subject's age at the time of the survey in 1976

fatheduc

subject's father's years of education

motheduc

subject's mother's years of education

weight

sampling weight

momdad14

indicator for whether subject lived with both mother and father at age 14

sinmom14

indicator for whether subject lived with single mom at age 14

step14

indicator for whehter subject lived with step-parent at age 14

reg661

indicator for whether subject lived in region 1 (New England) in 1966

reg662

indicator for whether subject lived in region 2 (Middle Atlantic) in 1966

reg663

indicator for whether subject lived in region 3 (East North Central) in 1966

reg664

indicator for whether subject lived in region 4 (West North Central) in 1966

reg665

indicator for whether subject lived in region 5 (South Atlantic) in 1966

reg666

indicator for whether subject lived in region 6 (East South Central) in 1966

reg667

indicator for whether subject lived in region 7 (West South Central) in 1966

reg668

indicator for whether subject lived in region 8 (Mountain) in 1966

reg669

indicator for whether subject lived in region 9 (Pacific) in 1966

south66

indicator for whether subject lived in South in 1966

black

indicator for whether subject's race is black

smsa

indicator for whether subject lived in SMSA in 1976

south

indicator for whether subject lived in the South in 1976

smsa66

indicator for whether subject lived in SMSA in 1966

wage

subject's wage in cents per hour in 1976

enroll

indicator for whether subject is enrolled in college in 1976

KWW

subject's score on the Knowledge of the World of Work (KWW) test in 1966

IQ

IQ-type test score collected from the high school of the subject.

married

indicator for whether the subject was married in 1976.

libcrd14

indicator for whether subject had library card at age 14.

exper

subject's years of labor force experience in 1976

lwage

subject's log wage in 1976

expersq

square of subject's years of labor force experience in 1976

region

region in which subject lived in 1976

Source

Card, D. Using Geographic Variation in College Proximity to Estimate the Return to Schooling. In Aspects of Labor Market Behavior: Essays in Honor of John Vanderkamp, eds. L.N. Christophides, E.K. Grant and R. Swidinsky. 201-222. National Longitudinal Survey of Young Men: https://www.nlsinfo.org/investigator/pages/login.jsp

Examples

data(card.data)

---

Conditional Likelihood Ratio Test

Description

CLR computes the conditional likelihood ratio test (Moreira, 2003) for the ivmodel object as well as the associated confidence interval.

Usage

CLR(ivmodel, beta0 = 0, alpha = 0.05)

Arguments

ivmodel	ivmodel object
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is 0
alpha	The significance level for hypothesis testing. Default is 0.05

Details

CLR.test computes the conditional likelihood ratio test for the instrumental variables model in ivmodel object, specifically for the parameter $\beta$. It also computes the $1 -\alpha$ confidence interval associated with it by inverting the test. The test is fully robust to weak instruments (Moreira 2003). We use the approximation suggested in Andrews et al. (2006) to evaluate the p value and the confidence interval.

Value

CLR returns a list containing the following components

test.stat	The value of the test statistic for testing the null hypothesis $H_0: \beta = \beta_0$ in ivmodel
p.value	The p value of the test under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel
ci	A matrix of two columns, each row contains an interval associated with the confidence interval
ci.info	A human-readable string describing the confidence interval

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Andrews, D. W. K., Moreira, M. J., and Stock, J. H. (2006). Optimal two-side invariant similar tests for instrumental variables regression. Econometrica 74, 715-752.
Moreira, M. J. (2003). A conditional likelihood ratio test for structural models. Econometrica 71, 1027-1048.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,c("nearc4","nearc2")]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
card.model2IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
CLR(card.model2IV,alpha=0.01)

---

Coefficients of the Fitted Model in the ivmodel Object

Description

This coef methods returns the point estimation, standard error, test statistic and p value for all specified k-Class estimation from an ivmodel object.

Usage

## S3 method for class 'ivmodel'
coef(object,...)

Arguments

object	ivmodel object.
...	Additional arguments to coef.

Value

A matrix summarizes all the k-Class estimations.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
coef(foo)

---

Exogenous Coefficients of the Fitted Model in the ivmodel Object

Description

This coefOther returns the point estimates, standard errors, test statistics and p values for the exogenous covariates associated with the outcome. It returns a list of matrices where each matrix is one of the k-Class estimates from an ivmodel object.

Usage

coefOther(ivmodel)

Arguments

ivmodel

ivmodel object.

Value

A list of matrices swhere each matrix summarizes the estimated coefficients from one of hte k-Class estimates.

Author(s)

Hyunseung Kang

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
coefOther(foo)

---

Confidence Intervals for the Fitted Model in ivmodel Object

Description

This confint methods returns a matrix of two columns, each row represents a confident interval for different IV approaches, which include k-Class, AR (Anderson and Rubin 1949) and CLR (Moreira 2003) estimations.

Usage

## S3 method for class 'ivmodel'
confint(object,parm,level=NULL,...)

Arguments

object	ivmodel object.
parm	Ignored for our code.
level	The confidence level.
...	Additional argument(s) for methods.

Value

A matrix, each row represents a confidence interval for different IV approaches.

Author(s)

Yag Jiang, Hyunseung Kang, and Dylan Small

References

Andrews, D. W. K., Moreira, M. J., and Stock, J. H. (2006). Optimal two-side invariant similar tests for instrumental variables regression. Econometrica 74, 715-752.
Moreira, M. J. (2003). A conditional likelihood ratio test for structural models. Econometrica 71, 1027-1048.
Fuller, W. (1977). Some properties of a modification of the limited information estimator. Econometrica, 45, 939-953.
Anderson, T.W. and Rubin, H. (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
confint(foo)

---

Plot randomization distributions of the Mahalanobis distance

Description

distributionBalancePlot displays the randomization distribution of the square root of the Mahalanobis distance across the treatment and/or instrument for different assignment mechanisms. This function supports complete randomization (displayed in black), block randomization (displayed in green), and Bernoulli trials for exposure (displayed in red) and instrument (displayed in blue). This function is used to create Figure 4 of Branson and Keele (2020).

Usage

distributionBalancePlot(X, D = NULL, Z = NULL, subclass = NULL,
complete = FALSE, blocked = FALSE, bernoulli = FALSE, perms = 1000)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
D	Indicator vector for a binary treatment (must contain 1 or 0 for each unit).
Z	Indicator vector for a binary instrument (must contain 1 or 0 for each unit).
subclass	Vector of subclasses (one for each unit). Subclasses can be numbers or characters, as long as there is one specified for each unit. Only needed if blocked = TRUE.
complete	If TRUE, displays the randomization distribution of the Mahalanobis distance under complete randomization.
blocked	If TRUE, displays the randomization distribution of the Mahalanobis distance under block randomization. Needs subclass specified.
bernoulli	If TRUE, displays the randomization distribution of the Mahalanobis distance under Bernoulli trials for the treatment and for the instrument.
perms	Number of permutations used to approximate the randomization distributions.

Value

Plot of randomization distributions of the square root of the Mahalanobis distance across the treatment and/or instrument for different assignment mechanisms.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
  data(icu.data)
  #the covariate matrix is
  X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
  #the treatment
  D = icu.data$icu_bed
  #the instrument
  Z = icu.data$open_bin
  #the subclass
  subclass = icu.data$site
  #make distribution plot of sqrt(MD) for
  #complete randomization, block randomization, and bernoulli trials
  #(just uncomment the code below)
  #distributionBalancePlot(X = X, D = D, Z = Z, subclass = subclass,
  #complete = TRUE, blocked = TRUE, bernoulli = TRUE, perms = 500)

---

Extract Model Fitted values in the ivmodel Object

Description

This fitted method returns the fitted values from k-Class estimators inside ivmodel.

Usage

## S3 method for class 'ivmodel'
fitted(object,...)

Arguments

object	ivmodel object.
...	Additional arguments to fitted.

Value

A matrix of fitted values from the k-Class estimations. Specifically, each column of the matrix represents predicted values of the outcome for each individual based on different estimates of the treatment effect from k-Class estimators. By default, one of the columns of the matrix is the predicted outcome when the treatment effect is estimated by ordinary least squares (OLS). Because OLS is generally biased in instrumental variables settings, the predictions will likely be biased. For consistent estimates, the predictions are estimates of E[Y | D,X]. In other words, they marginalize over the unmeasured confounder U and estimate the mean outcomes among all individuals with measured confounders X if they were to be assigned treatment value D. For example, in the Card study, if U represents the income of the study unit's parents which were not measured and X represents experience in years, the value of fitted for E[Y | D = 16, X = 4] is what the average log income among individuals who had 4 years of experience would be if they were assigned 16 years of education.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
fitted(foo)

---

Fuller-k Estimator

Description

Fuller computes the Fuller-k (Fuller 1977) estimate for the ivmodel object.

Usage

Fuller(ivmodel,
	      beta0 = 0, alpha = 0.05, b = 1,
	      manyweakSE = FALSE,heteroSE = FALSE,clusterID=NULL)

Arguments

ivmodel	ivmodel object.
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is 0.
alpha	The significance level for hypothesis testing. Default is 0.05.
b	Positive constant $b$ in Fuller-k estimator. Default is 1.
manyweakSE	Should many weak instrument (and heteroscedastic-robust) asymptotics in Hansen, Hausman and Newey (2008) be used to compute standard errors?
heteroSE	Should heteroscedastic-robust standard errors be used? Default is FALSE.
clusterID	If cluster-robust standard errors are desired, provide a vector of length that's identical to the sample size. For example, if n = 6 and clusterID = c(1,1,1,2,2,2), there would be two clusters where the first cluster is formed by the first three observations and the second cluster is formed by the last three observations. clusterID can be numeric, character, or factor.

Details

Fuller computes the Fuller-k estimate for the instrumental variables model in ivmodel, specifically for the parameter $beta$. The computation uses KClass with the value of $k = k_{LIML} - b/(n - L - p)$. It generates a point estimate, a standard error associated with the point estimate, a test statistic and a p value under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel along with a $1-\alpha$ confidence interval.

Value

Fuller returns a list containing the following components

k	The k value used when computing the Fuller estimate with the k-Class estimator.
point.est	Point estimate of $\beta$.
std.err	Standard error of the estimate.
test.stat	The value of the test statistic for testing the null hypothesis $H_0: \beta = \beta_0$ in ivmodel.
p.value	The p value of the test under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel.
ci	A matrix of one row by two columns specifying the confidence interval associated with the Fuller estimator.

Author(s)

Yang Jiang, Hyunseung Kang, Dylan Small

References

Fuller, W. (1977). Some properties of a modification of the limited information estimator. Econometrica, 45, 939-953.

See Also

See also ivmodel for details on the instrumental variables model. See also KClass for more information about the k-Class estimator.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,c("nearc4","nearc2")]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661",
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667",
		"reg668", "smsa66")
X=card.data[,Xname]
card.model2IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
Fuller(card.model2IV,alpha=0.01)

---

Get Covariate Mean Differences

Description

getCovMeanDiffs returns the covariate mean differences between two groups.

Usage

getCovMeanDiffs(X, indicator)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
indicator	Binary indicator vector (must contain 1 or 0 for each unit). For example, could be a binary treatment or instrument.

Value

Covariate mean differences between two groups.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
	data(icu.data)
	#the covariate matrix is
	X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
	#covariate mean differences across the treatment
	getCovMeanDiffs(X = X, indicator = icu.data$icu_bed)
	#covariate mean differences across the instrument
	getCovMeanDiffs(X = X, indicator = icu.data$open_bin)

---

Get Mahalanobis Distance

Description

getMD returns the Mahalanobis distance between two groups.

Usage

getMD(X, indicator, covX.inv = NULL)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
indicator	Binary indicator vector (must contain 1 or 0 for each unit). For example, could be a binary treatment or instrument.
covX.inv	Inverse of the covariate covariance matrix. Usually this is left as NULL, because getMD() will compute covX.inv for you. However, if getMD() is used many times (e.g., as in a permutation test), it can be computationally efficient to specify covX.inv beforehand.

Value

Mahalanobis distance between two groups.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
	data(icu.data)
	#the covariate matrix is
	X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
	#mahalanobis distance across the treatment
	getMD(X = X, indicator = icu.data$icu_bed)
	#mahalanobis distance across the instrument
	getMD(X = X, indicator = icu.data$open_bin)

---

Get Standardized Covariate Mean Differences

Description

getStandardizedCovMeanDiffs returns the standardized covariate mean differences between two groups.

Usage

getStandardizedCovMeanDiffs(X, indicator)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
indicator	Binary indicator vector (must contain 1 or 0 for each unit). For example, could be a binary treatment or instrument.

Value

Standardized covariate mean differences between two groups.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
	data(icu.data)
	#the covariate matrix is
	X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
	#standardized covariate mean differences across the treatment
	getStandardizedCovMeanDiffs(X = X, indicator = icu.data$icu_bed)
	#standardized covariate mean differences across the instrument
	getStandardizedCovMeanDiffs(X = X, indicator = icu.data$open_bin)

---

Pseudo-data based on Branson and Keele (2020)

Description

Data sampled with replacemenet from the original data from the (SPOT)light study used in Branson and Keele (2020). Also see Keele et al. (2018) for more details about the variables in this dataset.

Usage

data(icu.data)

Format

A data frame with 13011 observations on the following 18 variables.

age

Age of the patient in years.

male

Whether or not the patient is male; 1 if male and 0 otherwise.

sepsis_dx

Whether or not the patient is diagnosed with sepsis; 1 if so and 0 otherwise.

periarrest

Whether or not the patient is diagnosed with peri-arrest; 1 if so and 0 otherwise.

icnarc_score

The Intensive Care National Audit and Research Centre physiological score.

news_score

The National Health Service national early warning score.

sofa_score

The sequential organ failure assessment score.

v_cc1

Indicator for level of care at assessment (Level 0, normal ward care).

v_cc2

Indicator for level of care at assessment (Level 1, normal ward care).

v_cc4

Indicator for level of care at assessment (Level 2, care within a high dependency unit).

v_cc5

Indicator for level of care at assessment (Level 3, ICU care).

v_cc_r1

Indicator for recommended level of care at assessment (Level 0, normal ward care).

v_cc_r2

Indicator for recommended level of care after assessment (Level 1, normal ward care).

v_cc_r4

Indicator for recommended level of care after assessment (Level 2, care within a high dependency unit).

v_cc_r5

Indicator for recommended level of care after assessment (Level 3, ICU care).

open_bin

Binary instrument; 1 if the available number of ICU beds was less than 4, and 0 otherwise.

icu_bed

Binary treatment; 1 if admitted to an ICU bed.

site

ID for the hospital that the patient attended.

References

Keele, L. et al. (2018). Stronger instruments and refined covariate balance in an observational study of the effectiveness of prompt admission to intensive care units. Journal of the Royal Statistical Society: Series A (Statistics in Society).

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

data(icu.data)

---

Diagnostics of instrumental variable analysis

Description

Diagnostics of instrumental variable analysis

Usage

iv.diagnosis(Y, D, Z, X)

iv.diagnosis.plot(output, bias.ratio = TRUE, base_size = 15,
  text_size = 5)

Arguments

Y	A numeric vector of outcomes.
D	A vector of endogenous variables.
Z	A vector of instruments.
X	A vector, matrix or data frame of (exogenous) covariates.
output	Output from iv.diagnosis.
bias.ratio	Add bias ratios (text) to the plot?
base_size	size of the axis labels
text_size	size of the text (bias ratios)

Value

a list or data frame

x.mean1

Mean of X under Z = 1 (reported if Z is binary)

x.mean0

Mean of X under Z = 0 (reported if Z is binary)

coef

OLS coefficient of X ~ Z (reported if Z is not binary)

se

Standard error of OLS coefficient (reported if Z is not binary)

p.val

p-value of the independence of Z and X (Fisher's test if both are binary, logistic regression if Z is binary, linear regression if Z is continuous)

stand.diff

Standardized difference (reported if Z is binary)

bias.ratio

Bias ratio

bias.amplify

Amplification of bias ratio

bias.ols

Bias of OLS

bias.2sls

Bias of two stage least squares)

Functions

• iv.diagnosis.plot: IV diagnostic plot

Author(s)

Qingyuan Zhao

References

• Baiocchi, M., Cheng, J., & Small, D. S. (2014). Instrumental variable methods for causal inference. Statistics in Medicine, 33(13), 2297-2340.

• Jackson, J. W., & Swanson, S. A. (2015). Toward a clearer portrayal of confounding bias in instrumental variable applications. Epidemiology, 26(4), 498.

• Zhao, Q., & Small, D. S. (2018). Graphical diagnosis of confounding bias in instrumental variable analysis. Epidemiology, 29(4), e29–e31.

Examples

n <- 10000
Z <- rbinom(n, 1, 0.5)
X <- data.frame(matrix(c(rnorm(n), rbinom(n * 5, 1, 0.5)), n))
D <- rbinom(n, 1, plogis(Z + X[, 1] + X[, 2] + X[, 3]))
Y <- D + X[, 1] + X[, 2] + rnorm(n)
print(output <- iv.diagnosis(Y, D, Z, X))
iv.diagnosis.plot(output)

Z <- rnorm(n)
D <- rbinom(n, 1, plogis(Z + X[, 1] + X[, 2] + X[, 3]))
Y <- D + X[, 1] + X[, 2] + rnorm(n)
print(output <- iv.diagnosis(Y, D, Z, X)) ## stand.diff is not reported
iv.diagnosis.plot(output)

---

Fitting Instrumental Variables (IV) Models

Description

ivmodel fits an instrumental variables (IV) model with one endogenous variable and a continuous outcome. It carries out several IV regressions, diagnostics, and tests associated this IV model. It is robust to most data formats, including factor and character data, and can handle very large IV models efficiently.

Usage

ivmodel(Y, D, Z, X, intercept = TRUE,
        beta0 = 0, alpha = 0.05, k = c(0, 1),
        manyweakSE = FALSE, heteroSE = FALSE, clusterID = NULL,
        deltarange = NULL, na.action = na.omit)

Arguments

Y	A numeric vector of outcomes.
D	A vector of endogenous variables.
Z	A matrix or data frame of instruments.
X	A matrix or data frame of (exogenous) covariates.
intercept	Should the intercept be included? Default is TRUE and if so, you do not need to add a column of 1s in X.
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is $0$.
alpha	The significance level for hypothesis testing. Default is 0.05.
k	A numeric vector of k values for k-class estimation. Default is 0 (OLS) and 1 (TSLS).
manyweakSE	Should many weak instrument (and heteroscedastic-robust) asymptotics in Hansen, Hausman and Newey (2008) be used to compute standard errors? (Not supported for k ==0)
heteroSE	Should heteroscedastic-robust standard errors be used? Default is FALSE.
clusterID	If cluster-robust standard errors are desired, provide a vector of length that's identical to the sample size. For example, if n = 6 and clusterID = c(1,1,1,2,2,2), there would be two clusters where the first cluster is formed by the first three observations and the second cluster is formed by the last three observations. clusterID can be numeric, character, or factor.
deltarange	Range of $\delta$ for sensitivity analysis with the Anderson-Rubin (1949) test.
na.action	NA handling. There are na.fail, na.omit, na.exclude, na.pass available. Default is na.omit.

Details

Let $Y$, $D$, $X$, and $Z$ represent the outcome, endogenous variable, p dimensional exogenous covariates, and L dimensional instruments, respectively. Note that the intercept is a type of exogenous covariate and can be added to $X$ by specifying intercept as TRUE (the default behavior); the user does not have to manually add an intercept column in $X$. ivmodel assumes the following IV model

$Y = X \alpha + D \beta + \epsilon, E(\epsilon | X, Z) = 0$

and produces statistics for $\beta$. In particular, ivmodel computes the OLS, TSLS, k-class, limited information maximum likelihood (LIML), and Fuller-k (Fuller 1977) estimates of $\beta$ using KClass, LIML, and codeFuller. Also, ivmodel computes confidence intervals and hypothesis tests of the type $H_0: \beta = \beta_0$ versus $H_0: \beta \neq \beta_0$ for the said estimators as well as two weak-IV confidence intervals, Anderson and Rubin (Anderson and Rubin 1949) confidence interval (Anderson and Rubin 1949) and the conditional likelihood ratio confidence interval (Moreira 2003). Finally, the code also conducts a sensitivity analysis if $Z$ is one-dimensional (i.e. there is only one instrument) using the method in Jiang et al. (2015).

Some procedures (e.g. conditional likelihood ratio test, sensitivity analysis with Anderson-Rubin) assume an additional linear model

$D = Z \gamma + X \kappa + \xi, E(\xi | X, Z) = 0$

Value

ivmodel returns an object of class "ivmodel".

An object class "ivmodel" is a list containing the following components

n	Sample size.
L	Number of instruments.
p	Number of exogenous covariates (including intercept).
Y	Outcome, cleaned for use in future methods.
D	Treatment, cleaned for use in future methods.
Z	Instrument(s), cleaned for use in future methods.
X	Exogenous covariates (if provided), cleaned for use in future methods.
Yadj	Adjusted outcome, projecting out X.
Dadj	Adjusted treatment, projecting out X.
Zadj	Adjusted instrument(s), projecting out X.
ZadjQR	QR decomposition for adjusted instrument(s).
ZXQR	QR decomposition for concatenated matrix of Z and X.
alpha	Significance level for the hypothesis tests.
beta0	Null value of the hypothesis tests.
kClass	A list from KClass function.
LIML	A list from LIML function.
Fuller	A list from Fuller function.
AR	A list from AR.test.
CLR	A list from CLR.

In addition, if there is only one instrument, ivreg will generate an "ARsens" list within "ivmodel" object.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T. W. and Rubin, H. (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. Annals of Mathematical Statistics 20, 46-63.

Freeman G., Cowling B. J., Schooling C. M. (2013). Power and Sample Size Calculations for Mendelian Randomization Studies Using One Genetic Instrument. International Journal of Epidemiology 42(4), 1157-1163.

Fuller, W. (1977). Some properties of a modification of the limited information estimator. Econometrica, 45, 939-953.

Hansen, C., Hausman, J., and Newey, W. (2008) Estimation with many instrumental variables. Journal of Business and Economic Statistics 26(4), 398-422.

Moreira, M. J. (2003). A conditional likelihood ratio test for structural models. Econometrica 71, 1027-1048.

Sargan, J. D. (1958). The estimation of economic relationships using instrumental variables. Econometrica , 393-415.

Wang, X., Jiang, Y., Small, D. and Zhang, N. (2017), Sensitivity analysis and power for instrumental variable studies. Biometrics 74(4), 1150-1160.

See Also

See also KClass, LIML, Fuller, AR.test, and CLR for individual methods associated with ivmodel. For extracting the estimated effect of the exogenous covariates on the outcome, see coefOther. For sensitivity analysis with the AR test, see ARsens.test. ivmodel has vcov.ivmodel,model.matrix.ivmodel,summary.ivmodel, confint.ivmodel, fitted.ivmodel, residuals.ivmodel and coef.ivmodel methods associated with it.

Examples

data(card.data)
# One instrument #
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661",
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667",
		"reg668", "smsa66")
X=card.data[,Xname]
card.model1IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
card.model1IV

# Multiple instruments
Z = card.data[,c("nearc4","nearc2")]
card.model2IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
card.model2IV

---

Fitting Instrumental Variables (IV) Models

Description

ivmodelFormula fits an instrumental variables (IV) model with one endogenous variable and a continuous outcome. It carries out several IV regressions, diagnostics, and tests associated this IV model. It is robust to most data formats, including factor and character data, and can handle very large IV models efficiently.

Usage

ivmodelFormula(formula, data, subset,
        beta0=0,alpha=0.05,k=c(0,1), 
        manyweakSE = FALSE,
        heteroSE = FALSE, clusterID = NULL, 
        deltarange=NULL, na.action = na.omit)

Arguments

formula	a formula describing the model to be fitted. For example, the formula Y ~ D + X1 + X2 | Z1 + Z2 + X1 + X2 describes the mode where $Y = \alpha_0 + D \beta + X_{1} \alpha_1 + X_{2} \alpha_2 + \epsilon$ and $D = \gamma_0 + Z_{1} \gamma_1 + Z_{2} \gamma2 + X_{1} \kappa_1 + X_{2} \kappa_2 + \xi$ The outcome is Y, the endogenous variable is D. The exogenous covariates are X1 and X2. The instruments are Z1 and Z2. The formula environment follows the formula environment in the ivreg function in the AER package.
data	an optional data frame containing the variables in the model. By default the variables are taken from the environment which ivmodel is called from
subset	an index vector indicating which rows should be used.
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is $0$.
alpha	The significance level for hypothesis testing. Default is 0.05.
k	A numeric vector of k values for k-class estimation. Default is 0 (OLS) and 1 (TSLS).
manyweakSE	Should many weak instrument (and heteroscedastic-robust) asymptotics in Hansen, Hausman and Newey (2008) be used to compute standard errors? (Not supported for k ==0)
heteroSE	Should heteroscedastic-robust standard errors be used? Default is FALSE.
clusterID	If cluster-robust standard errors are desired, provide a vector of length that's identical to the sample size. For example, if n = 6 and clusterID = c(1,1,1,2,2,2), there would be two clusters where the first cluster is formed by the first three observations and the second cluster is formed by the last three observations. clusterID can be numeric, character, or factor.
deltarange	Range of $\delta$ for sensitivity analysis with the Anderson-Rubin (1949) test.
na.action	NA handling. There are na.fail, na.omit, na.exclude, na.pass available. Default is na.omit.

Details

Let $Y$, $D$, $X$, and $Z$ represent the outcome, endogenous variable, p dimensional exogenous covariates, and L dimensional instruments, respectively. ivmodel assumes the following IV model

$Y = X \alpha + D \beta + \epsilon, E(\epsilon | X, Z) = 0$

and produces statistics for $\beta$. In particular, ivmodel computes the OLS, TSLS, k-class, limited information maximum likelihood (LIML), and Fuller-k (Fuller 1977) estimates of $\beta$ using KClass, LIML, and codeFuller. Also, ivmodel computes confidence intervals and hypothesis tests of the type $H_0: \beta = \beta_0$ versus $H_0: \beta \neq \beta_0$ for the said estimators as well as two weak-IV confidence intervals, Anderson and Rubin (Anderson and Rubin 1949) confidence interval (Anderson and Rubin 1949) and the conditional likelihood ratio confidence interval (Moreira 2003). Finally, the code also conducts a sensitivity analysis if $Z$ is one-dimensional (i.e. there is only one instrument) using the method in Jiang et al. (2015).

Some procedures (e.g. conditional likelihood ratio test, sensitivity analysis with Anderson-Rubin) assume an additional linear model

$D = Z \gamma + X \kappa + \xi, E(\xi | X, Z) = 0$

Value

ivmodel returns an object of class "ivmodel".

An object class "ivmodel" is a list containing the following components

n	Sample size.
L	Number of instruments.
p	Number of exogenous covariates (including intercept).
Y	Outcome, cleaned for use in future methods.
D	Treatment, cleaned for use in future methods.
Z	Instrument(s), cleaned for use in future methods.
X	Exogenous covariates (if provided), cleaned for use in future methods.
Yadj	Adjusted outcome, projecting out X.
Dadj	Adjusted treatment, projecting out X.
Zadj	Adjusted instrument(s), projecting out X.
ZadjQR	QR decomposition for adjusted instrument(s).
ZXQR	QR decomposition for concatenated matrix of Z and X.
alpha	Significance level for the hypothesis tests.
beta0	Null value of the hypothesis tests.
kClass	A list from KClass function.
LIML	A list from LIML function.
Fuller	A list from Fuller function.
AR	A list from AR.test.
CLR	A list from CLR.

In addition, if there is only one instrument, ivreg will generate an "ARsens" list within "ivmodel" object.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Anderson, T. W. and Rubin, H. (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. Annals of Mathematical Statistics 20, 46-63.

Freeman G., Cowling B. J., Schooling C. M. (2013). Power and Sample Size Calculations for Mendelian Randomization Studies Using One Genetic Instrument. International Journal of Epidemiology 42(4), 1157-1163.

Fuller, W. (1977). Some properties of a modification of the limited information estimator. Econometrica, 45, 939-953.

Hansen, C., Hausman, J., and Newey, W. (2008) Estimation with many instrumental variables. Journal of Business and Economic Statistics 26(4), 398-422.

Moreira, M. J. (2003). A conditional likelihood ratio test for structural models. Econometrica 71, 1027-1048.

Sargan, J. D. (1958). The estimation of economic relationships using instrumental variables. Econometrica , 393-415.

Wang, X., Jiang, Y., Small, D. and Zhang, N. (2017), Sensitivity analysis and power for instrumental variable studies. Biometrics 74(4), 1150-1160.

See Also

See also KClass, LIML, Fuller, AR.test, and CLR for individual methods associated with ivmodel. For extracting the estimated effect of the exogenous covariates on the outcome, see coefOther. For sensitivity analysis with the AR test, see ARsens.test. ivmodel has vcov.ivmodel,model.matrix.ivmodel,summary.ivmodel, confint.ivmodel, fitted.ivmodel, residuals.ivmodel and coef.ivmodel methods associated with it.

Examples

data(card.data)
# One instrument #
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
card.model1IV = ivmodelFormula(lwage ~ educ + exper + expersq + black + 
                                south + smsa + reg661 + 
                                reg662 + reg663 + reg664 + 
                                reg665 + reg666 + reg667 + 
                                reg668 + smsa66 | nearc4 + 
                                exper + expersq + black + 
                                south + smsa + reg661 + 
                                reg662 + reg663 + reg664 + 
                                reg665 + reg666 + reg667 + 
                                reg668 + smsa66,data=card.data)
card.model1IV

# Multiple instruments
Z = card.data[,c("nearc4","nearc2")]
card.model2IV = ivmodelFormula(lwage ~ educ + exper + expersq + black + 
                                south + smsa + reg661 + 
                                reg662 + reg663 + reg664 + 
                                reg665 + reg666 + reg667 + 
                                reg668 + smsa66 | nearc4 + nearc2 +
                                exper + expersq + black + 
                                south + smsa + reg661 + 
                                reg662 + reg663 + reg664 + 
                                reg665 + reg666 + reg667 + 
                                reg668 + smsa66,data=card.data)
card.model2IV

---

Power calculation for IV models

Description

IVpower computes the power for one of the following tests: two stage least square estimates; Anderson-Rubin (1949) test; Sensitivity analysis.

Usage

IVpower(ivmodel, n = NULL, alpha = 0.05, beta = NULL, type = "TSLS", 
        deltarange = NULL, delta = NULL)

Arguments

ivmodel	ivmodel object.
n	number of sample size, if missing, will use the sample size from the input ivmodel object.
alpha	The significance level for hypothesis testing. Default is 0.05.
beta	True causal effect minus null hypothesis causal effect. If missing, will use the beta calculated from the input ivmodel object.
type	Determines which test will be used for power calculation. "TSLS" for two stage least square estimates; "AR" for Anderson-Rubin test; "ARsens" for sensitivity analysis.
deltarange	Range of sensitivity allowance. A numeric vector of length 2. If missing, will use the deltarange from the input ivmodel object.
delta	True value of sensitivity parameter when calculating the power. Usually take delta = 0 for the favorable situation or delta = NULL for unknown delta.

Details

IVpower computes the power for one of the following tests: two stage least square estimates; Anderson-Rubin (1949) test; Sensitivity analysis. The related value of parameters will be inferred from the input of ivmodel object.

Value

a power value for the specified type of test.

Author(s)

Yang Jiang, Hyunseung Kang, Dylan Small

References

Freeman G, Cowling BJ, Schooling CM (2013). Power and Sample Size Calculations for Mendelian Randomization Studies Using One Genetic Instrument. International journal of epidemiology, 42(4), 1157-1163.
Anderson, T.W. and Rubin, H. (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. Annals of Mathematical Statistics, 20, 46-63.
ang, X., Jiang, Y., Small, D. and Zhang, N (2017), Sensitivity analysis and power for instrumental variable studies, (under review of Biometrics).

See Also

See also ivmodel for details on the instrumental variables model. See also TSLS.power, AR.power, ARsens.power for details on the power calculation.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
card.model = ivmodel(Y=Y,D=D,Z=Z,X=X)

IVpower(card.model)
IVpower(card.model, n=10^4, type="AR")

---

Calculating minimum sample size for achieving a certain power

Description

IVsize calculates the minimum sample size needed for achieving a certain power in one of the following tests: two stage least square estimates; Anderson-Rubin (1949) test; Sensitivity analysis.

Usage

IVsize(ivmodel, power, alpha = 0.05, beta = NULL, type = "TSLS", 
       deltarange = NULL, delta = NULL)

Arguments

ivmodel	ivmodel object.
power	The power threshold to achieve.
alpha	The significance level for hypothesis testing. Default is 0.05.
beta	True causal effect minus null hypothesis causal effect. If missing, will use the beta calculated from the input ivmodel object.
type	Determines which test will be used for power calculation. "TSLS" for two stage least square estimates; "AR" for Anderson-Rubin test; "ARsens" for sensitivity analysis.
deltarange	Range of sensitivity allowance. A numeric vector of length 2. If missing, will use the deltarange from the input ivmodel object.
delta	True value of sensitivity parameter when calculating the power. Usually take delta = 0 for the favorable situation or delta = NULL for unknown delta.

Details

IVsize calculates the minimum sample size needed for achieving a certain power for one of the following tests: two stage least square estimates; Anderson-Rubin (1949) test; Sensitivity analysis. The related value of parameters will be inferred from the input of ivmodel object.

Value

minimum sample size needed for achieving a certain power

Author(s)

Yang Jiang, Hyunseung Kang, Dylan Small

References

Freeman G, Cowling BJ, Schooling CM (2013). Power and Sample Size Calculations for Mendelian Randomization Studies Using One Genetic Instrument. International journal of epidemiology, 42(4), 1157-1163.
Anderson, T.W. and Rubin, H. (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. Annals of Mathematical Statistics, 20, 46-63.
ang, X., Jiang, Y., Small, D. and Zhang, N (2017), Sensitivity analysis and power for instrumental variable studies, (under review of Biometrics).

See Also

See also ivmodel for details on the instrumental variables model. See also TSLS.size, AR.size, ARsens.size for calculation details.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
card.model = ivmodel(Y=Y,D=D,Z=Z,X=X, deltarange=c(-0.01, 0.01))

IVsize(card.model, power=0.8)
IVsize(card.model, power=0.8, type="AR")
IVsize(card.model, power=0.8, type="ARsens", deltarange=c(-0.01, 0.01))

---

k-Class Estimator

Description

KClass computes the k-Class estimate for the ivmodel object.

Usage

KClass(ivmodel,
       beta0 = 0, alpha = 0.05, k = c(0, 1),
       manyweakSE = FALSE, heteroSE = FALSE,clusterID = NULL)

Arguments

ivmodel	ivmodel object.
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is 0.
alpha	The significance level for hypothesis testing. Default is 0.05.
k	A vector of $k$ values for the k-Class estimator. Default is 0 (OLS) and 1 (TSLS).
manyweakSE	Should many weak instrument (and heteroscedastic-robust) asymptotics in Hansen, Hausman and Newey (2008) be used to compute standard errors? (Not supported for k=0)
heteroSE	Should heteroscedastic-robust standard errors be used? Default is FALSE.
clusterID	If cluster-robust standard errors are desired, provide a vector of length that's identical to the sample size. For example, if n = 6 and clusterID = c(1,1,1,2,2,2), there would be two clusters where the first cluster is formed by the first three observations and the second cluster is formed by the last three observations. clusterID can be numeric, character, or factor.

Details

KClass computes the k-Class estimate for the instrumental variables model in ivmodel, specifically for the parameter $\beta$. It generates a point estimate, a standard error associated with the point estimate, a test statistic and a p value under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel along with a $1-\alpha$ confidence interval.

Value

KClass returns a list containing the following components

k	A row matrix of k values supplied to KClass.
point.est	A row matrix of point estimates of $\beta$, with each row corresponding to the k values supplied.
std.err	A row matrix of standard errors of the estimates, with each row corresponding to the k values supplied.
test.stat	A row matrix of test statistics for testing the null hypothesis $H_0: \beta = \beta_0$ in ivmodel, with each row corresponding to the k values supplied.
p.value	A row matrix of p value of the test under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel, with each row corresponding to the k values supplied.
ci	A matrix of two columns specifying the confidence interval, with each row corresponding to the k values supplied.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,c("nearc4","nearc2")]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661",
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667",
		"reg668", "smsa66")
X=card.data[,Xname]
card.model2IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
KClass(card.model2IV,
          k=c(0,1,0.5))

## Not run: 
## The following code tests the mank weak IV standard error for LIML and Fuller.
example <- function(q = 10, rho1 = 0.5, n1 = 10000,
sigma.uv = 0.5, beta = 1, gamma = rep(1/sqrt(q), q)) {

    Sigma1 <- outer(1:q, 1:q, function(i, j) rho1^abs(i - j))

    library(MASS)
    Z1 <- mvrnorm(n1, rep(1, q), Sigma1)
    Z1 <- matrix(2 * as.numeric(Z1 > 0) - 1, nrow = n1)
    UV1 <- mvrnorm(n1, rep(0, 2), matrix(c(1, sigma.uv, sigma.uv, 1), 2))
    X1 <- Z1 
    Y1 <- X1 

    list(Z1 = Z1, X1 = X1, Y1 = Y1)

}

one.sim <- function(manyweakSE) {
    data <- example(q = 100, n1 = 200)
    fit <- ivmodel(data$Y1, data$X1, data$Z1, manyweakSE = manyweakSE)
    1 > coef(fit)[, 2] - 1.96 * coef(fit)[, 3] & 1 < coef(fit)[, 2] + 1.96 * coef(fit)[, 3]
}

res <- replicate(200, one.sim(TRUE))
apply(res, 1, mean)

res <- replicate(200, one.sim(FALSE))
apply(res, 1, mean)

## End(Not run)

---

Limited Information Maximum Likelihood Ratio (LIML) Estimator

Description

LIML computes the LIML estimate for the ivmodel object.

Usage

LIML(ivmodel,
     beta0 = 0, alpha = 0.05,
     manyweakSE = FALSE, heteroSE = FALSE,clusterID = NULL)

Arguments

ivmodel	ivmodel object.
beta0	Null value $\beta_0$ for testing null hypothesis $H_0: \beta = \beta_0$ in ivmodel. Default is 0.
alpha	The significance level for hypothesis testing. Default is 0.05.
manyweakSE	Should many weak instrument (and heteroscedastic-robust) asymptotics in Hansen, Hausman and Newey (2008) be used to compute standard errors?
heteroSE	Should heteroscedastic-robust standard errors be used? Default is FALSE.
clusterID	If cluster-robust standard errors are desired, provide a vector of length that's identical to the sample size. For example, if n = 6 and clusterID = c(1,1,1,2,2,2), there would be two clusters where the first cluster is formed by the first three observations and the second cluster is formed by the last three observations. clusterID can be numeric, character, or factor.

Details

LIML computes the LIML estimate for the instrumental variables model in ivmodel, specifically for the parameter $beta$. The computation uses KClass with the value of $k = k_{LIML}$, which is the smallest root of the equation

$det(L^T L - k L^T R_Z L) = 0$

where $L$ is a matrix of two columns, the first column consisting of the outcome vector, $Y$, and the second column consisting of the endogenous variable, $D$, and $R_Z = I - Z (Z^T Z)^{-1} Z^T$ with $Z$ being the matrix of instruments. LIML generates a point estimate, a standard error associated with the point estimate, a test statistic and a p value under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel along with a $1-\alpha$ confidence interval.

Value

LIML returns a list containing the following components

k	The k value for LIML.
point.est	Point estimate of $\beta$.
std.err	Standard error of the estimate.
test.stat	The value of the test statistic for testing the null hypothesis $H_0: \beta = \beta_0$ in ivmodel.
p.value	The p value of the test under the null hypothesis $H_0: \beta = \beta_0$ in ivmodel.
ci	A matrix of one row by two columns specifying the confidence interval associated with the Fuller estimator.

Author(s)

Yang Jiang, Hyunseung Kang, Dylan Small

See Also

See also ivmodel for details on the instrumental variables model. See also KClass for more information about the k-Class estimator.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,c("nearc4","nearc2")]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661",
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667",
		"reg668", "smsa66")
X=card.data[,Xname]
card.model2IV = ivmodel(Y=Y,D=D,Z=Z,X=X)
LIML(card.model2IV,alpha=0.01)

---

Extract Design Matrix for ivmodel Object

Description

This method extracts the design matrix inside ivmodel.

Usage

## S3 method for class 'ivmodel'
model.matrix(object,...)

Arguments

object	ivmodel object.
...	Additional arguments to fitted.

Value

A design matrix for the ivmodel object.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
model.matrix(foo)

---

Parameter Estimation from Ivmodel

Description

para computes the estimation of several parameters for the ivmodel object.

Usage

para(ivmodel)

Arguments

ivmodel

ivmodel object.

Details

para computes the coefficients of 1st and 2nd stage regression (gamma and beta). It also computes the covariance matrix of the error term of 1st and 2nd stage. (sigmau, sigmav, and rho)

Value

para returns a list containing the following components

gamma	The coefficient of IV in first stage, calculated by linear regression
beta	The TSLS estimator of the exposure effect
sigmau	Standard deviation of potential outcome under control (structural error for y).
sigmav	Standard deviation of error from regressing treatment on instruments
rho	Correlation between u (potential outcome under control) and v (error from regressing treatment on instrument).

Author(s)

Yang Jiang, Hyunseung Kang, Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
cardfit=ivmodel(Y=Y, D=D, Z=Z, X=X)
para(cardfit)

---

Perform a permutation test using the sum of absolute biases

Description

permTest.absBias performs a permutation test for complete randomization using the sum of absolute biases as a test statistic.

Usage

permTest.absBias(X, D = NULL, Z = NULL,
assignment = "complete",
perms = 1000, subclass = NULL)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
D	Indicator vector for a binary treatment (must contain 1 or 0 for each unit).
Z	Indicator vector for a binary instrument (must contain 1 or 0 for each unit).
assignment	Must be "complete", "block", or "bernoulli". Designates whether to test for complete randomization, block randomization, or Bernoulli trials.
subclass	Vector of subclasses (one for each unit). Subclasses can be numbers or characters, as long as there is one specified for each unit. Only needed if assignment = "block".
perms	Number of permutations used to approximate the permutation test.

Value

p-value testing whether or not an indicator (treatment or instrument) is as-if randomized under complete randomization (i.e., random permutations), block randomization (i.e., random permutations within subclasses), or Bernoulli trials.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
	data(icu.data)
	#the covariate matrix is
	X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
	#the treatment
	D = icu.data$icu_bed
	#the instrument
	Z = icu.data$open_bin
	#the subclass
	subclass = icu.data$site
	
	#can uncomment the following code for examples
	
	#permutation test for complete randomization (for the treatment)
	#permTest.absBias(X = X, D = D,
	#assignment = "complete", perms = 500)
	#permutation test for complete randomization (for the instrument)
	#permTest.absBias(X = X, D = D, Z = Z,
	#assignment = "complete", perms = 500)
	#permutation test for block randomization (for the treatment)
	#permTest.absBias(X = X, D = D,
	#assignment = "block", subclass = subclass, perms = 500)
	#permutation test for block randomization (for the instrument)
	#permTest.absBias(X = X, D = D, Z = Z,
	#assignment = "block",
	#subclass = subclass, perms = 500)
	#permutation test for bernoulli trials (for the treatment)
	#permTest.absBias(X = X, D = D,
	#assignment = "bernoulli", perms = 500)
	#permutation test for bernoulli randomization (for the instrument)
	#permTest.absBias(X = X, D = D, Z = Z,
	#assignment = "bernoulli", perms = 500)

---

Perform a permutation test using the Mahalanobis distance

Description

permTest.md performs a permutation test for complete randomization using the Mahalanobis distance as a test statistic.

Usage

permTest.md(X, indicator, assignment = "complete", perms = 1000, subclass = NULL)

Arguments

X	Covariate matrix (with units as rows and covariates as columns).
indicator	Binary indicator vector (must contain 1 or 0 for each unit). For example, could be a binary treatment or instrument.
assignment	Must be "complete", "block", or "bernoulli". Designates whether to test for complete randomization, block randomization, or Bernoulli trials.
subclass	Vector of subclasses (one for each unit). Subclasses can be numbers or characters, as long as there is one specified for each unit. Only needed if assignment = "block".
perms	Number of permutations used to approximate the permutation test.

Value

p-value testing whether or not an indicator (treatment or instrument) is as-if randomized under complete randomization (i.e., random permutations), block randomization (i.e., random permutations within subclasses), or Bernoulli trials.

Author(s)

Zach Branson and Luke Keele

References

Branson, Z. and Keele, L. (2020). Evaluating a Key Instrumental Variable Assumption Using Randomization Tests. American Journal of Epidemiology. To appear.

Examples

#load the data
	data(icu.data)
	#the covariate matrix is
	X = as.matrix(subset(icu.data, select = -c(open_bin, icu_bed)))
	#the treatment
	D = icu.data$icu_bed
	#the instrument
	Z = icu.data$open_bin
	#the subclass
	subclass = icu.data$site
	
	#can uncomment the following code for examples

	#permutation test for complete randomization (for the treatment)
	#permTest.md(X = X, indicator = D,
	#assignment = "complete", perms = 500)
	#permutation test for complete randomization (for the instrument)
	#permTest.md(X = X, indicator = Z,
	#assignment = "complete", perms = 500)
	#permutation test for block randomization (for the treatment)
	#permTest.md(X = X, indicator = D,
	#assignment = "block", subclass = subclass, perms = 500)
	#permutation test for block randomization (for the instrument)
	#permTest.md(X = X, indicator = Z,
	#assignment = "block", subclass = subclass, perms = 500)
	#permutation test for bernoulli trials (for the treatment)
	#permTest.md(X = X, indicator = D,
	#assignment = "bernoulli", perms = 500)
	#permutation test for bernoulli randomization (for the instrument)
	#permTest.md(X = X, indicator = Z,
	#assignment = "bernoulli", perms = 500)

---

Residuals from the Fitted Model in the ivmodel Object

Description

This function returns the residuals from the k-Class estimators inside the ivmodel object.

Usage

## S3 method for class 'ivmodel'
residuals(object,...)
## S3 method for class 'ivmodel'
resid(object,...)

Arguments

object	ivmodel object.
...	Additional arguments to residuals or resid.

Value

A matrix of residuals for each k-Class estimator. Specifically, each column of the matrix represents residuals for each individual based on different estimates of the treatment effect from k-Class estimators. By default, one of the columns of the matrix is the residuals when the treatment effect is estimated by ordinarly least squares (OLS). Because OLS is generally biased in instrumental variables settings, the residuals will likely be biased.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
resid(foo)
residuals(foo)

---

Power of TSLS Estimator

Description

TSLS.power computes the power of the asymptotic t-test of TSLS estimator.

Usage

TSLS.power(n, beta, rho_ZD, sigmau, sigmaDsq, alpha = 0.05)

Arguments

n	Sample size.
beta	True causal effect minus null hypothesis causal effect.
rho_ZD	Correlation between the IV Z and the exposure D.
sigmau	Standard deviation of potential outcome under control. (structural error for y)
sigmaDsq	The variance of the exposure D.
alpha	Significance level.

Details

The power formula is given in Freeman (2013).

Value

Power of the asymptotic t-test of TSLS estimator basd on given values of parameters.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Freeman G, Cowling BJ, Schooling CM (2013). Power and Sample Size Calculations for Mendelian Randomization Studies Using One Genetic Instrument. International journal of epidemiology, 42(4), 1157-1163.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

# Assume we calculate the power of asymptotic t-test of TSLS estimator
# in a study with one IV (l=1) and the only one exogenous variable is 
# the intercept (k=1). 

# Suppose the difference between the null hypothesis and true causal
# effect is 1 (beta=1).
# The sample size is 250 (n=250). 
# The correlation between the IV and exposure is .5 (rho_ZD= .5).
# The standard deviation of potential outcome is 1(sigmau= 1). 
# The variance of the exposure is 1 (sigmaDsq=1).
# The significance level for the study is alpha = .05.

# power of asymptotic t-test of TSLS estimator
TSLS.power(n=250, beta=1, rho_ZD=.5, sigmau=1, sigmaDsq=1, alpha = 0.05)

---

Sample Size Calculator for the Power of Asymptotic T-test

Description

TSLS.size computes the minimum sample size required for achieving certain power of asymptotic t-test of TSLS estimator.

Usage

TSLS.size(power, beta, rho_ZD, sigmau, sigmaDsq, alpha = 0.05)

Arguments

power	The desired power over a constant.
beta	True causal effect minus null hypothesis causal effect.
rho_ZD	Correlation between the IV Z and the exposure D.
sigmau	Standard deviation of potential outcome under control. (structural error for y)
sigmaDsq	The variance of the exposure D.
alpha	Significance level.

Details

The calculation is based on inverting the power formula given in Freeman (2013).

Value

Minimum sample size required for achieving certain power of asymptotic t-test of TSLS estimator.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

References

Freeman G, Cowling BJ, Schooling CM (2013). Power and Sample Size Calculations for Mendelian Randomization Studies Using One Genetic Instrument. International journal of epidemiology, 42(4), 1157-1163.

See Also

See also ivmodel for details on the instrumental variables model.

Examples

# Assume we performed an asymptotic t-test of TSLS estimator in a study 
# with one IV (l=1) and the only one exogenous variable is the intercept 
# (k=1). We want to calculate the minimum sample size needed for this 
# test to have an at least 0.8 power.

# Suppose the null hypothesis causal effect is 0 and the true causal 
# effect is 1 (beta=1-0=1).
# The correlation between the IV and exposure is .5 (rho_ZD= .5).
# The standard deviation of potential outcome is 1(sigmau= 1). 
# The variance of the exposure is 1 (sigmaDsq=1).
# The significance level for the study is alpha = .05.

### minimum sample size required for aysmptotic t-test
TSLS.size(power=.8, beta=1, rho_ZD=.5, sigmau=1, sigmaDsq=1, alpha =.05)

---

Calculate Variance-Covariance Matrix (i.e. Standard Error) for k-Class Estimators in the ivmodel Object

Description

This vcov method returns the variance-covariance matrix for all specified k-Class estimation from an ivmodel object.

Usage

## S3 method for class 'ivmodel'
vcov(object,...)

Arguments

object	ivmodel object.
...	Additional arguments to vcov.

Value

A matrix of standard error estimates for each k-Class estimator.

Author(s)

Yang Jiang, Hyunseung Kang, and Dylan Small

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
Y=card.data[,"lwage"]
D=card.data[,"educ"]
Z=card.data[,"nearc4"]
Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
        "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
		"reg668", "smsa66")
X=card.data[,Xname]
foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
vcov(foo)

---

Variance of Exogenous Coefficients of the Fitted Model in the ivmodel Object

Description

This vcovOther returns the estimated variances of the estimated coefficients for the exogenous covariates associated with the outcome. All the estimation is based on k-Class estimators.

Usage

vcovOther(ivmodel)

Arguments

ivmodel

ivmodel object.

Value

A matrix where each row represents a k-class estimator and each column represents one of the exogenous covariates. Each element is the estimated variance of the estimated coefficients.

Author(s)

Hyunseung Kang

See Also

See also ivmodel for details on the instrumental variables model.

Examples

data(card.data)
  Y=card.data[,"lwage"]
  D=card.data[,"educ"]
  Z=card.data[,"nearc4"]
  Xname=c("exper", "expersq", "black", "south", "smsa", "reg661", 
          "reg662", "reg663", "reg664", "reg665", "reg666", "reg667", 
          "reg668", "smsa66")
  X=card.data[,Xname]
  foo = ivmodel(Y=Y,D=D,Z=Z,X=X)
  vcovOther(foo)

---