tmle {tmleLite} | R Documentation |
Carries out a simplified version of targeted maximum likelihood estimation of marginal additive treatment effect of a binary point treatment on an outcome. This parameter is defined as E(E(Y|A=1,W) - E(Y|A=0,W)), where Y
is a continuous or binary outcome variable, A
is a binary treatment variable, (A=1 for treatment, A=0 for control), and W
is a matrix or dataframe of baseline covariates. The tmle
function is minimally called with arguments (Y,A,W)
. Missingness in the outcome is accounted for in the estimation procedure, if optional missingness indicator Delta
is supplied.
tmle(Y, A, W, Delta = rep(1, length(Y)), id = 1:length(Y), Q = NULL, g_A = NULL, g_M = NULL, wts = rep(1, length(Y)), DSAargs = NULL, family = "gaussian", epsilon = NULL)
Y |
continuous or binary outcome variable |
A |
binary treatment indicator, 1 - treatment, 0 - control |
W |
vector, matrix, or dataframe containing baseline covariates |
Delta |
indicator of missing outcome or treatment assignment. 1 - observed, 0 - missing |
id |
id identifying repeated measures |
Q |
E(Y|A,W), the Q portion of the likelihood, specified in one of three ways: NULL specifies DSA estimation of E(Y|A=a, W), with A forced into the model (default). matrix of values, one row per observation, three columns: E(Y|A=a,W), E(Y|A=1,W), E(Y|A=0,W). formula for estimation of E(Y|A, W), suitable for call to glm |
g_A |
binary treatment mechanism, specified in one of three ways: NULL defaults to DSA estimation of P(A=1|W) vector of values P(A=1|W), one entry per observation formula for estimation of P(A=1,W), suitable for call to glm |
g_M |
missingness mechanism, specified in one of three ways: NULL defaults to DSA estimation of P(Delta=1|W) vector of values P(Delta=1|W), one entry per observation formula for estimation of P(Delta=1,W), suitable for call to glm |
wts |
optional weights on observations. Defaults to unweighted |
DSAargs |
optional settings for DSA estimation. See DSA help files for further information. Default settings are:
maxsumofpow = 2, maxorderint = 2, maxsize=15, vfold = 5,
nsplits=1,
Dmove=FALSE,
Smove=FALSE. |
family |
family specification for working regression models, generally ‘gaussian’ for continuous outcomes (default), ‘binomial’ for binary outcomes. |
epsilon |
advanced option. Ordinarily this argument should not be specified, but can optionally be set to 0 to circumvent the targeting step. |
psi |
additive treatment effect estimate |
var |
variance of estimate, based on the influence curve |
pvalue |
two-sided p-value |
CI |
95% confidence interval |
epsilon |
MLE for coefficient used in targeting step |
Q |
initial estimate of Q portion of the likelihood. Q$coefficients are the coefficients for the model for Q selected by DSA or specified by the user. Q$Q is an nx3 matrix, where n is the number of observations. Columns contain targeted predicted values for Q(A,W),Q(1,W),Q(0,W) , respectively. Q$type is method for estimating Q , NULL - user supplied, ‘DSA’, or ‘glm’ |
g_A |
treatment mechanism estimate. A list with two items: g_A$coefficients the coefficients for the model for g_A selected by DSA or specified by the user. g_A$g1W contains values of P(A=1|W) for each observation |
g_M |
missingness mechanism estimate. A list with two items: g_M$coefficients the coefficients for the model for g_M selected by DSA or specified by the user. g_M$g1W contains values of P(Delta=1|A,W) for each observation |
Qcounterfactual |
targeted estimate of counterfactual outcomes Q(1,W),Q(0,W) |
Susan Gruber sgruber@berkeley.edu, in collaboration with Mark van der Laan.
1. Mark J. van der Laan and Daniel Rubin (2006), “Targeted Maximum Likelihood Learning”. The International Journal of Biostatistics, 2(1). http://www.bepress.com/ijb/vol2/iss1/11/
2. Susan Gruber and Mark J. van der Laan (2009), “Targeted Maximum Likelihood Estimation: A Gentle Introduction”. U.C. Berkeley Division of Biostatistics Working Paper Series. Working Paper 252. http://www.bepress.com/ucbbiostat/paper252
3. Mark J. van der Laan, Sherri Rose, Susan Gruber editors, (2009) “Readings in Targeted Maximum Likelihood Estimation” . U.C. Berkeley Division of Biostatistics Working Paper Series. Working Paper xxx. http://www.bepress.com/ucbbiostat
4. Sandra E. Sinisi and Mark J. van der Laan, (2004). “Loss-Based Cross-Validated Deletion/Substitution/Addition Algorithms in Estimation”. U.C. Berkeley Division of Biostatistics Working Paper Series. Working Paper 143. http://www.bepress.com/ucbbiostat/paper143
summary.tmle
,
estimate_Q
,
estimate_g
,
DSA
library(tmleLite) # generate data n <- 500 W <- matrix(rnorm(n*3), ncol=3) A <- rbinom(n,1, 1/(1+exp(-(.1*W[,1] - .1*W[,2] + .5*W[,3])))) Y <- A + 2*W[,1] + W[,3] + W[,2]^2 + rnorm(n) colnames(W) <- paste("W",1:3, sep="") # Example 1. Simplest function invocation # DSA called to estimate Q, g_A. # Because Delta is not supplied, all outcomes are known to be measured. # result1 <- tmle(Y,A,W) summary(result1) # Example 2: Binary outcome # DSA called to estimate Q # known treatment mechanism, g_A, is user supplied # A.ex2 <- rbinom(n,1,.5) Y.ex2 <- A.ex2 + 2*W[,1] + W[,3] + W[,2]^2 + rnorm(n) result2 <- tmle(Y=Y.ex2,A=A.ex2,W, g_A =rep(.5, length(Y))) summary(result2) # Example 3: # User-supplied (misspecified) model for Q, # DSA called to estimate g_A, g_M # approx. 20 Delta <- rbinom(n, 1, 1/(1+exp(-(1.7-1*W[,1])))) result3 <- tmle(Y,A,W, Delta=Delta, Q=Y~A+W1+W2+W3) summary(result3)