Changes/Bug Fixes in Version 2.0.1.1
-- Fixed fatal bug when bootstrapping ATT and ATC and more than 5% of outcomes are missing
-- And another bug when estimating Q with GLM, some missing outcomes, and cvQinit=FALSE
Changes/Bug Fixes in Version 2.0.1
-- Modified procedure for evaluating average effect among the treated (ATT) and controls (ATC) to improve finite sample bias and variance by trimming controls outside the distribution of W among the treated
-- Now returns EY1 and EY0 parameter estimates and inference
-- Fixed bug that crashed tmleMSM when setting "inference=FALSE"
Changes/Bug Fixes in Version 2.0.0
-- New option to specify observation weights for biased sampling, 2-stage designs (tmle and tmleMSM)
-- Option for Targeted bootstrap inference in addition to IC-based inference by setting B to a value >1
-- does not require additional calls to SL, so not too time-consuming
-- option to specify significance level other than alpha = 0.05
-- Ability to specify separate values for the number of cv folds for Q,g, g_Delta.
-- Other minor updates, see help files for details
Changes/Bug Fixes in Version 1.5.0.2
-- Minor bug fixes
-- Fixed bug when using dbarts to predict missingness probabilities
-- Covariates no longer prescreened when g.Deltaform is supplied
-- Automatic setting for V revised (see details in tmle function documentation)
Changes/Bug Fixes in Version 1.5.0-1
-- W can now include factor variables. They will be automatically converted to indicators via a call to model matrix.
-- Bounding g
-- g1W and g0W are now each bounded away from 0, but are allowed to take on the value 1.
-- default lower bound on g = 5/sqrt(n)/log(n) is highly recommended (manuscript in preparation)
-- prescreenW.g functionality changed to retain all covariates associated with Y by default, instead of only those associated with the residuals
-- setting a new argument, RESID = TRUE, allows replicating v1.4.0.1 behavior. That was sometimes problematic, and not recommended.
-- setting prescreenW.g = FALSE allows replicating behavior up until v 1.4.0, where all variables in W are used to model g
-- New arguments to better control estimation of conditional probability that outcome is observed.
-- g.Delta.SL.library - SL library to use for estimating P(Delta = 1 | A, W)
-- g.Delta.discreteSL - TRUE/FALSE to use discrete vs. ensemble SL
-- Two new diagnostics returned: Rsquared for initial estimate of Q (Mcfadden pseudo-Rsq for binary Y), AUC for propensity scores that condidtion on retained covariates in W.
-- If upgrading from package version 1.3.x, continue reading about substantial changes introduced in version 1.4.0.1
Changes/Bug Fixes in Version 1.4.0 and 1.4.0.1
-- Default settings have changed from the previous versions, and some new arguments.
-- New tuning parameters control estimation of g, and choice of targeting procedure
-- prescreenW.g = TRUE. Covariates included in the models for the components of g are now pre-screened by default to exclude those not associated with residual bias after the initial estimation of Q. (Set to FALSE to replicate behavior in older versions).
-- target.gWt = TRUE. The targeting procedure now incorporates into the weights when fitting epsilon by maximum likelihood, rather than in the denominator of the clever covariates. (Set to FALSE to replicate behavior in older versions.)
-- Q.SL.library and g.SL.library are now defined as SL.glm, SL.dbarts, and SL.glmnet, (formerly, SL.glm, SL.step, and SL.glm.interaction).
-- Qbounds that are not specified by the user are now set to 1% beyond the range of the observed Y (formerly 10%).
-- cvQinit is now set to TRUE so the initial estimate of Q returns cross-validated predictions by default (formerly FALSE)
-- alpha is now set to 0.9995 (formerly 0.995).
-- automatic = FALSE. Sets all tuning parameters to their default values, overriding user-specified values. Also data-adaptively sets V based on sample size ('help tmle' for more details).
-- Version 1.4.0.1 changes depends to suggests for dbarts and gam.
Changes/Bug Fixes in Version 1.3.0-2
-- W can now be passed in as a vector, matrix, or dataframe
-- Corrected small bug in variance calculation in tmleMSM function when outcomes are continuous
Changes/Bug Fixes in Version 1.3.0-1
-- Corrected small bug in oneStepATT function where bounds on g were initially ignored
Changes/Bug Fixes in Version 1.3.0
-- tmle function now returns the additive effect of treatment among the treated (ATT) and among the controls (ATC)
Changes/Bug Fixes in Version 1.2.0-5
-- Updated to be compatible with SuperLearner v2.0-21. SL.glm.interaction no longer masked.
-- Option to specify number of cross-validation folds for SL estimation of Q and g, default remains V = 5.
Changes/Bug Fixes in Version 1.2.0-4
-- Licences change to BSD_3_clause + file LICENSE | GPL-2
Changes/Bug Fixes in Version 1.2.0-3
-- Licences change to BSD | GPL-2
Changes/Bug Fixes in Version 1.2.0-2
-- Added cite to Journal of Statistical Software publication, "tmle: An R Package for Targeted Maximum Likelihood Estimation", 51(13), 1-35,
http://www.jstatsoft.org/v51/i13/
-- License change to BSD
Changes/Bug Fixes in Version 1.2.0-1
-- FEV dataset included withing the package
-- Update examples for tmleMSM()
Changes/Bug Fixes in Version 1.2.0
-- New tmleMSM() function for estimating parameters of an arbitry MSM. Type help(tmleMSM) for details.
-- Now works with SuperLearner versions 1.0 through 2.0-6
-- Default SuperLearner libraries for Q and g are now set to (SL.glm, SL.step, SL.glm.interaction),
all of which are included in the base installation of R.
-- Specifying SuperLearner screening algorithms within calls to SuperLearner is not supported.
-- If SuperLearner fails while estimating Q, the program quits instead of defaulting to main terms
regression with glm. But, if a call to SuperLearner fails when estimating g, the program prints a
warning message, and defaults to estimating g with main terms regression. This way, initial
estimate of Q, will be returned, and can be supplied to the next invocation of the function
without having to be be re-calculated.
-- Interaction terms now allowed in user-specified formulas for estimating Q and g
-- Factors are allowed in covariate matrix W, but will only work correctly when glm is used to estimate
Q and g. (Super Learner does not support factors. Conversion to dummy variables is recommended)