For ordinal multicategory treatments, setting
link = "br.logit" now uses brglm2::bracl() to
fit a bias-reduced ordinal regression model.
Added the vignette “Installing Supporting Packages” to explain
how to install the various packages that might be needed for
WeightIt to use certain methods, including when the package
is not on CRAN. See the vignette at
vignette("installing-packages").
Fixed a bug that would occur when a factor or character predictor
with a single level was passed to weightit().
Improved the code for entropy balancing, fixing a bug when using
s.weights with a continuous treatment and improving
messages when the optimization fails to converge. (#33)
Improved robustness of documentation to missing packages.
Updated the logo, thanks to Ben Stillerman.
Fixed a bug that would occur when the formula.tools
package was loaded, which would occur most commonly when
logistf was loaded. It would cause the error
The treatment and covariates must have the same number of units.
(#25)
Fixed a bug where the info component would not be
included in the output of weightit() when using
method = "super".
Added the ability to specify num.formula as a list
of formulas in weightitMSM(). This is primarily to get
around the fact that when stabilize = TRUE, a fully
saturated model with all treatments is used to compute the stabilization
factor, which, for many time points, is time-consuming and may be
impossible (especially if not all treatment combinations are observed).
Thanks to @maellecoursonnais for bringing up
this issue (#27).
ps.cont() has been retired since the same
functionality is available using weightit() with
method = "gbm" and in the twangContinuous
package.
With method = "energy", a new argument,
lambda, can be supplied, which puts a penalty on the square
of the weights to control the effective sample size. Typically this is
not needed but can help when the balancing is too aggressive.
With method = "energy", min.w can now
be negative, allowing for negative weights.
With method = "energy", dist.mat can
now be supplied as the name of a method to compute the distance matrix:
"scaled_euclidean", "mahalanobis", or
"euclidean".
Support for negative weights added to summary().
Negative weights are possible (though not by default) when using
method = "energy" or
method = "optweight".
Fixed a bug where glm() would fail to converge with
method = "ps" for binary treatments due to bad starting
values. (#31)
miss = "saem" can once again be used with
method = "ps" when missing values are present in the
covariates.
Fixed bugs with processing input formulas.
An error is now thrown if an incorrect link is
supplied with method = "ps".
The use of method = "twang" has been retired and
will now give an error message. Use method = "gbm" for
nearly identical functionality with more options, as detailed at
?method_gbm.
With multinomial treatments with link = "logit" (the
default), if the mclogit package is installed, it can be
requested for estimating the propensity score by setting the option
use.mclogit = TRUE, which uses
mclogit::mblogit(). It should give the same results as the
default, which uses mlogit, but can be faster and so is
recommended.
Added a plot() method for
summary.weightitMSM objects that functions just like
plot.summary.weightit() for each time point.
Fixed a bug in summary.weightit() where the labels
of the top weights were incorrect. Thanks to Adam Lilly.
Fixed a bug in sbps() when using a stochastic search
(i.e., full.search = FALSE or more than 8 moderator
levels). (#17)
Fixed a bug that would occur when all weights in a treatment
group were NA. Bad weights (i.e., all the same) now produce
a warning rather than an error so the weights can be diagnosed manually.
(#18)
Fixed a bug when using method = "energy" with
estimand = "ATE" and improved = TRUE (the
default). The between-treatment energy distance contribution was half of
what it should have been; this has now been corrected.
Added L1 median measure as a balance criterion. See
?stop.method for details.
Fixed a bug where logical treatments would yield an error. (#21)
Fixed a bug where Warning: Deprecated would appear
sometimes when purrr (part of the tidyverse)
was loaded. (#22) Thanks to MrFlick on StackOverflow for the solution.
Added support for estimating propensity scores using Bayesian
additive regression trees (BART) with method = "bart". This
method fits a BART model for the treatment using functions in the
dbarts package to estimate propensity scores that are used
in weights. Binary, multinomial, and continuous treatments are
supported. BART uses Bayesian priors for its hyperparameters, so no
hyperparameter tuning is necessary to get well-performing
predictions.
Fixed a bug when using method = "gbm" with
stop.method = "cv{#}".
Fixed a bug when setting estimand = "ATC" for
methods that produce a propensity score. In the past, the output
propensity score was the probability of being in the control group; now,
it is the probability of being in the treated group, as it is for all
other estimands. This does not affect the weights.
Setting method = "twang" is now deprecated. Use
method = "gbm" for improved performance and increased
functionality. method = "twang" relies on the
twang package; method = "gbm" calls
gbm directly.
Using method = "ebal" no longer requires the
ebal package. Instead, optim() is used, as it
has been with continuous treatments. Balance is a little better, but
some options have been removed.
When using method = "ebal" with continuous
treatments, a new argument, d.moments, can now be
specified. This controls the number of moments of the covariate and
treatment distributions that are constrained to be the same in the
weighted sample as they are in the original sample. Vegetabile et
al. (2020) recommend setting d.moments to at least 3 to
ensure generalizability and reduce bias due to effect
modification.
Made some minor changes to summary.weightit() and
plot.summary.weightit(). Fixed how negative entropy was
computed.
The option use.mnlogit in weightit()
with multi-category treatments and method = "ps" has been
removed because mnlogit appears uncooperative.
Fixed a bug (#16) when using method = "cbps" with
factor variables, thanks to @danielebottigliengo.
Fixed a bug when using binary factor treatments, thanks to Darren Stewart.
Cleaned up the documentation.
With method = "gbm", added the ability to tune
hyperparameters like interaction.depth and
distribution using the same criteria as is used to select
the optimal tree. A summary of the tuning results is included in
info in the weightit output object.
Fixed a bug where moments and int were
ignored unless both were specified.
Effective sample sizes now print only up to two digits (believe me, you don’t need three) and print more cleanly with whole numbers.
Fixed a bug when using by, thanks to @frankpopham.
(#11)
Fixed a bug when using weightitMSM with methods that
process int and moments (though you probably
shouldn’t use them anyway). Thanks to Sven Reiger.
Fixed a bug when using method = "npcbps" where
weights could be excessively small and mistaken for all being the same.
The weights now sum to the number of units.
Added support for energy balancing with
method = "energy". This method minimizes the energy
distance between samples, which is a multivariate distance measure. This
method uses code written specifically for WeightIt (i.e.,
it does not call a package specifically designed for energy balancing)
using the osqp package for the optimization (same as
optweight). See Huling & Mak (2020) for details on this
method. Also included is an option to require exact balance on moments
of the covariates while minimizing the energy distance. The method works
for binary and multinomial treatments with the ATE, ATT, or ATC.
Sampling weights are supported. Because the method requires the
calculation and manipulation of a distance matrix for all units, it can
be slow and/or memory intensive for large datasets.
Improvements to method = "gbm" and to
method = "super" with
SL.method = "method.balance". A new suite of
stop.methods are allowed. For binary treatments, these
include the energy distance, sample Mahalanobis distance, and pseudo-R2
of the weighted treatment model, among others. See
?stop.method for allowable options. In addition,
performance for both is quite a bit faster.
With multinomial treatments with link = "logit" (the
default), if the mnlogit package is installed, it can be
requested for estimating the propensity score by setting the option
use.mnlogit = TRUE. It should give the same results as the
default, which uses mlogit, but can be faster for large
datasets.
Added option estimand = "ATOS" for the “optimal
subset” treatment effect as described by Crump et al. (2009). This
estimand finds the subset of units who, with ATE weights applied, yields
a treatment effect with the lowest variance, assuming homoscedasticity
(and other assumptions). It is only available for binary treatments with
method = "ps". In general it makes more sense to use
estimand = "ATO" if you want a low-variance estimate and
don’t care about the target population, but I added this here for
completeness. It is available in get_w_from_ps() as
well.
make_full_rank() is now faster.
Cleaning up of some error messages.
Fixed a bug when using link = "log" for
method = "ps" with binary treatments.
Fixed a bug when using method = "cbps" with
continuous treatments and sampling weights. Previously the returned
weights included the sampling weights multiplied in; now they are
separated, as they are in all other scenarios and for all other
methods.
Improved processing of non-0/1 binary treatments, including for
method = "gbm". A guess will be made as to which treatment
is considered “treated”; this only affects produced propensity scores
but not weights.
Changed default value of at in trim()
from .99 to 0.
Added output for the number of weights equal to zero in
summary.weightit. This can be especially helpful when using
"optweight" or "energy" methods or when using
estimand = "ATOS".
Added support for entropy balancing
(method = "ebal") for continuous treatments as described by
Tübbicke (2020). Relies on hand-written code contributed by Stefan
Tübbicke rather than another R package. Sampling weights and base
weights are both supported as they are with binary and multi-category
treatments.
Added support for Balance SuperLearner as described by Pirracchio
and Carone (2018) with method = "super". Rather than using
NNLS to choose the optimal combination of predictions, you can now
optimize balance. To do so, set
SL.method = "method.balance". You will need to set an
argument to stop.method, which works identically to how it
does for method = "gbm". For example, for
stop.method = "es.max", the predicted values given will be
the combination of predicted values that minimizes the largest absolute
standardized mean difference of the covariates in the sample weighted
using the predicted values as propensity scores.
Changed some of the statistics displayed when using
summary(): the weight ratio is gone (because weights can be
0, which is not problematic but would explode the ratio), and the mean
absolute deviation and entropy of the weights are now present.
Added crayon for prettier printing of
summary() output.
Formula interfaces now accept poly(x, .) and other
matrix-generating functions of variables, including the
rms-class-generating functions from the rms
package (e.g., pol(), rcs(), etc.) (the
rms package must be loaded to use these latter ones) and
the basis-class-generating functions from the
splines package (i.e., bs() and
ns()). A bug in an early version of this was found by @ahinton-mmc.
Added support for marginal mean weighting through stratification
(MMWS) as described by Hong (2010, 2012) for weightit() and
get_w_from_ps() through the subclass argument
(see References at ?get_w_from_ps). With this method,
subclasses are formed based on the propensity score and weights are
computed based on the number of units in each subclass. MMWS can be used
with any method that produces a propensity score. The implementation
here ensures all subclasses have a least one member by filling in empty
subclasses with neighboring units.
Added stabilize option to
get_w_from_ps().
A new missing argument has been added to
weightit() to choose how missing data in the covariates is
handled. For most methods, only "ind" (i.e., missing
indicators with single-value imputation) is allowed, but for
"ps", "gbm", and "twang", other
methods are possible. For method = "ps", a stochastic
approximation of the EM algorithm (SAEM) can be used through the
misaem package by setting
missing = "saem".
For continuous treatments with the "ps",
"gbm", and "super" methods (i.e., where the
conditional density of the treatment needs to be estimated), the user
can now supply their own density as a string or function rather than
using the normal density or kernel density estimation. For example, to
use the density of the t-distribution with 3 degrees of freedom, one can
set density = "dt_3". T-distributions often work better
than normal distributions for extreme values of the treatment.
Some methods now have an info component in the
output object. This contains information that might be useful in
diagnosing or reporting the method. For example, when
method = "gbm", info contains the tree that
was used to compute the weights and the balance resulting from all the
trees, which can be plotted using plot(). When
method = "super", info contains the
coefficients in the stacking model and the cross-validation risk of each
of the component methods.
For method = "gbm", the best tree can be chosen
using cross validation rather than balance by setting
stop.method = "cv5", e.g., to do 5-fold
cross-validation.
For method = "gbm", a new optional argument
start.tree can be set to select the tree at which balance
begins to be computed. This can speed things up when you know that the
best tree is not within the first 100 trees, for example.
When using method = "gbm" with multi-category
treatments and estimands other than the ATE,
ATT, or ATC are used with standardized mean
differences as the stopping rule, the mean differences will be between
the weighted overall sample and each treatment group. Otherwise, some
efficiency improvements.
When using method = "ps" with multi-category
treatments, the use of use.mlogit = FALSE to request
multiple binary regressions instead of multinomial regression is now
documented and an associated bug is now fixed, thanks to @ahinton-mmc.
When use method = "super", one can now set
discrete = TRUE to use discrete SuperLearner instead of
stacked SuperLearner, but you probably shouldn’t.
moments and int can now be used with
method = "npcbps".
Performance enhancements.
Fixed bug when using weightit() inside another
function that passed a by argument explicitly. Also changed
the syntax for by; it must now either be a string (which
was always possible) or a one-sided formula with the stratifying
variable on the right-hand side. To use a variable that is not in
data, you must use the formula interface.
Fixed bug when trying to use ps with by
in weightit().
Added new sbps() function for estimating subgroup
balancing propensity score weights, including both the standard method
and a new smooth version.
Setting method = "gbm" and
method = "twang" will now do two different things.
method = "gbm" uses gbm and
cobalt functions to estimate the weights and is much
faster, while method = "twang" uses twang
functions to estimate the weights. The results are similar between the
two methods. Prior to this version, method = "gbm" and
method = "twang" both did what
method = "twang" does now.
Bug fixes when stabilize = TRUE, thanks to @ulriksartipy and Sven
Rieger.
Fixes for using base.weight argument with
method = "ebal". Now the supplied vector should have a
length equal to the number of units in the dataset (in contrast to its
use in ebalance, which requires a length equal to the
number of control units).
Restored dependency on cobalt for examples and
vignette.
When method = "ps" and the treatment is ordered
(i.e., ordinal), MASS::polr() is used to fit an ordinal
regression. Make the treatment un-ordered to to use multinomial
regression instead.
Added support for using bias-reduced fitting functions when
method = "ps" as provided by the brglm2
package. These can be accessed by changing the link to, for
example, "br.logit" or "br.probit". For
multinomial treatments, setting link = "br.logit" fits a
bias-reduced multinomial regression model using
brglm2::brmultinom(). This can be helpful when regular
maximum likelihood models fail to converge, though this may also be a
sign of lack of overlap.
Bug fixes. Functions now work better when used inside other
functions (e.g., lapply).
Behavior of weightit() in the presence of
non-NULL focal has changed. When
focal is specified, estimand is assumed to be
ATT. Previously, focal would be ignored unless
estimand = "ATT".
Processing of estimand and focal is
improved. Functions are smarter about guessing which group is the focal
group when one isn’t specified, especially with non-numeric treatments.
focal can now be used with estimand = "ATC" to
indicate which group is the control group, so "ATC" and
"ATT" now function more similarly.
Added function get_w_from_ps() to transform
propensity scores into weights (instead of having to go through
weightit()).
Added functions as.weightit() and
as.weightitMSM() to convert weights and treatments and
other components into weightit objects so that
summary.weightit() can be used on them.
Updated documentation to describe how missing data in the covariates is handled. Some bugs related to missing data have been fixed as well, thanks to Yong Hao Pua.
ps.cont() had the “z-transformed correlation”
options removed to simplify output. This function and its supporting
functions will be deprecated as soon as the new version of
twang is released.
When using method = "ps" or
method = "super" with continuous treatments, setting
use.kernel = TRUE and plot = TRUE, the plot is
now made with ggplot2 rather than the base R
plots.
Added plot.summary.weightit() to plot the
distribution of weights (a feature also in
optweight).
Removed dependency on cobalt temporarily, which
means the examples and vignette won’t run.
Added ggplot2 to Imports.
Fixed a bug when using the ps argument in
weightit().
Fixed a bug when setting include.obj = TRUE in
weightitMSM().
Added warnings for using certain methods with longitudinal treatments as they are not validated and may lead to incorrect inferences.
Added super method to estimate propensity scores
using the SuperLearner package.
Added optweight method to estimate weights using
optimization (but you should probably just use the
optweight package).
weightit() now uses the correct formula to estimate
weights for the ATO with multinomial treatments as described by Li &
Li (2018).
Added include.obj option in weightit()
and weightitMSM() to include the fitted object in the
output object for inspection. For example, with
method = "ps", the glm object containing the
propensity score model will be included in the output.
Rearranged the help pages. Each method now has its own
documentation page, linked from the weightit help
page.
Propensity scores are now included in the output for binary
treatments with gbm and cbps methods. Thanks
to @Blanch-Font
for the suggestion.
Other bug fixes and minor changes.
Added trim() function to trim weights.
Added ps.cont() function, which estimates
generalized propensity score weights for continuous treatments using
generalized boosted modeling, as in twang. This function
uses the same syntax as ps() in twang, and can
also be accessed using weightit() with
method = "gbm". Support functions were added to make it
compatible with twang functions for assessing balance
(e.g., summary, bal.table, plot).
Thanks to Donna Coffman for enlightening me about this method and
providing the code to implement it.
The input formula is now much more forgiving, allowing objects in
the environment to be included. The data argument to
weightit() is now optional. To simplify things, the output
object no longer contains a data field.
Under-the-hood changes to facilitate adding new features and debugging. Some aspects of the output objects have been slightly changed, but it shouldn’t affect use for most users.
Fixed a bug where variables would be thrown out when
method = "ebal".
Added new moments and int options for
some weightit() methods to easily specify moments and
interactions of covariates.
Fixed bug when using objects not in the data set in
weightit(). Behavior has changed to include transformed
covariates entered in formula in weightit()
output.
Fixed bug resulting from potentially colinearity when using
ebal or ebcw.
Added a vignette.
Edits to code and help files to protect against missing
CBPS package.
Corrected sampling weights functionality so they work correctly.
Also expanded sampling weights to be able to be used with all methods,
including those that do not natively allow for sampling weights (e.g.,
ATE).
Minor bug fixes and spelling corrections.
Added weightitMSM() function (and supporting
print() and summary() functions) to estimate
weights for marginal structural models with time-varying treatments and
covariates.
Fixed some bugs, including when using CBPS with continuous
treatments, and when using focal incorrectly.
Added method = "sbw" for stable balancing weights
(now removed and replaced with
method = "optweight")
Allowed for estimation of multinomial propensity scores using
multiple binary regressions if mlogit is not
installed
Allowed for estimation of multinomial CBPS using multiple binary CBPS for more than 4 groups
Added README and NEWS