WeightIt 2.1.0With method = "bart", use.offset can
now be set to TRUE to use the linear predictor of a GLM as
an offset in the BART model.
With method = "gbm", balance is now assessed on
multiple trees in parallel when a future
plan with more than one worker is set. This requires future.apply
to be installed.
Model formulas supplied to weightitMSM() can now
have random effects included. Note this is intended to be used in the
case of clustering, not for modeling longitudinal treatments in a single
model.
Fixed a bug in summary.weightitMSM() where mean
weights would be printed twice, one for each group.
In weightitMSM(), stabilize can now be
supplied the same arguments as num.formula, making
num.formula essentially redundant.
Stabilization formulas can now include random effects in both
weightit() and weightitMSM().
The stabilize component of weightit
objects has been renamed to stabilization to match
weightitMSM objects; this contains the stabilization
formula, if any. This also means the mean weights will be displayed in
the summary.weightit() output.
In weightit(), setting stabilize = TRUE
with a continuous treatment is accepted rather than being ignored with a
warning. It requests a marginal density model in the numerator, which
for a point treatment is the density the weights already divide by, so
the weights are unchanged; such an object is not reported as stabilized
and has no stabilization component. Supplying a formula
with terms in it, as in stabilize = ~ x1, does change the
weights. The same applies to any numerator that comes to exactly 1, in
weightit() and weightitMSM() alike.
Fixed a bug in which the name of a multi-category (i.e.,
factor or character) treatment was lost when
the treatment was processed. In weightitMSM(), this meant a
multi-category time point would either be left unnamed or fail with an
error about a replacement having length zero.
Fixed a bug in print.weightitMSM() in which the
stabilization factors would be run together with the line that follows
them.
Added a documentation page at ?method_ps describing
how weights are computed from propensity scores supplied to the
ps argument of weightit(), including the
subclass and density arguments that page accepts.
Fixed a bug in get_w_from_ps() (and
weightit()) with estimand = "ATOS" in which
only part of the candidate values of alpha were searched,
so the resulting subset depended on which treatment level was coded as
treated.
Fixed a bug in summary.weightit() in which the
maximum of the weight range ignored the sampling weights while the
minimum did not.
Fixed a bug in summary.weightitMSM() in which
weight.range = FALSE was ignored.
summary() on a weightitMSM object that
models censoring now includes an entry for each censoring model, placed
among the treatment entries in the order the models were fit and named
for its censoring indicator. Each covers the units still under
observation when that model was fit, since the censored units have a
weight of exactly 0 and are not part of the weighted sample.
plot() can select one by name as well as by position,
e.g. plot(summary(W), time = "C_2").
summary() on a weightitMSM object gains
a which.time argument, which restricts the summary to the
given models, given as a vector of positions in
formula.list or of treatment or censoring variable names;
omit it for all of them. Selecting a subset does not renumber anything,
so a model keeps the position it has in the whole sequence.
The headings in the printed summary of a weightitMSM
object now name the model, as in 1. Treatment: A_1 and
2. Censoring: C_2, where the number is the model’s position
in formula.list. Previously they were numbered by treatment
time point and did not say which variable they were about.
In plot() on a summary.weightitMSM
object, the time argument has been renamed
which.time for consistency with summary().
time continues to work.
For a censoring model, the effective sample size reported by
summary() is now measured against the units that model was
fit on, i.e., those still under observation entering it, rather than
against all units. With censoring at more than one time point, the
latter counted units that had already dropped out and were never
eligible.
Fixed a bug in plot() on a
summary.weightit object in which the weights of units with
no treatment value, i.e., those censored at an earlier time point, were
displayed in a facet of their own, and in which a censoring model’s
weights were displayed as though the censoring indicator were a
treatment.
Fixed a bug in plot.summary.weightit() in which the
bins argument was ignored.
Fixed a bug in method = "ebal" with a vector-valued
tols, which errored for binary and multi-category
treatments.
Fixed a bug in method = "ebal" with continuous
treatments in which supplying d.moments greater than any
entry of a per-covariate moments vector (e.g.,
moments = c(x1 = 2, x2 = 3) with
d.moments = 3) would hold only the means of the
covariates to their unweighted values, rather than the requested number
of moments; i.e., raising d.moments reduced the number of
moments held instead of increasing it.
With method = "bart", samplerOnly is
now ignored as documented rather than passed on to
dbarts::bart2(), where it would cause an error.
method = "gbm" no longer accepts
estimand = "ATOS", which was listed as available but too
slow to be practical.
After trim() or calibrate(), the
components used for M-estimation are now removed, so
glm_weightit() no longer computes standard errors from the
untrimmed or uncalibrated weights. Such objects now use
vcov = "HC0" by default; use vcov = "BS" or
"FWB" to account for trimming.
.cens() is now cobalt’s
.cens(), re-exported. cobalt 5.0.0 gained its own
copy so that cobalt::bal.tab() could recognize a censoring
model, and two identically named exports meant one masked the other on
library(WeightIt). There is now one function:
.cens() behaves exactly as before, and an indicator tagged
with either package attached is the same object and accepted by both.
cobalt 5.0.0 or later is required.
Fixed a bug in which only a bare .cens() marker was
recognized as one, so a censoring model written as
WeightIt::.cens(C) ~ x1 + x2 was named after the whole call
rather than after the indicator: the model was named
WeightIt::.cens(C) rather than C, which in
weightitMSM() is also how the time point is identified in
summary() and in cobalt::bal.tab(). The
weights were correct either way. What marks a censoring model is now the
indicator .cens() returns rather than the syntax that
produced it, so any way of naming the same function works, including
cobalt::.cens() and a local alias, as does supplying a
variable that was tagged ahead of time.
WeightIt 2.0.0Several new features have been added, described in more detail below. In particular, these are censoring weights estimation and random effects in propensity score models.
Added support for estimating inverse probability of censoring
weights (IPCW). Censoring is now its own treatment type, distinct from
binary, multi-category, and continuous treatments. Wrapping a censoring
indicator in .cens() on the left side of a model formula
requests it, as in
weightit(.cens(C) ~ x1 + x2, data = d, method = "glm"). The
convention is that the indicator is 1 for units that are censored and 0
for those still under observation; censored units receive a weight of
exactly 0, and the rest receive \(1/P(C = 0 |
X)\). .cens() is documented at
?.cens.
.cens() can also be called outside a formula to tag
a variable as a censoring indicator, which routes the lower-level
interfaces to the censoring path:
weightit.fit(covs, treat = .cens(C), method = "glm") and
get_w_from_ps(ps, treat = .cens(C)) estimate and compute
censoring weights, the latter returning (1 - C)/(1 - ps),
where ps is the probability of being censored.
Arguments that do not apply (estimand, focal,
treated, subclass, stabilize) are
ignored with a warning.
Each weighting method has a dedicated implementation for
censoring weights, rather than reusing the binary-treatment machinery.
Because inverse probability of censoring weights are only ever needed
for the units still under observation, these estimate weights for that
group alone, targeting the covariate distribution of the full at-risk
sample. This avoids the wasted (and sometimes infeasible) problem of
solving for weights among the censored units, which matters most when
few units are censored. Censoring weights are available for
method values "glm", "bart",
"cbps", "ebal", "energy",
"gbm", "ipt", "optweight",
"super", and "cfd", as well as
method = NULL (which yields complete-case weights) and a
supplied ps. They are not available for
method = "npcbps". M-estimation-based standard errors are
available for the same methods that support them for binary
treatments.
weightitMSM() supports censoring models interleaved
with the treatment models in formula.list in temporal
order, as in
list(A_1 ~ X_0, A_2 ~ X_1 + A_1, .cens(C_2) ~ X_1 + A_1, A_3 ~ X_2 + A_2).
Each model is fit only among the units still under observation at that
time point, missing values are permitted in later treatments for units
already censored, and the censoring weights are folded into the product
of weights across time points. Censoring is treated as a treatment type
rather than as a separate kind of model: the censoring indicators sit
among the treatments in treat.list, and their covariates in
covs.list, in the order the models were fit, and the
print() and summary() methods list them there
too, identifying each entry by its position in formula.list
and the variable modeled. The new at.risk component of the
output records which units were under observation when each model was
fit and is documented in the “Value” section of
?weightitMSM.
In weightitMSM(), censoring time points are
stabilized exactly as treatment time points are. With
stabilize = TRUE, the numerator of a censoring weight is a
model for that censoring indicator saturated in the preceding treatments
(an intercept when no treatment precedes it), and
num.formula adds stabilization factors to it as it does for
a treatment. When num.formula is supplied as a list, it
must therefore have one entry per entry of formula.list,
censoring entries included.
The numerator of a stabilized censoring weight is fit as a
censoring model, using the same method as the denominator, in both
weightit() and weightitMSM(). The stabilized
weight is \(P(C = 0 | V)/P(C = 0 | X)\)
for the units still under observation, where \(V\) are the stabilizing variables (usually
just prior treatments, if any), and exactly 0 for those
censored.
A censoring formula may have an empty right hand side, as in
.cens(C) ~ 1, requesting a marginal censoring model in
which censoring does not depend on the covariates. The weights are then
\(1/P(C = 0)\) for the units still
under observation and 0 for those censored. This works in both
weightit() and weightitMSM(), with any
method (see the note on empty formulas under Other new
features), and leaves the rest of the censoring machinery intact,
including the risk sets, the missing values permitted after censoring,
by, stabilize, and M-estimation.
Censoring weights are also available with
method = "cbps" and is.MSM.method = TRUE in
weightitMSM(), i.e., estimating one set of weights that
satisfies the balance conditions at all time points simultaneously. Each
condition is evaluated among the units still under observation at that
time point, so the covariates and treatments that are missing for
already-censored units never enter the estimating equations. The
treatment conditions use the cumulative weights, which include the
censoring factors, so they account for censoring; each censoring
condition uses its own factor.
glm_weightit(), lm_weightit(),
multinom_weightit(), ordinal_weightit(), and
coxph_weightit() now tolerate missing values in the model
variables for units with a weight of 0, which contribute neither to the
fit nor to its variance. This makes it possible to fit an outcome model
after censoring, where the outcome is unobserved for censored units.
Missing values in units with a nonzero weight continue to produce an
error. Relatedly, coxph_weightit() now allows weights of 0,
which survival::coxph() rejects; such units are omitted
from the model fit.
method = "glm" now supports multilevel
(mixed-effects) propensity score models. When the model
formula supplied to weightit() contains
lme4-style random effects terms (e.g.,
treat ~ x1 + x2 + (1 | school)), a multilevel model is fit
to estimate the propensity scores: lme4::glmer() for binary
treatments, lme4::lmer() for continuous treatments, and
mclogit::mblogit() for multi-category treatments. The
estimated propensity scores are cluster-specific (they include the
estimated random effects).
method = "bart" also supports multilevel propensity
score models. When the model formula supplied to
weightit() contains lme4-style random effects
terms (e.g., treat ~ x1 + x2 + (1 | school)), a multilevel
BART model is fit using stan4bart::stan4bart(), which
combines a BART sum-of-trees for the covariates with a Stan-estimated
random-effects component. The full flexibility of
lme4::glmer() random effects is available, including
multiple grouping factors and random slopes. bart2()-style
control arguments are translated to their stan4bart()
equivalents. The estimated propensity scores are
cluster-specific.
Added an re_ok field to the entries of
.weightit_methods indicating whether each method supports
random effects terms in the model formula.
multinom_weightit() and
ordinal_weightit() gain a br argument, which,
when set to TRUE, performs mean bias reduction by solving
the bias-reducing adjusted score equations of Firth (1993) instead of
the usual maximum likelihood score equations. This yields estimates with
smaller asymptotic bias that are always finite, even when the maximum
likelihood estimates are not (e.g., under separation or when an end
category of an ordinal outcome is unobserved). This mirrors
br in glm_weightit() but uses in-house
implementations rather than brglm2, so it requires no
additional dependencies and composes with M-estimation, cluster-robust
standard errors, and bootstrapping.
multinom_weightit(), the adjustment is that of
Kosmidis and Firth (2011); because the model uses the canonical link,
the adjusted score is the gradient of the log-likelihood penalized by
the Jeffreys invariant prior, and estimates align with
brglm2::brmultinom().ordinal_weightit(), the adjustment is the one
derived for cumulative link models by Kosmidis (2014) and works for all
supported links; the adjusted score equations are solved by quasi-Fisher
scoring, controlled by the new br.maxit and
br.tol components of control. Note this
differs from brglm2::bracl(), which fits an adjacent
category logit model rather than a cumulative link model (including with
parallel = TRUE, which constrains the adjacent category
logits rather than the cumulative ones to be parallel).With method = "glm" and a multi-category treatment,
bias-reduced propensity score models are now fit by WeightIt
itself rather than by brglm2. A br. prefix on
link requests mean bias reduction, as before, but it is now
handled by multinom_weightit()’s and
ordinal_weightit()’s machinery:
multi.method = "weightit" accepts "br.logit"
for unordered treatments and any of "br.logit",
"br.probit", "br.loglog",
"br.cloglog", and "br.cauchit" for ordered
ones. Because MASS::polr() cannot fit bias-reduced models,
supplying a br. link with
multi.method = "polr" uses the WeightIt
implementation as well. The practical gain is that M-estimation-based
standard errors are now available for bias-reduced propensity score
models, which they were not when brglm2 did the fitting;
brglm2 is also no longer needed for this.
M-estimation-based standard errors are now available when
by is supplied to weightit() and
weightitMSM(). Previously, specifying by
omitted the M-estimation components entirely; now the per-stratum
components are combined so that glm_weightit() and friends
produce standard errors asymptotically equivalent to those from
estimating the weights from a single model in which the by
variable is fully interacted with all the covariates (at all time points
for weightitMSM()). This works for methods that support
M-estimation and composes with stabilize.
link = "softplus" can be used with
method = "cbps" and method = "ipt" to use the
softplus link.
weightit.fit() gains a subset argument,
restricting which units are used to fit the model.
The formulas in formula.list supplied to
weightitMSM() can now have an empty right hand side (e.g.,
A_1 ~ 1), which requests a marginal model at that time
point. Previously this produced an error, even though
weightit() has always allowed it.
An empty model formula now works with every method.
With no covariates there is nothing for any method to model or balance,
and every method’s target is met by the same weights, but most methods
could not run at all on a covariate-free model, failing with errors from
the packages they call that said nothing about the cause. Those weights
are now computed by fitting an intercept-only GLM whatever
method is supplied, ignoring any method-specific arguments.
This is an implementation shortcut only: method is reported
as supplied, and the package for the requested method need not be
installed. weightit(), weightitMSM(), and
weightit.fit() all behave this way.
method = NULL, a user-defined method function,
a supplied ps, and a formula whose only terms are
lme4-style random effects (e.g.,
A ~ (1 | school)) are unaffected. Relatedly,
method = "cbps" with is.MSM.method = TRUE now
accepts a time point with no covariates, which contributes only its
intercept balance condition.
With continuous treatments, estimation of the weights works a bit
differently for method = "glm", "bart",
"gbm", and "super". Previously, the numerator
of the weights (corresponding to the marginal distribution of treatment)
was estimated using the argument to density (by default, a
normal distribution). Now, density only controls the
conditional distribution used to estimate the denominator, and
the numerator is estimated by marginalizing over the conditional
distribution. This yields a marginal distribution that is compatible
with the conditional distribution. In most cases, this will improve
balance, but estimation will be a bit slower. Note this makes results of
these methods no longer backward compatible. Also note that
method = "cbps" is not affected by this change, even though
it too estimates generalized propensity score weights.
The plot argument can no longer be used to visualize
the densities for continuous treatments.
With method = "gbm", method = "surr" is
no longer allowed, as it did nothing different from "ind".
It is now set to "ind" with a warning.
The link argument in ordinal_weightit()
no longer accepts "log" and "clog" as valid
links.
An argument passed to the subset argument of
weightit.fit() was previously silently ignored; it is now
honored.
nobs() on a coxph_weightit object now
returns the number of units contributing to the fit (i.e., those with a
nonzero weight), consistent with the other *_weightit()
methods. Previously the method was defined but never registered, so
dispatch fell through to survival:::nobs.coxph(), which
returns the number of events. AIC() and
BIC() are unaffected, as they read the "nobs"
attribute that logLik() carries, but the residual degrees
of freedom reported by anova() change accordingly.
Passing family = "multinomial" to
glm_weightit() now results in an error. Previously it would
route the input to multinom_weightit().
multi.method = "brmultinom" and
multi.method = "bracl" in weightit() with
method = "glm" are deprecated. They are accepted with a
warning and now use the equivalent in-house bias-reduced fitters. Note
the estimated propensity scores change for the former
"bracl": brglm2::bracl() fits an adjacent
category logit model, whereas ordinal_weightit() fits a
cumulative link (proportional odds) model.
For ordered multi-category treatments with
method = "glm" and multi.method = "weightit",
the link argument no longer accepts
"identity", "log", or
"clog".
method = "cbps" now warns when the just-identified
CBPS (over = FALSE) fails to solve its balance conditions.
Previously this was silent: optim()’s convergence flag only
reports whether the iteration limit or a line search failed, and the
code never checked whether the balance objective had actually been
driven to zero, so a fit that settled at a point where the conditions
were still unmet returned weights that did not balance the covariates
with no indication anything was wrong. This is most likely with a
continuous treatment and many covariates, and depends on the sample
rather than the specification alone. See the new Unsolvable Balance
Conditions section of ?method_cbps.
Fixed a bug in tidy() for
glm_weightit(), multinom_weightit(),
ordinal_weightit(), and coxph_weightit()
objects that mislabeled columns. With conf.int = TRUE the
two confidence limits were returned unnamed rather than as
conf.low and conf.high, and with
exponentiate = TRUE every name was shifted by one — the
test statistic was labeled std.error — because transforming
drops the standard-error column. The names are now taken from the
columns actually returned.
plot() on a summary.weightit object for
a continuous treatment no longer triggers a ggplot2 deprecation
warning about the size aesthetic for lines.
Fixed a bug in method = "glm" with a multi-category
treatment and multi.method = "glm" that made every
M-estimation-based standard error NA. That option fits one
one-vs-rest binomial model per treatment level, and the M-estimation
parts passed the treatment factor to each of those scores instead of the
level’s 0/1 indicator, so the score matrix (and hence the whole variance
matrix) came out NA. Note this failed silently:
glm_weightit() and friends returned a fit whose standard
errors were all NA rather than erroring.
Fixed a bug in method = "glm" with a continuous
treatment that made every M-estimation-based standard error an error.
The residual standard deviation was omitted from the M-estimation
parameter vector, leaving it one element short of the model matrix, so
vcov() on a glm_weightit() fit using such
weights failed with “non-conformable arguments”.
vcov = "HC0" and the other non-M-estimation variance types
were unaffected. The standard deviation used is now the
sampling-weighted root mean square of the residuals, which is the value
the M-estimation score for that parameter solves for.
multinom_weightit() and
ordinal_weightit() now warn when the optimization fails,
which previously passed silently. Two cases are detected:
optim() reporting non-convergence (e.g., from exhausting
maxit), and optim() reporting success at a
point that does not solve the estimating equations. The latter happens
when the optimization is started far from the solution, which can occur
when start is supplied from a diverging maximum likelihood
fit; the coefficients returned in that case were not a solution of
anything, and the score contributions did not sum to 0, so the standard
errors were invalid too. The internal starting values are unaffected:
multinom_weightit() starts from 0, and
ordinal_weightit() with br = TRUE retries from
a neutral point.
Fixed a bug in ordinal_weightit() with
link = "loglog" that made the gradient and every robust
variance matrix NaN. The "loglog"
mu.eta() function evaluated to NaN rather than
0 at a linear predictor of -Inf, which is where the lowest
threshold sits.
When br = TRUE in glm_weightit() and a
weightit object supplying M-estimation components is used,
the derivative of the outcome score with respect to the weighting
model’s coefficients is now computed by numeric differentiation. It
previously used a shortcut that assumes the outcome score is linear in
the weights, which the bias-reducing adjustment is not. Standard errors
change slightly.
Fixed a bug in method = "ebal" with
tols greater than 0 that could cause the FISTA solver to
stop before the requested balance tolerance was reliably achieved.
Thanks to Ivan Geshev for pointing it out.
Fixed a bug in method = "ebal" and
method = "energy" with tols greater than 0, a
continuous treatment, and sampling weights that could allow the
treatment-covariate correlation to slightly exceed the requested
tolerance.
Fixed a bug in method = "energy" with
estimand = "ATC" and moments greater than
0.
Fixed a bug in method = "energy" and
method = "cfd" with min.w set to
-Inf.
Fixed a bug in which method = "gbm" with
criterion set to a cross-validation specification (e.g.,
"cv5") would fail with an error.
Fixed a bug in method = "cbps" with
is.MSM.method = TRUE in which the balance conditions for a
multi-category treatment used the treatment levels of the last
time point rather than their own. This produced an error when the last
time point was binary and silently used the wrong level set when two
multi-category time points had the same number of levels.
Fixed a bug in method = "cbps" with
is.MSM.method = TRUE in which a time point at which every
unit was treated would compute the weights for all units as though none
were treated.
Fixed a bug in coxph_weightit() in which a
rank-deficient model (one with collinear predictors, so that some
coefficients are not estimable) would fail with “non-conformable
arguments”. The score contributions used for the variance had one column
per model matrix column, while the model matrix itself had been reduced
to its estimable columns. Such fits now succeed, reporting the estimable
coefficients and their standard errors and NA for the rest,
as glm_weightit() does.
Fixed a bug in weightitMSM() in which supplying a
single formula to num.formula would fail to correctly
incorporate its terms into the stabilization model at all but the first
time point.
Fixed a bug in which predict() on an
ordinal_weightit object with type = "link" or
type = "stdlv" and no newdata supplied could
return incorrect values due to a scaling error.
Fixed a bug in coxph_weightit() in which
vcov = "const" would always fail with the error “attempt to
set an attribute on NULL”. The model-based variance is now returned,
matching that of an unweighted survival::coxph()
fit.
Fixed a bug in ordinal_weightit() and
multinom_weightit() in which vcov = "const"
would always fail with the error “no applicable method for ‘vcov’
applied to an object of class "list"”. The model-based variance is now
returned, matching that of MASS::polr() and
mlogit::mlogit(), respectively.
Fixed a bug in coxph_weightit() in which a model
with no covariates (e.g., Surv(time, event) ~ 1) would fail
with the error “‘score’ residuals are not defined for a null model”.
Such a model now returns an object with no coefficients and an empty
variance matrix, which also makes anova() against a null
model work.
Fixed a bug in glm_weightit() and
lm_weightit() in which a model with collinear covariates
would fail, with the error “length of ‘dimnames’ [1] not equal to array
extent” when vcov = "const" and “the Hessian for the
outcome model could not be inverted” when vcov = "HC0" or
vcov = "asympt". As with stats::glm(), the
redundant coefficients are now NA, and the variance matrix
covers the estimable coefficients.
Fixed a bug in summary() for
glm_weightit, lm_weightit,
multinom_weightit, and ordinal_weightit
objects in which all standard errors (and the test statistics and
p-values derived from them) would be NA when any
coefficient was aliased due to collinear covariates.
Fixed a bug in coxph_weightit() in which a model fit
to data with no observed events would fail with the error “object
‘jacob’ not found” (or “replacement has length zero” for a model with no
covariates). As with survival::coxph(), the coefficients
are now NA, along with the variance matrix, and a warning
is thrown.
weights is now correctly ignored with a warning in
glm_weightit() and friends.
WeightIt 1.7.0coxph_weightit() can now estimate standard errors
that account for estimation of the weights using M-estimation as
described by Shu et
al. (2021).
weightit() with method = "ebal" now
accepts an optional tols argument, which can be used to set
a balance threshold for inexact balance. This uses a new algorithm for
optimizing with L1 regularization.
Fixed bugs in entropy balancing that made the M-estimation covariance slightly too small. Thanks to Chad Hazlett for pointing them out.
When using the tols argument with
method = "energy" with a continuous treatment, the
imbalance tolerances now yield correlations no larger than the imbalance
tolerance as computed with cobalt::col_w_corr().
Previously, they were allowed to be slightly too large.
glm_weightit(), multinom_weightit(),
ordinal_weightit(), and coxph_weightit() are
now each documented on their own page.
The Estimating Effects vignettes
(vignette("estimating-effects")) now describes how to use
the adrftools
package to estimate the effect of a continuous treatment.
arg is a new dependency for producing error messages and checking arguments.
Minor changes to plot.summary.weightit(), in
particular to remove a deprecation warning from
ggplot2.
In the summary.weightit() output, units with the
highest weights are now named according to their rownames in
data. summary.weightit.sbps() has been updated
to be more in line with summary.weightit().
WeightIt 1.6.0Added method = "cfd" for characteristic function
distance balancing as described by Santra, Chen, and Park
(2026). Energy balancing is a special case of this method.
Added support for the new version of osqp, which changes
some optional argument names and defaults for
method = "energy". These should not impact
results.
chk and crayon have been removed as dependencies in favor of cli.
WeightIt 1.5.1calibrate() with method = "isoreg" can
now be used with sampling weights.
With predict.ordinal_weightit() and
predict.multinom_weightit() with
type = "response", an optional level argument
can be supplied to request predictions solely for the given outcome
level.
When vcov is "BS" or "FWB"
in glm_weightit(), etc., any call to trim() or
calibrate() will be applied within each bootstrap
replication.
Fixed a bug when using bootstrapping on
glm_weightit(), etc., when the original call to
weightit() was called as WeightIt::weightit().
(#85)
Fixed a bug when using weightit() with only
categorical covariates. Thanks to @harshanperera85-commits and
Pengfei Guo for pointing it out. (#86)
When a variable is manually excluded from a formula, e.g.,
treat ~ . - re78, that variable no longer is included in
the covariate output.
WeightIt 1.5.0For predict.ordinal_weightit(), type
can now be "stdlv" to request predictions on the scale of
the standardized latent variable underlying the ordinal
responses.
moments can now be supplied as a named vector with a
different integer value for each covariate, e.g.,
moments = c(x1 = 2, x2 = 3) to balance just the first two
moments of x1 and the first three of
x2.
With method = "energy", a new tols
argument can be set to request inexact balance on covariate
moments.
Some arguments and documentation for them has changed, in
particular moments, int, and
subclass, which used to be named arguments of
weightit() and are now to be supplied through
.... They are documented only on the help pages of the
specific methods that allow them.
The covs component of the weightit()
output object now only includes the raw covariates, not any
transformations. This will affect cobalt::bal.tab()
output.
In summary.weightit(), a new argument,
weight.range can be supplied to specify whether the range
of weights in each group should be displayed.
Fixed a bug in which the output of bread() was off
by a factor of -1. This doesn’t affect its use in
sandwich::sandwich().
Fixed a bug in which bag.fraction for
method = "gbm" with binary and multi-category treatments
had a default of .5 instead of the stated 1.
Fixed bugs relating to method = "optweight";
sampling weights and the fractional weighted bootstrap can now be used
with it.
Fixed bugs related to coxph_weightit() with aliased
coefficients and multi-state models. Thanks to Mads Jeppe
Tarp-Johansen.
Fixed a bug in which messages about assuming which treatment level was “treated” were printed many times, including when bootstrapping.
Fixed a bug in which CBPS for continuous treatments would not yield correct balance in the presence of some extreme density estimates.
Typo fixes in vignettes and documentation.
WeightIt 1.4.0Entropy balancing works slightly differently when sampling
weights are supplied. The negative entropy between the estimated weights
and the product of the sampling weights and base weights (if any) is now
the quantity minimized in the optimization. Previously, the negative
entropy between the product of the sampling weights and estimated
weights and the base weights was minimized. The new behavior ensures
entropy balancing is consistent with mathematically equivalent methods
when it ought to be (i.e., CBPS and IPT for the ATT) and prevents
counter-intuitive results, like that the ESS after weighting could be
larger than that before weighting. Note this will cause results to
differ between this and previous versions of
WeightIt.
With method = "cbps", estimand can now
be set to "ATO" for binary and multi-category treatments.
For binary treatments with the default link, this will yield identical
weights to using method = "glm" with
estimand = "ATO".
Two new links can be supplied with method = "glm",
"cbps", and "ipt": "loglog" for
the log-log link and "clog" for the complementary log link.
The link argument can also now be supplied as a
link-glm object (e.g., the output of a call to
make.link()). This allows for more flexibility in the link
function used to estimate the propensity score.
A new solver argument can be supplied to
weightit() with method = "ebal" and with
method = "cbps" with over = FALSE (the
default); this argument controls whether to use
rootSolve::multiroot() or stats::optim() solve
the optimization problem for the weights. multiroot() is
used by default when rootSolve is installed as it is
quicker and more accurate.
For some methods, analytic formulas for the derivatives used in M-estimation are now used instead of relying on numeric differentiation. This increases speed and accuracy in computing standard errors that adjust for estimation of the weights. The formulas are benchmarked against numeric differentiation results.
Added new method argument to
calibrate() to support isotonic regression calibration as
described by van der Laan el
al. (2024).
Added clearer error and warning messages to several functions,
most notably glm_weightit() and friends, when missing
values are present in the model variables.
The update() method for glm_weightit
objects and friends is now a bit more sophisticated. When
data or s.weights are supplied, the
weightit object (if any) is refit before refitting the
glm_weightit model. This makes it easy to performing
bootstrapping by simply calling update() on a fitted object
with a new dataset or bootstrap weights.
The estfun() and bread() methods for
glm_weightit objects and friends now correctly extract the
estimating function and bread matrices to be used when computing the
sandwich covariance matrix using sandwich::sandwich().
estfun() (and thereby sandwich()) have an
optional asympt argument, which, controls whether the
asymptotic covariance matrix accounting for estimation of the weights is
used.
Improved estimation and convergence for
method = "cbps". In particular, for the over-identified
CBPS, the generalized inverse is now used only when the GMM weight
matrix is singular. Previously, it was always used.
Improved processing of estimand and
focal for binary treatments. weightit() is now
better at guessing which level of the treatment is considered “treated”,
and focal can be used to identify the focal group when
requesting the ATT or ATC. (#77)
Improved estimation of asymptotic and HC0 covariance matrices
after glm_weightit() and friends.
When using method = "super" with
SL.method = "method.balance", a new algorithm is used to
compute the optimal combination of predictions, which should yield
better performance. This may cause results to differ from past
versions.
Fixed a bug when using the fractional weighted bootstrap with
glm_weightit() and friends after entropy balancing. This
bug is still present with optimization-based weighting (i.e.,
method = "optweight"), so the fractional weighted bootstrap
is no longer allowed with this method.
Sampling weights can no longer be used with
method = "optweight" until a bug is sorted out.
Performance enhancements.
Added new tests for CBPS, IPT, and entropy balancing.
Updated the documentation for msmdata. Thanks to
@larsentom.
(#79)
WeightIt 1.3.2Fixes to tests for CRAN.
Improvements to weight calculation for continuous treatments with small densities.
WeightIt 1.3.1vcov(), summary(),
anova(), and confint() for
glm_weightit objects (and their relatives) now have a
vcov argument that can be used to specify how the variance
matrix is computed. This makes it possible to compute a variance matrix
different from the one specified in the model fitting call without
having to refit the model. anova() now displays which
variance matrix was used.
Added update() methods for
glm_weightit, multinom_weightit,
ordinal_weightit, and coxph_weightit objects
to update the model formula, dataset, or variance matrix. Updating the
dataset also refits the weightit object included, if
any.
anova() for glm_weightit objects gets
its own help page at help("anova.glm_weightit()").
Changed defaults with missing = "saem" for binary
and multi-category treatments to bypass a bug in misaem
code. (#71)
Preemptively fixed some bugs related to the use of
missing, including when missing is used with
by.
The missingness method (if any) is now included in the output of
weightit(), weightitMSM(), and
weightit.fit() and is printed when using the
print() method for these objects.
When missing = "saem", using
vcov = "FWB" in glm_weightit(), etc., now
appropriately results in an error. (#71)
model.matrix.ordinal_weightit() now excludes the
(Intercept) column.
Fixed a bug with predict.multinom_weightit() and
predict.ordinal_weightit() when the outcome was not
included in newdata.
Typo fixes in documentation.
WeightIt 1.3.0Added anova() methods for glm_weightit,
multinom_weightit, ordinal_weightit, and
coxph_weightit objects to perform Wald tests for comparing
nested models. The models do not have to be symbolically
nested.
Added the new user-facing object .weightit_methods,
which contains information on each method and the options allowed with
it. This is used within WeightIt for checking arguments but
can also be used by other package developers who call functions in
WeightIt. See help(".weightit_methods") for
details.
plot.weightit() can be used with
method = "optweight" to display the dual
variables.
missing no longer allows partial matching.
moments can now be set to 0 when
quantile is supplied to ensure balance on the quantiles
without the moments for the methods that accepts quantiles.
Thanks to @BERENZ for
the suggestion.
For ordinal_weightit objects, summary()
now has the option to omit thresholds from the output.
Fixed a bug in ordinal_weightit() where the Hessian
(and therefore the HC0 robust variance) were calculated incorrectly when
come coefficients were aliased (i.e., due to linearly dependent
predictors).
Fixed a bug in print.summary.glm_weightit() when
confidence intervals were requested. A new printing function is used
that produces slightly nicer tables.
Fixes to vignettes and tests to satisfy CRAN checks.
Minor bug, performance, and readability fixes.
WeightIt 1.2.0Added two new functions, multinom_weightit() and
ordinal_weightit() for multinomial logistic regression and
ordinal regression with capabilities to estimate a covariance matrix
that accounts for estimation of the weights using M-estimation.
Previously, multinomial logistic regression could be requested using
glm_weightit() with family = "multinomial";
this has been deprecated.
M-estimation can now be used for weighting with ordinal
regression for weights with multi-category ordered treatments with
method = "glm".
M-estimation can now be used with bias-reduced regression as
implemented in brglm2 for the propensity score
(method = "glm" with link = "br.{.}") and for
the outcome model (glm_weightit() with
br = TRUE). Thanks to Ioannis Kosmidis for supplying some
starter code to implement this.
For any weighting methods with continuous treatments that support
a density argument to specify the numerator and denominator
densities of the weights, density can now be specified as
"kernel" to request kernel density estimation. Previously,
this was requested by setting use.kernel = TRUE, which is
now deprecated.
Standard errors are now correctly computed when an offset is
included in glm_weightit(). Thanks to @zeynepbaskurt.
(#63)
Improved robustness of get_w_from_ps() to propensity
scores of 0 and 1.
Updates to weightit() with
method = "gbm":
use.offset is now tunable.class.stratify.cv is now set to
TRUE by default to stratify on treatment.plot() can be used on the output of
weightit() to display the results of the tuning process;
see help("plot.weightit") for details.distribution was not included in the
output when tuned.When using weightit() with
method = "super" for binary and multi-category treatments,
cross-validation now stratifies on treatment, as recommended by Phillips et
al. (2023).
Fixed a bug and clarified some error messages when using ordered
treatments with method = "glm". Thanks to Steve Worthington
for pointing them out.
Updated the help page of get_w_from_ps() to include
formulas for the weights.
WeightIt 1.1.0Added a new function, coxph_weightit(), for fitting
Cox proportional hazards models in the weighted sample, with the option
of accounting for estimation of the weights in computing standard errors
via bootstrapping. This function uses the summary() and
print() methods for glm_weightit objects,
which are different from those for coxph objects.
glm_weightit() gets a new print()
method that omits some invalid statistics displayed by the
print() method for glm objects and displays
the type of standard error estimated.
summary.glm_weightit() (which is also used for
coxph_weightit objects) gets a new argument,
transform, which can be used to transform the displayed
coefficients and confidence interval bounds (if requested), e.g., by
exponentiating them.
M-estimation is now supported for method = "glm"
with continuous treatments.
A new estimator is now used for method = "cbps" with
longitudinal treatments (i.e., using weightitMSM()).
Previously, the weights from CBPS applied to each time point were
multiplied together. Now, balance at all time points is optimized using
a single set of weights. This implementation is close to that described
by Huffman and van
Gameren (2018), not that of Imai and Ratkovic
(2015).
A new estimator is now used for method = "cbps" with
continuous treatments. The unconditional mean and variance are now
included as parameters to be estimated. For the just-identified CBPS,
this will typically improve balance, but results will depart from those
found using CBPS::CBPS().
For point treatments (i.e., using weightit()), the
stabilize argument has some new behavior. It can now be be
specified as a formula, and the stabilization factor is estimated
separately and included in the M-estimation if allowed. It can now only
be used when estimand = "ATE" (weights for other estimands
should not be stabilized).
For binary treatments with method = "glm",
link can now be specified as "flic" or
"flac" to use Firth corrected logistic regression as
implemented in the logistf package.
With method = "gbm", an error is now thrown if
criterion (formerly known as stop.method) is
supplied as anything other than a string.
For binary and continuous treatments with
method = "gbm", a new argument, use.offset,
can be supplied, which, if TRUE, uses the linear predictor
from a generalized linear model as an offset in the boosting model,
which can improve performance.
Added a section on conducting moderation analysis to the
estimating effect vignette
(vignette("estimating-effects")).
Fixed a bug when using M-estimation for sequential treatments
with weightitMSM() and stabilize = TRUE.
Standard errors incorrectly accounted for estimation of the
stabilization factor; they are now correct.
Fixed a bug when using method = "ipt" for the
ATE.
Fixed a bug when some coefficients were aliased for
glm_weightit(). Thanks to @kkwi5241.
Updated kernel balancing example in
method_user.
Improved warnings and errors for bad models throughout the package.
WeightIt 1.0.0Added a new function, glm_weightit() (along with
wrapper lm_weightit()) and associated methods for fitting
generalized linear models in the weighted sample, with the option of
accounting for estimation of the weights in computing standard errors
via M-estimation or two forms of bootstrapping.
glm_weightit() also supports multinomial logistic
regression in addition to all models supported by glm().
Cluster-robust standard errors are supported, and output is compatible
with any functions that accept glm() objects. Not all
weighting methods support M-estimation, but for those that do, a new
component is added to the weightit output object.
Currently, GLM propensity scores, entropy balancing, just-identified
CBPS, and inverse probability tilting (described below) support
M-estimation-based standard errors with
glm_weightit().
Added inverse probability tilting (IPT) as described by Graham,
Pinto, and Egel (2012), which can be requested by setting
method = "ipt". This is similar to entropy balancing and
CBPS in that it enforces exact balance and yields a propensity score,
but has some theoretical advantages to both methods. IPT does not rely
on any other packages and runs very quickly.
Estimating covariate balancing propensity score weights (i.e.,
method = "cbps") no longer depends on the CBPS
package. The default is now the just-identified versions of the method;
the over-identified version can be requested by setting
over = TRUE. The ATT for multi-category treatments is now
supported, as are arbitrary numbers of treatment groups
(CBPS only natively support up to 4 groups and only the ATE
for multi-category treatments). For binary treatments, generalized
linear models other than logistic regression are now supported (e.g.,
probit or Poisson regression).
New function calibrate() to apply Platt scaling to
calibrate propensity scores as recommended by Gutman et
al. (2024).
A new argument quantile can be supplied to
weightit() with all the methods that accept
moments and int ("ebal",
"cbps", "ipt", "npcbps",
"optweight", and "energy"). This allows one to
request balance on the quantiles of the covariates, which can add some
robustness as demonstrated by Beręsewicz (2023).
as.weightit() now has a method for
weightit.fit objects, which now have additional components
included in the output.
trim() now has a drop argument; setting
to TRUE sets the weights of all trimmed units to 0
(effectively dropping them).
When using weightit() with a continuous treatment
and a method that estimates the generalized propensity
score (e.g., "glm", "gbm",
"super"), sampling weights are now be incorporated into the
density when use.kernel = FALSE (the default) when supplied
to s.weights. Previously they were ignored in calculating
the density, but have always been and remain used in the modeling the
treatment (when allowed).
Fixed a bug when criterion was not specified when
using method = "gbm".
Fixed a bug when ps was supplied for continuous
treatments. Thanks to @taylordunn. (#53)
Warning messages now display immediately rather than at the end of evaluation.
The vignettes have been changed to use a slightly different
estimator for weighted g-computation. The estimated weights are no
longer to be included in the call to avg_comparisons(),
etc.; that is, they are only used to fit the outcome model. This makes
the estimators more consistent with other software, including
teffects ipwra in Stata, and most of the literature on
weighted g-computation. Note this will not effect any estimates for the
ATT or ATC and will only yield at most minor changes for the ATE. For
other estimands (e.g., ATO), the weights are still to be
included.
The word “multinomial” to describe treatments with more than two categories has been replaced with “multi-category” in all documentation and messages.
Transferred all help files to Roxygen and reorganized package scripts.
Reorganization of some functions.
WeightIt 0.14.2Fixed a bug when using estimand = "ATC" with
multi-category treatments. (#47)
Fixed a bug in the Estimating Effects vignette. (#46)
WeightIt 0.14.1cobalt version 4.5.1 or greater is now
required.
Fixed a bug when using balance Super Learner with
cobalt 4.5.1.
Added a section to the Estimating Effects vignette
(vignette("estimating-effects")) on estimating the effect
of a continuous treatment after weighting.
WeightIt 0.14.0Added energy balancing for continuous treatments, requested using
method = "energy", as described in Huling et
al. (2023). These weights minimize the distance covariance between
the treatment and covariates while maintaining representativeness. This
method supports exact balance constraints, distributional balance
constraints, and sampling weights. The implementation is similar to that
in the independenceWeights package. See
?method_energy for details.
Added a new vignette on estimating effects after weighting,
accessible using
vignette("estimating-effects", package = "WeightIt"). The
new workflow relies on the marginaleffects package. The
main vignette (vignette("WeightIt")) has been modernized as
well.
Added a new dataset, msmdata, to demonstrate
capabilities for longitudinal treatments. twang is no
longer a dependency.
Methods that use a balance criterion to select a tuning
parameter, in particular GBM and balance Super Learner, now rely on
cobalt’s bal.init() and
bal.compute() functionality, which adds new balance
criteria. The stop.method argument for these functions has
been renamed to criterion and
help("stop.method") has been removed; the same page is now
available at help("bal.compute", package = "cobalt"), which
describes the additional statistics available. This also fixes some bugs
that were present in some balance criteria.
Renamed method = "ps" to
method = "glm". "ps" continues to work as it
always had for back compatibility. "glm" is a more
descriptive name since many methods use propensity scores; what
distinguishes this method is that it uses generalized linear
models.
Using method = "ebcw" for empirical balancing
calibration weighting is no longer available because the
ATE package has been removed. Use
method = "ebal" for entropy balancing instead, which is
essentially identical.
Updated the trim() documentation to clarify the form
of trimming that is implemented (i.e., winsorizing). Suggested by David
Novgorodsky.
Fixed bugs when some s.weights are equal to zero
with method = "ebal", “cbps", and
"energy". Suggested by @statzhero. (#41)
Improved performance of method = "energy" for the
ATT.
Fixed a bug when using method = "energy" with
by.
With method = "energy", setting
int = TRUE automatically sets moments = 1 if
unspecified.
Errors and warnings have been updated to use
chk.
The missingness indicator approach now imputes the variable median rather than 0 for missing values. This will not change the performance of most methods, but change others, and doesn’t affect balance assessment.
WeightIt 0.13.1For ordinal multi-category 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.
WeightIt 0.13.0Fixed 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".
WeightIt 0.12.0The 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.
WeightIt 0.11.0Added 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.
WeightIt 0.10.2WeightIt 0.10.1With 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.
WeightIt 0.10.0Added 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".
WeightIt 0.9.0Added 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.
WeightIt 0.8.0Formula 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.
WeightIt 0.7.1Fixed 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().
WeightIt 0.7.0Added 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.
WeightIt 0.6.0Bug 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.
WeightIt 0.5.1Fixed 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.
WeightIt 0.5.0Added 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.
WeightIt 0.4.0Added 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".
WeightIt 0.3.2Added 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 potential collinearity when using
ebal or ebcw.
Added a vignette.
WeightIt 0.3.1Edits 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.
WeightIt 0.3.0Added 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.
WeightIt 0.2.0Added 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
WeightIt 0.1.0