The version involved a full rewrite and may not be backward compatible with prior versions. Basic functionality has not changed, but some of the more advanced functionality has changed.
Added a new function, optweightMV(), for performing
weighting for multivariate treatments. This takes the place of the old
functionality of optweight() to take in a list of formulas.
Documentation and other supporting text has been updated to emphasize
that the purpose of this function is for multivariate treatments, not
longitudinal treatments as previously documented. Note that
optweight() now throws an error if a list of formulas is
supplied. A lower-level version, optweightMV.fit(), is also
available.
check.tols() and check.targets() have
been renamed process_tols() and
process_targets(), respectively, and have some new
functionality. The stop argument has been removed.
process_targets() now accepts s.weights to
compute sampling-weighted target means.
norm can be set to "entropy" in
optweight(), etc., which minimizes the negative entropy of
the weights. This is equivalent to entropy balancing, which is
implemented more efficiently in WeightIt, but has support for
inexact balance and multivariate treatments.
norm can also be set to "log" in
optweight(), etc., which minimizes the sum of the negative
log of the weights. This is equivalent to nonparametric covariate
balancing propensity score (npCBPS) weighting, a version of which is
implemented in CBPS. The implementation here has support for
inexact balance and multivariate treatments.
Negative values are now allowed for the min.w
argument.
Different solvers can be used by supplying an argument to
solver. See ?optweight.fit() for defaults and
allowable options.
optweight.svy(), process_tols(), and
process_targets() can all be supplied with without a
formula, which uses all variables in the supplied data.
Documentation is now in roxygen2.
Added a new vignette
(vignette("optweight")).
Added b.weights argument to supply base weights.
When supplied, rather than minimizing the variance of the weights, the
squared distance from each base weight is minimized, mirroring the
functionality of the base.weights argument in
ebalance for entropy balancing. All norms now support base
weights. Omitting base weights is equivalent to setting them equal to
1.
Optimization proceeds differently with s.weights
supplied. First, the L\(\infinity\)
cannot be used with sampling weights. Second, the norm minimized is the
weighted norm of the difference between the estimated weights and base
weights, with the estimated weights not incorporating the sampling
weights. That is, the L2 norm minimizes \(\sum_i s_i(w_i-b_i)^2\), and the L1 norm
minimizes \(\sum_i s_i|w_i-b_i|\),
where \(s_i\) is the sampling weight
for unit \(i\), \(b_i\) is the base weight (1 by default),
and \(w_i\) are the weights to be
estimated. The weights used in the balance constraints (and ultimately
in effect estimation) are \(w^*_i=s_i
w_i\). An implication of this is that the ESS of the \(w^*_i\) is not maximized with the L2 norm.
This also ensures that the weighted bootstrap correctly accounts for
estimation of the weights.
summary() now displays the L2, L1, and L\(\infinity\) norms and the relative entropy
between the estimated weights and the base weights, and the number of
weights estimated to be 0. The L2 and L1 norms and relative entropy are
weighted by the s.weights if present.
polish is now TRUE by default for norms
other than "entropy"; this slightly improves
estimation.
Some default arguments to the solvers have changed.
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.
The returned covariates are now those without any transformations.
Updated the README.
Added a new logo.
The rosqp package is now osqp, and is faster with fewer bugs.
If focal is set, the estimand is automatically
changed to "ATT". In the past, focal would be
ignored unless estimand = "ATT".
Fixed some bugs with processing formula inputs. In particular,
functions can be used inside lapply() loops and nested
functions more gracefully.
Other bugs fixes and small changes.
Changed default min.w in
optweight.fit() and optweight.svy.fit() to
1E-8 from 0. This ensures all weights are nonzero, which can reduce bugs
in other functions that require nonzero weights (e.g,
jtools::summ() and survey::svyglm()).
Fixed warning that would occur when interactions were present in
the model formula in optweight().
Stable balancing weights have been discovered to be invalid for
longitudinal treatments, so attempting to use optweight()or
optweight.fit() with longitudinal treatments will now
produce an error. This can be overridden by setting
force = TRUE, though this is not recommended until further
research is done.
Added optweight.svy() and associated methods and
functions for estimating survey weights using optimization. These
weights when applied to the sample will yield a sample whose covariate
means are equal (within a specified tolerance) to given target
values.
Minor changes to check.targets(). It will now
produce the covariate means when the targets argument is
empty and will produce the previous empty output, a named vector of
NAs, when targets = NULL.
Changes to how dual variables are processed and displayed. Now,
each dual variable coming from optweight() represents the
change in the objective function corresponding to a 1-unit change in
tols. The reported duals are the sum of all the duals
affected by the constraint, so you can now reliably predict the change
in the objective function from a change in tols (it was
obscured and error-prone previously). The distinction between targeting
duals and balance duals is maintained.