teeMod
had a non-empty na.action, or the teeMod or
its weights argument had a dichotomy slotas.lmitt() finds StudySpecification
objects passed to
assigned()/a.()/z.()/adopters()
when the object isn’t explicitly passed to as.lmitt()summary.teeMod() is
dof.type=stata, which uses the number of clusters (or
observations, in the absence of clustering) less one. We suggest
specifying dof.type="IK" in summary.teeMod()
calls, which uses a routine adapted from Imbens and Kolesár (2016),
which adapts a routine from Bell and McCaffrey (2002). This routine has
been available in previous versions for summary.teeMod()
calls with type=CR2, but it has been updated so it better
matches the outputs of the dfadjust package maintained by
Kolesár. See the documentation for the internal function
.compute_IK_dof() for references and further details.cov_adj() when covariance adjustment model
formula includes transformations of variablesteeMod objects. CR2 corrections use a new, fast computation
that obviates the need for obtaining the spectral decomposition of
orthocomplements of cluster-specific projection matrices.teeMod objects now optionally
include methodological developments from a forthcoming paper from
Wasserman and Hansen:
type argument of
vcov_tee() determines the correction for the former, while
the cov_adj_rcorrect argument determines the latter. The
findings from the simulation study in the paper inform the default
arguments of vcov_tee() (and thus,
summary.teeMod()): an HC2/CR2 correction for the residuals
of the intent-to-treat effect estimates, and an HC1/CR1 correction or
the residuals from the covariance adjustment model fit.loco_residuals argument of vcov_tee()
indicates the residuals from effect estimation associated with
individuals in the overlapping units should be replaced with residuals
that use a leave-one-out estimate of the covariance adjustment model
parameters.summary.teeMod() now
reflect clustering: for CR0 and CR1 standard errors, the associated
degrees of freedom are one less than the number of clusters used for
estimation; for CR2 standard errors, degrees of freedom are computed
using the approach of Imbens and Kolesár (2016) (see documentation of
.compute_IK_dof() for citation) and leverage the fast
computational routine mentioned aboveStudySpecification are now
optional, though recommended.dichotomy argument of
lmitt()unit_of_assignment(), unitid(),
cluster(), uoa(), block(), and
forcing() no longer fail automatically when passed a
non-numeric or non-character variable. Now, they will first attempt to
convert the variable to a character variable.vcov_tee() no longer fails when
the environment in which vcov_tee() is called differs from
the environment in which the StudySpecification associated
with the teeMod is created.coefficients element, teeMod
objects now report estimates of mean quantities in the control condition
(response and, if applicable, predictions of response). See the
lmitt() man page for further details.etc() (effect of the treatment on controls)
and ato() (overlap-weighted effect) weighting functions.
atc is an alias for etc(), while
olw, owt, and pwt are aliases for
ato().ate()-type functions are called with
data arguments that do not include rows associated with all
units of assignment specified in the StudySpecification
object, the resulting weights reflect assignment probabilities across
all units of assignment in the StudySpecification, not only
those represented in data.teeMod objects to test multiple
outcome variables or different levels of a factor treatment variable can
pass those models to an mmm object from the
multcomp package, then pass the mmm object to
multcomp’s glht() function to obtain standard
errors estimated using vcov_tee(). This requires passing
vcov=vcov_tee to glht(). On a technical note,
**propertee** now “suggests” installing
multcomp.ate() and ett() no
longer produce 0’s or NA’s for StudySpecification objects
created with multiple columns specifying the blocking schemelmitt() fit due to NA’s in the covariates or
treatment assignment*_design is now *_spec
(e.g. rct_design is now rct_spec)Design objects are now StudySpecification
objectsdesign= argument to lmitt() is now
specification=.absorb = TRUE to lmitt without
specifying a block proceeds as if the entire sample is a single
block.dichotomy and moderator
variables in lmitt() could lead to errors due to too long
of a formula.lmitt(), weights calculation functions
ate() and ett(), and assignment vector
generation function assigned() now accept a
dichotomy argument that can be used for studies with
time-varying treatment assignment. The Design object,
unlike before, will not carry information about this dichotomization.
Instead, the information stored there reflecting when units were
assigned to treatment (if they were assigned to treatment) will be
leveraged to create inverse probability of assignment weights and
assignment indicators for datasets that have longitudinal data for the
study units.by
column is used to uniquely identify rows in the covariance adjustment or
effect estimation sample that cannot be distinguished with information
in the Design aloneDesign objects are created with a
tibblecov_adj() does not error with covariance adjustment
models fit with robustbase::glmrob()estfun.teeMod()
to account for a previously missing factor of sqrt(n / n_C) applied to
contributions to the covariance adjustment model estimating
equationsvcov_tee()
clusters units of assignment in small blocks, blocks with only one
treated or control unit, together.vcov_tee() scales estimating equations using different
constants than it did beforevcov_tee() were
liable to change from call to call as a result. This has been
fixed.vcov_tee() matrices lacking
sufficient degrees of freedom for estimation are returned as NA’s rather
than numeric zeros. This is a deviation from the sandwich
package that aims to provide clarity to results that may otherwise
appear as negative diagonal elements of the vcov matrixlmitt() is called with a blocked design and
absorb=TRUE, the block-centered assignment and, if
applicable, moderator and assignment:moderator interaction columns, are
no longer centered on the grand mean of the column. This ensures blocks
that do not satisfy positivity of the assignment variable (or positivity
within a factor level) do not contribute to effect estimationlmitt() now accepts references to formula objectsestfun.teeMod has been
improvedabsorb=TRUE estimates have been corrected in the case
when all observations in a stratum have 0 weights due to only treated or
control units of assignment existing in the stratumvcov_tee() can accept user-created variance estimation
functions that start with the prefix .vcov_; the
type argument should take the rest of the function name as
an inputrobustbase::glmrob) is now accommodatedvcov_tee() will return NA’s for the entries of the
covariance matrix that lack sufficient degrees of freedom for an
estimate. Informative warnings will accompany the matrix, further
indicating which standard errors have been NA’d out.ate() and
ett(), return weights of 0 rather than infinity for blocks
that contain treated units but no control units.SandwichLayer object before calling lmitt() or
called cov_adj() in the offset argument of the
lmitt() call. This has been corrected, and both ways return
the same variance estimates.robustbase::lmrob, are now
accommodated given the reformulated estimating equations in versions
v0.1.1 and later.model.matrix() in certain
cov_adj() calls has been resolved.teeMod objects’ matrix of estimating
equations based on user-specified ID columns or unit of assignment
ID’s.stats::update function can no longer be called on
teeMod objects.teeMod objects now have lmitt_call
slots.summary calls on teeMod objects accept
vcov.type arguments to specify the desired standard error
calculation shown in the output. Acceptable types follow the
documentation for vcov_tee.teeMod objects return more
comprehensible labels for ITT effect outputs.expand.model.frame calls on
teeMod objects and instead use the internal function
.expand.model.frame_teeMod when necessary.absorb argument. Previous versions
did not properly support this functionality. Valid standard errors under
absorption, however, have not been confirmed.
Need a high-speed mirror for your open-source project?
Contact our mirror admin team at info@clientvps.com.
This archive is provided as a free public service to the community.
Proudly supported by infrastructure from VPSPulse , RxServers , BuyNumber , UnitVPS , OffshoreName and secure payment technology by ArionPay.