- Change iterative targeting procedures in
est_tml()
to useglm2::glm2
instead ofstats::glm
to avoid issues with erratic IRLS in the latter; see https://journal.r-project.org/archive/2011/RJ-2011-012/ for details. - Changed default propensity score truncation bounds specified in
g_bounds
by an order of magnitude, fromc(0.001, 0.999)
toc(0.01, 0.99)
, to mitigate potential stability issues.
- Fixes bug in weighted TMLEs introduced during prior update to
est_tml()
.
- Added support for a semiparametric correction for outcome-dependent two-phase sampling designs with known or estimated sampling weights.
- Tightened sanity checks for estimation of natural direct and indirect effects
by requiring that EIF scores related to intermediate confounders uniformly be
zero when
Z = NULL
is specified. - The above required minor changes to
est_tml()
so as to avoid fluctuation models for the TMLE step from updating the intermediate confounder nuisance components (i.e.,q_prime_Z_one
,q_prime_Z_natural
) and causing numerical issues that violate the above internal checks.
- For user clarity, the name of the argument for providing externally computed
observation-level weights has changed (from
ext_weights
) tosvy_weights
. - Support for the natural direct and indirect effects has been added, requiring
the addition of the new internal argument
effect_type
across functions for estimation, includingcv_eif()
,est_onestep()
, andest_tml()
. WhenZ = NULL
is set inmedoutcon()
, a natural effect estimate corresponding to the argumenteffect
is returned instead of an interventional effect. - The
summary()
andprint()
methods have been updated to allow handling of natural effects and counterfactual means under arbitrary contrasts.
- An initial public release of this package, version 0.1.0, which includes support for external observation-level weights.