Balancing weights for a binary, multi-category or continuous treatment
under the method names and rules of WeightIt (Greifer 2024), with
no call into it: "glm" (a propensity model: any binomial link,
maximum likelihood, Firth's bias reduction, ridge, lasso or an elastic
net; multinomial or cumulative-link for more than two levels; the
density-ratio generalised propensity weight for a continuous treatment),
"ebal" (entropy balancing, exact moment balance), "ipt"
(inverse probability tilting) and "cbps" (the covariate balancing
propensity score, just- or over-identified). The weights each estimand
takes from a propensity score follow WeightIt's get_w_from_ps:
"ATE", "ATT", "ATC", "ATO", "ATM" and
"ATOS" for a binary treatment, "ATE", "ATT",
"ATC", "ATO" and "ATM" for a multi-category one.
Usage
morie_weightit(
data,
treatment,
covariates,
method = c("glm", "ebal", "ipt", "cbps"),
estimand = "ATE",
focal = NULL,
link = "logit",
ps_model = "mle",
lambda = 1,
alpha = 1,
sampling_weights = NULL,
stabilize = FALSE,
moments = 1L,
int = FALSE,
over = FALSE,
density = "normal",
bw = "nrd0",
adjust = 1,
ordinal = FALSE,
d_moments = 1L
)
# S3 method for class 'morie_weightit'
print(x, ...)Arguments
- data
A data frame.
- treatment
Name of the treatment column. Two distinct values make a binary treatment; a factor, character, or an integer-valued column with at most ten levels a multi-category one; any other numeric column a continuous one.
- covariates
Character vector of covariate names; non-numeric columns enter as dummies.
- method
One of
"glm","ebal","ipt","cbps".- estimand
The target: see Details for what each method and treatment type accept. Ignored for a continuous treatment.
- focal
For a multi-category treatment with
estimand"ATT"or"ATC", the level the others are weighted to.- link
The propensity model's link (
"logit","probit","cloglog","loglog","cauchit","log","clog","identity"); for a continuous treatment the Gaussian link.- ps_model
For a binary
"glm":"mle","firth","ridge","lasso"or"elasticnet".- lambda, alpha
The penalty and the elastic-net mixing weight of a penalised propensity model.
- sampling_weights
Optional sampling weights (a numeric vector or a column name); the balancing weights multiply them.
- stabilize
For a binary or multi-category treatment, multiply the weights by the marginal probability of the unit's level. A continuous treatment's weights are stabilised already.
- moments, int
For
"ebal", balance powers up tomomentsof the non-binary covariates and, withint, all pairwise products.- over
For
"cbps", the over-identified (GMM) estimator.- density, bw, adjust
For a continuous
"glm", the density of the standardised residuals ("normal","laplace","t_<df>","kernel"or a function) and the kernel bandwidth arguments.- ordinal
Treat a multi-category treatment as ordered and fit a cumulative-link model.
- d_moments
For a continuous
"ebal", balance powers up to this order of the treatment.- x
A
morie_weightitobject.- ...
Ignored.
Value
A list of class morie_weightit: weights; ps
(the propensity scores, a matrix of them for a multi-category
treatment, the fitted treatment for a continuous one, NULL for
"ebal"); treatment_type; method; estimand;
focal; ess (the effective sample size); and, for an
over-identified "cbps", objective, the GMM criterion as a
function of a propensity score vector.
Details
Each method's estimands: "glm" every one above;
"ebal" "ATE", "ATT", "ATC"; "ipt"
"ATE", "ATT", "ATC"; "cbps" (binary only)
"ATE", "ATT", "ATC", "ATO". A continuous
treatment takes "glm" or "ebal".
References
Greifer N (2024). WeightIt: Weighting for Covariate Balance in Observational Studies. Hainmueller J (2012). Entropy balancing for causal effects. Political Analysis 20(1). Graham BS, Pinto CC, Egel D (2012). Inverse probability tilting. Review of Economic Studies 79(3). Imai K, Ratkovic M (2014). Covariate balancing propensity score. JRSS-B 76(1). Kosmidis I, Firth D (2009). Bias reduction in exponential family nonlinear models. Biometrika 96(4).