get_w_from_ps: Get weights from propensity scores corresponding to different estimands

Description

Given a vector or matrix of propensity scores, output a vector of weights that target the provided estimand.

Usage

get_w_from_ps(ps,
              treat,
              estimand = "ATE",
              focal = NULL,
              treated = NULL)

Arguments

A vector, matrix, or data frame of propensity scores. See Details.

treat

A vector of treatment statuses for each individual. See Details.

estimand

The desired estimand that the weights should target. Current options include "ATE" (average treatment effect), "ATT" (average treatment effect on the treated), "ATC" (average treatment effect on the control), "ATO" (average treatment effect in the overlap), and "ATM" (average treatment effect in the matched sample). See method_ps for additional information and references.

focal

When the estimand is the ATT or ATC, which group should be consider the (focal) "treated" or "control" group, respectively. If not NULL and estimand is not "ATT" or "ATC", estimand will automatically be set to "ATT".

treated

When treatment is binary, the value of treat that is considered the "treated" group (i.e., the group for which the propensity scores are the probability of being in). If NULL, get_w_from_ps will attempt to figure it out on its own using some heuristics. This really only matters when treat has values other than 0 and 1 and when ps is given as a vector or an unnamed single-column matrix or data frame.

Value

A vector of weights.

Details

get_w_from_ps applies the formula for computing weights from propensity scores for the desired estimand. See the References section at method_ps for information on these estimands and the formulas.

ps can be entered a variety of ways. For binary treatments, when ps is entered as a vector or unnamed single-column matrix or data frame, get_w_from_ps has to know which value of treat corresponds to the "treated" group. For 0/1 variables, 1 will be considered treated. For other types of variables, get_w_from_ps will try to figure it out using heuristics, but it's safer to supply an argument to treated. When estimand is "ATT" or "ATC", supplying a value to focal is sufficient (for ATT, focal is the treated group, and for ATC, focal is the control group). When entered as a matrix or data frame, the columns must be named with the levels of the treatment, and it is assumed that each column corresponds to the probability of being in that treatment group. This is the safest way to supply ps unless treat is a 0/1 variable.

For multinomial treatments, ps can be entered as a vector or a matrix or data frame. When entered as a vector, it is assumed the value corresponds to the probability of being in the treatment actually received; this is only possible when the estimand is "ATE". Otherwise, ps must be entered as a named matrix or data frame as described above for binary treatments. When the estimand is "ATT" or "ATC", a value for focal must be specified.

get_w_from_ps is not compatible with continuous treatments.

Examples

Run this code

# NOT RUN {
library("cobalt")
data("lalonde", package = "cobalt")

ps.fit <- glm(treat ~ age + educ + race + married +
                nodegree + re74 + re75, data = lalonde,
              family = binomial)
ps <- ps.fit$fitted.values

w1 <- get_w_from_ps(ps, treat = lalonde$treat,
                    estimand = "ATT")

treatAB <- factor(ifelse(lalonde$treat == 1, "A", "B"))
w2 <- get_w_from_ps(ps, treat = treatAB,
                    estimand = "ATT", focal = "A")
all.equal(w1, w2)
w3 <- get_w_from_ps(ps, treat = treatAB,
                    estimand = "ATT", treated = "A")
all.equal(w1, w3)

#A multinomial example using GBM predicted probabilities
library(gbm)
T3 <- factor(sample(c("A", "B", "C"), nrow(lalonde), replace = TRUE))

gbm.fit <- gbm(T3 ~ age + educ + race + married +
                 nodegree + re74 + re75, data = lalonde,
               distribution = "multinomial", n.trees = 200,
               interaction.depth = 3)
ps.multi <- drop(predict(gbm.fit, type = "response",
                         n.trees = 200))
w <- get_w_from_ps(ps.multi, T3, estimand = "ATE")

# }