Heterogeneous treatment effects (CATE / uplift)

Per-unit conditional average treatment effect \(\tau(x) = E[Y \mid do(T = t_1), X = x] - E[Y \mid do(T = t_0), X = x]\), read in closed form off the fitted joint mixture. Under the model's default "ignorability" assumption this is the contrast of two component-gated conditional means; under "latent_confounder" it is the regime-gated within-class slope (the do-operator). The two coincide when treatment carries no information about the regime beyond X; their difference is proxy_confounding_gap().

Usage

proxy_cate(
  model,
  newdata,
  t1 = 1,
  t0 = 0,
  se = TRUE,
  se_method = c("delta", "mc"),
  level = 0.95,
  B = 200L,
  scale = c("link", "response"),
  threshold = 0.5,
  ...
)

Arguments

model

An uplift_model.

newdata

A data frame carrying the covariate columns.

t1, t0

The treated and control treatment values. Default 1 and 0.

se

Logical – compute standard errors and confidence intervals.

se_method

One of "delta" (closed form, the default) or "mc" (resampling).

level

Confidence level for the interval. Default 0.95.

B

Number of bootstrap refits when se_method = "mc". Default 200.

scale

One of "link" (the latent / continuous scale, the default) or "response". For a binary outcome the response scale reports the effect on the discretised predictive probability P(Y > threshold); for continuous and count outcomes the two scales coincide.

threshold

Decision threshold for the binary discretised predictive. Default 0.5.

...

Forwarded to fit_proxymix() inside the "mc" refits.

Value

A data.table::data.table with columns id, tau, se, ci_lo, ci_hi, overlap_flag.

Details

The default delta-method standard error is the within-component prediction variance, holding the regime gate fixed; it reduces to the ordinary least-squares standard error of the treatment effect at K = 1. Set se_method = "mc" for a resampling standard error that also reflects gate uncertainty.

See also

proxy_decide(), proxy_confounding_gap(), proxy_overlap()

Other decision: fit_uplift(), gmm_cf_mean(), gmm_cf_tail_prob(), gmm_cf_variance(), gmm_counterfactual(), gmm_intervene(), proxy_confounding_gap(), proxy_decide(), proxy_identification_report(), proxy_overlap(), proxy_policy_value(), proxy_predict(), proxy_regime_segments(), proxy_retrospective_uplift(), proxy_uplift(), uplift_identification(), uplift_model()

Examples

set.seed(1)
n <- 400L
x <- stats::rnorm(n)
t <- stats::rbinom(n, 1L, 0.5)
y <- 1 + 0.5 * x + (1 + x) * t + stats::rnorm(n, sd = 0.5)
dat <- data.frame(y = y, t = t, x = x)
m <- fit_uplift(dat, "y", "t", "x", N = 2L, regime = "sample",
                max_iter = 50L, seed = 1L)
proxy_cate(m, newdata = data.frame(x = c(-1, 0, 1)))
#>       id        tau         se      ci_lo     ci_hi overlap_flag
#>    <int>      <num>      <num>      <num>     <num>       <lgcl>
#> 1:     1 0.02760485 0.08010995 -0.1294078 0.1846175        FALSE
#> 2:     2 1.01893392 0.05416744  0.9127677 1.1251002        FALSE
#> 3:     3 2.01026300 0.07624457  1.8608264 2.1596996        FALSE