Assuming treatment never hurts, y(1) >= y(0) for every unit, pins
the lower bound at exactly zero and leaves only the upper bound to be
estimated. Unlike the worst-case bound, the arm bounds use every
(y, D) pair rather than only the pairs in the arm of interest, so
they stay informative even when one arm is nearly empty.
Details
Formula: \(upper = [E(y | D = 1) P(D = 1) + y_max P(D = 0)] -
[E(y | D = 0) P(D = 0) + y_min P(D = 1)]\), lower = 0.
References
Manski, C. F. (1997). Monotone treatment response. Econometrica 65(6), 1311-1334. doi:10.2307/2171738 . Binary-treatment form as equation (2.13) and the paragraph beneath it in Molinari, F. (2021), Handbook of Econometrics 7A (arXiv:2004.11751 pp. 18-19).
Examples
set.seed(1)
Bndpos(y = rnorm(40), D = rbinom(40, 1, 0.5), y_max = 3)
#> $lower
#> [1] 0
#>
#> $upper
#> [1] 2.643278
#>
#> $width
#> [1] 2.643278
#>
#> $estimate
#> [1] 1.321639
#>
#> $p_treated
#> [1] 0.45
#>
#> $n
#> [1] 40
#>
#> $method
#> [1] "Positive-only treatment bound"
#>
#> attr(,"class")
#> [1] "morie_rich_result" "list"
