
Imbens-Manski confidence interval for a partially identified parameter
Source:R/bounds_confidence.R
morie_bounds_confidence.RdGiven estimated bounds \([\hat L, \hat U]\) with standard errors, the
interval \([\hat L - c\,\sigma_L,\; \hat U + c\,\sigma_U]\) covers every
point of the identification region with probability \(1-\alpha\)
(uniformly over the region's width) when \(c\) solves
\(\Phi(c + \Delta) - \Phi(-c) = 1 - \alpha\) with
\(\Delta = (\hat U - \hat L)/\max(\sigma_L, \sigma_U)\) (Imbens & Manski
2004). The cutoff is non-increasing in \(\Delta\)
(Research.P11.im_cutoff_antitone), lies between the one-sided and
the two-sided normal quantiles (im_cutoff_between), and the
two-sided quantile over-covers whenever the region has positive width
(two_sided_overcovers). An interval built to cover the whole
region is at least as conservative for the parameter
(region_coverage_le); the function reports both. The cutoff the
root finder returns exists and is unique: the coverage is continuous and
strictly increasing in \(c\) with limits \(-1\) and \(1\)
(im_cutoff_exists, im_cutoff_unique).
Arguments
- lower, upper
Estimated bounds (e.g. from
morie_sentence_effect_boundswith bootstrap standard errors).- se_lower, se_upper
Their standard errors, positive.
- level
Confidence level.
Value
A list with cutoff (\(c\)), delta, interval
(the Imbens-Manski interval), region_interval (the interval that
covers the whole region, using the two-sided quantile), z_one_sided,
z_two_sided, coverage_two_sided (what the two-sided cutoff
would deliver for a point at one end, above level), and theorems.
Examples
morie_bounds_confidence(lower = 0.10, upper = 0.35, se_lower = 0.03, se_upper = 0.04)
#> $cutoff
#> [1] 1.644854
#>
#> $delta
#> [1] 6.25
#>
#> $level
#> [1] 0.95
#>
#> $interval
#> lower upper
#> 0.05065439 0.41579415
#>
#> $region_interval
#> lower upper
#> 0.04120108 0.42839856
#>
#> $z_one_sided
#> [1] 1.644854
#>
#> $z_two_sided
#> [1] 1.959964
#>
#> $coverage_two_sided
#> [1] 0.975
#>
#> $theorems
#> [1] "Research.P11.region_coverage_le" "Research.P11.coverage_strictMono_c"
#> [3] "Research.P11.im_cutoff_antitone" "Research.P11.im_cutoff_between"
#> [5] "Research.P11.two_sided_overcovers" "Research.P11.im_cutoff_exists"
#> [7] "Research.P11.im_cutoff_unique"
#>