Tipping-point analysis for missing-data sensitivity — tipping_point

How much would unobserved outcomes need to differ from observed ones for the treatment effect to become non-significant? The tipping point is the smallest shift \(\delta\) of the estimate that makes the two-sided 5\ \(\delta^* = \hat\tau - \mathrm{sign}(\hat\tau) z_{0.975} se\) (0 when the estimate is already non-significant); the grid reports adjusted estimates and p-values over delta_range. For the unmeasured-confounder tipping point see morie_sensitivity_tipping_point.

Usage

tipping_point_analysis(
  estimate,
  se,
  n_treated,
  n_control,
  delta_range = NULL,
  outcome_type = "continuous"
)

Arguments

estimate: Observed treatment effect.
se: Standard error of the estimate.
n_treated: Number of treated units.
n_control: Number of control units.
delta_range: Numeric vector of bias parameters (default seq(-3|est|, 3|est|, length.out = 101)).
outcome_type: "continuous" or "binary"; for a binary outcome the default delta grid is capped at the unit interval.

Value

A morie_tipping_point named-list.

Examples

str(tipping_point_analysis(0.5, 0.15, n_treated = 100, n_control = 100),
    max.level = 1)
#> List of 6
#>  $ delta_values      : num [1:101] -1.5 -1.47 -1.44 -1.41 -1.38 -1.35 -1.32 -1.29 -1.26 -1.23 ...
#>  $ adjusted_estimates: num [1:101] 2 1.97 1.94 1.91 1.88 1.85 1.82 1.79 1.76 1.73 ...
#>  $ adjusted_p_values : num [1:101] 1.48e-40 2.12e-39 2.92e-38 3.86e-37 4.91e-36 ...
#>  $ tipping_point     : num 0.206
#>  $ original_estimate : num 0.5
#>  $ interpretation    : chr "The observed estimate (0.5000) becomes non-significant when outcomes for missing data differ by delta = 0.2060."| __truncated__
#>  - attr(*, "class")= chr [1:2] "morie_tipping_point" "list"