attrition

First release. The package implements the estimators of Coppock, Gerber, Green, and Kern (2017) for randomized experiments with nonignorable missing outcomes.

Estimators

• estimator_ev(): worst-case (Manski 1990) bounds from a single round of data collection, with a joint Imbens-Manski (2004) confidence interval.
• estimator_ds(): the double-sampling bounds of the paper, with its analytic variance. A random sample of the initial nonrespondents is pursued a second time, and only the subjects who refuse twice get worst-case treatment.
• estimator_ds_sens() and sensitivity_ds(): the interpolation between worst-case bounds and ignorability, indexed by delta, the share of follow-up nonrespondents left unmodeled, and a sweep over delta for delta*, the smallest value at which the confidence interval reaches zero.
• estimator_trim(): trimming bounds on the effect among always-reporters. The design and the selection assumption are separate choices. Which response arguments are supplied picks the design, single sample (R) or double sampling (R1, Attempt, R2); monotonicity picks the assumption, "treatment_increases_response" (Lee 2009), "treatment_decreases_response", or "none", which gives the sharp bounds of Imai (2008, Proposition 1) under random assignment alone. All four combinations estimate. Standard errors come from Lee (2009, Proposition 3) where that derivation applies, which is the single-sample, one-group-trimmed case, and from a bootstrap resampled within treatment arm everywhere else; asking for analytic standard errors where they do not apply is an error rather than a silent substitution.
• estimator_ev(), estimator_ds(), estimator_ds_sens() and sensitivity_ds() take a strata argument for poststratification on a discrete covariate, which targets the same identification region and estimates it more precisely.

Output

• Every estimator returns the same six named elements in the same order, under broom's names: estimate_lower and estimate_upper, the two ends of the identification region; std.error_lower and std.error_upper, their standard errors; and conf.low and conf.high, the joint Imbens-Manski interval. estimator_trim() follows them with the intermediate quantities of the path taken.
• print(), summary() and tidy() methods for every result. summary() names the estimand and the assumptions that produced the numbers; for trimming bounds that includes the design, the direction assumed, how much of which group was trimmed, and where the standard errors came from. tidy() returns a three-row tibble, a bounds row carrying the whole vector and a row per endpoint, so DeclareDesign::declare_estimator() can select an endpoint with term. sensitivity_ds() returns a classed list whose print() reports delta* and whose tidy() returns the bounds at every value of delta.
• A formula interface on every estimator, estimator_ds(Y ~ Z, R1 = "R1", Attempt = "Attempt", R2 = "R2", ...), which is what declare_estimator(.method = ...) expects. Response, attempt and stratification arguments accept an unquoted column name, a quoted string, or a one-sided formula.

Guards