Skip to contents

Changelog

Source: NEWS.md

attrition 1.0.0

First release.

New features

  • Formula interface on all estimators: estimator_ev(Y ~ Z, R = "R", ...), estimator_ds(Y ~ Z, R1 = "R1", Attempt = "Attempt", R2 = "R2", ...), and likewise for estimator_trim(), estimator_ds_sens(), and sensitivity_ds(). Response, attempt, and stratification arguments accept either an unquoted column name or a quoted string. The formula form is what DeclareDesign::declare_estimator(.method = ...) expects.

  • tidy() methods for all estimator output classes (attrition_bounds, attrition_trim) via the generics package. Each method returns a three-row tibble with rows "bounds", "lower_bound", and "upper_bound", suitable for use with DeclareDesign. The "bounds" row carries the joint Imbens-Manski confidence interval in conf.low/conf.high and the bound point estimates in estimate.low/estimate.high; the individual rows carry point estimates and standard errors selectable via the term argument.

  • S3 classes on all estimator return values: estimator_ds() returns class c("attrition_ds", "attrition_bounds"), estimator_ev() returns c("attrition_ev", "attrition_bounds"), estimator_ds_sens() returns c("attrition_ds_sens", "attrition_bounds"), and estimator_trim() returns c("attrition_trim").

  • Stratified estimators now index results by name rather than position, preventing silent errors if the output vector order ever changes.

  • sensitivity_ds() no longer calls require() at runtime; all dependencies are declared via @importFrom.

  • reshape2 dependency replaced by tibble; generics added for tidy re-export.

  • estimator_trim() gains standard errors and confidence intervals, via a new se argument. se = "analytic" (the default) uses the closed-form asymptotic variance of Lee (2009), Proposition 3, and covers the single-stage R path. se = "bootstrap" resamples units within treatment arm and covers both that path and the weighted double-sampling path, for which no analytic variance exists in the literature. se = "none" returns bounds alone. The bound standard errors feed the existing Imbens-Manski machinery, so tidy.attrition_trim() now returns std.error, conf.low and conf.high instead of NA throughout.

    The analytic variance was validated three ways: Lee’s published term for the estimated trimming proportion agrees to machine precision with the algebraically distinct form used by Tauchmann’s Stata leebounds; the mean analytic standard error tracks the Monte Carlo standard deviation of the estimator across sample sizes from 1,000 to 64,000; and a conventional 95% interval around each bound endpoint covers the true population endpoint 95.4% and 94.9% of the time.

    Requesting se = "analytic" on the double-sampling path is an error rather than a silent substitution, since Lee’s derivation assumes i.i.d. sampling and trimming of one group only. A zero trimming proportion warns, because the bounds then collapse to a point on the boundary of the parameter space and Lee’s interior-point condition fails.

  • Worked examples on every exported function.

  • A second vignette, “Drawing the extreme value bounds”, which draws the imputation that estimator_ev() averages over using vayr and checks that a difference in means inside each imputed scenario recovers the two bound estimates. It also shows why the Imbens-Manski interval is narrower than stacking a confidence interval on each bound. vayr and estimatr are suggested packages and the vignette does not evaluate without them.

Bug fixes

  • The Imbens-Manski critical value was found by minimizing an absolute value with optim(..., method = "Brent", lower = 1, upper = 2). That interval only covers alpha near 0.05, so at other significance levels the optimizer returned a boundary value and the confidence intervals were silently wrong: at alpha = 0.01 it returned 2.000 against a correct 2.326, and at alpha = 0.001 it returned 2.000 against a correct 3.090. The search interval is now derived from alpha, and the root is found with uniroot() on the signed coverage excess rather than by minimizing its absolute value. Bound point estimates and variances are unaffected.

  • estimator_trim() reported every failure inside trimming_bounds() as a monotonicity violation, because it caught all errors and returned NA bounds. It now catches only a classed monotonicity condition.

  • The monotonicity violation message named the wrong group. Q < 0 means the treatment group is more likely to be missing than the control group.

  • trimming_bounds() built its weighted CDFs with a loop that counted backwards when a treatment group had one observed outcome, silently appending an NA, and threw an opaque error when a group had none. Both are now cumsum(), with an explicit check for empty groups.

  • sensitivity_ds() could not detect a delta* at the last point of the delta grid, reporting a genuine crossing near delta = 1 as no crossing at all.

  • minY, maxY, alpha, delta, and sims are validated. Previously a reversed minY/maxY returned a lower bound above the upper bound, an assumed support that did not cover the observed outcomes returned bounds that are not bounds, and a delta outside [0, 1] returned NaN variances.

  • estimator_ev(): fixed a copy-paste error where n1_c_s was incorrectly referenced as n1_c_c in the unstratified path.

  • estimator_ds() and estimator_ds_sens(): replaced a commented-out sentinel initialization (-99) with NA_real_ for c1a_t, c1r_t, c2a_t, c2r_t. These arguments were never used and have since been removed from the internal estimators altogether.