Skip to contents

Changelog

Source: NEWS.md

powerape 1.11.4

Documentation only: no result moves. The one change in the code is the year in a warning message (below).

  • One website for R and Stata. The package website (https://jespernwulff.github.io/powerape/) now documents the R package and the Stata command of the same name (https://github.com/jespernwulff/powerape-stata) side by side: Get started in R and in Stata, a worked example with five designs, the workflow vignettes with R and Stata code in tabs, a guide matching the R functions to the Stata subcommands, and how both implementations are validated. The tabs exist on the website only; the vignettes installed with the package render as before. The README links to the website and to the Stata command.
  • Vignettes. vignette("minimum-effect") points to the Stata guide. vignette("iv-designs") now says that the Stata margins inconsistency after cfprobit with a fractional first stage was fixed in the 14 Oct 2026 update of StataNow 19.5 (see the 1.11.3 upstream note); the vignette still described it as current.
  • Citation. Hancock and Feng appeared in Psychological Methods, 31(1), 174-199, in 2026; it was cited as 2025. The year is corrected in the DESCRIPTION, ?powerape, ?ape_robust, the README, the minimum-effect vignette (which now gives volume and pages) and the warning that ape_dgp_empirical() and ape_dgp_from_fit() give for pilots with fewer than 500 rows.

powerape 1.11.3

Wording and documentation only: no result changes, and every printed line of a probit world is unchanged.

  • signal under logit. signal is the share of the latent-index variance due to the nuisance covariates with the latent error variance fixed at 1 under both links. Under probit that is the covariates’ McKelvey-Zavoina R-squared; under logit, whose latent error variance is pi^2/3, the same signal describes a weaker world (a McKelvey-Zavoina R-squared of 0.033 at signal = 0.10, 0.115 at 0.30). The documentation called it a latent pseudo-R-squared for both links, and print() and power_statement() did the same. For logit worlds they now name the scale and add the logistic-scale McKelvey-Zavoina R-squared, which for an ape_dgp_from_fit() world built on a logit pilot is the value a fit statistic reports for that pilot. ?ape_dgp gives the conversion from a logit pilot’s R-squared to signal, and the minimum-effect vignette points to it.
  • Upstream note. The Stata margins inconsistency after cfprobit with a fractional-probit first stage (1.4.0 below; confirmed by StataCorp on 28 Aug 2026) was fixed in the 14 Oct 2026 update of StataNow 19.5 (whatsnew item 9); an updated Stata now agrees with the estimation-consistent generalized residuals powerape has always used. Nothing in powerape changes.

powerape 1.11.2

  • The equivalence-claim searches now invert the two-bound power. ape_n(claim = "equivalence") and the equivalence margin of ape_mde() stepped with the one-sided power map z + qnorm(p) that the detect and minimum-effect claims use. The equivalence claim needs both interval bounds inside the margin, so at a truth of 0 the first proposal landed at about 72% of the requirement, the Newton steps oscillated (most searches used all their iterations), and the answer overshot by about 2.4% on average. In roughly 3-5% of runs the exact power at the answer fell more than the confirmation tolerance below the goal, against 0.1-0.4% for the other claims. The proposal, the Newton step, and the confirmation push of both functions now invert the TOST power pnorm(d - z) + pnorm(r d - z) - 1 of both bounds (r is the far bound’s distance from the truth over the near one’s; at a truth of 0 this is z + qnorm((1 + p) / 2)). Measured against exact power in the saturated design (400 emulated runs each), the first proposal now sits within 0.1% of the exact requirement, searches take 1.1 steps on average, and breaches fall to 0-0.2%, at truths of 0 and .01. ape_robust() inherits the fix through its equivalence n_max and its MDE mode. The detect and minimum-effect claims are unchanged bit for bit (checked on seeded ape_n(), ape_mde() and ape_robust() calls, including a sample-size rule for alpha). The equivalence vignette’s required n changes; no number in the companion article comes from an equivalence search.
  • New tests from a mutation pilot. Six plausible one-line defects passed the 1.11.1 suite. New tests close each gap:
    • every analytic delta-method gradient (18 APE, AIE, and control-function Jacobians, both links) equals its finite-difference gradient;
    • the AIE separation check reads all four focal-by-moderator cells on the standard and IV routes, by replaying the engine’s draws;
    • se = "robust" equals sandwich::vcovHC(type = "HC0"), and the cluster sandwich equals vcovCL, to 1e-6 on a tightly converged fit;
    • the confirmation stage’s acceptance rule is tested with a mocked power_once() on both sides of its tolerance;
    • seeded calls leave the caller’s random-number stream untouched;
    • the engine draw realizes a non-exchangeable correlation matrix column by column.
    New regression tests also check the equivalence searches against exact power.

powerape 1.11.1

Documentation only; no code changed, so every result is identical to 1.11.0.

  • Unequal panel lengths do cost power. The panel-designs vignette and the 1.7.0 entry below said that a mean-preserving mixture of panel lengths showed no penalty against a balanced panel of the same average length. That was wrong: the comparison behind it (validation battery V18, 2,000 replications per design, run under the pre-1.8.0 interval convention) could not resolve a penalty of one to two points. The engine prices the design effect for unequal cluster sizes of Eldridge, Ashby & Kerry (2006): under a mixture of 2, 4, and 6 periods with mean 4, its clustered standard error exceeds the balanced panel’s by a factor of 1.0282 against the formula’s 1.0295 (battery V26, which replaces V18’s comparison). The vignette now says so, cites the formula, and names it among the validation anchors.

powerape 1.11.0

Fixes from the September 2026 blindspot audit. Numbers from designs without rare outcomes, small focal groups, or the affected options are unchanged: every seeded result of the article’s worked designs regenerates bit for bit, apart from ape_mde() (fixed below).

  • Separation is now detected, reported, and by default scored as a failed fit (new argument separation = c("fail", "keep") in ape_power(), ape_curve(), ape_n(), ape_mde(), and ape_robust(); DESIGN.md section 16). When a simulated study has a cell that identifies the effect – a level of a binary focal variable, or for AIE designs a focal-by-moderator cell – with no events or no non-events, the maximum-likelihood effect does not exist. Stata’s probit/logit drop the perfect predictor and margins reports the effect as not estimable; R’s glm() instead stops silently at a large finite coefficient whose delta-method standard error collapses to the other cell’s binomial SE. Through 1.10.0 the engine kept such replications, contrary to its documentation, and they usually counted as detections: power was overstated, and required sample sizes understated, in rare-outcome designs with small focal groups (by up to about half in the audit’s unbalanced examples), with failed 0.000 printed throughout. They now count as failed ("fail", the default, matching the documentation and Stata’s default analysis); separation = "keep" restores the 1.10.0 behavior (R’s glm() + marginaleffects analysis, Stata’s asis). Either way results report the share of simulated studies with separation (separated) and the average smallest identifying-cell count of events or non-events (min_cell); print methods and power_statement() state both, and the functions warn when the share exceeds 1% or the smallest cell averages fewer than 10 events. A backup check flags quasi-separated fits the cell check cannot see (a focal-side coefficient with an absurd standardized standard error). The exact-enumeration certificate (tests/testthat/helper-exact.R, battery V1) now enumerates either convention and is exercised where the conventions differ.
  • Panel and IV routes: a continuous focal is now centered in the simulated outcome index, as it already was in calibration, inversion, and truth evaluation. With a focal mean other than 0 the simulation used to run in a world with a different baseline and APE than the ones pinned and reported (e.g. an APE about 8% larger and a baseline of .365 instead of .300 at focal mean 2 on the panel route; for IV-AIE designs the moderator’s main effect was also displaced). Binary focals and mean-zero continuous focals are unaffected bit for bit; shifting the focal’s location now leaves every draw and estimate unchanged (unit-tested to 1e-6).
  • ape_mde() gains direction ("positive" or "negative", defaulting to the sign of the pinned effect): it used to search increases only and silently re-pinned a DGP pinned at a negative effect to a positive MDE. With a binary outcome the direction matters away from a 0.5 baseline (a decrease is harder to detect than an increase of the same size above 0.5, easier below it); the returned mde is signed, and print and power_statement() name the direction. ape_robust(mode = "mde") passes direction through and takes the MDE largest in magnitude as its worst case.
  • ape_mde() re-measures the standard error at its first proposal before searching. The pilot SE was taken at a reference effect (half the distance from the baseline to the nearer bound), and at large n – where the MDE lies far below that reference – the first candidate overshot by 2-4% at a 0.30 baseline (more at rarer baselines) and was accepted within the search tolerance without ever being trimmed. MDE results change accordingly (the confirmation stage still pushes only outward).
  • ape_robust()’s n_max is now the largest requirement across scenarios. It used to be solved in the scenario with the lowest simulated power at the user’s n; when every scenario’s power saturated near 1, that tie went to the first grid row (often the most favorable scenario), so n_max could be understated by about half. Scenarios are now ranked by a criterion that does not saturate (the normal-approximation requirement implied by each scenario’s simulated standard error), ape_n() runs in the least favorable scenario and in any within 5% of it (up to three), and the result reports the scenario that sets n_max (nmax_scenario) and every candidate searched (nmax_candidates). When a scenario’s implied effect sits on or beyond the claim boundary (possible with pin = "coefficients"), n_max is withheld with a warning. Ties in the worst-power scenario are broken by the same criterion. The documented nmax_power = 0.90 default is unchanged; the vignette now says so.
  • ape_dgp_from_fit()’s focal guard is structural. It checked the focal’s variable name while addressing its model-matrix column, so interactions of a logical or factor focal, transformed copies of the focal (I(dose^2) next to dose, log(dose), I(treat * female)), multi-column terms, and offsets were accepted silently – and the pinned “APE” was then a partial effect holding the focal’s other columns fixed, not the APE that margins or marginaleffects report for the same model. All of these are now refused with an explanation; a logical or two-level factor focal entering as a single main-effect dummy is accepted.
  • power_statement() describes the world it priced: the latent (Gaussian-copula) correlation actually specified (or that the regressors are independent – it used to assert “Gaussian-copula dependence” even without one), the covariates’ marginals, each panel variable’s within-unit persistence, and the empirical route’s resampling mode, plus the separation convention. ape_n() now stores its standard-error type (se), so statements for robust-SE sample-size results no longer describe model-based standard errors; print() shows it.

powerape 1.10.0

  • alpha_frontier(): the closed-form justified error rate at the detectability frontier – the level that minimizes the weighted combined Type I and Type II error rate (Mudge et al., 2012; Maier & Lakens, 2022) for the effect that is just detectable at the target power. Under the normal approximation the first-order condition collapses to phi(z) = prior * phi(qnorm(power)) / (m * cost) (m = 2 two-sided detection, m = 1 one-sided minimum-effect), so the level involves neither n nor the standard error: solve it once and price a whole MDE grid under one justified alpha (ape_mde(..., alpha = alpha_frontier(cost = 4))). Anchors: 0.0274 at Cohen’s 4:1 weighting and 80% power (two-sided); exactly 1 - power at equal costs (one-sided), where the frontier balances alpha = beta. Errors when the weighting admits no interior optimum and warns when the level is not below 0.5. Battery check V22 verifies the closed form against ape_alpha()’s simulation optimum on frontier-pinned draws.
  • ape_robust() gains the se argument of ape_power(): a design analyzed with heteroskedasticity-robust standard errors now prices its robustness sweep (and the n_max search and MDE mode) under the same standard errors instead of silently reverting to model-based ones. Panel and IV routes keep their fixed inference and warn, as in ape_power().
  • Validation battery V21 covers the empirical-covariates and pilot-model routes, which previously had unit tests but no battery item: 95% CI coverage at the nominal rate, estimator consistency, and concordance with the matched parametric world (a large pilot drawn from that world must reproduce its power within Monte Carlo error) for both ape_dgp_empirical() and ape_dgp_from_fit().
  • Input-validation hardening: a sesoi supplied alongside claim = "detect" (it breaks out the outcome table) is now checked for being a single positive number instead of silently corrupting that table; pa_var() warns when arguments irrelevant to the declared type are supplied (p for a normal variable, mean/sd for a binary one).
  • plot.powerape_curve() no longer forces isotonic smoothing when alpha is a sample-size rule – under an n-dependent level, power is legitimately non-monotone in n, and the smoother would misdraw it.
  • power_statement() no longer prints an empty covariate list for an empirical-route DGP whose pilot data supply only the focal column.

powerape 1.9.0

  • Justified error rates (vignette("justified-alpha"); DESIGN.md section 15). Two routes to an alpha that is chosen rather than inherited, both opt-in; the default stays alpha = 0.05 per claim.
    • Sample-size rules. alpha may be a function of n in every power function, for example alpha = function(n) alphaN::alphaN(n, BF = 3), the Bayes-factor calibration of Wulff and Taylor (2024) that lowers alpha as n grows. The rule is evaluated at each candidate n, so ape_n() designs jointly over the pair (alpha(n), n), and ape_curve(), ape_mde(), and ape_robust() use the level that belongs to their n. Results store the rule and the level it produced (alpha, alpha_rule; curves and search histories gain an alpha column), print methods say “by sample-size rule”, and power_statement() states it.
    • Error-cost optimization. ape_power() now keeps the estimate and standard error of every replication (keep_draws = TRUE), so power at any alpha is a re-threshold of one simulation, exact on those draws. power_at() exposes that curve. ape_alpha() minimizes or balances the weighted combined Type I and Type II error rate of Mudge et al. (2012) and Maier and Lakens (2022) for given error costs and prior odds; it reports the flat range of levels indistinguishable from the optimum at the run’s precision, confines the search to cap = 0.05 (raising alpha needs a justification of its own) while stating when the unconstrained optimum lies above the cap, and with size = TRUE simulates the claim-boundary world to report the REALIZED error rate at the chosen level, which analytic compromise-power tools cannot. print, plot, and power_statement() methods; a preregistration paragraph.
  • Verification: new unit tests (test-alpha.R: constant rules reproduce scalar runs bit for bit, power_at() exactness, a closed-form twin of the optimizer on synthetic draws, Cohen’s 4:1 balance returning .05/.20 at 80% power, JustifyAlpha concordance on the same draws, the realized boundary size) and battery check V20.
  • alphaN and JustifyAlpha join Suggests (examples and concordance tests only).

powerape 1.8.0

  • Breaking: error-rate convention. Every claim is now tested at alpha (new argument, default 0.05) in its conventional form. Detection stays the two-sided 5% test (95% interval, directional counting); the minimum-effect claim is a one-sided test at 5% and equivalence a TOST at 5%, both read off the 90% interval – the convention of Lakens (2017), Lakens et al. (2018), TOSTER, and Riesthuis (2024) for equivalence. Before 1.8.0 all three claims read the 95% interval, which made the SESOI claims one-sided tests at 2.5% (a TOST at 2.5%): internally consistent but stricter than the field’s convention without a stated reason. Consequences: minimum-effect and equivalence power rise and required n falls (Design A’s minimum-claim requirement moves from ~2,600 to ~2,100); detection results are unchanged. conf remains as an explicit override of the claim’s own interval level (conf = 0.95 on a SESOI claim reproduces the old numbers). The outcome distribution uses both intervals accordingly. power_statement() now names the claim’s error rate.
  • Verification: the TOSTER concordance test and battery V7 now audit the DEFAULT convention (powerape’s default must match TOSTER at TOSTER’s default alpha with no remapping), and a new battery check V19 measures size at the claim boundaries (.05 for the SESOI claims, .025 for directional detection). Why the old battery did not flag the issue: V1/V7/V13 verified that the engine implements its stated rule exactly – V7 even matched TOSTER at a remapped alpha = .025 and guarded that .05 “did not match” – so a convention choice could not be caught by a convention-matched check. The new checks pin the default to the field’s default.

powerape 1.7.0

  • Unbalanced panels (Wooldridge, 2019, J. Econometrics). ape_dgp_panel() accepts a distribution over panel lengths: a vector n_periods with probabilities p_periods (equal weights by default), or a scalar n_periods with a per-wave retention rate for monotone attrition. Mundlak means are computed over each unit’s observed periods and the estimating model gains period-count cohort indicators (the Wooldridge 2019 workhorse); selection is completely at random by construction, rho is the length-mixture average of the latent unit-effect share, and sample sizes still count units. Degenerate mixtures (retention = 1, single support) short-circuit to the balanced code path, bit-identically. retention is sweepable in ape_robust(). Validated as battery V18: balanced-reduction identity; Stata probit i.d ... i.T, vce(cluster) + margins agreement to ~5e-7 on coefficients and all seven decimals on the APE, with the 0.12% clustered-SE gap reproduced to 4e-8 from the observed-information bread; an unbalanced xtprobit ASF anchor (|diff| .0012); consistency and 95% coverage; and content anchors (power monotone in retention; the absence of an unequal-cluster penalty for a mean-preserving length mixture reported here was wrong, see 1.11.1).

powerape 1.6.0

  • Breaking: panel designs are probit-only. ape_dgp_panel(model = "logit") is now refused with a teaching error instead of building a world. Rationale: the exact ASF recovery that justifies the CRE panel route is probit-specific (a normal unit effect rescales the index by sqrt(1 + var_a)); a pooled CRE logit estimates only a quasi-ML approximation of the ASF (measured gap in the least favorable validated configuration: -0.0002, Monte Carlo SE 0.001 – small, but an approximation, and powerape does not price designs with an approximate estimand). Calibrated to the same baseline and probability-point target, the probit panel world is practically indistinguishable on the probability scale. The 1.2.1 Gauss-Hermite calibration machinery remains internally (and tested) for the mixture integrals; no public route reaches its logit branch.

powerape 1.5.1

  • Verification expansion (battery V12-V17, adversarially refereed): exact enumeration of the clustered panel decision rule (within half a Monte Carlo SE of simulation at 24 and 40 clusters); a size audit across all routes (nominal at the standard/panel routes and at the minimum-claim boundary; the small-G cluster over-rejection and the IV route’s conservative null quantified and decomposed); a recursive bivariate-probit anchor separating the binary-endogenous control function’s plim gap (-0.011) from the truth; an asymptotic-SE identity for the CF route; xtprobit agreement (pure-RE and Mundlak-augmented) to ~0.001; and bootstrap-vs-delta SE ratios within 2% on all three routes. New stochastic consistency/coverage test for the IV-AIE.
  • Documentation: ape_dgp_iv() now records two properties of the control-function estimator found in validation – the conservative null of the binary-endogenous generalized residual, and the observed-pairs vs product-measure ASF wedge for continuous endogenous focal variables (the margins convention, replicated exactly; a product-measure option is on the roadmap).

powerape 1.5.0

  • Minimum detectable effect: ape_mde() inverts the design question – the sample size is fixed (an archive of so many firms, a panel of so many units and waves) and the function finds the smallest APE or AIE the design reliably concludes at a target power. Detect and minimum-effect claims search over the assumed effect, re-running the APE/AIE inversion at every candidate; the equivalence claim holds the pinned truth and returns the tightest establishable margin instead. Every route works unchanged (parametric, empirical, pilot-model, panel with n in units, IV), and the answer is verified with the same conservative confirmation stage as ape_n() (pushed upward, never down, when the high-precision run falls short). Feasibility ceilings produce a teaching error reporting the largest attainable effect and the power available there.
  • Robustness for the MDE: ape_robust(mode = "mde") sweeps the contextual assumptions and reports the minimum detectable effect per scenario; the worst case is the largest MDE – the smallest effect the design finds even under the least favorable assumptions.
  • power_statement() renders ape_mde() results (“the minimum detectable APE at n = … was …”), units-aware for panel designs.
  • Validation: battery gains V11 – the MDE at the four-tool anchor (n = 712) returns the anchor’s target (~0.10), the exact-enumeration power of the decision rule at the returned MDE sits at the goal, and the analytic inverse from power.prop.test() agrees.
  • The minimum-effect and panel vignettes gain fixed-n sections demonstrating ape_mde().

powerape 1.4.0

  • IV designs via control-function probit: ape_dgp_iv() specifies a world with an endogenous focal variable – linear first stage for a continuous focal, probit first stage for a binary one, instruments with a friendly strength knob (iv_strength = the share of the focal’s (latent) variance the instruments explain), and endogeneity = the structural/first-stage error correlation. Estimation in the engine replicates Stata 18.5’s cfprobit exactly: two-step control function with stacked method-of-moments standard errors (no bootstrap; robust and cluster flavors verified against Stata to ~1e-7, including the convention details – observed score derivatives, generated regressors as fixed instruments in the first-stage Jacobian, no small-sample factors). The estimand stays the ASF-based APE/AIE, which margins after cfprobit targets with the control functions held fixed.
  • IV interaction designs: with a binary exogenous moderator, the main equation gains the focal-by-moderator term, the control function, and the control-function-by-moderator interaction (Stata’s interact() + mainonly()), and the first stage includes the moderator and the instrument-by-moderator interactions – the general IV requirement for interaction models. Validated against Stata’s exact syntax.
  • The estimator internals also cover fprobit and poisson first stages, matching cfprobit’s fit exactly. For fractional first stages we found Stata’s margins to be internally inconsistent with its own estimator (its _remake_cfs.ado rebuilds the control function with the binary-outcome branch, ignoring the fractional value); powerape uses the estimation-consistent generalized residuals, and the discrepancy is reproduced to 7 decimals in the validation battery.
  • Robust standard errors for the standard routes: ape_power(), ape_curve(), and ape_n() gain se = c("model", "robust") – heteroskedasticity-robust (HC0) sandwich matching sandwich::vcovHC(type = "HC0") exactly. Panel designs keep their unit-clustered SEs and IV designs their stacked robust sandwich (a warning explains this if se is passed there).
  • Validation: battery gains V10 (the seven-case Stata equivalence table, the fprobit attribution, engine consistency and coverage, and the “IV price” anchor – at zero endogeneity the CF estimator’s SE exceeds the exogenous estimator’s by about sqrt(1/iv_strength), the 2SLS variance logic).

powerape 1.3.0

  • Panel AIE designs (binary x binary): ape_dgp_panel() gains a moderator argument, and set_aie() pins the average interaction effect of a CRE panel world. All anchors – baseline, the two conditional-at-reference main-effect APEs, and the target AIE – are defined on the average structural function (unit effects integrated exactly, Mundlak means held fixed), so the panel AIE is the same population quantity a cross-sectional AIE design targets. The estimating model includes the interaction’s own Mundlak mean whenever the product term is time-varying (the mean of a product is not the product of means); inference is unit-clustered throughout, and ape_robust(), ape_curve(), ape_n(), and power_statement() all work unchanged.
  • Validation: the ASF double difference matches Stata (probit i.d##i.m ... , vce(cluster id) + margins + lincom) to all printed digits on a fixed panel, with the clustered-SE gap again fully attributed to Stata’s observed-Hessian bread (reproduced to 7 decimals); ginteff reproduces the double difference and model-based SE on the identical fit to 1e-6; exact reduction of the inversion to the cross-sectional case at rho = 0; estimator consistency and nominal clustered-CI coverage of the true AIE. Battery gains V9.
  • Continuous focal or moderator pairs in panel AIE designs are not yet supported (clean error); cross-sectional AIE keeps all four type pairs.
  • New vignette panel-designs: the CRE workflow end to end – the panel knobs, units-not-observations accounting, the panel-vs-cross-section comparison, rho robustness, and the panel AIE. The README now covers the panel route.

powerape 1.2.1

  • Fixed: for logit panel DGPs with a pure random-effects component (rho > 0 and cre_share < 1), calibration and true values integrated the unit effect with the probit closed form (index divided by sqrt(1 + var_a)), which over-attenuates a logistic kernel. A requested baseline .30 / APE .10 logit world at rho = .5, cre_share = 0 actually had baseline .269 and true APE .116, so simulated power for such designs was optimistic. The unit effect is now integrated by 20-node Gauss-Hermite quadrature for logit; probit keeps the exact closed form and is numerically unchanged. Realized-vs-requested calibration is regression-tested for both links.

powerape 1.2.0

  • Panel designs: ape_dgp_panel() specifies a correlated random effects (CRE) probit/logit world – Mundlak heterogeneity on observed unit means, latent unit-effect share rho (xtprobit convention), cre_share for the heterogeneity-regressor correlation, and per-variable within-unit persistence via pa_var(..., icc =) (icc = 1 gives unit-level, time-constant variables). Estimation in the engine is pooled probit/logit with Mundlak means and unit-clustered standard errors; the target remains the ASF-based APE, which the pooled CRE estimator recovers even though coefficients are attenuated. Sample-size arguments count units (clusters); each contributes n_periods observations.
  • Clustered delta-method inference: score-based cluster-robust sandwich matching R’s sandwich::vcovCL convention exactly (verified against marginaleffects with vcov = ~id); Stata’s vce(cluster) differs only by its observed-Hessian bread (~0.3% in the validation example, with the attribution reproduced to 7 decimals). A warning is issued below 30 clusters.
  • Validation: exact-enumeration reduction (no-Mundlak case), the Donner-Klar cluster design-effect anchor, clustered-CI coverage, and the documented “Mundlak insurance premium” (at rho = 0 the CRE estimator is correctly less powerful than a cross-sectional analysis of the same information). Battery gains V8.
  • Panel scope in this release: parametric route, balanced panels, no moderator (AIE) combination yet.

powerape 1.1.0

  • ape_n() gains a high-precision confirmation stage: by default (confirm = TRUE) the candidate n is re-measured at nsim_confirm = 4 * nsim replications and accepted only if the confirmed power is within max(0.005, 1.5 * MCSE) below the goal, pushing n upward otherwise (overshoot is never trimmed – the conservative direction). The returned power and MCSE come from the confirmation run; the search history gains a stage column. confirm = FALSE restores the fast single-stage search.
  • Validation battery added (tests/testthat/test-exact-power.R and the project-level battery): simulated power matches the exact finite-sample power of the engine’s decision rule – computed by triple binomial enumeration in the saturated case – to within Monte Carlo error, for detection, minimum-effect, and equivalence claims. Cross-checks against power.prop.test(), Stata power twoproportions, pwr (arcsine), and Hsieh’s logistic-regression formulas (powerMediation) agree.

powerape 1.0.0

  • Pilot-model route: ape_dgp_from_fit() lifts covariate rows and nuisance coefficients from a fitted binomial glm; the focal coefficient is re-solved in APE units by set_ape(). Optional baseline override.
  • Four vignettes: the minimum-effect workflow, the classic SESOI detection analysis, equivalence testing, and average interaction effects.
  • Validation: APE point estimates and delta-method SEs match Stata margins, dydx() to ~4e-7 on fixed data.
  • pkgdown site and hex logo.

powerape 0.3.0

  • Average interaction effects: set_aie() pins a moderated DGP from three explicit anchors (conditional-at-reference main-effect APEs plus the target AIE); the internal AIE estimator covers all four variable-type pairs and reproduces ginteff exactly (estimates and SEs) on fixed data.
  • ape_robust(): scenario sweeps over contextual assumptions with pin-the-APE / pin-the-coefficients modes, worst-case power, and the insurance-premium sample size n_max (Hancock & Feng, 2025).
  • power_statement(): renders any result as a citable methods paragraph.

powerape 0.2.0

  • Continuous (normal) focal variables: average-derivative APE with plateau-aware inversion and feasibility reporting.
  • Empirical-covariate route: ape_dgp_empirical() resamples pilot rows (focal jointly from a column, or independently as a pa_var).
  • Validation: APE estimates and SEs match marginaleffects on fixed data; simulated power matches Stata power twoproportions analytics.

powerape 0.1.0

  • Initial engine: probit/logit, binary focal variable, Gaussian-copula covariates, APE-unit inversion with feasibility checks, CI-based claims (detect, minimum, equivalence; Riesthuis, 2024), ape_power(), ape_curve(), ape_n(), Monte Carlo SEs throughout.