LoginSign Up
← Back to all sparks
F

fect

ANALYTICS
Velocity0.0

R package fect by xuyiqing — release notes from GitHub.

github.com/xuyiqing/fect ↗

A counterfactual estimator turning itself into a platform for multiple estimands

r-packagecausal-inferencepanel-dataapi-redesignreproducibility
◆Current state
fect implements counterfactual estimators for panel data with treatment effects — imputation-based fixed effects, interactive fixed effects, matrix completion. The 2026 releases move fast and bundle heavily: 2.1.0 rewrote complex fixed effect handling, 2.2.0 unified cross-validation under a single cv.method parameter and replaced method='gsynth' with an explicit time.component.from switch, 2.4.1 introduced a post-hoc estimand API, and 2.4.5 added group.fe for coarsened fixed effects plus a $sample slot exposing which cells entered estimation.
◆Where it's heading
The direction is separation of estimation from interpretation. Where the package once returned one effect from one fit, estimand() now dispatches typed estimands — ATT, cumulative ATT, APTT, log ATT — from any imputation fit, with effect() and att.cumu() soft-deprecated but byte-identical pending 3.0.0. Alongside that runs a transparency thread: the $sample matrix, out-of-sample comparison via fect_mspe(), and named component sources instead of opaque method aliases. The release notes are unusually precise about which results change and which do not.
◆Prediction
The soft-deprecation notice names 3.0.0 as the removal point for effect() and att.cumu(), so a major release consolidating on the estimand() dispatcher is the clearly signposted next step.

◆Recent moves

  1. 2mo ago

    Adds group.fe for coarsened fixed effects and a $sample slot

    Bundles four releases: group.fe absorbs additive fixed effects at a coarser level than the unit identifier with clustered SEs defaulting to match, and fect() now returns a $sample matrix marking every cell used in estimation. Both continue the package's push toward making its own internals inspectable rather than implicit.

    View source ↗
  2. 3mo ago

    Post-hoc estimand API decouples estimands from the fit

    ⚡ SPARK

    Introduces estimand() as a typed dispatcher returning alternative estimands from any imputation fit, and soft-deprecates effect() and att.cumu() against a 3.0.0 removal. This is the structural turn the trajectory has been building toward — the fit becomes a substrate, and what you ask of it becomes a separate decision.

    View source ↗
  3. 4mo ago

    Unified cross-validation and explicit control of time components

    Replaces method='gsynth' with an explicit time.component.from switch, collapses cv.treat and mask.method into a single cv.method parameter, adds k-fold CV for never-treated estimation, and adds fect_mspe() for out-of-sample model comparison. Retiring a method alias in favor of naming the actual modeling choice is the same clarity impulse that later produces estimand().

    View source ↗
  4. 7mo ago

    Rewrites complex fixed effect handling and fixes speed

    A terse release rewriting complex fixed effect handling alongside a speed fix. The fixed-effects machinery reworked here is what group.fe later builds on in 2.4.5.

    View source ↗
  5. 11mo ago

    Adds heterogeneous treatment effect plots and caps default cores

    Adds a type='hte' plot for heterogeneous treatment effects and caps default parallel cores, plus dependency-driven fixes. Modest next to the 2.x releases, and the last entry before the package's pace picks up sharply.

    View source ↗