Skip to contents

Affine Paper Replication

Source: vignettes/paper-affine-replication.Rmd
paper-affine-replication.Rmd

The study represented here

The installed replication layer represents the current affine paper variant of Orthogonal Inference for Conditional Z-Estimation without Cross-Fitting: A Distributional Nearest-Neighbor Approach. The snapshot is October 14, 2026; manifest.json pins the exact paper and source snapshot commits. It includes the bounded CATE and residualized-slope simulation campaign, the overhauled SIPP application, and accepted evidence for the current manuscript and supplement. The independent coupled-score manuscript is separate.

The paper’s scientific R files are bundled unchanged. Package launchers handle installation paths, dependencies, output directories and verification. Superseded CATE and IV study launchers are removed. General sim_cate_* and experimental sim_iv_* APIs remain synthetic benchmark tools; they do not implement this affine campaign.

Install the replication bundle

The bundle was introduced in development version 0.1.0.9000. Install the current master branch before following this guide; the initial v0.1.0 release does not contain the affine bundle.

install.packages("pak")
pak::pak("JakobJuergens/RdistNN@master")
bundle <- system.file("paper-replication/affine", package = "RdistNN")
list.files(bundle)
#>  [1] "configs"           "evidence.zip"      "helpers.R"        
#>  [4] "manifest.json"     "README.md"         "render_evidence.R"
#>  [7] "run_simulations.R" "run_sipp.R"        "source"           
#> [10] "verify.R"

Read README.md in that directory for the complete protocol and dependencies. manifest.json pins the paper commit, source and evidence hashes, production configuration and SIPP run. No paper checkout or native run cache is needed.

Recreate the accepted manuscript displays

This is the quickest way to reproduce the paper’s reported numerical displays. Install the optional rendering dependencies and launch the installed script:

install.packages(c("digest", "jsonlite", "ggplot2", "patchwork"))
system2(file.path(R.home("bin"), "Rscript"), c(
  shQuote(file.path(bundle, "render_evidence.R")),
  "--output-dir=affine-paper-displays"
))

Rendering uses R’s Cairo PDF device. With the CRAN build of R on macOS, install XQuartz to provide its X11 libraries. This requirement is described in the R for macOS installation notes.

The script verifies the compressed evidence, regenerates eight simulation tables, thirteen SIPP tables and six PDF figures, runs the upstream check routines, and compares all tables with accepted references. It performs no simulation or estimation. Generated display manifests record the input hashes; replication.json records the pinned paper identity and successful comparison. The unchanged renderers receive the pinned paper commit as source provenance through the portable launcher.

Historical transformed SIPP estimates remain saved evidence because the current supplement reports that distinct target. Their old fitting workflow is not included. The overhauled application uses raw dollars.

Run the current affine simulation campaign 

install.packages(c("digest", "jsonlite", "future", "future.apply"))
system2(file.path(R.home("bin"), "Rscript"), c(
  shQuote(file.path(bundle, "run_simulations.R")),
  "--mode=smoke", "--output-dir=affine-smoke"
))

Smoke runs are explicitly validation: two preflight draw families at n=128. They exercise both CATE and slope, full-sample/pooled-five-fold/oracle nuisances, DNN and Gaussian localization, and all three covariance routes. They do not reproduce the large-dictionary endpoint or constitute manuscript evidence.

The accepted production grid contains:

Profile Cells Main replications per cell
DNN C0/C1/C2 14 10,000
DNN/kernel comparison 9 10,000
M=3 endpoint, n=531,441 1 10,000
C0H/C2S settings 4 10,000

Both applications have point target 1/8 at X=(1/2,1/2). There are two reporting coordinates and four adjustment coordinates. First stages use clipped growing ridge series with penalty n^(-1/2). The primary DNN order is floor(n^(3/4)); C2 also uses multiplier 0.4. This study-specific schedule overrides the general package tuning default. The Gaussian comparison matches population DNN ESS.

The covariance routes are frozen-score jackknife, direct moment and direct squared-weight estimates. Diagnostics distinguish point targets, realized weighted truth, spatial drift, remainders and joint slope coefficient covariance. The shared seed schedule supports paired comparisons. The 2,516,400 rows contain 10,000 common independent draw families and 480,000 duplicated DNN rows across profiles.

Launch the full campaign on Linux with Python3 and sufficient resources:

system2(file.path(R.home("bin"), "Rscript"), c(
  shQuote(file.path(bundle, "run_simulations.R")),
  "--mode=production", "--output-dir=affine-production"
))

The original monitor enforces profile wall limits and a sampled 40-GiB RSS cap. The scientific runner uses four workers, eight-family chunks and one inner thread. The M3 endpoint is substantial. Source/dependency hashes, actual DLL identity, seed maps, schemas and process receipts accompany each run. The launcher audits and analyzes the completed four-profile campaign. Use --profiles=... for selected profiles and --resume for an existing production directory with an identical runtime freeze. Partial/stopped runs are not complete production. See the installed README for budgets and layout.

The historical production package pin is 40060f155cc4d84d59ee5af119f6a2534c569bab. Fresh runs record the package actually executed. A version label alone does not certify executable identity, and fresh results require review before promotion into the manuscript.

Run the overhauled SIPP application 

install.packages(c("hdm", "grf", "digest", "jsonlite"))
system2(file.path(R.home("bin"), "Rscript"), c(
  shQuote(file.path(bundle, "run_sipp.R")),
  "--mode=smoke", "--output-dir=sipp-smoke"
))

Use --mode=production and a new directory for the accepted specification: 9,915 complete households from hdm::pension, eligibility treatment, raw-dollar net financial assets, baseline adjustment, seven age-40 income profiles, standardized age/asinh-income reporting coordinates, 2,000-tree forests, five folds and seed 20260709.

The three prediction mechanisms are literal all-tree full-fit, own-arm OOB from the same full forests, and held-out five-fold predictions. Final DNN weights are common across mechanisms. Default s=397; fitted orders are 298,397,496,745. The eight-order exploratory grid 99,149,198,248,298,397,496,745 retains all 168 profile/mechanism/order rows.

The launcher fits and saves models/predictions, computes diagnostics, exports paired component covariance and frozen-score influence information, and runs localization sensitivity without refitting. Its output lives in runs/affine-current/, with separate stage logs and hashed run receipts. Smoke uses 300 households, 50 trees and two folds; its exports carry validation labels. Output directories are never silently overwritten.

These practical forests, unbounded dollar outcomes and data-selected reporting scales/profiles are outside the bounded-series primitive theorem. Frozen deletions do not refit nuisance learners. Order sensitivity does not estimate spatial bias or select an optimal order, and full-fit/OOB agreement does not verify a stability rate.