simulation-study skill

/simulation-study — Monte Carlo Simulation Study

Design and run a Monte Carlo experiment that characterizes an estimator's finite-sample behavior, then review it for the bugs that quietly invalidate simulation evidence.

Input: $ARGUMENTS — a description of the estimator(s) and DGP to study (e.g., "compare 2SLS vs LIML under weak instruments with heteroskedasticity"), or a pointer to an existing script/paper whose simulation you want to reproduce or extend.

Constraints

• Follow .claude/rules/simulation-conventions.md — the simulation contract (DGP, truth, estimand, MCSE, assumption regime) is non-negotiable.
• Declare the assumption regime in the script header and respect the firewall — an out-of-assumption run never supports a within-assumption claim (simulation-conventions.md §2).
• Follow .claude/rules/r-code-conventions.md for general R standards (header, library() at top, relative paths, numerical discipline).
• Save the script to scripts/R/ with a numbered, descriptive name (e.g., scripts/R/sim2slsvs_liml.R).
• Save outputs (per-rep raw tibble, summary table, figures) to output/.
• saveRDS() the per-replication raw results, not just the summary — re-aggregation and the review pass need them.
• Run the sim-reviewer agent on the generated script before presenting results, then address Critical/High findings.

Workflow Phases

Phase 0: Pre-Flight Report

Before writing any code, produce a Pre-Flight Report showing you have pinned down the experiment. This prevents the most common failure mode — a beautiful results table built on a mismatched estimand or a coverage-against-the-estimate bug.

## Pre-Flight Report — Simulation Design

**Research question:** [what finite-sample property is being demonstrated]
**Target estimand:** [ATT / ATE / coefficient θ — and how its TRUE value is computed from the DGP params]
**Maintained assumptions:** [the FULL list the estimator(s) under study require — A1 … An, every one]
**Regime:** [IN-ASSUMPTION — all hold | OUT-OF-ASSUMPTION — relaxes A[k] only, severity grid {…}, targeting pseudo-estimand …]
**Verification:** [per assumption, the checkable property of the DGP that establishes it — by construction or by an assertion]
**DGP:** [structure + the parameters that define it; what is held fixed vs. varied]
**Estimator grid:** [list each estimator + which estimand it targets + how it returns est/se/CI]
**Design grid:** [sample sizes, parameter values, scenarios to sweep]
**Replications R:** [value] → implied MCSE on coverage ≈ sqrt(0.95·0.05/R) = [value]
**Metrics:** bias, empirical SE, RMSE, coverage, size/power — each with MCSE
**Conventions read:** simulation-conventions.md, r-code-conventions.md

If the estimand or its true value is ambiguous, stop and ask before writing code.

If an assumption cannot be verified — you cannot name the property of the DGP that establishes it — the run is not IN-ASSUMPTION, and per the firewall no within-assumption claim may rest on it. Say so in the Pre-Flight Report rather than letting the header assert what was never checked.

Phase 1: The DGP

Write one parameterized function that returns a dataset. Compute and return (or store) the true target value from the parameters.

generate_data <- function(n, params) {
  # ... generate covariates, treatment, outcome from params ...
  list(data = df, truth = compute_truth(params))   # truth from params, never from an estimate
}

The header's Verified lines are earned here: every assumption the regime block claims holds by a check gets that check written into the script (a large-draw assertion, a condition number, a stopifnot() on the parameter bounds), run once at setup. An assumption whose verification exists only in the comment is asserted, not verified.

Phase 2: Estimator Grid

Each estimator is a function data -> list(est, se, cilo, cihi, converged). State the estimand each one targets; an estimator scored against a mismatched truth is a bug, not a finding.

Phase 3: Replication Engine

• set.seed(YYYYMMDD) once. For parallel reps use RNGkind("L'Ecuyer-CMRG") and furrr::furrr_options(seed = TRUE).
• One run = generate data → run every estimator → record a row per estimator with est, se, cilo, cihi, converged.
• Pre-allocate / bind results into a tibble of R × (#estimators) rows. Track non-convergence; never silently drop.

Phase 4: Metrics & Summary

Per estimator × scenario, against truth:

• Bias = mean(est) - truth (+ MCSE = sd(est)/sqrt(R))
• Empirical SE = sd(est); RMSE = sqrt(mean((est - truth)^2))
• Coverage = mean(cilo <= truth & truth <= cihi) (+ MCSE = sqrt(p(1-p)/R))
• Size / power = rejection rate under the null / alternative DGP
• Failures = count of non-converged reps

Build a tidy summary table; report MCSE next to every headline metric.

Phase 5: Figures

Use ggplot2 with the project theme: bias / coverage vs. sample size (or scenario), with reference lines (0 bias, nominal coverage). Transparent background, explicit dimensions (per r-code-conventions.md §4).

Phase 6: Save & Review

1. saveRDS() the raw per-rep tibble and the summary table to output/; also write the summary as .csv/.tex.
2. Run the review:

Delegate to the sim-reviewer agent:
   "Review the simulation script at scripts/R/[name].R"

The agent is read-only and returns its report; save it to qualityreports/[name]sim_review.md.

1. Address Critical/High findings (coverage-vs-truth, estimand mismatch, missing MCSE, dropped reps, an unstated or unverified regime) before presenting.
2. Apply the firewall to the presentation itself. Every claim you are about to make must cite a run whose regime can bear it — consistency, valid analytic standard errors, nominal coverage, and shipping a default require an IN-ASSUMPTION run and nothing else (simulation-conventions.md §2). Carry the regime in every caption — and per row wherever a severity grid mixes the two.

Script Structure

# ============================================================
# [Title] — Monte Carlo simulation
# Author: [project context]
# Purpose: [property being demonstrated]
# Estimand: [target + how truth is computed]
# Maintained assumptions: [A1 ... An — the FULL list the estimator requires]
# Regime: [IN-ASSUMPTION | OUT-OF-ASSUMPTION: relaxes A[k] only, severity ...,
#          targeting pseudo-estimand ...]
# Verified: [per assumption, the property that was actually checked, not asserted]
# Outputs: output/[name]_raw.rds, [name]_summary.{rds,csv}
# ============================================================

# 0. Setup ----
library(tidyverse)
library(furrr)            # parallel reps (optional)
plan(multisession)        # enable parallel workers; omit this line to run sequentially
RNGkind("L'Ecuyer-CMRG")
set.seed(20260531)        # once, YYYYMMDD (simulation-conventions.md §3)
R   <- 2000L              # MCSE on coverage near .95 ≈ 0.005
dir.create("output", recursive = TRUE, showWarnings = FALSE)

# 1. DGP ----
generate_data <- function(n, params) { ... }     # returns list(data, truth)

# 2. Estimators ----
estimators <- list(tsls = est_tsls, liml = est_liml)    # each -> est, s

Important

• State the regime, then respect the firewall. A DGP that violates a maintained assumption produces a table on which every other check here passes — and a within-assumption claim built on it is unsupported however clean the numbers look.
• The truth comes from the DGP, never from an estimate. Coverage is the CI containing the true parameter.
• No result without an MCSE. If two estimators differ by less than ~2× MCSE, say so.
• Save raw, not just summary. A number that exists only in the console cannot be audited or put on a slide.
• Count your failures. Silently dropped non-converged reps bias every metric.

Long-running simulations: use the Monitor tool

Large grids (many scenarios × large R) can run for many minutes. Background-launch via Bash with runinbackground: true, writing R stdout and stderr to a log (e.g. Rscript scripts/R/[name].R > output/[name].log 2>&1), and run the Monitor tool with a command that tails that log through grep --line-buffered, matching progress milestones (e.g. a progressr update) and failure signatures (Error, Execution halted), instead of polling with sleep. Monitor has no job-id parameter: the stdout of its own command is the event stream. See data-analysis/SKILL.md and the guide's Cost-Conscious Parallelism section.