skills/66-zheng-siyao-empirical-research-skills/R-optimizer/SKILL.md
R语言实证分析优化Skill。优化R代码效率、处理大规模面板数据、加速回归计算(并行化、向量化、向量化)。触发词:R语言优化/R加速/R性能优化/大规模数据处理/R optimization
npx skillsauth add brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research R-optimizerInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
本 Skill 针对经济学实证分析场景,优化 R 代码的执行效率和内存占用。涵盖:向量化操作、并行计算、大规模面板数据处理、回归加速。让原本需要数小时的回归在几分钟内完成。
实证分析的瓶颈往往不是算法,而是 I/O 和循环。优化顺序:
诊断慢点 → 向量化优化 → 并行化加速 → 内存优化 → 验证结果一致性
使用 profvis 定位瓶颈:
# 安装并使用 profvis
install.packages("profvis")
library(profvis)
profvis({
# 放入需要诊断的代码
df <- readstata13::read.dta13("data/raw/cgss2021.dta")
result <- lm(income ~ education + age + urban, data = df)
})
常见瓶颈定位:
| 代码模式 | 问题 | 优化方向 |
|----------|------|----------|
| for (i in 1:nrow(df)) | 行循环 | 向量化 |
| merge() 多次 | I/O 瓶颈 | data.table::merge |
| lm(y ~ x1 + x2 + ... + x100) | 大矩阵求逆 | 固定效应投影矩阵 |
| boot() 大量重复 | 单线程 | 并行化 |
用 data.table 替代 data.frame:
library(data.table)
# 读取速度:data.table 比 read.csv 快10-100倍
DT <- fread("data/raw/cgss2021.csv")
# 分组计算:避免 for 循环
DT[, .(mean_income = mean(income, na.rm = TRUE),
sd_income = sd(income, na.rm = TRUE)),
by = .(urban, year)]
# 滚动窗口:向量化替代循环
DT[, lag_income := shift(income, 1), by = id]
DT[, income_growth := income / lag_income - 1]
向量化替代循环示例:
# ❌ 慢:行循环
for (i in 2:nrow(df)) {
df$income_growth[i] <- (df$income[i] - df$income[i-1]) / df$income[i-1]
}
# ✅ 快:向量化
library(data.table)
DT <- as.data.table(df)
DT[, income_growth := income / shift(income, 1) - 1, by = id]
使用 future + furrr 实现并行:
# 安装
install.packages(c("future", "furrr", "doParallel"))
library(future)
library(furrr)
plan(multisession, workers = 6) # 使用6个核心
# 并行化回归:对多个因变量分别回归
library(furrr)
library(broom)
models <- c("income", "education", "employment")
results <- future_map(models, ~ {
lm(as.formula(paste(.x, "~ treatment + controls")), data = df) %>%
tidy() %>%
filter(term == "treatment")
})
并行 Bootstrap:
library(doParallel)
cl <- makeCluster(6)
registerDoParallel(cl)
boot_results <- foreach(i = 1:1000, .combine = rbind) %dopar% {
# Bootstrap 抽样
idx <- sample(nrow(df), replace = TRUE)
df_boot <- df[idx, ]
# 单次回归
coef(lm(income ~ treatment, data = df_boot))
}
stopCluster(cl)
用 fixest 替代 felm 或手动固定效应:
# 安装
install.packages("fixest")
library(fixest)
# 固定效应回归:比 lm 快了 10-100倍
# 原理:先做组内变换(within transformation),不构造大矩阵
# 双向固定效应(国家+年份)
feols(income ~ treatment | country + year,
data = DT,
cluster = ~country) # 聚类标准误
# 聚类在多个层面(双向聚类)
feols(income ~ treatment | country^year,
data = DT,
cluster = ~country + year)
# 多时点 DID(交错处理)
feols(log_gdp ~ treatment | country + year,
data = DT,
panel.id = ~country + year,
didsetup = TRUE) # 自动识别交错处理
# 输出带星标的回归表
etable(feols_model1, feols_model2,
digits = 3,
signifCode = FALSE,
tex = TRUE)
fixest vs lm 性能对比:
| 场景 | lm | felm (lfe) | fixest | |------|-----|------------|--------| | 100国家 × 30年面板 | 慢 | 快 | 最快 | | 聚类标准误(国家层) | ✅ | ✅ | ✅ | | 双向固定效应 | ✅ | ✅ | ✅ | | 并行计算 | ❌ | ❌ | ✅ |
减少数据拷贝:
# ❌ 每次操作都拷贝数据
df2 <- df[df$year > 2010, ]
df3 <- df2[df2$urban == 1, ]
# ✅ 用 data.table 原地修改
DT <- fread("data/raw/cgss2021.csv")
setkey(DT, year)
DT <- DT[J(2011:2021)] # 原地过滤,不拷贝
大型数据读写:
# ❌ read.csv 慢且占内存
df <- read.csv("data/raw/large_file.csv")
# ✅ fread 自动检测类型,速度快10倍
DT <- fread("data/raw/large_file.csv")
# ✅ 对超大型文件(>10GB)分块读取
DT <- fread("data/raw/large_file.csv",
nrows = 100000, # 先读10万行
skip = 1000000) # 跳过前100万行
优化后必须验证结果一致:
# 验证:向量化版本 vs 循环版本结果一致
original_result <- loop_version(df)
vectorized_result <- vectorized_version(df)
all.equal(original_result, vectorized_result) # 必须返回 TRUE
# 验证:并行版本 vs 单线程版本结果一致
set.seed(42)
parallel_boot <- parallel_bootstrap(df, nboot = 1000, ncores = 6)
set.seed(42)
sequential_boot <- sequential_bootstrap(df, nboot = 1000)
max_abs_diff <- max(abs(parallel_boot - sequential_boot))
print(paste("最大差异:", max_abs_diff)) # 必须 < 1e-6
# 优化代码保存规范
# 保存脚本:scripts/optimize_regression.R
# 主脚本:analysis/01_baseline_regression.R
# 输出:output/tables/regression_table1.tex
在 Claude Code 对话窗口输入:
/R-optimizer
或完整 Prompt:
用R-optimizer Skill优化本项目的R代码:使用data.table向量化处理、fixest固定效应回归、并行Bootstrap。确保优化后结果与优化前一致。
codebook-pass → 清洗后的数据用 R 优化脚本处理latex-table → fixest 的 etable() 直接输出 LaTeXdid-reviewer → DID 估计使用 fixest 的多时点功能tools
Recommend AND run open-source AI tools, agents, Claude Code / Codex skills, and MCP servers for any stage of a literature review — searching, reading, extracting, synthesizing, screening, citation-checking, and paper writing. Use when the user asks "what tool should I use to..." OR "install/run/use <tool> to ..." for research/lit-review work: automating a survey or related-work section, PDF→Markdown extraction for LLMs (MinerU/marker/docling), PRISMA / systematic review (ASReview), citation-backed Q&A over PDFs (PaperQA2), wiring papers into Claude/Cursor via MCP (arxiv/paper-search/zotero servers), or chatting with a Zotero library. Ships a launcher (scripts/litrun.py) that installs each tool in an isolated venv and runs it. Curated catalog of 70+ vetted projects. 支持中英文(用于「文献综述工具选型」与「一键安装/运行」)。
development
Route empirical-research requests through the Auto-Empirical Research Skills catalog when this whole repository is installed as one skill in Codex, CodeBuddy, Claude Code, or another IDE. Use to choose and load the right vendored AERS skill for causal inference, econometrics, replication, data acquisition, manuscript writing, peer review and referee responses, citation checking, de-AIGC editing, or full empirical-paper workflows without reading the entire repository at once.
documentation
Use when the project collects primary data or runs a field, lab, or survey experiment, before the intervention begins — write the pre-analysis plan, size the sample from a power calculation, and register with the AEA RCT Registry. Apply after the design is chosen in aer-identification and before any outcome data are seen.
tools
Guide economists to authoritative data sources with explicit, confirmed data specifications before retrieval; interfaces with Playwright MCP to navigate portals and extract real data, not articles about data.