The results of this package do not match the results of the test data with causal forest. How to explain it?

[zhangkaicr]

First of all, thank you for providing this package for heterogeneity analysis of treatment effects. I am currently studying this aspect.
Through learning, I understand the relevant content of causal forest. Referring to the tutorial content at the following website, there should be treatment effect heterogeneity in the indo_rct dataset.
https://www.r-bloggers.com/2024/09/explaining-a-causal-forest/

Considering that the test data is an RCT study, I set the code according to your study, but the ABC values ​​I got are all negative, which seems to be contrary to the results of the causal forest. I hope you can answer this question. Thank you.

`
library(medicaldata)
library(tidyverse)
library(precmed)

data("indo_rct")
head(indo_rct)

indo_rct$rx <- as.integer(indo_rct$rx)-1
indo_rct$outcome <- as.integer(indo_rct$outcome)-1

indo_rct <-
indo_rct |>
mutate_if(is.factor, function(v) as.integer(v) - 1L) |>
rename(male = gender)

使用线性回归分析治疗(trt)对结果(y)的影响

output_lm <- glm(outcome ~ rx, family = "binomial", data = indo_rct)

打印线性回归结果

output_lm

output_atefit <- atefit(response = "count",
data = indo_rct,
cate.model = outcome ~ age + male + pep + recpanc + type + difcan + psphinc + bsphinc + pdstent + train,
ps.model = rx ~ 1,
n.boot = 500,
seed = 999,
verbose = 1)

打印ATE估计结果

output_atefit

计算率比(rate ratio)

rate.ratio <- exp(output_atefit$log.rate.ratio$estimate)
rate.ratio

计算率比的95%置信区间

CI.rate.ratio <- exp(output_atefit$log.rate.ratio$estimate + c(-1, 1) * qnorm(0.975) * sqrt(output_atefit$log.rate.ratio$SE))
CI.rate.ratio

绘制ATE结果图

plot(output_atefit)

使用catecv函数进行交叉验证

output_catecv <- catecv(
response = "count", # 响应变量类型为计数
data = indo_rct, # 数据集
# 指定CATE评分方法
score.method = c("poisson", "contrastReg","boosting"),
# 指定CATE模型
cate.model = outcome ~ age + male + pep + recpanc + type + difcan + psphinc + bsphinc + pdstent + train,
ps.model = rx ~ 1, # 倾向性评分模型
initial.predictor.method = "poisson", # 初始预测方法
higher.y = FALSE, # 是否使用较高的y值
cv.n = 5, # 交叉验证折数
seed = 999, # 随机数种子
plot.gbmperf = FALSE, # 是否绘制GBM性能图
verbose = 1 # 显示进度信息
)

计算曲线间面积(ABC)

output_abc <- abc(x = output_catecv)
output_abc

计算每种方法的平均ABC

average_abc <- apply(output_abc, 1, mean)
average_abc`