绘制结构主题模型：如何实现时间维度的不连续性

现有实现代码 #

我正在使用R语言的stm包运行结构主题模型，模型包含faction_id与numeric_date（时间度量）的交互效应，已完成模型估计并绘制了两个派系第9号主题随时间的占比趋势图。现有代码如下：

# 拟合模型
model.fac.dat.int <- stm(
  documents = docs,
  vocab = vocab,
  K = 20,
  prevalence = ~ faction_id * s(numeric_date),
  max.em.its = 75,
  data = meta,
  reportevery = 50,
  verbose = TRUE,
  init.type = "Spectral"
)

# 估计第9号主题的效应
est.fac.dat.int <- 
  estimateEffect(
    formula = c(9) ~ faction_id * numeric_date,
    stmobj = model.fac.dat.int,
    metadata = meta,
    uncertainty = "None"
  )

# 绘制Greens派系的趋势线
plot(
  est.fac.dat.int,
  covariate = "numeric_date",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "Greens",
  linecol = "green",
  xlab = "Time",
  ylim = c(0, 0.1),
  printlegend = F
)

# 添加CDU派系的趋势线
plot(
  est.fac.dat.int,
  covariate = "numeric_date",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "CDU",
  linecol = "red",
  add = T,
  printlegend = F
)

需求与疑问 #

我希望在numeric_date=7000（选举日期）处的线性图中添加断点——理论上该时间点后折线会下移，当前图可能掩盖了这一效应，因此想要制作类似RDD的图。由于stm包没有专门的函数实现此场景，我考虑过使用rdd包，但不知如何与现有stm设置结合。请问是否应该分别估计numeric_date<7000和numeric_date>7000的效应，再将两段图合并？

解决方案 #

你提到的分时间段估计效应再合并绘图是完全可行的，同时还有一种更高效的方法：无需重新拟合STM模型，直接在estimateEffect中加入断点变量来捕捉时间断点的效应，以下是两种方法的具体实现：

方法1：分数据集估计效应并合并绘图

1. 拆分元数据

# 拆分数据集为选举前和选举后
meta_pre <- subset(meta, numeric_date < 7000)
meta_post <- subset(meta, numeric_date >= 7000)

2. 分别估计两个时间段的效应

# 选举前第9主题的效应
est_pre <- estimateEffect(
  formula = c(9) ~ faction_id * numeric_date,
  stmobj = model.fac.dat.int,
  metadata = meta_pre,
  uncertainty = "None"
)

# 选举后第9主题的效应
est_post <- estimateEffect(
  formula = c(9) ~ faction_id * numeric_date,
  stmobj = model.fac.dat.int,
  metadata = meta_post,
  uncertainty = "None"
)

3. 合并绘制断点图

# 先绘制选举前Greens的趋势
plot(
  est_pre,
  covariate = "numeric_date",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "Greens",
  linecol = "green",
  xlab = "Time",
  ylim = c(0, 0.1),
  xlim = range(meta$numeric_date),
  printlegend = F
)
# 添加选举后Greens的趋势
plot(
  est_post,
  covariate = "numeric_date",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "Greens",
  linecol = "green",
  add = T,
  printlegend = F
)

# 添加CDU选举前趋势
plot(
  est_pre,
  covariate = "numeric_date",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "CDU",
  linecol = "red",
  add = T,
  printlegend = F
)
# 添加CDU选举后趋势
plot(
  est_post,
  covariate = "numeric_date",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "CDU",
  linecol = "red",
  add = T,
  printlegend = F
)

# 添加垂直断点线
abline(v = 7000, lty = 2, col = "gray50")

方法2：在estimateEffect中加入断点变量（无需重新拟合模型）

通过构造断点虚拟变量及其交互项，直接在效应估计中捕捉时间断点的影响：

1. 构造断点变量

meta$post_election <- as.numeric(meta$numeric_date >= 7000)
# 构造日期与断点的交互项（以断点为中心调整日期）
meta$date_centered <- meta$numeric_date - 7000

2. 估计包含断点的效应

est_break <- estimateEffect(
  formula = c(9) ~ faction_id * date_centered * post_election,
  stmobj = model.fac.dat.int,
  metadata = meta,
  uncertainty = "None"
)

3. 绘制带断点的趋势图

# 绘制Greens的完整趋势（自动包含断点）
plot(
  est_break,
  covariate = "date_centered",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "Greens",
  linecol = "green",
  xlab = "Time (centered at election date)",
  ylim = c(0, 0.1),
  printlegend = F
)
# 添加CDU的趋势
plot(
  est_break,
  covariate = "date_centered",
  model = model.fac.dat.int,
  method = "continuous",
  moderator = "faction_id",
  moderator.value = "CDU",
  linecol = "red",
  add = T,
  printlegend = F
)
# 添加垂直断点线（对应x=0）
abline(v = 0, lty = 2, col = "gray50")

方法对比 #

• 方法1的优势是直观，完全对应你提出的分阶段思路，适合需要明确拆分时间段分析的场景；
• 方法2无需拆分数据集，通过模型公式直接纳入断点效应，更高效且能保持模型的整体性，同时可以直接检验断点交互项的显著性。

内容的提问来源于stack exchange，提问作者Elisa Benni