R语言：构建模型解释面板数据中行业组间FDI差异的方法

Hey there! Let's walk through how to build your panel data model in R step by step—since you're aiming to explain FDI differences across industry groups using both industry-specific and non-industry-specific variables, we'll cover everything from data prep to model selection and diagnostics.

First, install and load the packages you'll need. plm is the go-to for panel data analysis, dplyr helps with data cleaning, and lmtest/sandwich handle robust standard errors and model tests:

install.packages(c("plm", "dplyr", "lmtest", "sandwich"))
library(plm)
library(dplyr)
library(lmtest)
library(sandwich)

Make sure your data has two key identifiers: an industry group ID (like industry_id) and a time column (like year). This tells R it's dealing with panel data.

• Check for missing values first—you don't want those messing up your model:

  # Load your data (replace with your actual file path)
  panel_raw <- read.csv("your_fdi_data.csv")
  
  # Find rows with missing values
  missing_rows <- panel_raw %>% filter_all(any_vars(is.na(.)))
  print(missing_rows)
  

  Handle missing data by imputing values or dropping rows (just be cautious with dropping—only do it if the missingness is random).

• Convert your data to a panel-specific object so R recognizes its structure:

  panel_data <- pdata.frame(panel_raw, index = c("industry_id", "year"))
  

Now let's test different panel models, starting with basics and moving to more robust options. Your core formula will be:
FDI ~ RW + TFP + IY + CY + GDP + LP + IR + ER + PS

3.1 Pooled OLS (Baseline) #

This treats all observations as independent, which is simple but ignores industry-specific differences. We'll add clustered standard errors to account for industry-level correlation:

pooled_model <- plm(FDI ~ RW + TFP + IY + CY + GDP + LP + IR + ER + PS,
                    data = panel_data, model = "pooling")

# Get clustered robust standard errors (cluster by industry)
pooled_clustered <- coeftest(pooled_model, vcov = vcovHC(pooled_model, cluster = "group"))
summary(pooled_clustered)

3.2 Fixed Effects (FE) Model #

This controls for unobserved, time-invariant industry characteristics (like a sector's inherent attractiveness to FDI). Important note: If non-industry variables (IR, ER, PS) are identical across industries in the same year (e.g., national interest rates), the FE model will drop them—because there's no within-industry variation to estimate.

To work around this, add interaction terms between macro variables and industry-specific traits. This lets you estimate how different industries respond to macro changes:

# Basic FE model (may drop non-industry variables if they have no within-industry variation)
fe_model <- plm(FDI ~ RW + TFP + IY + CY + GDP + LP + IR + ER + PS,
                data = panel_data, model = "within")
summary(fe_model)

# FE model with interaction terms (captures industry-specific responses to macro variables)
fe_interact_model <- plm(FDI ~ RW + TFP + IY + CY + GDP + LP + IR*RW + ER*TFP + PS*IY,
                         data = panel_data, model = "within")
summary(fe_interact_model)

3.3 Random Effects (RE) Model #

This assumes unobserved industry effects are random rather than fixed. It lets you include time-invariant variables, but you need to test if this assumption holds:

re_model <- plm(FDI ~ RW + TFP + IY + CY + GDP + LP + IR + ER + PS,
                data = panel_data, model = "random")
summary(re_model)

3.4 Hausman Test (Choose FE vs RE) #

This test tells you whether fixed effects or random effects is the better fit. A significant p-value means you should go with FE:

hausman_test <- phtest(fe_model, re_model)
print(hausman_test)

Don't skip this—you need to ensure your results are reliable:

• Heteroskedasticity check: Use the Breusch-Pagan test. If significant, use robust standard errors:

  bptest(fe_model) # Test for heteroskedasticity
  fe_robust <- coeftest(fe_model, vcov = vcovHC(fe_model, type = "HC1")) # Robust SE
  

• Serial correlation check: Use the Breusch-Godfrey test. If present, use HAC robust standard errors:

  pbgtest(fe_model) # Test for serial correlation
  fe_hac <- coeftest(fe_model, vcov = vcovHAC(fe_model)) # HAC SE
  

• If non-industry variables are purely macro-level (no variation across industries), their direct effect can't be estimated in FE models—but interaction terms are a great alternative to show how industries react differently to macro changes.
• If your panel is unbalanced (some industries missing years), plm handles this automatically, but double-check that missingness isn't biased (e.g., no systematic drop-out of low-FDI industries).
• Consider adding a time trend (e.g., + year) to control for linear changes over time that affect all industries.

内容的提问来源于stack exchange，提问作者Restemayer