RStudio中逐步多元线性回归运行停滞问题求助

我有6个自变量和5个因变量，用R代码实现逐步多元线性回归：每次回归后移除p值>0.05的自变量，循环到仅保留p值<0.05的变量。代码对前3个因变量运行正常，但处理第4个因变量输出最终回归摘要时停滞；用示例数据测试时，仅处理第一个因变量就出现停滞。我已经用Excel手动完成所有回归并得到有效结果，附上示例数据和代码，请求排查问题。

示例数据 #

df <- structure(list(IndVar1 = c(114156L, 93447L, 113909L, 103983L, 
111403L, 93071L, 115366L, 98578L, 98487L, 105592L, 102727L, 108970L, 
112604L, 103223L, 95176L, 100051L), IndVar2 = c(131937L, 128287L, 
108688L, 94270L, 137057L, 100127L, 130482L, 109484L, 97535L, 
139141L, 122752L, 124523L, 111948L, 92846L, 104659L, 112592L), 
    IndVar3 = c(10967L, 9733L, 9485L, 10251L, 9426L, 6040L, 7413L, 
    6648L, 9862L, 6605L, 9364L, 8749L, 8621L, 10285L, 9700L, 
    8926L), IndVar4 = c(16260L, 13994L, 14936L, 11645L, 15831L, 
    11916L, 11227L, 14199L, 16051L, 15330L, 14205L, 13756L, 14175L, 
    12530L, 15572L, 15048L), IndVar5 = c(9772L, 8682L, 6849L, 
    10667L, 9223L, 7591L, 10003L, 7660L, 6431L, 8149L, 9111L, 
    10551L, 8448L, 10296L, 6291L, 10131L), IndVar6 = c(12643L, 
    15848L, 11522L, 11114L, 15946L, 16181L, 13360L, 11343L, 12682L, 
    13412L, 16175L, 11187L, 16494L, 14926L, 14569L, 12954L), 
    DepVar1 = c(1674605L, 1731694L, 1910052L, 1955383L, 1698939L, 
    1563494L, 1573640L, 1836676L, 1726679L, 1768668L, 1679431L, 
    1890720L, 1837590L, 1677332L, 1794722L, 1715671L), DepVar2 = c(12721L, 
    12169L, 9924L, 4358L, 9992L, 5290L, 11456L, 10956L, 7254L, 
    4089L, 12761L, 10360L, 4999L, 5387L, 8468L, 7446L), DepVar3 = c(2.55, 
    1.73, 1.08, 0.93, 1.88, 1.77, 2.62, 2.13, 2.32, 1.68, 0.99, 
    1.84, 1.4, 1.34, 0.93, 2), DepVar4 = c(42820L, 28394L, 38820L, 
    50756L, 30729L, 39528L, 30783L, 39551L, 38429L, 46858L, 29727L, 
    52171L, 37714L, 30164L, 43782L, 43410L), DepVar5 = c(2.7, 
    2.59, 1.93, 1.48, 2.03, 1.81, 1.63, 1.53, 2.7, 1.87, 2.33, 
    2.15, 2.09, 2.35, 2.48, 1.54)), class = "data.frame", row.names = c(NA, 
-16L))

原代码 #

# Function to perform stepwise regression
perform_stepwise_regression <- function(dependent_variable, independent_variables, data) {
  cat("======================================================================
")
  cat("Dependent Variable: ", dependent_variable, "
")
  cat("======================================================================
")
  
  # Fit initial model
  lm_model <- lm(as.formula(paste(dependent_variable, "~", paste(independent_variables, collapse = " + "))), data = df)
  
   while (TRUE) {
    # Check p-values
    p_values <- summary(lm_model)$coefficients[, 4]
    
    # Identify variables with the highest p-value > 0.05
    variable_to_remove <- names(p_values[p_values == max(p_values)])
    
    # Break the loop if no variable has a p-value > 0.05
    if (max(p_values) <= 0.05) {
      break
    }
    
    # Create a new formula excluding the variable
    predictors_to_include <- setdiff(names(lm_model$model)[-1], variable_to_remove)
    
    new_formula <- as.formula(paste(dependent_variable, "~", paste(predictors_to_include, collapse = " + ")))
    
    # Fit a new model
    lm_model <- lm(new_formula, data = df)
  }
  
  # Create the final formula using the single remaining independent variable
  final_formula <- as.formula(paste(dependent_variable, "~", predictors_to_include))
  
  # Fit the final model using the created formula
  final_model <- lm(final_formula, data = df)
  
 # Print the final summary in the console
  print(summary(lm_model))
}

# Example usage:
dependent_variables <- c("DepVar1", "DepVar2", "DepVar3", "DepVar4", "DepVar5")
independent_variables <- c("IndVar1", "IndVar2", "IndVar3", "IndVar4", "IndVar5", "IndVar6")

for (dependent_variable in dependent_variables) {
  perform_stepwise_regression(dependent_variable, independent_variables, df)
}

核心问题分析 #

• 截距项干扰死循环：原代码提取的p值包含截距项，当所有自变量p值都达标时，截距项的p值可能大于0.05，导致循环无法终止。
• 多变量移除逻辑错误：当多个自变量有相同最大p值时，variable_to_remove会变成向量，setdiff处理时可能意外删除多个变量，引发后续模型构建异常。
• 冗余的模型拟合：循环结束后已得到最终模型，无需重复拟合；且predictors_to_include在循环外可能未正确赋值（如初始模型就满足终止条件时）。

修复后的代码 #

# 逐步多元线性回归函数
perform_stepwise_regression <- function(dependent_variable, independent_variables, data) {
  cat("======================================================================\n")
  cat("因变量: ", dependent_variable, "\n")
  cat("======================================================================\n")
  
  # 初始化当前保留的自变量
  current_predictors <- independent_variables
  # 拟合初始模型
  lm_model <- lm(as.formula(paste(dependent_variable, "~", paste(current_predictors, collapse = " + "))), data = data)
  
  while (TRUE) {
    # 提取自变量的p值（排除截距项）
    coeff_summary <- summary(lm_model)$coefficients
    if (nrow(coeff_summary) <= 1) {  # 只剩截距项时终止
      break
    }
    p_values <- coeff_summary[-1, 4]
    
    max_p <- max(p_values, na.rm = TRUE)
    # 所有自变量p值都达标，终止循环
    if (max_p <= 0.05) {
      break
    }
    
    # 仅移除第一个p值最大的自变量
    variable_to_remove <- names(p_values)[p_values == max_p][1]
    # 更新保留的自变量列表
    current_predictors <- setdiff(current_predictors, variable_to_remove)
    
    # 拟合新模型
    lm_model <- lm(as.formula(paste(dependent_variable, "~", paste(current_predictors, collapse = " + "))), data = data)
  }
  
  # 输出最终模型结果
  print(summary(lm_model))
}

# 运行示例
dependent_variables <- c("DepVar1", "DepVar2", "DepVar3", "DepVar4", "DepVar5")
independent_variables <- c("IndVar1", "IndVar2", "IndVar3", "IndVar4", "IndVar5", "IndVar6")

for (dep_var in dependent_variables) {
  perform_stepwise_regression(dep_var, independent_variables, df)
}

修复说明 #

• 排除截距项干扰：提取p值时直接跳过截距项，只关注自变量的显著性。
• 单变量移除逻辑：当多个自变量p值相同时，仅移除第一个，避免一次删除多个变量导致模型不稳定。
• 清晰的变量追踪：用current_predictors直接管理当前保留的自变量，避免依赖模型内部的变量名，逻辑更直观。
• 增加边界终止条件：当模型只剩截距项时主动终止循环，彻底避免死循环。

内容的提问来源于stack exchange，提问作者Doctor K