关于R语言mice包插补泰坦尼克数据集缺失年龄的变量选择疑问

Hey there! Let's walk through how to choose variables for missing value imputation with the mice package, specifically for your Vanderbilt Titanic dataset where you're trying to fill in missing Age values.

The goal here is to pick variables that have a meaningful relationship with Age (the variable with missing values)—these will give the imputation model the best information to predict missing values. You also want to avoid redundant or highly correlated variables that can mess up model stability.

1. Analyze Relationships Between Age and Other Variables #

First, figure out which variables actually correlate with Age:

• Numeric variables: Use correlation coefficients to check the strength of linear relationships. For example, to compare Age and Fare:

  cor(training_data$Age, training_data$Fare, use = "complete.obs")
  

• Categorical variables (like your pclass factor): Use visualizations or statistical tests to spot differences. A boxplot will quickly show if Age varies by cabin class:

  boxplot(Age ~ pclass, data = training_data)
  

  For your Titanic data, pclass is a no-brainer here—first-class passengers were typically older than those in lower classes.

2. Avoid Multicollinearity #

If two variables are highly correlated (e.g., SibSp and Parch, both measuring family size), including both doesn't add new information and can make the imputation model unstable.

• For numeric variables, use correlation matrices to spot high correlations.
• For a formal check, use variance inflation factors (VIF) with car::vif() (you'll need to fit a preliminary regression model first). Pick one of the correlated variables instead of both.

3. Lean on Domain Logic (Titanic-Specific) #

Think about the real-world context of the Titanic dataset—these variables are likely to be useful for predicting Age:

• pclass: Cabin class directly ties to socioeconomic status, which correlates with age
• Sex: Historical data shows gender differences in passenger age distributions
• Fare: Tied to cabin class, so it indirectly reflects age trends
• SibSp/Parch: Passengers with children or siblings on board are likely to be in middle age
• Embarked: Port of embarkation can relate to passenger demographics and age

4. How to Specify Variables in mice #

By default, mice() uses all other variables to impute missing values, but you can customize this with the predictorMatrix parameter:

1. Generate a default predictor matrix to start with:

   pred_matrix <- mice::make.predictorMatrix(training_data)
   

   This is a square matrix where rows are variables to impute, columns are predictor variables. A 1 means the column variable is used to predict the row variable; 0 means it's not.
2. Adjust the matrix to only use your chosen variables for Age imputation. For example, if you want to use pclass, Sex, Fare, and SibSp:

   # Reset all entries to 0 first
   pred_matrix[,] <- 0
   # Allow selected variables to predict Age
   pred_matrix["Age", c("pclass", "Sex", "Fare", "SibSp")] <- 1
   # Ensure Age doesn't predict itself (already 0 by default, but good to confirm)
   pred_matrix["Age", "Age"] <- 0
   

3. Run mice with your custom matrix:

   imputed_data <- mice(training_data, predictorMatrix = pred_matrix, m = 5, seed = 123)
   

   Here, m=5 creates 5 imputed datasets (standard practice for multiple imputation), and seed=123 ensures your results are reproducible.

5. Validate Your Imputation #

After running the imputation, check if the results make sense:

• Plot the density of imputed Age values to compare with the observed data:

  densityplot(imputed_data, ~Age)
  

• Summarize the imputed values to check for outliers or unrealistic values:

  summary(imputed_data$imp$Age)
  

• Fit a model (like logistic regression for survival prediction) on each imputed dataset and use pool() to combine results—stable coefficients across datasets mean your imputation is reliable.

内容的提问来源于stack exchange，提问作者Richard Golz