变量间mice插补效果差异及优化方法咨询

Hey there! Let's work through your mice package imputation issues—those tricky variables 2 and 8 giving you trouble with CART, plus the PMM error. I've wrestled with similar problems before, so here's a breakdown of what might be happening and how to fix it:

Why Variables 2 and 8 Are Underperforming with CART Imputation #

First, let's unpack why these two variables aren't imputing well while the others are:

• Weak Correlations with Other Variables: CART builds prediction trees by splitting on variables that have predictive power for the missing values. If variables 2 and 8 don't have strong linear or non-linear relationships with your other 7 variables, the tree can't learn meaningful patterns to fill in the gaps. Grab a correlation matrix (use cor(data, use = "pairwise.complete.obs")) to check this—low correlations here are a big clue.
• Skewed or Outlier-Ridden Distributions: CART struggles with variables that have extreme outliers or highly skewed distributions. Since you have density plots, take a close look at the original non-missing data for these two variables—do they have long tails or rare extreme values? The tree might be overfitting to those outliers instead of capturing the typical value range.
• Insufficient Missing Data Sample Size: If variables 2 and 8 have way more missing values than the rest, or the number of non-missing observations is too small, CART won't have enough data points to build a reliable tree. Use md.pattern(data) to count missing values across all variables and confirm this.

Optimization Tips for CART Imputation #

Let's fix those underperforming variables:

• Tweak CART Hyperparameters: The default CART settings in mice might be overfitting or underfitting. Try adjusting tree depth and minimum leaf size to balance bias and variance. For example:

  # Limit tree depth and require at least 5 observations per leaf
  imp_cart <- mice(your_data, method = "cart", maxit = 5,
                   cartControl = list(maxdepth = 3, minbucket = 5))
  

• Add Auxiliary Variables (If Available): If you have other variables (even categorical ones not in your original 9) that relate to variables 2 and 8, throw them into the imputation model. More predictive signals will help the CART tree make smarter guesses.
• Transform Skewed Variables: If either variable is highly skewed, apply a transformation (log, square root, or Box-Cox) before imputation, then reverse-transform the imputed values afterward. This makes the distribution more manageable for CART to work with.

Fixing the Default PMM Error #

Now, about that PMM failure—here are the most common fixes:

• Check for Multicollinearity: PMM can crash if there's strong collinearity between variables (correlations >0.9). Use the correlation matrix I mentioned earlier to spot these pairs. You might need to remove one of the collinear variables or use PCA to reduce dimensionality first.
• Adjust the PMM Match Count: The default k (number of matches to draw from) might be too high for your dataset size. Try lowering it to 3 or 5:

  imp_pmm <- mice(your_data, method = "pmm", maxit = 5, pmm.k = 3)
  

• Zero Variance Check: If variables 2 or 8 have zero variance (all non-missing values are identical), PMM will throw an error. Run apply(your_data, 2, var, na.rm = TRUE) to check—if any variance is zero, you might need to exclude that variable or recode it to add some variation.

内容的提问来源于stack exchange，提问作者user196772