Best Practices

Quick reference for using MICE correctly.

Basic Setup

from imputation import MICE, configure_logging
from plotting.diagnostics import plot_chain_stats
from plotting.utils import md_pattern_like

# Enable logging
configure_logging(level='INFO')

# Check missing patterns
pattern = md_pattern_like(df)
print(pattern)

# Set random seed for reproducibility
import numpy as np
np.random.seed(42)

Method Selection

# Use different methods for different variables
method_dict = {
    'numeric_normal': 'pmm',
    'numeric_skewed': 'midas',
    'categorical': 'cart',
    'complex': 'rf'
}
mice.impute(method=method_dict)
Quick guide:

  • PMM: General numeric data

  • CART: Categorical or non-linear

  • RF: Complex patterns

  • MIDAS: Skewed or small samples

  • Sample: Initial imputation or simple cases

Always Check Convergence

from plotting.diagnostics import plot_chain_stats

plot_chain_stats(mice.chain_mean, mice.chain_var,
                 save_path='convergence.png')

Look for flat, stable chains in later iterations.

Compare Distributions

from plotting.diagnostics import stripplot, densityplot

missing_pattern = df.notna().astype(int)

stripplot(mice.imputed_datasets, missing_pattern)
densityplot(mice.imputed_datasets, missing_pattern)

Ensure imputed values are within reasonable range.

Proper Pooling

Always use Rubin’s rules:

# Correct
mice.fit('outcome ~ predictor')
pooled = mice.pool(summ=True)
print(pooled)

# Check FMI
print(f"Max FMI: {pooled['FMI'].max():.3f}")
Never:

  • Use only one imputed dataset

  • Average the imputed datasets

  • Use standard analysis on single imputation

Common Mistakes

Mistake 1: Too few imputations

# Don't do this
mice.impute(n_imputations=5)  # Often not enough

# Do this
mice.impute(n_imputations=20)  # Better

Mistake 2: Not checking convergence

# Always check
plot_chain_stats(mice.chain_mean, mice.chain_var)

Mistake 3: Using single imputation

# Don't do this
single_dataset = mice.imputed_datasets[0]
model = fit_model(single_dataset)  # Wrong!

# Do this
mice.fit('y ~ x')
pooled = mice.pool(summ=True)  # Correct

Mistake 4: Imputing after transformations

# Don't do this
df['log_income'] = np.log(df['income'])
mice = MICE(df)  # Imputes both income and log_income separately!

# Do this
mice = MICE(df[['income', 'other_vars']])
mice.impute(n_imputations=20)
# Then create transformations after imputation
for dataset in mice.imputed_datasets:
    dataset['log_income'] = np.log(dataset['income'])

Variable Selection

Include in imputation:

  • All variables in your analysis model

  • Variables that predict missingness

  • Variables correlated with incomplete variables

# If analyzing: income ~ age + education
# Impute: income, age, education, plus auxiliary variables

mice = MICE(df[['income', 'age', 'education',
                 'occupation', 'zip_code']])  # auxiliaries

Predictor Matrix

For most cases, use default (all variables predict each other):

mice.impute(n_imputations=20)  # Uses default predictor matrix

For custom control:

import numpy as np

predictor_matrix = pd.DataFrame(1, index=df.columns, columns=df.columns)
np.fill_diagonal(predictor_matrix.values, 0)

# Customize as needed
predictor_matrix.loc['var1', 'var2'] = 0

mice.impute(predictor_matrix=predictor_matrix)

See Predictor Matrices for details.

Performance Tips

For large datasets:

# Use faster methods
mice.impute(method='cart')  # Faster than RF

# Use quickpred to reduce predictors
from imputation.utils import quickpred
pred_matrix = quickpred(df, mincor=0.3)
mice.impute(predictor_matrix=pred_matrix)

For many variables:

# Automatic predictor selection
pred_matrix = quickpred(df, mincor=0.2, minpuc=0.1)
mice.impute(predictor_matrix=pred_matrix)

Essential Checklist

Before finalizing:

☐ Checked missing patterns ☐ Used appropriate methods ☐ Ran sufficient iterations (≥20) ☐ Created enough imputations (≥20) ☐ Checked convergence ☐ Compared observed vs imputed distributions ☐ Used proper pooling (Rubin’s rules) ☐ Set random seed for reproducibility

See Also