如何手动执行多重插补填补缺失数据？适配Pearson相关重测信度场景

Great questions! Let's tackle both your manual multiple imputation query and the specific test-retest reliability scenario step by step.

Multiple imputation works by creating several complete datasets, analyzing each, then pooling the results to account for uncertainty from missing values. Here's how to do it manually:

• Step 1: Map Missing Data Patterns
  First, document where missing values occur across your dataset. Classify the missingness type: Missing Completely at Random (MCAR), Missing at Random (MAR), or Missing Not at Random (MNAR). This guides your imputation approach—for example, cross-tabulate missing flags with other variables to spot correlations.
• Step 2: Choose an Imputation Method
  Pick a method aligned with your missingness type and data structure:
  • For numerical/ordered categorical data (like your 0-4 test scores), use bootstrap sampling to generate varied imputations. Instead of filling all missing values with a single mean/median, randomly draw values from the non-missing observations of the same variable (or same subgroup) for each imputation dataset.
  • For unordered categorical data, use mode-based bootstrap sampling similarly.
• Step 3: Generate Multiple Complete Datasets
  Aim for 5-10 datasets (enough to capture imputation uncertainty without redundant work). For each missing value slot, fill it with a unique sampled value in each dataset. For example, if a student has 6 missing test records, each dataset gets a different set of sampled scores for those slots.
• Step 4: Analyze Each Dataset
  Run your intended analysis (e.g., calculating Pearson correlations for test-retest reliability) on every complete dataset.
• Step 5: Pool Results
  Calculate the average of your key statistics (e.g., average correlation coefficient) across all datasets. Also compute the combined variance—this includes both the variance within individual datasets and the variance between datasets (to account for imputation uncertainty) to get a robust final estimate.

Your use case has specific constraints: 4 test items (0-4 scores), students with 2 to 8 test attempts, and a need to fill gaps to create full 8-attempt datasets for Pearson correlation-based reliability. Here's how to adapt the process:

1. Restructure Your Data #

Convert your data to a wide format: each row represents one student, with columns for each item in each test attempt (e.g., test1_item1, test1_item2, ..., test8_item4). Mark missing attempts with NA.

2. Define Imputation Logic Tailored to Test-Retest Reliability #

To preserve the student's true performance level and test-to-test variability:

• Student-Specific Imputation: For students with only 2 attempts, first calculate their mean score and standard deviation for each item across their existing tests. Use bootstrap sampling from their own non-missing item scores to generate the 6 missing attempts—this ensures imputed scores align with their observed performance.
• Subgroup Reference (If Needed): If a student's existing data is extremely limited (though you have 2 attempts here), group students by their total observed test scores (high/medium/low) and sample imputation values from peers in the same subgroup. This adds realistic variability while staying aligned with the student's ability level.

3. Generate Your Imputation Datasets #

Create 5-10 full datasets. For example:

• In Dataset 1, fill each missing item score with a random draw from the student's own observed scores for that item.
• In Dataset 2, draw from the corresponding subgroup's non-missing scores for that item.
• Repeat this variation across all datasets to capture different plausible imputation scenarios.

4. Calculate Test-Retest Reliability #

For each complete dataset:

1. Compute Pearson correlation coefficients between every pair of test attempts (e.g., test1 vs test2, test1 vs test3, ..., test7 vs test8).
2. Take the average of all these correlations as the dataset's test-retest reliability estimate.

• Pool the reliability estimates across all datasets: calculate the mean of the dataset averages, plus the combined variance to get a final reliability range.

Key Notes to Avoid Bias #

• Skip Single Imputation: Filling missing values once will underestimate uncertainty—multiple imputation is critical for reliable results.
• Respect Score Bounds: Since your scores are 0-4, never generate values outside this range. Bootstrap sampling from existing scores naturally enforces this, but if using distribution-based methods, round/clamp values to stay within bounds.
• Validate Imputations: Compare the distribution of imputed scores to non-missing scores (e.g., mean, standard deviation, frequency of each score) to ensure they don't skew your dataset.

内容的提问来源于stack exchange，提问作者wuijin