如何检验二次分析中缺失数据是否为完全随机缺失（MCAR）

Great question—testing for MCAR is a critical first step when dealing with missing data, especially with such a high missing rate (50%) and 10-year follow-up period. Below are practical, actionable methods tailored to your study context:

1. Compare Baseline & Trial-Period Characteristics Between Complete and Missing Groups #

MCAR means the probability of missing data is unrelated to any variables (observed or unobserved). A straightforward check is to split your sample into two groups: those with complete 10-year hypertension data, and those without. Then compare these groups on all variables you have available:

• For continuous variables (e.g., baseline age, baseline blood pressure), use independent samples t-tests (if normally distributed) or Mann-Whitney U tests (for non-normal data).
• For categorical variables (e.g., gender, intervention group, baseline comorbidities), use chi-square tests (or Fisher’s exact test if cell counts are small).
• Pro tip for your study: Don’t limit comparisons to baseline—also check trial-period variables (e.g., 2-year intervention adherence, in-trial blood pressure changes) since these might predict long-term follow-up loss.
• Critical note: If you test multiple variables, apply a multiple comparison correction (like Bonferroni) to avoid false positive results.

Example R code for group comparisons:

# Assume your dataset is named 'study_data', with a binary flag 'missing_10yr' (1 = missing 10yr hypertension data, 0 = complete)
# Continuous variable: Baseline age
t.test(study_data$baseline_age ~ study_data$missing_10yr)

# Categorical variable: Intervention group
chisq.test(table(study_data$intervention_group, study_data$missing_10yr))

2. Use Little’s MCAR Test (Global Statistical Test) #

Little’s test is a formal hypothesis test designed specifically to assess MCAR. The null hypothesis is that your data is MCAR; a non-significant p-value (>0.05) supports this assumption.

• It’s easy to implement in R using packages like mice or LittleMCAR.
• Caveat for your study: With 50% missing data, this test may have limited power or produce unstable results, especially if many variables have missing values. Use it as a complement to group comparisons, not the sole test.

Example R code with mice:

library(mice)
# Run Little's MCAR test on your dataset
mcar_test(study_data)

3. Visualize Missing Data Patterns #

Visualization helps you intuitively spot if missingness clusters with specific variables or groups:

• Use md.pattern() from the mice package to generate a matrix showing all missing data combinations in your dataset.
• Use aggr() from the VIM package to create a heatmap or bar chart highlighting which variables have the most missing data, and whether missingness correlates with other variables.

Example R code for visualization:

library(VIM)
# Generate a missing data aggregation plot
aggr(study_data, 
     col = c('navyblue', 'yellow'), 
     numbers = TRUE, 
     sortVars = TRUE, 
     labels = names(study_data), 
     cex.axis = 0.7, 
     gap = 3, 
     ylab = c("Missing Data", "Pattern"))

Additional Recommendations for Your Study #

• If tests suggest data is not MCAR (e.g., significant group differences on baseline variables), avoid complete-case analysis—it will bias your results. Instead, use methods for Missing at Random (MAR) like multiple imputation.
• Even if data is MCAR, 50% missingness means you’ll lose substantial statistical power with complete-case analysis. Multiple imputation is still a better approach to retain all available data.

内容的提问来源于stack exchange，提问作者Vincent