Missingness Mechanism Analysis

Missingness mechanism analysis relates each variable's missing-data indicator to the observed variables with group comparisons and logistic regression to identify predictors of missingness.

변수마다 결측 여부 지시변수를 만들고 관측된 다른 변수와 집단 비교·로지스틱 회귀로 관련성을 따져 결측을 예측하는 변수를 찾는 방법입니다.

Income is missing for 18% of my respondents. How can I tell which observed variables are related to whether income is missing?

응답자의 18%가 소득을 비워 뒀어요. 관측된 변수 중 무엇이 소득 결측과 관련 있는지 어떻게 알 수 있나요?

Create an indicator for missing income and compare every other variable between respondents with and without income using Welch t tests or chi-square tests, then fit a logistic regression of the indicator on those variables. Variables related to missingness show the data are not missing completely at random and should be used as auxiliary variables in imputation or FIML.

When to use Missingness Mechanism Analysis

• You are preparing multiple imputation or FIML and need to choose auxiliary variables.
• You want to describe how respondents with missing values differ from those without.
• Little's MCAR test was rejected and you want to know which variables drive it.

When to use something else

• You want a single omnibus test of MCAR; use Little's MCAR test instead.
• You need to distinguish MAR from MNAR; that cannot be done from the observed data, so use a sensitivity analysis such as a tipping-point analysis instead.

Assumptions and requirements

• Observations are independent.
• The logistic model uses rows where all other variables are observed.
• Many comparisons are reported without multiplicity correction and are exploratory.

What the analysis reports

• Welch t with Cohen's d or chi-square with Cramér's V for each target–predictor pair
• Logistic regression coefficients and p-values for numeric predictors, and the model likelihood-ratio chi-square
• List of predictors suggested as auxiliary variables

How to report it

Missingness on [variable] was related to [predictor] (Welch t = [value], p = [value], Cohen's d = [value]); a logistic regression of the missingness indicator on observed variables gave χ²([df]) = [value], p = [value].

Related methods

• Little's MCAR Test
• Missing Data Pattern Analysis
• Multiple Imputation (MICE) with Pooled Regression

Run Missingness Mechanism Analysis on Quorant