Concepts of MCAR, MAR, MNAR

• Missing completely at random (MCAR): the probability of being missing is the same for all cases
  • Cause of missing is unrelated to the data
• Missing at random (MAR): the probability of being missing only depends on the observed data
  • Cause of missing is unrelated to the missing values
• Missing not at random (MNAR): probability of being missing depends on the missing values themselves

Listwise deletion and pairwise deletion

• Listwise deletion (also called complete-case analysis): delete rows which contain one or more missing values
  • If data is MCAR, listwise deletion produces unbiased estimates of means, variances, and regression weights (if need to train a predictive model)
  • If data is not MCAR, listwise deletion can severely bias the above estimates.
• Pairwise deletion (also called available-case analysis)
  • Mean and variance of variable \(X\) are based on all cases with observed data on \(X\)
  • Covariance and correlation of \(X\) and \(Y\) is based on all data which both \(X\) and \(Y\) have non-missing values

Mean imputation

• Compared with the observed data, in the imputed data (observed + imputed values)
  • Standard deviations decrease
  • Correlation decreases
  • Means can be biased if the data is not MCAR.

Regression imputation

1. Build a regression model from the observed data
2. Impute the missing values in the response variable with the predicted values from the fitted regression

• The impute values are the most likely values under the model
  • However, it decreases the variance of the target variable
  • And it increases the correlations between the target and covariates
• Regression imputation, and its modern incarnations in machine learning is probably the most dangerous of all ad-hoc methods

Stochastic regression imputation

1. Build a regression model from the observed data
2. Impute a missing value in the response variable with the predicted value plus a random draw from the residual

• Preserves variance and correlation.
• Imputed values can exceed the range (e.g., a negative Ozone level). A more suitable model may resolve this.

LOCF and BOCF

• Last observation carried forward (LOCF) and baseline observation carried forward (BOCF) are for longitudinal data.

• LOCF can yield biased estimation even under MCAR.

Indicator method

• Not for imputation, but for building predictive models
• Only works for missing in covariates, not the target variables

Summary of ad-hoc imputation methods

• Note: the unbiasness of regression coefficients are assess with the variable containing missing values as the target variable

Multiple imputation creates \(m>1\) complete datasets

• Three steps of multiple imputation
  1. Imputation
  2. Analysis: train separate models
  3. Pooling: variance among \(m\) parameter estimates combines the conventional sampling variance (within-imputation variance) and the extra variance caused by the missing data (between-imputation variance)

Why using multiple imputation?

• It provides a mechanism to deal with the inherent uncertainty of the imputations
• It separate the solution of the missing data problem from the solution of the complete-data problem (train predictive models on complete data)

Multiple imputation example using the mice package

## Load the mice package
library(mice); 
## Impute 20 times, using preditive mean matching
imp <- mice(airquality, seed = 1, m = 20, print = FALSE)
## Fit linear regressions
fit <- with(imp, lm(Ozone ~ Wind + Temp + Solar.R))
## Pooled regression estimates
pander(summary(pool(fit)))

References

• Van Buuren, S. (2018). Flexible Imputation of Missing Data, 2nd Edition. CRC press.

  • https://stefvanbuuren.name/fimd/