An analysis is only as good as its data, and every researcher has struggled with dubious results because of missing data. In this article, I will cover three ways to deal with missing data.
Types of Missing Data
Understanding the nature of missing data is critical in determining what treatments can be applied to overcome the lack of data. Data can be missing in the following ways:
Missing Completely At Random (MCAR): When missing values are randomly distributed across all observations, then we consider the data to be missing completely at random. A quick check for this is to compare two parts of data – one with missing observations and the other without missing observations. On a t-test, if we do not find any difference in means between the two samples of data, we can assume the data to be MCAR.
Missing At Random (MAR): The key difference between MCAR and MAR is that under MAR the data is not missing randomly across all observations, but is missing randomly only within sub-samples of data. For example, if high school GPA data is missing randomly across all schools in a district, that data will be considered MCAR. However, if data is randomly missing for students in specific schools of the district, then the data is MAR.
Not Missing At Random (NMAR): When the missing data has a structure to it, we cannot treat it as missing at random. In the above example, if the data was missing for all students from specific schools, then the data cannot be treated as MAR.
1. Mean or Median Imputation
When data is missing at random, we can use list-wise or pair-wise deletion of the missing observations. However, there can be multiple reasons why this may not be the most feasible option:
There may not be enough observations with non-missing data to produce a reliable analysis
In predictive analytics, missing data can prevent the predictions for those observations which have missing data
External factors may require specific observations to be part of the analysis
In such cases, we impute values for missing data. A common technique is to use the mean or median of the non-missing observations. This can be useful in cases where the number of missing observations is low. However, for large number of missing values, using mean or median can result in loss of variation in data and it is better to use imputations. Depending upon the nature of the missing data, we use different techniques to impute data that have been described below.
2. Multivariate Imputation by Chained Equations (MICE)
MICE assumes that the missing data are Missing at Random (MAR). It imputes data on a variable-by-variable basis by specifying an imputation model per variable. MICE uses predictive mean matching (PMM) for continuous variables, logistic regressions for binary variables, bayesian polytomous regressions for factor variables, and proportional odds model for ordered variables to impute missing data.
To set up the data for MICE, it is important to note that the algorithm uses all the variables in the data for predictions. In this case, variables that may not be useful for predictions, like the ID variable, should be removed before implementing this algorithm.
Data$ID <- NULL
Secondly, as mentioned above, the algorithm treats different variables differently. So, all categorical variables should be treated as factor variables before implementing MICE.
You can also ignore some variables as predictors or skip a variable from being imputed using the MICE library in R. Additionally, the library also allows you to set a method of imputation discussed above depending upon the nature of the variable.
3. Random Forest
Random forest is a non-parametric imputation method applicable to various variable types that works well with both data missing at random and not missing at random. Random forest uses multiple decision trees to estimate missing values and outputs OOB (out of bag) imputation error estimates.
One caveat is that random forest works best with large datasets and using random forest on small datasets runs the risk of overfitting. The extent of overfitting leading to inaccurate imputations will depend upon how closely the distribution for predictor variables for non-missing data resembles the distribution of predictor variables for missing data. For example, if the distribution of race/ethnicity for non-missing data is similar to the distribution of race/ethnicity for missing data, overfitting is not likely to throw off results. However, if the two distributions differ, the accuracy of imputations will suffer.
The MICE library in R also allows imputations by random forest by setting the method to “rf”. The authors of the MICE library have provided an example on how to implement the random forest method here.
To sum up data imputations is tricky and should be done with care. It is important to understand the nature of the data that is missing when deciding which algorithm to use for imputations. While using the above algorithms, predictor variables should be set up carefully to avoid confusion in the methods implemented during imputation. Finally, you can test the quality of your imputations by normalized root mean square error (NRMSE) for continuous variables and proportion of falsely classified (PFC) for categorical variables.