Skip to main content

Table 2 Summary of methods compared in simulation

From: Multiple imputation of multiple multi-item scales when a full imputation model is infeasible

Method description

Assumptions

Comments

1. Exclude observations with any missing data from the analysis

Missing values are independent of Morisky score given the other variables

Complete case analysis

2. For partially observed scales, sum the observed values, weighted by (1/proportion of items observed). Exclude observations with completely missing scales

Partially observed items are MAR given other items in the scale and completely missing scales are MCAR

Effectively single imputation as the mean of observed items within a scale

3. For partially observed scales, set the score to missing. Multiply impute the scale sums from a multivariate normal model with Morisky score and age as covariates

Missingness is MAR, and this process is the same for missing scales or missing items within scales

Wasteful of observed data

4. For partially observed scales, sum the observed values, weighted by (1/proportion of items observed). For completely missing scales, multiply impute the scale sums from a multivariate normal model with Morisky score and age as covariates

Completely missing scales are MAR

Uses single imputation in the same way as approach 2

5. Multiply impute missing items based on the total of the other scale, and the other items within the scale for the item being imputed (with Morisky score and age as covariates). This requires the use of chained equations with linear regression imputation rather than a multivariate normal model

Missing at random for both variables, but that the regression for one item on the other scale items is the same as the regression on the other scale total

Proposed adaptation

6. Multiply impute all items using all other items via a multivariate normal model, including Morisky score and age as covariates

Multivariate normality

It is in some senses the benchmark