Combining proration and full information maximum likelihood in handling missing data in Likert scale items: A hybrid approach

Large-scale surveys are common in social and behavioral science research. Missing data often occur at item levels due to nonresponses or planned missing data designs. In practice, the item scores are typically aggregated into scale scores (i.e., sum or mean scores) for further analyses. Although several strategies to handle item-level missing data have been proposed, most of them are not easy to implement, especially for applied researchers. Using Monte Carlo simulations, we examined a practical hybrid approach to deal with item-level missing data in Likert scale items with a varying number of categories (i.e., four, five, and seven) and missing data mechanisms. Specifically, the examined approach first uses proration to calculate the scale scores for a participant if a certain proportion of item scores is available (a cutoff criterion of proration) and then use full information maximum likelihood to deal with missing data at the scale level when scale scores cannot be computed due to the selected proration cutoff criterion. Our simulation results showed that the hybrid approach was generally acceptable when the missing data were randomly spread over the items, even when they had different thresholds/means and loadings, with caution to be taken when the missingness is determined by one of the scale items. Based on the results, we recommend using the cutoff of 30% or 40% for proration when the sample size is small and the cutoff of 40% or 50% when the sample size is moderate or large.

Automated summary

Large-scale surveys in social and behavioral science research often face challenges with missing data at the item level. Through Monte Carlo simulations, a hybrid approach was proposed to handle missing data in Likert scale items, with recommendations to use a 30% or 40% cutoff for small sample sizes, and a 40% or 50% cutoff for moderate to large sample sizes.