Multilevel and time-series missing value imputation for combined survey and longitudinal context data

Comparative research examining relationships between individual-level survey response data and time-varying country context variables for political or socioeconomic characteristics is often complicated by missing values. Surveys and longitudinal context measures may be produced during alternative years and at differing frequencies. Observations may be intermittent or may only cover few consecutive years across a full longitudinal sequence. Statistical evaluations that do not impute values with consideration to data’s missingness characteristics may produce biased estimates. Model-based approaches for missing value imputation such as multiple imputation and time series imputation offer means through which imputed values may be produced given complex hierarchical and longitudinal relations. Using incomplete survey data for institutional trust measures from 554,104 respondents from twenty-seven Eastern European and Central Asian countries between 1993 and 2016, and corresponding longitudinal context descriptors of demographic, socioeconomic and political conditions, multilevel multiple imputation and time-series imputation methods were compared and evaluated. Where missingness is intermittent across the breadth of longitudinal sequence, time series imputation may produce convincing estimates for national-level variables’ values while understating uncertainty associated with imputation. When missing values are numerous and span tail ends of a sequence, multivariate multilevel multiple imputation with time variable fixed effects may produce better estimates for country-variables through incorporation of information derived from additional covariates and other countries’ concurrent trajectories. Multilevel multiple imputation models with random slopes for time variables were found to have beneficial qualities in that countries’ unique longitudinal trends are emphasized and fit while that effects of pooled observations and additional covariates contribute to estimation.