Other
Handling Item Nonresponse in Youth Surveys: Insights from the VACS.
Roba A, Annor F ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This is a methodological paper about survey data quality and statistical handling of missing responses in the Violence Against Children Surveys (VACS). It examines how different approaches to dealing with item nonresponse—such as listwise deletion and various imputation methods—affect prevalence estimates for violence indicators. The study finds that missingness is minimal and random across ten VACS conducted between 2018-2022, and that standard imputation is unnecessary for these surveys.
Key points
- Item nonresponse averaged only 0.56% across surveys, with 90% of respondents providing complete data for all variables
- Missingness was predominantly random and associated with country income level, education, and survey year, but not with sex or reported violence
- Missing data handling approaches (MCAR, NOMCAR, listwise deletion) produced similar prevalence estimates with no statistically significant differences
- Listwise deletion was least precise but differences were small; routine imputation was found unnecessary given minimal missingness
Limitations
This paper is a technical methodological study focused on statistical approaches to survey data quality rather than findings about children's violence experiences or outcomes. It does not evaluate interventions to improve child well-being or report substantive findings about violence prevalence or impacts on children.
Abstract
Abstract Background Although item nonresponse is typically minimal in surveys that use electronic data capture tools, the inclusion of “Don’t Know” or “Declined” response options can increase overall missingness. This study examines patterns of item nonresponse and evaluates how different missing data handling approaches affect prevalence estimates in the Violence Against Children Surveys (VACS). Methods We pooled data from ten VACS conducted between 2018 and 2022 and characterized missingness mechanisms and item nonresponse rates while identifying predictors of response propensity and missingness. Weighted prevalence estimates for 24 composite and single-item indicators were compared across four analytic approaches—listwise deletion (LDA), missing completely at random (MCAR), not missing completely at random (NOMCAR), and a sensitivity analysis—using an imputed dataset as the reference. Precision and accuracy were assessed using relative standard error (RSE) and relative bias. Results The overall survey response rate was 90% (n = 45,764). Item nonresponse averaged 0.56%, and 90% of respondents provided complete data for all 24 variables. Missingness was predominantly random (MCAR) and was associated with country income level, education, and survey year, but not with sex, perceived survey worthiness, response rates, or reported violence. Compared with the imputed reference (RSE = 4.16; relative bias = 0.0), listwise deletion produced the least precise estimates (RSE = 4.36). Sensitivity analyses yielded the largest relative bias (5.39%), whereas MCAR and NOMCAR approaches produced nearly identical estimates with comparable precision. Differences across methods were small and not statistically significant (p = 0.9996). Conclusions Given minimal and random missingness in key survey variables in the VACS, MCAR or NOMCAR approaches produce precise estimates, making routine imputation unnecessary. Further simulation studies might help determine thresholds at which imputation becomes beneficial.