Adverse childhood experiences, non-response and loss to follow-up: Findings from a prospective birth cohort and recommendations for addressing missing data

Abstract

Adverse childhood experiences have wide-ranging impacts on population health but are inherently difficult to study. Retrospective self-report is commonly used to identify exposure but adult population samples may be biased by non-response and loss to follow-up. We explored the implications of missing data for research on child abuse and neglect, domestic violence, parental mental illness and parental substance use. Using 15 waves of data collected over 28 years in a population-based birth cohort, the Australian Temperament Project, we examined the relationship between retrospective self-reports of adverse childhood experiences and parent- and cohort-responsiveness at other time points. We then compared prevalence estimates under complete case analysis, inverse probability-weighting using baseline auxiliary variables, multiple imputation using baseline auxiliary variables, multiple imputation using auxiliary variables from all waves, and multiple imputation using additional measures of participant responsiveness. Retrospective self-reports of adverse childhood experiences were strongly associated with non-response by both parents and cohort members at all observable time points. Biases in complete case estimates appeared large and inverse probability-weighting did not reduce them. Multiple imputation increased the estimated prevalence of any adverse childhood experiences from 30.0% to 36.9% with only baseline auxiliary variables, 39.7% with a larger set of auxiliary variables and 44.0% when measures of responsiveness were added. Close attention must be paid to missing data and non-response in research on adverse childhood experiences as data are unlikely to be missing at random. Common approaches may greatly underestimate their prevalence and compromise analysis of their causes and consequences. Sophisticated techniques using a wide range of auxiliary variables are critical in this field of research, including, where possible, measures of participant responsiveness.