Effect of Missing Data on Classification Error in Panel Surveys
Sensitive outcomes of surveys are plagued by wave nonresponse and measurement error (classification error for categorical outcomes). These types of error can lead to biased estimates and erroneous conclusions if they are not understood and addressed. The National Crime Victimization Survey (NCVS) is a nationally representative rotating panel survey with seven waves measuring property and violent crime victimization. Because not all crime is reported to the police, there is no gold standard measure of whether a respondent was victimized. For panel data, Markov Latent Class Analysis (MLCA) is a model-based approach that uses response patterns across interview waves to estimate false positive and false negative classification probabilities typically applied to complete data.
This article uses Full Information Maximum Likelihood (FIML) to include respondents with partial information in MLCA. The impact of including partial respondents in the MLCA is assessed for reduction of bias in the estimates, model specification differences, and variability in classification error estimates by comparing results from complete case and FIML MLCA models. The goal is to determine the potential of FIML to improve MLCA estimates of classification error. While we apply this process to the NCVS, the approach developed is general and can be applied to any panel survey.