Classification error in measuring sexual victimization among inmates

Abstract

All survey estimates are subject to measurement error - classification error in the case of categorical outcomes. This is especially true of sensitive outcomes such as sexual victimization. The National Inmate Survey (NIS), sponsored by the U.S. Bureau of Justice Statistics, is a nationally representative survey of inmates in prisons and jails which measures two types of sexual victimization - inmate-on-inmate and staff sexual misconduct with an inmate. This paper builds on the research of Berzofsky, Biemer, and Kalsbeek (2014) to present the results of a latent class analysis (LCA) designed to assess the measurement error in each type of sexual victimization. LCA uses multiple indicators of a construct embedded in the survey instrument to estimate the false positive and false negative probabilities in each indicator. Due to the rare nature of sexual victimization among inmates, our analysis combines data from the 2007-08 NIS and the 2009 NIS in order to achieve adequate precision in the results. One issue with LCA is how missing data for indicators and grouping variables are taken into account. Traditionally, if either type of variable was missing the model would use listwise deletion and remove the case from the model. Newer software, such as LatentGold, incorporates full information maximum likelihood (FIML) for dependent variables to utilize all records. This paper assesses the impact that the inclusion of cases with missing data have on the LCA estimates. Using MAR adjustments for missing data, we found evidence that inmates who do not respond to all indicators are more likely to be victims and more likely to not provide truthful responses for the items they do answer.