Hidden bias in the use of archival data

Nonresponses in archival data may violate the missing-at-random assumption in ways difficult to detect. Standard methods of comparing sociodemographics of respondents and nonrespondents are inappropriate when the units of analysis are not also the individuals who maintain the archival record. Under these circumstances, the distribution of missing data may be correlated with the dependent variable and traits of the record keepers. This will distort relationships, especially when listwise deletion of missing values is used in multivariate analysis. Data are used from a large clinical chart study of mentally ill patients to demonstrate the process of identifying hidden bias and the implications of such bias.