|
Abstract:
|
Zero-inflated models comprise an important subclass of finite mixture models useful for data containing many zeros. Recently, several researchers extended these models to the longitudinal data case. Such a problem can arise (e.g., in IT system monitoring) where occurrences of error-related events in some time intervals are tracked. An error event may not happen often, inflating the number of zero counts. But, in rare occasions, an error can occur a many times in an interval, producing a potential outlier. The existing methods for such models provide no particular protection against the presence of outliers, which can influence parameter estimates and inferences substantially. We present a robust estimation approach designed to downweight influential observations. Properties of the proposed method and its comparison to others are discussed.
|
