Estimation and Specification Analysis of Models of Dividend Behavior Based on Censored Panel Data.

Dividends move in discrete jumps. Moreover, some companies pay dividends, others do not. Both these aspects necessitate the use of limited dependent variable models in the analysis of dividend behavior. Models of dividend behavior usually ignore these problems and treat dividends as a continuous variable. This paper analyzes dividend behavior using panel data on 649 firms for 12 years (1976-87). The model used is a censored regression model which allows for firm-specific and time effects. It is estimated using the maximum likelihood method under three different error covariance specifications. Based on specification tests, it is argued that it is important to allow for the zero observations, industry effects, and firm-specific and time effects in the estimation of models of dividend behavior.