ABSTRACT
Clinical trials comparing treatments that follow patients over a period of time often suffer from dropouts. In some cases, these dropouts are treatment related. It may be due to clinical improvement or deterioration. Such dropouts contribute partial information regarding the effectiveness of the treatment and should be included in the analysis appropriately. The methods proposed in the literature for such situations, such as the last observation carried forward method, the regression prediction method, etc., require assumptions that may be weak. When treatment related dropouts occur the distribution of the observations often resemble a truncated normal. Traditionally, dropouts would be treated as missing, and imputation methods would be applied for the missing data. Here, as an alternative, a method for analyzing the data under a truncated multivariate normal distribution is proposed. An Expectation-Maximization (EM) algorithm is applied to simplify the estimation of the truncated normal likelihood and to utilize standard software (such as SAS) for the analysis. A data set from a collaborative study conducted by the National Institute of Mental Health is used to illustrate the method. Some of the traditional methods are compared.