Longitudinal survey studies allow researchers to describe and make inference about the dynamics of health changes in the aging process that can not be drawn from cross-sectional data. However, complicated sampling issues are involved in utilizing these longitudinal data, e.g. informative dropout and cohort heterogeneity. In addition, comparing to a subject from a more recent cohort, a subject from an earlier cohort must survive longer to be included in the study. This cause selection bias and leads to biased estimate of the population longitudinal trajectory. To adjust the bias we use a maximum likelihood joint modeling approach, in which the selection process and informative dropout is modeled by a Cox model with left truncation, and the change of functional status over age is modeled by a simple linear mixed model. EM algorithm is used to estimate model parameters.