Handling non-ignorable dropouts in longitudinal data: A conditional model based on a latent Markov heterogeneity structure