Latent variables in linear stochastic models : refletions on "maximum likelihood" and "partial least squares" methods