Abstract

Response models with latent parameters generally lack identifiability, which means that different combinations of values for their parameters may imply the same distribution for the observed data. If this happens, observed data do not allow us to distinguish between these alternative parameter values, and any attempt at statistical inference with respect to the model breaks down (San Martín, Volume Two, Chapter 8). Generally, identifiability problems are caused by an unbalance between the numbers of model equations and their unknown parameters. As both are fixed for a given response matrix, the only option left to make the model identifiable is to restrict the parameters further by adding extra equations to the system.