Lack of identifiability - scourge of statistical modelling
Late lunch talk by Trond Reitan
Lack of identifiability in a model is a problem that haunts statistical analysis. When using already tried and tested types of statistical analysis this is (normally) not a problem. But in analysis of various biological data there might be enough knowledge available that new and untried models are constructed. Personal experience suggests that when making new statistical models, identifiability problems pops up more often than what is comfortable. I will describe what identifiability (and lack thereof) is, give many examples of non-identifiable models and describe how to avoid or work around this problem, both in classic and Bayesian analysis.