In v4.2.5, options_.qz_criterium is initialised to 1+1e-6 everywhere it is used (check.m, dynare_sensitivity.m, osr.m, stoch_simul.m), except when estimating models with lik_init=1, in which case it is set to 1-1e-6. This means that if you estimate a model, save the steady-state, then run the mod-file again with the estimation commented out, on the second run check may produce an error, even though it didn’t on the first run (since check doesn’t write over options_.qz_criterium if it’s already set).
What’s the justification for options_.qz_criterium being less than 1 in the lik_init=1 case? I presume there is one since there’s an explicit error if options_.qz_criterium>=1.
What should one think of models which pass the 1-1e-6 check, but not the 1+1e-6 one?
For the time being I’ve forced dynare_estimation_init.m to use 1+1e-6, as every other initialisation does, but I don’t feel particularly comfortable in this at the moment…
Any advice/thoughts would be much appreciated.