I have tried to compare efficiency of the same hierarchical model but one with cholesky decomposition tweak while other using multi-normal. Agains expectations the multi-normal version shaved off 1000s per 2000 iterations and gave no divergencies while cholesky version was slover and gave 45 divergencies. Is this expected, what version is generally recommended - manual recommends cholecky but my results are contradictory. Perhaps I had to few parameters and lots of data. Both models are attached for those who may be interested.
cholesky version doa3.stan (2.8 KB)
multi-normal version doa5.stan (2.5 KB)