Generally ARMA modelling involves nonlinear optimization techniques. Based on an initial estimate estimation involves an iterative process until optimality is achieved. The error message is an indication of a problem encountered in the process. This might signal model inadequacy.

Generally ARMA modelling involves nonlinear optimization techniques. Based on an initial estimate estimation involves an iterative process until optimality is achieved. The error message is an indication of a problem encountered in the process. This might signal model inadequacy.

Although I am not an expert, but I sense some basic misunderstandings leading to the command msg. The parameters are too many. It clearly shows that Stata is beginning to choke. This appears to be putting lots of constants into variables. Using scalars is much better style and less wasteful of storage. You must rewrite your program so that you are working in terms of the log likelihood throughout. Your program calculates the likelihood directly, before logging it, and the numbers become far too big to handle.

*

Goodluck

I encountered very similar problem while computing '.tebalance overid' test for IPW treatment effect in STATA 16.0.

['could not calculate numerical derivatives -- discontinuous region with missing values encountered']

I found suprisingly easy solution: first, I identified the variable that made this problem. I realised that it was expresed in relatively high units comparing to my dependent treatment var ( the latter was binary 0-1 while the var causing problem was measured in thousands) I simply divided this var by 1000 that obviously didn't change antything in my model, but the problem was solved and the error message hasn't occured any more.

Ways to Diagnose the problem

I also had the error message "could not calculate numerical derivatives -- flat or discontinuous region encountered" for another nonlinear model. This error message is somehow general. As mentioned by Bazyli Czyzewski, too large values are one cause. I suggest here some simple tricks to identify easy to solve possible sources of the problem in your moptimize code.

i. Check that you have NO missing value in your dataset.

ii. Check if your likelihood does not face a value outside of the support of your pdf. A possible cause of this is if you have a parameters in your pdf that is restricted to take positive values. Make sure that these parameters are exponentiated in your program (obviously the resulting estimate will be interpreted as log of the parameter of interest). Another cause is if your pdf has an absolute value of a parameter but you ignore this perhaps because you believe that the estimated parameter will eventually take a positive value.

iii. print the log likelihood values of the first few observations to see on which observation you have a missing value. Manually calculate the loglikelihood of that observations (if you have a complicated pdf, calculate separately the different parts of your pdf and see which part if causing the error).