This will mean that the MLE will have the same result every time. Currently the differences between runs are small, but may confuse a user that doesn't understand numerical root-finding methods.
Alternative would be to round all results by an amount appropriate for the numerical precision of results. If random initial seed is causing differences above an beyond this, the this suggests there is something wrong with the model/code, so may be more helpful to identify issues than simple setting the seed to be fixed
This will mean that the MLE will have the same result every time. Currently the differences between runs are small, but may confuse a user that doesn't understand numerical root-finding methods.
Alternative would be to round all results by an amount appropriate for the numerical precision of results. If random initial seed is causing differences above an beyond this, the this suggests there is something wrong with the model/code, so may be more helpful to identify issues than simple setting the seed to be fixed