Inconsistent estimation of average marginal effects for mlogit

[David Ralf Simon]

Hi everyone!
I am running a multinomial logistic regression with mlogit and want to obtain the average marginal effects for all outcomes.
My estimation is the following:

mlogit EKC b4.Env_Degradation i.Region GDP POPURB INSSOC INDUSTRY TRADE FDI RENEWABLE, vce(cluster study)
margins, dydx (*) pr(out(1))
margins, dydx (*) pr(out(2))
margins, dydx (*) pr(out(3))

whereby EKC is a categorical variable with three different outcomes.
When I rerun the exact same code several times, I receive the same coefficients, but the standard errors are slightly different every time.

In a similar regression with slightly different control variables I sometimes receive the error: "Warning: variance matrix is nonsymmetric or highly singular", but when I rerun the exact same regression again, the estimation often works.

Is this a general problem in Stata when calculating marginal effects for mlogit? Or is there something wrong with my code?
I ran this in both Stata14 and Stata17 and standard errors were changing in both versions.

Thank you for your help!

[David Ralf Simon]

In case someone will encounter the same problem in the future.
Finally, I found the cause of this weirdly inconsistent estimation of standard errors. There is one dummy variable, which is roughly 98% 1, 2% 0. When I exclude this variable from the model, all regressions run perfectly fine and there are no more changing standard errors.

[Richard Williams]

As a sidelight, unless you are using an ancient version of Stata (13 or older), there is no need to specify results outcome by outcome. Stata 14+ will do them all at once for you. For more, see

https://www3.nd.edu/~rwilliam/xsoc73994/Margins05.pdf