Why is my margins command taking too long after melogit?

[Deepali godha]

Hello,

I have finalized a two-level logit model with melogit using Stata 15. The model has fixed effects at both levels and interactions but has no random effects. When I run my margins command, it is just taking too long. I ran a simple margins command which worked but the one that I really want is taking too long (hours)

melogit im i.matgr##i.anc  i.matgr##i.wlth i.matgr##i.rel  c.pc_A_c c.pc_B_c c.pc_C_c [pweight=pwt1]||SID:, cov(unstruct) from (a) vce(cluster SID)
margins matgr, vsquish post vce(unconditional)
* worked 
margins , dydx(matgr##anc) at (wlth=(0 1 2 3 4)) vsquish post vce(unconditional)
margins , dydx(matgr##anc) atmeans(c) at(wlth=(0 1 2 3 4) rel=2 rural=0) vsquish post vce(unconditional)

The above is just one of the commands I want to get the results for. I have tried several different ones. I thought that maybe the second margins command is asking too much so decided to fix some parameters in my third command but it has not helped so far. I understand that this being a multilevel model, margins will take more time but I am not sure how much time it should take.

I am now wondering if there is an error in my command. If not, is there some way to make this faster? I will really appreciate any understanding on this issue.

Thank you

[Deepali godha]

Please ignore the above post. It just takes time.

[Richard Williams]

If you can live without the standard errors, the nose option may speed things up.

[Deepali godha]

Thanks Richard- very helpful advice. Though I still need my CIs, it helps to get a quick insight.

[Weiwen Ng]

Deepali, it's not just meologit. Margins also runs slowly after latent class analysis. My guess is that when there are any sort of random effects involved, they take a longer time to produce standard errors. In your case, I can see you only have a random intercept, but you are asking for margins at quite a few different levels of -at-. Also, in your command as typed, you asked for unstructured covariance between random effects, and if you had more than one random effect, I bet that takes a lot of time as well (NB: in your command as typed, there's only a random intercept, so is unstructured covariance applicable?)

In latent class analysis, I am guessing that margins has to repeat its calculations over each latent class, which is what slows things down. My last margins command with 6 classes took maybe 30 minutes to run, which is not as slow as yours, but is still quite some time.

[Deepali godha]

Hmmm... That makes sense. Thanks for the detailed explanation Weiwen. I did get that standard errors were the main culprit behind, given that Richard's advice worked beautifully- but this helps in clarifying concepts. As for the unstructured variance, thanks for pointing that out. I just left it there from my previous commands- makes sense to remove it.