Multinomial logit model- marginal effects (how to get output in word)

[andy macrobarty]

Dear All,

I have a question regarding out of multinomial logit model and marginal effects. When I say marginal effects, I do not mean relative risk ratio (RRR).

I mean how to get an out for this function into word directly. Please assist me soon

global xlist JQ JR PBB3 FE Independent1 LSALES LFCF2 lntenure lnd
mlogit $ylist $xlist ib3.Industry11_num ib1.advisor11_num

outreg2 using myreg.doc, replace ctitle(Model 1)



how about these marginal effects, how to get word output?

margins, dydx(*) atmeans predict(pr outcome(1))
margins, dydx(*) atmeans predict(pr outcome(2))
margins, dydx(*) atmeans predict(pr outcome(3))
margins, dydx(*) atmeans predict(pr outcome(4))


Regards,
Andy

[John Ngombe]

Andy, try the following,
1st: margins, dydx(*) atmeans post predict(pr outcome(1)) *I have included "post" before predict so that u get marginal effects for outcome 1. U have to do so for all outcomes or else you will only get regression coeffs which are not marginal effects.
2nd: You can now use your outreg2, e.g., outreg2 using my_results, word
**follow the same procedure for each outcome, I have skipped the 'replace' option in the outreg2 command because if you don't, the final outcome will not have all the marginal effects added to one word document, they will be replacing the previous outcomes. Besides, maintain the same name of the document (i.e my_results in my example) so that you end up having one document with all marginal effects for each outcome except the base.

Best!

[andy macrobarty]

John, Thanks a lot for providing this and it really helped me and saved a lot of time. I have a last question.

Unlike, RRR or mlogit command, it does relate relative to the base category

This the margins for outcome 1 tells probability of select 1 outcome . 1% increase in sales leads to (10%* coefficient size of sales)increase in outcome 1


However, we say in RRR, relative to the base category. Am I right?

[John Ngombe]

Thus great Andy. For your question, even the marginal effects (MEs) are interpreted relative to the base category. For example, ME of 0.1, in your example would be interpreted as; keeping other things constant, an increase in sales (if variable sales is an independent variable) by 1 unit (I don't know the units for sales that you are using) would lead to an increase in likelihood/probability of XXXX choosing or being in 1st category more than in the base(I don't know your base) by 0.1. In short, MEs from discrete choice models are interpreted as probabilities and relative to the base category.

Best!

[andy macrobarty]

Now, I am getting confused again, as I thought marginal effects or even predicted probabilities is not relative to the base outcome as I can commute this as well

outcome 3 is my reference category

margins, dydx(*) atmeans predict(pr outcome(3))

http://www.statalist.org/forums/foru...inner-question


https://www.youtube.com/watch?v=iqypob4My4o&t=157s

it does not talk about reference category

[John Ngombe]

Andy, don't get confused...as long as category 3 is your reference, there is no need for you to compute its marginal effects because even the marginal effects computed by Stata are with reference to that base category. Notice that you don't even get coefficients for the base category in your mlogit output, which justifies why you don't need their marginal effects. Read more on this or listen carefully to one of Ann Katchova's videos you shared...for marginal effects, it's exactly as I explained above, they are relative to the base category. Marginal effects are not same as predicted probabilities, the later are the actual values of probabilities you will get for including parameter estimates into the Multinomial logistic cdf at every observation in your sample...Thanks.

[John Ngombe]

If those links don't mention the base category, it is for simplicity purposes...and if you do the same, it's not wrong, but then mention in your results that the base category was XXXXX. Good luck.

[andy macrobarty]

I agree with mlogit yes that they do not give a reference category. If that's the case, then why does significance changes sometimes for marginal effects than computing mlogit

Sorry, for disturbing you. I definitely need to read more on this

[John Ngombe]

I am not sure if I got your question. Statistical significance will change every time you change the reference category. One thing is that the standard errors for margins are not computed in the same way as the coefficients themselves. Standard errors are computed by the delta method. However, if you repeat the same commands, statistical significance will not change unless you change the reference category.

[andy macrobarty]

I got your idea. Last bit is whenever we mention marginal effects, do we say always refer as percentage points?

This implies that an increase of 1 unit in 3-year cash flow volatility raises the probability of selecting cash flow only by 3.6 percentage points. The probability of firms operating total property return in mercer, deloitte, pwc and rerais 0.27, 0.25, 0.31 and 0.35 higher than the group which is included. Are these both right?