Relative Risk Ratio option reports different results

[feifeibu]

Hi,

I used the 'rrr' option to get the relative risk ratios after 'runmlwin' (for a multinomial response model). I realize that it gives different results compared with 'exp()' and 'lincom'. The differences are not really negligible , for example 0.6376 vs. 0.6277. Any idea why? Thanks!

Feifei

[ChrisCharlton]

-runmlwin- calculates this with the following code:

Code: Select all

nlcom (eform: exp([`eq']`par'))

Could you try the equivalent manually with your estimates to see if you get the same results?

[feifeibu]

Hi Chris,

Thank you for replying so promptly. Strangely, 'runmlwin, rrr' and 'lincom (or nlcom)' give the some relative risk ratios if I use MQL1. The problem occurs if I use MCMC.

Code: Select all

use http://www.bristol.ac.uk/cmm/media/runmlwin/bang, clear
generate lc1 = (lc==1)
generate lc2 = (lc==2)
generate lc3plus = (lc>=3)

runmlwin use4 cons lc1 lc2 lc3plus, level2(district: cons) ///
level1(woman)  ///
discrete(distribution(multinomial) link(mlogit) denom(cons) basecategory(4)) ///
nopause

runmlwin, rrr
lincom lc1_1,rrr
nlcom (exp(_b[lc1_1]))

runmlwin use4 cons lc1 lc2 lc3plus, level2(district: cons) ///
level1(woman)  ///
discrete(distribution(multinomial) link(mlogit) denom(cons) basecategory(4)) ///
mcmc(orthogonal) initsprevious ///
nopause

runmlwin, rrr
lincom lc1_1,rrr
nlcom (exp(_b[lc1_1]))

[ChrisCharlton]

If you're using MCMC then the difference that you are seeing makes more sense. To calculate the rrr option for MCMC results we calculate the exp() of each value in the parameter chain and report the mean of this as the estimate, rather than taking the exp() of the mean of the original values as in your original calculation.

[feifeibu]

That explains everything. Thank you very much!

Feifei