Can anyone tell me how to calculate standard errors(and confidence intervals) for predicted probability after multivariable logistic regression for longitudinal/panel data (xtlogit) ? My y is a dicotomous variable 0,1, while I have 3 explanatory variables in the model. I do not need confidence intervals for coefficient nor for OR, but for the predicted probability I can obtain with the command predict y, pu0 (hope this predictions are right).
The 'predictnl' command with the se() option can be used to obtain these.
Thanks to both answers.
As regards Diwakar, I think I need nonlinear predictions and standard errors for predictions!
To both and possibly others: I can predict probability with the predict command, but standard errors only for the linear prediction.
As regards Nikolaj, I wonder the right nonlinear expression for non linear extimation after logistic regression for longitudinal/panel data.
Hi Laura,
This should do the trick:
predictnl pr1 = exp(xb(#1))/(1+exp(xb(#1))) , se(SERR1)
or, in the scale of xb:
predictnl xb = xb(#1) , se(SERR2)
or, maybe simpler:
predict xb, xb
predict SERR2, stdp
The confidence interval would then be:
gen pright = exp(xb + 1.96*SERR2))/(1+ (exp(xb + 1.96*SERR2)))
gen pleft = exp(xb - 1.96*SERR2))/(1+ (exp(xb - 1.96*SERR2)))
I find strategies like the ones suggested above much too error prone. Invest a little time in learning how to use Stata's -margins- command (see -help margins- and the corresponding manual entry, Mike Mitchell has a whole book devoted to -margins- and -marginsplot-: "Interpreting and Visualizing Regression Models Using Stata"). -margins- and -marginsplot- are very powerful, yet easy to use.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis