[Limdep Nlogit List] NLogit converging problems compared with STATA

Julian Sagebiel jsagebiel at gmx.de
Mon Sep 6 02:57:21 EST 2010


Dear all,

I estimated a Conditional Logit model with NLogit 4.0 and with Stata 10.1. The data comes from a choice experiment with 800 respondents facing 9 choice sets each, and contains two alternatives per choice set. For testing purposes I estimated a very simple model with no respect to panel data and case specific variables. The results are ok but differ slightly in Stata and NLogit. In fact the results from Stata seem better (LL=-4249.105 in NLogit and LL= -3778.1856 in Stata, Pseudo R^2 is also higher in Stata). I am wondering what could be the reason for this difference. It seems that with NLogit, the log likelihood function does not converge properly. In Stata I used several methods namely Newton Raphson, BHHH, DFP and BFGS and checked for tolerance level. Not surprisingly, they all displayed the same LL value. I also heard about NLogit facing problems with coded data, so that I compared results with codes 0,1,2 and with the actual values, 0,15,30 for two attributes, 0.02, 0.05
 ,0.1 for the third attribute aso.  I would be very happy to solve this issue as I am wondering that if NLogit does not converge properly with the conditional logit it will also have problems with the Latent Class model, which I cannot estimate with Stata.
I am looking forward for fruitful answers.  
                                                           
Thank you so much in Advance

Julian

-- 
GMX DSL SOMMER-SPECIAL: Surf & Phone Flat 16.000 für nur 19,99 Euro/mtl.!*
http://portal.gmx.net/de/go/dsl


More information about the Limdep mailing list