[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: L&F's fitstat command, contradictory results

From   John Reynolds <[email protected]>
To   [email protected]
Subject   st: L&F's fitstat command, contradictory results
Date   Thu, 10 Sep 2009 15:18:32 -0400

Dear Statalist members,

I am running two nested multinomial logit models where the DV has 3 categories. The "saved" model is a baseline model, the "current" model adds 12 interaction terms. When I run Long & Freese's fitstat command to compare the goodness of fit for the two models, I get opposite conclusions depending on whether I use change in the LRS or change in the BIC'. Change in the LRS suggests a significantly better fit for the current or interaciton model, but the change in BIC' "provides very strong support for the saved model." Have others of you run into this situation? Relevant output copied below, thanks for your insights or advice.

                              Current             Saved        Difference
Model:                          mlogit            mlogit
N:                               45593             45593                 0
Log-Lik Intercept Only      -46975.389        -46975.389             0.000
Log-Lik Full Model          -42533.017        -42583.047            50.029
D 85066.035(45533) 85166.093(45557) 100.059(24) LR 8884.743(58) 8784.685(34) 100.059(24)
Prob > LR                        0.000             0.000             0.000
BIC                        -403389.654       -403547.056           157.402
BIC'                         -8262.548         -8419.949           157.402
BIC used by Stata            85709.685         85552.284           157.402

Difference of  157.402 in BIC' provides very strong support for saved model.

John Reynolds
Prof. & Graduate Director, Sociology
Director, Pepper Institute on Aging & Public Policy
Florida State University
636 W. Call St., Room 207
Tallahassee, FL 32306-1121
850-644-8825[o], 850-320-2403[c]
[email protected]

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index