Comments and answers for "Tobit regression"
https://developer.ibm.com/answers/questions/217847/tobit-regression.html
The latest comments and answers for the question "Tobit regression"Answer by SystemAdmin
https://developer.ibm.com/answers/answers/217855/view.html
Since one model is nested in the other with one parameter different, the difference of the log likelihoods is asymptotically chi-squared with 1 df. However, in that simple case, you might as well go with the z statistic that comes directly from the output. It surprises me, though, that the two likelihoods are identical but the Wald statistics are very different. Perhaps that is a typo.<br>
<br>
Your ultimate decision, though, depends on the plausibility of the model and normal practice in your field. The statistics just provide evidence for your decision.<br>
<br>
Regards,<br>
Jon PeckWed, 31 Aug 2011 23:44:17 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217854/view.html
Hi,<br>
<br>
I would be immensely grateful if you explain this in a "Tobit regression/statistics for semi-dummies" manner! <br>
<br>
I have two Tobit regression models. (the first one with 4 independent variables and the other with 3 independent variables which excludes one of the variables in the first one). With the below Tobit regression results, how should I choose which model is better?<br>
First model:<br>
<br>
Scale: 0.2720<br>
Residual d.f.: 90<br>
Log likelihood: -50.968 D.f.: 6<br>
Wald statistic: 59.158 D.f.: 4<br>
<br>
Second model:<br>
<br>
Scale: 0.3005<br>
Residual d.f.: 91<br>
Log likelihood: -50.968 D.f.: 5<br>
Wald statistic: 35.090 D.f.: 3<br>
Many thanks.Wed, 31 Aug 2011 21:46:13 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217853/view.html
That is just the equivalent of the ordinary regression residual variance. It's significance would be irrelevant.<br>
<br>
HTH,<br>
Jon PeckWed, 06 Jul 2011 09:43:11 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217852/view.html
Hi again...<br>
<br>
Another Tobit regression question...I guess it is more statistics again... <br>
<br>
What does 'log (scale)' at the bottom of Tobit regression results table mean and how we can determine whether it is statistically significant or not? Also 'Scale' and 'Log likelihood'... Many thanks...Wed, 06 Jul 2011 00:27:54 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217851/view.html
Thanks very much... I brushed up my knowledge of statistics, revised my inital model and afterwards the choice of independent variables for the Tobit regression and now I get sensible results... Many thanks...Fri, 10 Jun 2011 21:29:02 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217850/view.html
Whatever distribution you choose is specifying the distribution of the dependent variable given the distribution parameters derived from the regression equation. So under normality, you are specifying that the error distribution is normal(0, sigma2) and hence that y is normal(X*beta, sigma2). That is the ordinary regression assumption.<br>
<br>
As for what to choose, why are you using TOBIT if your data are bounded in (0,1)? Why not use ordinary logistic regression?<br>
<br>
Tobit is intended for the case where you have an ordinary linear regression model but the data are censored at, typically, zero. In that case, the ordinary regression model is inappropriate, but tobit is a one-sided truncation.<br>
<br>
HTH,<br>
Jon PeckFri, 10 Jun 2011 00:02:12 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217849/view.html
Thanks very much for your reply. I know this is more a statistics question but I really need to get it right and would be very grateful for your help:<br>
<br>
1- If we choose Gaussian/normal for the dependant variable, is it assuming the normality of the error distribution for the Tobit regression or is it assuming that the dependant variable itself has normal distribution? <br>
<br>
2- How should I be sure which distribution is the best to choose for Tobit regression? For example I get much better z-values when I choose logistic distribution rather than normal. Does that mean it is better to use that one? <br>
<br>
(My dependant variable is a set of efficiency scores (between 0 to 1) for different decision making units. ) <br>
<br>
Thanks a lot.Thu, 09 Jun 2011 23:50:18 GMTSystemAdminAnswer by SystemAdmin
https://developer.ibm.com/answers/answers/217848/view.html
The distribution choice is about generalized linear models as with the Statistics GENLIN command. See the Algorithms for GENLIN if you want to get the mathematical details.<br>
<br>
With ordinary regression, you are typically assuming that the dependent variable is normally distributed with mean X*beta and a constant variance. You could alternatively assume that y is, for example, Poisson distributed with, again, a parameter X*beta. That would be Poisson regression for count data.<br>
<br>
Traditional Tobit models assume normality, but the SPSSINC TOBIT procedure (and underlying R package) allow you to choose other distributions.<br>
<br>
So, in summary, choosing normality for Tobit is no different from what you usually assume with linear regression.<br>
<br>
HTH,<br>
Jon PeckThu, 09 Jun 2011 18:36:55 GMTSystemAdmin