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

Re: st: AW: Tobit, negative predictesd values

From   Austin Nichols <>
Subject   Re: st: AW: Tobit, negative predictesd values
Date   Tue, 10 Nov 2009 11:00:33 -0500

If you want predictions in [0,1], a new model is probably in order:

On Tue, Nov 10, 2009 at 10:44 AM, Nick Cox <> wrote:
> In a different vein, but I think consistently with Maarten's suggestions:
> 1. You are fitting a hyperplane in some space. That could stay entirely above the origin but that's a tall order if many responses are at or very near zero. Perhaps your model could do with some (more?) curvature....
> 2. Otherwise put, if you believe that no predicted responses can plausibly be negative, this functional form won't ensure that.
> 3. Plot observed vs predicted and residual vs predicted to get some handles on where your model is misbehaving.
> Nick
> Maarten buis
> That is perfectly consistent with the logic behind the -tobit-
> model: It assumes that there is some latent variable (the ideal
> intensity of training), but this can only be realized when this
> ideal is larger than some cut-off point, in this case 0.
> The default for the predicted values are these ideal levels of
> training (the linear predictor). So, what you found is that
> some employers ideally would want to take training away from
> their employees. If you think that that is not a senisible
> interpretation then the -tobit- may not be the suitable model
> for your situation.
> ---  Solorzano Mosquera, Jenniffer wrote:
>> I estimated a tobit model having intensity labor training
>> as dependent variable and a group of firm characteristics
>> which are presumpted as determinants of that intensity.
> <snip>
>> However I've been looking and I found that heavy censoring
>> causes these kind of problems on predicted values and even
>> worse when high proportion of censored cases is the situation.

*   For searches and help try:

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