Censored data & Tobit Regressions

[Hina Sharma]

I am studying the impact of highways on deforestation. I am using annual data on forest cover (example provided below). I define my deforestation/forest loss variable as a change in forest cover and if the change is zero or positive my deforestation/forest loss variable is zero. Does this imply truncated or censored data and thus requiring the use of a Tobit regression? I am hesitant to use Tobit because of the problems of including fixed effects.

* Example generated by -dataex-. For more info, type help dataex
clear
forest_cover forest_change loss
 3.5    .  .
   3  -.5 .5
   1   -2  2
3.75 2.75  0
3.75    0  0
   5 1.25  0
   3   -2  2
   3    0  0
 2.5  -.5 .5
   3    3  0
end

[Andrew Musau]

It appears that you have uncensored data (forest_change) and you are doing the censoring yourself by creating the variable "loss". Thereafter, you ask for an estimation method to correct for the censoring. Most people who use these sample selection models can only wish that they had access to the uncensored data. Feel free to ignore my advice, but I recommend that you use the uncensored data. Granted that you want to study deforestation, but that is simply a framing issue. Stating that a kilometer of new road construction decreases reforestation by XX hectares is the same thing as saying that it increases deforestation by XX hectares. So use the data on forest change (positive, zero and negative) and frame your conclusions in terms of deforestation.

[Hina Sharma]

Thank you, appreciate the response. I agree looking at forest change is indeed the best and most efficient way of utilizing the data I have.