Hi,
I am using stata 16.

I am running an ivtobit model. Data is shown below. The deaths data is censored below the value of 9 so for all censored data i replaced the missing data with the number 9. Smoke and Fire are both measured in percentages.

Code:
* Example generated by -dataex-. To install: ssc install dataex
clear
input float panelid int deaths float(pred smoke fire)
1  40 111.82269           0             0
1  36 156.26273   3.7378564             0
1  41  66.95786    .5494103             0
1  33  53.73887           0             0
1  46  57.47389     .361936             0
1  33  60.35159 .0004871885             0
1  39  73.95678    3.254674             0
1  46  92.15742    3.379441             0.5783204
1  33  62.73242           0             0
1  34 71.866066           0             0
1  37  122.6314    3.287671             0
1  41 133.70306    6.575342             0
1  42  112.9622           0             0
1  35  74.42028    3.335207             0
1  41  63.43359    3.287683             0
1  41  77.39986   2.5153205             0
1  35 70.794304   1.8029382             0.0341351
1  37  56.74924    .5541319             0
1  41  74.91539    3.747238             0
1  33 119.90632           0             0
1  28   39.7854           0             0
1  32  205.2623    37.00338             0
The model I run takes into account the censoring.

Code:
ivtobit deaths temp precip pm25median (smoke = fire), ll(9) vce(cluster panelid)
I notice when i produce the fitted values using:

Code:
predict pred, xb
some of the fitted values are negative. Given the deaths data cannot be negative as well as the data relating to fire and smoke (as they are percentages) have I made a mistake with my use of the ivtobit model? Or is the linear prediction unable to provide fitted values that make sense in this case given this is a nonlinear model?

In addition my ivtobit results show that I reject the null of a non endogenous variable

Code:
Wald test of exogeneity (corr = 0): chi2(1) = 9.52        Prob > chi2 = 0.0020
But I was wondering do other diagnostic tests such as a Ramsey Reset test to look at the specification of my first stage work with non linear regressions like tobit or poisson?