1

I am working with a negative binomial regression. The data frame contains 38 predictors and 48 records. After variable selection I used only 8 variables. Finally, I got some good results. Here is the plot of modeled versus actual data (red: Test-sample and black Train): enter image description here

Now I want to go a step further: I want that all my estimated predictor coefficients (fitted by the NBGLM) stay non negative (In my model there is one variable with negative coefficient but for practical use I don't want this case).

Is there a version of the NBGLM where I can introduce thresholds for the parameter estimates?

Antoine
  • 5,740
  • 7
  • 29
  • 53
R_FF92
  • 251
  • 2
  • 9
  • Please provide more details and possibly a data example (raw data, plots etc.). – Tim Oct 21 '15 at 13:55
  • @Tim Hope this is edit is useful for you! – R_FF92 Oct 21 '15 at 14:14
  • *Why* do you need the constraints? I ask because this sounds like an example of "how to lie with statistics"... – Tim Oct 21 '15 at 14:24
  • I know what you mean because I want to change something after seeing the results of modelling. The predictors describe complexity levels and the dependent variable a price. The higher the complexity, the higher the price. In practical use we want to find a model which makes sense in the way that a higher complexity goes with a higher price for all predictors. So I wanted to try a modification of the model with paramter constraints even the statistics say something different...Does it make sense? – R_FF92 Oct 21 '15 at 14:51
  • OK, but maybe your assumptions are wrong, maybe it is not so (and the model is right)? – Tim Oct 21 '15 at 14:54
  • If I look at the Spearman Correlation of the predictor which have this negative estimate I see a positive correlation. What could be reasons for this? – R_FF92 Oct 21 '15 at 15:00
  • See e.g. http://stats.stackexchange.com/questions/116804/coefficient-changes-sign-when-adding-a-variable-in-logistic-regression – Tim Oct 21 '15 at 15:06
  • Thank you :) By the way: You prefer using only the test sample for checking predictive power or should I do cross validation also? – R_FF92 Oct 21 '15 at 15:12

0 Answers0