3

I am interested in developing a predictive multiple regression model which predicts a concentration of one compound based on the measured concentrations of several other compounds. Both the dependent and independent variables contain "non-detects", which results in occasional (~20% of total observations) left censored observations.

What methods, codes, or papers should I explore to develop this model?

Ryan Conway
  • 41
  • 2
  • 1
    An interesting problem. The measurements below the detection threshold in the response could maybe be modeled with a mixture model of a Bernoulli distribution and whatever distribution underlies the non-missing response. If there are no true zeroes, you could use a hurdle model instead. The values below the detection threshold in the explanatory variables could maybe be modeled by including an interaction with an indicator variable for whether the original variable is below the threshold. – Frans Rodenburg Nov 28 '19 at 06:51

0 Answers0