Open fedemolina opened 2 years ago
As soon as one has inequality restrictions, there is no longer a closed form solution for the estimated parameters. One could transform the parameter (e.g., replace beta_i by phi_i^2, estimate phi_i, and after the estimation calculate beta_i = \hat{phi}_i^2, which is always non-negative) and use nlsystemfit() for the estimation but nlsystemfit() is in an early development stage and frequently does not work well. Everyone is invited to contribute to the development of nlsystemfit().
The sistemfit package that it's used to estimate the aids model allow equality restrictions over coefficients (beta = 0).
Is it possible to add restrictions of inequality over coefficients (beta > 0)?
For example, glmnet package allow that by regularization but I guess that could bias the estimated coefficients and because of that standards errors could have no sense. Another example is used in the quantreg package via fnc method (Rb ≥qr)