Hi,
I am having trouble figuring out which global holds the Hessian (the information matrix) while estimating a model using maxlik.
As far I can figure out from the documentation _max_HessCov should hold the matrix, but it seems that _max_HessCov returns a scalar, while it should return a KxK matrix.
Please let me know if I am missing something here.
Thanks and Regards
Annesha
2 Answers
0
First, _max_covPar must equal 3 and then _max_HessCov will contain the covariance matrix of the parameters, not the Hessian. If you want the Hessian it will be stored in _max_FinalHess.
If _max_CovPar = 3, and _max_HessCov is a scalar missing value, then the Hessian failed to invert indicating a linear dependency in the model. To diagnose this problem check the eigenvalues of the Hessian stored in _max_FinalHess. The Hessian must be positive definite, i.e., no negative eigenvalues. If there are eigenvalues equal to or nearly equal to zero, then it is singular and you have a linear dependency. To analysis this see the final section in the article, Optimization with the Quasi-Newton Method.
0
Thanks for your answer.
I have couple more questions here then.
1. Does _max_FinalHess varies depending on what value I specify for _max_CovPar? It believe the answer is yes.
2. If yes then what are the three different matrices supplied by _max_FinalHess for
_max_CovPar = 1, _max_CovPar = 2, and _max_CovPar = 3.
I understand that _max_CovPar = 1 calculates covariance of the parameters by inverse of hessian, _max_CovPar = 2 calculates covariance of the parameters
by outer product of the score and _max_CovPar = 3 calculates covariance of the parameters by Hessian_inv*outer_product*Hessian_inv.
I look forward to your response.
Thanks
Annesha
Your Answer
2 Answers
First, _max_covPar must equal 3 and then _max_HessCov will contain the covariance matrix of the parameters, not the Hessian. If you want the Hessian it will be stored in _max_FinalHess.
If _max_CovPar = 3, and _max_HessCov is a scalar missing value, then the Hessian failed to invert indicating a linear dependency in the model. To diagnose this problem check the eigenvalues of the Hessian stored in _max_FinalHess. The Hessian must be positive definite, i.e., no negative eigenvalues. If there are eigenvalues equal to or nearly equal to zero, then it is singular and you have a linear dependency. To analysis this see the final section in the article, Optimization with the Quasi-Newton Method.
Thanks for your answer.
I have couple more questions here then.
1. Does _max_FinalHess varies depending on what value I specify for _max_CovPar? It believe the answer is yes.
2. If yes then what are the three different matrices supplied by _max_FinalHess for
_max_CovPar = 1, _max_CovPar = 2, and _max_CovPar = 3.
I understand that _max_CovPar = 1 calculates covariance of the parameters by inverse of hessian, _max_CovPar = 2 calculates covariance of the parameters
by outer product of the score and _max_CovPar = 3 calculates covariance of the parameters by Hessian_inv*outer_product*Hessian_inv.
I look forward to your response.
Thanks
Annesha