* @brief Compute the diagnostic statistics * */
| 33 | * |
| 34 | */ |
| 35 | AnyType stateToResult( |
| 36 | const Allocator &inAllocator, const ColumnVector &inCoef, |
| 37 | const ColumnVector &diagonal_of_inverse_of_hessian, |
| 38 | double logLikelihood, const Matrix &inHessian, |
| 39 | int nIter, const ColumnVector &stds) |
| 40 | { |
| 41 | MutableNativeColumnVector coef( |
| 42 | inAllocator.allocateArray<double>(inCoef.size())); |
| 43 | MutableNativeColumnVector std_err( |
| 44 | inAllocator.allocateArray<double>(inCoef.size())); |
| 45 | MutableNativeColumnVector waldZStats( |
| 46 | inAllocator.allocateArray<double>(inCoef.size())); |
| 47 | MutableNativeColumnVector waldPValues( |
| 48 | inAllocator.allocateArray<double>(inCoef.size())); |
| 49 | |
| 50 | for (Index i = 0; i < inCoef.size(); ++i) { |
| 51 | coef(i) = inCoef(i) / stds(i); |
| 52 | std_err(i) = std::sqrt(diagonal_of_inverse_of_hessian(i)) / stds(i); |
| 53 | waldZStats(i) = coef(i) / std_err(i); |
| 54 | waldPValues(i) = 2. * prob::cdf( prob::normal(), |
| 55 | -std::abs(waldZStats(i))); |
| 56 | } |
| 57 | |
| 58 | // Hessian being symmetric is updated as lower triangular matrix. |
| 59 | // We need to convert diagonal matrix to full-matrix before output |
| 60 | Matrix full_hessian = inHessian + inHessian.transpose(); |
| 61 | full_hessian.diagonal() /= 2; |
| 62 | for (Index i = 0; i < inCoef.size(); ++i) |
| 63 | for (Index j = 0; j < inCoef.size(); ++j) |
| 64 | full_hessian(i,j) *= stds(i) * stds(j); |
| 65 | |
| 66 | // Return all coefficients, standard errors, etc. in a tuple |
| 67 | AnyType tuple; |
| 68 | tuple << coef << logLikelihood << std_err << waldZStats << waldPValues |
| 69 | << full_hessian << nIter; |
| 70 | return tuple; |
| 71 | } |
| 72 | |
| 73 | // ------------------------------------------------------------ |
| 74 |