I am estimating a multivariate ordered probit model using a composite marginal likelihood (CML) approach. In other words, I replace the full likelihood function by a surrogate likelihood constructed from pair-wise bivariate ordered probit models. The code is written in MATLAB and I use the solver fmincon to maximize my likelihood under constraints.
The code works perfectly and I recover the true parameters in my empirical exercise but I am struggling now to compute the standard errors.

My CML estimator is asymptotically normally distributed. The asymptotic variance-covariance matrix for the estimator is given by the inverse of the Godambe sandwich information matrix (Godambe, 1960) defined by:

\begin{align*} & V_{CML}(\theta)= [H(\theta)]^{-1} J(\theta) [H(\theta)]^{-1} \quad \text{with} \\ & H(\theta) = E \Big[ -\frac{\partial^2 \: \: log \: L_{CML}(\theta) }{\partial \theta \: \: \partial \theta'} \Big] \\ & J(\theta) = E \Big[ \left( \frac{\partial \: \: log \: L_{CML}(\theta) }{\partial \theta } \right)^{'} \left( \frac{\partial \: \: log \: L_{CML}(\theta) }{\partial \theta } \right) \Big] \end{align*}

I tried to use H as the Hessian at the estimated parameter values and J as the Gradient vector. Both are given by fmincon. When I then take the square root of the diagonal elements, I find very low standard errors which does not make sense to me.
Note that I am simulating the data.

Am I doing something wrong?

Thank you



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.