Alternative algorithms for the estimation of dynamic factor, mimic and varying coefficient regression models (Q790575): Difference between revisions

A state-space model is considered consisting of two sets of equations: transition equations: \(x_ t=\phi_ tx_{t-1}+\gamma_ tz_ t+G_ tv_ t\); \(\phi_ t:\) (\(j\times j)\); \(\gamma_ t:\) (\(j\times k)\); \(G_ t:\) (jxm); \(x_ t:\) unobservable (\(j\times 1)\) state vector; \(z_ t:\) exogenous or lagged dependent (\(k\times 1)\) vector; \(v_ t:\) (\(m\times 1)\) disturbance. The measurement equations are: \(y_ t=\alpha_ tx_ t+\beta_ tz_ t+e_ t\), where \(y_ t\) is measured; \(e_ t:\) disturbance; \(\left( \begin{matrix} v_ t\\ e_ t\end{matrix} \right)\sim NI(0\), \(\left( \begin{matrix} Q_ t\\ 0\end{matrix} \begin{matrix} 0\\ R_ t\end{matrix} \right))\); and the parameters \(\phi\),\(\gamma\),G,\(\alpha\),\(\beta\),Q,R are assumed here time invariant. The log-likelihood function of the unknown parameters is: \(L(\theta)=-\sum^{T}_{t=1}(\log | H_ t| +\eta^ T_ tH_ t^{-1}\eta_ t)\) where the ''innovations'' \(\eta_ t=y_ t-E(y_ t| y_{t-1},...,y_ 1,z_ t,...,z_ 1)\), whereas the matrix \(H_ t\) denotes the variance of \(\eta_ t.\) A scoring method of estimation is given using first derivatives of L (a zig-zag approach is suggested: using k th iteration values \(\alpha^ k,\phi^ k,R^ k,Q^ k\) to determine \(\beta^ k\) and \(\gamma^ k\), these values being then used to find \(\alpha^{k+1},\phi^{k+1},R^{k+1},Q^{k+1}\), and so on until convergence). The EM consists of two steps: Estimation and maximization are iterated until convergence: more precisely, in the presence of unobservables, the estimation step constructs estimates of the sufficient statistics of the model conditional on the observables, unobservables being estimated based on the parameter values at the current step of the iteration whereas at the maximization step, the likelihood function is maximized based both on the observables and these estimated values. This EM method is presented under special linear restrictions on \(\alpha\),\(\beta\),\(\phi\),\(\gamma\). The max. step uses SUR calculations (regression), so neither first nor second derivatives need be computed. In the neighbourhood of the EM solution, the scoring method may be useful to pinpoint the maximum and to estimate the information matrix. An economic application is given (estimation of a common factor in wage rate data from several industries in Los Angeles).