Alternative Estimators of FIML Covariance Matrix: A Monte Carlo Study

Self-selection and direct estimation of across-regime correlation parameter

“…3 A relevant exception is represented by the ML approach to estimate simultaneously demand and supply equations in a disequilibrium model, where the observations belong, respectively, to a demand or to a supply function (see, among others, Maddala and Nelson, 1974). The choice of the regime does not depend on a third selection equation, but the likelihood function specified in this model, including coefficients and variances, does not include the across-regime covariance.…”

Systems estimation: A comparison of SAS, SHAZAM and TSP

“…(1974) parameter standard errors may be more accurate. See Calzolari and Panattoni (1988). There are some differences of 0.01 between the TSP statistics and the SAS SYSLIN/ VARDEF = N and SHAZAM statistics for some unit 1979 and 1982 samples and Sample A.…”

“…Berndt discusses both cases. For the Klein Model I structural form, the estimates of Calzolari and Panattoni (1988) are an independent standard (to which only the TSP manual refers) and are used as a basis for comparison. For this model, the FIML option can be used with the SAS PROC SYSLIN example in the manual (SAS/ETS, pp.856-862).…”

“…(Starting SHAZAM at the 3SLS estimates makes the consumption parameters slightly better, but the other parameters worse.) Both the relatively simple TSP FIML structural estimation program and the very complex TSP LSQ program for RRF estimates and standard errors below were provided for Calzolari and Panattoni (1988) in Table V. Est.…”

Measuring Economic Convergence

“…3. Calzolari and Panattoni (1988) have examined the relative magnitudes of the estimated covariance matrix obtained from the inverse Hessian with the outer product of the first derivative of the loglikelihood (and also by a GLS type procedure). They found that the outer product calculations generate standard errors double or triple the Hessian estimates, and they, in turn, are considerably larger than those computed with a GLS type matrix.…”