Extended multiple models selection algorithms based on iterative feasible generalized least squares (IFGLS) and expectation-maximization (EM) algorithm

Automated model selection has been used to bridge the gap between experts and end users since 1960s starting with Stepwise and recently with Autometrics for single equation. This extension of Autometrics for model selection was also developed for multiple equations by integrating it with seemingly unrelated regressions equations (SURE) and estimated using feasible generalized least squares (FGLS), known as SURE-Autometrics algorithm. However, SURE-Autometrics has not been estimated using maximum likelihood estimation (MLE). Therefore, in this study SUREAutometrics is improvised using two MLE methods, which are iterative feasible generalized least squares (IFGLS) and expectation-maximization (EM) algorithm, named as SURE(IFGLS)-Autometrics and SURE(EM)-Autometrics algorithms. Simulation and empirical studies are conducted in validating the performance of the two algorithms. In the simulation study, different sample sizes, strength of correlation among equations, size of general unrestricted model (GUMS), number of equations, significance levels and true specification models are incorporated by evaluating the percentages of finding the true GUMS. While in the empirical study, two empirical data sets which are national growth rates and water quality index (WQI) are assessed using root mean square error and geometric root mean square error where 18 models selection procedures of manual and automated approaches are compared. The simulation results indicated that performance of SURE(IFGLS)-Autometrics and SURE(EM)-Autometrics algorithms improved in conditions of large sample, strong correlation among equations, small GUMS, a smaller number of equations, tight significance level and in an empty model (without predictor variables). The empirical results for both algorithms performed well as compared to other models selection procedures, particularly using WQI data where the sample size is bigger and has good quality data. In conclusion, SURE(IFGLS)-Autometrics and SURE(EM)-Autometrics can be used as models selection algorithms. Additionally, both algorithms are suitable in improving performance of automated models selection procedures. General findings support the idea that automated procedures surpass the manual procedures.