THESIS TOPIC:
link to homepage
Model specification selection is always been of importance because of its multiple uses in social sciences. For example, in economics it can be used for policy implications, predictions etc. In other science like medical and environment science it is used to see the effect of different treatment on human being. This topic is always been under discussion and many books are written on it but it is still open for debate. There are numerous procedures to select the important variables/factors from the set of variables. Due to the improvements in computational ways i.e. computers most of them are available in different software. These automated procedures are time saving are easy to access for common researcher. This thesis discusses many of such automated specification search methods that are used from decades.
The most common procedures that are used for the model selection are Information based criteria like Akiake (1973) , Hannan (1979) ,Bayesian (1978) and their different version, different types of path reduction procedures like Autometrics (2009) (latest version of General to specific approach, various stepwise procedures i.e. forward, backward (1960). These methods are available almost in all commonly used software packages for social sciences e.g. SPSS, STATA, PCGIVE etc.
This study reviews the existing research, conducts an extensive simulation based comparative analysis of various Model selection strategies for different types of data and situations by considering the performance of these model selection methods. Their performance is judged on the basis of selection of true model, potency and gauge i.e. (analogous to power and size in statistical hypothesis testing) in different circumstances of sample size, parameterization.
In simulation analysis different types of data and models are used. Firstly, univariate static model is taken for the comparison to match up our results with the previous studies as it is found often in literature. The main target of this study is to analyze the performance of model selection procedures for the panel data as it is getting more attention of the researchers nowadays. For this purpose Autometrics and stepwise procedures are developed for panel data and then the performance of this procedure is evaluated on the basis how much they get the true model, significant variables and discard in significant variables.
The results of the simulation for the univariate static model are found to be consistent with the literature. A Bayesian information criterion (BIC) selects parsimonious models and is consistent. Akiake information criteria (AIC) perform well when there are more relevant variable than irrelevant but with high rate of selecting larger models. Autometrics did well when smaller values of parameter are used and is consistent. In the circumstances of panel data, as like univariate, no conclusive results can be inferred. Different procedures did well in one situation but have low or worst performance in other condition. However, overall Autometrics did well as it competes with BIC in consistency, and with AIC in small samples and smaller parameter values. Overall, it shows good potency and gauge i.e. around specified level, especially in random coefficient models. After Autometrics, stepwise procedures did well and then the information based procedures. At the end I reconsidered the factors explaining the investment for developing countries discussed in different theories and empirical research. A unique model is found, through Autometrics approach using random coefficient model, for all the countries in the sample through which policy decision may be improved.
Progress Report
AZ: Tasks
SUBGOALS: