Abstract:
To compare model selection criteria for linear regression with small sample size. The four model selection criteria in this comparison are composed of Corrected Akaike's Information Criterion (AIC[subscript c]), Corrected Schwarz's Information Criterion (SIC[subscript c]), Corrected Hannan and Quinn's Information Criterion (HQ[subscript c]) and Corrected Kullback's Information Criterion (KIC[subscript c]). The criterion of comparison is the Average of Mean Squares Error (AMSE) using Ratio of Different Average Mean Squares Error (RDAMSE) to compare the efficiency of these model selection criteria. The distribution of random errors are normal distribution with zero mean and variances are 1, 2, 4, 8 and 16, respectively. The sample sizes used in this study are 12, 15, 18, 21, 24, 27 and 30, respectively. The number of independent variables in regression model are 3, 5 and 6, respectively. The data for this experiment are generated through the Monte Carlo simulation technique, repeating 1,000 times for each case. The results of this research are as follow : The factors that affected AMSE of all model selection criteria are the number of independent variables, the variance of random errors and sample size, respectively. The AMSE vary with, most to least, respectively, the number of independent variables and the variance of random errors but AMSE is converse to sample size. These factors affected model selection criteria as follow. 1. In case of sample size is 12 to 15. For all the number of independent variables and all variance of random errors, the HQ[subscript c] is the best, the AIC[subscript c], the SIC[subscript c] and the KIC[subscript c], respectively. 2. In case of sample size is 18ํ For all the number of independent variables and all variance of random errors, the HQ[subscript c] and the AIC[subscript c] are the best, the KIC[subscript c] and the SIC[subscript c], respectively. The KIC[subscript c] give AMSE slightly lower than the SIC[subscript c]. 3. In case of sample size is 21 to 30ํ For all the number of independent variables and all variance of random errors, the AIC[subscript c] is the best, the HQ[subscript c], the SIC[subscript c] and the KIC[subscript c], respectively. Furthermore, the four model selection criteria efficiency is separated into two groups by RDAMSE. The first group is composed of the HQ[subscript c] and the AIC[subscript c]. Their RDAMSE are close and they are the best for all cases. The second group is composed of the SIC[subscript c] and the KIC[subscript c]. Their RDAMSE are close but their RDAMSE are clearly different from the first group for all cases.