Abstract:
The objective of this research is to compare the estimation of parameters (regression coefficients) in the multiple linear regression with small sample size. In this study, the estimation methods are the Ordinary Least Square method (OLS), Bootstrap method (BS), M-estimator method using Ramsays robust criteria (M/R) and M-estimator method using Tukeys robust criteria (M/T). They were compared by average of mean square error (AMSE). The comparison was done under several situations which are follows: a)The number of independent variables are 3 and 6 b)The sample sizes are 10,15,20,25 and 30 c) This study used three residual distributions which are normal distribution, scale-contaminated normal distribution and gamma distribution. For normal distribution, the means of 0 and 1, the standard deviations of 1, 2, and 3 are used. For scale-contaminated normal distribution, the scale factor of 10 the percent of contaminations of 25%, 50%, and 75% are used. For gamma distribution, the beta parameter of 1, the alpha parameter of 0.5,1,2 and 4 are used. The data for this research is simulated by using the Monte-Carlo simulation technique with 1,000 repetitions for each situation by S-PLUS 2000 program. The results of this research are as follows: In case that residuals have normal distribution and scale-contaminated normal distribution the AMSE of BS method is lower than the others when the sample sizes are small (lower than 20) but when the sample size and the number of independent variables are increased the AMSE of M/T method is lower than the others. But in case that residuals have gamma distribution, the AMSE of M/R method is lower than the others every levels of the sample size and the number of independent variables. In general, the AMSE varies with, most to least respectively, the variance of the residuals, the number of independent variables and the percent of contaminations of outliers in the residuals but coverse to sample size.