This paper discusses the problem of decision support systems in the organization. The procedure (linear combination) developed with the aim to combine some predicted results obtained with simulation of linear and nonlinear regression models (experts), multiple regression model, nonparametric regression model, and semi parametric regression model. This adjustment procedure enforce some statistical characteristics like the expected value of the gross production rate based on Cobb-Douglas production function is unbiased for the actual value, and the total weights(importance) of all models(experts) is equal to one. We used modeling and simulation techniques to generate our data and to apply the procedure.
In economics, the relationship between the production (outputs) and the production elements (inputs) known as production function. Usually decision makers in the organization and companies used some statistical and economics methods to support the production policy. Cobb-Douglas (CD) production function is one of the most useful tools to support the production policy, [1] and [9] used the CD function and found that public investment has a large contribution to production, [7] argued that public-sector capital has no effect on production after controlling for location characteristic, [2] and [5] pursued the insignificance of public capital. All of the literature above assumes a Cobb-Douglas (CD) production function to estimate productivity of inputs. One property of the CD functional form is that elasticity of an input is the estimated slope coefficient when the output and inputs are measured in logs, [6] used nonparametric perspective, which allows elasticities to vary across location and time, [15] use the CD functions proposed by [9]. [2] and [6] considers the gross product(GP) as responding variable, public capital (PUC), private capital (PRC), the employment rate (EMP), and the unemployment rate (UEPR) as explanatory variables, to predict of the gross product by using linear and nonlinear regression models. [13] they have propsed linear pooling method to combine the probability forecasts and proved the propsed method gave them more accurate results.
In this paper we used the same variables and methodology to predict gross production, but we use linear combination method to combine the results of the three experts (multiple linear regression, non-linear regression, and semi-parametric) as input variables, we choose these three types of models because we think there is linear relationship, non-linear relationship and mixture relationship between gross production and the predictor variables. It is based on our hypothesis that the linear combination method will give us more accurate predictive gross production since we believe this is a better way to take into account the results of multiple models accompanied by an appropriate weighting scheme. Furthermore, we prove that this method gives an unbiased estimator to the gross production. Our contribution is that we use linear combination method to pool three regression models for predicting the gross production. Linear combination is linear pooling of some sets of ordered pairs (importance, gross production), we have explored the prediction associated with different regression model can be combined into one final model via weighted linear combination will give more accurate results. Mathematically linear combination of the sequence 1 2
, ,..., n y y y each with mean is:
, where i a is the weight of i y ,and
Regression analysis is a technique used in data analysis; we use regression technique to predict the value of the response (dependent) variable given any value of the predictor (independent) variable. A general regression model is [15]: ( )
Where i=1, 2... n denoting an observation of a subject. y i is the response variable and
x i is a k 1 vector of independent variables. E (y i │x i ) is the expectation of y i conditional on x i , and e i is the error term. In this paper we will use the following types of regression model:
In this model, it is assumed that y i is linearly related with x i , so we can say it is linear regression model:
( )
Thus a linear regression model is written as:
Where α is the intercept and is k 1 vector of parameters. Under Gauss-Markov assumptions, the estimators of α and are Best Linear Unbiased Estimators (BLUE), and can be estimate by using Ordinary Least Squares method (OLS).
(3)
If we do not know the data generating process, it is very unlikely that a linear regression (parametric) model is exactly the appropriate model specification, so the estimators α and are not BLUE (best linear unbiased estimator). Instead of making assumptions for the functional form of E (y i │x i ), nonparametric regression methods do not require any presumptions about the underlying data generating process [10]. Let ( )
Where m (.) is some function of unspecified functional form. Some basic assumptions about m (.) are commonly made. Many methods have been devised to estimate the regression function m (.), but we will just consider a simple but effective estimate known as the Kernel regression estimate. Suppose we have a random sample (x 1 , y 1 ), (x 2 , y 2 ) …, (x n ,y n ). The Kernel regression estimate of m(x) =E (y│X=x) is given by:
Here, K is a nonnegative symmetric function that is not increasing as its argument gets away from zero, and d is a parameter called the smoothing parameter that is selected by the user to control the amount of smoothing. The estimator in equation ( 4) is called the Local-Constant Least-Squares, which can be interpreted as a weighted average of y i , where:
Where x js is the s th component of x j and d s the s th component of d. The Kernel function k (.) can take several forms. In this paper we used the Gaussian Kernel function is defined as:
The smoothing parameter d is generally the most important factor when performing nonparametric regression. The bandwidth is chosen to obtain a desirable (4) (5) (6) (7) trade-off between the bias and the variance of estimation, so we need a method that could balance the bias and variance of the resulting estimate.
We have used Leave-One-Out Cross-Validation to select the optimal bandwidth. This method is depending on the principle of selecting bandwidth that minimizes the sum squared error of the resulting estimates. We will try to find the minimum MSE, mathematically; we are trying to minimize the following function [15]:
Where ˆ( )
m x is the Leave-One-Out estimator of m (.) evaluated at x j . SAS software uses this method to find the optimal bandwidth values.
Nonparametric techniques are very flexible because they do not need any assumptions about functional form. However, there are cases in which the relationship between the dependent variable and some independent variables is known to be linear and the relation between the dependent variable and other independent variables remain undetermined. In this situation semi parametric models are developed to solve this problem. Semi parametric models have both parametric and nonparametric components [15].
Consider the semi parametric partially linear model: ( )
Where is a k 1 vector of parameters, z i is a 1 q vector of independent variables and e i is an additive error, is assumed to be uncorrelated with xi and z i . is the parametric part of the model and the unknown function m (.) is the nonparametric part. [8] [11] proposed an estimate of and m (z i ) as follows. Using ordinary least squares, the estimator of is:
Once we obtain ˆ, the nonparametric part m (z i ) is easy to estimate. From equation ( 8), we get: m
Then we can get the estimator of m (z i ) as follow:
Where d can be estimate similar to the nonparametric model? SAS software uses Cross-Validation to find the optimal bandwidth.
We will generate a random sample of size 30 observations for each explanatory variables: public capital (PUC), private capital (PRC), the employment rate (EMP), and the unemployment rate (UERP). The Cobb-Douglas (CD) production function is, ( , , , )
We can rewrite the CD function as,
The log-linear CD production function is >1, returns to scale are increasing.
We apply multi-linear regression model, covariance model, and variance component model to our generated data to find the estimated values by using SAS procedure proc reg [12] and the output are listed in Tables 1, follow: The results of Table 1 are based on the model in [12].
We used this model to find the estimated values of the intercept and elasticity values by using generated data and SAS procedure (proc gam). The output from proc gam is listed in following table, Table 2
We used this model to find the estimates values of Elasticity by using SAS procedure (proc gam) and we considered the independent variable (Public Capital) is linearly related to the gross production. The output from proc gam is listed as follows in Table 3.
We will use the estimated models (experts), multiple parametric regression, nonparametric regression, and Semiparametric regression to predict o f the log-gross production, for sample of size equal to 10 observations, and by using the values of the columns(Parametric 1 From Table 5 we can find that the correlation coefficient between the predicted values of the linear combination(Y) and the actual variable of gross production is (0.954) that means the linear combination method is more associated with the actual values than the expert1, expert2, and expert3.this strongest association supports the suggestion that the predicted value of the gross production which associated with the different experts that combined is better than taking the predictive values of each expert individually. The average difference is (0.261, 0.959, 0.146, 0.103) of the three regression models (parametric, non-parametric, and semi-parametric) respectively and linear combination procedure. The proposed method (linear combination) is more accurate than regression models because the difference (actual and Y) in table 4 is more less than the other models, so this means that the proposed method give us unbiased estimator.
Table 5: Correlation Matrix
In this paper, table 5(correlation matrix) shown that the correlation coefficients between the linear combination method, parametric model, nonparametric model and semi-parametric and the actual values of the gross production are (0.954, 0.457, 0.935, and -0.015), because we believe the public capital (PUC) is linearly related to gross production, so we choose it as parametric variable and that is why the
correlation between y 3 and actual value is -0.015. Linear combination method is more associated with the actual value than the others experts (regression model), i.e. linear combination method reflects the experts views to find the most fitted predicted value to the actual value. Finally as forecasters often wish to provide an accurate predicted value, so we will use the linear combination method to predict of the gross production.