Linear regression

of one random variable $ \mathbf Y = ( Y ^ {(} 1) \dots Y ^ {(} m) ) ^ \prime $ on another $ \mathbf X = ( X ^ {(} 1) \dots X ^ {(} p) ) ^ \prime $

An $ m $- dimensional vector form, linear in $ \mathbf x $, supposed to be the conditional mean (given $ \mathbf X = \mathbf x $) of the random vector $ \mathbf Y $. The corresponding equations

$$ \tag{* } y ^ {(} k) ( \mathbf x , \mathbf b ) = {\mathsf E} ( Y ^ {(} k) \mid \mathbf X = \mathbf x ) = \ \sum _ { j= } 0 ^ { p } b _ {kj} x ^ {(} j) , $$

$$ x ^ {(} 0) \equiv 1 ,\ k = 1 \dots m, $$

are called the linear regression equations of $ \mathbf Y $ on $ \mathbf X $, and the parameters $ b _ {kj} $ are called the regression coefficients (see also Regression), $ \mathbf X $ is an observable parameter (not necessarily random), on which the mean of the resulting function (response) $ \mathbf Y ( \mathbf X ) $ under investigation depends.

In addition, the linear regression of $ Y ^ {(} k) $ on $ \mathbf X $ is frequently also understood to be the "best" (in a well-defined sense) linear approximation of $ Y ^ {(} k) $ by means of $ \mathbf X $, or even the result of the best (in a well-defined sense) smoothing of a system of experimental points ( "observations" ) $ ( Y _ {i} ^ {(} k) , \mathbf X _ {i} ) $, $ i = 1 \dots n $, by means of a hyperplane in the space $ ( Y ^ {(} k) , \mathbf X ) $, in situations when the interpretation of these points as samples from a corresponding general population need not be allowable. With such a definition one has to distinguish different versions of linear regression, depending on the choice of the method of computing the errors of the linear approximation of $ Y ^ {(} k) $ by means of $ \mathbf X $( or depending on the actual choice of a criterion for the amount of smoothing). The most widespread criteria for the quality of the approximation of $ Y ^ {(} k) $ by means of linear combinations of $ \mathbf X $( linear smoothing of the points $ ( Y _ {i} ^ {(} k) , \mathbf X _ {i} ) $) are:

$$ Q _ {1} ( \mathbf b ) = {\mathsf E} \left \{ \omega ^ {2} ( \mathbf X ) \cdot \left ( Y ^ {(} k) ( \mathbf X ) - \sum _ { j= } 0 ^ { p } b _ {kj} X ^ {(} j) \right ) ^ {2} \right \} , $$

$$ \widetilde{Q} _ {1} ( \mathbf b ) = \sum _ { i= } 1 ^ { n } \omega _ {i} ^ {2} \left ( Y _ {i} ^ {(} k) - \sum _ { j= } 0 ^ { p } b _ {kj} X _ {i} ^ {(} j) \right ) ^ {2} , $$

$$ Q _ {2} ( \mathbf b ) = {\mathsf E} \left \{ \omega ( \mathbf X ) \left | Y ^ {(} k) ( \mathbf X ) - \sum _ { j= } 0 ^ { p } b _ {kj} X ^ {(} j) \right | \right \} , $$

$$ \widetilde{Q} _ {2} ( \mathbf b ) = \sum _ { j= } 1 ^ { n } \omega _ {i} \left | Y _ {i} ^ {(} k) - \sum _ { j= } 0 ^ { p } b _ {kj} X _ {i} ^ {(} j) \right | , $$

$$ Q _ {3} ( \mathbf b ) = {\mathsf E} \left \{ \omega ^ {2} ( \mathbf X ) \cdot \rho ^ {2} \left ( Y ^ {(} k) ( \mathbf X ) , \sum _ { j= } 0 ^ { p } b _ {kj} X ^ {(} j) \right ) \right \} , $$

$$ \widetilde{Q} _ {3} ( \mathbf b ) = \sum _ { i= } 1 ^ { n } \omega _ {i} ^ {2} \cdot \rho ^ {2} \left ( Y _ {i} ^ {(} k) , \sum _ { j= } 0 ^ { p } b _ {kj} X _ {i} ^ {(} j) \right ) . $$

In these relations the choice of "weights" $ \omega ( \mathbf X ) $ or $ \omega _ {i} $ depends on the nature of the actual scheme under investigation. For example, if the $ Y ^ {(} k) ( \mathbf X ) $ are interpreted as random variables with known variances $ {\mathsf D} Y ^ {(} k) ( \mathbf X ) $( or with known estimates of them), then $ \omega ^ {2} ( \mathbf X ) = [ {\mathsf D} Y ^ {(} k) ( \mathbf X ) ] ^ {-} 1 $. In the last two criteria the "discrepancies" of the approximation or the smoothing are measured by the distances $ \rho ( \cdot , \cdot ) $ from $ Y ^ {(} k) ( \mathbf X ) $ or $ Y _ {i} ^ {(} k) $ to the required hyperplane of regression. If the coefficients $ b _ {kj} $ are determined by minimizing the quantities $ Q _ {1} ( \mathbf b ) $ or $ \widetilde{Q} _ {1} ( \mathbf b ) $, then the linear regression is said to be least squares or $ L _ {2} $; if the criteria $ Q _ {2} ( \mathbf b ) $ and $ \widetilde{Q} _ {2} ( \mathbf b ) $ are used, the linear regression is said to be minimal absolute deviations or $ L _ {1} $; if the criteria $ Q _ {3} ( \mathbf b ) $ and $ \widetilde{Q} _ {3} ( \mathbf b ) $ are used, it is said to be minimum $ \rho $- distance.

In certain cases, linear regression in the classical sense (*) is the same as linear regression defined by using functionals of the type $ Q _ {i} $. Thus, if the vector $ ( \mathbf X ^ \prime , Y ^ {(} k) ) $ is subject to a multi-dimensional normal law, then the regression of $ Y ^ {(} k) $ on $ \mathbf X $ in the sense of (*) is linear and is the same as least squares or minimum mean squares linear regression (for $ \omega ( \mathbf X ) \equiv 1 $).

References

[1]	Yu.V. Linnik, "Methode der kleinste Quadraten in moderner Darstellung" , Deutsch. Verlag Wissenschaft. (1961) (Translated from Russian)
[2]	H. Cramér, "Mathematical methods of statistics" , Princeton Univ. Press (1946)
[3]	M.G. Kendall, A. Stuart, "The advanced theory of statistics" , 2. Inference and relationship , Macmillan (1979)
[4]	C.R. Rao, "Linear statistical inference and its applications" , Wiley (1965)