nag_correg_linregm_coeffs_const (g02cg) performs a multiple linear regression on a set of variables whose means, sums of squares and cross-products of deviations from means, and Pearson product-moment correlation coefficients are given.

Syntax

[result, coef, con, rinv, c, ifail] = g02cg(n, xbar, ssp, r, 'k1', k1)

[result, coef, con, rinv, c, ifail] = nag_correg_linregm_coeffs_const(n, xbar, ssp, r, 'k1', k1)

Note: the interface to this routine has changed since earlier releases of the toolbox:

At Mark 23:

k was removed from the interface

Description

nag_correg_linregm_coeffs_const (g02cg) fits a curve of the form

y = a + b_{1} x_{1} + b_{2} x_{2} + \dots + b_{k} x_{k}

to the data points

\begin{matrix} (x_{11}, x_{21}, \dots, x_{k 1}, y_{1}) \\ (x_{12}, x_{22}, \dots, x_{k 2}, y_{2}) \\ ⋮ \\ (x_{1 n}, x_{2 n}, \dots, x_{k n}, y_{n}) \end{matrix}

such that

y_{i} = a + b_{1} x_{1 i} + b_{2} x_{2 i} + \dots + b_{k} x_{k i} + e_{i}, i = 1, 2, \dots, n .

The function calculates the regression coefficients,

b_{1}, b_{2}, \dots, b_{k}

, the regression constant,

a

, and various other statistical quantities by minimizing

\sum_{i = 1}^{n} e_{i}^{2} .

The actual data values

(x_{1 i}, x_{2 i}, \dots, x_{k i}, y_{i})

are not provided as input to the function. Instead, input consists of:

(i)	The number of cases, $n$ , on which the regression is based.
(ii)	The total number of variables, dependent and independent, in the regression, $(k + 1)$ .
(iii)	The number of independent variables in the regression, $k$ .
(iv)	The means of all $k + 1$ variables in the regression, both the independent variables $(x_{1}, x_{2}, \dots, x_{k})$ and the dependent variable $(y)$ , which is the $(k + 1)$ th variable: i.e., ${\bar{x}}_{1}, {\bar{x}}_{2}, \dots, {\bar{x}}_{k}, \bar{y}$ .
(v)	The $(k + 1)$ by $(k + 1)$ matrix [ $S_{i j}$ ] of sums of squares and cross-products of deviations from means of all the variables in the regression; the terms involving the dependent variable, $y$ , appear in the $(k + 1)$ th row and column.
(vi)	The $(k + 1)$ by $(k + 1)$ matrix [ $R_{i j}$ ] of the Pearson product-moment correlation coefficients for all the variables in the regression; the correlations involving the dependent variable, $y$ , appear in the $(k + 1)$ th row and column.

The quantities calculated are:

(a)

The inverse of the

k

k

partition of the matrix of correlation coefficients, [

R_{i j}

], involving only the independent variables. The inverse is obtained using an accurate method which assumes that this sub-matrix is positive definite.

(b)

The modified inverse matrix,

C = [c_{i j}]

, where

c_{i j} = \frac{R_{i j} r_{i j}}{S_{i j}}, i, j = 1, 2, \dots, k,

where

r_{i j}

is the

(i, j)

th element of the inverse matrix of [

R_{i j}

] as described in (a) above. Each element of

C

is thus the corresponding element of the matrix of correlation coefficients multiplied by the corresponding element of the inverse of this matrix, divided by the corresponding element of the matrix of sums of squares and cross-products of deviations from means.

(c)

The regression coefficients:

b_{i} = \sum_{j = i}^{k} c_{i j} S_{j (k + 1)}, i = 1, 2, \dots, k,

where

S_{j (k + 1)}

is the sum of cross-products of deviations from means for the independent variable

x_{j}

and the dependent variable

y

(d)

The sum of squares attributable to the regression,

S S R

, the sum of squares of deviations about the regression,

S S D

, and the total sum of squares,

S S T

$S S T = S_{(k + 1) (k + 1)}$ , the sum of squares of deviations from the mean for the dependent variable, $y$ ;
$S S R = \sum_{j = 1}^{k} b_{j} S_{j (k + 1)}; S S D = S S T - S S R$

(e)

The degrees of freedom attributable to the regression,

D F R

, the degrees of freedom of deviations about the regression,

D F D

, and the total degrees of freedom,

D F T

D F R = k; D F D = n - k - 1; D F T = n - 1 .

(f)

The mean square attributable to the regression,

M S R

, and the mean square of deviations about the regression,

M S D

M S R = S S R / D F R; M S D = S S D / D F D .

(g)

The

F

values for the analysis of variance:

F = M S R / M S D .

(h)

The standard error estimate:

s = \sqrt{M S D} .

(i)

The coefficient of multiple correlation,

R

, the coefficient of multiple determination,

R^{2}

and the coefficient of multiple determination corrected for the degrees of freedom,

{\bar{R}}^{2}

;

R = \sqrt{1 - \frac{S S D}{S S T}}; R^{2} = 1 - \frac{S S D}{S S T}; {\bar{R}}^{2} = 1 - \frac{S S D \times D F T}{S S T \times D F D} .

(j)

The standard error of the regression coefficients:

s e (b_{i}) = \sqrt{M S D \times c_{i i}}, i = 1, 2, \dots, k .

(k)

The

t

values for the regression coefficients:

t (b_{i}) = \frac{b_{i}}{s e (b_{i})}, i = 1, 2, \dots, k .

(l)

The regression constant,

a

, its standard error,

s e (a)

, and its

t

value,

t (a)

a = \bar{y} - \sum_{i = 1}^{k} b_{i} {\bar{x}}_{i}; s e (a) = \sqrt{M S D \times (\frac{1}{n} + \sum_{i = 1}^{k} \sum_{j = 1}^{k} {\bar{x}}_{i} c_{i j} {\bar{x}}_{j})}; t (a) = \frac{a}{s e (a)} .

References

Draper N R and Smith H (1985) Applied Regression Analysis (2nd Edition) Wiley

Parameters

Compulsory Input Parameters

1: $n$ – int64int32nag_int scalar: The number of cases $n$ , used in calculating the sums of squares and cross-products and correlation coefficients.
2: $xbar (k1)$ – double array: $xbar (i)$ must be set to ${\bar{x}}_{i}$ , the mean value of the $i$ th variable, for $i = 1, 2, \dots, k + 1$ ; the mean of the dependent variable must be contained in $xbar (k + 1)$ .
3: $ssp (ldssp, k1)$ – double array: ldssp, the first dimension of the array, must satisfy the constraint $ldssp \geq k1$ .
$ssp (i, j)$ must be set to $S_{i j}$ , the sum of cross-products of deviations from means for the $i$ th and $j$ th variables, for $i = 1, 2, \dots, k + 1$ and $j = 1, 2, \dots, k + 1$ ; terms involving the dependent variable appear in row $k + 1$ and column $k + 1$ .
4: $r (ldr, k1)$ – double array: ldr, the first dimension of the array, must satisfy the constraint $ldr \geq k1$ .
$r (i, j)$ must be set to $R_{i j}$ , the Pearson product-moment correlation coefficient for the $i$ th and $j$ th variables, for $i = 1, 2, \dots, k + 1$ and $j = 1, 2, \dots, k + 1$ ; terms involving the dependent variable appear in row $k + 1$ and column $k + 1$ .

Optional Input Parameters

1: $k1$ – int64int32nag_int scalar: Default: the dimension of the array xbar and the first dimension of the arrays ssp, r and the second dimension of the arrays ssp, r. (An error is raised if these dimensions are not equal.)
The total number of variables, independent and dependent, $(k + 1)$ , in the regression.

Constraint: $2 \leq k1 < n$ .

Output Parameters

1: $result (13)$ – double array

The following information:

$result (1)$	$S S R$ , the sum of squares attributable to the regression;
$result (2)$	$D F R$ , the degrees of freedom attributable to the regression;
$result (3)$	$M S R$ , the mean square attributable to the regression;
$result (4)$	$F$ , the $F$ value for the analysis of variance;
$result (5)$	$S S D$ , the sum of squares of deviations about the regression;
$result (6)$	$D F D$ , the degrees of freedom of deviations about the regression;
$result (7)$	$M S D$ , the mean square of deviations about the regression;
$result (8)$	$S S T$ , the total sum of squares;
$result (9)$	$D F T$ , the total degrees of freedom;
$result (10)$	$s$ , the standard error estimate;
$result (11)$	$R$ , the coefficient of multiple correlation;
$result (12)$	$R^{2}$ , the coefficient of multiple determination;
$result (13)$	${\bar{R}}^{2}$ , the coefficient of multiple determination corrected for the degrees of freedom.

2: $coef (ldcoef, 3)$ – double array

For

i = 1, 2, \dots, k

, the following information:

$coef (i, 1)$: $b_{i}$ , the regression coefficient for the $i$ th variable.
$coef (i, 2)$: $s e (b_{i})$ , the standard error of the regression coefficient for the $i$ th variable.
$coef (i, 3)$: $t (b_{i})$ , the $t$ value of the regression coefficient for the $i$ th variable.

3: $con (3)$ – double array

The following information:

$con (1)$	$a$ , the regression constant;
$con (2)$	$s e (a)$ , the standard error of the regression constant;
$con (3)$	$t (a)$ , the $t$ value for the regression constant.

4: $rinv (ldrinv, k)$ – double array

k = k1 - 1

The inverse of the matrix of correlation coefficients for the independent variables; that is, the inverse of the matrix consisting of the first

k

rows and columns of r.

5: $c (ldc, k)$ – double array

k = k1 - 1

The modified inverse matrix, where

$c (i, j) = r (i, j) \times rinv (i, j) / ssp (i, j)$ , for $i = 1, 2, \dots, k$ and $j = 1, 2, \dots, k$ .

6: $ifail$ – int64int32nag_int scalar

ifail = 0

unless the function detects an error (see Error Indicators and Warnings).

Error Indicators and Warnings

Errors or warnings detected by the function:

$ifail = 1$

On entry,

k1 < 2

$ifail = 2$

On entry,

k1 \neq (k + 1)

$ifail = 3$

On entry,

n \leq k1

$ifail = 4$

On entry,	$ldssp < k1$ ,
or	$ldr < k1$ ,
or	$ldcoef < k$ ,
or	$ldrinv < k$ ,
or	$ldc < k$ ,
or	$ldwkz < k$ .

$ifail = 5$: The $k$ by $k$ partition of the matrix $R$ which is to be inverted is not positive definite.

$ifail = 6$: The refinement following the actual inversion fails, indicating that the $k$ by $k$ partition of the matrix $R$ , which is to be inverted, is ill-conditioned. The use of nag_correg_linregm_fit (g02da), which employs a different numerical technique, may avoid this difficulty (an extra ‘variable’ representing the constant term must be introduced for nag_correg_linregm_fit (g02da)).

$ifail = 7$: Unexpected error in nag_linsys_real_posdef_solve_ref (f04ab).

$ifail = - 99$: An unexpected error has been triggered by this routine. Please contact NAG.

$ifail = - 399$: Your licence key may have expired or may not have been installed correctly.

$ifail = - 999$: Dynamic memory allocation failed.

Accuracy

The accuracy of any regression function is almost entirely dependent on the accuracy of the matrix inversion method used. In nag_correg_linregm_coeffs_const (g02cg), it is the matrix of correlation coefficients rather than that of the sums of squares and cross-products of deviations from means that is inverted; this means that all terms in the matrix for inversion are of a similar order, and reduces the scope for computational error. For details on absolute accuracy, the relevant section of the document describing the inversion function used, nag_linsys_real_posdef_solve_ref (f04ab), should be consulted. nag_correg_linregm_fit (g02da) uses a different method, based on nag_linsys_real_gen_lsqsol (f04am), and that function may well prove more reliable numerically. It does not handle missing values, nor does it provide the same output as this function. (In particular it is necessary to include explicitly the constant in the regression equation as another ‘variable’.)

If, in calculating

F

t (a)

, or any of the

t (b_{i})

(see Description), the numbers involved are such that the result would be outside the range of numbers which can be stored by the machine, then the answer is set to the largest quantity which can be stored as a double variable, by means of a call to nag_machine_real_largest (x02al).

Further Comments

The time taken by nag_correg_linregm_coeffs_const (g02cg) depends on

k

This function assumes that the matrix of correlation coefficients for the independent variables in the regression is positive definite; it fails if this is not the case.

This correlation matrix will in fact be positive definite whenever the correlation matrix and the sums of squares and cross-products (of deviations from means) matrix have been formed either without regard to missing values, or by eliminating completely any cases involving missing values, for any variable. If, however, these matrices are formed by eliminating cases with missing values from only those calculations involving the variables for which the values are missing, no such statement can be made, and the correlation matrix may or may not be positive definite. You should be aware of the possible dangers of using correlation matrices formed in this way (see the G02 Chapter Introduction), but if they nevertheless wish to carry out regression using such matrices, this function is capable of handling the inversion of such matrices provided they are positive definite.

If a matrix is positive definite, its subsequent re-organisation by either nag_correg_linregm_service_select (g02ce) or nag_correg_linregm_service_reorder (g02cf) will not affect this property, and the new matrix can safely be used in this function. Thus correlation matrices produced by any of nag_correg_coeffs_pearson (g02ba), nag_correg_coeffs_pearson_miss_case (g02bb), nag_correg_coeffs_pearson_subset (g02bg) or nag_correg_coeffs_pearson_subset_miss_case (g02bh), even if subsequently modified by either nag_correg_linregm_service_select (g02ce) or nag_correg_linregm_service_reorder (g02cf), can be handled by this function.

It should be noted that in forming the sums of squares and cross-products matrix and the correlation matrix a column of constants should not be added to the data as an additional ‘variable’ in order to obtain a constant term in the regression. This function automatically calculates the regression constant,

a

, and any attempt to insert such a ‘dummy variable’ is likely to cause the function to fail.

It should also be noted that the function requires the dependent variable to be the last of the

k + 1

variables whose statistics are provided as input to the function. If this variable is not correctly positioned in the original data, the means, standard deviations, sums of squares and cross-products of deviations from means, and correlation coefficients can be manipulated by using nag_correg_linregm_service_select (g02ce) or nag_correg_linregm_service_reorder (g02cf) to reorder the variables as necessary.

Example

This example reads in the means, sums of squares and cross-products of deviations from means, and correlation coefficients for three variables. A multiple linear regression is then performed with the third and final variable as the dependent variable. Finally the results are printed.

Open in the MATLAB editor: g02cg_example

function g02cg_example


fprintf('g02cg example results\n\n');

% Data
n    = int64(5);
k    = 2;
xbar = [  5.4;     5.8;      2.8];
ssp  = [ 99.2,   -57.6,      6.4;
        -57.6,   102.8,    -29.2;
          6.4,   -29.2,     14.8   ];
r    = [  1,     -0.5704,    0.167;
         -0.5704, 1,        -0.7486;
          0.167, -0.7486,    1     ];
% Reordering
korder = [int64(1); 3; 2];

% Display data
disp('Means:');
disp(xbar);
disp('Sums of squares and cross-products about means:');
disp(ssp);
disp('Correlation coefficients:');
disp(r);

% Fit multiple linear regression model
[result, coeff, con, rinv, c, ifail] = ...
  g02cg(n, xbar, ssp, r);

% Display results
fprintf('  Variable    Coef      Std err  t-value\n');
disp([[1:k]' coeff]);
  
fprintf('Analysis of regression table :-\n\n');
fprintf('      Source        Sum of squares  DF   Mean square   F-value\n');
fprintf('Due to regression %11.3f%8d%13.3f%12.3f\n', result(1:4));
fprintf('About  regression %11.3f%8d%13.3f\n', result(5:7));
fprintf('Total             %11.3f%8d\n\n', result(8:9));
fprintf('Standard error of estimate = %8.4f\n', result(10));
fprintf('Multiple correlation (R)   = %8.4f\n', result(11));
fprintf('Determination (R squared)  = %8.4f\n', result(12));
fprintf('Corrected R squared        = %8.4f\n\n', result(13));

disp('Inverse of correlation matrix of independent variables:');
disp(rinv);
disp('Modified inverse matrix:');
disp(c);

g02cg example results

Means:
    5.4000
    5.8000
    2.8000

Sums of squares and cross-products about means:
   99.2000  -57.6000    6.4000
  -57.6000  102.8000  -29.2000
    6.4000  -29.2000   14.8000

Correlation coefficients:
    1.0000   -0.5704    0.1670
   -0.5704    1.0000   -0.7486
    0.1670   -0.7486    1.0000

  Variable    Coef      Std err  t-value
    1.0000   -0.1488    0.1937   -0.7683
    2.0000   -0.3674    0.1903   -1.9309

Analysis of regression table :-

      Source        Sum of squares  DF   Mean square   F-value
Due to regression       9.777       2        4.888       1.946
About  regression       5.023       2        2.512
Total                  14.800       4

Standard error of estimate =   1.5848
Multiple correlation (R)   =   0.8128
Determination (R squared)  =   0.6606
Corrected R squared        =   0.3212

Inverse of correlation matrix of independent variables:
    1.4823    0.8455
    0.8455    1.4823

Modified inverse matrix:
    0.0149    0.0084
    0.0084    0.0144

PDF version (NAG web site, 64-bit version, 64-bit version)

Chapter Contents

Chapter Introduction

NAG Toolbox