how to calculate singular values in collin test to detect multicollinearity

I am a bit confused as to how we find the singular values and therefore condition index number to detect multicollinearity in multiple linear regression analysis. Some mathematicians say the singular values are the square roots of the eigenvalues of the correlation matrix of the predictors of a model. While others says we use covariance matrix instead. Again some math publications said The singular values are the square roots of the eigenvalues of the square matrix X'X of multiple linear regression model. Then I treid using all three methods but when I cross checked with MATLAB results using collintest it does not match with either of my calculations. it does not explain how we go the output. can someone explain it to me?

[Umar]

Hi Nafisa,

That is a very good question that you ask. In order to calculate singular values and the condition index for multicollinearity detection in multiple linear regression, you can use the svd function to compute the singular value decomposition of the design matrix X. The singular values are the square roots of the eigenvalues of X'X.

Here's a simple example:

X = [1 2 3; 4 5 6; 7 8 9]; % Example design matrix

[U, S, V] = svd(X); % Singular value decomposition

singular_values = diag(S); % Extract singular values

condition_index = max(singular_values) / min(singular_values); % Calculate condition index

By computing the singular values and condition index using the above method, you can effectively detect multicollinearity in your regression model. This approach aligns with the mathematical principles behind singular value decomposition and provides accurate results for assessing multicollinearity.

[Umar]

Furthermore, regarding the methods for finding singular values, there are variations in mathematical formulations. Some sources suggest calculating the singular values as the square roots of the eigenvalues of the correlation matrix of predictors. This approach examines the relationships between predictors by analyzing their correlations. On the other hand, using the covariance matrix instead of the correlation matrix provides insights into how variables vary together, which can also be helpful in detecting multicollinearity. The singular values obtained from this method offer a different perspective on the interdependence of predictors. Additionally, some publications mention computing singular values as the square roots of the eigenvalues of the square matrix X'X, where X represents the design matrix in multiple linear regression. This method focuses on the structure of the predictor variables within the regression model. When comparing your calculations with MATLAB results using collintest, discrepancies may arise due to various factors such as rounding errors, algorithm differences, or specific implementation details. It's crucial to understand how MATLAB processes and interprets these calculations to interpret its output accurately.

[N/A]

Hello Umar, thankyou for your clarification. Also I was wondering if I could use Belsley collinearity diagnostics to get the same output because according to your given example I am getting diference answers for singular values. In other words I am gettjng different singular values when using svd versus when using collin test. Again thankyou!

[Umar]

Hi Nafisa, Glad to help, to answer your question regarding Belsely collinearity diagnostics, we have to understand the concept of comparing the singular values obtained from SVD (Singular Value Decomposition) with those from the Belsley collinearity diagnostics in Matlab, differences may arise due to the nature of the methods. SVD directly computes the singular values of a matrix, while collinearity diagnostics like the collin test in Matlab focus on assessing multicollinearity in regression models rather than directly computing singular values. It's essential to understand the specific purpose and methodology of each approach to interpret the results correctly. If you seek singular values, SVD is the appropriate method, whereas collinearity diagnostics are more suitable for assessing multicollinearity in regression analysis.

Hope this help clarifies to resolve your problem.

[N/A]

Below I have attached a dataset of boston house prices and when I tried calculating condition indices for each singular value, dividing the largest singular value by each of the singular values individually using your code. However I am getting different answer as to real one. I have also attached the example page. https://stataiml.com/posts/42_condition_index_r/#input-dataset

Answer 1

[Umar]

Hi Nafisa,

In order to help you further with your problem, can you share the matlab code which is causing error, I have to review it in order to share my detailed thoughts. Hope that should not be a problem.

[N/A]

Absolutely,

data =  readtable('boston house prices.xlsx', 'VariableNamingRule','preserve');

x1 = data.CRIM;
x2 = data.ZN;
x3 = data.INDUS;
x4 = data.CHAS;
x5 = data.NOX;
x6 = data.RM;
x7 = data.AGE;
x8 = data.DIS;
x9 = data.RAD;
x10 = data.TAX;
x11 = data.PTRATIO;
x12 = data.B;
x13 = data.LSTAT;

Predictors = [x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13];

[U, S, V] = svd(Predictors); % Singular value decomposition
singular_values = diag(S) % Extract singular values
condition_index = (max(singular_values) ./ singular_values)

Answer 2

[Umar]

Hi Nafisa,

Due to incorrect calculation of the condition indices formula. The condition index should be the ratio of the largest singular value to the smallest singular value, not to each singular value individually. So, in order to correctly calculate the condition indices, modify the calculation as follows:

condition_indices = singular_values(1) ./ singular_values;

By modifying it will ensure that the largest singular value divided by each singular value individually, which gives us the correct condition indices for the predictors.

Hope this will help resolve your problem.

Answer 3

[Paul]

Hi NAFISA,

Using the example from the the doc collintest

load Data_Canada

Output from collintest

[sValues,condInx] = collintest(Data);

Variance Decomposition

 sValue  condIdx   Var1    Var2    Var3    Var4    Var5  
---------------------------------------------------------
 2.1748    1      0.0012  0.0018  0.0003  0.0000  0.0001 
 0.4789   4.5413  0.0261  0.0806  0.0035  0.0006  0.0012 
 0.1602  13.5795  0.3386  0.3802  0.0811  0.0011  0.0137 
 0.1211  17.9617  0.6138  0.5276  0.1918  0.0004  0.0193 
 0.0248  87.8245  0.0202  0.0099  0.7233  0.9979  0.9658 

The same output for the sValue and condIndx can be found by (not including various error checks)...

Scale Data so that each column has unit magnitude

sData = Data./vecnorm(Data,2,1);

Take the svd of the scaled data.

[~,S,V] = svd(sData);
S = diag(S);

Compute the indices

idx = max(S)./S;
table(S,idx)

ans = 5x2 table
       S         idx  
    ________    ______

      2.1748         1
     0.47889    4.5413
     0.16015    13.579
     0.12108    17.962
    0.024763    87.825

Hopefully that makes sense in the context of what collintest is supposed to do. I have zero knowledge of that function.

[Umar]

Hi Paul,

I highly appreciate your efforts to resolve this issue. It was a great teamwork effort. Happy 4th of July.

[N/A]

Hello Umar, I udnerstand singular values from svd function in matlab is different from singular values from collin test. My question is how do I calculate the singular values from collin test and from svd function theoretically.

[Paul]

This answer that you're commenting on showed exactly how to compute the sValue output from collintest.

First, scale the columns of the data matrix so that they all have unit 2-norm. Then take the svd of the scaled data matrix.

Also,

a) one can take the svd of a non-square matrix, as was shown in this answer

b) one cannot calcuate the singular values of X by examining the condition number of X

c) The singular values of X are not the eigenvalues of X'*X; they are the sqrt of those eigenvalues.

Illustrating points (a) and (c)

rng(100)
X = rand(3,2);  % non-square matrix
[svd(X) sqrt(sort(eig(X'*X),'descend'))]

ans = 2x2

1.0771 1.0771 0.3483 0.3483

<mw-icon class=""></mw-icon>

<mw-icon class=""></mw-icon>

[N/A]

Hello Paul, Thank you for helping me with this. However I need some more assistance in calculating the variance decomposition proportions in collin test theoretically. Below I am attaching a dataset below

where the predictors are Age, Weight, RunTime, RestPulse, RunPulse and MaxPulse.

[Paul]

Why not just use the variance decomposition proportions that are output from collintest? Also, that doc age discusses how those proportions are calculated.

[N/A]

Hello Paul, I need to understand theoreteically how it is calculated. I did not quite inderstand when you said doc age

[Paul]

I meant the doc page. Here is the link

Answer 4

[Umar]

Hi Nafisa,

You asked how do I calculate the singular values from collin test and from svd function theoretically.

To answer this question, regarding collinearity test, you can calculate singular values by examining the condition number of a matrix. The condition number is the ratio of the largest to the smallest singular value. Higher condition numbers indicate a higher degree of collinearity. While using the Singular Value Decomposition (SVD) function in Matlab, you can directly compute the singular values of a matrix.

[N/A]

Hello Umar, I did not quite understand when you said by examining the condition number of a matrix. It would be better if you can provide a formula or something. Also you said that While using the Singular Value Decomposition (SVD) function in Matlab, I can directly compute the singular values of a matrix.To compute the singular values I need to have a square matrix so I tried finding the eigenvalues of X'X where X isthe design matrix and when I did that I am getting different eigenvalues.

Answer 5

[Umar]

Hi Nafisa,

Glad to hear back from you. Please see my answers to your comments below.

Comment#1:did not quite understand when you said by examining the condition number of a matrix. It would be better if you can provide a formula or something.

Answer:The condition number of a matrix measures its sensitivity to changes in input, where a high condition number indicates potential numerical instability. In Matlab, you can compute the condition number of a matrix A using the cond() function: cond(A).The condition number of a matrix indicates how sensitive the matrix is to changes in its input values.Here is an example demonstrating how to compute the condition number of a matrix A in Matlab by creating a 2x2 matrix A and then use the cond() function to calculate its condition number. The result is displayed using disp().

>> % Define a matrix A A = [1, 2; 3, 4];

>> % Calculate the condition number of matrix A condition_number = cond(A);

>> disp(['The condition number of matrix A is: ', num2str(condition_number)]);

This information is valuable in assessing the stability and accuracy of numerical computations involving the matrix A.

Comment#2:Also you said that While using the Singular Value Decomposition (SVD) function in Matlab, I can directly compute the singular values of a matrix

Answer: Yes, you can directly compute the singular values of a matrix A using the svd() function: [U, S, V] = svd(A). Ensure that A is a square matrix for SVD computation.

Comment#3: To compute the singular values I need to have a square matrix so I tried finding the eigenvalues of X'X where X isthe design matrix and when I did that I am getting different eigenvalues.

Answer: When finding the eigenvalues of X'X, ensure that X is properly defined and that X'X results in a square matrix. To compute eigenvalues, you can use the eig() function: eigenvalues = eig(X' * X). Ensure that X is correctly defined to match the expected behavior.