I am working on a spatial linear regression and I can tell there is collinearity between covariates. Can I use PCA (Principal Component Analysis) images instead of original covariates to estimate the dependent variable? I am assuming PC1=Variable 1, PC2=Variable 2, etc. Or are there any other methods to solve the collinearity problem?
$\begingroup$
$\endgroup$
7
gung - Reinstate Monica
148k89 gold badges406 silver badges716 bronze badges
asked Aug 25, 2014 at 16:10
-
1$\begingroup$ What is your goal in fitting the regression model? Collinearity is only a problem in some cases. $\endgroup$– shadowtalkerCommented Aug 25, 2014 at 16:17
-
$\begingroup$ I am trying to estimate the response variable from covariates using spatial linear regression. $\endgroup$– Kaleab WoldemariamCommented Aug 25, 2014 at 16:24
-
1$\begingroup$ To be more specific: are you interested in the magnitudes and directions of the coefficients, or making accurate predictions, or both? $\endgroup$– shadowtalkerCommented Aug 25, 2014 at 16:34
-
$\begingroup$ @ssdecontrol .Both the coefficients and sign of the predictors as well as accurate estimation of response variable using these PC's of covariates. $\endgroup$– Kaleab WoldemariamCommented Aug 25, 2014 at 16:45
-
$\begingroup$ Thanks for clarifying, that will help me write a more helpful answer. $\endgroup$– shadowtalkerCommented Aug 25, 2014 at 16:54
|
Show 2 more comments
1 Answer
$\begingroup$
$\endgroup$
There is a method called partial least square that is something very close to what you are trying to do.
https://en.wikipedia.org/wiki/Partial_least_squares_regression
The choice to use the PCA transformation of the data can lead to a better estimation of the output $y$ but to understand the role of the original variables will be more difficult.
I suggest you to start with the lasso estimator https://en.wikipedia.org/wiki/Lasso_%28statistics%29#Lasso_method
