1
$\begingroup$

I have a table which contains about 80 variables. I will give one variable as input and as output I want to have all the variables that are related to the input variable.

Example: I have hospital data which contains variables such as patient name, blood pressure, sugar, heart rate, ECG, "heart patient" (binary yes/no variable), etc. If I give "heart patient" as input, then as output I would get ECG, heart rate, blood pressure, etc.

So I need to find variables which are related to the "heart patient". And the rest will be ignored.

I am using principle component analysis (PCA) to solve this. The problem is that PCA gives me the number of variables which best represent the data set, not the variables that are related to a given variable.

Please suggest me some idea or keywords that help me to proceed. Also, could you give me some reading which would improve my PCA understanding. I have already read Making sense of principal component analysis, eigenvectors & eigenvalues.

$\endgroup$
1
  • $\begingroup$ no comments , no answers , Is there a problem in my question ? $\endgroup$
    – munjal007
    Mar 12, 2015 at 17:19

2 Answers 2

1
$\begingroup$

I don't immediately see how PCA would be useful here, since PCA is more of an approach for characterizing an entire dataset in one go than a way of characterizing all the variables in terms of a single specially chosen variable.

I think the simplest approach to this problem is to, given the input variable, calculate its correlation (Kendall's τ is a nice general-purpose notion of correlation) with every other variable. Sort these correlations by absolute value to make a list of the variables in order of how closely related they are to the input variable. Choose a threshold absolute correlation, or take the n most closely related, and you have a subset of the variables that are more related to the input variable than the other variables are.

$\endgroup$
0
0
$\begingroup$

What PCA does is "squeeze" the variance of the original variables into a smaller set of variables (the principal components). The problem you describe requires a Feature Selection algorithm or a simple correlation analysis between the input variables and your chosen output. If you insist (or you're required) to do the task using PCA, then search the literature for "feature selection using pca". However, I suggest you don't go in the latter direction if you're just starting to use PCA.

$\endgroup$
3
  • $\begingroup$ "Feature selection using PCA" does not take an "input variable" into account either. So how would it be useful here? $\endgroup$
    – amoeba
    Mar 12, 2015 at 18:10
  • 1
    $\begingroup$ You're right. That particular paper won't give you an output-specific ranking of your variables (output is what you refer to as "input variable"; It is in fact an output, as the variables would like to select would be inputs to it, if you had a model). As far as I know, no ordinary PCA method can do that, atleast, not in a straightforward way, so you should look elsewhere. $\endgroup$
    – amjams
    Mar 13, 2015 at 13:40
  • $\begingroup$ I am not the OP :) But yes, I agree. $\endgroup$
    – amoeba
    Mar 13, 2015 at 13:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.