Your analysis is very interesting one.  Your using logistic regression model with 19 variables.  First you have run the analysis with suitable model (Binary, Multiple or ordinal) and then find the significant variables.  Then fit the model.

All the best.

Hi Dr. Senthilvel,

I would be deploying ordinal regression as the outcome is either "SAT" or "D-SAT" for the survey data.

I can add variable one by one and measure the AIC values and the null and residual deviance. Based on this i can add or drop variables however i was looking for some correlation test for these categorical variables. 

Thanks.

To check for correlations among pairs of categorical variables, you can use a chi-square test of association, or related tests (Fisher exact, G-test)

Hi,

First, what is your dependent variable? Based on that you need to decide which regression model be used. Second, perform bivariate analysis and check as which predictors show significant relationships, particularly with the dependent variable.  Third, please reduce the number of predictors as much as possible. This will ease your work and help in indicating the original predictors responsible for variation in the dependent variable. 

I guess it helps !!

Suppose you put all 19 variables into the model. You can use something called VIF or Tolerance to determine how all of the 19 variables correlate with each other in the model.

It is very possible that you could have 2+ factors in the model that correlate with one other factor. Where as a series of correlation studies will only compare one to another. So, you will miss these correlations. The software I use gives me VIF for each term in the model. It doesn't care if my factors are continuous or categorical. It automatically calculates VIF. Usually, you want the VIF to be less than 10.00. Some will say 5.00.

If the correlation is too high, your standard error will be larger than it should be. Thus, you will have some terms that look insignificant, but really are significant, just hidden because of the VIF.

SAS uses tolerance. Tolerance is 1/VIF. 

There is some confusion i believe. as per the below post,

"https://www.researchgate.net/post/How_to_test_multicollinearity_in_logistic_regression/1"

the forum has also suggested the use of VIF for logistic however as per this:

https://communities.sas.com/t5/SAS-Statistical-Procedures/multicollinearity-in-Logistic-Regression/m-p/291474#M15491

suggested is VIF is for linear. so lot of confusion building here. 

Hi, 

Yes, in this analysis you have use the stepwise analysis.  Then only you will get the correct influencing variables (significant variables) with high R - Square values.

All the best to you.

Hi, 

Seems like there is confusion on the outcome. I am working on a logistic regression hence there is no way i will get the R- squared values. 

Logistic regression will give me log odds, AIC, SC, deviance and null deviance so these would be measuring criteria not R-squared which is when we run a linear model. 

Hi, Shivi Bhatia.  Your approach is a reasonable one.

Good ways to choose what variables to include in a model include AIC, BIC (which will tend to choose a model with fewer parameters than AIC), AICc (which I think helps when there are fewer observations).

People also use a stepwise approach where they look at the p-value for each term.

But there is nothing that can substitute for examining the data, seeing which independent variables correlate with what other variables, seeing which independent variables correlate with the dependent variable, and most importantly --- if possible --- using logic to determine which variables make the most sense to include in the model.

By the way, you can use various pseudo R-squared statistics for logistic regression, although their use and interpretation are controversial.  For example, if you look at Table 5 below, I reported the marginal increase in the pseudo R-squared value for each term included in the final model.  I think reporting these can help the reader understand how important each term is, even if it is an imperfect measurement.

http://www.joe.org/joe/2016june/rb1.php

Thanks Salvatore for the suggestion. I will follow the approach you have mentioned, even though this could be a time consuming one but a more robust one. 

I would also look into the material that you have shared. Thanks again for your help. 

You're welcome.  But I wasn't really suggesting you take a different approach.  I think what you outlined originally is a good way!

VIF is for you "X-variables". It doesn't matter if you use logistic regression, linear regression or poisson regression. VIF only looks at your X-matrix.  

Md Kamrul Hasan , Spearman's correlation makes sense for categorical data only if 1) the data are ordinal; or 2) there are only two levels in the categorical variable.

Hi all, A naive question. I am running a mixed Poisson model in stata, then i wanted to test for multi-collinearity. I discovered that stata command allows vif only for regress command from some earlier discussions. so I ran the model with regress command instead of mepoisson and took out the clusturing levels (but certainly did not use the coefficients or CI from the model). then i ran the vif command and found one category of two variables each had VIF>10. I removed the variable with highest VIF reran and found vif for none of categories of other variables is >5. now what should be ideal should I remove the entire variable (with multiple categories) that has the highest VIF or I have to create dummy variable for each of the categories then remove the variable of the category that has high VIF? Please suggest