2
$\begingroup$

I have been working on a dataset pertaining to 'churn analysis'. I have been trying to demonstrate whether the customers that are being charged more are also the ones that churn more or not. My dataset consists of as many as 21 variables and has more than 3.3K records. There are 4 columns that refer to the charges that are being imposed on the customers, separately for day, evening, night, and international. 85% of the dataset is classified into 'False' category, i.e. the customers that did not churn, and the rest into its 'True' counterpart. I analyzed the effect of higher charges simply by plotting the PDFs (employed Histograms too but that obviously wasn't a good choice considering the imbalanced dataset that I have). Then I added the corresponding records in all these 4 variables and made one single variable for the total charges. Following is what I found out (of course on a sample);

enter image description here

However, I'm now thinking whether multicollinearity might have influenced such a trend that I depicted or not? I also want to know whether I need to take into consideration the other variables in the dataset or not? If yes, then in what ways, use what techniques, find out what using them, and most importantly, why?

Below is a glimpse of the dataset that I have been working on;

enter image description here

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

You can seek a model churn ~ number_of_charges + international_plan + ... and include any number of predictors. Each predictor will increase or decrease the probability of churn. Churn is a logical variable, so it would be a classification model or a logistic regression.

You would have to address multicollinearity and overfitting somehow, with possibly regularized regression. For example cv.glmnet(x, y, family = "binomial") or train(formula, data, method = "glmnet", family = "binomial").

I can anticipate a challenge: the customers that didn't churn might have a longer history and have accumulated more charges, etc. This could confound your results unless you can find a way to address.

Recommend also you think about those histograms. If you're scaling them somehow to plot two histograms on the same plot, it may be problematic to make a statement about the probability. The height of a histogram or a pdf is not the same as probability.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you @Arthur but the question that I have been working upon is simple; "whether the customers that are being charged more are also the ones that churn more or not". And the question that I asked too was pretty straightforward; "whether multicollinearity might have influenced such a trend that I depicted or not?" Alongside some supplementary questions. I did not get the answer to either of these. Could you please elaborate on my questions, perhaps you might have misinterpreted my questions earlier :) $\endgroup$
    – Ritik P. Nayak
    Feb 8, 2022 at 19:11
  • $\begingroup$ My answer is "yes." It's my opinion that in order to do an EDA with multiple variables, you should fit a simple model. A model is the only way to understand the simultaneous action of a variety of predictor variables. Plots and correlation coefficients only test pairwise relationships. $\endgroup$
    – Arthur
    Feb 8, 2022 at 19:30

Your Answer

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

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