电信客户流失Logistic回归系数异常及变量显著性问题咨询

I'm working on building a logistic regression model to predict telecom customer churn, using a dataset that includes variables like Account_Age, Current_Bill_Amount, Avg_Days_Delinquent, complaints, and Avg_Calls.

The complaints variable has 6 distinct categories, so I converted it into 5 dummy variables, using 'pricing' complaints as the reference category. From my initial data exploration, customers with Call Quality or Billing Problem complaints have drastically higher churn rates—both in absolute numbers and as a proportion of their group—while other complaint types seem to have little impact on churn.

But my model results are throwing me for a loop with two key issues:

1. Why is "Call Quality" not significant despite an 81% churn rate? #

In my model output, the Call Quality dummy variable has a p-value of 0.527, marking it as statistically insignificant. But in the raw data, customers with call quality complaints have an 81% churn rate—this is a massive mismatch between what the data shows and what the model is telling me. What could be causing this disconnect?

2. Why do significant complaint variables have negative coefficients? #

Three of the complaint dummy variables came out as statistically significant (p < 0.05): Billing Problem, Check Account, and Moving, with coefficients of -1.0033, -2.5675, and -2.1132 respectively. It makes intuitive sense that customers with these complaints would be more likely to churn, so I expected positive coefficients. Instead, they're negative. How do I interpret this, and what might be driving it?

Here's the code I used to fit the model:

import statsmodels.api as sm
logReg=sm.Logit(Y_train,X_train)
logistic_regression=logReg.fit()
logistic_regression.summary()

内容的提问来源于stack exchange，提问作者Vikrant Arora