Hey there! Let’s tackle your logistic regression issue head-on. The version mismatch between Python 3.5 (your friend’s setup) and 3.6 (your Spyder environment) is almost certainly playing a role here—often tied to differences in statsmodels or pandas versions, not just Python itself. Here’s how to diagnose, fix, and optimize your code:

Step 1: Diagnose the Root Cause #

First, let’s narrow down why the error is happening:

• Check statsmodels versions: Ask your friend to run import statsmodels; print(statsmodels.__version__) and compare it to yours. Python 3.5 typically pairs with older statsmodels versions (e.g., 0.8.x), while 3.6 can run newer releases (0.9.x+) that have API changes.
• Pinpoint the error message: Even if you didn’t share it, common issues include:

  ValueError: Singular matrix: This happens when your dummy variables create perfect multicollinearity (e.g., keeping all categories for a variable instead of dropping one baseline category).
  Attribute errors: Newer statsmodels versions may have renamed parameters or methods.

Step 2: Fix the Code for Python 3.6 #

Here are the most likely fixes based on version differences:

Fix 1: Resolve Multicollinearity in Dummy Variables #

Older pandas versions (used with Python 3.5) might have handled dummy variables differently, but in newer versions, you need to explicitly drop a baseline category to avoid singular matrix errors:

import pandas as pd
# Replace this (risky for multicollinearity)
X_dummies = pd.get_dummies(your_data[categorical_columns])
# With this (drops first category to eliminate perfect correlation)
X_dummies = pd.get_dummies(your_data[categorical_columns], drop_first=True)

Fix 2: Add a Constant Term (If Missing) #

Newer statsmodels versions require you to manually add a constant term to your feature matrix—older versions might have done this automatically:

import statsmodels.api as sm
# Add constant to your feature matrix
X = sm.add_constant(X_dummies)
# Now fit the model
model = sm.Logit(your_target_variable, X).fit()

Fix 3: Use the Formula API (Simpler & More Robust) #

The statsmodels formula API handles dummy variables automatically, reducing manual errors and version-related bugs. It’s cleaner and less prone to issues across versions:

import statsmodels.formula.api as smf
# Define your formula (C() denotes categorical variables)
formula = "target ~ C(category_col1) + C(category_col2) + numeric_col1 + numeric_col2"
# Fit the model
model = smf.logit(formula, data=your_data).fit()
# Print results
print(model.summary())

Step 3: Optimization Tips #

To avoid version-related headaches and improve your workflow:

• Lock your dependencies: Create a requirements.txt file to ensure everyone uses the same package versions. Example:

  pandas==1.1.5
  statsmodels==0.12.2
  numpy==1.19.5
  

  Install with pip install -r requirements.txt to replicate the working environment.
• Validate your data: After creating dummy variables, check for multicollinearity with:

  # Print correlation matrix (look for values close to 1.0)
  print(X_dummies.corr().abs())
  

• Add error handling: Catch and print detailed error messages to debug faster:

  try:
      model = sm.Logit(y, X).fit()
      print(model.summary())
  except Exception as e:
      print(f"Error details: {str(e)}")
      print(f"Feature matrix shape: {X.shape}")
      print(f"Feature columns: {X.columns.tolist()}")
  

内容的提问来源于stack exchange，提问作者Sebastian