Why a Regression Coefficient Can Turn Negative When the Bivariate Relationship Is Clearly Positive

One plot that instantly breaks every intuition you had about regression coefficients.

Take a good look at that green line sloping down.

That single line is the clearest proof you’ll ever see that regression coefficients can lie — beautifully — when multicollinearity is severe.

Why Can a Coefficient Become Negative in Multiple Linear Regression

When Its Bivariate Relationship Is Clearly Positive?

One of the deepest and most frequently misunderstood concepts in multiple linear regression:

The regression coefficient is not the simple bivariate slope.

It is the partial effect of that variable while holding all other predictors constant.

When severe multicollinearity exists, "holding all other predictors constant" becomes a counterfactual — often physically impossible — scenario. In such cases, the sign of the coefficient can completely reverse.

We demonstrate this using the classic Canadian FuelConsumptionCo2 dataset.

import numpy as np

import pandas as pd

import matplotlib.pyplot as plt

from sklearn.linear_model import LinearRegression

from sklearn.metrics import r2_score

import warnings

warnings.filterwarnings('ignore')



%matplotlib inline

Load the dataset

df = pd.read_csv("/home/pyrz-tech/Desktop/MachineLearning/FuelConsumptionCo2.csv")

df.head()

Select features and target

features = ['ENGINESIZE', 'CYLINDERS', 'FUELCONSUMPTION_COMB',

            'FUELCONSUMPTION_CITY', 'FUELCONSUMPTION_HWY']

target = 'CO2EMISSIONS'



X = df[features]

y = df[target]

Train/test split

np.random.seed(42)

msk = np.random.rand(len(df)) < 0.8

train = df[msk]

test = df[~msk]



print(f"Train size: {len(train)}, Test size: {len(test)}")

Train size: 848, Test size: 219

Fit multiple linear regression

reg = LinearRegression()

reg.fit(train[features], train[target])



print("Coefficients:", reg.coef_.round(3))

print("Intercept:   ", reg.intercept_.round(3))

Coefficients: [ 10.962   7.25  -10.937  12.227   8.104]

Intercept:    65.076

Model performance

pred = reg.predict(test[features])

r2 = r2_score(test[target], pred)

print(f"R² on test set: {r2:.4f}")

R² on test set: 0.8712

The diagnostic plot that reveals everything

plt.figure(figsize=(13, 8))



colors = ['blue', 'orange', 'green', 'red', 'purple']

labels = features



for i, col in enumerate(features):

    plt.scatter(test[col], test[target], color=colors[i], alpha=0.6, label=labels[i])



# The magic (technically "wrong") line that draws the exact multivariate coefficients

plt.plot(test[features].values,

         reg.coef_ * test[features].values + reg.intercept_,

         linewidth=3)



plt.xlabel('Feature value', fontsize=12)

plt.ylabel('CO₂ Emissions', fontsize=12)

plt.title('Each line = exact partial effect of that feature in the multivariate model', fontsize=14)

plt.legend()

plt.grid(alpha=0.3)

plt.show()

Why the sign reversal happens

The three fuel consumption variables are nearly perfectly collinear:

FUELCONSUMPTION_COMB ≈ 0.55 × CITY + 0.45 × HWY

In real data, you almost never observe a change in CITY or HWY while COMB stays fixed.

The regression coefficient for HWY answers the question:

"What happens to CO₂ if I increase HWY while holding COMB, CITY, and all other variables constant?"

That scenario is physically almost impossible → the coefficient becomes negative purely as a mathematical compensation mechanism.

This is known as sign reversal due to severe multicollinearity (or the suppressor effect).

Key takeaways

1. High R² + unexpected coefficient signs → almost always severe multicollinearity


2. In severe multicollinearity, coefficients lose causal interpretability


3. The plot above is one of the clearest diagnostic and teaching tools for this phenomenon

From now on, whenever you see a coefficient with the “wrong” sign, don’t argue.


Just draw this plot.

The discussion ends in 5 seconds.

If this just saved you hours of confusion, hit that ❤️ — it helps more people find it!