Linear regression is often the first model we learn and the one we misuse the most.
Not because it’s complicated, but because it’s too easy to run.
A few lines of code, a decent score, and we move on.
But linear regression isn’t just a model.
It quietly assumes a lot about your data.
And if those assumptions don’t hold, your results might still look fine, just not reliable.
Before trusting a linear regression model, there are four things worth checking:
1. Linearity
Are you actually modeling a straight-line relationship?
If the data is curved and you force a line through it, the model will miss patterns.
A quick scatter plot or residual plot usually tells the story.
2. Independent Observations
Your data points shouldn’t influence each other.
This becomes important in time-based data, where yesterday affects today.
If errors are connected, your model is learning patterns it shouldn’t.
3. Normality of Errors
The residuals (errors) should roughly follow a normal distribution.
This matters if you care about interpreting the model, not just predicting.
4. Homoscedasticity
In simple terms, your model should be equally wrong across all values.
If errors grow or shrink in certain regions, the model becomes biased.
The tricky part is that none of these checks stop your model from running.
You’ll still get predictions.
You’ll still get metrics.
But without checking these, you’re not really using linear regression, you’re just fitting a line and hoping it works.
A small habit that changes everything:
Before tuning the model, spend a few minutes validating these assumptions.
When was the last time you actually checked them before trusting your model?
