This wiki page Simple linear regression has formulas to calculate $\alpha$ and $\beta$. Could anyone tell me how to derive the formulas in weighted case?


2 Answers 2


Think of ordinary least squares (OLS) as a "black box" to minimize

$$\sum_{i=1}^n (y_i - (\alpha 1 + \beta x_i))^2$$

for a data table whose $i^\text{th}$ row is the tuple $(1, x_i, y_i)$.

When there are weights, necessarily positive, we can write them as $w_i^2$. By definition, weighted least squares minimizes

$$\sum_{i=1}^n w_i^2(y_i - (\alpha 1 + \beta x_i))^2$$

$$=\sum_{i=1}^n (w_i y_i - (\alpha w_i + \beta w_i x_i))^2 .$$

But that's exactly what the OLS black box is minimizing when given the data table consisting of the "weighted" tuples $(w_i, w_i x_i, w_i y_i)$. So, applying the OLS formulas to these weighted tuples gives the formulas you seek.


The answer from whuber gives the intuition behind the maths, which is nice to have, but I could still not figure the formulas (i.e. where should I put the weights). After some search on the web, I found these slides which give the following:

You want to minimize the following error: $$\sum_{i=1}^n (w_i(y_i - (\alpha + \beta x_i)))^2$$

Then, the optimal pair $(\hat\alpha,\hat\beta)$ is: $$\hat\alpha = \overline y_w - \hat\beta \overline x_w$$ $$\hat\beta = \frac{\sum_{i=1}^n w_i(x_i-\overline x_w)(y_i-\overline y_w)}{\sum_{i=1}^n w_i(x_i-\overline x_w)^2}$$

Where $\overline x_w$ and $\overline y_w$ are the weighted means:

$$\overline x_w = \frac{\sum_{i=1}^n w_ix_i}{\sum_{i=1}^n w_i}$$ $$\overline y_w = \frac{\sum_{i=1}^n w_iy_i}{\sum_{i=1}^n w_i}$$


Your Answer

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

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