In a simple linear regression

$$y=\alpha+\beta x + \varepsilon,$$

if $x$ is binary we can show that the slope coefficient ($\beta$) is equal to the difference in the outcome variable between levels of the predictor, i.e.

$$E(y\mid x) = \alpha+\beta x$$ so $$E(y\mid x=0) = \alpha$$ and $$E(y\mid x=1) = \alpha+\beta$$.

We then have: $$E(y\mid x=1) - E(y\mid x=0) = \beta.$$

However, in a multiple regression this usually is not the case. When we control other variables, some of the variation in $y$ originally explained by $x$ may be partitioned into the new covariate, which changes the estimate on $\beta$ so that the original relationship of $E(y\mid x=1) - E(y\mid x=0) = \beta$ will no longer be true.

I understand this intuitively and can show it with data, but I'm not sure how to express it mathematically with formulas (?).

An illustration using data in R:

# generate data
outcome    <- rnorm(n=100, mean = 10, sd = 1)
original_x <- rbinom(n=100, 1, prob=.5)
new_x      <- rnorm(n=100, mean = 10, sd = 10)

# Simple Linear Regression
coef(glm(outcome~original_x))[2]  # 0.08680985
mean(outcome[original_x==1])-mean(outcome[original_x==0]) # 0.08680985

# Multiple Regression
coef(glm(outcome~original_x + new_x))[2] # 0.1112377 

1 Answer 1


Going to change up the notation a bit here, to match what you really did in your code. We are assuming that $E(Y|X=x) = x\beta$, so, given $X = x$,

$$Y = x\beta + \epsilon$$

where $\beta = (\alpha, \beta_1)^T$, and $x$ is an $n \times 2$ matrix with observations as rows. Implicitly, we are also assuming that $x$ is full-rank.

As I'm sure you know, the OLS estimate for $\beta$ is

$$\hat{\beta}_{OLS} = (x^Tx)^{-1}x^Ty$$

where $y$ is our observed response. What happens if we add a new regressor? Let the vector of observations for this new regressor be $x_a$. Then, we can define a new design matrix $x_{new} = \begin{pmatrix} x & x_a\end{pmatrix}$ which is $n \times 3$. Our new regression coefficients are

$$\hat{\beta}_{new} = (x_{new}^Tx_{new})^{-1}x_{new}^Ty = \begin{pmatrix} x^Tx & x^Tx_a\\ x_a^Tx & x_a^Tx_a\end{pmatrix}^{-1}\begin{pmatrix}x^Ty \\ x_a^Ty \end{pmatrix}$$

where we're now estimating three coefficients, namely $\beta = (\alpha, \beta_1, \beta_2)$. Notice, that the cross terms $x^Tx_a$ and $x_a^Tx$ are in a sense "responsible" for why the estimates for $\alpha$ and $\beta_1$ change from $\hat{\beta}$ to $\hat{\beta}_{new}$.

From here, I'd like to make a qualification to your statement. In most cases, you are right; your coefficient estimates change. In the case when the columns of $x$ are orthogonal to $x_a$, however, the vectors $x^Tx_a$ and $x_a^Tx$ are both matrices of different sizes, with all elements zero. Further algebraic trickery (taking the inverse of a $2 \times 2$ block matrix, which I won't do here) will reveal that

$$\hat{\beta}_{new} = \begin{pmatrix} (x^Tx)^{-1}(x^Ty)\\ (x_a^Tx_a)^{-1}x_a^Ty \end{pmatrix}$$

Therefore, in this case, each regressor "stays in its own lane"; thus, the addition or removal of a certain regressor will not change the estimates for the others.

  • $\begingroup$ Thank you very much! I believe this is what I need. I think I know what you mean by this last part, that if the original predictor x and the new predictor xa were perfectly orthogonal then theoretically the coefficient on x wouldn't change, but to be clear could you clarify what you meant about "are both equal to the zero vector (of different lengths!)"? $\endgroup$
    – Vanesh
    Dec 13, 2020 at 1:52
  • $\begingroup$ I'm just saying that $x^Tx_a$ is the zero vector of size $3 \times 1$ and $x_a^Tx$ is the transpose of the zero vector of size $3 \times 1$. I've changed the wording to be clear. $\endgroup$
    – user303375
    Dec 13, 2020 at 1:59
  • $\begingroup$ Thank you, that helps! $\endgroup$
    – Vanesh
    Dec 13, 2020 at 2:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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