What are the most relaxed assumptions to get consistency of the linear regression estimates with $p$ variables?

The most basic assumptions that I know are in White (1984):

1) The model is correct

2) $X'\epsilon/n = op(1)$, with $\epsilon = Y - X\beta$

3) $X'X/n - M_n= op(1)$, with $M_n = Op(1)$ and uniformly positive definite

Assumption (3) implies that $(X'X/n)^{-1}$ exists asymptotically and is $Op(1)$. Then $$\hat{\beta} - \beta = (X'X/n)^{-1}(X'\epsilon/n) = Op(1)op(1) = op(1) $$

Did anyone since White came up with more general assumptions?


1 Answer 1


Maybe not more general, but different.

Let $A = [X'X]^{-1}X'$, and assume

$$E[\epsilon_k | X] = 0, k = 1,..., n$$

Then, $\hat{\beta} = \beta + op(1)$ if one of the following hold:

(1) $\epsilon_k$ are iid, and $\max_{i,j} | A_{ij}| = O(n^{-1})$

(2) $\epsilon_k$ are iid with finite variance, and $\max_{i,j} | A_{ij}| = o((n \log \log n)^{-1/2})$

(3) $\epsilon_k$ are iid with finite variance, and $(X'X)^{-1} = o(1)$

(4) $\epsilon_k$ are uniformly integrable, pairwise independent, $\max_{i,j} | A_{ij}| = o(1)$, $\sum_{i,j} | A_{ij}| = O(1)$

Note: Condition 4 is true if the probability of $X'X'$ being invertible tends to one, and $\limsup_{n\to\infty} n [\lambda_{min}(X'X)]^{-1} < \infty$, where $\lambda_{min}(A)$ is the smallest eigenvalue of $A$.

Proofs: Miao and Xu, 2012; Wang & Rao 1984


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.