# When is OLS unbiased and what is the relation with weak orthogonality? [duplicate]

So I have a question about the unbiasedness of the OLS estimator. It is unbiased when $E\{\epsilon|X\} = 0$ and some other assumptions, where $X$ is the regressor. Is it still unbiased if I relax the requirement and say that $E\{\epsilon\}=0$ and $E\{X\epsilon\} = 0$? I read a paper and it is refered to as the weak orthogonality. However, I do not know if these two conditions: $E{\epsilon}=0$ and $E{X\epsilon} = 0$ can lead to unbiasedness, or just consistency?

• Are you considering $X$ to be random, or fixed? If $X$ is fixed then $E(\epsilon) = 0 \Rightarrow E(X \epsilon) = X E(\epsilon) = 0$. Commented Oct 28, 2015 at 23:15
• No, $X$ is random. I am not sure if these conditions lead to consistency or unbiasedness. Commented Oct 28, 2015 at 23:23
• Do you have an alternate model in mind besides $Y = X\beta + \epsilon$ for the true distribution of $Y|X$? Are we omitting a predictor or otherwise mis-specifying the model? Or are you defining the estimand to be $E( (X^T X)^{-1} X^T y)$ (ie, the population OLS estimate for covariates $X$). Commented Oct 28, 2015 at 23:40
• Yes the estimator should be $(X^TX)^{-1}X^Ty$. No we are not omitting a predictor. I am curious as to when the OLS is unbiased, is $E\{\epsilon|X\}=0$ sufficient or not? Commented Oct 29, 2015 at 0:11
• Because I think the conditions $E\{\epsilon\} = 0$ and $E\{X\epsilon\} = 0$ are weaker conditions, I don't know if they still lead to unbiasedness or just consistency... Commented Oct 29, 2015 at 0:13

The OLS estimate is defined by $$\hat \beta = \left(X^T X \right)^{-1}X^Ty$$ and we wish to find out when $E(\hat \beta) = \beta$ under the model $Y = X \beta + \epsilon$ with both $X$ and $\epsilon$ drawn from distribution $P$.

$$E (\hat \beta) = E \left( \left( X^T X \right)^{-1}X^T \left( X \beta + \epsilon \right) \right)$$

$$= E \left( \left( X^T X \right)^{-1} \left(X^T X \right) \beta + \left( X^T X \right)^{-1}X^T \epsilon \right)$$

$$= E \left( \beta \right) + E \left( \left( X^T X \right)^{-1}X^T \epsilon \right)$$ so we need kill off the second term somehow. Clearly $E(\epsilon |X)=0$ would do it, as would just assuming that the second term has expectation zero, which is to say that the moore-penrose psuedoinverse of $X$ times $\epsilon$ has zero expectation.

• Your derivation seems to be claiming that $E \left( \left( X^T X \right)^{-1}X^T \epsilon \right)=E(\epsilon)$, which is not true. Commented Oct 29, 2015 at 7:06
• Not only is that assertion untrue, the conclusion is false as well. The weaker conditions do not imply the OLS estimator is unbiased. (In fact, additional conditions are needed just to make sure the OLS estimator is a well defined random variable! In particular, $X$ cannot be discrete.)
– whuber
Commented Oct 29, 2015 at 16:49
• @whuber: Interesting regarding the conditions for $\hat \beta$ to exist as a random variable. I would have guessed that the only condition would need to be that $X$ has full column rank almost surely. Do you have a reference? Commented Oct 29, 2015 at 17:26
• @ChristophHanck: yup! Good catch. See edits. Commented Oct 29, 2015 at 18:55
• @When $X$ is discrete, there is a positive chance that all the $X_i$ are identical, in which case the OLS estimator is undefined. You need to assume $X$ is continuous in order to avoid this possibility.
– whuber
Commented Oct 29, 2015 at 19:09