Bias
Also known as: Estimator bias, Statistical bias
Bias of an estimator $W$ for $\theta$ is $\mathrm{Bias}\theta(W)=E\theta[W]-\theta$. Unbiased means $E_\theta[W]=\theta$ for all $\theta$ in the model.
MSE $=\mathrm{Var}+\mathrm{Bias}^2$, so a little bias can buy lower variance. Normal $\hat\sigma^2_{\mathrm{MLE}}$ has bias $-\sigma^2/n$; bias-corrected $S^2$ is unbiased. For OLS slopes, zero-mean errors and exogeneity ($\mathrm{Cov}(X,\varepsilon)=0$) deliver unbiasedness.