Model Misspecification
Choosing the wrong specification: the wrong functional form, or the wrong set of variables. The lecture's two variable-choice errors are adding an irrelevant variable (coefficients stay unbiased, only precision suffers) and omitting a relevant correlated variable (bias).
When to use
Spot an irrelevant variable by a failed t-test and drop it (after checking it is not just collinear); fix an omission, the more dangerous error, by adding the variable back. When unsure, keep a borderline variable, since bias is worse than extra variance.
See Lecture 9 and Lecture 10 and Lecture 13.