Bias and high-dimensional adjustment in observational studies of peer effects