Instrumental variable estimators can be severely biased in finite samples when the degree of overidentification is high or when the instruments are weakly correlated with the endogenous regressors. This paper proposes an estimator based on the use of the principal components of the instruments as a means of dealing with these issues. By promoting parsimony, the proposed estimator can exhibit considerably lower bias, often without giving up asymptotic efficiency. To make the estimator operational, a simple but flexible rule to select the relevant components for estimation is suggested. Simulation evidence shows that this approach yields significant finite sample improvements over other instrumental variable estimators.