Partial Least Squares (PLS) has become an increasingly popular approach to testing research models with multiple proposed causality links. Moreover, recent interest in the specification of constructs in a formative manner has accentuated this tendency, given the purported ability of PLS to handle this methodological development. While a review of the literature reveals an extensive use of PLS in this capacity, there is neither theoretical nor empirical evidence supporting this property of the technique. An examination of the inner workings of PLS shows several limitations of PLS when used in 'formative' (Mode B) estimation, and compares it to linear regression and covariance-based approaches. Results from Monte Carlo simulations comparing the performance of PLS and covariance-based techniques in estimating models with formatively specified constructs in either exogenous or endogenous positions reveals important biases for PLS, but not for covariance-based SEM. The results are discussed and recommendations for researchers are proposed.