Information systems researchers have recently begun to propose models that include formatively specified constructs, and largely rely on partial least squares (PLS) to estimate the parameters of interest in those models. In this research, we focus on those cases where the formatively specified constructs are endogenous to other constructs in the research model in addition to their own manifest indicators, which are quite common in published research in the discipline, and analyze whether PLS is a valid statistical technique for estimating those models. Although there is evidence that covariance-based approaches can accurately estimate them, this is the first research that examines whether PLS can indeed do so. Through a theoretical analysis based on the inner workings of the PLS algorithm, which is later validated and extended through a series of Monte Carlo simulations, we conclude that is not the case. Specifically, estimates obtained from PLS are capturing something other than the relationship of interest when the formatively specified constructs are endogenous to others in the model. We show how our results apply more generally to a class of models, and discuss implications for future research practice.
Researchers in a number of disciplines, including Information Systems, have argued that much of past research may have incorrectly specified the relationship between latent variables and indicators as reflective when an understanding of a construct and its measures indicates that a formative specification would have been warranted. Coupled with the posited severe biasing effects of construct misspecification on structural parameters, these two assertions would lead to concluding that an important portion of our literature is largely invalid. While we do not delve into the issue of when one specification should be employed over another, our work here contends that construct misspecification, but with a particular exception, does not lead to severely biased estimates. We argue, and show through extensive simulations, that a lack of attention to the metric in which relationships are expressed is responsible for the current belief in the negative effects of misspecification.