MEASUREMENT ERROR IN PLS, REGRESSION AND CB-SEM

Partial Least Squares (PLS) has become a very popular statistical technique for analyzing causal path models involving multiple indicator data in MIS research. Proponents of the use of PLS view it as a relatively easy to use “second-generation” statistical analysis technique that, like covariance-based structural equation modeling (CB-SEM), has several advantages over regression. While it isn’t frequently explicitly stated, most users seem to believe that PLS somehow takes measurement indicator errors into account, perhaps similarly to the way CB-SEM techniques do. We conducted an extensive Monte Carlo simulation study to compare the accuracy of path estimates for regression, PLS and CBSEM (employing LISREL) under a variety of conditions. We found that PLS provides path estimates that are much closer to regression than to LISREL. Further, we observed that under most circumstances the amount of bias in the PLS (and regression) estimates could be corrected by applying Nunnally and Bernstein’s (xxxx) formula for correcting attenuation. We conclude that our results offer strong evidence that 1) PLS does not take indicator measurement error into account, and 2) under most circumstances the accuracy of PLS (or of regression) path estimates can be improved by applying Nunnally and Bernstein’s formula for correcting attenuation. There is also the suggestion that under certain circumstances, PLS may over-estimate path values.