The Saga of PLS

by Gaston Sanchez

The Renaissance of Herman’s PLS

The PLS Regression of scandinavian chemetricians was the perfect match to the French multivariate analysis school also known as Analyse des Données: both favoring the same core concepts of projections, inductive style, interpretable tools, and graphical displays. With a long tradition of Multidmensional Analysis methods, French data analysts have had their preferences for tools with a heavy use of geometry, algebra, and projection methods. Rooted in the seminal work of Jean-Paul Benzécri during the mid-1960s, data analysis à la Française privileged data analysis with a decisive inclination to exploratory and descriptive approaches—similar to the spirit of John Tuckey’s Exploratory Data Analysis (EDA). One limitation of the French vision, however, was the lack of more explanatory and causal modeling techniques that matched the geometric spirit without privileging inferential procedures based on unrealistic assumptions about the data (like the “soft modeling” ideals of Herman Wold).

PLS Regression was a revelation to Michel Tenenhaus, and he was immediately caught up studying and researching the method. Soon after, he started to prepare a book on that topic. “During the writing of this book,” Michel remembers, “I had many interrogations on many points.” Invited by Svante Wold, Michel spent some time in Sweden to discuss various issues. Michel asked a number of questions about PLS which Svante answered and clarified. But not all them, “some formulas in SIMCA-P are still secret!” Michel admits. Among the copious bibliographic sources Svante gave Michel, there was a green hardcover book that Svante’s father, jointly with Karl Jöreskog, edited in 1982: Systems Under Indirect Observation, now considered one of the classical references about the PLS framework of Herman Wold.

Being a professor in the prestigious Business School HEC, Michel asked Svante about PLS applications in management. But Svante didn’t have much material around management or business applications, so he referred Michel to Swedish professor and consultant Clases Fornell, Business and Management professor at University of Michigan, and also an expert in Customer Satisfaction Measurement.

Customer Satisfaction

Claes Fornell had been introduced to PLS around the late 1970s by American Professor and morphometrician Fred Bookstein, while both were teaching at the University of Michigan. In turn, Fred Bookstein was an expert in morphometrics who had had the opportunity to learn about Herman Wold’s methods and even collaborate with him. In fact, a couple of geometrical interpretations of PLS-PM were provided by Bookstein (1980, 1982). Among one of the most cited PLS papers on marketing applications is the 1982 article “Exit Voice”, written by both Fornell and Bookstein in the Journal of Marketing Research. Claes found PLS very interesting and convenient, and he even edited a two volume book published under the title A Second Generation of Multivariate Analysis (Fornell, 1982). And he also taught those methods to students in the Business PhD Program at Michigan University.

Claes Fornell used the PLS Path Modeling approach for Customer Satisfaction, and he, like Svante, used it not only for academic research but also for consulting purposes. In this case, the method was employed by the firm Claes Fornell Group (CFG). The main application of PLS-PM by Claes Fornell was on Customer Satisfaction. Among his entrepeneurial research projects he developed the Swedish Customer Satisfaction Barometer in 1989. And then he started the project for developing the American Customer Satisfaction Index (ACSI) in 1994.

Michel Tenenhaus traveled to Michigan to meet with Claes Fornell who talked to him and presented him PLS applications in management. Claes showed Michel applications in Customer Satisfaction and talked about the PLS algorithm. But something didn’t click with Tenenhaus. Although Fornell was talking about Partial Least Squares, his version was not exactly the one used by the chemometricians. There was obviously something that Tenenhaus didn’t quite get. After all the research Michel had done on PLS regression, he was totally perplexed by the method Fornell was describing to him. According to Michel this was “a very unpleasant situation”. Had Svante forgotten to mention something else? Was Michel missing something important?

There was clearly something wrong. Under the same term “Partial Least Squares,” the version of the chemometricians was not at all the same version Claes Fornell had shown Michel. Strangely though, since both methods had the PLS label and Wold was the last name of the main author for both versions. Where did Tenenhaus’s confusion come from?

American Influence

Jointly with Claes Fornell, there was another leading expert of PLS-PM who was part of the very reduced group of American researchers making intensive use of the PLS approach: Wynne Chin, professor of Decision and Information Science.

In the early 1990s, Chin found himself in the middle of a crossroad. On its way to becoming a leading figure in the research field of Management Information Systems (MIS), Wynne Chin had also been developing the software program PLS-Graph, named after the little known—at the time—data modeling approach of Partial Least Squares Path Modeling.

Basically, research on the field of Management Information Systems is concerned with the problems related around management and usage of information technology and information systems: the use of Information Technology systems, how they are developed, used, and applied in organizations and industries. Among the quantitative tools for studying these problems, MIS researchers use a set of data modeling methods better known under the umbrella term of Structural Equation Models. Historically, this type of models have been borrowed, used and adopted from other disciplines, mainly from social and behavorial sciences rooted in psychometrics and causal modeling literature.

In his early 30s, Chin was in that singular spot that makes someone being a skilled master of some unique tradition, and at the same time being one of the few possessors of such knowledge. It looked like if Wynne was in a privileged power play position… except for the fact that almost no one else seemed to care much about that obscure PLS methodology. Trained in biophysics and chemical engineering, Wynne Chin enrolled to the MBA program at the University of Michigan in 1980, and then he continued with a PhD in Computers and Information Systems. He got his first contact with PLS-PM from the seminars taught by Claes Fornell. But more important, Chin intermittently worked over a period of ten years developing a series of computer programs that eventually would led to the final version of his PLS-Graph software. In the early 1980s he helped with the programming for a version on the then still in use main frame computers. Shortly after the introduction of the first IBM PCs, he also started to develop a version for MS-DOS.

Wynne Chin considerably helped spread PLS-PM among MIS researchers, and he even released for free his PLS-Graph, under the condition that it would only be used for academica purposes, and that it wouldn’t be extended or developed wihtout his approval—something that didn’t happened. Although the development of PLS-Graph has considerably slowed down, at the time of this writing, it is still available previous contact with Wynne Chin.

Before Chin’s software was available, the other complete computer program for PLS-PM was Lohmöller’s LVPLS. Distributed at nominal charge, LVPLS was not a commercial software but an academic one for research-purposes. This short supply of PLS software was a substantial difference compared to the software available for covariance-based Structural Equation Models. By that time the most popular software was LISREL, developed by Karl Jöreskog and Dag Sörbom. At the end the 1980s LISREL was already in its version VI with its microcomputer PC-based program. Version 7 was launched in 1989 integrated within the larger software SPSS, and it already had a graphic interface intended to make it ease for users.

Since none of the PLS versions had a graphical interface, Wynne Chin started to work on a software consistent with Lohmöller’s PC version around 1987. Although it was working on Windows 2.0 back then, the reliable version required Windows 3.0—mainly for Wynne Chin’s own use. According to Wynne, the first official presentation of his software was in 1988. However, the so-called PLS approach remained largely unknown eclipsed by the more commercially successful LISREL, widely available and converted in the gold-standard for SEM at that time. The PLS method was so unique that it suffered from a lack of interest in the best of the cases, and from multiple misundertandings in most occasions.

The first half of the 1990s decade also coincided with the passing away of Jan-Bernd Lohmöller and Herman Wold. “At that point, it was clear no one else was supporting the use of PLS,” Chin recalled. Watching how the method was fading away at an alarming rate, Chin made what he believed was the most appropriate decision at that time. As a last resort to keep alive the PLS approach among his colleagues, he decided to share PLS-Graph within the MIS community where he kept it “officially free for academic research” purposes.

The Path Modeling Renaissance

After the initial confusion Michel had with PLS Path Modeling, he soon realized about the attractive modeling possibilities the method provided. Greatly interested in this method, he began a long trip that would eventually rescued PLS-PM from its ostracism. In 1999 Michel Tenenhaus and Alain Morineau organized an international symposium in Paris bringing together users of both methodologies: PLS Regression as well as PLS Path Modeling. Given the produced enthusiasm and interest of the participants, it was decided to organize a PLS Symposium every two years dedicated to all PLS methods.

Likewise, the Customer Satisfaction movement of the late 1990s inspired European members to proposed a similar alternative to the American Customer Satisfaction Index, but this time adapted to the European geography and economy. A consortium of universities and researchers was formed and the ESIS project was born with the goal, among other purposes, to continue the development of PLS-PM software. This also sparked a fire of interest for the methodology, bringing a revival and mass difusion.

Next chapter