Modeling Formative Constructs in PLS

Validating Formative Partial Least Squares (PLS) Models

• The study examines formative constructs, differing from reflective constructs
• Formative constructs are under-researched in IS and Management, lacking methodological guidelines
• 29% of constructs are misspecified as reflective instead of formative
• Guidelines for formative constructs estimation are fragmented
• The paper provides a methodological review for identifying and evaluating formative models
• It outlines the theoretical differences between formative and reflective constructs
• A guideline for estimating formative measurement and structural models is included
• An illustration of a simplified model with three formative constructs is presented, using SmartPLS
• Key terms include Partial Least Squares (PLS), formative constructs, pure formative PLS model, validation, BVIT, and SmartPLS

Abstract

• The issue of formative constructs, as opposed to reflective ones, is gaining traction among IS and Management researchers
• Researchers generally believe formative constructs have been understudied and that methodological literature is lacking to guide researchers on how such constructs should be developed and estimated
• A survey of IS research revealed that approximately 29% of constructs were misspecified as reflective rather than formative
• The paper aims to guide researchers through formative model identification and evaluation, outlining theoretical constructs and providing estimation guidelines
• A simplified model with three formative constructs is included for illustration using SmartPLS

Introduction

• The issue of formative constructs gains momentum among IS and Management researchers
• Methodological literature is lacking for guiding researchers on how to develop and estimate formative constructs
• Previous reviews indicate that many constructs were misspecified as reflective, instead of formative
• The paper provides a review of formative model identification and evaluation, focusing on formative construct assessment
• The paper includes a methodological review of formative, PLS-based, model identification and evaluation

Formative vs. Reflective Constructs

• Formative constructs are caused by measures
• Reflective constructs are manifestations of the construct
• Formative items should not highly correlate, as the items are distinct
• Formative indicators should reflect different aspects of the construct and their antecedents and consequences may not be the same
• In formative measures, changing a measure may only impact that measure in the construct, not all measures simultaneously and conversely
• Researchers often make mistakes in identifying reflective rather than formative latent constructs

PLS in SEM

• Two types of Structural Equation Modeling (SEM) exist: covariance-based (CovSEM) and component-based (PLS)
• Covariance-based SEM is traditionally used and involves estimating model parameters to minimize the difference between calculated and observed covariance matrices
• PLS is a distribution-free approach
• PLS models are two-step approaches containing a measurement and structural model
• PLS is simpler and assumes lesser requirements on the measurement scale and sample size
• PLS is more heuristically used for exploratory research

Identifying Formative or Reflective Constructs

• Criteria outlined for identifying formative or reflective constructs, based on previous research
• The causality flows from indicators to the construct for a formative construct
• Formative constructs do not require highly correlated items (compared to those reflective), however, correlations between items in a construct matter
• Formative measures do not have to correlate
• An absence of interchangeability or common theme between items points toward a formative construct
• Antecedents and consequences of measures do not necessarily have to be the same for a formative construct

PLS Formative Model Estimation

• The paper focuses on the use of PLS for formative models, distinguishing the model into measurement (outer) and structural (inner) models
• The measurement model (outer) defines relations between measurement variables and latent variables
• The structural model (inner) defines relations between latent variables

Evaluation of the Measurement Model

• Model evaluation includes content validity, construct reliability, and construct validity
• Content validity ensures the model's indicators encompass all aspects of a given construct
• Construct reliability tests the internal consistency of the measurement model, measuring multicollinearity
• Construct validity tests if indicators accurately measure the intentions of the model, testing discriminant and convergent validity, and possible external and/or nomological
• PLS method measures, like standard deviation, correlation, t-test or bootstrapping to assess validity

Evaluation of the Structural Model

• Explains variance of dependent construct (R-square) and examines effect sizes
• Model's ability to predict is tested by the Stone-Geisser Q² test using blindfolding procedures
• The paper utilizes explanatory power, examining R-square and effect sizes
• Predictive power of the model is determined using path coefficients and re-sampling techniques (e.g., bootstrapping)
• These techniques assess if the model significantly impacts the dependent variable (evaluating if constructs make predictions)

Multi-Group Analysis

• Investigates potential differences across two or more different groups affecting parameter estimates
• This procedure is used to test for potential differences of impacts across different groups, such as IT or business managers.

Discussion and Conclusions

• The paper synthesizes existing knowledge on formative constructs, offering a practical and illustrative empirical approach
• Limitations are acknowledged, primarily the omission of the disturbance/error term with a suggestion for future work
• The illustration, composed of three constructs, shows validity due to following the formative construct guidelines
• The paper promotes formative constructs as appropriate in certain situations and promotes more accurate evaluation/guidelines