Dirk Temme

Incorporating Latent Variables into Discrete Choice Models - A Simultaneous Estimation Approach Using SEM Software

Integrated choice and latent variable (ICLV) models represent a promising new class of models which merge classic choice models with the structural equation approach (SEM) for latent variables. Despite their conceptual appeal, applications of ICLV models in marketing remain rare. We extend previous ICLV applications by first estimating a multinomial choice model and, second, by estimating hierarchical relations between latent variables. An empirical study on travel mode choice clearly demonstrates the value of ICLV models to enhance the understanding of choice processes. In addition to the usually studied directly observable variables such as travel time, we show how abstract motivations such as power and hedonism as well as attitudes such as a desire for flexibility impact on travel mode choice. Furthermore, we show that it is possible to estimate such a complex ICLV model with the widely available structural equation modeling package Mplus. This finding is likely to encourage more widespread application of this appealing model class in the marketing field.

A Comparison of Current PLS Path Modeling Software: Features, Ease-of-Use, and Performance

After years of stagnancy, PLS path modeling has recently attracted renewed interest from applied researchers in marketing. At the same time, the availability of software alternatives to Lohmoller’s LVPLS package has considerably increased (PLS-Graph, PLS-GUI, SPAD-PLS, SmartPLS). To help the user to make an informed decision, the existing programs are reviewed with regard to requirements, methodological options, and ease-of-use; their strengths and weaknesses are identified. Furthermore, estimation results for different simulated data sets, each focusing on a specific issue (sign changes and bootstrapping, missing data, and multi-collinearity), are compared.

Gruppenvergleiche bei hypothetischen Konstrukten — Die Prüfung der Übereinstimmung von Messmodellen mit der Strukturgleichungsmethodik

ZusammenfassungBei einem Vergleich von Gruppen anhand hypothetischer Konstrukte muss gewährleistet sein, dass die Messmodelle in den Gruppen gleich sind. Andernfalls besteht die Gefahr, dass falsche Rückschlüsse über die Unterschiede auf der latenten Ebene gezogen werden. Dieser Beitrag bietet einen state-of-the-Art zur Überprüfung der Messinvarianz mit der Mehrgruppenanalyse konfirmatorischer Faktormodelle. Ein Schwerpunkt liegt dabei auf der Ermittlung nichtinvarianter Indikatoren, wobei unterschiedliche Ansätze zur Identifikation latenter skalen sowie die wesentlichen Teststrategien verglichen werden. Die empirische studie zur Invarianz eines Messmodells der psychologischen Markenstärke („Brand Potential Index“) zeigt, dass bei einem Vergleich loyaler und nichtloyaler Konsumenten die Indikatoren Kaufabsicht und Weiterempfehlungsbereitschaft nichtinvariant sind. Aufgrund der Messinvarianz der übrigen Indikatoren sind aber sowohl Vergleiche auf der latenten Ebene als auch direkte Mittelwertvergleiche für die invarianten Indikatoren zulässig.SummaryComparing groups (e.g., customer segments) with respect to hypothetical constructs requires that the measurement models are equivalent across groups. Otherwise, conclusions drawn from the observed indicators regarding differences at the latent level (mean differences, differences in the structural relations) might be severly distorted. However, with the exception of intercultural studies, at present researchers in business administration do not pay enough attention to this issue. This article provides a state of the art on how to apply multi-group confirmatory factor analysis to assess measurement invariance. The required steps in the simultaneous analysis of the indicator means and variances/covariances are described, placing special emphasis on how to determine noninvariant indicators (“partial measurement invariance”). Here, different approaches to the identification of latent scales as well as the main test strategies are discussed. The procedure is demonstrated considering the construct brand strength (“Brand Potential Index”, BPI®) introduced by GfK Market Research as an example. The empirical study shows that the indicators buying intention and intention to recommend the brand are noninvariant across loyal and non-loyal consumers. Since the remaining seven BPI® indicators are invariant, comparisons at the latent variable level as well as for the invariant indicators are meaningful. Observed group differences in buying intention overestimate the underlying group differences in perceived brand attractiveness. Possible statistical as well as substantial explanations for these findings are discussed.

Probleme der Validierung mit Strukturgleichungsmodellen

Dieser Beitrag setzt sich mit der Leistungsfahigkeit von Strukturgleichungsmodellen bei der Validitatsprufung von Messmodellen fur hypothetische Konstrukte auseinander und geht auf ausgewahlte Problembereiche bei der gangigen Anwendung dieser Methodik fur die Skalenkonstruktion ein. Insbesondere werden mit der Kontrolle verschiedener Arten von Methodeneffekten Alternativen zur Elimination von Indikatoren ausschlieslich auf Basis statistischer Kriterien (z. B. interne Konsistenz) aufgezeigt.

Die Spezifikation und Identifikation formativer Messmodelle der Marketingforschung in Kovarianzstrukturanalysen

Die Diskussion um die korrekte Spezifikation von Messmodellen im Marketing hat gezeigt, dass viele theoretische Konzepte formativer statt reflektierender Natur sind. Orientiert sich die Skalenentwicklung für diese Konstrukte an reflektierenden Messmodellen, so kann dies zu stark verzerrten Ergebnissen führen. Während für reflektierend gemessene Konstrukte die Kovarianzstrukturanalyse das Standardinstrument für die Schätzung darstellt, scheint sich der Partial-Least-SquaresAnsatz (PLS) als dominierendes Analysetool für Strukturgleichungsmodelle mit formativen Konstrukten zu etablieren. Ein wichtiger Treiber dieser Entwicklung ist vermutlich die weit verbreitete Auffassung, der „LISREL“-Ansatz sei nicht in der Lage, Modelle mit derartigen Konstrukten zu schätzen. Zwar werfen formative Messmodelle im Rahmen der Kovarianzstrukturanalyse ganz spezifische Identifikationsprobleme auf, diese lassen sich jedoch vielfach lösen. Aufgrund der Tatsache, dass der „LISREL“-Ansatz – trotz gewisser Nachteile gegenüber PLS – spezifische Stärken (z. B. bei der Modellbeurteilung) besitzt, sollten Forscher den Einsatz dieser Methodik auch bei Strukturgleichungsmodellen mit formativen Konstrukten in das Kalkül ziehen. Um sie bei der Beurteilung der Identifikation ihrer Modelle zu unterstützen, werden Identifikationsregeln für drei verschiedene Modelltypen behandelt, deren Einhaltung leicht anhand grafischer Kriterien überprüft werden kann. Eine empirische Anwendung demonstriert, dass der „LISREL“-Ansatz selbst dann nützlich sein kann, wenn nicht alle Parameter des ursprünglichen Modells identifiziert sind.

Assessing measurement invariance of ordinal indicators in cross-national research

Meaningful cross-national comparisons of scales require that the indicators used to operationalize the underlying constructs (e.g., attitudes, values) are measurement invariant across countries. Linear multi-group confirmatory factor (MGCF) analysis is arguably the most common method to assess measurement invariance. Although, strictly speaking, this method assumes continuous variables, in empirical studies typically a covariance matrix for ordinal items (e.g., Likert-type scales) is analyzed. Simulation studies have indeed shown that single-group confirmatory factor analysis is relatively robust against violating the assumption of continuous variables if categorization is based on at least five answer categories and the data does not show excessive skewness and/or kurtosis. New simulation evidence, however, has revealed that these results do not necessarily carry over to multiple groups. These insights and the availability of robust WLS estimators which are considerably less demanding with respect to the required sample size than the full WLS approach strongly advocate the use of appropriate estimation methods for ordinally scaled variables. This paper contributes to comparative cross-cultural research by proposing a procedure for testing measurement equivalence based on the MGCF model for ordinal indicators. The procedure is applied to a cross-national study on attitudes towards a specific advertisement.

A Monte Carlo study of structural equation models for finite mixtures

Empirical applications of structural equation modeling (SEM) typically rest on the assumption that the analysed sample is homogenous with respect to the underlying structural model or that homogenous subsamples have been formed based on a priori knowledge. However, researchers often are ignorant about the true causes of heterogeneity and thus risk to produce misleading results. Using a sequential procedure of cluster analysis in combination with multi-group SEM has been shown to be inappropriate to solve the problem of unobserved heterogeneity. Recently, two encouraging approaches have been developed in this regard: (1) Finite mixtures of structural equation models and (2) hierarchical Bayesian estimation. In this paper, we focus exclusively on the MECOSA approach to finite normal mixtures subject to conditional mean and covariance structures. Since not much is known about the performance of MECOSA, which is both a specific odel and a software, we present the results of an extensive Monte Carlo simulation. It was found that MECOSA performed best where homogenous groups were present in the data in equal proportions and in conjunction with rather large differences in parameters across the groups. MECOSA performed worse when the proportions were unequal and parameters were relatively close together across groups. Of the three estimation methods available in MECOSA the two-stage minimum distance estimation (MDE) in general performed worse than the alternative EM algorithms (EM and EMG). This effect was especially pronounced under conditions of close parameters and unequal group proportions. Above that, for these conditions the modified likelihood ratio test turned out to be inappropriate in the three groups case.

Drivers and impediments of consumer online information search: Self-controlled versus agent-based search in a high involvement context

INTRODUCTIONOne of the most enduring sequences of research effort in consumer behaviour has been thestudy of information search behaviour prior to purchase (Beatty and Smith, 1987; Moorthy etal., 1997; Punj and Staelin, 1983; Srinivasan and Ratchford, 1991). The goal has been toinvestigate drivers and impediments of the search activity as well as search intensity.Since we can observe an increasing use of the WWW as a source of product purchaseinformation, many of these older studies risk to become partially outdated. This is, becausethe new electronic medium promises to considerably reduce traditional search cost relevant inoffline markets (Alba et al., 1997; Bakos, 1997), offers an exciting amount of new productinformation sources and efficiently supports the search process through agent technology(Hauble and Trifts, 2000). Within a few years, the Internet has evolved as a major source ofproduct information retrieval and it is expected to play an even bigger role in the buyingprocess once 2

Choice Modeling and SEM

The paper shows how two dominant methodological approaches in quantitative marketing research of the last decades can be integrated. Combining the strengths of covariance structure analysis to control for measurement errors and the ability to predict choice behavior via the Multinomial Logit (MNL) model creates a powerful hybrid approach – the Integrated Choice and Latent variable (ICLV) model – for marketing research. We document the basic features of this approach and present an example which illustrates how the ICLV model can be used to explain travel mode choice. The hybrid modeling framework provides several advantages: (1) it gives a more realistic and comprehensive representation of the choice process taking place in the consumer’s “black box”; (2) it provides greater explanatory power; (3) it helps to remedy the biasing effect of neglecting important latent variables to explain choice behavior, thus allowing for a more accurate assessment of how marketing influences customers’ choice behavior.

Structural Equation Models for Finite Mixtures — Simulation Results and Empirical Applications

Unobserved heterogeneity is a serious but often neglected problem in structural equation modelling (SEM) challenging the validity of many empirical results. Recently, a finite mixture approach to SEM has been proposed to resolve this problem but until now only a few studies analyse the performance of the relevant software. The contribution of this paper is twofold: First, results from a Monte Carlo study into the properties of the program system MECOSA are presented. Second, an empirical application to data from a large-scale consumer survey in the fast moving consumer goods industry is described.