Published online ahead of print May 12, 2009 Information Systems Research informs ® Articles in Advance, pp. 1–27 y doi 10.1287/isre.1080.0224 issn 1047-7047 eissn 1526-5536 © 2009 INFORMS .org. ma ms file or The . Research Note ibers Toward a Causal Interpretation from Observational missions@inf subscr per Data: A New Bayesian Networks Method for to Structural Models with Latent Variables policy institutional this Zhiqiang (Eric) Zheng to School of Management, University of Texas at Dallas, Richardson, Texas 75083, le [email protected] Paul A. Pavlou ailab v Fox School of Business and Management, Temple University, Philadelphia, Pennsylvania 19122, regarding a [email protected] made ecause a fundamental attribute of a good theory is causality, the Information Systems (IS) literature has is strived to infer causality from empirical data, typically seeking causal interpretations from longitudinal, questions B experimental, and panel data that include time precedence. However, such data are not always obtainable and y observational (cross-sectional, nonexperimental) data are often the only data available. To infer causality from an which observational data that are common in empirical IS research, this study develops a new data analysis method that integrates the Bayesian networks (BN) and structural equation modeling (SEM) literatures. send Similar to SEM techniques (e.g., LISREL and PLS), the proposed Bayesian Networks for Latent Variables ersion, (BN-LV) method tests both the measurement model and the structural model. The method operates in two v stages: First, it inductively identifies the most likely LVs from measurement items without prespecifying a Please measurement model. Second, it compares all the possible structural models among the identified LVs in an . ance exploratory (automated) fashion and it discovers the most likely causal structure. By exploring the causal struc- site tural model that is not restricted to linear relationships, BN-LV contributes to the empirical IS literature by Adv overcoming three SEM limitations (Lee et al. 1997)—lack of causality inference, restrictive model structure, and in lack of nonlinearities. Moreover, BN-LV extends the BN literature by (1) overcoming the problem of latent vari- able identification using observed (raw) measurement items as the only inputs, and (2) enabling the use of author’s ticles ordinal and discrete (Likert-type) data, which are commonly used in empirical IS studies. Ar the The BN-LV method is first illustrated and tested with actual empirical data to demonstrate how it can help reconcile competing hypotheses in terms of the direction of causality in a structural model. Second, we conduct this a comprehensive simulation study to demonstrate the effectiveness of BN-LV compared to existing techniques to in the SEM and BN literatures. The advantages of BN-LV in terms of measurement model construction and including structural model discovery are discussed. ight , yr Key words: causality; Bayesian networks; structural equation modeling; observational data; Bayesian graphs History: Vallabh Sambamurthy, Senior Editor and Associate Editor. This paper was received on October 25, cop ebsite 2006, and was with the authors 20 months for 2 revisions. Published online in Articles in Advance. w holds other y 1. Introduction the requisite attention. Similar to most other disci- an Because a fundamental attribute of a good theory plines (e.g., Mitchell and James 2001, Shugan 2007), the on IS discipline tends to avoid issues of causality because INFORMS is causality (Bagozzi 1980), causality inference (X causes Y ) is deemed invaluable in the social and of the difficulty in inferring causal relationships from posted behavioral sciences in general and information sys- data, and because causality is only inferred from pure yright: be tems (IS) research in particular. However, despite the theory. This is partly because of the fact that causality enhanced sophistication of IS studies in terms of the- inference requires strict conditions. Though there is no Cop not ory and empirical testing, causality has not received consensus on the necessary and sufficient conditions 1 Zheng and Pavlou: New Bayesian Networks Method for Structural Models with Latent Variables 2 Information Systems Research, Articles in Advance, pp. 1–27, © 2009 INFORMS y .org. ma for inferring causality, Popper’s (1959) three condi- (3) help spawn future research in refining SEM-based ms file or tions for inferring causality are generally accepted: methods that render causal interpretations. The (1) X precedes Y ; (2) X and Y are related; and (3) no According to Lee et al. (1997), SEM methods have . confounding factors explain the X → Y relationship. three key limitations: lack of causality inference, To satisfy these strict conditions, researchers need to restrictive model structure, and lack of nonlinearities. ibers use longitudinal, experimental, or panel data with First, though SEM was originally designed to model missions@inf time precedence between variables X and Y to account causal relationships, causality has gradually faded subscr per for confounds and reverse causality (Allison 2005).1 away from SEM studies (Pearl 2000). In fact, a review to However, it is often impossible to obtain such data of the literature suggests that SEM studies do not in IS research (Mithas et al. 2006, p. 223) and obser- attempt to infer causality, and most SEM (Cartwright policy vational (cross-sectional, nonexperimental) data are 1995) and IS researchers (Gefen et al. 2000) believe institutional that SEM models cannot infer causality.3 The inability this often the only data available. Therefore, our objective to is to develop a method to help infer causality using for causal inference has forced IS researchers to refrain le observational data that are commonly used in empiri- from even discussing issues of causality in IS studies. ailab cal IS research. Second, most SEM studies specify one model struc- v regarding a Following the literature that maintains that “near” ture and use data to confirm or disconfirm this spe- 4 (versus “absolute”) causality inference is possible from cific structure by operating in a confirmatory mode. made observational data (e.g., Granger 1986, Holland 1986), This prevents the automated exploration of alterna- is tive or equivalent models. Chin (1998) warns that questions we develop a new data analysis method built on the y Bayesian networks (BN) and structural equation mod- overlooking equivalent models is common in SEM an studies, and Breckler (1990) showed that only 1 of 72 which eling (SEM) literature that offers a causal interpreta- tion to relationships among latent variables (LVs) in published SEM studies even suggests the possibility send structural equation models.2 Our proposed method of alternative models. Third, SEM only encodes linear ersion, relationships among constructs, essentially ignoring v (termed BN-LV—Bayesian networks for latent vari- the possibility of nonlinear relationships.5 Please ables) encodes the relationships among LVs in a graph- . ance ical model as conditional probabilities, it accounts To address these three SEM limitations, we devel- site oped the BN-LV method, which has three key Adv for potential confounds, and it discovers the most properties: First, it encodes the relationships among in likely causal structure from observational data. The proposed BN-LV method seeks to (1) sensitize IS constructs as conditional probabilities that, according to author’s Druzdzel and Simon (1993), can offer a causal interpre- ticles researchers about the importance of causality and Ar tation (as opposed to SEM, which uses correlation that the present the possibility to infer causal relationships from data, (2) offer a method to IS researchers to help this 3 SEM techniques are primarily based on linear equations (e.g., PLS) to infer causality among constructs from observational or covariance structures (e.g., LISREL). Additional conditions such data while overcoming key SEM limitations, and including as isolation of competing hypotheses (Cook and Campbell 1979) or ight , temporal ordering (Bollen 1989) are deemed necessary. yr 1 Even with longitudinal data that have time precedence, it is not 4 To the best of our knowledge, no SEM techniques allow researchers cop ebsite readily known which variable precedes which. It is often impos- to automate the process of examining alternative models. Manual w sible to know when a person formed certain perceptions (e.g., examination of alternative models becomes virtually impossible for holds perceived usefulness and perceived ease of use), even if these vari- complex models with multiple constructs. other ables are measured in different periods. Thus, even data with time 5 There have been attempts to incorporate interaction and nonlinear y precedence may not correspond to the actual timing of a person’s effects in SEM (e.g., Kenny and Judd 1984). However, as acknowl- an perceptions. edged by Kenny and Judd (1984, p. 209), their method is prelim- on 2 From a theoretical point of view, because the proposed BN-LV inary because it only deals with a single nonlinearity (quadratic INFORMS method uses observational, cross-sectional data, it only addresses function). Existing approaches need to prespecify the exact form of two of Popper’s (1959) three conditions for inferring causality, the nonlinear relationship. A general approach with unknown non- posted excluding the condition that X must precede Y . Therefore, it is not linearities still remains