Discovering unobserved heterogeneity in structural equation models to avert validity threats
Summary (9 min read)
Introduction
- Assuming that data in empirical studies are homogeneous and represent a single population is often unrealistic in the social and behavioral sciences, such as information systems, management, and marketing (Rust and Verhoef 2005; Wedel and Kamakura 2000).
- In contrast, inexperienced users show a strong positive relationship between PEOU and IU and a weak, or nonsignificant, relationship between PU and IU .
- (1) While IS studies now routinely address observed heterogeneity by introducing moderators, a priori groupings, contextual factors, and control variables in their research models, they have not considered unobserved heterogeneity in their data.
- Only recently has a method been proposed to detect unobserved heterogeneity in PLS path models: finite mixture partial least squares (FIMIX-PLS;.
Concept of Heterogeneity and its Treatment in IS Research
- Researchers can obtain different parameter estimates when they consider differences among observations relative to when they overlook them.
- Heterogeneity among observations is not necessarily captured by variables that are preconceived by the researcher and specified by existing theory, as it can exist beyond these previously identified variables (Jedidi et al. 1997).
- As a consequence, it is necessary to differentiate between the following two types of heterogeneity: (1) observed heterogeneity when subpopulations are defined a priori based on known variables and (2) unobserved heterogeneity when the subpopulations in the data are unknown (Lubke and Muthén 2005).
Treatment of Heterogeneity in IS Research
- Given the complexity of the social and behavioral phenomena tackled in IS research, heterogeneity is likely to exist in samples that are used to develop, test, and refine models.
- If this heterogeneity is not uncovered and controlled, the heterogeneity can bias results and conclusions (e.g., Ansari et al. 2000; Johns 2006).
- Unobserved heterogeneity is receiving increasing attention in related disciplines (e.g., marketing, where scholars study similar complex phenomena pertaining to consumer choices and preferences, the alignment of firm-level marketing strategies, interorganizational relationships, and the business value of tangible and intangible resources) to safeguard against biases and probe the underlying reasons for unobserved heterogeneity (e.g., Rigdon et al. 2010).
- This enhances the likelihood of obtaining valid results as well as of generating greater theoretical contributions.
- Methodologists in marketing, econometrics, and psychology have proposed advances to uncover unobserved heterogeneity in various approaches—for instance, regression analysis (DeSarbo and Cron 1988; Späth.
Unobserved Heterogeneity in the Structural Model
- In the context of SEM, heterogeneity can affect the structural model, the measurement model (formative and reflective), or both (e.g., Ansari et al. 2000; Qureshi and Compeau 2009).
- 2 (2) If a researcher overlooks unobserved heterogeneity and obtains a nonsignificant relationship between the constructs as the overall sample estimate, this researcher may overgeneralize the nonsignificant finding, which exists only in the second segment, thereby leading to a Type II error with respect to the first segment.
- The R² decreases in all situations, implying an 2This does not mean that there will be a Type I error in general (i.e., for both segments) but only with respect to segment 2 where the true effect is zero.
Unobserved Heterogeneity in the Measurement Model
- Measurement model specification requires the consideration of the nature of the relationship between constructs and measures.
- The direction of causality is from the indicators to the construct and the weights of formative indicators represent the importance of each indicator in explaining the variance of the construct (Edwards and Lambert 2007; Petter et al.
- In contrast, when a construct’s measures are formative, unobserved heterogeneity can lead to differences in the formative indicators’ weights across groups.
- As a consequence, the interpretation of the weights estimated using the overall sample is misleading, and the formative measures based on the overall sample represent neither the first group nor the second.
- Given this bias in the formative measures for service quality, the relationship between service quality and other constructs (e.g., customer satisfaction) is also likely to be biased.
Implications of Unobserved Heterogeneity for Model Validity
- If unobserved heterogeneity characterizes the data and results are based on the overall sample, the estimated model lacks validity because it will not uncover the true effects of the underlying groups.
- Unobserved heterogeneity is a threat to internal validity because contextual or group variables that affect results are overlooked, thereby resulting in an incomplete model.
- In addition, unobserved heterogeneity threatens statistical conclusion validity.
- Experienced users might have a different understanding of a system’s usefulness compared to inexperienced users, thereby leading to different correlation patterns for the PU construct’s indicators.
- Both threats are regularly addressed when testing for ME/I in multigroup models (i.e., observed heterogeneity) (see Steenkamp and Baumgartner 1998; Vandenberg and Lance.
In contrast, unobserved heterogeneity typically does not affect
- Content validity because the constructs’ measures are normally the same across groups and are grounded in theory.
- An increase in the value of a formative measure’s error term due to unobserved heterogeneity can lead to misinterpretations, as a high error term is typically associated with the construct measure’s incompleteness (Diamantopoulos et al. 2008).
- Finally, if unobserved heterogeneity is not uncovered, there is a threat to external validity (i.e., the ability to generalize findings beyond the current population and context) because the overall sample results are not representative of the underlying groups.
- Next, the authors present an overview of methods to uncover unobserved heterogeneity in structural equation models that researchers can apply to overcome threats to validity due to unobserved heterogeneity.
Uncovering Heterogeneity in Structural Equation Models
- The authors first synthesize and compare different methods in SEM (i.e., CB-SEM and PLS path modeling) to uncover observed and unobserved heterogeneity.
- Given the objectives of their paper, the authors focus primarily on methods in SEM to uncover unobserved heterogeneity.
Existing Methods to Uncover Observed Heterogeneity in SEM
- SEM methods to address observed heterogeneity are now commonly applied in the social and behavioral sciences, including information systems.
- The first category of methods identifies homogenous groups of observations (e.g., individuals) a priori based on grouping variables (e.g., psychographic or socio-demographic).
- The second category of methods aims at identifying moderating factors that explain heterogeneity in specific structural model relationships.
- Uncovering observed heterogeneity with both types of methods requires a priori knowledge about differences across groups.
CB-SEM Methods to Uncover Unobserved Heterogeneity
- In CB-SEM, the following two primary methods have been developed to uncover unobserved heterogeneity: (1) finite mixture models that extend multigroup CB-SEM (Arminger et al.
- Several applications and simulation studies (e.g., Arminger et al.
- In contrast to finite mixture models, hierarchical Bayesian models for CB-SEM, which were developed by Ansari et al. (2000), do not assume heterogeneity among a defined number of groups of individuals but estimate unobserved heterogeneity at the individual4 level using a random coefficients model.
2003) to extend the method to dichotomous variables and
- Missing data and evaluate the performance of these methods.
- While both the finite mixture and the hierarchical Bayesian CB-SEM models have been the subject of extensive methodological research, finite mixture models have been applied in empirical CB-SEM research to a greater extent.
- An increasing number of applications, especially in the marketing, econometrics, and sociology literatures, have utilized finite mixture models to uncover unobserved heterogeneity, thereby improving theoretical and practical implications (e.g., Bart et al.
PLS Path Modeling Methods to Uncover Unobserved Heterogeneity
- Multiple PLS segmentation methods have been proposed.
- The authors draw on Sarstedt’s (2008) review of these methods to identify the following key PLS segmentation methods:.
1. The PATHMOX (path modeling segmentation tree) algorithm (Sánchez 2009; Sánchez and Aluja 2006).5
- This algorithm requires the a priori specification of explanatory variables that are not used as indicators in the PLS path model to discover segments.
- It captures heterogeneity by estimating the probabilities of segment memberships for each observation in order to optimize the likelihood function.
- While FIMIX-PLS is generally applicable to PLS path models regardless of whether the latent variables are measured reflectively or formatively, it does not account for the heterogeneity in the measurement models.
- First, based on an assessment of the benefits and limitations of these methods, Sarstedt (2008, p. 152) concludes: “To sum up, FIMIX-PLS can presently be viewed as the most comprehensive and commonly used approach to capture heterogeneity in PLS path modeling.”.
Partial Least Squares–Prediction-Oriented Segmentation (PLS-POS)
- The latter feature of PLS-POS ensures continuous improvement of the objective criterion throughout the iterations of the algorithm (hill-climbing approach) and provides the ability to uncover very small niche segments.
- The segmentation objective in a PLS path model is to form homogenous groups of observations with increased predictive power (R² of the endogenous latent variables) of the group-specific path model estimates (compared to the overall sample model).
- The extension will be made available with the next release of SmartPLS.
- In the next section, the authors detail the comprehensive simulation experiments they conducted to evaluate whether the differences in the capabilities of FIMIX-PLS and PLS-POS noted in Table 4 hold empirically.
Simulations of PLS-POS and FIMIX-PLS Performance
- The authors conducted experiments with simulated data that define the true group-specific PLS parameters a priori.
- The authors assessed the performance of PLS-POS and FIMIX-PLS based on the differences between the true parameters and those estimated by each method.
- Subsequently, the authors compared the performance of PLS-POS and FIMIX-PLS in recovering the true parameter estimates.
Model Specification
- Consistent with most simulation studies on PLS path models (e.g., Chin et al. 2003), the authors specified a direct effects path model that includes four exogenous latent variables and one endogenous variable.
- The authors specified two versions of the path model: model 1 uses reflective measures for the exogenous and endogenous latent variables , while model 2 uses formative measures for the exogenous latent variables and reflective measures for the endogenous latent variables .
- The authors results were generally stable for these more complex models as well.
- Conversely, for group 2, p4 has a high true parameter value, while the path coefficients p1 to p3 have lower true values.
- For group 1, the measurement weights w1 and w3 have high true values, while weights w2 and w4 have low true values.
Factor Design of the Simulations
- The authors selection of experimental factors and their levels was informed by criteria that were shown to influence PLS path modeling or segmentation results in prior simulation studies.
- When the R² value in both groups is .85, the overall sample that combines the two groups has a R² value of .425 because of unobserved heterogeneity.
- (7) Measurement model heterogeneity—that is, the groupspecific differences in formative measurement weights (.25, .50, .75).
- The number of factors and the number of factor levels systematically increase the complexity of the PLS segmentation task.
Data Generation
- Simulation studies in PLS path modeling require that data generated for the indicators (manifest variables) match the true values of the model.
- Previous studies on PLS path modeling (e.g., Chin et al.
- 12The unequal condition has one segment with 80% and one with 20% of the total sample size.
- Becker et al./Discovering Unobserved Heterogeneity in SEM in the measurement model.
- Data for the formative indicators must first be generated to compute the latent variable scores for formative constructs.
Performance Assessment
- The objectives of their simulation experiments were to (1) assess PLS-POS and FIMIX-PLS in terms of their respective abilities to recover true group-specific parameters, (2) compare PLS-POS and FIMIX-PLS based on the assessment of their parameter recovery, and (3) identify the relative effects of the design factors on the parameter recovery of PLS-POS and FIMIX-PLS.
- The authors knew the true parameters of each factorial combination (i.e., the R², path coefficients, outer weights, and loadings) a priori based on the parameter settings for the data generation.
- The smaller the differences between the true values and the segmentation method’s parameter estimates, the better the parameter recovery.
- MAB values close to zero indicate near perfect parameter recovery.
- Finally, to understand the relative importance of the design factors, the authors evaluated parameter recovery (i.e., the path coefficient’s MAB) using a mixed-effects ANOVA model with the two segmentation methods (PLS-POS and FIMIX-PLS; within-subjects factor) and the eight design factors (between-subjects factors).
Results for Model 1: Reflective Measures
- Table 5 presents the results for the ANOVA with MAB as the dependent variable.
- The direct effect and all of the interaction effects of reliability are nonsignificant.
- The between-subjects effects identify the factors that influenced MAB for both segmentation methods.
- In general, the method has a significant and substantial impact on the parameter recovery for the reflective model.
WithinSubjects
- Df = degrees of freedom well as between measurement model heterogeneity and multicollinearity are significant and substantial but have very little impact compared to the factors discussed earlier, also known as Note.
- For the within-subjects effects, the method’s effect on MAB is significant and substantial.
- The method also significantly and substantially interacts with heterogeneity in both the structural model and the measurement model.
- While the MAB for PLS-POS is always below .05, thereby indicating good parameter recovery, the MAB for FIMIX-PLS increases when measurement model heterogeneity becomes higher and structural model heterogeneity becomes lower.
- Becker et al./Discovering Unobserved Heterogeneity in SEM.
Results for Model 2: Formative Measures
- Table 7 presents the results for the ANOVA in model 2 (formative measures) with MAB as the dependent variable.
- Again, for the sake of space and simplicity, Table 7 presents the direct effects, all two-way interactions with the method, and all other interactions that have significant and substantial effects (partial η² of more than .02).
- For the between-subjects effects, all of the direct effects on MAB are significant, but again, the effect of relative segment size (partial η² = .012) on MAB is not substantial.
- Moreover, the interaction effects between the structural model and measurement model heterogeneity as 680 MIS Quarterly Vol. 37 No. 3/September 2013.
- Becker et al./Discovering Unobserved Heterogeneity in SEM.
Segmentation
- Method Desired Criteria for a PLS Segmentation Method Ability to detect heterogeneity in reflective measures Ability to detect heterogeneity in formative measures Ability to detect heterogeneity in the structural model Maximizes group-specific R² of endogenous latent variables (prediction orientation).
- Ability to handle nonnormal data FIMIX-PLS Hahn et al. situations with very high structural model heterogeneity regardless of the measurement model heterogeneity and also in situations where the measurement model heterogeneity is low and the structural model heterogeneity is at moderate levels.
- Therefore, as the results in Figures 4a and 4b reveal, the parameter recovery ability of a segmentation method cannot be assessed independently for these two types of heterogeneity.
- It is worth noting that the interaction effect between method and data distribution is not substantial for either model 1 (reflective measures) or model 2 (formative measures).
- This might explain this initially surprising result.
Summary of Results
- Overall, the authors can conclude that the use of either PLS-POS or FIMIX-PLS is better for reducing biases in parameter estimates and avoiding inferential errors than ignoring unobserved heterogeneity in PLS path models.
- A notable exception is when there is low structural model heterogeneity and high formative measurement model heterogeneity; in this condition, FIMIX-PLS produces results that are even more biased than those resulting from ignoring heterogeneity and estimating the model at the overall sample level.
- FIMIX-PLS becomes more effective when there is high multicollinearity in the formative measures, while PLSPOS consistently performs well.
- The strongly correlated formative measures become closer to a homogenous reflective measurement of the construct.
- Thus, the simulation experiments provide an empirical assessment of the segmentation criteria associated with PLS-POS and FIMIX-PLS (Table 9).
A Process for Unobserved Heterogeneity Discovery
- Given the availability of methods to uncover unobserved heterogeneity, as discussed in the two previous sections, researchers working with SEM face the following two major questions: when to investigate unobserved heterogeneity and 684 MIS Quarterly Vol. 37 No. 3/September 2013.
- Becker et al./Discovering Unobserved Heterogeneity in SEM how to apply methods for uncovering unobserved heterogeneity and defining segments.
- The authors address these questions by proposing a UHD process and also by identifying how this process can be applied given the research objective (i.e., purely testing a model or testing and elaborating a model; Colquitt and Zapata-Phelan 2007).
How to Apply the UHD Process
- When selecting an appropriate UHD method, researchers have to determine whether they are interested in evaluating unobserved heterogeneity associated with latent segments or individual-level estimates (e.g., hierarchical Bayesian approach, fixed effects, and random effects).
- In contrast, if the objective is to examine unobserved heterogeneity for individual-level estimates, the described UHD process does not apply because the methods have different assumptions and objectives and require different data (i.e., several observations per individual).
- The UHD process for the discovery of latent segments consists of the following three stages:.
Selecting an Appropriate UHD Method (Stage 1 of the UHD Process)
- As discussed earlier, the methodological options for analyzing unobserved heterogeneity involving CB-SEM cover two conceptually different approaches (i.e., latent segment analysis and individual-level estimate correction).
- For latent segment analysis, the appropriate UHD choice is the finite mixture model as no model-based clustering alternative is available.
- For analyses involving PLS path modeling, there are no methods available that address unobserved heterogeneity associated with individual-level estimates.
- Latent segments in PLS path modeling can be uncovered using one of the two methods the authors present in this paper (i.e., FIMIX-PLS and PLSPOS).
- Therefore, researchers should choose FIMIX-PLS if their models include only reflective measures and heterogeneity is expected to affect only the structural model and not the measurement model.
Applying the UHD Method to Define Segments (Stage 2 of the UHD Process)
- After choosing the appropriate method for uncovering unobserved heterogeneity, the researcher has to apply the method to evaluate whether significant unobserved heterogeneity is present in the model and to define the number of segments to retain from the data.
- Determining the correct number of segments is important as under- or over-segmentation leads to biased results and misinterpretations.
- The second stage of the UHD process focuses on (1) defining with heuristics a range of statistically well-fitting segments and (2) evaluating the segments based on theoretical considerations.
- The steps in this stage emphasize that researchers (1) evaluate the plausibility of segments by connecting the segmentation solution to theory and (2) avoid capitalizing on data idiosyncrasies to improve the explained variance or significance of parameters.
- In mixture models, these heuristics include model-selection criteria that are well known from the model-selection literature (e.g., AIC, BIC, and CAIC) and can also be used to approximate the best fitting number of segments (Andrews and Currim 2003a; Sarstedt et al. 2011a).
In contrast, model-based clustering methods, such as PLSPOS, are not based on the mixture model concept and do not
- Becker et al./Discovering Unobserved Heterogeneity in SEM provide model-selection criteria.
- Researchers should not rely purely on heuristics (e.g., model-selection criteria in finite-mixture modeling or the explained variance per segment in PLS-POS) to retain the best fitting number of segments because past studies have shown heuristics to have a low probability of finding the true number of segments.
Validating the Segmentation Results (Stage 3 of the UHD Process)
- In the final stage of the UHD process, researchers should validate the segmentation results, including the number of segments, with external data not used in the estimation process.
- Becker et al./Discovering Unobserved Heterogeneity in SEM random splits to compare the stability of segmentation results (Jedidi et al. 1997), or (3) collect additional data (e.g., in a follow-up study) to evaluate the results and find new explanatory variables that match segments better to explain heterogeneity (i.e., make them accessible).
- Furthermore, repeating the segmentation study on a different population (i.e., sample) and testing the proposed explanatory variables (i.e., moderators or grouping variables) in follow-up studies increases the generalizability of the results.
When to Apply Methods to Uncover Unobserved Heterogeneity
- Given a model that is grounded in substantive theory, the complexity of the social and behavioral phenomena examined in IS research makes it plausible there will be heterogeneity in any sample that is used to test and refine the model.
- Hence, theory testers apply the UHD process to evaluate validity threats due to unobserved heterogeneity.
- If the research objective is theory testing and elaboration (i.e., expanders; Colquitt and Zapata-Phelan 2007), uncovering unobserved heterogeneity not only serves as a validity check but can also guide researchers to identify variables explaining the uncovered segments and to integrate these variables to expand the model/theory.
- Becker et al./Discovering Unobserved Heterogeneity in SEM contrast, the other segment shows a strong positive relationship between PEOU and IU and a weak, or nonsignificant, relationship between PU and IU .
- The researcher concludes that these two identified segments (i.e., users emphasizing PU or PEOU) are theoretically plausible (i.e., within TAM, it is reasonable that there are different users who emphasize different system characteristics) and conceptually important for the theory.
Limitations and Future Research
- While their study makes contributions, it has its limitations and opens up avenues for future research.
- First, the validity and generalizability of simulation studies are limited by the choice of design factors and factor levels.
- The analysis of all factorlevel combinations of the two PLS path models entailed 126,720 simulated segmentation runs for assessing the performance of PLS-POS and FIMIX-PLS.
In addition, researchers may want to assess the effect of unobserved heterogeneity in models that do not comply with the recursive nature of models imposed by PLS path models.
- If heterogeneity affects non-recursive relationships, it might have a strong impact on the ability of both PLS segmentation methods (FIMIX-PLS and PLS-POS) to uncover unobserved heterogeneity.
- Third, this research does not focus on the parameter settings of the methods or the time needed to arrive at the final segmentation solution.
- The authors simulations suggest that PLS-POS is more time consuming than FIMIX-PLS.
- In absolute terms, PLS-POS works within acceptable timeframes.
Conclusion
- The authors differentiated between observed and unobserved heterogeneity and showed why unobserved heterogeneity biases structural equation model estimates, leads to Type I and Type II errors, and is a threat to different types of validity (i.e., internal, instrumental, statistical conclusion, and external).
- The authors findings also reveal that unobserved heterogeneity in formative measures and in the structural model should be evaluated collectively.
- As FIMIX-PLS does not uncover heterogeneity in measurement models, PLS-POS should be applied for discovering unobserved heterogeneity if PLS path models include formative measures.
- Moreover, in situations in which the researcher discovers anomalies that must be resolved through theoretical elaboration, theory development is significantly enhanced by abduction.
- Using the presented methods in PLS path modeling and CB-SEM within the UHD process is a possible way to achieve this goal.
Did you find this useful? Give us your feedback
Citations
6,220 citations
5,191 citations
3,251 citations
1,906 citations
1,842 citations
References
272,030 citations
27,798 citations
21,880 citations
21,693 citations
16,513 citations
Related Papers (5)
Frequently Asked Questions (9)
Q2. What is the advantage of applying the parametric finite mixture regression concept to PLS path models?
The advantage of applying the parametric finite mixture regression concept to PLS path models is that it offers segment retention criteria (e.g., AIC, BIC, and CAIC; Hahn et al.
Q3. What are the main reasons why finite mixture models are being used in empirical CB-S?
An increasing number of applications, especially in the marketing, econometrics, and sociology literatures, have utilized finite mixture models to uncover unobserved heterogeneity, thereby improving theoretical and practical implications (e.g., Bart et al.
Q4. What methods are used to uncover unobserved heterogeneity in the data?
finite mixture models address unobserved heterogeneity in the data by grouping observations and estimating group-specific 3There are several methods to uncover both observed and unobserved heterogeneity in other methodological contexts—for example, regression analysis (DeSarbo and Cron 1988; Späth 1979; Wedel and DeSarbo1994), panel data models (Allenby and Rossi 1998; Popkowski Leszczyc and Bass 1998), and conjoint analysis (DeSarbo et al.
Q5. What is the effect of the measurement model heterogeneity on MAB?
In contrast, parameter recovery for FIMIX-PLS decreases with decreasing structural model heterogeneity or increasing measurement model heterogeneity.
Q6. What is the effect of unobserved heterogeneity on the parameter estimates?
The results show that unobserved heterogeneity biases the parameter estimates, decreases the R², and increases the risk of Type The authorand Type II errors.
Q7. What are the examples of moderators or contextual factors considered in IS research?
Examples of such moderators or contextual factors considered in IS research include individual cultural differences (e.g., individualism versus collectivism; Srite and Karahanna 2006), individual demographic differences (e.g., gender, income levels, and education; Hsieh et al.
Q8. What is the reason why the results of one study are not generalized to other studies?
The observation of inconsistent, conflicting, or mixed findings in the 12 meta-analyses in Table A1 (Appendix A) also show that the results of one study often cannot be generalized to other studies (indicating low external validity) with unobserved heterogeneity being one of the plausible reasons.
Q9. What is the main difference between formative and reflective measures?
While recent research has discussed ME/I in formative measures (Diamantopoulos and Papadopoulos 2010), it is important to uncover formative indicator weight differences due to unobserved heterogeneity in order to avoid ambiguous interpretations.