Week 2 Report

International Journal of Psychology International Journal of Psychology, 2019 Vol. 54, No. 1, 88–92, DOI: 10.1002/ijop.12442

Opinion leadership types or continuous opinion leadership traits?

Timo Gnambs

Educational Measurement, Leibniz-Institute for Educational Trajectories, Bamberg, Germany

O pinion leadership is typically conceptualised as a continuous personality trait. However, many authors adhere tothe view of qualitatively different opinion leadership types and apply arbitrary criteria to split continuous trait scores into two groups (i.e., opinion leaders vs. non-leaders). The present study is the first to empirically evaluate this approach. A sample of N = 3812 adults (67% women) was administered a validated opinion leadership scale. Finite mixture models examined whether the latent trait distribution can be represented by a set of discrete trait levels that reflected distinct opinion leadership types. The results did not give support to a discrete typology that distinguished leaders from non-leaders. Rather, opinion leadership was best characterised as a continuous trait.

Keywords: Personality type; Social influence; Mixture modelling; Non-parametric factor analysis.

Individual differences in social influence determine the degree to which people can shape attitudes, decisions and overt behaviours of their friends, family members and co-workers. A trait reflecting the ability to informally influence others and thereby promote the diffusion of new ideas and trends in a social group is opinion leadership (cf. Batinic, Appel, & Gnambs, 2016; Flynn, Goldsmith, & Eastman, 1996). Opinion leadership is a central con- cept in such diverse fields as marketing, political research or health communication that has attracted worldwide interest (see Weimann, Tustin, van Vuuren, & Joubert, 2007). Recently, a fresh impetus to opinion leadership research around the world has resulted from the increas- ing popularity of social media and lead to numerous studies highlighting peer influences on discussion boards or social networking sites (e.g., Weeks, Ardèvol-Abreu, & de Zúñiga, 2017). Even international public opinion surveys such as the representative Eurobarometer1 sur- veys conducted each month in the European member states routinely include measures to stratify attitudes on current topics by levels of opinion leadership. Although many of these studies administered different instruments that varied with regard to the precise construct defini- tions and the breadth of the operationalised constructs (Trepte & Scherer, 2010), the scales shared a common focus and described individuals that informally influence

Correspondence should be addressed to Timo Gnambs, Leibniz-Institute for Educational Trajectories, Wilhelmsplatz 3, 96047 Bamberg, Germany. (E-mail: [email protected]).

1http://ec.europa.eu/COMMFrontOffice/publicopinion/

their social peer group. Moreover, in line with typical personality traits all these scales conceptualised opinion leadership as a continuous trait.

In contrast, some authors adhered to the idea of qual- itatively different opinion leadership types that distin- guished opinion leaders from non-leaders (e.g., Chan & Misra, 1990; Goldsmith & Flynn, 1994; Lyons & Henderson, 2005; Vernette, 2004). Thus, they used the instruments constructed for measuring continuous opin- ion leadership traits and, subsequently, applied some arbitrary criterion to split the sample into two artifi- cial groups that supposedly distinguished opinion lead- ers from non-leaders. Some authors used the sample’s mean or the theoretical mean based on the response scale to divide the sample into two groups (Chan & Misra, 1990). Others split their sample in such a way as to place a certain percentage of their sample into the group of opinion leaders (Goldsmith & Flynn, 1994; Lyons & Henderson, 2005; Vernette, 2004). They argued on the- oretical accounts that only a minority of the population (e.g., 10%) is expected to act as opinion leader and, thus, should be viewed as such in a given sample. Common to these studies is that they created two artificial groups based on some arbitrary criterion without any psychome- tric evaluation if, indeed, the sample at hand represented a mixture of two distinct subpopulations.

© 2017 International Union of Psychological Science

OPINION LEADERSHIP TYPES 89

As highlighted by several methodological studies (e.g., Bissonnette, Ickes, Bernstein, & Knowles, 1990; Rucker, McShane, & Preacher, 2015) these post-hoc classifica- tions can introduce a non-negligible bias into research findings. Discretising continuous variables to create arti- ficial groups not only results in a loss of power to identify meaningful effects but more seriously can also increase Type I errors by identifying spurious effects that do not actually exist. Therefore, it is important to empiri- cally evaluate whether distinct subpopulations of respon- dents can be identified before creating opinion leadership groups. To address this research gap, the present study examined the responses from a large sample of adults on a validated opinion leadership scale. Finite mixture mod- els (Hallquist & Wright, 2014) were estimated to identify potential subgroups of respondents that might reflect the opinion leadership dichotomy of leader versus non-leader. Moreover, age differences between different opinion lead- ership types were studied for different scoring schemes to highlight the danger of drawing misleading conclusions depending on the way the opinion leadership groups were created.

METHOD

Participants

The respondents were members of a commercial mar- ket research panel from Germany who repeatedly partic- ipate at anonymous web-based surveys in exchange for minor incentives (e.g., gift coupons). The present sample included 1212 men, 2496 women and 104 individuals that did not report their gender. They were between 14 and 85 years old (M = 30.97, SD= 11.88). Most respondents had an educational level equivalent to university entrance qualifications (45%) or had already obtained a university degree (24%).

Instruments

Generalised opinion leadership was measured with nine items (e.g., “I usually succeed if I want to convince some- one about something.”) on 5-point response scales from 1 (do not agree at all) to 5 (agree completely) (Gnambs & Batinic, 2011a). Previous research demonstrated good psychometric properties of this instrument including a unidimensional factor structure, high test–retest reliabil- ity and good construct validity (e.g., Batinic et al., 2016; Gnambs & Batinic, 2011b, 2012). In the present sample, the scale score (M = 2.93, SD= 0.57) had coefficient α and ω hierarchical reliabilities of .84 and .71, respectively.

Statistical analyses

The administered items were scaled using a graded response model (Samejima, 1969) that estimated a

single latent factor representing the opinion leadership trait. Subgroups of respondents that might reflect dif- ferent opinion leadership types were identified using a non-parametric factor approach (see Hallquist & Wright, 2014). Thus, finite mixture models with two or more latent classes were specified that estimated the probability of belonging to a given latent class for each respondent. Across the different classes strict measurement invariance of the latent factor was enforced. Moreover, the factor variances in each class were fixed to zero. In this way, the latent trait distribution can be represented by a set of dis- crete levels along the trait continuum with homogenous respondents within each class that reflect distinct opinion leadership types. Subsequently, the assumption of homo- geneity within classes was relaxed by estimating different variances within class. This resulted in a semi-parametric factor model that allows for a potential non-normal trait distribution (see Hallquist & Wright, 2014). The number of subgroups (i.e., latent classes) were identified by a stepwise procedure comparing models with different numbers of classes based on the Bayesian information criterion (BIC) where lower values indicate models that more closely approximate the empirical data. Moreover, adjusted likelihood ratio tests (Lo, Mendel, & Rubin, 2001) were used to compare models with a given number of classes to one with one class less; statistical significant results would indicate more support for the model with more classes over the model with fewer classes. Finally, the sample size of each class was used to guide decisions on the practical relevance of a given class solution. A two-class solution with higher opinion leadership scores in the smaller class would support the assumption of qualitatively distinct opinion leadership types (i.e., leaders vs. non-leaders), whereas multiple classes with approximately linear increasing mean opinion leadership scores would rather fall in line with a continuous trait representation. All mixture models were estimated in Mplus 7 (Muthén & Muthén, 1998–2012) with a robust maximum likelihood estimator.

RESULTS

The factor score distribution of the opinion leadership trait for the entire sample without specifying any latent classes (see Figure 1) exhibited an approximately normal shape and revealed no evidence of multiple local max- ima that might indicate two or more qualitatively different respondent types. In the next step, various non-parametric factor models were estimated that specified between two and eight latent classes. The respective model fit indices are summarised in Table 1. The model with two classes exhibited an inferior fit as indicated by the significant (p< .05) likelihood-ratio test and the large BIC as com- pared to models with more latent classes. Thus, there was no support for two qualitatively distinct opinion

© 2017 International Union of Psychological Science

90 GNAMBS

Figure 1. Factor score distribution of opinion leadership (left panel) and mean factor scores with proportions of sample size in latent classes (right panel).

TABLE 1 Fit indices for various non-parametric factor models

Model logLik Number of parameters BIC LRT-p

Smallest class proportion

0 classa −37,524.3 37 75,353.0 — — 1 class −41,996.2 44 84,354.6 — — 2 classes −39,012.2 46 78,402.9 <.001 .50 3 classes −38,075.1 48 76,545.2 <.001 .21 4 classes −37,696.2 50 75,803.8 .29 .04 5 classes −37,471.9 52 75,371.7 <.001 .02 6 classes −37,379.0 54 75,202.4 <.001 <.01 7 classes −37,348.9 56 75,158.6 .05 <.01 8 classes −37,341.7 58 75,160.7 .01 <.01

Note: BIC=Bayesian information criterion; logLik= logarithm of the model likelihood; LRT-p= p-value associated with the adjusted likelihood ratio test (Lo et al., 2001). aFactor model with no latent classes and estimated latent factor variance.

leadership types. To examine whether the respondents might be grouped into three or more discrete levels along the trait continuum, I also examined the three to eight class models. However, the different model fit indices did not converge to a common solution. The likelihood ratio test identified no superior fit of the four-class model as compared to the three-class model (p= .29). How- ever, the BIC of the latter was rather large as compared to models with more latent classes and, moreover, indi- cated a worse fit than the initial model without any latent classes (ΔBIC= 1192.2). The best fit in terms of the BIC was achieved by a model with seven classes. How- ever, in this model many classes were rather small; only three classes had class proportions exceeding 5 %. More importantly, the latent factor means in the seven latent classes exhibited an approximately linear increase of the opinion leadership trait (see Figure 1). Thus, the assump- tion that opinion leadership would be better represented by qualitatively distinct types rather than a continuous trait yielded no support. Finally, the previous analyses were replicated by relaxing the homogeneity assumption

within classes and estimating different variances for each class. Fit indices for one to three class models favoured the two class solution with BIC1 = 75,219.5, BIC2 = 75,088.1 and BIC3 = 75,101.6, respectively. However, the second class was quite small (1.4% of the sample) and, in con- trast to opinion leadership theory, had a smaller mean (M =−0.33, SD= 5.63) than the larger class (M = 0.00, SD= 1.18). Again, the analyses did not support the assumption of a distinct group characterised by particu- larly high levels of opinion leadership.

To demonstrate the consequences of using arbitrary cutoffs for the derivation of opinion leadership types, three classification schemes were adopted. In line with prevalent practice (see Vernette, 2004), the top 5, 10 or 15% scorers on the opinion leadership scale were classified as opinion leaders, whereas the remaining samples were considered non-leaders. Given that opin- ion leadership is subject to pronounced age differences (Batinic et al., 2016), the standardised mean difference in the respondents’ age was calculated for the three scor- ing schemes. If only the top 5% of the respondents were

© 2017 International Union of Psychological Science

OPINION LEADERSHIP TYPES 91

considered as opinion leaders, the age difference between the two groups would amount to Cohen’s d = .20, 95% CI [.05, .34]. In contrast, using either the top 10 or 15% scorers as opinion leaders would lead to age differ- ences of d = .09, 95% CI [.01, .19] and d =−.02, 95% CI [−.06, .10], respectively. Thus, depending on the arbitrar- ily chosen classification scheme researchers would draw different conclusions regarding age differences between opinion leaders and non-leaders.

DISCUSSION

Contemporary views on opinion leadership consider the concept as a continuous trait of inter-individual differ- ences in social influence (cf. Batinic et al., 2016; Flynn et al., 1996; Weimann et al., 2007). In practice, how- ever, many researchers operate with an implicit typol- ogy and try to classify their respondents into two distinct groups that distinguish opinion leaders from non-leaders. Thus, they try to identify subgroups of particularly influ- ential individuals within a sample. Despite the intuitive appeal of this approach (e.g., allowing for an easy com- munication of group comparisons on key variables to the general public), it is unknown whether empirical data actually reflects distinct opinion leadership types. There- fore, the present study applied finite mixture models to identify homogenous subgroups of respondents that fall in line with the hypothesised opinion leadership dichotomy. However, despite being based on a large sample includ- ing over 3000 respondents and using a validated instru- ment these analyses found no support for discrete opinion leadership types. Rather, opinion leadership was best con- ceptualised as a continuous trait. These findings also align with several previous taxometric analyses showing that individual differences in personality are typically contin- uous rather than categorical (e.g., Foster & Campbell, 2007; Marcus, Lilienfeld, Edens, & Poythress, 2006). Therefore, previous attempts deriving two groups reflect- ing leaders and non-leaders should be viewed with due caution because there is little evidence that these sub- groups actually exist. Moreover, given that discretising continuous scores can introduce substantial biases into statistical results (e.g., Bissonnette et al., 1990; Rucker et al., 2015), the practice of deriving post-hoc opinion leadership groups without appropriate psychometric eval- uations should be abandoned.

In conclusion, the study failed to provide support for qualitatively distinct opinion leadership types but sug- gested a continuous opinion leadership trait. Future stud- ies are encouraged to extend this line of research to related instruments measuring other variants of opinion leader- ship (Flynn et al., 1996; Weimann et al., 2007). However, it is expected that the reported findings will generalise to many of these scales because they exhibit strong conver- gent validities with the administered instrument (Gnambs

& Batinic, 2011b). Importantly, researchers adhering to the concept of different opinion leadership types are well advised to apply appropriate psychometric models (see Hallquist & Wright, 2014) to identify the hypothesised types before applying arbitrary criteria to create artificial groups for which empirical support may not be available. Particularly, international, cross-cultural studies need to demonstrate that comparable opinion leadership types exist in all samples before deriving conclusions on the determinants and consequences of opinion leadership in different countries.

Manuscript received January 2017 Revised manuscript accepted June 2017

First published online July 2017

REFERENCES

Batinic, B., Appel, M., & Gnambs, T. (2016). Examining individual differences in interpersonal influence: On the psychometric properties of the Generalized Opinion Lead- ership Scale (GOLS). Journal of Psychology: Interdisci- plinary and Applied, 150, 88–101. https://doi.org/10.1080/ 00223980.2015.1009415.

Bissonnette, V., Ickes, W., Bernstein, I., & Knowles, E. (1990). Personality moderating variables: A warning about statistical artifact and a comparison of analytic techniques. Journal of Personality, 58, 567–587. https://doi.org/10.1111/j.1467- 6494.1990.tb00243.x.

Chan, K. K., & Misra, S. (1990). Characteristics of the opinion leader: A new dimension. Journal of Advertising, 19, 53–60. https://doi.org/10.1080/00913367.1990.10673192.

Flynn, L. R., Goldsmith, R. E., & Eastman, J. K. (1996). Opin- ion leadership and opinion seekers: Two new measurement scales. Journal of the Academy of Marketing Science, 24, 137–147. https://doi.org/10.1177/0092070396242004.

Foster, J. D., & Campbell, W. K. (2007). Are there such things as “narcissists” in social psychology? A taxometric analysis of the Narcissistic Personality Inventory. Personality and Individual Differences, 43, 1321–1332. https://doi.org/10 .1016/j.paid.2007.04.003.

Gnambs, T., & Batinic, B. (2011a). Evaluation of measurement precision with Rasch-type models: The case of the short Generalized Opinion Leadership Scale. Personality and Indi- vidual Differences, 50, 53–58. https://doi.org/10.1016/j.paid .2010.08.021.

Gnambs, T., & Batinic, B. (2011b). Convergent and discriminant validity of opinion leadership: Multitrait-multimethod anal- ysis across measurement occasion and informant type. Jour- nal of Individual Differences, 32, 94–102. https://doi.org/10 .1027/1614-0001/a000040.

Gnambs, T., & Batinic, B. (2012). A personality-competence model of opinion leadership. Psychology and Marketing, 29, 606–621. https://doi.org/10.1002/mar.20547.

Goldsmith, R. E., & Flynn, L. (1994). Opinion leadership for vacation travel services. Advances in Business Studies, 4, 281–284.

Hallquist, M. N., & Wright, A. G. C. (2014). Mixture mod- eling methods for the assessment of normal and abnormal

© 2017 International Union of Psychological Science

92 GNAMBS

personality I: Cross-sectional models. Journal of Personality Assessment, 96, 256–268. https://doi.org/10.1080/00223891 .2013.845201.

Lo, Y., Mendel, N., & Rubin, D. B. (2001). Testing the num- ber of components in a normal mixture. Biometrika, 88, 767–778. https://doi.org/10.1093/biomet/88.3.767.

Lyons, B., & Henderson, K. (2005). Opinion leadership in computer-mediated environments. Journal of Consumer Behaviour, 4, 319–329. https://doi.org/10.1002/cb.22.

Marcus, D. K., Lilienfeld, S. O., Edens, J. F., & Poythress, N. G. (2006). Is antisocial personality disorder continuous or categorical? A taxometric analysis. Psychological Medicine, 36, 1571–1581. https://doi.org/10.1017/ S0033291706008245.

Muthén, L. K., & Muthén, B. O. (1998–2012). Mplus user’s guide (7th ed.). Los Angeles, CA: Muthén & Muthén.

Rucker, D. D., McShane, B. B., & Preacher, K. J. (2015). A researcher’s guide to regression, discretization, and median splits of continuous variables. Journal of Consumer Psy- chology, 25, 666–678. https://doi.org/10.1016/j.jcps.2015 .04.004.

Samejima, F. (1969). Estimation of a latent ability using a response pattern of graded scores. Psychometrika, 34(Suppl. 1), 1–97. https://doi.org/10.1007/BF03372160.

Trepte, S., & Scherer, H. (2010). Opinion leaders – Do they know more than others about their area of interest? Commu- nications, 35, 119–140. https://doi.org/10.1515/comm.2010 .007.

Vernette, É. (2004). Targeting women’s clothing fashion opinion leaders in media planning: An application of magazines. Journal of Advertising Research, 44, 90–107. https://doi.org/ 10.1017/S0021849904040061.

Weeks, B. E., Ardèvol-Abreu, A., & de Zúñiga, H. G. (2017). Online influence? Social media use, opinion leadership, and political persuasion. International Journal of Public Opin- ion Research, 29, 214–239. https://doi.org/10.1093/ijpor/ edv050.

Weimann, G., Tustin, D. H., van Vuuren, D., & Joubert, J. P. R. (2007). Looking for opinion leaders: Traditional vs. modern measures in traditional societies. International Journal of Public Opinion Research, 19, 173–190. https://doi.org/10 .1093/ijpor/edm005.

© 2017 International Union of Psychological Science

Copyright of International Journal of Psychology is the property of John Wiley & Sons, Inc. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.